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/dma-direct.h>
32 #include <linux/iommu-helper.h>
33 #include <linux/iommu.h>
34 #include <linux/delay.h>
35 #include <linux/amd-iommu.h>
36 #include <linux/notifier.h>
37 #include <linux/export.h>
38 #include <linux/irq.h>
39 #include <linux/msi.h>
40 #include <linux/dma-contiguous.h>
41 #include <linux/irqdomain.h>
42 #include <linux/percpu.h>
43 #include <linux/iova.h>
44 #include <asm/irq_remapping.h>
45 #include <asm/io_apic.h>
47 #include <asm/hw_irq.h>
48 #include <asm/msidef.h>
49 #include <asm/proto.h>
50 #include <asm/iommu.h>
54 #include "amd_iommu_proto.h"
55 #include "amd_iommu_types.h"
56 #include "irq_remapping.h"
58 #define AMD_IOMMU_MAPPING_ERROR 0
60 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
62 #define LOOP_TIMEOUT 100000
64 /* IO virtual address start page frame number */
65 #define IOVA_START_PFN (1)
66 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
68 /* Reserved IOVA ranges */
69 #define MSI_RANGE_START (0xfee00000)
70 #define MSI_RANGE_END (0xfeefffff)
71 #define HT_RANGE_START (0xfd00000000ULL)
72 #define HT_RANGE_END (0xffffffffffULL)
75 * This bitmap is used to advertise the page sizes our hardware support
76 * to the IOMMU core, which will then use this information to split
77 * physically contiguous memory regions it is mapping into page sizes
80 * 512GB Pages are not supported due to a hardware bug
82 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
84 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
86 /* List of all available dev_data structures */
87 static LIST_HEAD(dev_data_list);
88 static DEFINE_SPINLOCK(dev_data_list_lock);
90 LIST_HEAD(ioapic_map);
92 LIST_HEAD(acpihid_map);
95 * Domain for untranslated devices - only allocated
96 * if iommu=pt passed on kernel cmd line.
98 const struct iommu_ops amd_iommu_ops;
100 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
101 int amd_iommu_max_glx_val = -1;
103 static const struct dma_map_ops amd_iommu_dma_ops;
106 * general struct to manage commands send to an IOMMU
112 struct kmem_cache *amd_iommu_irq_cache;
114 static void update_domain(struct protection_domain *domain);
115 static int protection_domain_init(struct protection_domain *domain);
116 static void detach_device(struct device *dev);
117 static void iova_domain_flush_tlb(struct iova_domain *iovad);
120 * Data container for a dma_ops specific protection domain
122 struct dma_ops_domain {
123 /* generic protection domain information */
124 struct protection_domain domain;
127 struct iova_domain iovad;
130 static struct iova_domain reserved_iova_ranges;
131 static struct lock_class_key reserved_rbtree_key;
133 /****************************************************************************
137 ****************************************************************************/
139 static inline int match_hid_uid(struct device *dev,
140 struct acpihid_map_entry *entry)
142 const char *hid, *uid;
144 hid = acpi_device_hid(ACPI_COMPANION(dev));
145 uid = acpi_device_uid(ACPI_COMPANION(dev));
151 return strcmp(hid, entry->hid);
154 return strcmp(hid, entry->hid);
156 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
159 static inline u16 get_pci_device_id(struct device *dev)
161 struct pci_dev *pdev = to_pci_dev(dev);
163 return PCI_DEVID(pdev->bus->number, pdev->devfn);
166 static inline int get_acpihid_device_id(struct device *dev,
167 struct acpihid_map_entry **entry)
169 struct acpihid_map_entry *p;
171 list_for_each_entry(p, &acpihid_map, list) {
172 if (!match_hid_uid(dev, p)) {
181 static inline int get_device_id(struct device *dev)
186 devid = get_pci_device_id(dev);
188 devid = get_acpihid_device_id(dev, NULL);
193 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
195 return container_of(dom, struct protection_domain, domain);
198 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
200 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
201 return container_of(domain, struct dma_ops_domain, domain);
204 static struct iommu_dev_data *alloc_dev_data(u16 devid)
206 struct iommu_dev_data *dev_data;
209 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
213 dev_data->devid = devid;
215 spin_lock_irqsave(&dev_data_list_lock, flags);
216 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
217 spin_unlock_irqrestore(&dev_data_list_lock, flags);
219 ratelimit_default_init(&dev_data->rs);
224 static struct iommu_dev_data *search_dev_data(u16 devid)
226 struct iommu_dev_data *dev_data;
229 spin_lock_irqsave(&dev_data_list_lock, flags);
230 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
231 if (dev_data->devid == devid)
238 spin_unlock_irqrestore(&dev_data_list_lock, flags);
243 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
245 *(u16 *)data = alias;
249 static u16 get_alias(struct device *dev)
251 struct pci_dev *pdev = to_pci_dev(dev);
252 u16 devid, ivrs_alias, pci_alias;
254 /* The callers make sure that get_device_id() does not fail here */
255 devid = get_device_id(dev);
256 ivrs_alias = amd_iommu_alias_table[devid];
257 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
259 if (ivrs_alias == pci_alias)
265 * The IVRS is fairly reliable in telling us about aliases, but it
266 * can't know about every screwy device. If we don't have an IVRS
267 * reported alias, use the PCI reported alias. In that case we may
268 * still need to initialize the rlookup and dev_table entries if the
269 * alias is to a non-existent device.
271 if (ivrs_alias == devid) {
272 if (!amd_iommu_rlookup_table[pci_alias]) {
273 amd_iommu_rlookup_table[pci_alias] =
274 amd_iommu_rlookup_table[devid];
275 memcpy(amd_iommu_dev_table[pci_alias].data,
276 amd_iommu_dev_table[devid].data,
277 sizeof(amd_iommu_dev_table[pci_alias].data));
283 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
284 "for device %s[%04x:%04x], kernel reported alias "
285 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
286 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
287 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
288 PCI_FUNC(pci_alias));
291 * If we don't have a PCI DMA alias and the IVRS alias is on the same
292 * bus, then the IVRS table may know about a quirk that we don't.
294 if (pci_alias == devid &&
295 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
296 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
297 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
298 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
305 static struct iommu_dev_data *find_dev_data(u16 devid)
307 struct iommu_dev_data *dev_data;
308 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
310 dev_data = search_dev_data(devid);
312 if (dev_data == NULL) {
313 dev_data = alloc_dev_data(devid);
315 if (translation_pre_enabled(iommu))
316 dev_data->defer_attach = true;
322 struct iommu_dev_data *get_dev_data(struct device *dev)
324 return dev->archdata.iommu;
326 EXPORT_SYMBOL(get_dev_data);
329 * Find or create an IOMMU group for a acpihid device.
331 static struct iommu_group *acpihid_device_group(struct device *dev)
333 struct acpihid_map_entry *p, *entry = NULL;
336 devid = get_acpihid_device_id(dev, &entry);
338 return ERR_PTR(devid);
340 list_for_each_entry(p, &acpihid_map, list) {
341 if ((devid == p->devid) && p->group)
342 entry->group = p->group;
346 entry->group = generic_device_group(dev);
348 iommu_group_ref_get(entry->group);
353 static bool pci_iommuv2_capable(struct pci_dev *pdev)
355 static const int caps[] = {
358 PCI_EXT_CAP_ID_PASID,
362 for (i = 0; i < 3; ++i) {
363 pos = pci_find_ext_capability(pdev, caps[i]);
371 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
373 struct iommu_dev_data *dev_data;
375 dev_data = get_dev_data(&pdev->dev);
377 return dev_data->errata & (1 << erratum) ? true : false;
381 * This function checks if the driver got a valid device from the caller to
382 * avoid dereferencing invalid pointers.
384 static bool check_device(struct device *dev)
388 if (!dev || !dev->dma_mask)
391 devid = get_device_id(dev);
395 /* Out of our scope? */
396 if (devid > amd_iommu_last_bdf)
399 if (amd_iommu_rlookup_table[devid] == NULL)
405 static void init_iommu_group(struct device *dev)
407 struct iommu_group *group;
409 group = iommu_group_get_for_dev(dev);
413 iommu_group_put(group);
416 static int iommu_init_device(struct device *dev)
418 struct iommu_dev_data *dev_data;
419 struct amd_iommu *iommu;
422 if (dev->archdata.iommu)
425 devid = get_device_id(dev);
429 iommu = amd_iommu_rlookup_table[devid];
431 dev_data = find_dev_data(devid);
435 dev_data->alias = get_alias(dev);
437 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
438 struct amd_iommu *iommu;
440 iommu = amd_iommu_rlookup_table[dev_data->devid];
441 dev_data->iommu_v2 = iommu->is_iommu_v2;
444 dev->archdata.iommu = dev_data;
446 iommu_device_link(&iommu->iommu, dev);
451 static void iommu_ignore_device(struct device *dev)
456 devid = get_device_id(dev);
460 alias = get_alias(dev);
462 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
463 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
465 amd_iommu_rlookup_table[devid] = NULL;
466 amd_iommu_rlookup_table[alias] = NULL;
469 static void iommu_uninit_device(struct device *dev)
471 struct iommu_dev_data *dev_data;
472 struct amd_iommu *iommu;
475 devid = get_device_id(dev);
479 iommu = amd_iommu_rlookup_table[devid];
481 dev_data = search_dev_data(devid);
485 if (dev_data->domain)
488 iommu_device_unlink(&iommu->iommu, dev);
490 iommu_group_remove_device(dev);
496 * We keep dev_data around for unplugged devices and reuse it when the
497 * device is re-plugged - not doing so would introduce a ton of races.
