2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/platform_device.h>
25 #include <linux/pci-ats.h>
26 #include <linux/bitmap.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/scatterlist.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/iommu-helper.h>
32 #include <linux/iommu.h>
33 #include <linux/delay.h>
34 #include <linux/amd-iommu.h>
35 #include <linux/notifier.h>
36 #include <linux/export.h>
37 #include <linux/irq.h>
38 #include <linux/msi.h>
39 #include <linux/dma-contiguous.h>
40 #include <linux/irqdomain.h>
41 #include <linux/percpu.h>
42 #include <linux/iova.h>
43 #include <asm/irq_remapping.h>
44 #include <asm/io_apic.h>
46 #include <asm/hw_irq.h>
47 #include <asm/msidef.h>
48 #include <asm/proto.h>
49 #include <asm/iommu.h>
53 #include "amd_iommu_proto.h"
54 #include "amd_iommu_types.h"
55 #include "irq_remapping.h"
57 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
59 #define LOOP_TIMEOUT 100000
61 /* IO virtual address start page frame number */
62 #define IOVA_START_PFN (1)
63 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
64 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
66 /* Reserved IOVA ranges */
67 #define MSI_RANGE_START (0xfee00000)
68 #define MSI_RANGE_END (0xfeefffff)
69 #define HT_RANGE_START (0xfd00000000ULL)
70 #define HT_RANGE_END (0xffffffffffULL)
73 * This bitmap is used to advertise the page sizes our hardware support
74 * to the IOMMU core, which will then use this information to split
75 * physically contiguous memory regions it is mapping into page sizes
78 * 512GB Pages are not supported due to a hardware bug
80 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
82 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
84 /* List of all available dev_data structures */
85 static LIST_HEAD(dev_data_list);
86 static DEFINE_SPINLOCK(dev_data_list_lock);
88 LIST_HEAD(ioapic_map);
90 LIST_HEAD(acpihid_map);
92 #define FLUSH_QUEUE_SIZE 256
94 struct flush_queue_entry {
95 unsigned long iova_pfn;
97 struct dma_ops_domain *dma_dom;
103 struct flush_queue_entry *entries;
106 DEFINE_PER_CPU(struct flush_queue, flush_queue);
109 * Domain for untranslated devices - only allocated
110 * if iommu=pt passed on kernel cmd line.
112 static const struct iommu_ops amd_iommu_ops;
114 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
115 int amd_iommu_max_glx_val = -1;
117 static struct dma_map_ops amd_iommu_dma_ops;
120 * This struct contains device specific data for the IOMMU
122 struct iommu_dev_data {
123 struct list_head list; /* For domain->dev_list */
124 struct list_head dev_data_list; /* For global dev_data_list */
125 struct protection_domain *domain; /* Domain the device is bound to */
126 u16 devid; /* PCI Device ID */
127 u16 alias; /* Alias Device ID */
128 bool iommu_v2; /* Device can make use of IOMMUv2 */
129 bool passthrough; /* Device is identity mapped */
133 } ats; /* ATS state */
134 bool pri_tlp; /* PASID TLB required for
136 u32 errata; /* Bitmap for errata to apply */
140 * general struct to manage commands send to an IOMMU
146 struct kmem_cache *amd_iommu_irq_cache;
148 static void update_domain(struct protection_domain *domain);
149 static int protection_domain_init(struct protection_domain *domain);
150 static void detach_device(struct device *dev);
153 * Data container for a dma_ops specific protection domain
155 struct dma_ops_domain {
156 /* generic protection domain information */
157 struct protection_domain domain;
160 struct iova_domain iovad;
163 static struct iova_domain reserved_iova_ranges;
164 static struct lock_class_key reserved_rbtree_key;
166 /****************************************************************************
170 ****************************************************************************/
172 static inline int match_hid_uid(struct device *dev,
173 struct acpihid_map_entry *entry)
175 const char *hid, *uid;
177 hid = acpi_device_hid(ACPI_COMPANION(dev));
178 uid = acpi_device_uid(ACPI_COMPANION(dev));
184 return strcmp(hid, entry->hid);
187 return strcmp(hid, entry->hid);
189 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
192 static inline u16 get_pci_device_id(struct device *dev)
194 struct pci_dev *pdev = to_pci_dev(dev);
196 return PCI_DEVID(pdev->bus->number, pdev->devfn);
199 static inline int get_acpihid_device_id(struct device *dev,
200 struct acpihid_map_entry **entry)
202 struct acpihid_map_entry *p;
204 list_for_each_entry(p, &acpihid_map, list) {
205 if (!match_hid_uid(dev, p)) {
214 static inline int get_device_id(struct device *dev)
219 devid = get_pci_device_id(dev);
221 devid = get_acpihid_device_id(dev, NULL);
226 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
228 return container_of(dom, struct protection_domain, domain);
231 static struct iommu_dev_data *alloc_dev_data(u16 devid)
233 struct iommu_dev_data *dev_data;
236 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
240 dev_data->devid = devid;
242 spin_lock_irqsave(&dev_data_list_lock, flags);
243 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
244 spin_unlock_irqrestore(&dev_data_list_lock, flags);
249 static struct iommu_dev_data *search_dev_data(u16 devid)
251 struct iommu_dev_data *dev_data;
254 spin_lock_irqsave(&dev_data_list_lock, flags);
255 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
256 if (dev_data->devid == devid)
263 spin_unlock_irqrestore(&dev_data_list_lock, flags);
268 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
270 *(u16 *)data = alias;
274 static u16 get_alias(struct device *dev)
276 struct pci_dev *pdev = to_pci_dev(dev);
277 u16 devid, ivrs_alias, pci_alias;
279 /* The callers make sure that get_device_id() does not fail here */
280 devid = get_device_id(dev);
281 ivrs_alias = amd_iommu_alias_table[devid];
282 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
284 if (ivrs_alias == pci_alias)
290 * The IVRS is fairly reliable in telling us about aliases, but it
291 * can't know about every screwy device. If we don't have an IVRS
292 * reported alias, use the PCI reported alias. In that case we may
293 * still need to initialize the rlookup and dev_table entries if the
294 * alias is to a non-existent device.
296 if (ivrs_alias == devid) {
297 if (!amd_iommu_rlookup_table[pci_alias]) {
298 amd_iommu_rlookup_table[pci_alias] =
299 amd_iommu_rlookup_table[devid];
300 memcpy(amd_iommu_dev_table[pci_alias].data,
301 amd_iommu_dev_table[devid].data,
302 sizeof(amd_iommu_dev_table[pci_alias].data));
308 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
309 "for device %s[%04x:%04x], kernel reported alias "
310 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
311 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
312 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
313 PCI_FUNC(pci_alias));
316 * If we don't have a PCI DMA alias and the IVRS alias is on the same
317 * bus, then the IVRS table may know about a quirk that we don't.
319 if (pci_alias == devid &&
320 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
321 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
322 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
323 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
330 static struct iommu_dev_data *find_dev_data(u16 devid)
332 struct iommu_dev_data *dev_data;
334 dev_data = search_dev_data(devid);
336 if (dev_data == NULL)
337 dev_data = alloc_dev_data(devid);
342 static struct iommu_dev_data *get_dev_data(struct device *dev)
344 return dev->archdata.iommu;
348 * Find or create an IOMMU group for a acpihid device.
350 static struct iommu_group *acpihid_device_group(struct device *dev)
352 struct acpihid_map_entry *p, *entry = NULL;
355 devid = get_acpihid_device_id(dev, &entry);
357 return ERR_PTR(devid);
359 list_for_each_entry(p, &acpihid_map, list) {
360 if ((devid == p->devid) && p->group)
361 entry->group = p->group;
365 entry->group = generic_device_group(dev);
370 static bool pci_iommuv2_capable(struct pci_dev *pdev)
372 static const int caps[] = {
375 PCI_EXT_CAP_ID_PASID,
379 for (i = 0; i < 3; ++i) {
380 pos = pci_find_ext_capability(pdev, caps[i]);
388 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
390 struct iommu_dev_data *dev_data;
392 dev_data = get_dev_data(&pdev->dev);
394 return dev_data->errata & (1 << erratum) ? true : false;
398 * This function checks if the driver got a valid device from the caller to
399 * avoid dereferencing invalid pointers.
401 static bool check_device(struct device *dev)
405 if (!dev || !dev->dma_mask)
408 devid = get_device_id(dev);
412 /* Out of our scope? */
413 if (devid > amd_iommu_last_bdf)
416 if (amd_iommu_rlookup_table[devid] == NULL)
422 static void init_iommu_group(struct device *dev)
424 struct iommu_group *group;
426 group = iommu_group_get_for_dev(dev);
430 iommu_group_put(group);
433 static int iommu_init_device(struct device *dev)
435 struct iommu_dev_data *dev_data;
438 if (dev->archdata.iommu)
441 devid = get_device_id(dev);
445 dev_data = find_dev_data(devid);
449 dev_data->alias = get_alias(dev);
451 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
452 struct amd_iommu *iommu;
454 iommu = amd_iommu_rlookup_table[dev_data->devid];
455 dev_data->iommu_v2 = iommu->is_iommu_v2;
458 dev->archdata.iommu = dev_data;
460 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
466 static void iommu_ignore_device(struct device *dev)
471 devid = get_device_id(dev);
475 alias = get_alias(dev);
477 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
478 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
480 amd_iommu_rlookup_table[devid] = NULL;
481 amd_iommu_rlookup_table[alias] = NULL;
484 static void iommu_uninit_device(struct device *dev)
487 struct iommu_dev_data *dev_data;
489 devid = get_device_id(dev);
493 dev_data = search_dev_data(devid);
497 if (dev_data->domain)
500 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
503 iommu_group_remove_device(dev);
506 dev->archdata.dma_ops = NULL;
509 * We keep dev_data around for unplugged devices and reuse it when the
510 * device is re-plugged - not doing so would introduce a ton of races.
