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/pci-ats.h>
25 #include <linux/bitmap.h>
26 #include <linux/slab.h>
27 #include <linux/debugfs.h>
28 #include <linux/scatterlist.h>
29 #include <linux/dma-mapping.h>
30 #include <linux/iommu-helper.h>
31 #include <linux/iommu.h>
32 #include <linux/delay.h>
33 #include <linux/amd-iommu.h>
34 #include <linux/notifier.h>
35 #include <linux/export.h>
36 #include <linux/irq.h>
37 #include <linux/msi.h>
38 #include <linux/dma-contiguous.h>
39 #include <linux/irqdomain.h>
40 #include <linux/percpu.h>
41 #include <asm/irq_remapping.h>
42 #include <asm/io_apic.h>
44 #include <asm/hw_irq.h>
45 #include <asm/msidef.h>
46 #include <asm/proto.h>
47 #include <asm/iommu.h>
51 #include "amd_iommu_proto.h"
52 #include "amd_iommu_types.h"
53 #include "irq_remapping.h"
55 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
57 #define LOOP_TIMEOUT 100000
60 * This bitmap is used to advertise the page sizes our hardware support
61 * to the IOMMU core, which will then use this information to split
62 * physically contiguous memory regions it is mapping into page sizes
65 * 512GB Pages are not supported due to a hardware bug
67 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
69 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
71 /* List of all available dev_data structures */
72 static LIST_HEAD(dev_data_list);
73 static DEFINE_SPINLOCK(dev_data_list_lock);
75 LIST_HEAD(ioapic_map);
77 LIST_HEAD(acpihid_map);
80 * Domain for untranslated devices - only allocated
81 * if iommu=pt passed on kernel cmd line.
83 static const struct iommu_ops amd_iommu_ops;
85 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
86 int amd_iommu_max_glx_val = -1;
88 static struct dma_map_ops amd_iommu_dma_ops;
91 * This struct contains device specific data for the IOMMU
93 struct iommu_dev_data {
94 struct list_head list; /* For domain->dev_list */
95 struct list_head dev_data_list; /* For global dev_data_list */
96 struct protection_domain *domain; /* Domain the device is bound to */
97 u16 devid; /* PCI Device ID */
98 bool iommu_v2; /* Device can make use of IOMMUv2 */
99 bool passthrough; /* Device is identity mapped */
103 } ats; /* ATS state */
104 bool pri_tlp; /* PASID TLB required for
106 u32 errata; /* Bitmap for errata to apply */
110 * general struct to manage commands send to an IOMMU
116 struct kmem_cache *amd_iommu_irq_cache;
118 static void update_domain(struct protection_domain *domain);
119 static int protection_domain_init(struct protection_domain *domain);
120 static void detach_device(struct device *dev);
123 * For dynamic growth the aperture size is split into ranges of 128MB of
124 * DMA address space each. This struct represents one such range.
126 struct aperture_range {
128 spinlock_t bitmap_lock;
130 /* address allocation bitmap */
131 unsigned long *bitmap;
132 unsigned long offset;
133 unsigned long next_bit;
136 * Array of PTE pages for the aperture. In this array we save all the
137 * leaf pages of the domain page table used for the aperture. This way
138 * we don't need to walk the page table to find a specific PTE. We can
139 * just calculate its address in constant time.
145 * Data container for a dma_ops specific protection domain
147 struct dma_ops_domain {
148 /* generic protection domain information */
149 struct protection_domain domain;
151 /* size of the aperture for the mappings */
152 unsigned long aperture_size;
154 /* aperture index we start searching for free addresses */
155 u32 __percpu *next_index;
157 /* address space relevant data */
158 struct aperture_range *aperture[APERTURE_MAX_RANGES];
161 /****************************************************************************
165 ****************************************************************************/
167 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
169 return container_of(dom, struct protection_domain, domain);
172 static struct iommu_dev_data *alloc_dev_data(u16 devid)
174 struct iommu_dev_data *dev_data;
177 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
181 dev_data->devid = devid;
183 spin_lock_irqsave(&dev_data_list_lock, flags);
184 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
185 spin_unlock_irqrestore(&dev_data_list_lock, flags);
190 static struct iommu_dev_data *search_dev_data(u16 devid)
192 struct iommu_dev_data *dev_data;
195 spin_lock_irqsave(&dev_data_list_lock, flags);
196 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
197 if (dev_data->devid == devid)
204 spin_unlock_irqrestore(&dev_data_list_lock, flags);
209 static struct iommu_dev_data *find_dev_data(u16 devid)
211 struct iommu_dev_data *dev_data;
213 dev_data = search_dev_data(devid);
215 if (dev_data == NULL)
216 dev_data = alloc_dev_data(devid);
221 static inline int match_hid_uid(struct device *dev,
222 struct acpihid_map_entry *entry)
224 const char *hid, *uid;
226 hid = acpi_device_hid(ACPI_COMPANION(dev));
227 uid = acpi_device_uid(ACPI_COMPANION(dev));
233 return strcmp(hid, entry->hid);
236 return strcmp(hid, entry->hid);
238 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
241 static inline u16 get_pci_device_id(struct device *dev)
243 struct pci_dev *pdev = to_pci_dev(dev);
245 return PCI_DEVID(pdev->bus->number, pdev->devfn);
248 static inline int get_acpihid_device_id(struct device *dev,
249 struct acpihid_map_entry **entry)
251 struct acpihid_map_entry *p;
253 list_for_each_entry(p, &acpihid_map, list) {
254 if (!match_hid_uid(dev, p)) {
263 static inline int get_device_id(struct device *dev)
268 devid = get_pci_device_id(dev);
270 devid = get_acpihid_device_id(dev, NULL);
275 static struct iommu_dev_data *get_dev_data(struct device *dev)
277 return dev->archdata.iommu;
281 * Find or create an IOMMU group for a acpihid device.
283 static struct iommu_group *acpihid_device_group(struct device *dev)
285 struct acpihid_map_entry *p, *entry = NULL;
288 devid = get_acpihid_device_id(dev, &entry);
290 return ERR_PTR(devid);
292 list_for_each_entry(p, &acpihid_map, list) {
293 if ((devid == p->devid) && p->group)
294 entry->group = p->group;
298 entry->group = generic_device_group(dev);
303 static bool pci_iommuv2_capable(struct pci_dev *pdev)
305 static const int caps[] = {
308 PCI_EXT_CAP_ID_PASID,
312 for (i = 0; i < 3; ++i) {
313 pos = pci_find_ext_capability(pdev, caps[i]);
321 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
323 struct iommu_dev_data *dev_data;
325 dev_data = get_dev_data(&pdev->dev);
327 return dev_data->errata & (1 << erratum) ? true : false;
331 * This function actually applies the mapping to the page table of the
334 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
335 struct unity_map_entry *e)
339 for (addr = e->address_start; addr < e->address_end;
341 if (addr < dma_dom->aperture_size)
342 __set_bit(addr >> PAGE_SHIFT,
343 dma_dom->aperture[0]->bitmap);
348 * Inits the unity mappings required for a specific device
350 static void init_unity_mappings_for_device(struct device *dev,
351 struct dma_ops_domain *dma_dom)
353 struct unity_map_entry *e;
356 devid = get_device_id(dev);
360 list_for_each_entry(e, &amd_iommu_unity_map, list) {
361 if (!(devid >= e->devid_start && devid <= e->devid_end))
363 alloc_unity_mapping(dma_dom, e);
368 * This function checks if the driver got a valid device from the caller to
369 * avoid dereferencing invalid pointers.
371 static bool check_device(struct device *dev)
375 if (!dev || !dev->dma_mask)
378 devid = get_device_id(dev);
382 /* Out of our scope? */
383 if (devid > amd_iommu_last_bdf)
386 if (amd_iommu_rlookup_table[devid] == NULL)
392 static void init_iommu_group(struct device *dev)
394 struct dma_ops_domain *dma_domain;
395 struct iommu_domain *domain;
396 struct iommu_group *group;
398 group = iommu_group_get_for_dev(dev);
402 domain = iommu_group_default_domain(group);
406 dma_domain = to_pdomain(domain)->priv;
408 init_unity_mappings_for_device(dev, dma_domain);
410 iommu_group_put(group);
413 static int iommu_init_device(struct device *dev)
415 struct iommu_dev_data *dev_data;
418 if (dev->archdata.iommu)
421 devid = get_device_id(dev);
425 dev_data = find_dev_data(devid);
429 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
430 struct amd_iommu *iommu;
432 iommu = amd_iommu_rlookup_table[dev_data->devid];
433 dev_data->iommu_v2 = iommu->is_iommu_v2;
436 dev->archdata.iommu = dev_data;
438 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
444 static void iommu_ignore_device(struct device *dev)
449 devid = get_device_id(dev);
453 alias = amd_iommu_alias_table[devid];
455 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
456 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
458 amd_iommu_rlookup_table[devid] = NULL;
459 amd_iommu_rlookup_table[alias] = NULL;
462 static void iommu_uninit_device(struct device *dev)
465 struct iommu_dev_data *dev_data;
467 devid = get_device_id(dev);
471 dev_data = search_dev_data(devid);
475 if (dev_data->domain)
478 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
481 iommu_group_remove_device(dev);
484 dev->archdata.dma_ops = NULL;
487 * We keep dev_data around for unplugged devices and reuse it when the
488 * device is re-plugged - not doing so would introduce a ton of races.
