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
);
108 static atomic_t queue_timer_on
;
109 static struct timer_list queue_timer
;
112 * Domain for untranslated devices - only allocated
113 * if iommu=pt passed on kernel cmd line.
115 static const struct iommu_ops amd_iommu_ops
;
117 static ATOMIC_NOTIFIER_HEAD(ppr_notifier
);
118 int amd_iommu_max_glx_val
= -1;
120 static struct dma_map_ops amd_iommu_dma_ops
;
123 * This struct contains device specific data for the IOMMU
125 struct iommu_dev_data
{
126 struct list_head list
; /* For domain->dev_list */
127 struct list_head dev_data_list
; /* For global dev_data_list */
128 struct protection_domain
*domain
; /* Domain the device is bound to */
129 u16 devid
; /* PCI Device ID */
130 u16 alias
; /* Alias Device ID */
131 bool iommu_v2
; /* Device can make use of IOMMUv2 */
132 bool passthrough
; /* Device is identity mapped */
136 } ats
; /* ATS state */
137 bool pri_tlp
; /* PASID TLB required for
139 u32 errata
; /* Bitmap for errata to apply */
140 bool use_vapic
; /* Enable device to use vapic mode */
144 * general struct to manage commands send to an IOMMU
150 struct kmem_cache
*amd_iommu_irq_cache
;
152 static void update_domain(struct protection_domain
*domain
);
153 static int protection_domain_init(struct protection_domain
*domain
);
154 static void detach_device(struct device
*dev
);
157 * Data container for a dma_ops specific protection domain
159 struct dma_ops_domain
{
160 /* generic protection domain information */
161 struct protection_domain domain
;
164 struct iova_domain iovad
;
167 static struct iova_domain reserved_iova_ranges
;
168 static struct lock_class_key reserved_rbtree_key
;
170 /****************************************************************************
174 ****************************************************************************/
176 static inline int match_hid_uid(struct device
*dev
,
177 struct acpihid_map_entry
*entry
)
179 const char *hid
, *uid
;
181 hid
= acpi_device_hid(ACPI_COMPANION(dev
));
182 uid
= acpi_device_uid(ACPI_COMPANION(dev
));
188 return strcmp(hid
, entry
->hid
);
191 return strcmp(hid
, entry
->hid
);
193 return (strcmp(hid
, entry
->hid
) || strcmp(uid
, entry
->uid
));
196 static inline u16
get_pci_device_id(struct device
*dev
)
198 struct pci_dev
*pdev
= to_pci_dev(dev
);
200 return PCI_DEVID(pdev
->bus
->number
, pdev
->devfn
);
203 static inline int get_acpihid_device_id(struct device
*dev
,
204 struct acpihid_map_entry
**entry
)
206 struct acpihid_map_entry
*p
;
208 list_for_each_entry(p
, &acpihid_map
, list
) {
209 if (!match_hid_uid(dev
, p
)) {
218 static inline int get_device_id(struct device
*dev
)
223 devid
= get_pci_device_id(dev
);
225 devid
= get_acpihid_device_id(dev
, NULL
);
230 static struct protection_domain
*to_pdomain(struct iommu_domain
*dom
)
232 return container_of(dom
, struct protection_domain
, domain
);
235 static struct dma_ops_domain
* to_dma_ops_domain(struct protection_domain
*domain
)
237 BUG_ON(domain
->flags
!= PD_DMA_OPS_MASK
);
238 return container_of(domain
, struct dma_ops_domain
, domain
);
241 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
243 struct iommu_dev_data
*dev_data
;
246 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
250 dev_data
->devid
= devid
;
252 spin_lock_irqsave(&dev_data_list_lock
, flags
);
253 list_add_tail(&dev_data
->dev_data_list
, &dev_data_list
);
254 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
259 static struct iommu_dev_data
*search_dev_data(u16 devid
)
261 struct iommu_dev_data
*dev_data
;
264 spin_lock_irqsave(&dev_data_list_lock
, flags
);
265 list_for_each_entry(dev_data
, &dev_data_list
, dev_data_list
) {
266 if (dev_data
->devid
== devid
)
273 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
278 static int __last_alias(struct pci_dev
*pdev
, u16 alias
, void *data
)
280 *(u16
*)data
= alias
;
284 static u16
get_alias(struct device
*dev
)
286 struct pci_dev
*pdev
= to_pci_dev(dev
);
287 u16 devid
, ivrs_alias
, pci_alias
;
289 /* The callers make sure that get_device_id() does not fail here */
290 devid
= get_device_id(dev
);
291 ivrs_alias
= amd_iommu_alias_table
[devid
];
292 pci_for_each_dma_alias(pdev
, __last_alias
, &pci_alias
);
294 if (ivrs_alias
== pci_alias
)
300 * The IVRS is fairly reliable in telling us about aliases, but it
301 * can't know about every screwy device. If we don't have an IVRS
302 * reported alias, use the PCI reported alias. In that case we may
303 * still need to initialize the rlookup and dev_table entries if the
304 * alias is to a non-existent device.
306 if (ivrs_alias
== devid
) {
307 if (!amd_iommu_rlookup_table
[pci_alias
]) {
308 amd_iommu_rlookup_table
[pci_alias
] =
309 amd_iommu_rlookup_table
[devid
];
310 memcpy(amd_iommu_dev_table
[pci_alias
].data
,
311 amd_iommu_dev_table
[devid
].data
,
312 sizeof(amd_iommu_dev_table
[pci_alias
].data
));
318 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
319 "for device %s[%04x:%04x], kernel reported alias "
320 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias
), PCI_SLOT(ivrs_alias
),
321 PCI_FUNC(ivrs_alias
), dev_name(dev
), pdev
->vendor
, pdev
->device
,
322 PCI_BUS_NUM(pci_alias
), PCI_SLOT(pci_alias
),
323 PCI_FUNC(pci_alias
));
326 * If we don't have a PCI DMA alias and the IVRS alias is on the same
327 * bus, then the IVRS table may know about a quirk that we don't.
329 if (pci_alias
== devid
&&
330 PCI_BUS_NUM(ivrs_alias
) == pdev
->bus
->number
) {
331 pci_add_dma_alias(pdev
, ivrs_alias
& 0xff);
332 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
333 PCI_SLOT(ivrs_alias
), PCI_FUNC(ivrs_alias
),
340 static struct iommu_dev_data
*find_dev_data(u16 devid
)
342 struct iommu_dev_data
*dev_data
;
344 dev_data
= search_dev_data(devid
);
346 if (dev_data
== NULL
)
347 dev_data
= alloc_dev_data(devid
);
352 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
354 return dev
->archdata
.iommu
;
358 * Find or create an IOMMU group for a acpihid device.
360 static struct iommu_group
*acpihid_device_group(struct device
*dev
)
362 struct acpihid_map_entry
*p
, *entry
= NULL
;
365 devid
= get_acpihid_device_id(dev
, &entry
);
367 return ERR_PTR(devid
);
369 list_for_each_entry(p
, &acpihid_map
, list
) {
370 if ((devid
== p
->devid
) && p
->group
)
371 entry
->group
= p
->group
;
375 entry
->group
= generic_device_group(dev
);
380 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
382 static const int caps
[] = {
385 PCI_EXT_CAP_ID_PASID
,
389 for (i
= 0; i
< 3; ++i
) {
390 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
398 static bool pdev_pri_erratum(struct pci_dev
*pdev
, u32 erratum
)
400 struct iommu_dev_data
*dev_data
;
402 dev_data
= get_dev_data(&pdev
->dev
);
404 return dev_data
->errata
& (1 << erratum
) ? true : false;
408 * This function checks if the driver got a valid device from the caller to
409 * avoid dereferencing invalid pointers.
411 static bool check_device(struct device
*dev
)
415 if (!dev
|| !dev
->dma_mask
)
418 devid
= get_device_id(dev
);
422 /* Out of our scope? */
423 if (devid
> amd_iommu_last_bdf
)
426 if (amd_iommu_rlookup_table
[devid
] == NULL
)
432 static void init_iommu_group(struct device
*dev
)
434 struct iommu_group
*group
;
436 group
= iommu_group_get_for_dev(dev
);
440 iommu_group_put(group
);
443 static int iommu_init_device(struct device
*dev
)
445 struct iommu_dev_data
*dev_data
;
448 if (dev
->archdata
.iommu
)
451 devid
= get_device_id(dev
);
455 dev_data
= find_dev_data(devid
);
459 dev_data
->alias
= get_alias(dev
);
461 if (dev_is_pci(dev
) && pci_iommuv2_capable(to_pci_dev(dev
))) {
462 struct amd_iommu
*iommu
;
464 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
465 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
468 dev
->archdata
.iommu
= dev_data
;
470 iommu_device_link(amd_iommu_rlookup_table
[dev_data
->devid
]->iommu_dev
,
476 static void iommu_ignore_device(struct device
*dev
)
481 devid
= get_device_id(dev
);
485 alias
= get_alias(dev
);
487 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
488 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
490 amd_iommu_rlookup_table
[devid
] = NULL
;
491 amd_iommu_rlookup_table
[alias
] = NULL
;
494 static void iommu_uninit_device(struct device
*dev
)
497 struct iommu_dev_data
*dev_data
;
499 devid
= get_device_id(dev
);
503 dev_data
= search_dev_data(devid
);
507 if (dev_data
->domain
)
510 iommu_device_unlink(amd_iommu_rlookup_table
[dev_data
->devid
]->iommu_dev
,
513 iommu_group_remove_device(dev
);
516 dev
->archdata
.dma_ops
= NULL
;
519 * We keep dev_data around for unplugged devices and reuse it when the
520 * device is re-plugged - not doing so would introduce a ton of races.