514 /****************************************************************************
516 * Interrupt handling functions
518 ****************************************************************************/
520 static void dump_dte_entry(u16 devid)
524 for (i = 0; i < 4; ++i)
525 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
526 amd_iommu_dev_table[devid].data[i]);
529 static void dump_command(unsigned long phys_addr)
531 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
534 for (i = 0; i < 4; ++i)
535 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
538 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
540 int type, devid, domid, flags;
541 volatile u32 *event = __evt;
546 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
547 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
548 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
549 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
550 address = (u64)(((u64)event[3]) << 32) | event[2];
553 /* Did we hit the erratum? */
554 if (++count == LOOP_TIMEOUT) {
555 pr_err("AMD-Vi: No event written to event log\n");
562 printk(KERN_ERR "AMD-Vi: Event logged [");
565 case EVENT_TYPE_ILL_DEV:
566 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
567 "address=0x%016llx flags=0x%04x]\n",
568 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
570 dump_dte_entry(devid);
572 case EVENT_TYPE_IO_FAULT:
573 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
574 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
575 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
576 domid, address, flags);
578 case EVENT_TYPE_DEV_TAB_ERR:
579 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
580 "address=0x%016llx flags=0x%04x]\n",
581 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
584 case EVENT_TYPE_PAGE_TAB_ERR:
585 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
586 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
588 domid, address, flags);
590 case EVENT_TYPE_ILL_CMD:
591 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
592 dump_command(address);
594 case EVENT_TYPE_CMD_HARD_ERR:
595 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
596 "flags=0x%04x]\n", address, flags);
598 case EVENT_TYPE_IOTLB_INV_TO:
599 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
600 "address=0x%016llx]\n",
601 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
604 case EVENT_TYPE_INV_DEV_REQ:
605 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
606 "address=0x%016llx flags=0x%04x]\n",
607 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
611 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
614 memset(__evt, 0, 4 * sizeof(u32));
617 static void iommu_poll_events(struct amd_iommu *iommu)
621 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
622 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
624 while (head != tail) {
625 iommu_print_event(iommu, iommu->evt_buf + head);
626 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
629 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
632 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
634 struct amd_iommu_fault fault;
636 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
637 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
641 fault.address = raw[1];
642 fault.pasid = PPR_PASID(raw[0]);
643 fault.device_id = PPR_DEVID(raw[0]);
644 fault.tag = PPR_TAG(raw[0]);
645 fault.flags = PPR_FLAGS(raw[0]);
647 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
650 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
654 if (iommu->ppr_log == NULL)
657 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
658 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
660 while (head != tail) {
665 raw = (u64 *)(iommu->ppr_log + head);
668 * Hardware bug: Interrupt may arrive before the entry is
669 * written to memory. If this happens we need to wait for the
672 for (i = 0; i < LOOP_TIMEOUT; ++i) {
673 if (PPR_REQ_TYPE(raw[0]) != 0)
678 /* Avoid memcpy function-call overhead */
683 * To detect the hardware bug we need to clear the entry
686 raw[0] = raw[1] = 0UL;
688 /* Update head pointer of hardware ring-buffer */
689 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
690 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
692 /* Handle PPR entry */
693 iommu_handle_ppr_entry(iommu, entry);
695 /* Refresh ring-buffer information */
696 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
697 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
701 irqreturn_t amd_iommu_int_thread(int irq, void *data)
703 struct amd_iommu *iommu = (struct amd_iommu *) data;
704 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
706 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
707 /* Enable EVT and PPR interrupts again */
708 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
709 iommu->mmio_base + MMIO_STATUS_OFFSET);
711 if (status & MMIO_STATUS_EVT_INT_MASK) {
712 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
713 iommu_poll_events(iommu);
716 if (status & MMIO_STATUS_PPR_INT_MASK) {
717 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
718 iommu_poll_ppr_log(iommu);
722 * Hardware bug: ERBT1312
723 * When re-enabling interrupt (by writing 1
724 * to clear the bit), the hardware might also try to set
725 * the interrupt bit in the event status register.
726 * In this scenario, the bit will be set, and disable
727 * subsequent interrupts.
729 * Workaround: The IOMMU driver should read back the
730 * status register and check if the interrupt bits are cleared.
731 * If not, driver will need to go through the interrupt handler
732 * again and re-clear the bits
734 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
739 irqreturn_t amd_iommu_int_handler(int irq, void *data)
741 return IRQ_WAKE_THREAD;
744 /****************************************************************************
746 * IOMMU command queuing functions
748 ****************************************************************************/
750 static int wait_on_sem(volatile u64 *sem)
754 while (*sem == 0 && i < LOOP_TIMEOUT) {
759 if (i == LOOP_TIMEOUT) {
760 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
767 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
768 struct iommu_cmd *cmd,
773 target = iommu->cmd_buf + tail;
774 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
776 /* Copy command to buffer */
777 memcpy(target, cmd, sizeof(*cmd));
779 /* Tell the IOMMU about it */
780 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
783 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
785 WARN_ON(address & 0x7ULL);
787 memset(cmd, 0, sizeof(*cmd));
788 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
789 cmd->data[1] = upper_32_bits(__pa(address));
791 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
794 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
796 memset(cmd, 0, sizeof(*cmd));
797 cmd->data[0] = devid;
798 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
801 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
802 size_t size, u16 domid, int pde)
807 pages = iommu_num_pages(address, size, PAGE_SIZE);
812 * If we have to flush more than one page, flush all
813 * TLB entries for this domain
815 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
819 address &= PAGE_MASK;
821 memset(cmd, 0, sizeof(*cmd));
822 cmd->data[1] |= domid;
823 cmd->data[2] = lower_32_bits(address);
824 cmd->data[3] = upper_32_bits(address);
825 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
826 if (s) /* size bit - we flush more than one 4kb page */
827 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
828 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
829 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
832 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
833 u64 address, size_t size)
838 pages = iommu_num_pages(address, size, PAGE_SIZE);
843 * If we have to flush more than one page, flush all
844 * TLB entries for this domain
846 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
850 address &= PAGE_MASK;
852 memset(cmd, 0, sizeof(*cmd));
853 cmd->data[0] = devid;
854 cmd->data[0] |= (qdep & 0xff) << 24;
855 cmd->data[1] = devid;
856 cmd->data[2] = lower_32_bits(address);
857 cmd->data[3] = upper_32_bits(address);
858 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
860 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
863 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
864 u64 address, bool size)
866 memset(cmd, 0, sizeof(*cmd));
868 address &= ~(0xfffULL);
870 cmd->data[0] = pasid;
871 cmd->data[1] = domid;
872 cmd->data[2] = lower_32_bits(address);
873 cmd->data[3] = upper_32_bits(address);
874 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
875 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
877 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
878 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
881 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
882 int qdep, u64 address, bool size)
884 memset(cmd, 0, sizeof(*cmd));
886 address &= ~(0xfffULL);
888 cmd->data[0] = devid;
889 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
890 cmd->data[0] |= (qdep & 0xff) << 24;
891 cmd->data[1] = devid;
892 cmd->data[1] |= (pasid & 0xff) << 16;
893 cmd->data[2] = lower_32_bits(address);
894 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
895 cmd->data[3] = upper_32_bits(address);
897 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
898 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
901 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
902 int status, int tag, bool gn)
904 memset(cmd, 0, sizeof(*cmd));
906 cmd->data[0] = devid;
908 cmd->data[1] = pasid;
909 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
911 cmd->data[3] = tag & 0x1ff;
912 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
914 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
917 static void build_inv_all(struct iommu_cmd *cmd)
919 memset(cmd, 0, sizeof(*cmd));
920 CMD_SET_TYPE(cmd, CMD_INV_ALL);
923 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
925 memset(cmd, 0, sizeof(*cmd));
926 cmd->data[0] = devid;
927 CMD_SET_TYPE(cmd, CMD_INV_IRT);
931 * Writes the command to the IOMMUs command buffer and informs the
932 * hardware about the new command.
934 static int iommu_queue_command_sync(struct amd_iommu *iommu,
935 struct iommu_cmd *cmd,
938 u32 left, tail, head, next_tail;
942 spin_lock_irqsave(&iommu->lock, flags);
944 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
945 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
946 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
947 left = (head - next_tail) % CMD_BUFFER_SIZE;
950 struct iommu_cmd sync_cmd;
951 volatile u64 sem = 0;
954 build_completion_wait(&sync_cmd, (u64)&sem);
955 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
957 spin_unlock_irqrestore(&iommu->lock, flags);
959 if ((ret = wait_on_sem(&sem)) != 0)
965 copy_cmd_to_buffer(iommu, cmd, tail);
967 /* We need to sync now to make sure all commands are processed */
968 iommu->need_sync = sync;
970 spin_unlock_irqrestore(&iommu->lock, flags);
975 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
977 return iommu_queue_command_sync(iommu, cmd, true);
981 * This function queues a completion wait command into the command
984 static int iommu_completion_wait(struct amd_iommu *iommu)
986 struct iommu_cmd cmd;
987 volatile u64 sem = 0;
990 if (!iommu->need_sync)
993 build_completion_wait(&cmd, (u64)&sem);
995 ret = iommu_queue_command_sync(iommu, &cmd, false);
999 return wait_on_sem(&sem);
1002 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1004 struct iommu_cmd cmd;
1006 build_inv_dte(&cmd, devid);
1008 return iommu_queue_command(iommu, &cmd);
1011 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1015 for (devid = 0; devid <= 0xffff; ++devid)
1016 iommu_flush_dte(iommu, devid);
1018 iommu_completion_wait(iommu);
1022 * This function uses heavy locking and may disable irqs for some time. But
1023 * this is no issue because it is only called during resume.