492 #ifdef CONFIG_AMD_IOMMU_STATS
495 * Initialization code for statistics collection
498 DECLARE_STATS_COUNTER(compl_wait);
499 DECLARE_STATS_COUNTER(cnt_map_single);
500 DECLARE_STATS_COUNTER(cnt_unmap_single);
501 DECLARE_STATS_COUNTER(cnt_map_sg);
502 DECLARE_STATS_COUNTER(cnt_unmap_sg);
503 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
504 DECLARE_STATS_COUNTER(cnt_free_coherent);
505 DECLARE_STATS_COUNTER(cross_page);
506 DECLARE_STATS_COUNTER(domain_flush_single);
507 DECLARE_STATS_COUNTER(domain_flush_all);
508 DECLARE_STATS_COUNTER(alloced_io_mem);
509 DECLARE_STATS_COUNTER(total_map_requests);
510 DECLARE_STATS_COUNTER(complete_ppr);
511 DECLARE_STATS_COUNTER(invalidate_iotlb);
512 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
513 DECLARE_STATS_COUNTER(pri_requests);
515 static struct dentry *stats_dir;
516 static struct dentry *de_fflush;
518 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
520 if (stats_dir == NULL)
523 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
527 static void amd_iommu_stats_init(void)
529 stats_dir = debugfs_create_dir("amd-iommu", NULL);
530 if (stats_dir == NULL)
533 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
534 &amd_iommu_unmap_flush);
536 amd_iommu_stats_add(&compl_wait);
537 amd_iommu_stats_add(&cnt_map_single);
538 amd_iommu_stats_add(&cnt_unmap_single);
539 amd_iommu_stats_add(&cnt_map_sg);
540 amd_iommu_stats_add(&cnt_unmap_sg);
541 amd_iommu_stats_add(&cnt_alloc_coherent);
542 amd_iommu_stats_add(&cnt_free_coherent);
543 amd_iommu_stats_add(&cross_page);
544 amd_iommu_stats_add(&domain_flush_single);
545 amd_iommu_stats_add(&domain_flush_all);
546 amd_iommu_stats_add(&alloced_io_mem);
547 amd_iommu_stats_add(&total_map_requests);
548 amd_iommu_stats_add(&complete_ppr);
549 amd_iommu_stats_add(&invalidate_iotlb);
550 amd_iommu_stats_add(&invalidate_iotlb_all);
551 amd_iommu_stats_add(&pri_requests);
556 /****************************************************************************
558 * Interrupt handling functions
560 ****************************************************************************/
562 static void dump_dte_entry(u16 devid)
566 for (i = 0; i < 4; ++i)
567 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
568 amd_iommu_dev_table[devid].data[i]);
571 static void dump_command(unsigned long phys_addr)
573 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
576 for (i = 0; i < 4; ++i)
577 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
580 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
582 int type, devid, domid, flags;
583 volatile u32 *event = __evt;
588 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
589 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
590 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
591 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
592 address = (u64)(((u64)event[3]) << 32) | event[2];
595 /* Did we hit the erratum? */
596 if (++count == LOOP_TIMEOUT) {
597 pr_err("AMD-Vi: No event written to event log\n");
604 printk(KERN_ERR "AMD-Vi: Event logged [");
607 case EVENT_TYPE_ILL_DEV:
608 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
609 "address=0x%016llx flags=0x%04x]\n",
610 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
612 dump_dte_entry(devid);
614 case EVENT_TYPE_IO_FAULT:
615 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
616 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
617 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
618 domid, address, flags);
620 case EVENT_TYPE_DEV_TAB_ERR:
621 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
622 "address=0x%016llx flags=0x%04x]\n",
623 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
626 case EVENT_TYPE_PAGE_TAB_ERR:
627 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
628 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
629 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
630 domid, address, flags);
632 case EVENT_TYPE_ILL_CMD:
633 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
634 dump_command(address);
636 case EVENT_TYPE_CMD_HARD_ERR:
637 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
638 "flags=0x%04x]\n", address, flags);
640 case EVENT_TYPE_IOTLB_INV_TO:
641 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
642 "address=0x%016llx]\n",
643 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
646 case EVENT_TYPE_INV_DEV_REQ:
647 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
648 "address=0x%016llx flags=0x%04x]\n",
649 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
653 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
656 memset(__evt, 0, 4 * sizeof(u32));
659 static void iommu_poll_events(struct amd_iommu *iommu)
663 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
664 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
666 while (head != tail) {
667 iommu_print_event(iommu, iommu->evt_buf + head);
668 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
671 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
674 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
676 struct amd_iommu_fault fault;
678 INC_STATS_COUNTER(pri_requests);
680 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
681 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
685 fault.address = raw[1];
686 fault.pasid = PPR_PASID(raw[0]);
687 fault.device_id = PPR_DEVID(raw[0]);
688 fault.tag = PPR_TAG(raw[0]);
689 fault.flags = PPR_FLAGS(raw[0]);
691 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
694 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
698 if (iommu->ppr_log == NULL)
701 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
702 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
704 while (head != tail) {
709 raw = (u64 *)(iommu->ppr_log + head);
712 * Hardware bug: Interrupt may arrive before the entry is
713 * written to memory. If this happens we need to wait for the
716 for (i = 0; i < LOOP_TIMEOUT; ++i) {
717 if (PPR_REQ_TYPE(raw[0]) != 0)
722 /* Avoid memcpy function-call overhead */
727 * To detect the hardware bug we need to clear the entry
730 raw[0] = raw[1] = 0UL;
732 /* Update head pointer of hardware ring-buffer */
733 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
734 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
736 /* Handle PPR entry */
737 iommu_handle_ppr_entry(iommu, entry);
739 /* Refresh ring-buffer information */
740 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
741 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
745 irqreturn_t amd_iommu_int_thread(int irq, void *data)
747 struct amd_iommu *iommu = (struct amd_iommu *) data;
748 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
750 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
751 /* Enable EVT and PPR interrupts again */
752 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
753 iommu->mmio_base + MMIO_STATUS_OFFSET);
755 if (status & MMIO_STATUS_EVT_INT_MASK) {
756 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
757 iommu_poll_events(iommu);
760 if (status & MMIO_STATUS_PPR_INT_MASK) {
761 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
762 iommu_poll_ppr_log(iommu);
766 * Hardware bug: ERBT1312
767 * When re-enabling interrupt (by writing 1
768 * to clear the bit), the hardware might also try to set
769 * the interrupt bit in the event status register.
770 * In this scenario, the bit will be set, and disable
771 * subsequent interrupts.
773 * Workaround: The IOMMU driver should read back the
774 * status register and check if the interrupt bits are cleared.
775 * If not, driver will need to go through the interrupt handler
776 * again and re-clear the bits
778 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
783 irqreturn_t amd_iommu_int_handler(int irq, void *data)
785 return IRQ_WAKE_THREAD;
788 /****************************************************************************
790 * IOMMU command queuing functions
792 ****************************************************************************/
794 static int wait_on_sem(volatile u64 *sem)
798 while (*sem == 0 && i < LOOP_TIMEOUT) {
803 if (i == LOOP_TIMEOUT) {
804 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
811 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
812 struct iommu_cmd *cmd,
817 target = iommu->cmd_buf + tail;
818 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
820 /* Copy command to buffer */
821 memcpy(target, cmd, sizeof(*cmd));
823 /* Tell the IOMMU about it */
824 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
827 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
829 WARN_ON(address & 0x7ULL);
831 memset(cmd, 0, sizeof(*cmd));
832 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
833 cmd->data[1] = upper_32_bits(__pa(address));
835 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
838 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
840 memset(cmd, 0, sizeof(*cmd));
841 cmd->data[0] = devid;
842 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
845 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
846 size_t size, u16 domid, int pde)
851 pages = iommu_num_pages(address, size, PAGE_SIZE);
856 * If we have to flush more than one page, flush all
857 * TLB entries for this domain
859 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
863 address &= PAGE_MASK;
865 memset(cmd, 0, sizeof(*cmd));
866 cmd->data[1] |= domid;
867 cmd->data[2] = lower_32_bits(address);
868 cmd->data[3] = upper_32_bits(address);
869 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
870 if (s) /* size bit - we flush more than one 4kb page */
871 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
872 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
873 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
876 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
877 u64 address, size_t size)
882 pages = iommu_num_pages(address, size, PAGE_SIZE);
887 * If we have to flush more than one page, flush all
888 * TLB entries for this domain
890 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
894 address &= PAGE_MASK;
896 memset(cmd, 0, sizeof(*cmd));
897 cmd->data[0] = devid;
898 cmd->data[0] |= (qdep & 0xff) << 24;
899 cmd->data[1] = devid;
900 cmd->data[2] = lower_32_bits(address);
901 cmd->data[3] = upper_32_bits(address);
902 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
904 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
907 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
908 u64 address, bool size)
910 memset(cmd, 0, sizeof(*cmd));
912 address &= ~(0xfffULL);
914 cmd->data[0] = pasid;
915 cmd->data[1] = domid;
916 cmd->data[2] = lower_32_bits(address);
917 cmd->data[3] = upper_32_bits(address);
918 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
919 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
921 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
922 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
925 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
926 int qdep, u64 address, bool size)
928 memset(cmd, 0, sizeof(*cmd));
930 address &= ~(0xfffULL);
932 cmd->data[0] = devid;
933 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
934 cmd->data[0] |= (qdep & 0xff) << 24;
935 cmd->data[1] = devid;
936 cmd->data[1] |= (pasid & 0xff) << 16;
937 cmd->data[2] = lower_32_bits(address);
938 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
939 cmd->data[3] = upper_32_bits(address);
941 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
942 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
945 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
946 int status, int tag, bool gn)
948 memset(cmd, 0, sizeof(*cmd));
950 cmd->data[0] = devid;
952 cmd->data[1] = pasid;
953 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
955 cmd->data[3] = tag & 0x1ff;
956 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
958 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
961 static void build_inv_all(struct iommu_cmd *cmd)
963 memset(cmd, 0, sizeof(*cmd));
964 CMD_SET_TYPE(cmd, CMD_INV_ALL);
967 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
969 memset(cmd, 0, sizeof(*cmd));
970 cmd->data[0] = devid;
971 CMD_SET_TYPE(cmd, CMD_INV_IRT);
975 * Writes the command to the IOMMUs command buffer and informs the
976 * hardware about the new command.
978 static int iommu_queue_command_sync(struct amd_iommu *iommu,
979 struct iommu_cmd *cmd,
982 u32 left, tail, head, next_tail;
986 spin_lock_irqsave(&iommu->lock, flags);
988 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
989 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
990 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
991 left = (head - next_tail) % CMD_BUFFER_SIZE;
994 struct iommu_cmd sync_cmd;
995 volatile u64 sem = 0;
998 build_completion_wait(&sync_cmd, (u64)&sem);
999 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1001 spin_unlock_irqrestore(&iommu->lock, flags);
1003 if ((ret = wait_on_sem(&sem)) != 0)
1009 copy_cmd_to_buffer(iommu, cmd, tail);
1011 /* We need to sync now to make sure all commands are processed */
1012 iommu->need_sync = sync;
1014 spin_unlock_irqrestore(&iommu->lock, flags);
1019 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1021 return iommu_queue_command_sync(iommu, cmd, true);
1025 * This function queues a completion wait command into the command
1026 * buffer of an IOMMU
1028 static int iommu_completion_wait(struct amd_iommu *iommu)
1030 struct iommu_cmd cmd;
1031 volatile u64 sem = 0;
1034 if (!iommu->need_sync)
1037 build_completion_wait(&cmd, (u64)&sem);
1039 ret = iommu_queue_command_sync(iommu, &cmd, false);
1043 return wait_on_sem(&sem);
1046 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1048 struct iommu_cmd cmd;
1050 build_inv_dte(&cmd, devid);
1052 return iommu_queue_command(iommu, &cmd);
1055 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1059 for (devid = 0; devid <= 0xffff; ++devid)
1060 iommu_flush_dte(iommu, devid);
1062 iommu_completion_wait(iommu);
1066 * This function uses heavy locking and may disable irqs for some time. But
1067 * this is no issue because it is only called during resume.