501 /****************************************************************************
503 * Interrupt handling functions
505 ****************************************************************************/
507 static void dump_dte_entry(u16 devid)
511 for (i = 0; i < 4; ++i)
512 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
513 amd_iommu_dev_table[devid].data[i]);
516 static void dump_command(unsigned long phys_addr)
518 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
521 for (i = 0; i < 4; ++i)
522 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
525 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
526 u64 address, int flags)
528 struct iommu_dev_data *dev_data = NULL;
529 struct pci_dev *pdev;
531 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
534 dev_data = get_dev_data(&pdev->dev);
536 if (dev_data && __ratelimit(&dev_data->rs)) {
537 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
538 domain_id, address, flags);
539 } else if (printk_ratelimit()) {
540 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
541 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
542 domain_id, address, flags);
549 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
551 int type, devid, domid, flags;
552 volatile u32 *event = __evt;
557 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
558 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
559 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
560 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
561 address = (u64)(((u64)event[3]) << 32) | event[2];
564 /* Did we hit the erratum? */
565 if (++count == LOOP_TIMEOUT) {
566 pr_err("AMD-Vi: No event written to event log\n");
573 if (type == EVENT_TYPE_IO_FAULT) {
574 amd_iommu_report_page_fault(devid, domid, address, flags);
577 printk(KERN_ERR "AMD-Vi: Event logged [");
581 case EVENT_TYPE_ILL_DEV:
582 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
583 "address=0x%016llx flags=0x%04x]\n",
584 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 dump_dte_entry(devid);
588 case EVENT_TYPE_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 event[0]=0x%08x "
622 "event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
623 type, event[0], event[1], event[2], event[3]);
626 memset(__evt, 0, 4 * sizeof(u32));
629 static void iommu_poll_events(struct amd_iommu *iommu)
633 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
634 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
636 while (head != tail) {
637 iommu_print_event(iommu, iommu->evt_buf + head);
638 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
641 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
644 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
646 struct amd_iommu_fault fault;
648 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
649 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
653 fault.address = raw[1];
654 fault.pasid = PPR_PASID(raw[0]);
655 fault.device_id = PPR_DEVID(raw[0]);
656 fault.tag = PPR_TAG(raw[0]);
657 fault.flags = PPR_FLAGS(raw[0]);
659 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
662 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
666 if (iommu->ppr_log == NULL)
669 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
670 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
672 while (head != tail) {
677 raw = (u64 *)(iommu->ppr_log + head);
680 * Hardware bug: Interrupt may arrive before the entry is
681 * written to memory. If this happens we need to wait for the
684 for (i = 0; i < LOOP_TIMEOUT; ++i) {
685 if (PPR_REQ_TYPE(raw[0]) != 0)
690 /* Avoid memcpy function-call overhead */
695 * To detect the hardware bug we need to clear the entry
698 raw[0] = raw[1] = 0UL;
700 /* Update head pointer of hardware ring-buffer */
701 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
702 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
704 /* Handle PPR entry */
705 iommu_handle_ppr_entry(iommu, entry);
707 /* Refresh ring-buffer information */
708 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
709 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
713 #ifdef CONFIG_IRQ_REMAP
714 static int (*iommu_ga_log_notifier)(u32);
716 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
718 iommu_ga_log_notifier = notifier;
722 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
724 static void iommu_poll_ga_log(struct amd_iommu *iommu)
726 u32 head, tail, cnt = 0;
728 if (iommu->ga_log == NULL)
731 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
732 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
734 while (head != tail) {
738 raw = (u64 *)(iommu->ga_log + head);
741 /* Avoid memcpy function-call overhead */
744 /* Update head pointer of hardware ring-buffer */
745 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
746 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
748 /* Handle GA entry */
749 switch (GA_REQ_TYPE(log_entry)) {
751 if (!iommu_ga_log_notifier)
754 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
755 __func__, GA_DEVID(log_entry),
758 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
759 pr_err("AMD-Vi: GA log notifier failed.\n");
766 #endif /* CONFIG_IRQ_REMAP */
768 #define AMD_IOMMU_INT_MASK \
769 (MMIO_STATUS_EVT_INT_MASK | \
770 MMIO_STATUS_PPR_INT_MASK | \
771 MMIO_STATUS_GALOG_INT_MASK)
773 irqreturn_t amd_iommu_int_thread(int irq, void *data)
775 struct amd_iommu *iommu = (struct amd_iommu *) data;
776 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
778 while (status & AMD_IOMMU_INT_MASK) {
779 /* Enable EVT and PPR and GA interrupts again */
780 writel(AMD_IOMMU_INT_MASK,
781 iommu->mmio_base + MMIO_STATUS_OFFSET);
783 if (status & MMIO_STATUS_EVT_INT_MASK) {
784 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
785 iommu_poll_events(iommu);
788 if (status & MMIO_STATUS_PPR_INT_MASK) {
789 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
790 iommu_poll_ppr_log(iommu);
793 #ifdef CONFIG_IRQ_REMAP
794 if (status & MMIO_STATUS_GALOG_INT_MASK) {
795 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
796 iommu_poll_ga_log(iommu);
801 * Hardware bug: ERBT1312
802 * When re-enabling interrupt (by writing 1
803 * to clear the bit), the hardware might also try to set
804 * the interrupt bit in the event status register.
805 * In this scenario, the bit will be set, and disable
806 * subsequent interrupts.
808 * Workaround: The IOMMU driver should read back the
809 * status register and check if the interrupt bits are cleared.
810 * If not, driver will need to go through the interrupt handler
811 * again and re-clear the bits
813 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
818 irqreturn_t amd_iommu_int_handler(int irq, void *data)
820 return IRQ_WAKE_THREAD;
823 /****************************************************************************
825 * IOMMU command queuing functions
827 ****************************************************************************/
829 static int wait_on_sem(volatile u64 *sem)
833 while (*sem == 0 && i < LOOP_TIMEOUT) {
838 if (i == LOOP_TIMEOUT) {
839 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
846 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
847 struct iommu_cmd *cmd)
851 target = iommu->cmd_buf + iommu->cmd_buf_tail;
853 iommu->cmd_buf_tail += sizeof(*cmd);
854 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
856 /* Copy command to buffer */
857 memcpy(target, cmd, sizeof(*cmd));
859 /* Tell the IOMMU about it */
860 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
863 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
865 u64 paddr = iommu_virt_to_phys((void *)address);
867 WARN_ON(address & 0x7ULL);
869 memset(cmd, 0, sizeof(*cmd));
870 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
871 cmd->data[1] = upper_32_bits(paddr);
873 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
876 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
878 memset(cmd, 0, sizeof(*cmd));
879 cmd->data[0] = devid;
880 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
883 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
884 size_t size, u16 domid, int pde)
889 pages = iommu_num_pages(address, size, PAGE_SIZE);
894 * If we have to flush more than one page, flush all
895 * TLB entries for this domain
897 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
901 address &= PAGE_MASK;
903 memset(cmd, 0, sizeof(*cmd));
904 cmd->data[1] |= domid;
905 cmd->data[2] = lower_32_bits(address);
906 cmd->data[3] = upper_32_bits(address);
907 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
908 if (s) /* size bit - we flush more than one 4kb page */
909 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
910 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
911 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
914 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
915 u64 address, size_t size)
920 pages = iommu_num_pages(address, size, PAGE_SIZE);
925 * If we have to flush more than one page, flush all
926 * TLB entries for this domain
928 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
932 address &= PAGE_MASK;
934 memset(cmd, 0, sizeof(*cmd));
935 cmd->data[0] = devid;
936 cmd->data[0] |= (qdep & 0xff) << 24;
937 cmd->data[1] = devid;
938 cmd->data[2] = lower_32_bits(address);
939 cmd->data[3] = upper_32_bits(address);
940 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
942 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
945 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
946 u64 address, bool size)
948 memset(cmd, 0, sizeof(*cmd));
950 address &= ~(0xfffULL);
952 cmd->data[0] = pasid;
953 cmd->data[1] = domid;
954 cmd->data[2] = lower_32_bits(address);
955 cmd->data[3] = upper_32_bits(address);
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
957 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
959 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
960 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
963 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
964 int qdep, u64 address, bool size)
966 memset(cmd, 0, sizeof(*cmd));
968 address &= ~(0xfffULL);
970 cmd->data[0] = devid;
971 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
972 cmd->data[0] |= (qdep & 0xff) << 24;
973 cmd->data[1] = devid;
974 cmd->data[1] |= (pasid & 0xff) << 16;
975 cmd->data[2] = lower_32_bits(address);
976 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
977 cmd->data[3] = upper_32_bits(address);
979 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
980 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
983 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
984 int status, int tag, bool gn)
986 memset(cmd, 0, sizeof(*cmd));
988 cmd->data[0] = devid;
990 cmd->data[1] = pasid;
991 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
993 cmd->data[3] = tag & 0x1ff;
994 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
996 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
999 static void build_inv_all(struct iommu_cmd *cmd)
1001 memset(cmd, 0, sizeof(*cmd));
1002 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1005 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1007 memset(cmd, 0, sizeof(*cmd));
1008 cmd->data[0] = devid;
1009 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1013 * Writes the command to the IOMMUs command buffer and informs the
1014 * hardware about the new command.
1016 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1017 struct iommu_cmd *cmd,
1020 unsigned int count = 0;
1021 u32 left, next_tail;
1023 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1025 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1028 /* Skip udelay() the first time around */
1030 if (count == LOOP_TIMEOUT) {
1031 pr_err("AMD-Vi: Command buffer timeout\n");
1038 /* Update head and recheck remaining space */
1039 iommu->cmd_buf_head = readl(iommu->mmio_base +
1040 MMIO_CMD_HEAD_OFFSET);
1045 copy_cmd_to_buffer(iommu, cmd);
1047 /* Do we need to make sure all commands are processed? */
1048 iommu->need_sync = sync;
1053 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1054 struct iommu_cmd *cmd,
1057 unsigned long flags;
1060 spin_lock_irqsave(&iommu->lock, flags);
1061 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1062 spin_unlock_irqrestore(&iommu->lock, flags);
1067 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1069 return iommu_queue_command_sync(iommu, cmd, true);
1073 * This function queues a completion wait command into the command
1074 * buffer of an IOMMU
1076 static int iommu_completion_wait(struct amd_iommu *iommu)
1078 struct iommu_cmd cmd;
1079 unsigned long flags;
1082 if (!iommu->need_sync)
1086 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1088 spin_lock_irqsave(&iommu->lock, flags);
1092 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1096 ret = wait_on_sem(&iommu->cmd_sem);
1099 spin_unlock_irqrestore(&iommu->lock, flags);
1104 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1106 struct iommu_cmd cmd;
1108 build_inv_dte(&cmd, devid);
1110 return iommu_queue_command(iommu, &cmd);
1113 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1117 for (devid = 0; devid <= 0xffff; ++devid)
1118 iommu_flush_dte(iommu, devid);
1120 iommu_completion_wait(iommu);
1124 * This function uses heavy locking and may disable irqs for some time. But
1125 * this is no issue because it is only called during resume.
1127 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1131 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1132 struct iommu_cmd cmd;
1133 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1135 iommu_queue_command(iommu, &cmd);
1138 iommu_completion_wait(iommu);
1141 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1143 struct iommu_cmd cmd;
1145 build_inv_all(&cmd);
1147 iommu_queue_command(iommu, &cmd);
1148 iommu_completion_wait(iommu);
1151 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1153 struct iommu_cmd cmd;
1155 build_inv_irt(&cmd, devid);
1157 iommu_queue_command(iommu, &cmd);
1160 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1164 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1165 iommu_flush_irt(iommu, devid);
1167 iommu_completion_wait(iommu);
1170 void iommu_flush_all_caches(struct amd_iommu *iommu)
1172 if (iommu_feature(iommu, FEATURE_IA)) {
1173 amd_iommu_flush_all(iommu);
1175 amd_iommu_flush_dte_all(iommu);
1176 amd_iommu_flush_irt_all(iommu);
1177 amd_iommu_flush_tlb_all(iommu);
1182 * Command send function for flushing on-device TLB
1184 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1185 u64 address, size_t size)
1187 struct amd_iommu *iommu;
1188 struct iommu_cmd cmd;
1191 qdep = dev_data->ats.qdep;
1192 iommu = amd_iommu_rlookup_table[dev_data->devid];
1194 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1196 return iommu_queue_command(iommu, &cmd);
1200 * Command send function for invalidating a device table entry
1202 static int device_flush_dte(struct iommu_dev_data *dev_data)
1204 struct amd_iommu *iommu;
1208 iommu = amd_iommu_rlookup_table[dev_data->devid];
1209 alias = dev_data->alias;
1211 ret = iommu_flush_dte(iommu, dev_data->devid);
1212 if (!ret && alias != dev_data->devid)
1213 ret = iommu_flush_dte(iommu, alias);
1217 if (dev_data->ats.enabled)
1218 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1224 * TLB invalidation function which is called from the mapping functions.