524 /****************************************************************************
526 * Interrupt handling functions
528 ****************************************************************************/
530 static void dump_dte_entry(u16 devid
)
534 for (i
= 0; i
< 4; ++i
)
535 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
536 amd_iommu_dev_table
[devid
].data
[i
]);
539 static void dump_command(unsigned long phys_addr
)
541 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
544 for (i
= 0; i
< 4; ++i
)
545 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
548 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
550 int type
, devid
, domid
, flags
;
551 volatile u32
*event
= __evt
;
556 type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
557 devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
558 domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
559 flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
560 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
563 /* Did we hit the erratum? */
564 if (++count
== LOOP_TIMEOUT
) {
565 pr_err("AMD-Vi: No event written to event log\n");
572 printk(KERN_ERR
"AMD-Vi: Event logged [");
575 case EVENT_TYPE_ILL_DEV
:
576 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
577 "address=0x%016llx flags=0x%04x]\n",
578 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
580 dump_dte_entry(devid
);
582 case EVENT_TYPE_IO_FAULT
:
583 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
584 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
585 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
586 domid
, address
, flags
);
588 case EVENT_TYPE_DEV_TAB_ERR
:
589 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
590 "address=0x%016llx flags=0x%04x]\n",
591 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
594 case EVENT_TYPE_PAGE_TAB_ERR
:
595 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
596 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
598 domid
, address
, flags
);
600 case EVENT_TYPE_ILL_CMD
:
601 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
602 dump_command(address
);
604 case EVENT_TYPE_CMD_HARD_ERR
:
605 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
606 "flags=0x%04x]\n", address
, flags
);
608 case EVENT_TYPE_IOTLB_INV_TO
:
609 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
610 "address=0x%016llx]\n",
611 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
614 case EVENT_TYPE_INV_DEV_REQ
:
615 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
616 "address=0x%016llx flags=0x%04x]\n",
617 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
621 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
624 memset(__evt
, 0, 4 * sizeof(u32
));
627 static void iommu_poll_events(struct amd_iommu
*iommu
)
631 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
632 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
634 while (head
!= tail
) {
635 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
636 head
= (head
+ EVENT_ENTRY_SIZE
) % EVT_BUFFER_SIZE
;
639 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
642 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u64
*raw
)
644 struct amd_iommu_fault fault
;
646 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
647 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
651 fault
.address
= raw
[1];
652 fault
.pasid
= PPR_PASID(raw
[0]);
653 fault
.device_id
= PPR_DEVID(raw
[0]);
654 fault
.tag
= PPR_TAG(raw
[0]);
655 fault
.flags
= PPR_FLAGS(raw
[0]);
657 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
660 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
664 if (iommu
->ppr_log
== NULL
)
667 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
668 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
670 while (head
!= tail
) {
675 raw
= (u64
*)(iommu
->ppr_log
+ head
);
678 * Hardware bug: Interrupt may arrive before the entry is
679 * written to memory. If this happens we need to wait for the
682 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
683 if (PPR_REQ_TYPE(raw
[0]) != 0)
688 /* Avoid memcpy function-call overhead */
693 * To detect the hardware bug we need to clear the entry
696 raw
[0] = raw
[1] = 0UL;
698 /* Update head pointer of hardware ring-buffer */
699 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
700 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
702 /* Handle PPR entry */
703 iommu_handle_ppr_entry(iommu
, entry
);
705 /* Refresh ring-buffer information */
706 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
707 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
711 #ifdef CONFIG_IRQ_REMAP
712 static int (*iommu_ga_log_notifier
)(u32
);
714 int amd_iommu_register_ga_log_notifier(int (*notifier
)(u32
))
716 iommu_ga_log_notifier
= notifier
;
720 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier
);
722 static void iommu_poll_ga_log(struct amd_iommu
*iommu
)
724 u32 head
, tail
, cnt
= 0;
726 if (iommu
->ga_log
== NULL
)
729 head
= readl(iommu
->mmio_base
+ MMIO_GA_HEAD_OFFSET
);
730 tail
= readl(iommu
->mmio_base
+ MMIO_GA_TAIL_OFFSET
);
732 while (head
!= tail
) {
736 raw
= (u64
*)(iommu
->ga_log
+ head
);
739 /* Avoid memcpy function-call overhead */
742 /* Update head pointer of hardware ring-buffer */
743 head
= (head
+ GA_ENTRY_SIZE
) % GA_LOG_SIZE
;
744 writel(head
, iommu
->mmio_base
+ MMIO_GA_HEAD_OFFSET
);
746 /* Handle GA entry */
747 switch (GA_REQ_TYPE(log_entry
)) {
749 if (!iommu_ga_log_notifier
)
752 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
753 __func__
, GA_DEVID(log_entry
),
756 if (iommu_ga_log_notifier(GA_TAG(log_entry
)) != 0)
757 pr_err("AMD-Vi: GA log notifier failed.\n");
764 #endif /* CONFIG_IRQ_REMAP */
766 #define AMD_IOMMU_INT_MASK \
767 (MMIO_STATUS_EVT_INT_MASK | \
768 MMIO_STATUS_PPR_INT_MASK | \
769 MMIO_STATUS_GALOG_INT_MASK)
771 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
773 struct amd_iommu
*iommu
= (struct amd_iommu
*) data
;
774 u32 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
776 while (status
& AMD_IOMMU_INT_MASK
) {
777 /* Enable EVT and PPR and GA interrupts again */
778 writel(AMD_IOMMU_INT_MASK
,
779 iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
781 if (status
& MMIO_STATUS_EVT_INT_MASK
) {
782 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
783 iommu_poll_events(iommu
);
786 if (status
& MMIO_STATUS_PPR_INT_MASK
) {
787 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
788 iommu_poll_ppr_log(iommu
);
791 #ifdef CONFIG_IRQ_REMAP
792 if (status
& MMIO_STATUS_GALOG_INT_MASK
) {
793 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
794 iommu_poll_ga_log(iommu
);
799 * Hardware bug: ERBT1312
800 * When re-enabling interrupt (by writing 1
801 * to clear the bit), the hardware might also try to set
802 * the interrupt bit in the event status register.
803 * In this scenario, the bit will be set, and disable
804 * subsequent interrupts.
806 * Workaround: The IOMMU driver should read back the
807 * status register and check if the interrupt bits are cleared.
808 * If not, driver will need to go through the interrupt handler
809 * again and re-clear the bits
811 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
816 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
818 return IRQ_WAKE_THREAD
;
821 /****************************************************************************
823 * IOMMU command queuing functions
825 ****************************************************************************/
827 static int wait_on_sem(volatile u64
*sem
)
831 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
836 if (i
== LOOP_TIMEOUT
) {
837 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
844 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
845 struct iommu_cmd
*cmd
,
850 target
= iommu
->cmd_buf
+ tail
;
851 tail
= (tail
+ sizeof(*cmd
)) % CMD_BUFFER_SIZE
;
853 /* Copy command to buffer */
854 memcpy(target
, cmd
, sizeof(*cmd
));
856 /* Tell the IOMMU about it */
857 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
860 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
862 WARN_ON(address
& 0x7ULL
);
864 memset(cmd
, 0, sizeof(*cmd
));
865 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
866 cmd
->data
[1] = upper_32_bits(__pa(address
));
868 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
871 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
873 memset(cmd
, 0, sizeof(*cmd
));
874 cmd
->data
[0] = devid
;
875 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
878 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
879 size_t size
, u16 domid
, int pde
)
884 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
889 * If we have to flush more than one page, flush all
890 * TLB entries for this domain
892 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
896 address
&= PAGE_MASK
;
898 memset(cmd
, 0, sizeof(*cmd
));
899 cmd
->data
[1] |= domid
;
900 cmd
->data
[2] = lower_32_bits(address
);
901 cmd
->data
[3] = upper_32_bits(address
);
902 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
903 if (s
) /* size bit - we flush more than one 4kb page */
904 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
905 if (pde
) /* PDE bit - we want to flush everything, not only the PTEs */
906 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
909 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
910 u64 address
, size_t size
)
915 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
920 * If we have to flush more than one page, flush all
921 * TLB entries for this domain
923 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
927 address
&= PAGE_MASK
;
929 memset(cmd
, 0, sizeof(*cmd
));
930 cmd
->data
[0] = devid
;
931 cmd
->data
[0] |= (qdep
& 0xff) << 24;
932 cmd
->data
[1] = devid
;
933 cmd
->data
[2] = lower_32_bits(address
);
934 cmd
->data
[3] = upper_32_bits(address
);
935 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
937 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
940 static void build_inv_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
941 u64 address
, bool size
)
943 memset(cmd
, 0, sizeof(*cmd
));
945 address
&= ~(0xfffULL
);
947 cmd
->data
[0] = pasid
;
948 cmd
->data
[1] = domid
;
949 cmd
->data
[2] = lower_32_bits(address
);
950 cmd
->data
[3] = upper_32_bits(address
);
951 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
952 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
954 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
955 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
958 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
959 int qdep
, u64 address
, bool size
)
961 memset(cmd
, 0, sizeof(*cmd
));
963 address
&= ~(0xfffULL
);
965 cmd
->data
[0] = devid
;
966 cmd
->data
[0] |= ((pasid
>> 8) & 0xff) << 16;
967 cmd
->data
[0] |= (qdep
& 0xff) << 24;
968 cmd
->data
[1] = devid
;
969 cmd
->data
[1] |= (pasid
& 0xff) << 16;
970 cmd
->data
[2] = lower_32_bits(address
);
971 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
972 cmd
->data
[3] = upper_32_bits(address
);
974 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
975 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
978 static void build_complete_ppr(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
979 int status
, int tag
, bool gn
)
981 memset(cmd
, 0, sizeof(*cmd
));
983 cmd
->data
[0] = devid
;
985 cmd
->data
[1] = pasid
;
986 cmd
->data
[2] = CMD_INV_IOMMU_PAGES_GN_MASK
;
988 cmd
->data
[3] = tag
& 0x1ff;
989 cmd
->data
[3] |= (status
& PPR_STATUS_MASK
) << PPR_STATUS_SHIFT
;
991 CMD_SET_TYPE(cmd
, CMD_COMPLETE_PPR
);
994 static void build_inv_all(struct iommu_cmd
*cmd
)
996 memset(cmd
, 0, sizeof(*cmd
));
997 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
1000 static void build_inv_irt(struct iommu_cmd
*cmd
, u16 devid
)
1002 memset(cmd
, 0, sizeof(*cmd
));
1003 cmd
->data
[0] = devid
;
1004 CMD_SET_TYPE(cmd
, CMD_INV_IRT
);
1008 * Writes the command to the IOMMUs command buffer and informs the
1009 * hardware about the new command.