1069 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1073 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1074 struct iommu_cmd cmd;
1075 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1077 iommu_queue_command(iommu, &cmd);
1080 iommu_completion_wait(iommu);
1083 static void iommu_flush_all(struct amd_iommu *iommu)
1085 struct iommu_cmd cmd;
1087 build_inv_all(&cmd);
1089 iommu_queue_command(iommu, &cmd);
1090 iommu_completion_wait(iommu);
1093 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1095 struct iommu_cmd cmd;
1097 build_inv_irt(&cmd, devid);
1099 iommu_queue_command(iommu, &cmd);
1102 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1106 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1107 iommu_flush_irt(iommu, devid);
1109 iommu_completion_wait(iommu);
1112 void iommu_flush_all_caches(struct amd_iommu *iommu)
1114 if (iommu_feature(iommu, FEATURE_IA)) {
1115 iommu_flush_all(iommu);
1117 iommu_flush_dte_all(iommu);
1118 iommu_flush_irt_all(iommu);
1119 iommu_flush_tlb_all(iommu);
1124 * Command send function for flushing on-device TLB
1126 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1127 u64 address, size_t size)
1129 struct amd_iommu *iommu;
1130 struct iommu_cmd cmd;
1133 qdep = dev_data->ats.qdep;
1134 iommu = amd_iommu_rlookup_table[dev_data->devid];
1136 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1138 return iommu_queue_command(iommu, &cmd);
1142 * Command send function for invalidating a device table entry
1144 static int device_flush_dte(struct iommu_dev_data *dev_data)
1146 struct amd_iommu *iommu;
1150 iommu = amd_iommu_rlookup_table[dev_data->devid];
1151 alias = amd_iommu_alias_table[dev_data->devid];
1153 ret = iommu_flush_dte(iommu, dev_data->devid);
1154 if (!ret && alias != dev_data->devid)
1155 ret = iommu_flush_dte(iommu, alias);
1159 if (dev_data->ats.enabled)
1160 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1166 * TLB invalidation function which is called from the mapping functions.
1167 * It invalidates a single PTE if the range to flush is within a single
1168 * page. Otherwise it flushes the whole TLB of the IOMMU.
1170 static void __domain_flush_pages(struct protection_domain *domain,
1171 u64 address, size_t size, int pde)
1173 struct iommu_dev_data *dev_data;
1174 struct iommu_cmd cmd;
1177 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1179 for (i = 0; i < amd_iommus_present; ++i) {
1180 if (!domain->dev_iommu[i])
1184 * Devices of this domain are behind this IOMMU
1185 * We need a TLB flush
1187 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1190 list_for_each_entry(dev_data, &domain->dev_list, list) {
1192 if (!dev_data->ats.enabled)
1195 ret |= device_flush_iotlb(dev_data, address, size);
1201 static void domain_flush_pages(struct protection_domain *domain,
1202 u64 address, size_t size)
1204 __domain_flush_pages(domain, address, size, 0);
1207 /* Flush the whole IO/TLB for a given protection domain */
1208 static void domain_flush_tlb(struct protection_domain *domain)
1210 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1213 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1214 static void domain_flush_tlb_pde(struct protection_domain *domain)
1216 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1219 static void domain_flush_complete(struct protection_domain *domain)
1223 for (i = 0; i < amd_iommus_present; ++i) {
1224 if (!domain->dev_iommu[i])
1228 * Devices of this domain are behind this IOMMU
1229 * We need to wait for completion of all commands.
1231 iommu_completion_wait(amd_iommus[i]);
1237 * This function flushes the DTEs for all devices in domain
1239 static void domain_flush_devices(struct protection_domain *domain)
1241 struct iommu_dev_data *dev_data;
1243 list_for_each_entry(dev_data, &domain->dev_list, list)
1244 device_flush_dte(dev_data);
1247 /****************************************************************************
1249 * The functions below are used the create the page table mappings for
1250 * unity mapped regions.
1252 ****************************************************************************/
1255 * This function is used to add another level to an IO page table. Adding
1256 * another level increases the size of the address space by 9 bits to a size up
1259 static bool increase_address_space(struct protection_domain *domain,
1264 if (domain->mode == PAGE_MODE_6_LEVEL)
1265 /* address space already 64 bit large */
1268 pte = (void *)get_zeroed_page(gfp);
1272 *pte = PM_LEVEL_PDE(domain->mode,
1273 virt_to_phys(domain->pt_root));
1274 domain->pt_root = pte;
1276 domain->updated = true;
1281 static u64 *alloc_pte(struct protection_domain *domain,
1282 unsigned long address,
1283 unsigned long page_size,
1290 BUG_ON(!is_power_of_2(page_size));
1292 while (address > PM_LEVEL_SIZE(domain->mode))
1293 increase_address_space(domain, gfp);
1295 level = domain->mode - 1;
1296 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1297 address = PAGE_SIZE_ALIGN(address, page_size);
1298 end_lvl = PAGE_SIZE_LEVEL(page_size);
1300 while (level > end_lvl) {
1305 if (!IOMMU_PTE_PRESENT(__pte)) {
1306 page = (u64 *)get_zeroed_page(gfp);
1310 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1312 if (cmpxchg64(pte, __pte, __npte)) {
1313 free_page((unsigned long)page);
1318 /* No level skipping support yet */
1319 if (PM_PTE_LEVEL(*pte) != level)
1324 pte = IOMMU_PTE_PAGE(*pte);
1326 if (pte_page && level == end_lvl)
1329 pte = &pte[PM_LEVEL_INDEX(level, address)];
1336 * This function checks if there is a PTE for a given dma address. If
1337 * there is one, it returns the pointer to it.
1339 static u64 *fetch_pte(struct protection_domain *domain,
1340 unsigned long address,
1341 unsigned long *page_size)
1346 if (address > PM_LEVEL_SIZE(domain->mode))
1349 level = domain->mode - 1;
1350 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1351 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1356 if (!IOMMU_PTE_PRESENT(*pte))
1360 if (PM_PTE_LEVEL(*pte) == 7 ||
1361 PM_PTE_LEVEL(*pte) == 0)
1364 /* No level skipping support yet */
1365 if (PM_PTE_LEVEL(*pte) != level)
1370 /* Walk to the next level */
1371 pte = IOMMU_PTE_PAGE(*pte);
1372 pte = &pte[PM_LEVEL_INDEX(level, address)];
1373 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1376 if (PM_PTE_LEVEL(*pte) == 0x07) {
1377 unsigned long pte_mask;
1380 * If we have a series of large PTEs, make
1381 * sure to return a pointer to the first one.
1383 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1384 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1385 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1392 * Generic mapping functions. It maps a physical address into a DMA
1393 * address space. It allocates the page table pages if necessary.
1394 * In the future it can be extended to a generic mapping function
1395 * supporting all features of AMD IOMMU page tables like level skipping
1396 * and full 64 bit address spaces.
1398 static int iommu_map_page(struct protection_domain *dom,
1399 unsigned long bus_addr,
1400 unsigned long phys_addr,
1402 unsigned long page_size)
1407 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1408 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1410 if (!(prot & IOMMU_PROT_MASK))
1413 count = PAGE_SIZE_PTE_COUNT(page_size);
1414 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1419 for (i = 0; i < count; ++i)
1420 if (IOMMU_PTE_PRESENT(pte[i]))
1424 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1425 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1427 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1429 if (prot & IOMMU_PROT_IR)
1430 __pte |= IOMMU_PTE_IR;
1431 if (prot & IOMMU_PROT_IW)
1432 __pte |= IOMMU_PTE_IW;
1434 for (i = 0; i < count; ++i)
1442 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1443 unsigned long bus_addr,
1444 unsigned long page_size)
1446 unsigned long long unmapped;
1447 unsigned long unmap_size;
1450 BUG_ON(!is_power_of_2(page_size));
1454 while (unmapped < page_size) {
1456 pte = fetch_pte(dom, bus_addr, &unmap_size);
1461 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1462 for (i = 0; i < count; i++)
1466 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1467 unmapped += unmap_size;
1470 BUG_ON(unmapped && !is_power_of_2(unmapped));
1475 /****************************************************************************
1477 * The next functions belong to the address allocator for the dma_ops
1478 * interface functions. They work like the allocators in the other IOMMU
1479 * drivers. Its basically a bitmap which marks the allocated pages in
1480 * the aperture. Maybe it could be enhanced in the future to a more
1481 * efficient allocator.
1483 ****************************************************************************/
1486 * The address allocator core functions.
1488 * called with domain->lock held
1492 * Used to reserve address ranges in the aperture (e.g. for exclusion
1495 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1496 unsigned long start_page,
1499 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1501 if (start_page + pages > last_page)
1502 pages = last_page - start_page;
1504 for (i = start_page; i < start_page + pages; ++i) {
1505 int index = i / APERTURE_RANGE_PAGES;
1506 int page = i % APERTURE_RANGE_PAGES;
1507 __set_bit(page, dom->aperture[index]->bitmap);
1512 * This function is used to add a new aperture range to an existing
1513 * aperture in case of dma_ops domain allocation or address allocation
1516 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1517 bool populate, gfp_t gfp)
1519 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1520 unsigned long i, old_size, pte_pgsize;
1521 struct aperture_range *range;
1522 struct amd_iommu *iommu;
1523 unsigned long flags;
1525 #ifdef CONFIG_IOMMU_STRESS
1529 if (index >= APERTURE_MAX_RANGES)
1532 range = kzalloc(sizeof(struct aperture_range), gfp);
1536 range->bitmap = (void *)get_zeroed_page(gfp);
1540 range->offset = dma_dom->aperture_size;
1542 spin_lock_init(&range->bitmap_lock);
1545 unsigned long address = dma_dom->aperture_size;
1546 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1547 u64 *pte, *pte_page;
1549 for (i = 0; i < num_ptes; ++i) {
1550 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1555 range->pte_pages[i] = pte_page;
1557 address += APERTURE_RANGE_SIZE / 64;
1561 spin_lock_irqsave(&dma_dom->domain.lock, flags);
1563 /* First take the bitmap_lock and then publish the range */
1564 spin_lock(&range->bitmap_lock);
1566 old_size = dma_dom->aperture_size;
1567 dma_dom->aperture[index] = range;
1568 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1570 /* Reserve address range used for MSI messages */
1571 if (old_size < MSI_ADDR_BASE_LO &&
1572 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1573 unsigned long spage;
1576 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1577 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1579 dma_ops_reserve_addresses(dma_dom, spage, pages);
1582 /* Initialize the exclusion range if necessary */
1583 for_each_iommu(iommu) {
1584 if (iommu->exclusion_start &&
1585 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1586 && iommu->exclusion_start < dma_dom->aperture_size) {
1587 unsigned long startpage;
1588 int pages = iommu_num_pages(iommu->exclusion_start,
1589 iommu->exclusion_length,
1591 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1592 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1597 * Check for areas already mapped as present in the new aperture
1598 * range and mark those pages as reserved in the allocator. Such
1599 * mappings may already exist as a result of requested unity
1600 * mappings for devices.