1225 * It invalidates a single PTE if the range to flush is within a single
1226 * page. Otherwise it flushes the whole TLB of the IOMMU.
1228 static void __domain_flush_pages(struct protection_domain *domain,
1229 u64 address, size_t size, int pde)
1231 struct iommu_dev_data *dev_data;
1232 struct iommu_cmd cmd;
1235 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1237 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1238 if (!domain->dev_iommu[i])
1242 * Devices of this domain are behind this IOMMU
1243 * We need a TLB flush
1245 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1248 list_for_each_entry(dev_data, &domain->dev_list, list) {
1250 if (!dev_data->ats.enabled)
1253 ret |= device_flush_iotlb(dev_data, address, size);
1259 static void domain_flush_pages(struct protection_domain *domain,
1260 u64 address, size_t size)
1262 __domain_flush_pages(domain, address, size, 0);
1265 /* Flush the whole IO/TLB for a given protection domain */
1266 static void domain_flush_tlb(struct protection_domain *domain)
1268 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1271 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1272 static void domain_flush_tlb_pde(struct protection_domain *domain)
1274 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1277 static void domain_flush_complete(struct protection_domain *domain)
1281 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1282 if (domain && !domain->dev_iommu[i])
1286 * Devices of this domain are behind this IOMMU
1287 * We need to wait for completion of all commands.
1289 iommu_completion_wait(amd_iommus[i]);
1295 * This function flushes the DTEs for all devices in domain
1297 static void domain_flush_devices(struct protection_domain *domain)
1299 struct iommu_dev_data *dev_data;
1301 list_for_each_entry(dev_data, &domain->dev_list, list)
1302 device_flush_dte(dev_data);
1305 /****************************************************************************
1307 * The functions below are used the create the page table mappings for
1308 * unity mapped regions.
1310 ****************************************************************************/
1313 * This function is used to add another level to an IO page table. Adding
1314 * another level increases the size of the address space by 9 bits to a size up
1317 static bool increase_address_space(struct protection_domain *domain,
1322 if (domain->mode == PAGE_MODE_6_LEVEL)
1323 /* address space already 64 bit large */
1326 pte = (void *)get_zeroed_page(gfp);
1330 *pte = PM_LEVEL_PDE(domain->mode,
1331 iommu_virt_to_phys(domain->pt_root));
1332 domain->pt_root = pte;
1334 domain->updated = true;
1339 static u64 *alloc_pte(struct protection_domain *domain,
1340 unsigned long address,
1341 unsigned long page_size,
1348 BUG_ON(!is_power_of_2(page_size));
1350 while (address > PM_LEVEL_SIZE(domain->mode))
1351 increase_address_space(domain, gfp);
1353 level = domain->mode - 1;
1354 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1355 address = PAGE_SIZE_ALIGN(address, page_size);
1356 end_lvl = PAGE_SIZE_LEVEL(page_size);
1358 while (level > end_lvl) {
1363 if (!IOMMU_PTE_PRESENT(__pte)) {
1364 page = (u64 *)get_zeroed_page(gfp);
1368 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1370 /* pte could have been changed somewhere. */
1371 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1372 free_page((unsigned long)page);
1377 /* No level skipping support yet */
1378 if (PM_PTE_LEVEL(*pte) != level)
1383 pte = IOMMU_PTE_PAGE(*pte);
1385 if (pte_page && level == end_lvl)
1388 pte = &pte[PM_LEVEL_INDEX(level, address)];
1395 * This function checks if there is a PTE for a given dma address. If
1396 * there is one, it returns the pointer to it.
1398 static u64 *fetch_pte(struct protection_domain *domain,
1399 unsigned long address,
1400 unsigned long *page_size)
1405 if (address > PM_LEVEL_SIZE(domain->mode))
1408 level = domain->mode - 1;
1409 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1410 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1415 if (!IOMMU_PTE_PRESENT(*pte))
1419 if (PM_PTE_LEVEL(*pte) == 7 ||
1420 PM_PTE_LEVEL(*pte) == 0)
1423 /* No level skipping support yet */
1424 if (PM_PTE_LEVEL(*pte) != level)
1429 /* Walk to the next level */
1430 pte = IOMMU_PTE_PAGE(*pte);
1431 pte = &pte[PM_LEVEL_INDEX(level, address)];
1432 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1435 if (PM_PTE_LEVEL(*pte) == 0x07) {
1436 unsigned long pte_mask;
1439 * If we have a series of large PTEs, make
1440 * sure to return a pointer to the first one.
1442 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1443 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1444 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1451 * Generic mapping functions. It maps a physical address into a DMA
1452 * address space. It allocates the page table pages if necessary.
1453 * In the future it can be extended to a generic mapping function
1454 * supporting all features of AMD IOMMU page tables like level skipping
1455 * and full 64 bit address spaces.
1457 static int iommu_map_page(struct protection_domain *dom,
1458 unsigned long bus_addr,
1459 unsigned long phys_addr,
1460 unsigned long page_size,
1467 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1468 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1470 if (!(prot & IOMMU_PROT_MASK))
1473 count = PAGE_SIZE_PTE_COUNT(page_size);
1474 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1479 for (i = 0; i < count; ++i)
1480 if (IOMMU_PTE_PRESENT(pte[i]))
1484 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1485 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1487 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1489 if (prot & IOMMU_PROT_IR)
1490 __pte |= IOMMU_PTE_IR;
1491 if (prot & IOMMU_PROT_IW)
1492 __pte |= IOMMU_PTE_IW;
1494 for (i = 0; i < count; ++i)
1502 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1503 unsigned long bus_addr,
1504 unsigned long page_size)
1506 unsigned long long unmapped;
1507 unsigned long unmap_size;
1510 BUG_ON(!is_power_of_2(page_size));
1514 while (unmapped < page_size) {
1516 pte = fetch_pte(dom, bus_addr, &unmap_size);
1521 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1522 for (i = 0; i < count; i++)
1526 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1527 unmapped += unmap_size;
1530 BUG_ON(unmapped && !is_power_of_2(unmapped));
1535 /****************************************************************************
1537 * The next functions belong to the address allocator for the dma_ops
1538 * interface functions.
1540 ****************************************************************************/
1543 static unsigned long dma_ops_alloc_iova(struct device *dev,
1544 struct dma_ops_domain *dma_dom,
1545 unsigned int pages, u64 dma_mask)
1547 unsigned long pfn = 0;
1549 pages = __roundup_pow_of_two(pages);
1551 if (dma_mask > DMA_BIT_MASK(32))
1552 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1553 IOVA_PFN(DMA_BIT_MASK(32)), false);
1556 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1557 IOVA_PFN(dma_mask), true);
1559 return (pfn << PAGE_SHIFT);
1562 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1563 unsigned long address,
1566 pages = __roundup_pow_of_two(pages);
1567 address >>= PAGE_SHIFT;
1569 free_iova_fast(&dma_dom->iovad, address, pages);
1572 /****************************************************************************
1574 * The next functions belong to the domain allocation. A domain is
1575 * allocated for every IOMMU as the default domain. If device isolation
1576 * is enabled, every device get its own domain. The most important thing
1577 * about domains is the page table mapping the DMA address space they
1580 ****************************************************************************/
1583 * This function adds a protection domain to the global protection domain list
1585 static void add_domain_to_list(struct protection_domain *domain)
1587 unsigned long flags;
1589 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1590 list_add(&domain->list, &amd_iommu_pd_list);
1591 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1595 * This function removes a protection domain to the global
1596 * protection domain list
1598 static void del_domain_from_list(struct protection_domain *domain)
1600 unsigned long flags;
1602 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1603 list_del(&domain->list);
1604 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1607 static u16 domain_id_alloc(void)
1609 unsigned long flags;
1612 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1613 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1615 if (id > 0 && id < MAX_DOMAIN_ID)
1616 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1619 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1624 static void domain_id_free(int id)
1626 unsigned long flags;
1628 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1629 if (id > 0 && id < MAX_DOMAIN_ID)
1630 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1631 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1634 #define DEFINE_FREE_PT_FN(LVL, FN) \
1635 static void free_pt_##LVL (unsigned long __pt) \
1643 for (i = 0; i < 512; ++i) { \
1644 /* PTE present? */ \
1645 if (!IOMMU_PTE_PRESENT(pt[i])) \
1649 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1650 PM_PTE_LEVEL(pt[i]) == 7) \
1653 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1656 free_page((unsigned long)pt); \
1659 DEFINE_FREE_PT_FN(l2, free_page)
1660 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1661 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1662 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1663 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1665 static void free_pagetable(struct protection_domain *domain)
1667 unsigned long root = (unsigned long)domain->pt_root;
1669 switch (domain->mode) {
1670 case PAGE_MODE_NONE:
1672 case PAGE_MODE_1_LEVEL:
1675 case PAGE_MODE_2_LEVEL:
1678 case PAGE_MODE_3_LEVEL:
1681 case PAGE_MODE_4_LEVEL:
1684 case PAGE_MODE_5_LEVEL:
1687 case PAGE_MODE_6_LEVEL:
1695 static void free_gcr3_tbl_level1(u64 *tbl)
1700 for (i = 0; i < 512; ++i) {
1701 if (!(tbl[i] & GCR3_VALID))
1704 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1706 free_page((unsigned long)ptr);
1710 static void free_gcr3_tbl_level2(u64 *tbl)
1715 for (i = 0; i < 512; ++i) {
1716 if (!(tbl[i] & GCR3_VALID))
1719 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1721 free_gcr3_tbl_level1(ptr);
1725 static void free_gcr3_table(struct protection_domain *domain)
1727 if (domain->glx == 2)
1728 free_gcr3_tbl_level2(domain->gcr3_tbl);
1729 else if (domain->glx == 1)
1730 free_gcr3_tbl_level1(domain->gcr3_tbl);
1732 BUG_ON(domain->glx != 0);
1734 free_page((unsigned long)domain->gcr3_tbl);
1737 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1739 domain_flush_tlb(&dom->domain);
1740 domain_flush_complete(&dom->domain);
1743 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1745 struct dma_ops_domain *dom;
1747 dom = container_of(iovad, struct dma_ops_domain, iovad);
1749 dma_ops_domain_flush_tlb(dom);
1753 * Free a domain, only used if something went wrong in the
1754 * allocation path and we need to free an already allocated page table
1756 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1761 del_domain_from_list(&dom->domain);
1763 put_iova_domain(&dom->iovad);
1765 free_pagetable(&dom->domain);
1768 domain_id_free(dom->domain.id);
1774 * Allocates a new protection domain usable for the dma_ops functions.