1011 static int __iommu_queue_command_sync(struct amd_iommu
*iommu
,
1012 struct iommu_cmd
*cmd
,
1015 u32 left
, tail
, head
, next_tail
;
1019 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
1020 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
1021 next_tail
= (tail
+ sizeof(*cmd
)) % CMD_BUFFER_SIZE
;
1022 left
= (head
- next_tail
) % CMD_BUFFER_SIZE
;
1025 struct iommu_cmd sync_cmd
;
1030 build_completion_wait(&sync_cmd
, (u64
)&iommu
->cmd_sem
);
1031 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
1033 if ((ret
= wait_on_sem(&iommu
->cmd_sem
)) != 0)
1039 copy_cmd_to_buffer(iommu
, cmd
, tail
);
1041 /* We need to sync now to make sure all commands are processed */
1042 iommu
->need_sync
= sync
;
1047 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
1048 struct iommu_cmd
*cmd
,
1051 unsigned long flags
;
1054 spin_lock_irqsave(&iommu
->lock
, flags
);
1055 ret
= __iommu_queue_command_sync(iommu
, cmd
, sync
);
1056 spin_unlock_irqrestore(&iommu
->lock
, flags
);
1061 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
1063 return iommu_queue_command_sync(iommu
, cmd
, true);
1067 * This function queues a completion wait command into the command
1068 * buffer of an IOMMU
1070 static int iommu_completion_wait(struct amd_iommu
*iommu
)
1072 struct iommu_cmd cmd
;
1073 unsigned long flags
;
1076 if (!iommu
->need_sync
)
1080 build_completion_wait(&cmd
, (u64
)&iommu
->cmd_sem
);
1082 spin_lock_irqsave(&iommu
->lock
, flags
);
1086 ret
= __iommu_queue_command_sync(iommu
, &cmd
, false);
1090 ret
= wait_on_sem(&iommu
->cmd_sem
);
1093 spin_unlock_irqrestore(&iommu
->lock
, flags
);
1098 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
1100 struct iommu_cmd cmd
;
1102 build_inv_dte(&cmd
, devid
);
1104 return iommu_queue_command(iommu
, &cmd
);
1107 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
1111 for (devid
= 0; devid
<= 0xffff; ++devid
)
1112 iommu_flush_dte(iommu
, devid
);
1114 iommu_completion_wait(iommu
);
1118 * This function uses heavy locking and may disable irqs for some time. But
1119 * this is no issue because it is only called during resume.
1121 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
1125 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
1126 struct iommu_cmd cmd
;
1127 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
1129 iommu_queue_command(iommu
, &cmd
);
1132 iommu_completion_wait(iommu
);
1135 static void iommu_flush_all(struct amd_iommu
*iommu
)
1137 struct iommu_cmd cmd
;
1139 build_inv_all(&cmd
);
1141 iommu_queue_command(iommu
, &cmd
);
1142 iommu_completion_wait(iommu
);
1145 static void iommu_flush_irt(struct amd_iommu
*iommu
, u16 devid
)
1147 struct iommu_cmd cmd
;
1149 build_inv_irt(&cmd
, devid
);
1151 iommu_queue_command(iommu
, &cmd
);
1154 static void iommu_flush_irt_all(struct amd_iommu
*iommu
)
1158 for (devid
= 0; devid
<= MAX_DEV_TABLE_ENTRIES
; devid
++)
1159 iommu_flush_irt(iommu
, devid
);
1161 iommu_completion_wait(iommu
);
1164 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
1166 if (iommu_feature(iommu
, FEATURE_IA
)) {
1167 iommu_flush_all(iommu
);
1169 iommu_flush_dte_all(iommu
);
1170 iommu_flush_irt_all(iommu
);
1171 iommu_flush_tlb_all(iommu
);
1176 * Command send function for flushing on-device TLB
1178 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
1179 u64 address
, size_t size
)
1181 struct amd_iommu
*iommu
;
1182 struct iommu_cmd cmd
;
1185 qdep
= dev_data
->ats
.qdep
;
1186 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1188 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
1190 return iommu_queue_command(iommu
, &cmd
);
1194 * Command send function for invalidating a device table entry
1196 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
1198 struct amd_iommu
*iommu
;
1202 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1203 alias
= dev_data
->alias
;
1205 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
1206 if (!ret
&& alias
!= dev_data
->devid
)
1207 ret
= iommu_flush_dte(iommu
, alias
);
1211 if (dev_data
->ats
.enabled
)
1212 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
1218 * TLB invalidation function which is called from the mapping functions.
1219 * It invalidates a single PTE if the range to flush is within a single
1220 * page. Otherwise it flushes the whole TLB of the IOMMU.
1222 static void __domain_flush_pages(struct protection_domain
*domain
,
1223 u64 address
, size_t size
, int pde
)
1225 struct iommu_dev_data
*dev_data
;
1226 struct iommu_cmd cmd
;
1229 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
1231 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1232 if (!domain
->dev_iommu
[i
])
1236 * Devices of this domain are behind this IOMMU
1237 * We need a TLB flush
1239 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
1242 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1244 if (!dev_data
->ats
.enabled
)
1247 ret
|= device_flush_iotlb(dev_data
, address
, size
);
1253 static void domain_flush_pages(struct protection_domain
*domain
,
1254 u64 address
, size_t size
)
1256 __domain_flush_pages(domain
, address
, size
, 0);
1259 /* Flush the whole IO/TLB for a given protection domain */
1260 static void domain_flush_tlb(struct protection_domain
*domain
)
1262 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
1265 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1266 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
1268 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
1271 static void domain_flush_complete(struct protection_domain
*domain
)
1275 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1276 if (domain
&& !domain
->dev_iommu
[i
])
1280 * Devices of this domain are behind this IOMMU
1281 * We need to wait for completion of all commands.
1283 iommu_completion_wait(amd_iommus
[i
]);
1289 * This function flushes the DTEs for all devices in domain
1291 static void domain_flush_devices(struct protection_domain
*domain
)
1293 struct iommu_dev_data
*dev_data
;
1295 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1296 device_flush_dte(dev_data
);
1299 /****************************************************************************
1301 * The functions below are used the create the page table mappings for
1302 * unity mapped regions.
1304 ****************************************************************************/
1307 * This function is used to add another level to an IO page table. Adding
1308 * another level increases the size of the address space by 9 bits to a size up
1311 static bool increase_address_space(struct protection_domain
*domain
,
1316 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1317 /* address space already 64 bit large */
1320 pte
= (void *)get_zeroed_page(gfp
);
1324 *pte
= PM_LEVEL_PDE(domain
->mode
,
1325 virt_to_phys(domain
->pt_root
));
1326 domain
->pt_root
= pte
;
1328 domain
->updated
= true;
1333 static u64
*alloc_pte(struct protection_domain
*domain
,
1334 unsigned long address
,
1335 unsigned long page_size
,
1342 BUG_ON(!is_power_of_2(page_size
));
1344 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1345 increase_address_space(domain
, gfp
);
1347 level
= domain
->mode
- 1;
1348 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1349 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1350 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1352 while (level
> end_lvl
) {
1357 if (!IOMMU_PTE_PRESENT(__pte
)) {
1358 page
= (u64
*)get_zeroed_page(gfp
);
1362 __npte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1364 if (cmpxchg64(pte
, __pte
, __npte
)) {
1365 free_page((unsigned long)page
);
1370 /* No level skipping support yet */
1371 if (PM_PTE_LEVEL(*pte
) != level
)
1376 pte
= IOMMU_PTE_PAGE(*pte
);
1378 if (pte_page
&& level
== end_lvl
)
1381 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1388 * This function checks if there is a PTE for a given dma address. If
1389 * there is one, it returns the pointer to it.
1391 static u64
*fetch_pte(struct protection_domain
*domain
,
1392 unsigned long address
,
1393 unsigned long *page_size
)
1398 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1401 level
= domain
->mode
- 1;
1402 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1403 *page_size
= PTE_LEVEL_PAGE_SIZE(level
);
1408 if (!IOMMU_PTE_PRESENT(*pte
))
1412 if (PM_PTE_LEVEL(*pte
) == 7 ||
1413 PM_PTE_LEVEL(*pte
) == 0)
1416 /* No level skipping support yet */
1417 if (PM_PTE_LEVEL(*pte
) != level
)
1422 /* Walk to the next level */
1423 pte
= IOMMU_PTE_PAGE(*pte
);
1424 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1425 *page_size
= PTE_LEVEL_PAGE_SIZE(level
);
1428 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1429 unsigned long pte_mask
;
1432 * If we have a series of large PTEs, make
1433 * sure to return a pointer to the first one.