1025 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1029 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1030 struct iommu_cmd cmd;
1031 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1033 iommu_queue_command(iommu, &cmd);
1036 iommu_completion_wait(iommu);
1039 static void iommu_flush_all(struct amd_iommu *iommu)
1041 struct iommu_cmd cmd;
1043 build_inv_all(&cmd);
1045 iommu_queue_command(iommu, &cmd);
1046 iommu_completion_wait(iommu);
1049 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1051 struct iommu_cmd cmd;
1053 build_inv_irt(&cmd, devid);
1055 iommu_queue_command(iommu, &cmd);
1058 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1062 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1063 iommu_flush_irt(iommu, devid);
1065 iommu_completion_wait(iommu);
1068 void iommu_flush_all_caches(struct amd_iommu *iommu)
1070 if (iommu_feature(iommu, FEATURE_IA)) {
1071 iommu_flush_all(iommu);
1073 iommu_flush_dte_all(iommu);
1074 iommu_flush_irt_all(iommu);
1075 iommu_flush_tlb_all(iommu);
1080 * Command send function for flushing on-device TLB
1082 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1083 u64 address, size_t size)
1085 struct amd_iommu *iommu;
1086 struct iommu_cmd cmd;
1089 qdep = dev_data->ats.qdep;
1090 iommu = amd_iommu_rlookup_table[dev_data->devid];
1092 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1094 return iommu_queue_command(iommu, &cmd);
1098 * Command send function for invalidating a device table entry
1100 static int device_flush_dte(struct iommu_dev_data *dev_data)
1102 struct amd_iommu *iommu;
1106 iommu = amd_iommu_rlookup_table[dev_data->devid];
1107 alias = dev_data->alias;
1109 ret = iommu_flush_dte(iommu, dev_data->devid);
1110 if (!ret && alias != dev_data->devid)
1111 ret = iommu_flush_dte(iommu, alias);
1115 if (dev_data->ats.enabled)
1116 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1122 * TLB invalidation function which is called from the mapping functions.
1123 * It invalidates a single PTE if the range to flush is within a single
1124 * page. Otherwise it flushes the whole TLB of the IOMMU.
1126 static void __domain_flush_pages(struct protection_domain *domain,
1127 u64 address, size_t size, int pde)
1129 struct iommu_dev_data *dev_data;
1130 struct iommu_cmd cmd;
1133 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1135 for (i = 0; i < amd_iommus_present; ++i) {
1136 if (!domain->dev_iommu[i])
1140 * Devices of this domain are behind this IOMMU
1141 * We need a TLB flush
1143 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1146 list_for_each_entry(dev_data, &domain->dev_list, list) {
1148 if (!dev_data->ats.enabled)
1151 ret |= device_flush_iotlb(dev_data, address, size);
1157 static void domain_flush_pages(struct protection_domain *domain,
1158 u64 address, size_t size)
1160 __domain_flush_pages(domain, address, size, 0);
1163 /* Flush the whole IO/TLB for a given protection domain */
1164 static void domain_flush_tlb(struct protection_domain *domain)
1166 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1169 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1170 static void domain_flush_tlb_pde(struct protection_domain *domain)
1172 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1175 static void domain_flush_complete(struct protection_domain *domain)
1179 for (i = 0; i < amd_iommus_present; ++i) {
1180 if (domain && !domain->dev_iommu[i])
1184 * Devices of this domain are behind this IOMMU
1185 * We need to wait for completion of all commands.
1187 iommu_completion_wait(amd_iommus[i]);
1193 * This function flushes the DTEs for all devices in domain
1195 static void domain_flush_devices(struct protection_domain *domain)
1197 struct iommu_dev_data *dev_data;
1199 list_for_each_entry(dev_data, &domain->dev_list, list)
1200 device_flush_dte(dev_data);
1203 /****************************************************************************
1205 * The functions below are used the create the page table mappings for
1206 * unity mapped regions.
1208 ****************************************************************************/
1211 * This function is used to add another level to an IO page table. Adding
1212 * another level increases the size of the address space by 9 bits to a size up
1215 static bool increase_address_space(struct protection_domain *domain,
1220 if (domain->mode == PAGE_MODE_6_LEVEL)
1221 /* address space already 64 bit large */
1224 pte = (void *)get_zeroed_page(gfp);
1228 *pte = PM_LEVEL_PDE(domain->mode,
1229 virt_to_phys(domain->pt_root));
1230 domain->pt_root = pte;
1232 domain->updated = true;
1237 static u64 *alloc_pte(struct protection_domain *domain,
1238 unsigned long address,
1239 unsigned long page_size,
1246 BUG_ON(!is_power_of_2(page_size));
1248 while (address > PM_LEVEL_SIZE(domain->mode))
1249 increase_address_space(domain, gfp);
1251 level = domain->mode - 1;
1252 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1253 address = PAGE_SIZE_ALIGN(address, page_size);
1254 end_lvl = PAGE_SIZE_LEVEL(page_size);
1256 while (level > end_lvl) {
1261 if (!IOMMU_PTE_PRESENT(__pte)) {
1262 page = (u64 *)get_zeroed_page(gfp);
1266 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1268 if (cmpxchg64(pte, __pte, __npte)) {
1269 free_page((unsigned long)page);
1274 /* No level skipping support yet */
1275 if (PM_PTE_LEVEL(*pte) != level)
1280 pte = IOMMU_PTE_PAGE(*pte);
1282 if (pte_page && level == end_lvl)
1285 pte = &pte[PM_LEVEL_INDEX(level, address)];
1292 * This function checks if there is a PTE for a given dma address. If
1293 * there is one, it returns the pointer to it.
1295 static u64 *fetch_pte(struct protection_domain *domain,
1296 unsigned long address,
1297 unsigned long *page_size)
1302 if (address > PM_LEVEL_SIZE(domain->mode))
1305 level = domain->mode - 1;
1306 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1307 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1312 if (!IOMMU_PTE_PRESENT(*pte))
1316 if (PM_PTE_LEVEL(*pte) == 7 ||
1317 PM_PTE_LEVEL(*pte) == 0)
1320 /* No level skipping support yet */
1321 if (PM_PTE_LEVEL(*pte) != level)
1326 /* Walk to the next level */
1327 pte = IOMMU_PTE_PAGE(*pte);
1328 pte = &pte[PM_LEVEL_INDEX(level, address)];
1329 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1332 if (PM_PTE_LEVEL(*pte) == 0x07) {
1333 unsigned long pte_mask;
1336 * If we have a series of large PTEs, make
1337 * sure to return a pointer to the first one.
1339 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1340 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1341 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1348 * Generic mapping functions. It maps a physical address into a DMA
1349 * address space. It allocates the page table pages if necessary.
1350 * In the future it can be extended to a generic mapping function
1351 * supporting all features of AMD IOMMU page tables like level skipping
1352 * and full 64 bit address spaces.
1354 static int iommu_map_page(struct protection_domain *dom,
1355 unsigned long bus_addr,
1356 unsigned long phys_addr,
1357 unsigned long page_size,
1364 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1365 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1367 if (!(prot & IOMMU_PROT_MASK))
1370 count = PAGE_SIZE_PTE_COUNT(page_size);
1371 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1376 for (i = 0; i < count; ++i)
1377 if (IOMMU_PTE_PRESENT(pte[i]))
1381 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1382 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1384 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1386 if (prot & IOMMU_PROT_IR)
1387 __pte |= IOMMU_PTE_IR;
1388 if (prot & IOMMU_PROT_IW)
1389 __pte |= IOMMU_PTE_IW;
1391 for (i = 0; i < count; ++i)
1399 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1400 unsigned long bus_addr,
1401 unsigned long page_size)
1403 unsigned long long unmapped;
1404 unsigned long unmap_size;
1407 BUG_ON(!is_power_of_2(page_size));
1411 while (unmapped < page_size) {
1413 pte = fetch_pte(dom, bus_addr, &unmap_size);
1418 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1419 for (i = 0; i < count; i++)
1423 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1424 unmapped += unmap_size;
1427 BUG_ON(unmapped && !is_power_of_2(unmapped));
1432 /****************************************************************************
1434 * The next functions belong to the address allocator for the dma_ops
1435 * interface functions.