1602 for (i = dma_dom->aperture[index]->offset;
1603 i < dma_dom->aperture_size;
1605 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1606 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1609 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1613 update_domain(&dma_dom->domain);
1615 spin_unlock(&range->bitmap_lock);
1617 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
1622 update_domain(&dma_dom->domain);
1624 free_page((unsigned long)range->bitmap);
1631 static dma_addr_t dma_ops_aperture_alloc(struct dma_ops_domain *dom,
1632 struct aperture_range *range,
1633 unsigned long pages,
1634 unsigned long dma_mask,
1635 unsigned long boundary_size,
1636 unsigned long align_mask,
1639 unsigned long offset, limit, flags;
1643 offset = range->offset >> PAGE_SHIFT;
1644 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1645 dma_mask >> PAGE_SHIFT);
1648 if (!spin_trylock_irqsave(&range->bitmap_lock, flags))
1651 spin_lock_irqsave(&range->bitmap_lock, flags);
1654 address = iommu_area_alloc(range->bitmap, limit, range->next_bit,
1655 pages, offset, boundary_size, align_mask);
1656 if (address == -1) {
1657 /* Nothing found, retry one time */
1658 address = iommu_area_alloc(range->bitmap, limit,
1659 0, pages, offset, boundary_size,
1665 range->next_bit = address + pages;
1667 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1670 domain_flush_tlb(&dom->domain);
1671 domain_flush_complete(&dom->domain);
1677 static unsigned long dma_ops_area_alloc(struct device *dev,
1678 struct dma_ops_domain *dom,
1680 unsigned long align_mask,
1683 unsigned long boundary_size, mask;
1684 unsigned long address = -1;
1690 mask = dma_get_seg_boundary(dev);
1693 start = this_cpu_read(*dom->next_index);
1695 /* Sanity check - is it really necessary? */
1696 if (unlikely(start > APERTURE_MAX_RANGES)) {
1698 this_cpu_write(*dom->next_index, 0);
1701 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1702 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1704 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1705 struct aperture_range *range;
1708 index = (start + i) % APERTURE_MAX_RANGES;
1710 range = dom->aperture[index];
1712 if (!range || range->offset >= dma_mask)
1715 address = dma_ops_aperture_alloc(dom, range, pages,
1716 dma_mask, boundary_size,
1718 if (address != -1) {
1719 address = range->offset + (address << PAGE_SHIFT);
1720 this_cpu_write(*dom->next_index, index);
1725 if (address == -1 && first) {
1735 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1736 struct dma_ops_domain *dom,
1738 unsigned long align_mask,
1741 unsigned long address = -1;
1743 while (address == -1) {
1744 address = dma_ops_area_alloc(dev, dom, pages,
1745 align_mask, dma_mask);
1747 if (address == -1 && alloc_new_range(dom, false, GFP_ATOMIC))
1751 if (unlikely(address == -1))
1752 address = DMA_ERROR_CODE;
1754 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1760 * The address free function.
1762 * called with domain->lock held
1764 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1765 unsigned long address,
1768 unsigned i = address >> APERTURE_RANGE_SHIFT;
1769 struct aperture_range *range = dom->aperture[i];
1770 unsigned long flags;
1772 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1774 #ifdef CONFIG_IOMMU_STRESS
1779 if (amd_iommu_unmap_flush) {
1780 domain_flush_tlb(&dom->domain);
1781 domain_flush_complete(&dom->domain);
1784 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1786 spin_lock_irqsave(&range->bitmap_lock, flags);
1787 if (address + pages > range->next_bit)
1788 range->next_bit = address + pages;
1789 bitmap_clear(range->bitmap, address, pages);
1790 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1794 /****************************************************************************
1796 * The next functions belong to the domain allocation. A domain is
1797 * allocated for every IOMMU as the default domain. If device isolation
1798 * is enabled, every device get its own domain. The most important thing
1799 * about domains is the page table mapping the DMA address space they
1802 ****************************************************************************/
1805 * This function adds a protection domain to the global protection domain list
1807 static void add_domain_to_list(struct protection_domain *domain)
1809 unsigned long flags;
1811 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1812 list_add(&domain->list, &amd_iommu_pd_list);
1813 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1817 * This function removes a protection domain to the global
1818 * protection domain list
1820 static void del_domain_from_list(struct protection_domain *domain)
1822 unsigned long flags;
1824 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1825 list_del(&domain->list);
1826 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1829 static u16 domain_id_alloc(void)
1831 unsigned long flags;
1834 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1835 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1837 if (id > 0 && id < MAX_DOMAIN_ID)
1838 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1841 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1846 static void domain_id_free(int id)
1848 unsigned long flags;
1850 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1851 if (id > 0 && id < MAX_DOMAIN_ID)
1852 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1853 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1856 #define DEFINE_FREE_PT_FN(LVL, FN) \
1857 static void free_pt_##LVL (unsigned long __pt) \
1865 for (i = 0; i < 512; ++i) { \
1866 /* PTE present? */ \
1867 if (!IOMMU_PTE_PRESENT(pt[i])) \
1871 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1872 PM_PTE_LEVEL(pt[i]) == 7) \
1875 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1878 free_page((unsigned long)pt); \
1881 DEFINE_FREE_PT_FN(l2, free_page)
1882 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1883 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1884 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1885 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1887 static void free_pagetable(struct protection_domain *domain)
1889 unsigned long root = (unsigned long)domain->pt_root;
1891 switch (domain->mode) {
1892 case PAGE_MODE_NONE:
1894 case PAGE_MODE_1_LEVEL:
1897 case PAGE_MODE_2_LEVEL:
1900 case PAGE_MODE_3_LEVEL:
1903 case PAGE_MODE_4_LEVEL:
1906 case PAGE_MODE_5_LEVEL:
1909 case PAGE_MODE_6_LEVEL:
1917 static void free_gcr3_tbl_level1(u64 *tbl)
1922 for (i = 0; i < 512; ++i) {
1923 if (!(tbl[i] & GCR3_VALID))
1926 ptr = __va(tbl[i] & PAGE_MASK);
1928 free_page((unsigned long)ptr);
1932 static void free_gcr3_tbl_level2(u64 *tbl)
1937 for (i = 0; i < 512; ++i) {
1938 if (!(tbl[i] & GCR3_VALID))
1941 ptr = __va(tbl[i] & PAGE_MASK);
1943 free_gcr3_tbl_level1(ptr);
1947 static void free_gcr3_table(struct protection_domain *domain)
1949 if (domain->glx == 2)
1950 free_gcr3_tbl_level2(domain->gcr3_tbl);
1951 else if (domain->glx == 1)
1952 free_gcr3_tbl_level1(domain->gcr3_tbl);
1954 BUG_ON(domain->glx != 0);
1956 free_page((unsigned long)domain->gcr3_tbl);
1960 * Free a domain, only used if something went wrong in the
1961 * allocation path and we need to free an already allocated page table
1963 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1970 free_percpu(dom->next_index);
1972 del_domain_from_list(&dom->domain);
1974 free_pagetable(&dom->domain);
1976 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1977 if (!dom->aperture[i])
1979 free_page((unsigned long)dom->aperture[i]->bitmap);
1980 kfree(dom->aperture[i]);
1986 static int dma_ops_domain_alloc_apertures(struct dma_ops_domain *dma_dom,
1989 int ret, i, apertures;
1991 apertures = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1994 for (i = apertures; i < max_apertures; ++i) {
1995 ret = alloc_new_range(dma_dom, false, GFP_KERNEL);
2004 * Allocates a new protection domain usable for the dma_ops functions.
2005 * It also initializes the page table and the address allocator data
2006 * structures required for the dma_ops interface
2008 static struct dma_ops_domain *dma_ops_domain_alloc(void)
2010 struct dma_ops_domain *dma_dom;
2013 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2017 if (protection_domain_init(&dma_dom->domain))
2020 dma_dom->next_index = alloc_percpu(u32);
2021 if (!dma_dom->next_index)
2024 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2025 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2026 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2027 dma_dom->domain.priv = dma_dom;
2028 if (!dma_dom->domain.pt_root)
2031 add_domain_to_list(&dma_dom->domain);
2033 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2037 * mark the first page as allocated so we never return 0 as
2038 * a valid dma-address. So we can use 0 as error value
2040 dma_dom->aperture[0]->bitmap[0] = 1;
2042 for_each_possible_cpu(cpu)
2043 *per_cpu_ptr(dma_dom->next_index, cpu) = 0;
2048 dma_ops_domain_free(dma_dom);
2054 * little helper function to check whether a given protection domain is a
2057 static bool dma_ops_domain(struct protection_domain *domain)
2059 return domain->flags & PD_DMA_OPS_MASK;
2062 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2067 if (domain->mode != PAGE_MODE_NONE)
2068 pte_root = virt_to_phys(domain->pt_root);
2070 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2071 << DEV_ENTRY_MODE_SHIFT;
2072 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2074 flags = amd_iommu_dev_table[devid].data[1];
2077 flags |= DTE_FLAG_IOTLB;
2079 if (domain->flags & PD_IOMMUV2_MASK) {
2080 u64 gcr3 = __pa(domain->gcr3_tbl);
2081 u64 glx = domain->glx;
2084 pte_root |= DTE_FLAG_GV;
2085 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2087 /* First mask out possible old values for GCR3 table */
2088 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2091 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2094 /* Encode GCR3 table into DTE */
2095 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2098 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2101 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2105 flags &= ~(0xffffUL);
2106 flags |= domain->id;
2108 amd_iommu_dev_table[devid].data[1] = flags;
2109 amd_iommu_dev_table[devid].data[0] = pte_root;
2112 static void clear_dte_entry(u16 devid)
2114 /* remove entry from the device table seen by the hardware */
2115 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2116 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2118 amd_iommu_apply_erratum_63(devid);
2121 static void do_attach(struct iommu_dev_data *dev_data,
2122 struct protection_domain *domain)
2124 struct amd_iommu *iommu;
2128 iommu = amd_iommu_rlookup_table[dev_data->devid];
2129 alias = amd_iommu_alias_table[dev_data->devid];
2130 ats = dev_data->ats.enabled;
2132 /* Update data structures */
2133 dev_data->domain = domain;
2134 list_add(&dev_data->list, &domain->dev_list);
2136 /* Do reference counting */
2137 domain->dev_iommu[iommu->index] += 1;
2138 domain->dev_cnt += 1;
2140 /* Update device table */
2141 set_dte_entry(dev_data->devid, domain, ats);
2142 if (alias != dev_data->devid)
2143 set_dte_entry(alias, domain, ats);
2145 device_flush_dte(dev_data);
2148 static void do_detach(struct iommu_dev_data *dev_data)
2150 struct amd_iommu *iommu;
2154 * First check if the device is still attached. It might already
2155 * be detached from its domain because the generic
2156 * iommu_detach_group code detached it and we try again here in
2157 * our alias handling.