1435 *page_size
= pte_mask
= PTE_PAGE_SIZE(*pte
);
1436 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1437 pte
= (u64
*)(((unsigned long)pte
) & pte_mask
);
1444 * Generic mapping functions. It maps a physical address into a DMA
1445 * address space. It allocates the page table pages if necessary.
1446 * In the future it can be extended to a generic mapping function
1447 * supporting all features of AMD IOMMU page tables like level skipping
1448 * and full 64 bit address spaces.
1450 static int iommu_map_page(struct protection_domain
*dom
,
1451 unsigned long bus_addr
,
1452 unsigned long phys_addr
,
1453 unsigned long page_size
,
1460 BUG_ON(!IS_ALIGNED(bus_addr
, page_size
));
1461 BUG_ON(!IS_ALIGNED(phys_addr
, page_size
));
1463 if (!(prot
& IOMMU_PROT_MASK
))
1466 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1467 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, gfp
);
1472 for (i
= 0; i
< count
; ++i
)
1473 if (IOMMU_PTE_PRESENT(pte
[i
]))
1477 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1478 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1480 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1482 if (prot
& IOMMU_PROT_IR
)
1483 __pte
|= IOMMU_PTE_IR
;
1484 if (prot
& IOMMU_PROT_IW
)
1485 __pte
|= IOMMU_PTE_IW
;
1487 for (i
= 0; i
< count
; ++i
)
1495 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1496 unsigned long bus_addr
,
1497 unsigned long page_size
)
1499 unsigned long long unmapped
;
1500 unsigned long unmap_size
;
1503 BUG_ON(!is_power_of_2(page_size
));
1507 while (unmapped
< page_size
) {
1509 pte
= fetch_pte(dom
, bus_addr
, &unmap_size
);
1514 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1515 for (i
= 0; i
< count
; i
++)
1519 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1520 unmapped
+= unmap_size
;
1523 BUG_ON(unmapped
&& !is_power_of_2(unmapped
));
1528 /****************************************************************************
1530 * The next functions belong to the address allocator for the dma_ops
1531 * interface functions.
1533 ****************************************************************************/
1536 static unsigned long dma_ops_alloc_iova(struct device
*dev
,
1537 struct dma_ops_domain
*dma_dom
,
1538 unsigned int pages
, u64 dma_mask
)
1540 unsigned long pfn
= 0;
1542 pages
= __roundup_pow_of_two(pages
);
1544 if (dma_mask
> DMA_BIT_MASK(32))
1545 pfn
= alloc_iova_fast(&dma_dom
->iovad
, pages
,
1546 IOVA_PFN(DMA_BIT_MASK(32)));
1549 pfn
= alloc_iova_fast(&dma_dom
->iovad
, pages
, IOVA_PFN(dma_mask
));
1551 return (pfn
<< PAGE_SHIFT
);
1554 static void dma_ops_free_iova(struct dma_ops_domain
*dma_dom
,
1555 unsigned long address
,
1558 pages
= __roundup_pow_of_two(pages
);
1559 address
>>= PAGE_SHIFT
;
1561 free_iova_fast(&dma_dom
->iovad
, address
, pages
);
1564 /****************************************************************************
1566 * The next functions belong to the domain allocation. A domain is
1567 * allocated for every IOMMU as the default domain. If device isolation
1568 * is enabled, every device get its own domain. The most important thing
1569 * about domains is the page table mapping the DMA address space they
1572 ****************************************************************************/
1575 * This function adds a protection domain to the global protection domain list
1577 static void add_domain_to_list(struct protection_domain
*domain
)
1579 unsigned long flags
;
1581 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1582 list_add(&domain
->list
, &amd_iommu_pd_list
);
1583 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1587 * This function removes a protection domain to the global
1588 * protection domain list
1590 static void del_domain_from_list(struct protection_domain
*domain
)
1592 unsigned long flags
;
1594 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1595 list_del(&domain
->list
);
1596 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1599 static u16
domain_id_alloc(void)
1601 unsigned long flags
;
1604 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1605 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1607 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1608 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1611 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1616 static void domain_id_free(int id
)
1618 unsigned long flags
;
1620 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1621 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1622 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1623 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1626 #define DEFINE_FREE_PT_FN(LVL, FN) \
1627 static void free_pt_##LVL (unsigned long __pt) \
1635 for (i = 0; i < 512; ++i) { \
1636 /* PTE present? */ \
1637 if (!IOMMU_PTE_PRESENT(pt[i])) \
1641 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1642 PM_PTE_LEVEL(pt[i]) == 7) \
1645 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1648 free_page((unsigned long)pt); \
1651 DEFINE_FREE_PT_FN(l2
, free_page
)
1652 DEFINE_FREE_PT_FN(l3
, free_pt_l2
)
1653 DEFINE_FREE_PT_FN(l4
, free_pt_l3
)
1654 DEFINE_FREE_PT_FN(l5
, free_pt_l4
)
1655 DEFINE_FREE_PT_FN(l6
, free_pt_l5
)
1657 static void free_pagetable(struct protection_domain
*domain
)
1659 unsigned long root
= (unsigned long)domain
->pt_root
;
1661 switch (domain
->mode
) {
1662 case PAGE_MODE_NONE
:
1664 case PAGE_MODE_1_LEVEL
:
1667 case PAGE_MODE_2_LEVEL
:
1670 case PAGE_MODE_3_LEVEL
:
1673 case PAGE_MODE_4_LEVEL
:
1676 case PAGE_MODE_5_LEVEL
:
1679 case PAGE_MODE_6_LEVEL
:
1687 static void free_gcr3_tbl_level1(u64
*tbl
)
1692 for (i
= 0; i
< 512; ++i
) {
1693 if (!(tbl
[i
] & GCR3_VALID
))
1696 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1698 free_page((unsigned long)ptr
);
1702 static void free_gcr3_tbl_level2(u64
*tbl
)
1707 for (i
= 0; i
< 512; ++i
) {
1708 if (!(tbl
[i
] & GCR3_VALID
))
1711 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1713 free_gcr3_tbl_level1(ptr
);
1717 static void free_gcr3_table(struct protection_domain
*domain
)
1719 if (domain
->glx
== 2)
1720 free_gcr3_tbl_level2(domain
->gcr3_tbl
);
1721 else if (domain
->glx
== 1)
1722 free_gcr3_tbl_level1(domain
->gcr3_tbl
);
1724 BUG_ON(domain
->glx
!= 0);
1726 free_page((unsigned long)domain
->gcr3_tbl
);
1730 * Free a domain, only used if something went wrong in the
1731 * allocation path and we need to free an already allocated page table
1733 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1738 del_domain_from_list(&dom
->domain
);
1740 put_iova_domain(&dom
->iovad
);
1742 free_pagetable(&dom
->domain
);
1748 * Allocates a new protection domain usable for the dma_ops functions.
1749 * It also initializes the page table and the address allocator data
1750 * structures required for the dma_ops interface
1752 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1754 struct dma_ops_domain
*dma_dom
;
1756 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1760 if (protection_domain_init(&dma_dom
->domain
))
1763 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
1764 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1765 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1766 if (!dma_dom
->domain
.pt_root
)
1769 init_iova_domain(&dma_dom
->iovad
, PAGE_SIZE
,
1770 IOVA_START_PFN
, DMA_32BIT_PFN
);
1772 /* Initialize reserved ranges */
1773 copy_reserved_iova(&reserved_iova_ranges
, &dma_dom
->iovad
);
1775 add_domain_to_list(&dma_dom
->domain
);
1780 dma_ops_domain_free(dma_dom
);
1786 * little helper function to check whether a given protection domain is a
1789 static bool dma_ops_domain(struct protection_domain
*domain
)
1791 return domain
->flags
& PD_DMA_OPS_MASK
;
1794 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1799 if (domain
->mode
!= PAGE_MODE_NONE
)
1800 pte_root
= virt_to_phys(domain
->pt_root
);
1802 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1803 << DEV_ENTRY_MODE_SHIFT
;
1804 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1806 flags
= amd_iommu_dev_table
[devid
].data
[1];
1809 flags
|= DTE_FLAG_IOTLB
;
1811 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1812 u64 gcr3
= __pa(domain
->gcr3_tbl
);
1813 u64 glx
= domain
->glx
;
1816 pte_root
|= DTE_FLAG_GV
;
1817 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1819 /* First mask out possible old values for GCR3 table */
1820 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1823 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1826 /* Encode GCR3 table into DTE */
1827 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1830 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1833 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1837 flags
&= ~(0xffffUL
);
1838 flags
|= domain
->id
;
1840 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1841 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1844 static void clear_dte_entry(u16 devid
)
1846 /* remove entry from the device table seen by the hardware */
1847 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1848 amd_iommu_dev_table
[devid
].data
[1] &= DTE_FLAG_MASK
;
1850 amd_iommu_apply_erratum_63(devid
);
1853 static void do_attach(struct iommu_dev_data
*dev_data
,
1854 struct protection_domain
*domain
)
1856 struct amd_iommu
*iommu
;
1860 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1861 alias
= dev_data
->alias
;
1862 ats
= dev_data
->ats
.enabled
;
1864 /* Update data structures */
1865 dev_data
->domain
= domain
;
1866 list_add(&dev_data
->list
, &domain
->dev_list
);
1868 /* Do reference counting */
1869 domain
->dev_iommu
[iommu
->index
] += 1;
1870 domain
->dev_cnt
+= 1;
1872 /* Update device table */
1873 set_dte_entry(dev_data
->devid
, domain
, ats
);
1874 if (alias
!= dev_data
->devid
)
1875 set_dte_entry(alias
, domain
, ats
);
1877 device_flush_dte(dev_data
);
1880 static void do_detach(struct iommu_dev_data
*dev_data
)
1882 struct amd_iommu
*iommu
;
1886 * First check if the device is still attached. It might already
1887 * be detached from its domain because the generic
1888 * iommu_detach_group code detached it and we try again here in
1889 * our alias handling.