1775 * It also initializes the page table and the address allocator data
1776 * structures required for the dma_ops interface
1778 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1780 struct dma_ops_domain *dma_dom;
1782 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1786 if (protection_domain_init(&dma_dom->domain))
1789 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1790 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1791 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1792 if (!dma_dom->domain.pt_root)
1795 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1797 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1800 /* Initialize reserved ranges */
1801 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1803 add_domain_to_list(&dma_dom->domain);
1808 dma_ops_domain_free(dma_dom);
1814 * little helper function to check whether a given protection domain is a
1817 static bool dma_ops_domain(struct protection_domain *domain)
1819 return domain->flags & PD_DMA_OPS_MASK;
1822 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1828 if (domain->mode != PAGE_MODE_NONE)
1829 pte_root = iommu_virt_to_phys(domain->pt_root);
1831 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1832 << DEV_ENTRY_MODE_SHIFT;
1833 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1835 flags = amd_iommu_dev_table[devid].data[1];
1838 flags |= DTE_FLAG_IOTLB;
1841 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1843 if (iommu_feature(iommu, FEATURE_EPHSUP))
1844 pte_root |= 1ULL << DEV_ENTRY_PPR;
1847 if (domain->flags & PD_IOMMUV2_MASK) {
1848 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1849 u64 glx = domain->glx;
1852 pte_root |= DTE_FLAG_GV;
1853 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1855 /* First mask out possible old values for GCR3 table */
1856 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1859 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1862 /* Encode GCR3 table into DTE */
1863 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1866 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1869 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1873 flags &= ~DEV_DOMID_MASK;
1874 flags |= domain->id;
1876 amd_iommu_dev_table[devid].data[1] = flags;
1877 amd_iommu_dev_table[devid].data[0] = pte_root;
1880 static void clear_dte_entry(u16 devid)
1882 /* remove entry from the device table seen by the hardware */
1883 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1884 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1886 amd_iommu_apply_erratum_63(devid);
1889 static void do_attach(struct iommu_dev_data *dev_data,
1890 struct protection_domain *domain)
1892 struct amd_iommu *iommu;
1896 iommu = amd_iommu_rlookup_table[dev_data->devid];
1897 alias = dev_data->alias;
1898 ats = dev_data->ats.enabled;
1900 /* Update data structures */
1901 dev_data->domain = domain;
1902 list_add(&dev_data->list, &domain->dev_list);
1904 /* Do reference counting */
1905 domain->dev_iommu[iommu->index] += 1;
1906 domain->dev_cnt += 1;
1908 /* Update device table */
1909 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
1910 if (alias != dev_data->devid)
1911 set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
1913 device_flush_dte(dev_data);
1916 static void do_detach(struct iommu_dev_data *dev_data)
1918 struct amd_iommu *iommu;
1922 * First check if the device is still attached. It might already
1923 * be detached from its domain because the generic
1924 * iommu_detach_group code detached it and we try again here in
1925 * our alias handling.
1927 if (!dev_data->domain)
1930 iommu = amd_iommu_rlookup_table[dev_data->devid];
1931 alias = dev_data->alias;
1933 /* decrease reference counters */
1934 dev_data->domain->dev_iommu[iommu->index] -= 1;
1935 dev_data->domain->dev_cnt -= 1;
1937 /* Update data structures */
1938 dev_data->domain = NULL;
1939 list_del(&dev_data->list);
1940 clear_dte_entry(dev_data->devid);
1941 if (alias != dev_data->devid)
1942 clear_dte_entry(alias);
1944 /* Flush the DTE entry */
1945 device_flush_dte(dev_data);
1949 * If a device is not yet associated with a domain, this function does
1950 * assigns it visible for the hardware
1952 static int __attach_device(struct iommu_dev_data *dev_data,
1953 struct protection_domain *domain)
1958 * Must be called with IRQs disabled. Warn here to detect early
1961 WARN_ON(!irqs_disabled());
1964 spin_lock(&domain->lock);
1967 if (dev_data->domain != NULL)
1970 /* Attach alias group root */
1971 do_attach(dev_data, domain);
1978 spin_unlock(&domain->lock);
1984 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1986 pci_disable_ats(pdev);
1987 pci_disable_pri(pdev);
1988 pci_disable_pasid(pdev);
1991 /* FIXME: Change generic reset-function to do the same */
1992 static int pri_reset_while_enabled(struct pci_dev *pdev)
1997 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2001 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2002 control |= PCI_PRI_CTRL_RESET;
2003 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2008 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2013 /* FIXME: Hardcode number of outstanding requests for now */
2015 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2017 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2019 /* Only allow access to user-accessible pages */
2020 ret = pci_enable_pasid(pdev, 0);
2024 /* First reset the PRI state of the device */
2025 ret = pci_reset_pri(pdev);
2030 ret = pci_enable_pri(pdev, reqs);
2035 ret = pri_reset_while_enabled(pdev);
2040 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2047 pci_disable_pri(pdev);
2048 pci_disable_pasid(pdev);
2053 /* FIXME: Move this to PCI code */
2054 #define PCI_PRI_TLP_OFF (1 << 15)
2056 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2061 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2065 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2067 return (status & PCI_PRI_TLP_OFF) ? true : false;
2071 * If a device is not yet associated with a domain, this function
2072 * assigns it visible for the hardware
2074 static int attach_device(struct device *dev,
2075 struct protection_domain *domain)
2077 struct pci_dev *pdev;
2078 struct iommu_dev_data *dev_data;
2079 unsigned long flags;
2082 dev_data = get_dev_data(dev);
2084 if (!dev_is_pci(dev))
2085 goto skip_ats_check;
2087 pdev = to_pci_dev(dev);
2088 if (domain->flags & PD_IOMMUV2_MASK) {
2089 if (!dev_data->passthrough)
2092 if (dev_data->iommu_v2) {
2093 if (pdev_iommuv2_enable(pdev) != 0)
2096 dev_data->ats.enabled = true;
2097 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2098 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2100 } else if (amd_iommu_iotlb_sup &&
2101 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2102 dev_data->ats.enabled = true;
2103 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2107 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2108 ret = __attach_device(dev_data, domain);
2109 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2112 * We might boot into a crash-kernel here. The crashed kernel
2113 * left the caches in the IOMMU dirty. So we have to flush
2114 * here to evict all dirty stuff.
2116 domain_flush_tlb_pde(domain);
2122 * Removes a device from a protection domain (unlocked)
2124 static void __detach_device(struct iommu_dev_data *dev_data)
2126 struct protection_domain *domain;
2129 * Must be called with IRQs disabled. Warn here to detect early
2132 WARN_ON(!irqs_disabled());
2134 if (WARN_ON(!dev_data->domain))
2137 domain = dev_data->domain;
2139 spin_lock(&domain->lock);
2141 do_detach(dev_data);
2143 spin_unlock(&domain->lock);
2147 * Removes a device from a protection domain (with devtable_lock held)
2149 static void detach_device(struct device *dev)
2151 struct protection_domain *domain;
2152 struct iommu_dev_data *dev_data;
2153 unsigned long flags;
2155 dev_data = get_dev_data(dev);
2156 domain = dev_data->domain;
2158 /* lock device table */
2159 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2160 __detach_device(dev_data);
2161 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2163 if (!dev_is_pci(dev))
2166 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2167 pdev_iommuv2_disable(to_pci_dev(dev));
2168 else if (dev_data->ats.enabled)
2169 pci_disable_ats(to_pci_dev(dev));
2171 dev_data->ats.enabled = false;
2174 static int amd_iommu_add_device(struct device *dev)
2176 struct iommu_dev_data *dev_data;
2177 struct iommu_domain *domain;
2178 struct amd_iommu *iommu;
2181 if (!check_device(dev) || get_dev_data(dev))
2184 devid = get_device_id(dev);
2188 iommu = amd_iommu_rlookup_table[devid];
2190 ret = iommu_init_device(dev);
2192 if (ret != -ENOTSUPP)
2193 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2196 iommu_ignore_device(dev);
2197 dev->dma_ops = &dma_direct_ops;
2200 init_iommu_group(dev);
2202 dev_data = get_dev_data(dev);
2206 if (iommu_pass_through || dev_data->iommu_v2)
2207 iommu_request_dm_for_dev(dev);
2209 /* Domains are initialized for this device - have a look what we ended up with */
2210 domain = iommu_get_domain_for_dev(dev);
2211 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2212 dev_data->passthrough = true;
2214 dev->dma_ops = &amd_iommu_dma_ops;
2217 iommu_completion_wait(iommu);
2222 static void amd_iommu_remove_device(struct device *dev)
2224 struct amd_iommu *iommu;
2227 if (!check_device(dev))
2230 devid = get_device_id(dev);
2234 iommu = amd_iommu_rlookup_table[devid];
2236 iommu_uninit_device(dev);
2237 iommu_completion_wait(iommu);
2240 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2242 if (dev_is_pci(dev))
2243 return pci_device_group(dev);
2245 return acpihid_device_group(dev);
2248 /*****************************************************************************
2250 * The next functions belong to the dma_ops mapping/unmapping code.
2252 *****************************************************************************/
2255 * In the dma_ops path we only have the struct device. This function
2256 * finds the corresponding IOMMU, the protection domain and the
2257 * requestor id for a given device.
2258 * If the device is not yet associated with a domain this is also done
2261 static struct protection_domain *get_domain(struct device *dev)
2263 struct protection_domain *domain;
2264 struct iommu_domain *io_domain;
2266 if (!check_device(dev))
2267 return ERR_PTR(-EINVAL);
2269 domain = get_dev_data(dev)->domain;
2270 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2271 get_dev_data(dev)->defer_attach = false;
2272 io_domain = iommu_get_domain_for_dev(dev);
2273 domain = to_pdomain(io_domain);
2274 attach_device(dev, domain);
2277 return ERR_PTR(-EBUSY);
2279 if (!dma_ops_domain(domain))
2280 return ERR_PTR(-EBUSY);
2285 static void update_device_table(struct protection_domain *domain)
2287 struct iommu_dev_data *dev_data;
2289 list_for_each_entry(dev_data, &domain->dev_list, list) {
2290 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2291 dev_data->iommu_v2);
2293 if (dev_data->devid == dev_data->alias)
2296 /* There is an alias, update device table entry for it */
2297 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2298 dev_data->iommu_v2);
2302 static void update_domain(struct protection_domain *domain)
2304 if (!domain->updated)
2307 update_device_table(domain);
2309 domain_flush_devices(domain);
2310 domain_flush_tlb_pde(domain);
2312 domain->updated = false;
2315 static int dir2prot(enum dma_data_direction direction)
2317 if (direction == DMA_TO_DEVICE)
2318 return IOMMU_PROT_IR;
2319 else if (direction == DMA_FROM_DEVICE)
2320 return IOMMU_PROT_IW;
2321 else if (direction == DMA_BIDIRECTIONAL)
2322 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2328 * This function contains common code for mapping of a physically
2329 * contiguous memory region into DMA address space. It is used by all
2330 * mapping functions provided with this IOMMU driver.