2159 if (!dev_data->domain)
2162 iommu = amd_iommu_rlookup_table[dev_data->devid];
2163 alias = amd_iommu_alias_table[dev_data->devid];
2165 /* decrease reference counters */
2166 dev_data->domain->dev_iommu[iommu->index] -= 1;
2167 dev_data->domain->dev_cnt -= 1;
2169 /* Update data structures */
2170 dev_data->domain = NULL;
2171 list_del(&dev_data->list);
2172 clear_dte_entry(dev_data->devid);
2173 if (alias != dev_data->devid)
2174 clear_dte_entry(alias);
2176 /* Flush the DTE entry */
2177 device_flush_dte(dev_data);
2181 * If a device is not yet associated with a domain, this function does
2182 * assigns it visible for the hardware
2184 static int __attach_device(struct iommu_dev_data *dev_data,
2185 struct protection_domain *domain)
2190 * Must be called with IRQs disabled. Warn here to detect early
2193 WARN_ON(!irqs_disabled());
2196 spin_lock(&domain->lock);
2199 if (dev_data->domain != NULL)
2202 /* Attach alias group root */
2203 do_attach(dev_data, domain);
2210 spin_unlock(&domain->lock);
2216 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2218 pci_disable_ats(pdev);
2219 pci_disable_pri(pdev);
2220 pci_disable_pasid(pdev);
2223 /* FIXME: Change generic reset-function to do the same */
2224 static int pri_reset_while_enabled(struct pci_dev *pdev)
2229 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2233 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2234 control |= PCI_PRI_CTRL_RESET;
2235 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2240 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2245 /* FIXME: Hardcode number of outstanding requests for now */
2247 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2249 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2251 /* Only allow access to user-accessible pages */
2252 ret = pci_enable_pasid(pdev, 0);
2256 /* First reset the PRI state of the device */
2257 ret = pci_reset_pri(pdev);
2262 ret = pci_enable_pri(pdev, reqs);
2267 ret = pri_reset_while_enabled(pdev);
2272 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2279 pci_disable_pri(pdev);
2280 pci_disable_pasid(pdev);
2285 /* FIXME: Move this to PCI code */
2286 #define PCI_PRI_TLP_OFF (1 << 15)
2288 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2293 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2297 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2299 return (status & PCI_PRI_TLP_OFF) ? true : false;
2303 * If a device is not yet associated with a domain, this function
2304 * assigns it visible for the hardware
2306 static int attach_device(struct device *dev,
2307 struct protection_domain *domain)
2309 struct pci_dev *pdev;
2310 struct iommu_dev_data *dev_data;
2311 unsigned long flags;
2314 dev_data = get_dev_data(dev);
2316 if (!dev_is_pci(dev))
2317 goto skip_ats_check;
2319 pdev = to_pci_dev(dev);
2320 if (domain->flags & PD_IOMMUV2_MASK) {
2321 if (!dev_data->passthrough)
2324 if (dev_data->iommu_v2) {
2325 if (pdev_iommuv2_enable(pdev) != 0)
2328 dev_data->ats.enabled = true;
2329 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2330 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2332 } else if (amd_iommu_iotlb_sup &&
2333 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2334 dev_data->ats.enabled = true;
2335 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2339 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2340 ret = __attach_device(dev_data, domain);
2341 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2344 * We might boot into a crash-kernel here. The crashed kernel
2345 * left the caches in the IOMMU dirty. So we have to flush
2346 * here to evict all dirty stuff.
2348 domain_flush_tlb_pde(domain);
2354 * Removes a device from a protection domain (unlocked)
2356 static void __detach_device(struct iommu_dev_data *dev_data)
2358 struct protection_domain *domain;
2361 * Must be called with IRQs disabled. Warn here to detect early
2364 WARN_ON(!irqs_disabled());
2366 if (WARN_ON(!dev_data->domain))
2369 domain = dev_data->domain;
2371 spin_lock(&domain->lock);
2373 do_detach(dev_data);
2375 spin_unlock(&domain->lock);
2379 * Removes a device from a protection domain (with devtable_lock held)
2381 static void detach_device(struct device *dev)
2383 struct protection_domain *domain;
2384 struct iommu_dev_data *dev_data;
2385 unsigned long flags;
2387 dev_data = get_dev_data(dev);
2388 domain = dev_data->domain;
2390 /* lock device table */
2391 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2392 __detach_device(dev_data);
2393 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2395 if (!dev_is_pci(dev))
2398 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2399 pdev_iommuv2_disable(to_pci_dev(dev));
2400 else if (dev_data->ats.enabled)
2401 pci_disable_ats(to_pci_dev(dev));
2403 dev_data->ats.enabled = false;
2406 static int amd_iommu_add_device(struct device *dev)
2408 struct iommu_dev_data *dev_data;
2409 struct iommu_domain *domain;
2410 struct amd_iommu *iommu;
2413 if (!check_device(dev) || get_dev_data(dev))
2416 devid = get_device_id(dev);
2420 iommu = amd_iommu_rlookup_table[devid];
2422 ret = iommu_init_device(dev);
2424 if (ret != -ENOTSUPP)
2425 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2428 iommu_ignore_device(dev);
2429 dev->archdata.dma_ops = &nommu_dma_ops;
2432 init_iommu_group(dev);
2434 dev_data = get_dev_data(dev);
2438 if (iommu_pass_through || dev_data->iommu_v2)
2439 iommu_request_dm_for_dev(dev);
2441 /* Domains are initialized for this device - have a look what we ended up with */
2442 domain = iommu_get_domain_for_dev(dev);
2443 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2444 dev_data->passthrough = true;
2446 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2449 iommu_completion_wait(iommu);
2454 static void amd_iommu_remove_device(struct device *dev)
2456 struct amd_iommu *iommu;
2459 if (!check_device(dev))
2462 devid = get_device_id(dev);
2466 iommu = amd_iommu_rlookup_table[devid];
2468 iommu_uninit_device(dev);
2469 iommu_completion_wait(iommu);
2472 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2474 if (dev_is_pci(dev))
2475 return pci_device_group(dev);
2477 return acpihid_device_group(dev);
2480 /*****************************************************************************
2482 * The next functions belong to the dma_ops mapping/unmapping code.
2484 *****************************************************************************/
2487 * In the dma_ops path we only have the struct device. This function
2488 * finds the corresponding IOMMU, the protection domain and the
2489 * requestor id for a given device.
2490 * If the device is not yet associated with a domain this is also done
2493 static struct protection_domain *get_domain(struct device *dev)
2495 struct protection_domain *domain;
2496 struct iommu_domain *io_domain;
2498 if (!check_device(dev))
2499 return ERR_PTR(-EINVAL);
2501 io_domain = iommu_get_domain_for_dev(dev);
2505 domain = to_pdomain(io_domain);
2506 if (!dma_ops_domain(domain))
2507 return ERR_PTR(-EBUSY);
2512 static void update_device_table(struct protection_domain *domain)
2514 struct iommu_dev_data *dev_data;
2516 list_for_each_entry(dev_data, &domain->dev_list, list)
2517 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2520 static void update_domain(struct protection_domain *domain)
2522 if (!domain->updated)
2525 update_device_table(domain);
2527 domain_flush_devices(domain);
2528 domain_flush_tlb_pde(domain);
2530 domain->updated = false;
2534 * This function fetches the PTE for a given address in the aperture
2536 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2537 unsigned long address)
2539 struct aperture_range *aperture;
2540 u64 *pte, *pte_page;
2542 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2546 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2548 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2550 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2552 pte += PM_LEVEL_INDEX(0, address);
2554 update_domain(&dom->domain);
2560 * This is the generic map function. It maps one 4kb page at paddr to
2561 * the given address in the DMA address space for the domain.
2563 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2564 unsigned long address,
2570 WARN_ON(address > dom->aperture_size);
2574 pte = dma_ops_get_pte(dom, address);
2576 return DMA_ERROR_CODE;
2578 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2580 if (direction == DMA_TO_DEVICE)
2581 __pte |= IOMMU_PTE_IR;
2582 else if (direction == DMA_FROM_DEVICE)
2583 __pte |= IOMMU_PTE_IW;
2584 else if (direction == DMA_BIDIRECTIONAL)
2585 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2591 return (dma_addr_t)address;
2595 * The generic unmapping function for on page in the DMA address space.
2597 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2598 unsigned long address)
2600 struct aperture_range *aperture;
2603 if (address >= dom->aperture_size)
2606 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2610 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2614 pte += PM_LEVEL_INDEX(0, address);
2616 WARN_ON_ONCE(!*pte);
2622 * This function contains common code for mapping of a physically
2623 * contiguous memory region into DMA address space. It is used by all
2624 * mapping functions provided with this IOMMU driver.
2625 * Must be called with the domain lock held.
2627 static dma_addr_t __map_single(struct device *dev,
2628 struct dma_ops_domain *dma_dom,
2635 dma_addr_t offset = paddr & ~PAGE_MASK;
2636 dma_addr_t address, start, ret;
2638 unsigned long align_mask = 0;
2641 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2644 INC_STATS_COUNTER(total_map_requests);
2647 INC_STATS_COUNTER(cross_page);
2650 align_mask = (1UL << get_order(size)) - 1;
2652 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2655 if (address == DMA_ERROR_CODE)
2659 for (i = 0; i < pages; ++i) {
2660 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2661 if (ret == DMA_ERROR_CODE)
2669 ADD_STATS_COUNTER(alloced_io_mem, size);
2671 if (unlikely(amd_iommu_np_cache)) {
2672 domain_flush_pages(&dma_dom->domain, address, size);
2673 domain_flush_complete(&dma_dom->domain);
2681 for (--i; i >= 0; --i) {
2683 dma_ops_domain_unmap(dma_dom, start);
2686 dma_ops_free_addresses(dma_dom, address, pages);
2688 return DMA_ERROR_CODE;
2692 * Does the reverse of the __map_single function. Must be called with
2693 * the domain lock held too
2695 static void __unmap_single(struct dma_ops_domain *dma_dom,
2696 dma_addr_t dma_addr,
2700 dma_addr_t flush_addr;
2701 dma_addr_t i, start;
2704 if ((dma_addr == DMA_ERROR_CODE) ||
2705 (dma_addr + size > dma_dom->aperture_size))
2708 flush_addr = dma_addr;
2709 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2710 dma_addr &= PAGE_MASK;
2713 for (i = 0; i < pages; ++i) {
2714 dma_ops_domain_unmap(dma_dom, start);
2718 SUB_STATS_COUNTER(alloced_io_mem, size);
2720 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2724 * The exported map_single function for dma_ops.