1891 if (!dev_data
->domain
)
1894 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1895 alias
= dev_data
->alias
;
1897 /* decrease reference counters */
1898 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1899 dev_data
->domain
->dev_cnt
-= 1;
1901 /* Update data structures */
1902 dev_data
->domain
= NULL
;
1903 list_del(&dev_data
->list
);
1904 clear_dte_entry(dev_data
->devid
);
1905 if (alias
!= dev_data
->devid
)
1906 clear_dte_entry(alias
);
1908 /* Flush the DTE entry */
1909 device_flush_dte(dev_data
);
1913 * If a device is not yet associated with a domain, this function does
1914 * assigns it visible for the hardware
1916 static int __attach_device(struct iommu_dev_data
*dev_data
,
1917 struct protection_domain
*domain
)
1922 * Must be called with IRQs disabled. Warn here to detect early
1925 WARN_ON(!irqs_disabled());
1928 spin_lock(&domain
->lock
);
1931 if (dev_data
->domain
!= NULL
)
1934 /* Attach alias group root */
1935 do_attach(dev_data
, domain
);
1942 spin_unlock(&domain
->lock
);
1948 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
1950 pci_disable_ats(pdev
);
1951 pci_disable_pri(pdev
);
1952 pci_disable_pasid(pdev
);
1955 /* FIXME: Change generic reset-function to do the same */
1956 static int pri_reset_while_enabled(struct pci_dev
*pdev
)
1961 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
1965 pci_read_config_word(pdev
, pos
+ PCI_PRI_CTRL
, &control
);
1966 control
|= PCI_PRI_CTRL_RESET
;
1967 pci_write_config_word(pdev
, pos
+ PCI_PRI_CTRL
, control
);
1972 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
1977 /* FIXME: Hardcode number of outstanding requests for now */
1979 if (pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE
))
1981 reset_enable
= pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_ENABLE_RESET
);
1983 /* Only allow access to user-accessible pages */
1984 ret
= pci_enable_pasid(pdev
, 0);
1988 /* First reset the PRI state of the device */
1989 ret
= pci_reset_pri(pdev
);
1994 ret
= pci_enable_pri(pdev
, reqs
);
1999 ret
= pri_reset_while_enabled(pdev
);
2004 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
2011 pci_disable_pri(pdev
);
2012 pci_disable_pasid(pdev
);
2017 /* FIXME: Move this to PCI code */
2018 #define PCI_PRI_TLP_OFF (1 << 15)
2020 static bool pci_pri_tlp_required(struct pci_dev
*pdev
)
2025 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2029 pci_read_config_word(pdev
, pos
+ PCI_PRI_STATUS
, &status
);
2031 return (status
& PCI_PRI_TLP_OFF
) ? true : false;
2035 * If a device is not yet associated with a domain, this function
2036 * assigns it visible for the hardware
2038 static int attach_device(struct device
*dev
,
2039 struct protection_domain
*domain
)
2041 struct pci_dev
*pdev
;
2042 struct iommu_dev_data
*dev_data
;
2043 unsigned long flags
;
2046 dev_data
= get_dev_data(dev
);
2048 if (!dev_is_pci(dev
))
2049 goto skip_ats_check
;
2051 pdev
= to_pci_dev(dev
);
2052 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2053 if (!dev_data
->passthrough
)
2056 if (dev_data
->iommu_v2
) {
2057 if (pdev_iommuv2_enable(pdev
) != 0)
2060 dev_data
->ats
.enabled
= true;
2061 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2062 dev_data
->pri_tlp
= pci_pri_tlp_required(pdev
);
2064 } else if (amd_iommu_iotlb_sup
&&
2065 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
2066 dev_data
->ats
.enabled
= true;
2067 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2071 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2072 ret
= __attach_device(dev_data
, domain
);
2073 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2076 * We might boot into a crash-kernel here. The crashed kernel
2077 * left the caches in the IOMMU dirty. So we have to flush
2078 * here to evict all dirty stuff.
2080 domain_flush_tlb_pde(domain
);
2086 * Removes a device from a protection domain (unlocked)
2088 static void __detach_device(struct iommu_dev_data
*dev_data
)
2090 struct protection_domain
*domain
;
2093 * Must be called with IRQs disabled. Warn here to detect early
2096 WARN_ON(!irqs_disabled());
2098 if (WARN_ON(!dev_data
->domain
))
2101 domain
= dev_data
->domain
;
2103 spin_lock(&domain
->lock
);
2105 do_detach(dev_data
);
2107 spin_unlock(&domain
->lock
);
2111 * Removes a device from a protection domain (with devtable_lock held)
2113 static void detach_device(struct device
*dev
)
2115 struct protection_domain
*domain
;
2116 struct iommu_dev_data
*dev_data
;
2117 unsigned long flags
;
2119 dev_data
= get_dev_data(dev
);
2120 domain
= dev_data
->domain
;
2122 /* lock device table */
2123 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2124 __detach_device(dev_data
);
2125 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2127 if (!dev_is_pci(dev
))
2130 if (domain
->flags
& PD_IOMMUV2_MASK
&& dev_data
->iommu_v2
)
2131 pdev_iommuv2_disable(to_pci_dev(dev
));
2132 else if (dev_data
->ats
.enabled
)
2133 pci_disable_ats(to_pci_dev(dev
));
2135 dev_data
->ats
.enabled
= false;
2138 static int amd_iommu_add_device(struct device
*dev
)
2140 struct iommu_dev_data
*dev_data
;
2141 struct iommu_domain
*domain
;
2142 struct amd_iommu
*iommu
;
2145 if (!check_device(dev
) || get_dev_data(dev
))
2148 devid
= get_device_id(dev
);
2152 iommu
= amd_iommu_rlookup_table
[devid
];
2154 ret
= iommu_init_device(dev
);
2156 if (ret
!= -ENOTSUPP
)
2157 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2160 iommu_ignore_device(dev
);
2161 dev
->archdata
.dma_ops
= &nommu_dma_ops
;
2164 init_iommu_group(dev
);
2166 dev_data
= get_dev_data(dev
);
2170 if (iommu_pass_through
|| dev_data
->iommu_v2
)
2171 iommu_request_dm_for_dev(dev
);
2173 /* Domains are initialized for this device - have a look what we ended up with */
2174 domain
= iommu_get_domain_for_dev(dev
);
2175 if (domain
->type
== IOMMU_DOMAIN_IDENTITY
)
2176 dev_data
->passthrough
= true;
2178 dev
->archdata
.dma_ops
= &amd_iommu_dma_ops
;
2181 iommu_completion_wait(iommu
);
2186 static void amd_iommu_remove_device(struct device
*dev
)
2188 struct amd_iommu
*iommu
;
2191 if (!check_device(dev
))
2194 devid
= get_device_id(dev
);
2198 iommu
= amd_iommu_rlookup_table
[devid
];
2200 iommu_uninit_device(dev
);
2201 iommu_completion_wait(iommu
);
2204 static struct iommu_group
*amd_iommu_device_group(struct device
*dev
)
2206 if (dev_is_pci(dev
))
2207 return pci_device_group(dev
);
2209 return acpihid_device_group(dev
);
2212 /*****************************************************************************
2214 * The next functions belong to the dma_ops mapping/unmapping code.
2216 *****************************************************************************/
2218 static void __queue_flush(struct flush_queue
*queue
)
2220 struct protection_domain
*domain
;
2221 unsigned long flags
;
2224 /* First flush TLB of all known domains */
2225 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
2226 list_for_each_entry(domain
, &amd_iommu_pd_list
, list
)
2227 domain_flush_tlb(domain
);
2228 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
2230 /* Wait until flushes have completed */
2231 domain_flush_complete(NULL
);
2233 for (idx
= 0; idx
< queue
->next
; ++idx
) {
2234 struct flush_queue_entry
*entry
;
2236 entry
= queue
->entries
+ idx
;
2238 free_iova_fast(&entry
->dma_dom
->iovad
,
2242 /* Not really necessary, just to make sure we catch any bugs */
2243 entry
->dma_dom
= NULL
;
2249 static void queue_flush_all(void)
2253 for_each_possible_cpu(cpu
) {
2254 struct flush_queue
*queue
;
2255 unsigned long flags
;
2257 queue
= per_cpu_ptr(&flush_queue
, cpu
);
2258 spin_lock_irqsave(&queue
->lock
, flags
);
2259 if (queue
->next
> 0)
2260 __queue_flush(queue
);
2261 spin_unlock_irqrestore(&queue
->lock
, flags
);
2265 static void queue_flush_timeout(unsigned long unsused
)
2267 atomic_set(&queue_timer_on
, 0);
2271 static void queue_add(struct dma_ops_domain
*dma_dom
,
2272 unsigned long address
, unsigned long pages
)
2274 struct flush_queue_entry
*entry
;
2275 struct flush_queue
*queue
;
2276 unsigned long flags
;
2279 pages
= __roundup_pow_of_two(pages
);
2280 address
>>= PAGE_SHIFT
;
2282 queue
= get_cpu_ptr(&flush_queue
);
2283 spin_lock_irqsave(&queue
->lock
, flags
);
2285 if (queue
->next
== FLUSH_QUEUE_SIZE
)
2286 __queue_flush(queue
);
2288 idx
= queue
->next
++;
2289 entry
= queue
->entries
+ idx
;
2291 entry
->iova_pfn
= address
;
2292 entry
->pages
= pages
;
2293 entry
->dma_dom
= dma_dom
;
2295 spin_unlock_irqrestore(&queue
->lock
, flags
);
2297 if (atomic_cmpxchg(&queue_timer_on
, 0, 1) == 0)
2298 mod_timer(&queue_timer
, jiffies
+ msecs_to_jiffies(10));
2300 put_cpu_ptr(&flush_queue
);
2305 * In the dma_ops path we only have the struct device. This function
2306 * finds the corresponding IOMMU, the protection domain and the
2307 * requestor id for a given device.