2331 * Must be called with the domain lock held.
2333 static dma_addr_t __map_single(struct device *dev,
2334 struct dma_ops_domain *dma_dom,
2337 enum dma_data_direction direction,
2340 dma_addr_t offset = paddr & ~PAGE_MASK;
2341 dma_addr_t address, start, ret;
2346 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2349 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2350 if (address == AMD_IOMMU_MAPPING_ERROR)
2353 prot = dir2prot(direction);
2356 for (i = 0; i < pages; ++i) {
2357 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2358 PAGE_SIZE, prot, GFP_ATOMIC);
2367 if (unlikely(amd_iommu_np_cache)) {
2368 domain_flush_pages(&dma_dom->domain, address, size);
2369 domain_flush_complete(&dma_dom->domain);
2377 for (--i; i >= 0; --i) {
2379 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2382 domain_flush_tlb(&dma_dom->domain);
2383 domain_flush_complete(&dma_dom->domain);
2385 dma_ops_free_iova(dma_dom, address, pages);
2387 return AMD_IOMMU_MAPPING_ERROR;
2391 * Does the reverse of the __map_single function. Must be called with
2392 * the domain lock held too
2394 static void __unmap_single(struct dma_ops_domain *dma_dom,
2395 dma_addr_t dma_addr,
2399 dma_addr_t i, start;
2402 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2403 dma_addr &= PAGE_MASK;
2406 for (i = 0; i < pages; ++i) {
2407 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2411 if (amd_iommu_unmap_flush) {
2412 dma_ops_free_iova(dma_dom, dma_addr, pages);
2413 domain_flush_tlb(&dma_dom->domain);
2414 domain_flush_complete(&dma_dom->domain);
2416 pages = __roundup_pow_of_two(pages);
2417 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2422 * The exported map_single function for dma_ops.
2424 static dma_addr_t map_page(struct device *dev, struct page *page,
2425 unsigned long offset, size_t size,
2426 enum dma_data_direction dir,
2427 unsigned long attrs)
2429 phys_addr_t paddr = page_to_phys(page) + offset;
2430 struct protection_domain *domain;
2431 struct dma_ops_domain *dma_dom;
2434 domain = get_domain(dev);
2435 if (PTR_ERR(domain) == -EINVAL)
2436 return (dma_addr_t)paddr;
2437 else if (IS_ERR(domain))
2438 return AMD_IOMMU_MAPPING_ERROR;
2440 dma_mask = *dev->dma_mask;
2441 dma_dom = to_dma_ops_domain(domain);
2443 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2447 * The exported unmap_single function for dma_ops.
2449 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2450 enum dma_data_direction dir, unsigned long attrs)
2452 struct protection_domain *domain;
2453 struct dma_ops_domain *dma_dom;
2455 domain = get_domain(dev);
2459 dma_dom = to_dma_ops_domain(domain);
2461 __unmap_single(dma_dom, dma_addr, size, dir);
2464 static int sg_num_pages(struct device *dev,
2465 struct scatterlist *sglist,
2468 unsigned long mask, boundary_size;
2469 struct scatterlist *s;
2472 mask = dma_get_seg_boundary(dev);
2473 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2474 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2476 for_each_sg(sglist, s, nelems, i) {
2479 s->dma_address = npages << PAGE_SHIFT;
2480 p = npages % boundary_size;
2481 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2482 if (p + n > boundary_size)
2483 npages += boundary_size - p;
2491 * The exported map_sg function for dma_ops (handles scatter-gather
2494 static int map_sg(struct device *dev, struct scatterlist *sglist,
2495 int nelems, enum dma_data_direction direction,
2496 unsigned long attrs)
2498 int mapped_pages = 0, npages = 0, prot = 0, i;
2499 struct protection_domain *domain;
2500 struct dma_ops_domain *dma_dom;
2501 struct scatterlist *s;
2502 unsigned long address;
2505 domain = get_domain(dev);
2509 dma_dom = to_dma_ops_domain(domain);
2510 dma_mask = *dev->dma_mask;
2512 npages = sg_num_pages(dev, sglist, nelems);
2514 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2515 if (address == AMD_IOMMU_MAPPING_ERROR)
2518 prot = dir2prot(direction);
2520 /* Map all sg entries */
2521 for_each_sg(sglist, s, nelems, i) {
2522 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2524 for (j = 0; j < pages; ++j) {
2525 unsigned long bus_addr, phys_addr;
2528 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2529 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2530 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2538 /* Everything is mapped - write the right values into s->dma_address */
2539 for_each_sg(sglist, s, nelems, i) {
2540 s->dma_address += address + s->offset;
2541 s->dma_length = s->length;
2547 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2548 dev_name(dev), npages);
2550 for_each_sg(sglist, s, nelems, i) {
2551 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2553 for (j = 0; j < pages; ++j) {
2554 unsigned long bus_addr;
2556 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2557 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2565 free_iova_fast(&dma_dom->iovad, address, npages);
2572 * The exported map_sg function for dma_ops (handles scatter-gather
2575 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2576 int nelems, enum dma_data_direction dir,
2577 unsigned long attrs)
2579 struct protection_domain *domain;
2580 struct dma_ops_domain *dma_dom;
2581 unsigned long startaddr;
2584 domain = get_domain(dev);
2588 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2589 dma_dom = to_dma_ops_domain(domain);
2590 npages = sg_num_pages(dev, sglist, nelems);
2592 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2596 * The exported alloc_coherent function for dma_ops.
2598 static void *alloc_coherent(struct device *dev, size_t size,
2599 dma_addr_t *dma_addr, gfp_t flag,
2600 unsigned long attrs)
2602 u64 dma_mask = dev->coherent_dma_mask;
2603 struct protection_domain *domain = get_domain(dev);
2604 bool is_direct = false;
2607 if (IS_ERR(domain)) {
2608 if (PTR_ERR(domain) != -EINVAL)
2613 virt_addr = dma_direct_alloc(dev, size, dma_addr, flag, attrs);
2614 if (!virt_addr || is_direct)
2618 dma_mask = *dev->dma_mask;
2620 *dma_addr = __map_single(dev, to_dma_ops_domain(domain),
2621 virt_to_phys(virt_addr), PAGE_ALIGN(size),
2622 DMA_BIDIRECTIONAL, dma_mask);
2623 if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2628 dma_direct_free(dev, size, virt_addr, *dma_addr, attrs);
2633 * The exported free_coherent function for dma_ops.
2635 static void free_coherent(struct device *dev, size_t size,
2636 void *virt_addr, dma_addr_t dma_addr,
2637 unsigned long attrs)
2639 struct protection_domain *domain = get_domain(dev);
2641 size = PAGE_ALIGN(size);
2643 if (!IS_ERR(domain)) {
2644 struct dma_ops_domain *dma_dom = to_dma_ops_domain(domain);
2646 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2649 dma_direct_free(dev, size, virt_addr, dma_addr, attrs);
2653 * This function is called by the DMA layer to find out if we can handle a
2654 * particular device. It is part of the dma_ops.
2656 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2658 if (!dma_direct_supported(dev, mask))
2660 return check_device(dev);
2663 static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2665 return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2668 static const struct dma_map_ops amd_iommu_dma_ops = {
2669 .alloc = alloc_coherent,
2670 .free = free_coherent,
2671 .map_page = map_page,
2672 .unmap_page = unmap_page,
2674 .unmap_sg = unmap_sg,
2675 .dma_supported = amd_iommu_dma_supported,
2676 .mapping_error = amd_iommu_mapping_error,
2679 static int init_reserved_iova_ranges(void)
2681 struct pci_dev *pdev = NULL;
2684 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2686 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2687 &reserved_rbtree_key);
2689 /* MSI memory range */
2690 val = reserve_iova(&reserved_iova_ranges,
2691 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2693 pr_err("Reserving MSI range failed\n");
2697 /* HT memory range */
2698 val = reserve_iova(&reserved_iova_ranges,
2699 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2701 pr_err("Reserving HT range failed\n");
2706 * Memory used for PCI resources
2707 * FIXME: Check whether we can reserve the PCI-hole completly
2709 for_each_pci_dev(pdev) {
2712 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2713 struct resource *r = &pdev->resource[i];
2715 if (!(r->flags & IORESOURCE_MEM))
2718 val = reserve_iova(&reserved_iova_ranges,
2722 pr_err("Reserve pci-resource range failed\n");
2731 int __init amd_iommu_init_api(void)
2735 ret = iova_cache_get();
2739 ret = init_reserved_iova_ranges();
2743 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2746 #ifdef CONFIG_ARM_AMBA
2747 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2751 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2758 int __init amd_iommu_init_dma_ops(void)
2760 swiotlb = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2764 * In case we don't initialize SWIOTLB (actually the common case
2765 * when AMD IOMMU is enabled and SME is not active), make sure there
2766 * are global dma_ops set as a fall-back for devices not handled by
2767 * this driver (for example non-PCI devices). When SME is active,
2768 * make sure that swiotlb variable remains set so the global dma_ops
2769 * continue to be SWIOTLB.
2772 dma_ops = &dma_direct_ops;
2774 if (amd_iommu_unmap_flush)
2775 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2777 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2783 /*****************************************************************************
2785 * The following functions belong to the exported interface of AMD IOMMU
2787 * This interface allows access to lower level functions of the IOMMU
2788 * like protection domain handling and assignement of devices to domains
2789 * which is not possible with the dma_ops interface.