1437 ****************************************************************************/
1440 static unsigned long dma_ops_alloc_iova(struct device *dev,
1441 struct dma_ops_domain *dma_dom,
1442 unsigned int pages, u64 dma_mask)
1444 unsigned long pfn = 0;
1446 pages = __roundup_pow_of_two(pages);
1448 if (dma_mask > DMA_BIT_MASK(32))
1449 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1450 IOVA_PFN(DMA_BIT_MASK(32)));
1453 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1455 return (pfn << PAGE_SHIFT);
1458 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1459 unsigned long address,
1462 pages = __roundup_pow_of_two(pages);
1463 address >>= PAGE_SHIFT;
1465 free_iova_fast(&dma_dom->iovad, address, pages);
1468 /****************************************************************************
1470 * The next functions belong to the domain allocation. A domain is
1471 * allocated for every IOMMU as the default domain. If device isolation
1472 * is enabled, every device get its own domain. The most important thing
1473 * about domains is the page table mapping the DMA address space they
1476 ****************************************************************************/
1479 * This function adds a protection domain to the global protection domain list
1481 static void add_domain_to_list(struct protection_domain *domain)
1483 unsigned long flags;
1485 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1486 list_add(&domain->list, &amd_iommu_pd_list);
1487 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1491 * This function removes a protection domain to the global
1492 * protection domain list
1494 static void del_domain_from_list(struct protection_domain *domain)
1496 unsigned long flags;
1498 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1499 list_del(&domain->list);
1500 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1503 static u16 domain_id_alloc(void)
1505 unsigned long flags;
1508 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1509 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1511 if (id > 0 && id < MAX_DOMAIN_ID)
1512 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1515 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1520 static void domain_id_free(int id)
1522 unsigned long flags;
1524 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1525 if (id > 0 && id < MAX_DOMAIN_ID)
1526 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1527 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1530 #define DEFINE_FREE_PT_FN(LVL, FN) \
1531 static void free_pt_##LVL (unsigned long __pt) \
1539 for (i = 0; i < 512; ++i) { \
1540 /* PTE present? */ \
1541 if (!IOMMU_PTE_PRESENT(pt[i])) \
1545 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1546 PM_PTE_LEVEL(pt[i]) == 7) \
1549 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1552 free_page((unsigned long)pt); \
1555 DEFINE_FREE_PT_FN(l2, free_page)
1556 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1557 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1558 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1559 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1561 static void free_pagetable(struct protection_domain *domain)
1563 unsigned long root = (unsigned long)domain->pt_root;
1565 switch (domain->mode) {
1566 case PAGE_MODE_NONE:
1568 case PAGE_MODE_1_LEVEL:
1571 case PAGE_MODE_2_LEVEL:
1574 case PAGE_MODE_3_LEVEL:
1577 case PAGE_MODE_4_LEVEL:
1580 case PAGE_MODE_5_LEVEL:
1583 case PAGE_MODE_6_LEVEL:
1591 static void free_gcr3_tbl_level1(u64 *tbl)
1596 for (i = 0; i < 512; ++i) {
1597 if (!(tbl[i] & GCR3_VALID))
1600 ptr = __va(tbl[i] & PAGE_MASK);
1602 free_page((unsigned long)ptr);
1606 static void free_gcr3_tbl_level2(u64 *tbl)
1611 for (i = 0; i < 512; ++i) {
1612 if (!(tbl[i] & GCR3_VALID))
1615 ptr = __va(tbl[i] & PAGE_MASK);
1617 free_gcr3_tbl_level1(ptr);
1621 static void free_gcr3_table(struct protection_domain *domain)
1623 if (domain->glx == 2)
1624 free_gcr3_tbl_level2(domain->gcr3_tbl);
1625 else if (domain->glx == 1)
1626 free_gcr3_tbl_level1(domain->gcr3_tbl);
1628 BUG_ON(domain->glx != 0);
1630 free_page((unsigned long)domain->gcr3_tbl);
1634 * Free a domain, only used if something went wrong in the
1635 * allocation path and we need to free an already allocated page table
1637 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1642 del_domain_from_list(&dom->domain);
1644 put_iova_domain(&dom->iovad);
1646 free_pagetable(&dom->domain);
1652 * Allocates a new protection domain usable for the dma_ops functions.
1653 * It also initializes the page table and the address allocator data
1654 * structures required for the dma_ops interface
1656 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1658 struct dma_ops_domain *dma_dom;
1660 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1664 if (protection_domain_init(&dma_dom->domain))
1667 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1668 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1669 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1670 dma_dom->domain.priv = dma_dom;
1671 if (!dma_dom->domain.pt_root)
1674 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1675 IOVA_START_PFN, DMA_32BIT_PFN);
1677 /* Initialize reserved ranges */
1678 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1680 add_domain_to_list(&dma_dom->domain);
1685 dma_ops_domain_free(dma_dom);
1691 * little helper function to check whether a given protection domain is a
1694 static bool dma_ops_domain(struct protection_domain *domain)
1696 return domain->flags & PD_DMA_OPS_MASK;
1699 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1704 if (domain->mode != PAGE_MODE_NONE)
1705 pte_root = virt_to_phys(domain->pt_root);
1707 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1708 << DEV_ENTRY_MODE_SHIFT;
1709 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1711 flags = amd_iommu_dev_table[devid].data[1];
1714 flags |= DTE_FLAG_IOTLB;
1716 if (domain->flags & PD_IOMMUV2_MASK) {
1717 u64 gcr3 = __pa(domain->gcr3_tbl);
1718 u64 glx = domain->glx;
1721 pte_root |= DTE_FLAG_GV;
1722 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1724 /* First mask out possible old values for GCR3 table */
1725 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1728 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1731 /* Encode GCR3 table into DTE */
1732 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1735 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1738 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1742 flags &= ~(0xffffUL);
1743 flags |= domain->id;
1745 amd_iommu_dev_table[devid].data[1] = flags;
1746 amd_iommu_dev_table[devid].data[0] = pte_root;
1749 static void clear_dte_entry(u16 devid)
1751 /* remove entry from the device table seen by the hardware */
1752 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1753 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1755 amd_iommu_apply_erratum_63(devid);
1758 static void do_attach(struct iommu_dev_data *dev_data,
1759 struct protection_domain *domain)
1761 struct amd_iommu *iommu;
1765 iommu = amd_iommu_rlookup_table[dev_data->devid];
1766 alias = dev_data->alias;
1767 ats = dev_data->ats.enabled;
1769 /* Update data structures */
1770 dev_data->domain = domain;
1771 list_add(&dev_data->list, &domain->dev_list);
1773 /* Do reference counting */
1774 domain->dev_iommu[iommu->index] += 1;
1775 domain->dev_cnt += 1;
1777 /* Update device table */
1778 set_dte_entry(dev_data->devid, domain, ats);
1779 if (alias != dev_data->devid)
1780 set_dte_entry(alias, domain, ats);
1782 device_flush_dte(dev_data);
1785 static void do_detach(struct iommu_dev_data *dev_data)
1787 struct amd_iommu *iommu;
1791 * First check if the device is still attached. It might already
1792 * be detached from its domain because the generic
1793 * iommu_detach_group code detached it and we try again here in
1794 * our alias handling.
1796 if (!dev_data->domain)
1799 iommu = amd_iommu_rlookup_table[dev_data->devid];
1800 alias = dev_data->alias;
1802 /* decrease reference counters */
1803 dev_data->domain->dev_iommu[iommu->index] -= 1;
1804 dev_data->domain->dev_cnt -= 1;
1806 /* Update data structures */
1807 dev_data->domain = NULL;
1808 list_del(&dev_data->list);
1809 clear_dte_entry(dev_data->devid);
1810 if (alias != dev_data->devid)
1811 clear_dte_entry(alias);
1813 /* Flush the DTE entry */
1814 device_flush_dte(dev_data);
1818 * If a device is not yet associated with a domain, this function does
1819 * assigns it visible for the hardware
1821 static int __attach_device(struct iommu_dev_data *dev_data,
1822 struct protection_domain *domain)
1827 * Must be called with IRQs disabled. Warn here to detect early
1830 WARN_ON(!irqs_disabled());
1833 spin_lock(&domain->lock);
1836 if (dev_data->domain != NULL)
1839 /* Attach alias group root */
1840 do_attach(dev_data, domain);
1847 spin_unlock(&domain->lock);
1853 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1855 pci_disable_ats(pdev);
1856 pci_disable_pri(pdev);
1857 pci_disable_pasid(pdev);
1860 /* FIXME: Change generic reset-function to do the same */
1861 static int pri_reset_while_enabled(struct pci_dev *pdev)
1866 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1870 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1871 control |= PCI_PRI_CTRL_RESET;
1872 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1877 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1882 /* FIXME: Hardcode number of outstanding requests for now */
1884 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
1886 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
1888 /* Only allow access to user-accessible pages */
1889 ret = pci_enable_pasid(pdev, 0);
1893 /* First reset the PRI state of the device */
1894 ret = pci_reset_pri(pdev);
1899 ret = pci_enable_pri(pdev, reqs);
1904 ret = pri_reset_while_enabled(pdev);
1909 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1916 pci_disable_pri(pdev);
1917 pci_disable_pasid(pdev);
1922 /* FIXME: Move this to PCI code */
1923 #define PCI_PRI_TLP_OFF (1 << 15)
1925 static bool pci_pri_tlp_required(struct pci_dev *pdev)
1930 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1934 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
1936 return (status & PCI_PRI_TLP_OFF) ? true : false;
1940 * If a device is not yet associated with a domain, this function
1941 * assigns it visible for the hardware
1943 static int attach_device(struct device *dev,
1944 struct protection_domain *domain)
1946 struct pci_dev *pdev;
1947 struct iommu_dev_data *dev_data;
1948 unsigned long flags;
1951 dev_data = get_dev_data(dev);
1953 if (!dev_is_pci(dev))
1954 goto skip_ats_check;
1956 pdev = to_pci_dev(dev);
1957 if (domain->flags & PD_IOMMUV2_MASK) {
1958 if (!dev_data->passthrough)
1961 if (dev_data->iommu_v2) {
1962 if (pdev_iommuv2_enable(pdev) != 0)
1965 dev_data->ats.enabled = true;
1966 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1967 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
1969 } else if (amd_iommu_iotlb_sup &&
1970 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1971 dev_data->ats.enabled = true;
1972 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1976 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1977 ret = __attach_device(dev_data, domain);
1978 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1981 * We might boot into a crash-kernel here. The crashed kernel
1982 * left the caches in the IOMMU dirty. So we have to flush
1983 * here to evict all dirty stuff.