2308 * If the device is not yet associated with a domain this is also done
2311 static struct protection_domain
*get_domain(struct device
*dev
)
2313 struct protection_domain
*domain
;
2315 if (!check_device(dev
))
2316 return ERR_PTR(-EINVAL
);
2318 domain
= get_dev_data(dev
)->domain
;
2319 if (!dma_ops_domain(domain
))
2320 return ERR_PTR(-EBUSY
);
2325 static void update_device_table(struct protection_domain
*domain
)
2327 struct iommu_dev_data
*dev_data
;
2329 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
2330 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2332 if (dev_data
->devid
== dev_data
->alias
)
2335 /* There is an alias, update device table entry for it */
2336 set_dte_entry(dev_data
->alias
, domain
, dev_data
->ats
.enabled
);
2340 static void update_domain(struct protection_domain
*domain
)
2342 if (!domain
->updated
)
2345 update_device_table(domain
);
2347 domain_flush_devices(domain
);
2348 domain_flush_tlb_pde(domain
);
2350 domain
->updated
= false;
2353 static int dir2prot(enum dma_data_direction direction
)
2355 if (direction
== DMA_TO_DEVICE
)
2356 return IOMMU_PROT_IR
;
2357 else if (direction
== DMA_FROM_DEVICE
)
2358 return IOMMU_PROT_IW
;
2359 else if (direction
== DMA_BIDIRECTIONAL
)
2360 return IOMMU_PROT_IW
| IOMMU_PROT_IR
;
2365 * This function contains common code for mapping of a physically
2366 * contiguous memory region into DMA address space. It is used by all
2367 * mapping functions provided with this IOMMU driver.
2368 * Must be called with the domain lock held.
2370 static dma_addr_t
__map_single(struct device
*dev
,
2371 struct dma_ops_domain
*dma_dom
,
2374 enum dma_data_direction direction
,
2377 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2378 dma_addr_t address
, start
, ret
;
2383 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2386 address
= dma_ops_alloc_iova(dev
, dma_dom
, pages
, dma_mask
);
2387 if (address
== DMA_ERROR_CODE
)
2390 prot
= dir2prot(direction
);
2393 for (i
= 0; i
< pages
; ++i
) {
2394 ret
= iommu_map_page(&dma_dom
->domain
, start
, paddr
,
2395 PAGE_SIZE
, prot
, GFP_ATOMIC
);
2404 if (unlikely(amd_iommu_np_cache
)) {
2405 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2406 domain_flush_complete(&dma_dom
->domain
);
2414 for (--i
; i
>= 0; --i
) {
2416 iommu_unmap_page(&dma_dom
->domain
, start
, PAGE_SIZE
);
2419 domain_flush_tlb(&dma_dom
->domain
);
2420 domain_flush_complete(&dma_dom
->domain
);
2422 dma_ops_free_iova(dma_dom
, address
, pages
);
2424 return DMA_ERROR_CODE
;
2428 * Does the reverse of the __map_single function. Must be called with
2429 * the domain lock held too
2431 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2432 dma_addr_t dma_addr
,
2436 dma_addr_t flush_addr
;
2437 dma_addr_t i
, start
;
2440 flush_addr
= dma_addr
;
2441 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2442 dma_addr
&= PAGE_MASK
;
2445 for (i
= 0; i
< pages
; ++i
) {
2446 iommu_unmap_page(&dma_dom
->domain
, start
, PAGE_SIZE
);
2450 if (amd_iommu_unmap_flush
) {
2451 dma_ops_free_iova(dma_dom
, dma_addr
, pages
);
2452 domain_flush_tlb(&dma_dom
->domain
);
2453 domain_flush_complete(&dma_dom
->domain
);
2455 queue_add(dma_dom
, dma_addr
, pages
);
2460 * The exported map_single function for dma_ops.
2462 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2463 unsigned long offset
, size_t size
,
2464 enum dma_data_direction dir
,
2465 unsigned long attrs
)
2467 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2468 struct protection_domain
*domain
;
2469 struct dma_ops_domain
*dma_dom
;
2472 domain
= get_domain(dev
);
2473 if (PTR_ERR(domain
) == -EINVAL
)
2474 return (dma_addr_t
)paddr
;
2475 else if (IS_ERR(domain
))
2476 return DMA_ERROR_CODE
;
2478 dma_mask
= *dev
->dma_mask
;
2479 dma_dom
= to_dma_ops_domain(domain
);
2481 return __map_single(dev
, dma_dom
, paddr
, size
, dir
, dma_mask
);
2485 * The exported unmap_single function for dma_ops.
2487 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2488 enum dma_data_direction dir
, unsigned long attrs
)
2490 struct protection_domain
*domain
;
2491 struct dma_ops_domain
*dma_dom
;
2493 domain
= get_domain(dev
);
2497 dma_dom
= to_dma_ops_domain(domain
);
2499 __unmap_single(dma_dom
, dma_addr
, size
, dir
);
2502 static int sg_num_pages(struct device
*dev
,
2503 struct scatterlist
*sglist
,
2506 unsigned long mask
, boundary_size
;
2507 struct scatterlist
*s
;
2510 mask
= dma_get_seg_boundary(dev
);
2511 boundary_size
= mask
+ 1 ? ALIGN(mask
+ 1, PAGE_SIZE
) >> PAGE_SHIFT
:
2512 1UL << (BITS_PER_LONG
- PAGE_SHIFT
);
2514 for_each_sg(sglist
, s
, nelems
, i
) {
2517 s
->dma_address
= npages
<< PAGE_SHIFT
;
2518 p
= npages
% boundary_size
;
2519 n
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2520 if (p
+ n
> boundary_size
)
2521 npages
+= boundary_size
- p
;
2529 * The exported map_sg function for dma_ops (handles scatter-gather
2532 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2533 int nelems
, enum dma_data_direction direction
,
2534 unsigned long attrs
)
2536 int mapped_pages
= 0, npages
= 0, prot
= 0, i
;
2537 struct protection_domain
*domain
;
2538 struct dma_ops_domain
*dma_dom
;
2539 struct scatterlist
*s
;
2540 unsigned long address
;
2543 domain
= get_domain(dev
);
2547 dma_dom
= to_dma_ops_domain(domain
);
2548 dma_mask
= *dev
->dma_mask
;
2550 npages
= sg_num_pages(dev
, sglist
, nelems
);
2552 address
= dma_ops_alloc_iova(dev
, dma_dom
, npages
, dma_mask
);
2553 if (address
== DMA_ERROR_CODE
)
2556 prot
= dir2prot(direction
);
2558 /* Map all sg entries */
2559 for_each_sg(sglist
, s
, nelems
, i
) {
2560 int j
, pages
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2562 for (j
= 0; j
< pages
; ++j
) {
2563 unsigned long bus_addr
, phys_addr
;
2566 bus_addr
= address
+ s
->dma_address
+ (j
<< PAGE_SHIFT
);
2567 phys_addr
= (sg_phys(s
) & PAGE_MASK
) + (j
<< PAGE_SHIFT
);
2568 ret
= iommu_map_page(domain
, bus_addr
, phys_addr
, PAGE_SIZE
, prot
, GFP_ATOMIC
);
2576 /* Everything is mapped - write the right values into s->dma_address */
2577 for_each_sg(sglist
, s
, nelems
, i
) {
2578 s
->dma_address
+= address
+ s
->offset
;
2579 s
->dma_length
= s
->length
;
2585 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2586 dev_name(dev
), npages
);
2588 for_each_sg(sglist
, s
, nelems
, i
) {
2589 int j
, pages
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2591 for (j
= 0; j
< pages
; ++j
) {
2592 unsigned long bus_addr
;
2594 bus_addr
= address
+ s
->dma_address
+ (j
<< PAGE_SHIFT
);
2595 iommu_unmap_page(domain
, bus_addr
, PAGE_SIZE
);
2603 free_iova_fast(&dma_dom
->iovad
, address
, npages
);
2610 * The exported map_sg function for dma_ops (handles scatter-gather
2613 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2614 int nelems
, enum dma_data_direction dir
,
2615 unsigned long attrs
)
2617 struct protection_domain
*domain
;
2618 struct dma_ops_domain
*dma_dom
;
2619 unsigned long startaddr
;
2622 domain
= get_domain(dev
);
2626 startaddr
= sg_dma_address(sglist
) & PAGE_MASK
;
2627 dma_dom
= to_dma_ops_domain(domain
);
2628 npages
= sg_num_pages(dev
, sglist
, nelems
);
2630 __unmap_single(dma_dom
, startaddr
, npages
<< PAGE_SHIFT
, dir
);
2634 * The exported alloc_coherent function for dma_ops.