2791 *****************************************************************************/
2793 static void cleanup_domain(struct protection_domain *domain)
2795 struct iommu_dev_data *entry;
2796 unsigned long flags;
2798 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2800 while (!list_empty(&domain->dev_list)) {
2801 entry = list_first_entry(&domain->dev_list,
2802 struct iommu_dev_data, list);
2803 __detach_device(entry);
2806 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2809 static void protection_domain_free(struct protection_domain *domain)
2814 del_domain_from_list(domain);
2817 domain_id_free(domain->id);
2822 static int protection_domain_init(struct protection_domain *domain)
2824 spin_lock_init(&domain->lock);
2825 mutex_init(&domain->api_lock);
2826 domain->id = domain_id_alloc();
2829 INIT_LIST_HEAD(&domain->dev_list);
2834 static struct protection_domain *protection_domain_alloc(void)
2836 struct protection_domain *domain;
2838 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2842 if (protection_domain_init(domain))
2845 add_domain_to_list(domain);
2855 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2857 struct protection_domain *pdomain;
2858 struct dma_ops_domain *dma_domain;
2861 case IOMMU_DOMAIN_UNMANAGED:
2862 pdomain = protection_domain_alloc();
2866 pdomain->mode = PAGE_MODE_3_LEVEL;
2867 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2868 if (!pdomain->pt_root) {
2869 protection_domain_free(pdomain);
2873 pdomain->domain.geometry.aperture_start = 0;
2874 pdomain->domain.geometry.aperture_end = ~0ULL;
2875 pdomain->domain.geometry.force_aperture = true;
2878 case IOMMU_DOMAIN_DMA:
2879 dma_domain = dma_ops_domain_alloc();
2881 pr_err("AMD-Vi: Failed to allocate\n");
2884 pdomain = &dma_domain->domain;
2886 case IOMMU_DOMAIN_IDENTITY:
2887 pdomain = protection_domain_alloc();
2891 pdomain->mode = PAGE_MODE_NONE;
2897 return &pdomain->domain;
2900 static void amd_iommu_domain_free(struct iommu_domain *dom)
2902 struct protection_domain *domain;
2903 struct dma_ops_domain *dma_dom;
2905 domain = to_pdomain(dom);
2907 if (domain->dev_cnt > 0)
2908 cleanup_domain(domain);
2910 BUG_ON(domain->dev_cnt != 0);
2915 switch (dom->type) {
2916 case IOMMU_DOMAIN_DMA:
2917 /* Now release the domain */
2918 dma_dom = to_dma_ops_domain(domain);
2919 dma_ops_domain_free(dma_dom);
2922 if (domain->mode != PAGE_MODE_NONE)
2923 free_pagetable(domain);
2925 if (domain->flags & PD_IOMMUV2_MASK)
2926 free_gcr3_table(domain);
2928 protection_domain_free(domain);
2933 static void amd_iommu_detach_device(struct iommu_domain *dom,
2936 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2937 struct amd_iommu *iommu;
2940 if (!check_device(dev))
2943 devid = get_device_id(dev);
2947 if (dev_data->domain != NULL)
2950 iommu = amd_iommu_rlookup_table[devid];
2954 #ifdef CONFIG_IRQ_REMAP
2955 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2956 (dom->type == IOMMU_DOMAIN_UNMANAGED))
2957 dev_data->use_vapic = 0;
2960 iommu_completion_wait(iommu);
2963 static int amd_iommu_attach_device(struct iommu_domain *dom,
2966 struct protection_domain *domain = to_pdomain(dom);
2967 struct iommu_dev_data *dev_data;
2968 struct amd_iommu *iommu;
2971 if (!check_device(dev))
2974 dev_data = dev->archdata.iommu;
2976 iommu = amd_iommu_rlookup_table[dev_data->devid];
2980 if (dev_data->domain)
2983 ret = attach_device(dev, domain);
2985 #ifdef CONFIG_IRQ_REMAP
2986 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2987 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2988 dev_data->use_vapic = 1;
2990 dev_data->use_vapic = 0;
2994 iommu_completion_wait(iommu);
2999 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3000 phys_addr_t paddr, size_t page_size, int iommu_prot)
3002 struct protection_domain *domain = to_pdomain(dom);
3006 if (domain->mode == PAGE_MODE_NONE)
3009 if (iommu_prot & IOMMU_READ)
3010 prot |= IOMMU_PROT_IR;
3011 if (iommu_prot & IOMMU_WRITE)
3012 prot |= IOMMU_PROT_IW;
3014 mutex_lock(&domain->api_lock);
3015 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3016 mutex_unlock(&domain->api_lock);
3021 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3024 struct protection_domain *domain = to_pdomain(dom);
3027 if (domain->mode == PAGE_MODE_NONE)
3030 mutex_lock(&domain->api_lock);
3031 unmap_size = iommu_unmap_page(domain, iova, page_size);
3032 mutex_unlock(&domain->api_lock);
3034 domain_flush_tlb_pde(domain);
3035 domain_flush_complete(domain);
3040 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3043 struct protection_domain *domain = to_pdomain(dom);
3044 unsigned long offset_mask, pte_pgsize;
3047 if (domain->mode == PAGE_MODE_NONE)
3050 pte = fetch_pte(domain, iova, &pte_pgsize);
3052 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3055 offset_mask = pte_pgsize - 1;
3056 __pte = *pte & PM_ADDR_MASK;
3058 return (__pte & ~offset_mask) | (iova & offset_mask);
3061 static bool amd_iommu_capable(enum iommu_cap cap)
3064 case IOMMU_CAP_CACHE_COHERENCY:
3066 case IOMMU_CAP_INTR_REMAP:
3067 return (irq_remapping_enabled == 1);
3068 case IOMMU_CAP_NOEXEC:
3075 static void amd_iommu_get_resv_regions(struct device *dev,
3076 struct list_head *head)
3078 struct iommu_resv_region *region;
3079 struct unity_map_entry *entry;
3082 devid = get_device_id(dev);
3086 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3090 if (devid < entry->devid_start || devid > entry->devid_end)
3093 length = entry->address_end - entry->address_start;
3094 if (entry->prot & IOMMU_PROT_IR)
3096 if (entry->prot & IOMMU_PROT_IW)
3097 prot |= IOMMU_WRITE;
3099 region = iommu_alloc_resv_region(entry->address_start,
3103 pr_err("Out of memory allocating dm-regions for %s\n",
3107 list_add_tail(®ion->list, head);
3110 region = iommu_alloc_resv_region(MSI_RANGE_START,
3111 MSI_RANGE_END - MSI_RANGE_START + 1,
3115 list_add_tail(®ion->list, head);
3117 region = iommu_alloc_resv_region(HT_RANGE_START,
3118 HT_RANGE_END - HT_RANGE_START + 1,
3119 0, IOMMU_RESV_RESERVED);
3122 list_add_tail(®ion->list, head);
3125 static void amd_iommu_put_resv_regions(struct device *dev,
3126 struct list_head *head)
3128 struct iommu_resv_region *entry, *next;
3130 list_for_each_entry_safe(entry, next, head, list)
3134 static void amd_iommu_apply_resv_region(struct device *dev,
3135 struct iommu_domain *domain,
3136 struct iommu_resv_region *region)
3138 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3139 unsigned long start, end;
3141 start = IOVA_PFN(region->start);
3142 end = IOVA_PFN(region->start + region->length - 1);
3144 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3147 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3150 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3151 return dev_data->defer_attach;
3154 const struct iommu_ops amd_iommu_ops = {
3155 .capable = amd_iommu_capable,
3156 .domain_alloc = amd_iommu_domain_alloc,
3157 .domain_free = amd_iommu_domain_free,
3158 .attach_dev = amd_iommu_attach_device,
3159 .detach_dev = amd_iommu_detach_device,
3160 .map = amd_iommu_map,
3161 .unmap = amd_iommu_unmap,
3162 .map_sg = default_iommu_map_sg,
3163 .iova_to_phys = amd_iommu_iova_to_phys,
3164 .add_device = amd_iommu_add_device,
3165 .remove_device = amd_iommu_remove_device,
3166 .device_group = amd_iommu_device_group,
3167 .get_resv_regions = amd_iommu_get_resv_regions,
3168 .put_resv_regions = amd_iommu_put_resv_regions,
3169 .apply_resv_region = amd_iommu_apply_resv_region,
3170 .is_attach_deferred = amd_iommu_is_attach_deferred,
3171 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3174 /*****************************************************************************
3176 * The next functions do a basic initialization of IOMMU for pass through
3179 * In passthrough mode the IOMMU is initialized and enabled but not used for
3180 * DMA-API translation.
3182 *****************************************************************************/
3184 /* IOMMUv2 specific functions */
3185 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3187 return atomic_notifier_chain_register(&ppr_notifier, nb);
3189 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3191 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3193 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3195 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3197 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3199 struct protection_domain *domain = to_pdomain(dom);
3200 unsigned long flags;
3202 spin_lock_irqsave(&domain->lock, flags);
3204 /* Update data structure */
3205 domain->mode = PAGE_MODE_NONE;
3206 domain->updated = true;
3208 /* Make changes visible to IOMMUs */
3209 update_domain(domain);
3211 /* Page-table is not visible to IOMMU anymore, so free it */
3212 free_pagetable(domain);
3214 spin_unlock_irqrestore(&domain->lock, flags);
3216 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3218 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3220 struct protection_domain *domain = to_pdomain(dom);
3221 unsigned long flags;
3224 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3227 /* Number of GCR3 table levels required */
3228 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3231 if (levels > amd_iommu_max_glx_val)
3234 spin_lock_irqsave(&domain->lock, flags);
3237 * Save us all sanity checks whether devices already in the
3238 * domain support IOMMUv2. Just force that the domain has no
3239 * devices attached when it is switched into IOMMUv2 mode.