2726 static dma_addr_t map_page(struct device *dev, struct page *page,
2727 unsigned long offset, size_t size,
2728 enum dma_data_direction dir,
2729 struct dma_attrs *attrs)
2731 phys_addr_t paddr = page_to_phys(page) + offset;
2732 struct protection_domain *domain;
2735 INC_STATS_COUNTER(cnt_map_single);
2737 domain = get_domain(dev);
2738 if (PTR_ERR(domain) == -EINVAL)
2739 return (dma_addr_t)paddr;
2740 else if (IS_ERR(domain))
2741 return DMA_ERROR_CODE;
2743 dma_mask = *dev->dma_mask;
2745 return __map_single(dev, domain->priv, paddr, size, dir, false,
2750 * The exported unmap_single function for dma_ops.
2752 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2753 enum dma_data_direction dir, struct dma_attrs *attrs)
2755 struct protection_domain *domain;
2757 INC_STATS_COUNTER(cnt_unmap_single);
2759 domain = get_domain(dev);
2763 __unmap_single(domain->priv, dma_addr, size, dir);
2767 * The exported map_sg function for dma_ops (handles scatter-gather
2770 static int map_sg(struct device *dev, struct scatterlist *sglist,
2771 int nelems, enum dma_data_direction dir,
2772 struct dma_attrs *attrs)
2774 struct protection_domain *domain;
2776 struct scatterlist *s;
2778 int mapped_elems = 0;
2781 INC_STATS_COUNTER(cnt_map_sg);
2783 domain = get_domain(dev);
2787 dma_mask = *dev->dma_mask;
2789 for_each_sg(sglist, s, nelems, i) {
2792 s->dma_address = __map_single(dev, domain->priv,
2793 paddr, s->length, dir, false,
2796 if (s->dma_address) {
2797 s->dma_length = s->length;
2803 return mapped_elems;
2806 for_each_sg(sglist, s, mapped_elems, i) {
2808 __unmap_single(domain->priv, s->dma_address,
2809 s->dma_length, dir);
2810 s->dma_address = s->dma_length = 0;
2817 * The exported map_sg function for dma_ops (handles scatter-gather
2820 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2821 int nelems, enum dma_data_direction dir,
2822 struct dma_attrs *attrs)
2824 struct protection_domain *domain;
2825 struct scatterlist *s;
2828 INC_STATS_COUNTER(cnt_unmap_sg);
2830 domain = get_domain(dev);
2834 for_each_sg(sglist, s, nelems, i) {
2835 __unmap_single(domain->priv, s->dma_address,
2836 s->dma_length, dir);
2837 s->dma_address = s->dma_length = 0;
2842 * The exported alloc_coherent function for dma_ops.
2844 static void *alloc_coherent(struct device *dev, size_t size,
2845 dma_addr_t *dma_addr, gfp_t flag,
2846 struct dma_attrs *attrs)
2848 u64 dma_mask = dev->coherent_dma_mask;
2849 struct protection_domain *domain;
2852 INC_STATS_COUNTER(cnt_alloc_coherent);
2854 domain = get_domain(dev);
2855 if (PTR_ERR(domain) == -EINVAL) {
2856 page = alloc_pages(flag, get_order(size));
2857 *dma_addr = page_to_phys(page);
2858 return page_address(page);
2859 } else if (IS_ERR(domain))
2862 size = PAGE_ALIGN(size);
2863 dma_mask = dev->coherent_dma_mask;
2864 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2867 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2869 if (!gfpflags_allow_blocking(flag))
2872 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2879 dma_mask = *dev->dma_mask;
2881 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2882 size, DMA_BIDIRECTIONAL, true, dma_mask);
2884 if (*dma_addr == DMA_ERROR_CODE)
2887 return page_address(page);
2891 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2892 __free_pages(page, get_order(size));
2898 * The exported free_coherent function for dma_ops.
2900 static void free_coherent(struct device *dev, size_t size,
2901 void *virt_addr, dma_addr_t dma_addr,
2902 struct dma_attrs *attrs)
2904 struct protection_domain *domain;
2907 INC_STATS_COUNTER(cnt_free_coherent);
2909 page = virt_to_page(virt_addr);
2910 size = PAGE_ALIGN(size);
2912 domain = get_domain(dev);
2916 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2919 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2920 __free_pages(page, get_order(size));
2924 * This function is called by the DMA layer to find out if we can handle a
2925 * particular device. It is part of the dma_ops.
2927 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2929 return check_device(dev);
2932 static int set_dma_mask(struct device *dev, u64 mask)
2934 struct protection_domain *domain;
2935 int max_apertures = 1;
2937 domain = get_domain(dev);
2939 return PTR_ERR(domain);
2941 if (mask == DMA_BIT_MASK(64))
2943 else if (mask > DMA_BIT_MASK(32))
2947 * To prevent lock contention it doesn't make sense to allocate more
2948 * apertures than online cpus
2950 if (max_apertures > num_online_cpus())
2951 max_apertures = num_online_cpus();
2953 if (dma_ops_domain_alloc_apertures(domain->priv, max_apertures))
2954 dev_err(dev, "Can't allocate %d iommu apertures\n",
2960 static struct dma_map_ops amd_iommu_dma_ops = {
2961 .alloc = alloc_coherent,
2962 .free = free_coherent,
2963 .map_page = map_page,
2964 .unmap_page = unmap_page,
2966 .unmap_sg = unmap_sg,
2967 .dma_supported = amd_iommu_dma_supported,
2968 .set_dma_mask = set_dma_mask,
2971 int __init amd_iommu_init_api(void)
2975 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2978 #ifdef CONFIG_ARM_AMBA
2979 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2986 int __init amd_iommu_init_dma_ops(void)
2988 swiotlb = iommu_pass_through ? 1 : 0;
2992 * In case we don't initialize SWIOTLB (actually the common case
2993 * when AMD IOMMU is enabled), make sure there are global
2994 * dma_ops set as a fall-back for devices not handled by this
2995 * driver (for example non-PCI devices).
2998 dma_ops = &nommu_dma_ops;
3000 amd_iommu_stats_init();
3002 if (amd_iommu_unmap_flush)
3003 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3005 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3010 /*****************************************************************************
3012 * The following functions belong to the exported interface of AMD IOMMU
3014 * This interface allows access to lower level functions of the IOMMU
3015 * like protection domain handling and assignement of devices to domains
3016 * which is not possible with the dma_ops interface.
3018 *****************************************************************************/
3020 static void cleanup_domain(struct protection_domain *domain)
3022 struct iommu_dev_data *entry;
3023 unsigned long flags;
3025 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3027 while (!list_empty(&domain->dev_list)) {
3028 entry = list_first_entry(&domain->dev_list,
3029 struct iommu_dev_data, list);
3030 __detach_device(entry);
3033 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3036 static void protection_domain_free(struct protection_domain *domain)
3041 del_domain_from_list(domain);
3044 domain_id_free(domain->id);
3049 static int protection_domain_init(struct protection_domain *domain)
3051 spin_lock_init(&domain->lock);
3052 mutex_init(&domain->api_lock);
3053 domain->id = domain_id_alloc();
3056 INIT_LIST_HEAD(&domain->dev_list);
3061 static struct protection_domain *protection_domain_alloc(void)
3063 struct protection_domain *domain;
3065 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3069 if (protection_domain_init(domain))
3072 add_domain_to_list(domain);
3082 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
3084 struct protection_domain *pdomain;
3085 struct dma_ops_domain *dma_domain;
3088 case IOMMU_DOMAIN_UNMANAGED:
3089 pdomain = protection_domain_alloc();
3093 pdomain->mode = PAGE_MODE_3_LEVEL;
3094 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3095 if (!pdomain->pt_root) {
3096 protection_domain_free(pdomain);
3100 pdomain->domain.geometry.aperture_start = 0;
3101 pdomain->domain.geometry.aperture_end = ~0ULL;
3102 pdomain->domain.geometry.force_aperture = true;
3105 case IOMMU_DOMAIN_DMA:
3106 dma_domain = dma_ops_domain_alloc();
3108 pr_err("AMD-Vi: Failed to allocate\n");
3111 pdomain = &dma_domain->domain;
3113 case IOMMU_DOMAIN_IDENTITY:
3114 pdomain = protection_domain_alloc();
3118 pdomain->mode = PAGE_MODE_NONE;
3124 return &pdomain->domain;
3127 static void amd_iommu_domain_free(struct iommu_domain *dom)
3129 struct protection_domain *domain;
3134 domain = to_pdomain(dom);
3136 if (domain->dev_cnt > 0)
3137 cleanup_domain(domain);
3139 BUG_ON(domain->dev_cnt != 0);
3141 if (domain->mode != PAGE_MODE_NONE)
3142 free_pagetable(domain);
3144 if (domain->flags & PD_IOMMUV2_MASK)
3145 free_gcr3_table(domain);
3147 protection_domain_free(domain);
3150 static void amd_iommu_detach_device(struct iommu_domain *dom,
3153 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3154 struct amd_iommu *iommu;
3157 if (!check_device(dev))
3160 devid = get_device_id(dev);
3164 if (dev_data->domain != NULL)
3167 iommu = amd_iommu_rlookup_table[devid];
3171 iommu_completion_wait(iommu);
3174 static int amd_iommu_attach_device(struct iommu_domain *dom,
3177 struct protection_domain *domain = to_pdomain(dom);
3178 struct iommu_dev_data *dev_data;
3179 struct amd_iommu *iommu;
3182 if (!check_device(dev))
3185 dev_data = dev->archdata.iommu;
3187 iommu = amd_iommu_rlookup_table[dev_data->devid];
3191 if (dev_data->domain)
3194 ret = attach_device(dev, domain);
3196 iommu_completion_wait(iommu);
3201 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3202 phys_addr_t paddr, size_t page_size, int iommu_prot)
3204 struct protection_domain *domain = to_pdomain(dom);
3208 if (domain->mode == PAGE_MODE_NONE)
3211 if (iommu_prot & IOMMU_READ)
3212 prot |= IOMMU_PROT_IR;
3213 if (iommu_prot & IOMMU_WRITE)
3214 prot |= IOMMU_PROT_IW;
3216 mutex_lock(&domain->api_lock);
3217 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3218 mutex_unlock(&domain->api_lock);
3223 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3226 struct protection_domain *domain = to_pdomain(dom);
3229 if (domain->mode == PAGE_MODE_NONE)
3232 mutex_lock(&domain->api_lock);
3233 unmap_size = iommu_unmap_page(domain, iova, page_size);
3234 mutex_unlock(&domain->api_lock);
3236 domain_flush_tlb_pde(domain);
3241 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3244 struct protection_domain *domain = to_pdomain(dom);
3245 unsigned long offset_mask, pte_pgsize;
3248 if (domain->mode == PAGE_MODE_NONE)
3251 pte = fetch_pte(domain, iova, &pte_pgsize);
3253 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3256 offset_mask = pte_pgsize - 1;