2636 static void *alloc_coherent(struct device
*dev
, size_t size
,
2637 dma_addr_t
*dma_addr
, gfp_t flag
,
2638 unsigned long attrs
)
2640 u64 dma_mask
= dev
->coherent_dma_mask
;
2641 struct protection_domain
*domain
;
2642 struct dma_ops_domain
*dma_dom
;
2645 domain
= get_domain(dev
);
2646 if (PTR_ERR(domain
) == -EINVAL
) {
2647 page
= alloc_pages(flag
, get_order(size
));
2648 *dma_addr
= page_to_phys(page
);
2649 return page_address(page
);
2650 } else if (IS_ERR(domain
))
2653 dma_dom
= to_dma_ops_domain(domain
);
2654 size
= PAGE_ALIGN(size
);
2655 dma_mask
= dev
->coherent_dma_mask
;
2656 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2659 page
= alloc_pages(flag
| __GFP_NOWARN
, get_order(size
));
2661 if (!gfpflags_allow_blocking(flag
))
2664 page
= dma_alloc_from_contiguous(dev
, size
>> PAGE_SHIFT
,
2671 dma_mask
= *dev
->dma_mask
;
2673 *dma_addr
= __map_single(dev
, dma_dom
, page_to_phys(page
),
2674 size
, DMA_BIDIRECTIONAL
, dma_mask
);
2676 if (*dma_addr
== DMA_ERROR_CODE
)
2679 return page_address(page
);
2683 if (!dma_release_from_contiguous(dev
, page
, size
>> PAGE_SHIFT
))
2684 __free_pages(page
, get_order(size
));
2690 * The exported free_coherent function for dma_ops.
2692 static void free_coherent(struct device
*dev
, size_t size
,
2693 void *virt_addr
, dma_addr_t dma_addr
,
2694 unsigned long attrs
)
2696 struct protection_domain
*domain
;
2697 struct dma_ops_domain
*dma_dom
;
2700 page
= virt_to_page(virt_addr
);
2701 size
= PAGE_ALIGN(size
);
2703 domain
= get_domain(dev
);
2707 dma_dom
= to_dma_ops_domain(domain
);
2709 __unmap_single(dma_dom
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2712 if (!dma_release_from_contiguous(dev
, page
, size
>> PAGE_SHIFT
))
2713 __free_pages(page
, get_order(size
));
2717 * This function is called by the DMA layer to find out if we can handle a
2718 * particular device. It is part of the dma_ops.
2720 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2722 return check_device(dev
);
2725 static struct dma_map_ops amd_iommu_dma_ops
= {
2726 .alloc
= alloc_coherent
,
2727 .free
= free_coherent
,
2728 .map_page
= map_page
,
2729 .unmap_page
= unmap_page
,
2731 .unmap_sg
= unmap_sg
,
2732 .dma_supported
= amd_iommu_dma_supported
,
2735 static int init_reserved_iova_ranges(void)
2737 struct pci_dev
*pdev
= NULL
;
2740 init_iova_domain(&reserved_iova_ranges
, PAGE_SIZE
,
2741 IOVA_START_PFN
, DMA_32BIT_PFN
);
2743 lockdep_set_class(&reserved_iova_ranges
.iova_rbtree_lock
,
2744 &reserved_rbtree_key
);
2746 /* MSI memory range */
2747 val
= reserve_iova(&reserved_iova_ranges
,
2748 IOVA_PFN(MSI_RANGE_START
), IOVA_PFN(MSI_RANGE_END
));
2750 pr_err("Reserving MSI range failed\n");
2754 /* HT memory range */
2755 val
= reserve_iova(&reserved_iova_ranges
,
2756 IOVA_PFN(HT_RANGE_START
), IOVA_PFN(HT_RANGE_END
));
2758 pr_err("Reserving HT range failed\n");
2763 * Memory used for PCI resources
2764 * FIXME: Check whether we can reserve the PCI-hole completly
2766 for_each_pci_dev(pdev
) {
2769 for (i
= 0; i
< PCI_NUM_RESOURCES
; ++i
) {
2770 struct resource
*r
= &pdev
->resource
[i
];
2772 if (!(r
->flags
& IORESOURCE_MEM
))
2775 val
= reserve_iova(&reserved_iova_ranges
,
2779 pr_err("Reserve pci-resource range failed\n");
2788 int __init
amd_iommu_init_api(void)
2790 int ret
, cpu
, err
= 0;
2792 ret
= iova_cache_get();
2796 ret
= init_reserved_iova_ranges();
2800 for_each_possible_cpu(cpu
) {
2801 struct flush_queue
*queue
= per_cpu_ptr(&flush_queue
, cpu
);
2803 queue
->entries
= kzalloc(FLUSH_QUEUE_SIZE
*
2804 sizeof(*queue
->entries
),
2806 if (!queue
->entries
)
2809 spin_lock_init(&queue
->lock
);
2812 err
= bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2815 #ifdef CONFIG_ARM_AMBA
2816 err
= bus_set_iommu(&amba_bustype
, &amd_iommu_ops
);
2820 err
= bus_set_iommu(&platform_bus_type
, &amd_iommu_ops
);
2826 for_each_possible_cpu(cpu
) {
2827 struct flush_queue
*queue
= per_cpu_ptr(&flush_queue
, cpu
);
2829 kfree(queue
->entries
);
2835 int __init
amd_iommu_init_dma_ops(void)
2837 setup_timer(&queue_timer
, queue_flush_timeout
, 0);
2838 atomic_set(&queue_timer_on
, 0);
2840 swiotlb
= iommu_pass_through
? 1 : 0;
2844 * In case we don't initialize SWIOTLB (actually the common case
2845 * when AMD IOMMU is enabled), make sure there are global
2846 * dma_ops set as a fall-back for devices not handled by this
2847 * driver (for example non-PCI devices).
2850 dma_ops
= &nommu_dma_ops
;
2852 if (amd_iommu_unmap_flush
)
2853 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2855 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2861 /*****************************************************************************
2863 * The following functions belong to the exported interface of AMD IOMMU
2865 * This interface allows access to lower level functions of the IOMMU
2866 * like protection domain handling and assignement of devices to domains
2867 * which is not possible with the dma_ops interface.
2869 *****************************************************************************/
2871 static void cleanup_domain(struct protection_domain
*domain
)
2873 struct iommu_dev_data
*entry
;
2874 unsigned long flags
;
2876 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2878 while (!list_empty(&domain
->dev_list
)) {
2879 entry
= list_first_entry(&domain
->dev_list
,
2880 struct iommu_dev_data
, list
);
2881 __detach_device(entry
);
2884 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2887 static void protection_domain_free(struct protection_domain
*domain
)
2892 del_domain_from_list(domain
);
2895 domain_id_free(domain
->id
);
2900 static int protection_domain_init(struct protection_domain
*domain
)
2902 spin_lock_init(&domain
->lock
);
2903 mutex_init(&domain
->api_lock
);
2904 domain
->id
= domain_id_alloc();
2907 INIT_LIST_HEAD(&domain
->dev_list
);
2912 static struct protection_domain
*protection_domain_alloc(void)
2914 struct protection_domain
*domain
;
2916 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2920 if (protection_domain_init(domain
))
2923 add_domain_to_list(domain
);
2933 static struct iommu_domain
*amd_iommu_domain_alloc(unsigned type
)
2935 struct protection_domain
*pdomain
;
2936 struct dma_ops_domain
*dma_domain
;
2939 case IOMMU_DOMAIN_UNMANAGED
:
2940 pdomain
= protection_domain_alloc();
2944 pdomain
->mode
= PAGE_MODE_3_LEVEL
;
2945 pdomain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2946 if (!pdomain
->pt_root
) {
2947 protection_domain_free(pdomain
);
2951 pdomain
->domain
.geometry
.aperture_start
= 0;
2952 pdomain
->domain
.geometry
.aperture_end
= ~0ULL;
2953 pdomain
->domain
.geometry
.force_aperture
= true;
2956 case IOMMU_DOMAIN_DMA
:
2957 dma_domain
= dma_ops_domain_alloc();
2959 pr_err("AMD-Vi: Failed to allocate\n");
2962 pdomain
= &dma_domain
->domain
;
2964 case IOMMU_DOMAIN_IDENTITY
:
2965 pdomain
= protection_domain_alloc();
2969 pdomain
->mode
= PAGE_MODE_NONE
;
2975 return &pdomain
->domain
;
2978 static void amd_iommu_domain_free(struct iommu_domain
*dom
)
2980 struct protection_domain
*domain
;
2981 struct dma_ops_domain
*dma_dom
;
2983 domain
= to_pdomain(dom
);
2985 if (domain
->dev_cnt
> 0)
2986 cleanup_domain(domain
);
2988 BUG_ON(domain
->dev_cnt
!= 0);
2993 switch (dom
->type
) {
2994 case IOMMU_DOMAIN_DMA
:
2996 * First make sure the domain is no longer referenced from the
3001 /* Now release the domain */
3002 dma_dom
= to_dma_ops_domain(domain
);
3003 dma_ops_domain_free(dma_dom
);
3006 if (domain
->mode
!= PAGE_MODE_NONE
)
3007 free_pagetable(domain
);
3009 if (domain
->flags
& PD_IOMMUV2_MASK
)
3010 free_gcr3_table(domain
);
3012 protection_domain_free(domain
);
3017 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
3020 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
3021 struct amd_iommu
*iommu
;
3024 if (!check_device(dev
))
3027 devid
= get_device_id(dev
);
3031 if (dev_data
->domain
!= NULL
)
3034 iommu
= amd_iommu_rlookup_table
[devid
];
3038 #ifdef CONFIG_IRQ_REMAP
3039 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
) &&
3040 (dom
->type
== IOMMU_DOMAIN_UNMANAGED
))
3041 dev_data
->use_vapic
= 0;
3044 iommu_completion_wait(iommu
);
3047 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
3050 struct protection_domain
*domain
= to_pdomain(dom
);
3051 struct iommu_dev_data
*dev_data
;
3052 struct amd_iommu
*iommu
;
3055 if (!check_device(dev
))
3058 dev_data