3242 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3246 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3247 if (domain->gcr3_tbl == NULL)
3250 domain->glx = levels;
3251 domain->flags |= PD_IOMMUV2_MASK;
3252 domain->updated = true;
3254 update_domain(domain);
3259 spin_unlock_irqrestore(&domain->lock, flags);
3263 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3265 static int __flush_pasid(struct protection_domain *domain, int pasid,
3266 u64 address, bool size)
3268 struct iommu_dev_data *dev_data;
3269 struct iommu_cmd cmd;
3272 if (!(domain->flags & PD_IOMMUV2_MASK))
3275 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3278 * IOMMU TLB needs to be flushed before Device TLB to
3279 * prevent device TLB refill from IOMMU TLB
3281 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3282 if (domain->dev_iommu[i] == 0)
3285 ret = iommu_queue_command(amd_iommus[i], &cmd);
3290 /* Wait until IOMMU TLB flushes are complete */
3291 domain_flush_complete(domain);
3293 /* Now flush device TLBs */
3294 list_for_each_entry(dev_data, &domain->dev_list, list) {
3295 struct amd_iommu *iommu;
3299 There might be non-IOMMUv2 capable devices in an IOMMUv2
3302 if (!dev_data->ats.enabled)
3305 qdep = dev_data->ats.qdep;
3306 iommu = amd_iommu_rlookup_table[dev_data->devid];
3308 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3309 qdep, address, size);
3311 ret = iommu_queue_command(iommu, &cmd);
3316 /* Wait until all device TLBs are flushed */
3317 domain_flush_complete(domain);
3326 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3329 return __flush_pasid(domain, pasid, address, false);
3332 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3335 struct protection_domain *domain = to_pdomain(dom);
3336 unsigned long flags;
3339 spin_lock_irqsave(&domain->lock, flags);
3340 ret = __amd_iommu_flush_page(domain, pasid, address);
3341 spin_unlock_irqrestore(&domain->lock, flags);
3345 EXPORT_SYMBOL(amd_iommu_flush_page);
3347 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3349 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3353 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3355 struct protection_domain *domain = to_pdomain(dom);
3356 unsigned long flags;
3359 spin_lock_irqsave(&domain->lock, flags);
3360 ret = __amd_iommu_flush_tlb(domain, pasid);
3361 spin_unlock_irqrestore(&domain->lock, flags);
3365 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3367 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3374 index = (pasid >> (9 * level)) & 0x1ff;
3380 if (!(*pte & GCR3_VALID)) {
3384 root = (void *)get_zeroed_page(GFP_ATOMIC);
3388 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3391 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3399 static int __set_gcr3(struct protection_domain *domain, int pasid,
3404 if (domain->mode != PAGE_MODE_NONE)
3407 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3411 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3413 return __amd_iommu_flush_tlb(domain, pasid);
3416 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3420 if (domain->mode != PAGE_MODE_NONE)
3423 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3429 return __amd_iommu_flush_tlb(domain, pasid);
3432 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3435 struct protection_domain *domain = to_pdomain(dom);
3436 unsigned long flags;
3439 spin_lock_irqsave(&domain->lock, flags);
3440 ret = __set_gcr3(domain, pasid, cr3);
3441 spin_unlock_irqrestore(&domain->lock, flags);
3445 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3447 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3449 struct protection_domain *domain = to_pdomain(dom);
3450 unsigned long flags;
3453 spin_lock_irqsave(&domain->lock, flags);
3454 ret = __clear_gcr3(domain, pasid);
3455 spin_unlock_irqrestore(&domain->lock, flags);
3459 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3461 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3462 int status, int tag)
3464 struct iommu_dev_data *dev_data;
3465 struct amd_iommu *iommu;
3466 struct iommu_cmd cmd;
3468 dev_data = get_dev_data(&pdev->dev);
3469 iommu = amd_iommu_rlookup_table[dev_data->devid];
3471 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3472 tag, dev_data->pri_tlp);
3474 return iommu_queue_command(iommu, &cmd);
3476 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3478 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3480 struct protection_domain *pdomain;
3482 pdomain = get_domain(&pdev->dev);
3483 if (IS_ERR(pdomain))
3486 /* Only return IOMMUv2 domains */
3487 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3490 return &pdomain->domain;
3492 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3494 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3496 struct iommu_dev_data *dev_data;
3498 if (!amd_iommu_v2_supported())
3501 dev_data = get_dev_data(&pdev->dev);
3502 dev_data->errata |= (1 << erratum);
3504 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3506 int amd_iommu_device_info(struct pci_dev *pdev,
3507 struct amd_iommu_device_info *info)
3512 if (pdev == NULL || info == NULL)
3515 if (!amd_iommu_v2_supported())
3518 memset(info, 0, sizeof(*info));
3520 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3522 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3524 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3526 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3528 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3532 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3533 max_pasids = min(max_pasids, (1 << 20));
3535 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3536 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3538 features = pci_pasid_features(pdev);
3539 if (features & PCI_PASID_CAP_EXEC)
3540 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3541 if (features & PCI_PASID_CAP_PRIV)
3542 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3547 EXPORT_SYMBOL(amd_iommu_device_info);
3549 #ifdef CONFIG_IRQ_REMAP
3551 /*****************************************************************************
3553 * Interrupt Remapping Implementation
3555 *****************************************************************************/
3557 static struct irq_chip amd_ir_chip;
3559 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3563 dte = amd_iommu_dev_table[devid].data[2];
3564 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3565 dte |= iommu_virt_to_phys(table->table);
3566 dte |= DTE_IRQ_REMAP_INTCTL;
3567 dte |= DTE_IRQ_TABLE_LEN;
3568 dte |= DTE_IRQ_REMAP_ENABLE;
3570 amd_iommu_dev_table[devid].data[2] = dte;
3573 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3575 struct irq_remap_table *table = NULL;
3576 struct amd_iommu *iommu;
3577 unsigned long flags;
3580 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3582 iommu = amd_iommu_rlookup_table[devid];
3586 table = irq_lookup_table[devid];
3590 alias = amd_iommu_alias_table[devid];
3591 table = irq_lookup_table[alias];
3593 irq_lookup_table[devid] = table;
3594 set_dte_irq_entry(devid, table);
3595 iommu_flush_dte(iommu, devid);
3599 /* Nothing there yet, allocate new irq remapping table */
3600 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3604 /* Initialize table spin-lock */
3605 spin_lock_init(&table->lock);
3608 /* Keep the first 32 indexes free for IOAPIC interrupts */
3609 table->min_index = 32;
3611 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3612 if (!table->table) {
3618 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3619 memset(table->table, 0,
3620 MAX_IRQS_PER_TABLE * sizeof(u32));
3622 memset(table->table, 0,
3623 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3628 for (i = 0; i < 32; ++i)
3629 iommu->irte_ops->set_allocated(table, i);
3632 irq_lookup_table[devid] = table;
3633 set_dte_irq_entry(devid, table);
3634 iommu_flush_dte(iommu, devid);
3635 if (devid != alias) {
3636 irq_lookup_table[alias] = table;
3637 set_dte_irq_entry(alias, table);
3638 iommu_flush_dte(iommu, alias);
3642 iommu_completion_wait(iommu);
3645 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3650 static int alloc_irq_index(u16 devid, int count, bool align)
3652 struct irq_remap_table *table;
3653 int index, c, alignment = 1;
3654 unsigned long flags;
3655 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3660 table = get_irq_table(devid, false);
3665 alignment = roundup_pow_of_two(count);
3667 spin_lock_irqsave(&table->lock, flags);
3669 /* Scan table for free entries */
3670 for (index = ALIGN(table->min_index, alignment), c = 0;
3671 index < MAX_IRQS_PER_TABLE;) {
3672 if (!iommu->irte_ops->is_allocated(table, index)) {
3676 index = ALIGN(index + 1, alignment);
3682 iommu->irte_ops->set_allocated(table, index - c + 1);
3694 spin_unlock_irqrestore(&table->lock, flags);
3699 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3700 struct amd_ir_data *data)
3702 struct irq_remap_table *table;
3703 struct amd_iommu *iommu;
3704 unsigned long flags;
3705 struct irte_ga *entry;
3707 iommu = amd_iommu_rlookup_table[devid];
3711 table = get_irq_table(devid, false);
3715 spin_lock_irqsave(&table->lock, flags);
3717 entry = (struct irte_ga *)table->table;
3718 entry = &entry[index];
3719 entry->lo.fields_remap.valid = 0;
3720 entry->hi.val = irte->hi.val;
3721 entry->lo.val = irte->lo.val;
3722 entry->lo.fields_remap.valid = 1;
3726 spin_unlock_irqrestore(&table->lock, flags);
3728 iommu_flush_irt(iommu, devid);
3729 iommu_completion_wait(iommu);
3734 static int modify_irte(u16 devid, int index, union irte *irte)
3736 struct irq_remap_table *table;
3737 struct amd_iommu *iommu;
3738 unsigned long flags;
3740 iommu = amd_iommu_rlookup_table[devid];
3744 table = get_irq_table(devid, false);
3748 spin_lock_irqsave(&table->lock, flags);
3749 table->table[index] = irte->val;
3750 spin_unlock_irqrestore(&table->lock, flags);
3752 iommu_flush_irt(iommu, devid);
3753 iommu_completion_wait(iommu);
3758 static void free_irte(u16 devid, int index)
3760 struct irq_remap_table *table;
3761 struct amd_iommu *iommu;
3762 unsigned long flags;
3764 iommu = amd_iommu_rlookup_table[devid];
3768 table = get_irq_table(devid, false);
3772 spin_lock_irqsave(&table->lock, flags);
3773 iommu->irte_ops->clear_allocated(table, index);
3774 spin_unlock_irqrestore(&table->lock, flags);
3776 iommu_flush_irt(iommu, devid);
3777 iommu_completion_wait(iommu);
3780 static void irte_prepare(void *entry,
3781 u32 delivery_mode, u32 dest_mode,
3782 u8 vector, u32 dest_apicid, int devid)
3784 union irte *irte = (union irte *) entry;
3787 irte->fields.vector = vector;
3788 irte->fields.int_type = delivery_mode;
3789 irte->fields.destination = dest_apicid;
3790 irte->fields.dm = dest_mode;
3791 irte->fields.valid = 1;
3794 static void irte_ga_prepare(void *entry,
3795 u32 delivery_mode, u32 dest_mode,
3796 u8 vector, u32 dest_apicid, int devid)
3798 struct irte_ga *irte = (struct irte_ga *) entry;
3802 irte->lo.fields_remap.int_type = delivery_mode;
3803 irte->lo.fields_remap.dm = dest_mode;
3804 irte->hi.fields.vector = vector;
3805 irte->lo.fields_remap.destination = dest_apicid;
3806 irte->lo.fields_remap.valid = 1;
3809 static void irte_activate(void *entry, u16 devid, u16 index)
3811 union irte *irte = (union irte *) entry;
3813 irte->fields.valid = 1;
3814 modify_irte(devid, index, irte);
3817 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3819 struct irte_ga *irte = (struct irte_ga *) entry;
3821 irte->lo.fields_remap.valid = 1;
3822 modify_irte_ga(devid, index, irte, NULL);
3825 static void irte_deactivate(void *entry, u16 devid, u16 index)
3827 union irte *irte = (union irte *) entry;
3829 irte->fields.valid = 0;
3830 modify_irte(devid, index, irte);
3833 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3835 struct irte_ga *irte = (struct irte_ga *) entry;
3837 irte->lo.fields_remap.valid = 0;
3838 modify_irte_ga(devid, index, irte, NULL);
3841 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3842 u8 vector, u32 dest_apicid)
3844 union irte *irte = (union irte *) entry;
3846 irte->fields.vector = vector;
3847 irte->fields.destination = dest_apicid;
3848 modify_irte(devid, index, irte);
3851 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3852 u8 vector, u32 dest_apicid)
3854 struct irte_ga *irte = (struct irte_ga *) entry;
3855 struct iommu_dev_data *dev_data = search_dev_data(devid);
3857 if (!dev_data || !dev_data->use_vapic ||