1985 domain_flush_tlb_pde(domain);
1991 * Removes a device from a protection domain (unlocked)
1993 static void __detach_device(struct iommu_dev_data *dev_data)
1995 struct protection_domain *domain;
1998 * Must be called with IRQs disabled. Warn here to detect early
2001 WARN_ON(!irqs_disabled());
2003 if (WARN_ON(!dev_data->domain))
2006 domain = dev_data->domain;
2008 spin_lock(&domain->lock);
2010 do_detach(dev_data);
2012 spin_unlock(&domain->lock);
2016 * Removes a device from a protection domain (with devtable_lock held)
2018 static void detach_device(struct device *dev)
2020 struct protection_domain *domain;
2021 struct iommu_dev_data *dev_data;
2022 unsigned long flags;
2024 dev_data = get_dev_data(dev);
2025 domain = dev_data->domain;
2027 /* lock device table */
2028 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2029 __detach_device(dev_data);
2030 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2032 if (!dev_is_pci(dev))
2035 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2036 pdev_iommuv2_disable(to_pci_dev(dev));
2037 else if (dev_data->ats.enabled)
2038 pci_disable_ats(to_pci_dev(dev));
2040 dev_data->ats.enabled = false;
2043 static int amd_iommu_add_device(struct device *dev)
2045 struct iommu_dev_data *dev_data;
2046 struct iommu_domain *domain;
2047 struct amd_iommu *iommu;
2050 if (!check_device(dev) || get_dev_data(dev))
2053 devid = get_device_id(dev);
2057 iommu = amd_iommu_rlookup_table[devid];
2059 ret = iommu_init_device(dev);
2061 if (ret != -ENOTSUPP)
2062 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2065 iommu_ignore_device(dev);
2066 dev->archdata.dma_ops = &nommu_dma_ops;
2069 init_iommu_group(dev);
2071 dev_data = get_dev_data(dev);
2075 if (iommu_pass_through || dev_data->iommu_v2)
2076 iommu_request_dm_for_dev(dev);
2078 /* Domains are initialized for this device - have a look what we ended up with */
2079 domain = iommu_get_domain_for_dev(dev);
2080 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2081 dev_data->passthrough = true;
2083 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2086 iommu_completion_wait(iommu);
2091 static void amd_iommu_remove_device(struct device *dev)
2093 struct amd_iommu *iommu;
2096 if (!check_device(dev))
2099 devid = get_device_id(dev);
2103 iommu = amd_iommu_rlookup_table[devid];
2105 iommu_uninit_device(dev);
2106 iommu_completion_wait(iommu);
2109 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2111 if (dev_is_pci(dev))
2112 return pci_device_group(dev);
2114 return acpihid_device_group(dev);
2117 /*****************************************************************************
2119 * The next functions belong to the dma_ops mapping/unmapping code.
2121 *****************************************************************************/
2123 static void __queue_flush(struct flush_queue *queue)
2125 struct protection_domain *domain;
2126 unsigned long flags;
2129 /* First flush TLB of all known domains */
2130 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
2131 list_for_each_entry(domain, &amd_iommu_pd_list, list)
2132 domain_flush_tlb(domain);
2133 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
2135 /* Wait until flushes have completed */
2136 domain_flush_complete(NULL);
2138 for (idx = 0; idx < queue->next; ++idx) {
2139 struct flush_queue_entry *entry;
2141 entry = queue->entries + idx;
2143 free_iova_fast(&entry->dma_dom->iovad,
2147 /* Not really necessary, just to make sure we catch any bugs */
2148 entry->dma_dom = NULL;
2154 static void queue_add(struct dma_ops_domain *dma_dom,
2155 unsigned long address, unsigned long pages)
2157 struct flush_queue_entry *entry;
2158 struct flush_queue *queue;
2159 unsigned long flags;
2162 pages = __roundup_pow_of_two(pages);
2163 address >>= PAGE_SHIFT;
2165 queue = get_cpu_ptr(&flush_queue);
2166 spin_lock_irqsave(&queue->lock, flags);
2168 if (queue->next == FLUSH_QUEUE_SIZE)
2169 __queue_flush(queue);
2171 idx = queue->next++;
2172 entry = queue->entries + idx;
2174 entry->iova_pfn = address;
2175 entry->pages = pages;
2176 entry->dma_dom = dma_dom;
2178 spin_unlock_irqrestore(&queue->lock, flags);
2179 put_cpu_ptr(&flush_queue);
2184 * In the dma_ops path we only have the struct device. This function
2185 * finds the corresponding IOMMU, the protection domain and the
2186 * requestor id for a given device.
2187 * If the device is not yet associated with a domain this is also done
2190 static struct protection_domain *get_domain(struct device *dev)
2192 struct protection_domain *domain;
2193 struct iommu_domain *io_domain;
2195 if (!check_device(dev))
2196 return ERR_PTR(-EINVAL);
2198 io_domain = iommu_get_domain_for_dev(dev);
2202 domain = to_pdomain(io_domain);
2203 if (!dma_ops_domain(domain))
2204 return ERR_PTR(-EBUSY);
2209 static void update_device_table(struct protection_domain *domain)
2211 struct iommu_dev_data *dev_data;
2213 list_for_each_entry(dev_data, &domain->dev_list, list)
2214 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2217 static void update_domain(struct protection_domain *domain)
2219 if (!domain->updated)
2222 update_device_table(domain);
2224 domain_flush_devices(domain);
2225 domain_flush_tlb_pde(domain);
2227 domain->updated = false;
2231 * This function contains common code for mapping of a physically
2232 * contiguous memory region into DMA address space. It is used by all
2233 * mapping functions provided with this IOMMU driver.
2234 * Must be called with the domain lock held.
2236 static dma_addr_t __map_single(struct device *dev,
2237 struct dma_ops_domain *dma_dom,
2243 dma_addr_t offset = paddr & ~PAGE_MASK;
2244 dma_addr_t address, start, ret;
2249 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2252 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2253 if (address == DMA_ERROR_CODE)
2256 if (direction == DMA_TO_DEVICE)
2257 prot = IOMMU_PROT_IR;
2258 else if (direction == DMA_FROM_DEVICE)
2259 prot = IOMMU_PROT_IW;
2260 else if (direction == DMA_BIDIRECTIONAL)
2261 prot = IOMMU_PROT_IW | IOMMU_PROT_IR;
2264 for (i = 0; i < pages; ++i) {
2265 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2266 PAGE_SIZE, prot, GFP_ATOMIC);
2275 if (unlikely(amd_iommu_np_cache)) {
2276 domain_flush_pages(&dma_dom->domain, address, size);
2277 domain_flush_complete(&dma_dom->domain);
2285 for (--i; i >= 0; --i) {
2287 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2290 domain_flush_tlb(&dma_dom->domain);
2291 domain_flush_complete(&dma_dom->domain);
2293 dma_ops_free_iova(dma_dom, address, pages);
2295 return DMA_ERROR_CODE;
2299 * Does the reverse of the __map_single function. Must be called with
2300 * the domain lock held too
2302 static void __unmap_single(struct dma_ops_domain *dma_dom,
2303 dma_addr_t dma_addr,
2307 dma_addr_t flush_addr;
2308 dma_addr_t i, start;
2311 flush_addr = dma_addr;
2312 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2313 dma_addr &= PAGE_MASK;
2316 for (i = 0; i < pages; ++i) {
2317 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2321 if (amd_iommu_unmap_flush) {
2322 dma_ops_free_iova(dma_dom, dma_addr, pages);
2323 domain_flush_tlb(&dma_dom->domain);
2324 domain_flush_complete(&dma_dom->domain);
2326 queue_add(dma_dom, dma_addr, pages);
2331 * The exported map_single function for dma_ops.
2333 static dma_addr_t map_page(struct device *dev, struct page *page,
2334 unsigned long offset, size_t size,
2335 enum dma_data_direction dir,
2336 struct dma_attrs *attrs)
2338 phys_addr_t paddr = page_to_phys(page) + offset;
2339 struct protection_domain *domain;
2342 domain = get_domain(dev);
2343 if (PTR_ERR(domain) == -EINVAL)
2344 return (dma_addr_t)paddr;
2345 else if (IS_ERR(domain))
2346 return DMA_ERROR_CODE;
2348 dma_mask = *dev->dma_mask;
2350 return __map_single(dev, domain->priv, paddr, size, dir, dma_mask);
2354 * The exported unmap_single function for dma_ops.
2356 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2357 enum dma_data_direction dir, struct dma_attrs *attrs)
2359 struct protection_domain *domain;
2361 domain = get_domain(dev);
2365 __unmap_single(domain->priv, dma_addr, size, dir);
2369 * The exported map_sg function for dma_ops (handles scatter-gather
2372 static int map_sg(struct device *dev, struct scatterlist *sglist,
2373 int nelems, enum dma_data_direction dir,
2374 struct dma_attrs *attrs)
2376 struct protection_domain *domain;
2378 struct scatterlist *s;
2380 int mapped_elems = 0;
2383 domain = get_domain(dev);
2387 dma_mask = *dev->dma_mask;
2389 for_each_sg(sglist, s, nelems, i) {
2392 s->dma_address = __map_single(dev, domain->priv,
2393 paddr, s->length, dir, dma_mask);
2395 if (s->dma_address) {
2396 s->dma_length = s->length;
2402 return mapped_elems;
2405 for_each_sg(sglist, s, mapped_elems, i) {
2407 __unmap_single(domain->priv, s->dma_address,
2408 s->dma_length, dir);
2409 s->dma_address = s->dma_length = 0;
2416 * The exported map_sg function for dma_ops (handles scatter-gather
2419 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2420 int nelems, enum dma_data_direction dir,
2421 struct dma_attrs *attrs)
2423 struct protection_domain *domain;
2424 struct scatterlist *s;
2427 domain = get_domain(dev);
2431 for_each_sg(sglist, s, nelems, i) {
2432 __unmap_single(domain->priv, s->dma_address,
2433 s->dma_length, dir);
2434 s->dma_address = s->dma_length = 0;
2439 * The exported alloc_coherent function for dma_ops.