3257 __pte = *pte & PM_ADDR_MASK;
3259 return (__pte & ~offset_mask) | (iova & offset_mask);
3262 static bool amd_iommu_capable(enum iommu_cap cap)
3265 case IOMMU_CAP_CACHE_COHERENCY:
3267 case IOMMU_CAP_INTR_REMAP:
3268 return (irq_remapping_enabled == 1);
3269 case IOMMU_CAP_NOEXEC:
3276 static void amd_iommu_get_dm_regions(struct device *dev,
3277 struct list_head *head)
3279 struct unity_map_entry *entry;
3282 devid = get_device_id(dev);
3286 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3287 struct iommu_dm_region *region;
3289 if (devid < entry->devid_start || devid > entry->devid_end)
3292 region = kzalloc(sizeof(*region), GFP_KERNEL);
3294 pr_err("Out of memory allocating dm-regions for %s\n",
3299 region->start = entry->address_start;
3300 region->length = entry->address_end - entry->address_start;
3301 if (entry->prot & IOMMU_PROT_IR)
3302 region->prot |= IOMMU_READ;
3303 if (entry->prot & IOMMU_PROT_IW)
3304 region->prot |= IOMMU_WRITE;
3306 list_add_tail(®ion->list, head);
3310 static void amd_iommu_put_dm_regions(struct device *dev,
3311 struct list_head *head)
3313 struct iommu_dm_region *entry, *next;
3315 list_for_each_entry_safe(entry, next, head, list)
3319 static const struct iommu_ops amd_iommu_ops = {
3320 .capable = amd_iommu_capable,
3321 .domain_alloc = amd_iommu_domain_alloc,
3322 .domain_free = amd_iommu_domain_free,
3323 .attach_dev = amd_iommu_attach_device,
3324 .detach_dev = amd_iommu_detach_device,
3325 .map = amd_iommu_map,
3326 .unmap = amd_iommu_unmap,
3327 .map_sg = default_iommu_map_sg,
3328 .iova_to_phys = amd_iommu_iova_to_phys,
3329 .add_device = amd_iommu_add_device,
3330 .remove_device = amd_iommu_remove_device,
3331 .device_group = amd_iommu_device_group,
3332 .get_dm_regions = amd_iommu_get_dm_regions,
3333 .put_dm_regions = amd_iommu_put_dm_regions,
3334 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3337 /*****************************************************************************
3339 * The next functions do a basic initialization of IOMMU for pass through
3342 * In passthrough mode the IOMMU is initialized and enabled but not used for
3343 * DMA-API translation.
3345 *****************************************************************************/
3347 /* IOMMUv2 specific functions */
3348 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3350 return atomic_notifier_chain_register(&ppr_notifier, nb);
3352 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3354 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3356 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3358 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3360 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3362 struct protection_domain *domain = to_pdomain(dom);
3363 unsigned long flags;
3365 spin_lock_irqsave(&domain->lock, flags);
3367 /* Update data structure */
3368 domain->mode = PAGE_MODE_NONE;
3369 domain->updated = true;
3371 /* Make changes visible to IOMMUs */
3372 update_domain(domain);
3374 /* Page-table is not visible to IOMMU anymore, so free it */
3375 free_pagetable(domain);
3377 spin_unlock_irqrestore(&domain->lock, flags);
3379 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3381 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3383 struct protection_domain *domain = to_pdomain(dom);
3384 unsigned long flags;
3387 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3390 /* Number of GCR3 table levels required */
3391 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3394 if (levels > amd_iommu_max_glx_val)
3397 spin_lock_irqsave(&domain->lock, flags);
3400 * Save us all sanity checks whether devices already in the
3401 * domain support IOMMUv2. Just force that the domain has no
3402 * devices attached when it is switched into IOMMUv2 mode.
3405 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3409 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3410 if (domain->gcr3_tbl == NULL)
3413 domain->glx = levels;
3414 domain->flags |= PD_IOMMUV2_MASK;
3415 domain->updated = true;
3417 update_domain(domain);
3422 spin_unlock_irqrestore(&domain->lock, flags);
3426 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3428 static int __flush_pasid(struct protection_domain *domain, int pasid,
3429 u64 address, bool size)
3431 struct iommu_dev_data *dev_data;
3432 struct iommu_cmd cmd;
3435 if (!(domain->flags & PD_IOMMUV2_MASK))
3438 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3441 * IOMMU TLB needs to be flushed before Device TLB to
3442 * prevent device TLB refill from IOMMU TLB
3444 for (i = 0; i < amd_iommus_present; ++i) {
3445 if (domain->dev_iommu[i] == 0)
3448 ret = iommu_queue_command(amd_iommus[i], &cmd);
3453 /* Wait until IOMMU TLB flushes are complete */
3454 domain_flush_complete(domain);
3456 /* Now flush device TLBs */
3457 list_for_each_entry(dev_data, &domain->dev_list, list) {
3458 struct amd_iommu *iommu;
3462 There might be non-IOMMUv2 capable devices in an IOMMUv2
3465 if (!dev_data->ats.enabled)
3468 qdep = dev_data->ats.qdep;
3469 iommu = amd_iommu_rlookup_table[dev_data->devid];
3471 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3472 qdep, address, size);
3474 ret = iommu_queue_command(iommu, &cmd);
3479 /* Wait until all device TLBs are flushed */
3480 domain_flush_complete(domain);
3489 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3492 INC_STATS_COUNTER(invalidate_iotlb);
3494 return __flush_pasid(domain, pasid, address, false);
3497 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3500 struct protection_domain *domain = to_pdomain(dom);
3501 unsigned long flags;
3504 spin_lock_irqsave(&domain->lock, flags);
3505 ret = __amd_iommu_flush_page(domain, pasid, address);
3506 spin_unlock_irqrestore(&domain->lock, flags);
3510 EXPORT_SYMBOL(amd_iommu_flush_page);
3512 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3514 INC_STATS_COUNTER(invalidate_iotlb_all);
3516 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3520 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3522 struct protection_domain *domain = to_pdomain(dom);
3523 unsigned long flags;
3526 spin_lock_irqsave(&domain->lock, flags);
3527 ret = __amd_iommu_flush_tlb(domain, pasid);
3528 spin_unlock_irqrestore(&domain->lock, flags);
3532 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3534 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3541 index = (pasid >> (9 * level)) & 0x1ff;
3547 if (!(*pte & GCR3_VALID)) {
3551 root = (void *)get_zeroed_page(GFP_ATOMIC);
3555 *pte = __pa(root) | GCR3_VALID;
3558 root = __va(*pte & PAGE_MASK);
3566 static int __set_gcr3(struct protection_domain *domain, int pasid,
3571 if (domain->mode != PAGE_MODE_NONE)
3574 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3578 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3580 return __amd_iommu_flush_tlb(domain, pasid);
3583 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3587 if (domain->mode != PAGE_MODE_NONE)
3590 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3596 return __amd_iommu_flush_tlb(domain, pasid);
3599 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3602 struct protection_domain *domain = to_pdomain(dom);
3603 unsigned long flags;
3606 spin_lock_irqsave(&domain->lock, flags);
3607 ret = __set_gcr3(domain, pasid, cr3);
3608 spin_unlock_irqrestore(&domain->lock, flags);
3612 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3614 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3616 struct protection_domain *domain = to_pdomain(dom);
3617 unsigned long flags;
3620 spin_lock_irqsave(&domain->lock, flags);
3621 ret = __clear_gcr3(domain, pasid);
3622 spin_unlock_irqrestore(&domain->lock, flags);
3626 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3628 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3629 int status, int tag)
3631 struct iommu_dev_data *dev_data;
3632 struct amd_iommu *iommu;
3633 struct iommu_cmd cmd;
3635 INC_STATS_COUNTER(complete_ppr);
3637 dev_data = get_dev_data(&pdev->dev);
3638 iommu = amd_iommu_rlookup_table[dev_data->devid];
3640 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3641 tag, dev_data->pri_tlp);
3643 return iommu_queue_command(iommu, &cmd);
3645 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3647 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3649 struct protection_domain *pdomain;
3651 pdomain = get_domain(&pdev->dev);
3652 if (IS_ERR(pdomain))
3655 /* Only return IOMMUv2 domains */
3656 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3659 return &pdomain->domain;
3661 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3663 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3665 struct iommu_dev_data *dev_data;
3667 if (!amd_iommu_v2_supported())
3670 dev_data = get_dev_data(&pdev->dev);
3671 dev_data->errata |= (1 << erratum);
3673 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3675 int amd_iommu_device_info(struct pci_dev *pdev,
3676 struct amd_iommu_device_info *info)
3681 if (pdev == NULL || info == NULL)
3684 if (!amd_iommu_v2_supported())
3687 memset(info, 0, sizeof(*info));
3689 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3691 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3693 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3695 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3697 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3701 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3702 max_pasids = min(max_pasids, (1 << 20));
3704 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3705 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3707 features = pci_pasid_features(pdev);
3708 if (features & PCI_PASID_CAP_EXEC)
3709 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3710 if (features & PCI_PASID_CAP_PRIV)
3711 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3716 EXPORT_SYMBOL(amd_iommu_device_info);
3718 #ifdef CONFIG_IRQ_REMAP
3720 /*****************************************************************************
3722 * Interrupt Remapping Implementation
3724 *****************************************************************************/
3742 u16 devid; /* Device ID for IRTE table */
3743 u16 index; /* Index into IRTE table*/
3746 struct amd_ir_data {
3747 struct irq_2_irte irq_2_irte;