3858 !irte->lo.fields_remap.guest_mode) {
3859 irte->hi.fields.vector = vector;
3860 irte->lo.fields_remap.destination = dest_apicid;
3861 modify_irte_ga(devid, index, irte, NULL);
3865 #define IRTE_ALLOCATED (~1U)
3866 static void irte_set_allocated(struct irq_remap_table *table, int index)
3868 table->table[index] = IRTE_ALLOCATED;
3871 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3873 struct irte_ga *ptr = (struct irte_ga *)table->table;
3874 struct irte_ga *irte = &ptr[index];
3876 memset(&irte->lo.val, 0, sizeof(u64));
3877 memset(&irte->hi.val, 0, sizeof(u64));
3878 irte->hi.fields.vector = 0xff;
3881 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3883 union irte *ptr = (union irte *)table->table;
3884 union irte *irte = &ptr[index];
3886 return irte->val != 0;
3889 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3891 struct irte_ga *ptr = (struct irte_ga *)table->table;
3892 struct irte_ga *irte = &ptr[index];
3894 return irte->hi.fields.vector != 0;
3897 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3899 table->table[index] = 0;
3902 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3904 struct irte_ga *ptr = (struct irte_ga *)table->table;
3905 struct irte_ga *irte = &ptr[index];
3907 memset(&irte->lo.val, 0, sizeof(u64));
3908 memset(&irte->hi.val, 0, sizeof(u64));
3911 static int get_devid(struct irq_alloc_info *info)
3915 switch (info->type) {
3916 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3917 devid = get_ioapic_devid(info->ioapic_id);
3919 case X86_IRQ_ALLOC_TYPE_HPET:
3920 devid = get_hpet_devid(info->hpet_id);
3922 case X86_IRQ_ALLOC_TYPE_MSI:
3923 case X86_IRQ_ALLOC_TYPE_MSIX:
3924 devid = get_device_id(&info->msi_dev->dev);
3934 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3936 struct amd_iommu *iommu;
3942 devid = get_devid(info);
3944 iommu = amd_iommu_rlookup_table[devid];
3946 return iommu->ir_domain;
3952 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3954 struct amd_iommu *iommu;
3960 switch (info->type) {
3961 case X86_IRQ_ALLOC_TYPE_MSI:
3962 case X86_IRQ_ALLOC_TYPE_MSIX:
3963 devid = get_device_id(&info->msi_dev->dev);
3967 iommu = amd_iommu_rlookup_table[devid];
3969 return iommu->msi_domain;
3978 struct irq_remap_ops amd_iommu_irq_ops = {
3979 .prepare = amd_iommu_prepare,
3980 .enable = amd_iommu_enable,
3981 .disable = amd_iommu_disable,
3982 .reenable = amd_iommu_reenable,
3983 .enable_faulting = amd_iommu_enable_faulting,
3984 .get_ir_irq_domain = get_ir_irq_domain,
3985 .get_irq_domain = get_irq_domain,
3988 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3989 struct irq_cfg *irq_cfg,
3990 struct irq_alloc_info *info,
3991 int devid, int index, int sub_handle)
3993 struct irq_2_irte *irte_info = &data->irq_2_irte;
3994 struct msi_msg *msg = &data->msi_entry;
3995 struct IO_APIC_route_entry *entry;
3996 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4001 data->irq_2_irte.devid = devid;
4002 data->irq_2_irte.index = index + sub_handle;
4003 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4004 apic->irq_dest_mode, irq_cfg->vector,
4005 irq_cfg->dest_apicid, devid);
4007 switch (info->type) {
4008 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4009 /* Setup IOAPIC entry */
4010 entry = info->ioapic_entry;
4011 info->ioapic_entry = NULL;
4012 memset(entry, 0, sizeof(*entry));
4013 entry->vector = index;
4015 entry->trigger = info->ioapic_trigger;
4016 entry->polarity = info->ioapic_polarity;
4017 /* Mask level triggered irqs. */
4018 if (info->ioapic_trigger)
4022 case X86_IRQ_ALLOC_TYPE_HPET:
4023 case X86_IRQ_ALLOC_TYPE_MSI:
4024 case X86_IRQ_ALLOC_TYPE_MSIX:
4025 msg->address_hi = MSI_ADDR_BASE_HI;
4026 msg->address_lo = MSI_ADDR_BASE_LO;
4027 msg->data = irte_info->index;
4036 struct amd_irte_ops irte_32_ops = {
4037 .prepare = irte_prepare,
4038 .activate = irte_activate,
4039 .deactivate = irte_deactivate,
4040 .set_affinity = irte_set_affinity,
4041 .set_allocated = irte_set_allocated,
4042 .is_allocated = irte_is_allocated,
4043 .clear_allocated = irte_clear_allocated,
4046 struct amd_irte_ops irte_128_ops = {
4047 .prepare = irte_ga_prepare,
4048 .activate = irte_ga_activate,
4049 .deactivate = irte_ga_deactivate,
4050 .set_affinity = irte_ga_set_affinity,
4051 .set_allocated = irte_ga_set_allocated,
4052 .is_allocated = irte_ga_is_allocated,
4053 .clear_allocated = irte_ga_clear_allocated,
4056 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4057 unsigned int nr_irqs, void *arg)
4059 struct irq_alloc_info *info = arg;
4060 struct irq_data *irq_data;
4061 struct amd_ir_data *data = NULL;
4062 struct irq_cfg *cfg;
4068 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4069 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4073 * With IRQ remapping enabled, don't need contiguous CPU vectors
4074 * to support multiple MSI interrupts.
4076 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4077 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4079 devid = get_devid(info);
4083 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4087 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4088 if (get_irq_table(devid, true))
4089 index = info->ioapic_pin;
4093 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4095 index = alloc_irq_index(devid, nr_irqs, align);
4098 pr_warn("Failed to allocate IRTE\n");
4100 goto out_free_parent;
4103 for (i = 0; i < nr_irqs; i++) {
4104 irq_data = irq_domain_get_irq_data(domain, virq + i);
4105 cfg = irqd_cfg(irq_data);
4106 if (!irq_data || !cfg) {
4112 data = kzalloc(sizeof(*data), GFP_KERNEL);
4116 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4117 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4119 data->entry = kzalloc(sizeof(struct irte_ga),
4126 irq_data->hwirq = (devid << 16) + i;
4127 irq_data->chip_data = data;
4128 irq_data->chip = &amd_ir_chip;
4129 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4130 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4136 for (i--; i >= 0; i--) {
4137 irq_data = irq_domain_get_irq_data(domain, virq + i);
4139 kfree(irq_data->chip_data);
4141 for (i = 0; i < nr_irqs; i++)
4142 free_irte(devid, index + i);
4144 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4148 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4149 unsigned int nr_irqs)
4151 struct irq_2_irte *irte_info;
4152 struct irq_data *irq_data;
4153 struct amd_ir_data *data;
4156 for (i = 0; i < nr_irqs; i++) {
4157 irq_data = irq_domain_get_irq_data(domain, virq + i);
4158 if (irq_data && irq_data->chip_data) {
4159 data = irq_data->chip_data;
4160 irte_info = &data->irq_2_irte;
4161 free_irte(irte_info->devid, irte_info->index);
4166 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4169 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4170 struct amd_ir_data *ir_data,
4171 struct irq_2_irte *irte_info,
4172 struct irq_cfg *cfg);
4174 static int irq_remapping_activate(struct irq_domain *domain,
4175 struct irq_data *irq_data, bool reserve)
4177 struct amd_ir_data *data = irq_data->chip_data;
4178 struct irq_2_irte *irte_info = &data->irq_2_irte;
4179 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4180 struct irq_cfg *cfg = irqd_cfg(irq_data);
4185 iommu->irte_ops->activate(data->entry, irte_info->devid,
4187 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4191 static void irq_remapping_deactivate(struct irq_domain *domain,
4192 struct irq_data *irq_data)
4194 struct amd_ir_data *data = irq_data->chip_data;
4195 struct irq_2_irte *irte_info = &data->irq_2_irte;
4196 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4199 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4203 static const struct irq_domain_ops amd_ir_domain_ops = {
4204 .alloc = irq_remapping_alloc,
4205 .free = irq_remapping_free,
4206 .activate = irq_remapping_activate,
4207 .deactivate = irq_remapping_deactivate,
4210 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4212 struct amd_iommu *iommu;
4213 struct amd_iommu_pi_data *pi_data = vcpu_info;
4214 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4215 struct amd_ir_data *ir_data = data->chip_data;
4216 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4217 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4218 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4221 * This device has never been set up for guest mode.
4222 * we should not modify the IRTE
4224 if (!dev_data || !dev_data->use_vapic)
4227 pi_data->ir_data = ir_data;
4230 * SVM tries to set up for VAPIC mode, but we are in
4231 * legacy mode. So, we force legacy mode instead.
4233 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4234 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4236 pi_data->is_guest_mode = false;
4239 iommu = amd_iommu_rlookup_table[irte_info->devid];
4243 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4244 if (pi_data->is_guest_mode) {
4246 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4247 irte->hi.fields.vector = vcpu_pi_info->vector;
4248 irte->lo.fields_vapic.ga_log_intr = 1;
4249 irte->lo.fields_vapic.guest_mode = 1;
4250 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4252 ir_data->cached_ga_tag = pi_data->ga_tag;
4255 struct irq_cfg *cfg = irqd_cfg(data);
4259 irte->hi.fields.vector = cfg->vector;
4260 irte->lo.fields_remap.guest_mode = 0;
4261 irte->lo.fields_remap.destination = cfg->dest_apicid;
4262 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4263 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4266 * This communicates the ga_tag back to the caller
4267 * so that it can do all the necessary clean up.
4269 ir_data->cached_ga_tag = 0;
4272 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4276 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4277 struct amd_ir_data *ir_data,
4278 struct irq_2_irte *irte_info,
4279 struct irq_cfg *cfg)
4283 * Atomically updates the IRTE with the new destination, vector
4284 * and flushes the interrupt entry cache.
4286 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4287 irte_info->index, cfg->vector,
4291 static int amd_ir_set_affinity(struct irq_data *data,
4292 const struct cpumask *mask, bool force)
4294 struct amd_ir_data *ir_data = data->chip_data;
4295 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4296 struct irq_cfg *cfg = irqd_cfg(data);
4297 struct irq_data *parent = data->parent_data;
4298 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4304 ret = parent->chip->irq_set_affinity(parent, mask, force);
4305 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4308 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4310 * After this point, all the interrupts will start arriving
4311 * at the new destination. So, time to cleanup the previous
4312 * vector allocation.
4314 send_cleanup_vector(cfg);
4316 return IRQ_SET_MASK_OK_DONE;
4319 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4321 struct amd_ir_data *ir_data = irq_data->chip_data;
4323 *msg = ir_data->msi_entry;
4326 static struct irq_chip amd_ir_chip = {
4328 .irq_ack = ir_ack_apic_edge,
4329 .irq_set_affinity = amd_ir_set_affinity,
4330 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4331 .irq_compose_msi_msg = ir_compose_msi_msg,
4334 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4336 struct fwnode_handle *fn;
4338 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4341 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4342 irq_domain_free_fwnode(fn);
4343 if (!iommu->ir_domain)
4346 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4347 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4353 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4355 unsigned long flags;
4356 struct amd_iommu *iommu;
4357 struct irq_remap_table *irt;
4358 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4359 int devid = ir_data->irq_2_irte.devid;
4360 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4361 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4363 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4364 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4367 iommu = amd_iommu_rlookup_table[devid];
4371 irt = get_irq_table(devid, false);
4375 spin_lock_irqsave(&irt->lock, flags);
4377 if (ref->lo.fields_vapic.guest_mode) {
4379 ref->lo.fields_vapic.destination = cpu;
4380 ref->lo.fields_vapic.is_run = is_run;
4384 spin_unlock_irqrestore(&irt->lock, flags);
4386 iommu_flush_irt(iommu, devid);
4387 iommu_completion_wait(iommu);
4390 EXPORT_SYMBOL(amd_iommu_update_ga);