2441 static void *alloc_coherent(struct device *dev, size_t size,
2442 dma_addr_t *dma_addr, gfp_t flag,
2443 struct dma_attrs *attrs)
2445 u64 dma_mask = dev->coherent_dma_mask;
2446 struct protection_domain *domain;
2449 domain = get_domain(dev);
2450 if (PTR_ERR(domain) == -EINVAL) {
2451 page = alloc_pages(flag, get_order(size));
2452 *dma_addr = page_to_phys(page);
2453 return page_address(page);
2454 } else if (IS_ERR(domain))
2457 size = PAGE_ALIGN(size);
2458 dma_mask = dev->coherent_dma_mask;
2459 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2462 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2464 if (!gfpflags_allow_blocking(flag))
2467 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2474 dma_mask = *dev->dma_mask;
2476 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2477 size, DMA_BIDIRECTIONAL, dma_mask);
2479 if (*dma_addr == DMA_ERROR_CODE)
2482 return page_address(page);
2486 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2487 __free_pages(page, get_order(size));
2493 * The exported free_coherent function for dma_ops.
2495 static void free_coherent(struct device *dev, size_t size,
2496 void *virt_addr, dma_addr_t dma_addr,
2497 struct dma_attrs *attrs)
2499 struct protection_domain *domain;
2502 page = virt_to_page(virt_addr);
2503 size = PAGE_ALIGN(size);
2505 domain = get_domain(dev);
2509 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2512 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2513 __free_pages(page, get_order(size));
2517 * This function is called by the DMA layer to find out if we can handle a
2518 * particular device. It is part of the dma_ops.
2520 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2522 return check_device(dev);
2525 static struct dma_map_ops amd_iommu_dma_ops = {
2526 .alloc = alloc_coherent,
2527 .free = free_coherent,
2528 .map_page = map_page,
2529 .unmap_page = unmap_page,
2531 .unmap_sg = unmap_sg,
2532 .dma_supported = amd_iommu_dma_supported,
2535 static int init_reserved_iova_ranges(void)
2537 struct pci_dev *pdev = NULL;
2540 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2541 IOVA_START_PFN, DMA_32BIT_PFN);
2543 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2544 &reserved_rbtree_key);
2546 /* MSI memory range */
2547 val = reserve_iova(&reserved_iova_ranges,
2548 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2550 pr_err("Reserving MSI range failed\n");
2554 /* HT memory range */
2555 val = reserve_iova(&reserved_iova_ranges,
2556 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2558 pr_err("Reserving HT range failed\n");
2563 * Memory used for PCI resources
2564 * FIXME: Check whether we can reserve the PCI-hole completly
2566 for_each_pci_dev(pdev) {
2569 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2570 struct resource *r = &pdev->resource[i];
2572 if (!(r->flags & IORESOURCE_MEM))
2575 val = reserve_iova(&reserved_iova_ranges,
2579 pr_err("Reserve pci-resource range failed\n");
2588 int __init amd_iommu_init_api(void)
2590 int ret, cpu, err = 0;
2592 ret = iova_cache_get();
2596 ret = init_reserved_iova_ranges();
2600 for_each_possible_cpu(cpu) {
2601 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2603 queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
2604 sizeof(*queue->entries),
2606 if (!queue->entries)
2609 spin_lock_init(&queue->lock);
2612 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2615 #ifdef CONFIG_ARM_AMBA
2616 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2620 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2626 for_each_possible_cpu(cpu) {
2627 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2629 kfree(queue->entries);
2635 int __init amd_iommu_init_dma_ops(void)
2637 swiotlb = iommu_pass_through ? 1 : 0;
2641 * In case we don't initialize SWIOTLB (actually the common case
2642 * when AMD IOMMU is enabled), make sure there are global
2643 * dma_ops set as a fall-back for devices not handled by this
2644 * driver (for example non-PCI devices).
2647 dma_ops = &nommu_dma_ops;
2649 if (amd_iommu_unmap_flush)
2650 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2652 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2658 /*****************************************************************************
2660 * The following functions belong to the exported interface of AMD IOMMU
2662 * This interface allows access to lower level functions of the IOMMU
2663 * like protection domain handling and assignement of devices to domains
2664 * which is not possible with the dma_ops interface.
2666 *****************************************************************************/
2668 static void cleanup_domain(struct protection_domain *domain)
2670 struct iommu_dev_data *entry;
2671 unsigned long flags;
2673 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2675 while (!list_empty(&domain->dev_list)) {
2676 entry = list_first_entry(&domain->dev_list,
2677 struct iommu_dev_data, list);
2678 __detach_device(entry);
2681 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2684 static void protection_domain_free(struct protection_domain *domain)
2689 del_domain_from_list(domain);
2692 domain_id_free(domain->id);
2697 static int protection_domain_init(struct protection_domain *domain)
2699 spin_lock_init(&domain->lock);
2700 mutex_init(&domain->api_lock);
2701 domain->id = domain_id_alloc();
2704 INIT_LIST_HEAD(&domain->dev_list);
2709 static struct protection_domain *protection_domain_alloc(void)
2711 struct protection_domain *domain;
2713 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2717 if (protection_domain_init(domain))
2720 add_domain_to_list(domain);
2730 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2732 struct protection_domain *pdomain;
2733 struct dma_ops_domain *dma_domain;
2736 case IOMMU_DOMAIN_UNMANAGED:
2737 pdomain = protection_domain_alloc();
2741 pdomain->mode = PAGE_MODE_3_LEVEL;
2742 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2743 if (!pdomain->pt_root) {
2744 protection_domain_free(pdomain);
2748 pdomain->domain.geometry.aperture_start = 0;
2749 pdomain->domain.geometry.aperture_end = ~0ULL;
2750 pdomain->domain.geometry.force_aperture = true;
2753 case IOMMU_DOMAIN_DMA:
2754 dma_domain = dma_ops_domain_alloc();
2756 pr_err("AMD-Vi: Failed to allocate\n");
2759 pdomain = &dma_domain->domain;
2761 case IOMMU_DOMAIN_IDENTITY:
2762 pdomain = protection_domain_alloc();
2766 pdomain->mode = PAGE_MODE_NONE;
2772 return &pdomain->domain;
2775 static void amd_iommu_domain_free(struct iommu_domain *dom)
2777 struct protection_domain *domain;
2782 domain = to_pdomain(dom);
2784 if (domain->dev_cnt > 0)
2785 cleanup_domain(domain);
2787 BUG_ON(domain->dev_cnt != 0);
2789 if (domain->mode != PAGE_MODE_NONE)
2790 free_pagetable(domain);
2792 if (domain->flags & PD_IOMMUV2_MASK)
2793 free_gcr3_table(domain);
2795 protection_domain_free(domain);
2798 static void amd_iommu_detach_device(struct iommu_domain *dom,
2801 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2802 struct amd_iommu *iommu;
2805 if (!check_device(dev))
2808 devid = get_device_id(dev);
2812 if (dev_data->domain != NULL)
2815 iommu = amd_iommu_rlookup_table[devid];
2819 iommu_completion_wait(iommu);
2822 static int amd_iommu_attach_device(struct iommu_domain *dom,
2825 struct protection_domain *domain = to_pdomain(dom);
2826 struct iommu_dev_data *dev_data;
2827 struct amd_iommu *iommu;
2830 if (!check_device(dev))
2833 dev_data = dev->archdata.iommu;
2835 iommu = amd_iommu_rlookup_table[dev_data->devid];
2839 if (dev_data->domain)
2842 ret = attach_device(dev, domain);
2844 iommu_completion_wait(iommu);
2849 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2850 phys_addr_t paddr, size_t page_size, int iommu_prot)
2852 struct protection_domain *domain = to_pdomain(dom);
2856 if (domain->mode == PAGE_MODE_NONE)
2859 if (iommu_prot & IOMMU_READ)
2860 prot |= IOMMU_PROT_IR;
2861 if (iommu_prot & IOMMU_WRITE)
2862 prot |= IOMMU_PROT_IW;
2864 mutex_lock(&domain->api_lock);
2865 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
2866 mutex_unlock(&domain->api_lock);
2871 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2874 struct protection_domain *domain = to_pdomain(dom);
2877 if (domain->mode == PAGE_MODE_NONE)
2880 mutex_lock(&domain->api_lock);
2881 unmap_size = iommu_unmap_page(domain, iova, page_size);
2882 mutex_unlock(&domain->api_lock);
2884 domain_flush_tlb_pde(domain);
2889 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2892 struct protection_domain *domain = to_pdomain(dom);
2893 unsigned long offset_mask, pte_pgsize;
2896 if (domain->mode == PAGE_MODE_NONE)
2899 pte = fetch_pte(domain, iova, &pte_pgsize);
2901 if (!pte || !IOMMU_PTE_PRESENT(*pte))
2904 offset_mask = pte_pgsize - 1;
2905 __pte = *pte & PM_ADDR_MASK;
2907 return (__pte & ~offset_mask) | (iova & offset_mask);
2910 static bool amd_iommu_capable(enum iommu_cap cap)
2913 case IOMMU_CAP_CACHE_COHERENCY:
2915 case IOMMU_CAP_INTR_REMAP:
2916 return (irq_remapping_enabled == 1);
2917 case IOMMU_CAP_NOEXEC:
2924 static void amd_iommu_get_dm_regions(struct device *dev,
2925 struct list_head *head)
2927 struct unity_map_entry *entry;
2930 devid = get_device_id(dev);
2934 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
2935 struct iommu_dm_region *region;
2937 if (devid < entry->devid_start || devid > entry->devid_end)
2940 region = kzalloc(sizeof(*region), GFP_KERNEL);
2942 pr_err("Out of memory allocating dm-regions for %s\n",
2947 region->start = entry->address_start;
2948 region->length = entry->address_end - entry->address_start;
2949 if (entry->prot & IOMMU_PROT_IR)
2950 region->prot |= IOMMU_READ;
2951 if (entry->prot & IOMMU_PROT_IW)
2952 region->prot |= IOMMU_WRITE;
2954 list_add_tail(®ion->list, head);
2958 static void amd_iommu_put_dm_regions(struct device *dev,
2959 struct list_head *head)
2961 struct iommu_dm_region *entry, *next;
2963 list_for_each_entry_safe(entry, next, head, list)
2967 static void amd_iommu_apply_dm_region(struct device *dev,
2968 struct iommu_domain *domain,
2969 struct iommu_dm_region *region)
2971 struct protection_domain *pdomain = to_pdomain(domain);
2972 struct dma_ops_domain *dma_dom = pdomain->priv;
2973 unsigned long start, end;
2975 start = IOVA_PFN(region->start);
2976 end = IOVA_PFN(region->start + region->length);
2978 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
2981 static const struct iommu_ops amd_iommu_ops = {
2982 .capable = amd_iommu_capable,
2983 .domain_alloc = amd_iommu_domain_alloc,
2984 .domain_free = amd_iommu_domain_free,
2985 .attach_dev = amd_iommu_attach_device,
2986 .detach_dev = amd_iommu_detach_device,
2987 .map = amd_iommu_map,
2988 .unmap = amd_iommu_unmap,
2989 .map_sg = default_iommu_map_sg,
2990 .iova_to_phys = amd_iommu_iova_to_phys,
2991 .add_device = amd_iommu_add_device,
2992 .remove_device = amd_iommu_remove_device,
2993 .device_group = amd_iommu_device_group,
2994 .get_dm_regions = amd_iommu_get_dm_regions,
2995 .put_dm_regions = amd_iommu_put_dm_regions,
2996 .apply_dm_region = amd_iommu_apply_dm_region,
2997 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3000 /*****************************************************************************
3002 * The next functions do a basic initialization of IOMMU for pass through
3005 * In passthrough mode the IOMMU is initialized and enabled but not used for
3006 * DMA-API translation.