3748 union irte irte_entry;
3750 struct msi_msg msi_entry;
3754 static struct irq_chip amd_ir_chip;
3756 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3757 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3758 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3759 #define DTE_IRQ_REMAP_ENABLE 1ULL
3761 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3765 dte = amd_iommu_dev_table[devid].data[2];
3766 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3767 dte |= virt_to_phys(table->table);
3768 dte |= DTE_IRQ_REMAP_INTCTL;
3769 dte |= DTE_IRQ_TABLE_LEN;
3770 dte |= DTE_IRQ_REMAP_ENABLE;
3772 amd_iommu_dev_table[devid].data[2] = dte;
3775 #define IRTE_ALLOCATED (~1U)
3777 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3779 struct irq_remap_table *table = NULL;
3780 struct amd_iommu *iommu;
3781 unsigned long flags;
3784 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3786 iommu = amd_iommu_rlookup_table[devid];
3790 table = irq_lookup_table[devid];
3794 alias = amd_iommu_alias_table[devid];
3795 table = irq_lookup_table[alias];
3797 irq_lookup_table[devid] = table;
3798 set_dte_irq_entry(devid, table);
3799 iommu_flush_dte(iommu, devid);
3803 /* Nothing there yet, allocate new irq remapping table */
3804 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3808 /* Initialize table spin-lock */
3809 spin_lock_init(&table->lock);
3812 /* Keep the first 32 indexes free for IOAPIC interrupts */
3813 table->min_index = 32;
3815 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3816 if (!table->table) {
3822 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3827 for (i = 0; i < 32; ++i)
3828 table->table[i] = IRTE_ALLOCATED;
3831 irq_lookup_table[devid] = table;
3832 set_dte_irq_entry(devid, table);
3833 iommu_flush_dte(iommu, devid);
3834 if (devid != alias) {
3835 irq_lookup_table[alias] = table;
3836 set_dte_irq_entry(alias, table);
3837 iommu_flush_dte(iommu, alias);
3841 iommu_completion_wait(iommu);
3844 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3849 static int alloc_irq_index(u16 devid, int count)
3851 struct irq_remap_table *table;
3852 unsigned long flags;
3855 table = get_irq_table(devid, false);
3859 spin_lock_irqsave(&table->lock, flags);
3861 /* Scan table for free entries */
3862 for (c = 0, index = table->min_index;
3863 index < MAX_IRQS_PER_TABLE;
3865 if (table->table[index] == 0)
3872 table->table[index - c + 1] = IRTE_ALLOCATED;
3882 spin_unlock_irqrestore(&table->lock, flags);
3887 static int modify_irte(u16 devid, int index, union irte irte)
3889 struct irq_remap_table *table;
3890 struct amd_iommu *iommu;
3891 unsigned long flags;
3893 iommu = amd_iommu_rlookup_table[devid];
3897 table = get_irq_table(devid, false);
3901 spin_lock_irqsave(&table->lock, flags);
3902 table->table[index] = irte.val;
3903 spin_unlock_irqrestore(&table->lock, flags);
3905 iommu_flush_irt(iommu, devid);
3906 iommu_completion_wait(iommu);
3911 static void free_irte(u16 devid, int index)
3913 struct irq_remap_table *table;
3914 struct amd_iommu *iommu;
3915 unsigned long flags;
3917 iommu = amd_iommu_rlookup_table[devid];
3921 table = get_irq_table(devid, false);
3925 spin_lock_irqsave(&table->lock, flags);
3926 table->table[index] = 0;
3927 spin_unlock_irqrestore(&table->lock, flags);
3929 iommu_flush_irt(iommu, devid);
3930 iommu_completion_wait(iommu);
3933 static int get_devid(struct irq_alloc_info *info)
3937 switch (info->type) {
3938 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3939 devid = get_ioapic_devid(info->ioapic_id);
3941 case X86_IRQ_ALLOC_TYPE_HPET:
3942 devid = get_hpet_devid(info->hpet_id);
3944 case X86_IRQ_ALLOC_TYPE_MSI:
3945 case X86_IRQ_ALLOC_TYPE_MSIX:
3946 devid = get_device_id(&info->msi_dev->dev);
3956 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3958 struct amd_iommu *iommu;
3964 devid = get_devid(info);
3966 iommu = amd_iommu_rlookup_table[devid];
3968 return iommu->ir_domain;
3974 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3976 struct amd_iommu *iommu;
3982 switch (info->type) {
3983 case X86_IRQ_ALLOC_TYPE_MSI:
3984 case X86_IRQ_ALLOC_TYPE_MSIX:
3985 devid = get_device_id(&info->msi_dev->dev);
3989 iommu = amd_iommu_rlookup_table[devid];
3991 return iommu->msi_domain;
4000 struct irq_remap_ops amd_iommu_irq_ops = {
4001 .prepare = amd_iommu_prepare,
4002 .enable = amd_iommu_enable,
4003 .disable = amd_iommu_disable,
4004 .reenable = amd_iommu_reenable,
4005 .enable_faulting = amd_iommu_enable_faulting,
4006 .get_ir_irq_domain = get_ir_irq_domain,
4007 .get_irq_domain = get_irq_domain,
4010 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4011 struct irq_cfg *irq_cfg,
4012 struct irq_alloc_info *info,
4013 int devid, int index, int sub_handle)
4015 struct irq_2_irte *irte_info = &data->irq_2_irte;
4016 struct msi_msg *msg = &data->msi_entry;
4017 union irte *irte = &data->irte_entry;
4018 struct IO_APIC_route_entry *entry;
4020 data->irq_2_irte.devid = devid;
4021 data->irq_2_irte.index = index + sub_handle;
4023 /* Setup IRTE for IOMMU */
4025 irte->fields.vector = irq_cfg->vector;
4026 irte->fields.int_type = apic->irq_delivery_mode;
4027 irte->fields.destination = irq_cfg->dest_apicid;
4028 irte->fields.dm = apic->irq_dest_mode;
4029 irte->fields.valid = 1;
4031 switch (info->type) {
4032 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4033 /* Setup IOAPIC entry */
4034 entry = info->ioapic_entry;
4035 info->ioapic_entry = NULL;
4036 memset(entry, 0, sizeof(*entry));
4037 entry->vector = index;
4039 entry->trigger = info->ioapic_trigger;
4040 entry->polarity = info->ioapic_polarity;
4041 /* Mask level triggered irqs. */
4042 if (info->ioapic_trigger)
4046 case X86_IRQ_ALLOC_TYPE_HPET:
4047 case X86_IRQ_ALLOC_TYPE_MSI:
4048 case X86_IRQ_ALLOC_TYPE_MSIX:
4049 msg->address_hi = MSI_ADDR_BASE_HI;
4050 msg->address_lo = MSI_ADDR_BASE_LO;
4051 msg->data = irte_info->index;
4060 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4061 unsigned int nr_irqs, void *arg)
4063 struct irq_alloc_info *info = arg;
4064 struct irq_data *irq_data;
4065 struct amd_ir_data *data;
4066 struct irq_cfg *cfg;
4072 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4073 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4077 * With IRQ remapping enabled, don't need contiguous CPU vectors
4078 * to support multiple MSI interrupts.
4080 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4081 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4083 devid = get_devid(info);
4087 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4091 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4092 if (get_irq_table(devid, true))
4093 index = info->ioapic_pin;
4097 index = alloc_irq_index(devid, nr_irqs);
4100 pr_warn("Failed to allocate IRTE\n");
4101 goto out_free_parent;
4104 for (i = 0; i < nr_irqs; i++) {
4105 irq_data = irq_domain_get_irq_data(domain, virq + i);
4106 cfg = irqd_cfg(irq_data);
4107 if (!irq_data || !cfg) {
4113 data = kzalloc(sizeof(*data), GFP_KERNEL);
4117 irq_data->hwirq = (devid << 16) + i;
4118 irq_data->chip_data = data;
4119 irq_data->chip = &amd_ir_chip;
4120 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4121 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4127 for (i--; i >= 0; i--) {
4128 irq_data = irq_domain_get_irq_data(domain, virq + i);
4130 kfree(irq_data->chip_data);
4132 for (i = 0; i < nr_irqs; i++)
4133 free_irte(devid, index + i);
4135 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4139 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4140 unsigned int nr_irqs)
4142 struct irq_2_irte *irte_info;
4143 struct irq_data *irq_data;
4144 struct amd_ir_data *data;
4147 for (i = 0; i < nr_irqs; i++) {
4148 irq_data = irq_domain_get_irq_data(domain, virq + i);
4149 if (irq_data && irq_data->chip_data) {
4150 data = irq_data->chip_data;
4151 irte_info = &data->irq_2_irte;
4152 free_irte(irte_info->devid, irte_info->index);
4156 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4159 static void irq_remapping_activate(struct irq_domain *domain,
4160 struct irq_data *irq_data)
4162 struct amd_ir_data *data = irq_data->chip_data;
4163 struct irq_2_irte *irte_info = &data->irq_2_irte;
4165 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4168 static void irq_remapping_deactivate(struct irq_domain *domain,
4169 struct irq_data *irq_data)
4171 struct amd_ir_data *data = irq_data->chip_data;
4172 struct irq_2_irte *irte_info = &data->irq_2_irte;
4176 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4179 static struct irq_domain_ops amd_ir_domain_ops = {
4180 .alloc = irq_remapping_alloc,
4181 .free = irq_remapping_free,
4182 .activate = irq_remapping_activate,
4183 .deactivate = irq_remapping_deactivate,
4186 static int amd_ir_set_affinity(struct irq_data *data,
4187 const struct cpumask *mask, bool force)
4189 struct amd_ir_data *ir_data = data->chip_data;
4190 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4191 struct irq_cfg *cfg = irqd_cfg(data);
4192 struct irq_data *parent = data->parent_data;
4195 ret = parent->chip->irq_set_affinity(parent, mask, force);
4196 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4200 * Atomically updates the IRTE with the new destination, vector
4201 * and flushes the interrupt entry cache.
4203 ir_data->irte_entry.fields.vector = cfg->vector;
4204 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4205 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4208 * After this point, all the interrupts will start arriving
4209 * at the new destination. So, time to cleanup the previous
4210 * vector allocation.
4212 send_cleanup_vector(cfg);
4214 return IRQ_SET_MASK_OK_DONE;
4217 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4219 struct amd_ir_data *ir_data = irq_data->chip_data;
4221 *msg = ir_data->msi_entry;
4224 static struct irq_chip amd_ir_chip = {
4225 .irq_ack = ir_ack_apic_edge,
4226 .irq_set_affinity = amd_ir_set_affinity,
4227 .irq_compose_msi_msg = ir_compose_msi_msg,
4230 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4232 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4233 if (!iommu->ir_domain)
4236 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4237 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);