3008 *****************************************************************************/
3010 /* IOMMUv2 specific functions */
3011 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3013 return atomic_notifier_chain_register(&ppr_notifier, nb);
3015 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3017 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3019 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3021 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3023 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3025 struct protection_domain *domain = to_pdomain(dom);
3026 unsigned long flags;
3028 spin_lock_irqsave(&domain->lock, flags);
3030 /* Update data structure */
3031 domain->mode = PAGE_MODE_NONE;
3032 domain->updated = true;
3034 /* Make changes visible to IOMMUs */
3035 update_domain(domain);
3037 /* Page-table is not visible to IOMMU anymore, so free it */
3038 free_pagetable(domain);
3040 spin_unlock_irqrestore(&domain->lock, flags);
3042 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3044 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3046 struct protection_domain *domain = to_pdomain(dom);
3047 unsigned long flags;
3050 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3053 /* Number of GCR3 table levels required */
3054 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3057 if (levels > amd_iommu_max_glx_val)
3060 spin_lock_irqsave(&domain->lock, flags);
3063 * Save us all sanity checks whether devices already in the
3064 * domain support IOMMUv2. Just force that the domain has no
3065 * devices attached when it is switched into IOMMUv2 mode.
3068 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3072 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3073 if (domain->gcr3_tbl == NULL)
3076 domain->glx = levels;
3077 domain->flags |= PD_IOMMUV2_MASK;
3078 domain->updated = true;
3080 update_domain(domain);
3085 spin_unlock_irqrestore(&domain->lock, flags);
3089 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3091 static int __flush_pasid(struct protection_domain *domain, int pasid,
3092 u64 address, bool size)
3094 struct iommu_dev_data *dev_data;
3095 struct iommu_cmd cmd;
3098 if (!(domain->flags & PD_IOMMUV2_MASK))
3101 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3104 * IOMMU TLB needs to be flushed before Device TLB to
3105 * prevent device TLB refill from IOMMU TLB
3107 for (i = 0; i < amd_iommus_present; ++i) {
3108 if (domain->dev_iommu[i] == 0)
3111 ret = iommu_queue_command(amd_iommus[i], &cmd);
3116 /* Wait until IOMMU TLB flushes are complete */
3117 domain_flush_complete(domain);
3119 /* Now flush device TLBs */
3120 list_for_each_entry(dev_data, &domain->dev_list, list) {
3121 struct amd_iommu *iommu;
3125 There might be non-IOMMUv2 capable devices in an IOMMUv2
3128 if (!dev_data->ats.enabled)
3131 qdep = dev_data->ats.qdep;
3132 iommu = amd_iommu_rlookup_table[dev_data->devid];
3134 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3135 qdep, address, size);
3137 ret = iommu_queue_command(iommu, &cmd);
3142 /* Wait until all device TLBs are flushed */
3143 domain_flush_complete(domain);
3152 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3155 return __flush_pasid(domain, pasid, address, false);
3158 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3161 struct protection_domain *domain = to_pdomain(dom);
3162 unsigned long flags;
3165 spin_lock_irqsave(&domain->lock, flags);
3166 ret = __amd_iommu_flush_page(domain, pasid, address);
3167 spin_unlock_irqrestore(&domain->lock, flags);
3171 EXPORT_SYMBOL(amd_iommu_flush_page);
3173 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3175 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3179 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3181 struct protection_domain *domain = to_pdomain(dom);
3182 unsigned long flags;
3185 spin_lock_irqsave(&domain->lock, flags);
3186 ret = __amd_iommu_flush_tlb(domain, pasid);
3187 spin_unlock_irqrestore(&domain->lock, flags);
3191 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3193 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3200 index = (pasid >> (9 * level)) & 0x1ff;
3206 if (!(*pte & GCR3_VALID)) {
3210 root = (void *)get_zeroed_page(GFP_ATOMIC);
3214 *pte = __pa(root) | GCR3_VALID;
3217 root = __va(*pte & PAGE_MASK);
3225 static int __set_gcr3(struct protection_domain *domain, int pasid,
3230 if (domain->mode != PAGE_MODE_NONE)
3233 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3237 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3239 return __amd_iommu_flush_tlb(domain, pasid);
3242 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3246 if (domain->mode != PAGE_MODE_NONE)
3249 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3255 return __amd_iommu_flush_tlb(domain, pasid);
3258 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3261 struct protection_domain *domain = to_pdomain(dom);
3262 unsigned long flags;
3265 spin_lock_irqsave(&domain->lock, flags);
3266 ret = __set_gcr3(domain, pasid, cr3);
3267 spin_unlock_irqrestore(&domain->lock, flags);
3271 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3273 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3275 struct protection_domain *domain = to_pdomain(dom);
3276 unsigned long flags;
3279 spin_lock_irqsave(&domain->lock, flags);
3280 ret = __clear_gcr3(domain, pasid);
3281 spin_unlock_irqrestore(&domain->lock, flags);
3285 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3287 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3288 int status, int tag)
3290 struct iommu_dev_data *dev_data;
3291 struct amd_iommu *iommu;
3292 struct iommu_cmd cmd;
3294 dev_data = get_dev_data(&pdev->dev);
3295 iommu = amd_iommu_rlookup_table[dev_data->devid];
3297 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3298 tag, dev_data->pri_tlp);
3300 return iommu_queue_command(iommu, &cmd);
3302 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3304 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3306 struct protection_domain *pdomain;
3308 pdomain = get_domain(&pdev->dev);
3309 if (IS_ERR(pdomain))
3312 /* Only return IOMMUv2 domains */
3313 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3316 return &pdomain->domain;
3318 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3320 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3322 struct iommu_dev_data *dev_data;
3324 if (!amd_iommu_v2_supported())
3327 dev_data = get_dev_data(&pdev->dev);
3328 dev_data->errata |= (1 << erratum);
3330 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3332 int amd_iommu_device_info(struct pci_dev *pdev,
3333 struct amd_iommu_device_info *info)
3338 if (pdev == NULL || info == NULL)
3341 if (!amd_iommu_v2_supported())
3344 memset(info, 0, sizeof(*info));
3346 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3348 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3350 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3352 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3354 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3358 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3359 max_pasids = min(max_pasids, (1 << 20));
3361 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3362 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3364 features = pci_pasid_features(pdev);
3365 if (features & PCI_PASID_CAP_EXEC)
3366 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3367 if (features & PCI_PASID_CAP_PRIV)
3368 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3373 EXPORT_SYMBOL(amd_iommu_device_info);
3375 #ifdef CONFIG_IRQ_REMAP
3377 /*****************************************************************************
3379 * Interrupt Remapping Implementation
3381 *****************************************************************************/
3399 u16 devid; /* Device ID for IRTE table */
3400 u16 index; /* Index into IRTE table*/
3403 struct amd_ir_data {
3404 struct irq_2_irte irq_2_irte;
3405 union irte irte_entry;
3407 struct msi_msg msi_entry;
3411 static struct irq_chip amd_ir_chip;
3413 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3414 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3415 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3416 #define DTE_IRQ_REMAP_ENABLE 1ULL
3418 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3422 dte = amd_iommu_dev_table[devid].data[2];
3423 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3424 dte |= virt_to_phys(table->table);
3425 dte |= DTE_IRQ_REMAP_INTCTL;
3426 dte |= DTE_IRQ_TABLE_LEN;
3427 dte |= DTE_IRQ_REMAP_ENABLE;
3429 amd_iommu_dev_table[devid].data[2] = dte;
3432 #define IRTE_ALLOCATED (~1U)
3434 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3436 struct irq_remap_table *table = NULL;
3437 struct amd_iommu *iommu;
3438 unsigned long flags;
3441 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3443 iommu = amd_iommu_rlookup_table[devid];
3447 table = irq_lookup_table[devid];
3451 alias = amd_iommu_alias_table[devid];
3452 table = irq_lookup_table[alias];
3454 irq_lookup_table[devid] = table;
3455 set_dte_irq_entry(devid, table);
3456 iommu_flush_dte(iommu, devid);
3460 /* Nothing there yet, allocate new irq remapping table */
3461 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3465 /* Initialize table spin-lock */
3466 spin_lock_init(&table->lock);
3469 /* Keep the first 32 indexes free for IOAPIC interrupts */
3470 table->min_index = 32;
3472 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3473 if (!table->table) {
3479 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3484 for (i = 0; i < 32; ++i)
3485 table->table[i] = IRTE_ALLOCATED;
3488 irq_lookup_table[devid] = table;
3489 set_dte_irq_entry(devid, table);
3490 iommu_flush_dte(iommu, devid);
3491 if (devid != alias) {
3492 irq_lookup_table[alias] = table;
3493 set_dte_irq_entry(alias, table);
3494 iommu_flush_dte(iommu, alias);
3498 iommu_completion_wait(iommu);
3501 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3506 static int alloc_irq_index(u16 devid, int count)
3508 struct irq_remap_table *table;
3509 unsigned long flags;
3512 table = get_irq_table(devid, false);
3516 spin_lock_irqsave(&table->lock, flags);
3518 /* Scan table for free entries */
3519 for (c = 0, index = table->min_index;
3520 index < MAX_IRQS_PER_TABLE;
3522 if (table->table[index] == 0)
3529 table->table[index - c + 1] = IRTE_ALLOCATED;
3539 spin_unlock_irqrestore(&table->lock, flags);
3544 static int modify_irte(u16 devid, int index, union irte irte)
3546 struct irq_remap_table *table;
3547 struct amd_iommu *iommu;
3548 unsigned long flags;
3550 iommu = amd_iommu_rlookup_table[devid];
3554 table = get_irq_table(devid, false);
3558 spin_lock_irqsave(&table->lock, flags);
3559 table->table[index] = irte.val;
3560 spin_unlock_irqrestore(&table->lock, flags);
3562 iommu_flush_irt(iommu, devid);
3563 iommu_completion_wait(iommu);
3568 static void free_irte(u16 devid, int index)
3570 struct irq_remap_table *table;
3571 struct amd_iommu *iommu;
3572 unsigned long flags;
3574 iommu = amd_iommu_rlookup_table[devid];
3578 table = get_irq_table(devid, false);
3582 spin_lock_irqsave(&table->lock, flags);
3583 table->table[index] = 0;
3584 spin_unlock_irqrestore(&table->lock, flags);
3586 iommu_flush_irt(iommu, devid);
3587 iommu_completion_wait(iommu);
3590 static int get_devid(struct irq_alloc_info *info)
3594 switch (info->type) {
3595 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3596 devid = get_ioapic_devid(info->ioapic_id);
3598 case X86_IRQ_ALLOC_TYPE_HPET:
3599 devid = get_hpet_devid(info->hpet_id);
3601 case X86_IRQ_ALLOC_TYPE_MSI:
3602 case X86_IRQ_ALLOC_TYPE_MSIX:
3603 devid = get_device_id(&info->msi_dev->dev);
3613 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3615 struct amd_iommu *iommu;
3621 devid = get_devid(info);
3623 iommu = amd_iommu_rlookup_table[devid];
3625 return iommu->ir_domain;
3631 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3633 struct amd_iommu *iommu;
3639 switch (info->type) {
3640 case X86_IRQ_ALLOC_TYPE_MSI:
3641 case X86_IRQ_ALLOC_TYPE_MSIX:
3642 devid = get_device_id(&info->msi_dev->dev);
3646 iommu = amd_iommu_rlookup_table[devid];
3648 return iommu->msi_domain;
3657 struct irq_remap_ops amd_iommu_irq_ops = {
3658 .prepare = amd_iommu_prepare,
3659 .enable = amd_iommu_enable,
3660 .disable = amd_iommu_disable,
3661 .reenable = amd_iommu_reenable,
3662 .enable_faulting = amd_iommu_enable_faulting,
3663 .get_ir_irq_domain = get_ir_irq_domain,
3664 .get_irq_domain = get_irq_domain,
3667 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3668 struct irq_cfg *irq_cfg,
3669 struct irq_alloc_info *info,
3670 int devid, int index, int sub_handle)
3672 struct irq_2_irte *irte_info = &data->irq_2_irte;
3673 struct msi_msg *msg = &data->msi_entry;
3674 union irte *irte = &data->irte_entry;
3675 struct IO_APIC_route_entry *entry;
3677 data->irq_2_irte.devid = devid;
3678 data->irq_2_irte.index = index + sub_handle;
3680 /* Setup IRTE for IOMMU */
3682 irte->fields.vector = irq_cfg->vector;
3683 irte->fields.int_type = apic->irq_delivery_mode;
3684 irte->fields.destination = irq_cfg->dest_apicid;
3685 irte->fields.dm = apic->irq_dest_mode;
3686 irte->fields.valid = 1;
3688 switch (info->type) {
3689 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3690 /* Setup IOAPIC entry */
3691 entry = info->ioapic_entry;
3692 info->ioapic_entry = NULL;
3693 memset(entry, 0, sizeof(*entry));
3694 entry->vector = index;
3696 entry->trigger = info->ioapic_trigger;
3697 entry->polarity = info->ioapic_polarity;
3698 /* Mask level triggered irqs. */
3699 if (info->ioapic_trigger)
3703 case X86_IRQ_ALLOC_TYPE_HPET:
3704 case X86_IRQ_ALLOC_TYPE_MSI:
3705 case X86_IRQ_ALLOC_TYPE_MSIX:
3706 msg->address_hi = MSI_ADDR_BASE_HI;
3707 msg->address_lo = MSI_ADDR_BASE_LO;
3708 msg->data = irte_info->index;
3717 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3718 unsigned int nr_irqs, void *arg)
3720 struct irq_alloc_info *info = arg;
3721 struct irq_data *irq_data;
3722 struct amd_ir_data *data;
3723 struct irq_cfg *cfg;
3729 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3730 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3734 * With IRQ remapping enabled, don't need contiguous CPU vectors
3735 * to support multiple MSI interrupts.
3737 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3738 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3740 devid = get_devid(info);
3744 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3748 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3749 if (get_irq_table(devid, true))
3750 index = info->ioapic_pin;
3754 index = alloc_irq_index(devid, nr_irqs);
3757 pr_warn("Failed to allocate IRTE\n");
3758 goto out_free_parent;
3761 for (i = 0; i < nr_irqs; i++) {
3762 irq_data = irq_domain_get_irq_data(domain, virq + i);
3763 cfg = irqd_cfg(irq_data);
3764 if (!irq_data || !cfg) {
3770 data = kzalloc(sizeof(*data), GFP_KERNEL);
3774 irq_data->hwirq = (devid << 16) + i;
3775 irq_data->chip_data = data;
3776 irq_data->chip = &amd_ir_chip;
3777 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3778 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3784 for (i--; i >= 0; i--) {
3785 irq_data = irq_domain_get_irq_data(domain, virq + i);
3787 kfree(irq_data->chip_data);
3789 for (i = 0; i < nr_irqs; i++)
3790 free_irte(devid, index + i);
3792 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3796 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3797 unsigned int nr_irqs)
3799 struct irq_2_irte *irte_info;
3800 struct irq_data *irq_data;
3801 struct amd_ir_data *data;
3804 for (i = 0; i < nr_irqs; i++) {
3805 irq_data = irq_domain_get_irq_data(domain, virq + i);
3806 if (irq_data && irq_data->chip_data) {
3807 data = irq_data->chip_data;
3808 irte_info = &data->irq_2_irte;
3809 free_irte(irte_info->devid, irte_info->index);
3813 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3816 static void irq_remapping_activate(struct irq_domain *domain,
3817 struct irq_data *irq_data)
3819 struct amd_ir_data *data = irq_data->chip_data;
3820 struct irq_2_irte *irte_info = &data->irq_2_irte;
3822 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3825 static void irq_remapping_deactivate(struct irq_domain *domain,
3826 struct irq_data *irq_data)
3828 struct amd_ir_data *data = irq_data->chip_data;
3829 struct irq_2_irte *irte_info = &data->irq_2_irte;
3833 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3836 static struct irq_domain_ops amd_ir_domain_ops = {
3837 .alloc = irq_remapping_alloc,
3838 .free = irq_remapping_free,
3839 .activate = irq_remapping_activate,
3840 .deactivate = irq_remapping_deactivate,
3843 static int amd_ir_set_affinity(struct irq_data *data,
3844 const struct cpumask *mask, bool force)
3846 struct amd_ir_data *ir_data = data->chip_data;
3847 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3848 struct irq_cfg *cfg = irqd_cfg(data);
3849 struct irq_data *parent = data->parent_data;
3852 ret = parent->chip->irq_set_affinity(parent, mask, force);
3853 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3857 * Atomically updates the IRTE with the new destination, vector
3858 * and flushes the interrupt entry cache.
3860 ir_data->irte_entry.fields.vector = cfg->vector;
3861 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
3862 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
3865 * After this point, all the interrupts will start arriving
3866 * at the new destination. So, time to cleanup the previous
3867 * vector allocation.
3869 send_cleanup_vector(cfg);
3871 return IRQ_SET_MASK_OK_DONE;
3874 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3876 struct amd_ir_data *ir_data = irq_data->chip_data;
3878 *msg = ir_data->msi_entry;
3881 static struct irq_chip amd_ir_chip = {
3882 .irq_ack = ir_ack_apic_edge,
3883 .irq_set_affinity = amd_ir_set_affinity,
3884 .irq_compose_msi_msg = ir_compose_msi_msg,
3887 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3889 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
3890 if (!iommu->ir_domain)
3893 iommu->ir_domain->parent = arch_get_ir_parent_domain();
3894 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);