= dev
->archdata
.iommu
;
3060 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3064 if (dev_data
->domain
)
3067 ret
= attach_device(dev
, domain
);
3069 #ifdef CONFIG_IRQ_REMAP
3070 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
)) {
3071 if (dom
->type
== IOMMU_DOMAIN_UNMANAGED
)
3072 dev_data
->use_vapic
= 1;
3074 dev_data
->use_vapic
= 0;
3078 iommu_completion_wait(iommu
);
3083 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
3084 phys_addr_t paddr
, size_t page_size
, int iommu_prot
)
3086 struct protection_domain
*domain
= to_pdomain(dom
);
3090 if (domain
->mode
== PAGE_MODE_NONE
)
3093 if (iommu_prot
& IOMMU_READ
)
3094 prot
|= IOMMU_PROT_IR
;
3095 if (iommu_prot
& IOMMU_WRITE
)
3096 prot
|= IOMMU_PROT_IW
;
3098 mutex_lock(&domain
->api_lock
);
3099 ret
= iommu_map_page(domain
, iova
, paddr
, page_size
, prot
, GFP_KERNEL
);
3100 mutex_unlock(&domain
->api_lock
);
3105 static size_t amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
3108 struct protection_domain
*domain
= to_pdomain(dom
);
3111 if (domain
->mode
== PAGE_MODE_NONE
)
3114 mutex_lock(&domain
->api_lock
);
3115 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3116 mutex_unlock(&domain
->api_lock
);
3118 domain_flush_tlb_pde(domain
);
3123 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3126 struct protection_domain
*domain
= to_pdomain(dom
);
3127 unsigned long offset_mask
, pte_pgsize
;
3130 if (domain
->mode
== PAGE_MODE_NONE
)
3133 pte
= fetch_pte(domain
, iova
, &pte_pgsize
);
3135 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3138 offset_mask
= pte_pgsize
- 1;
3139 __pte
= *pte
& PM_ADDR_MASK
;
3141 return (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3144 static bool amd_iommu_capable(enum iommu_cap cap
)
3147 case IOMMU_CAP_CACHE_COHERENCY
:
3149 case IOMMU_CAP_INTR_REMAP
:
3150 return (irq_remapping_enabled
== 1);
3151 case IOMMU_CAP_NOEXEC
:
3158 static void amd_iommu_get_dm_regions(struct device
*dev
,
3159 struct list_head
*head
)
3161 struct unity_map_entry
*entry
;
3164 devid
= get_device_id(dev
);
3168 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
3169 struct iommu_dm_region
*region
;
3171 if (devid
< entry
->devid_start
|| devid
> entry
->devid_end
)
3174 region
= kzalloc(sizeof(*region
), GFP_KERNEL
);
3176 pr_err("Out of memory allocating dm-regions for %s\n",
3181 region
->start
= entry
->address_start
;
3182 region
->length
= entry
->address_end
- entry
->address_start
;
3183 if (entry
->prot
& IOMMU_PROT_IR
)
3184 region
->prot
|= IOMMU_READ
;
3185 if (entry
->prot
& IOMMU_PROT_IW
)
3186 region
->prot
|= IOMMU_WRITE
;
3188 list_add_tail(®ion
->list
, head
);
3192 static void amd_iommu_put_dm_regions(struct device
*dev
,
3193 struct list_head
*head
)
3195 struct iommu_dm_region
*entry
, *next
;
3197 list_for_each_entry_safe(entry
, next
, head
, list
)
3201 static void amd_iommu_apply_dm_region(struct device
*dev
,
3202 struct iommu_domain
*domain
,
3203 struct iommu_dm_region
*region
)
3205 struct dma_ops_domain
*dma_dom
= to_dma_ops_domain(to_pdomain(domain
));
3206 unsigned long start
, end
;
3208 start
= IOVA_PFN(region
->start
);
3209 end
= IOVA_PFN(region
->start
+ region
->length
);
3211 WARN_ON_ONCE(reserve_iova(&dma_dom
->iovad
, start
, end
) == NULL
);
3214 static const struct iommu_ops amd_iommu_ops
= {
3215 .capable
= amd_iommu_capable
,
3216 .domain_alloc
= amd_iommu_domain_alloc
,
3217 .domain_free
= amd_iommu_domain_free
,
3218 .attach_dev
= amd_iommu_attach_device
,
3219 .detach_dev
= amd_iommu_detach_device
,
3220 .map
= amd_iommu_map
,
3221 .unmap
= amd_iommu_unmap
,
3222 .map_sg
= default_iommu_map_sg
,
3223 .iova_to_phys
= amd_iommu_iova_to_phys
,
3224 .add_device
= amd_iommu_add_device
,
3225 .remove_device
= amd_iommu_remove_device
,
3226 .device_group
= amd_iommu_device_group
,
3227 .get_dm_regions
= amd_iommu_get_dm_regions
,
3228 .put_dm_regions
= amd_iommu_put_dm_regions
,
3229 .apply_dm_region
= amd_iommu_apply_dm_region
,
3230 .pgsize_bitmap
= AMD_IOMMU_PGSIZES
,
3233 /*****************************************************************************
3235 * The next functions do a basic initialization of IOMMU for pass through
3238 * In passthrough mode the IOMMU is initialized and enabled but not used for
3239 * DMA-API translation.
3241 *****************************************************************************/
3243 /* IOMMUv2 specific functions */
3244 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3246 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3248 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3250 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3252 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3254 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3256 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3258 struct protection_domain
*domain
= to_pdomain(dom
);
3259 unsigned long flags
;
3261 spin_lock_irqsave(&domain
->lock
, flags
);
3263 /* Update data structure */
3264 domain
->mode
= PAGE_MODE_NONE
;
3265 domain
->updated
= true;
3267 /* Make changes visible to IOMMUs */
3268 update_domain(domain
);
3270 /* Page-table is not visible to IOMMU anymore, so free it */
3271 free_pagetable(domain
);
3273 spin_unlock_irqrestore(&domain
->lock
, flags
);
3275 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3277 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3279 struct protection_domain
*domain
= to_pdomain(dom
);
3280 unsigned long flags
;
3283 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3286 /* Number of GCR3 table levels required */
3287 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3290 if (levels
> amd_iommu_max_glx_val
)
3293 spin_lock_irqsave(&domain
->lock
, flags
);
3296 * Save us all sanity checks whether devices already in the
3297 * domain support IOMMUv2. Just force that the domain has no
3298 * devices attached when it is switched into IOMMUv2 mode.
3301 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3305 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3306 if (domain
->gcr3_tbl
== NULL
)
3309 domain
->glx
= levels
;
3310 domain
->flags
|= PD_IOMMUV2_MASK
;
3311 domain
->updated
= true;
3313 update_domain(domain
);
3318 spin_unlock_irqrestore(&domain
->lock
, flags
);
3322 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3324 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3325 u64 address
, bool size
)
3327 struct iommu_dev_data
*dev_data
;
3328 struct iommu_cmd cmd
;
3331 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3334 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3337 * IOMMU TLB needs to be flushed before Device TLB to
3338 * prevent device TLB refill from IOMMU TLB
3340 for (i
= 0; i
< amd_iommus_present
; ++i
) {
3341 if (domain
->dev_iommu
[i
] == 0)
3344 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3349 /* Wait until IOMMU TLB flushes are complete */
3350 domain_flush_complete(domain
);
3352 /* Now flush device TLBs */
3353 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3354 struct amd_iommu
*iommu
;
3358 There might be non-IOMMUv2 capable devices in an IOMMUv2
3361 if (!dev_data
->ats
.enabled
)
3364 qdep
= dev_data
->ats
.qdep
;
3365 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3367 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3368 qdep
, address
, size
);
3370 ret
= iommu_queue_command(iommu
, &cmd
);
3375 /* Wait until all device TLBs are flushed */
3376 domain_flush_complete(domain
);
3385 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3388 return __flush_pasid(domain
, pasid
, address
, false);
3391 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3394 struct protection_domain
*domain
= to_pdomain(dom
);
3395 unsigned long flags
;
3398 spin_lock_irqsave(&domain
->lock
, flags
);
3399 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3400 spin_unlock_irqrestore(&domain
->lock
, flags
);
3404 EXPORT_SYMBOL(amd_iommu_flush_page
);
3406 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3408 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3412 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3414 struct protection_domain
*domain
= to_pdomain(dom
);
3415 unsigned long flags
;
3418 spin_lock_irqsave(&domain
->lock
, flags
);
3419 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3420 spin_unlock_irqrestore(&domain
->lock
, flags
);
3424 EXPORT_SYMBOL(amd_iommu_flush_tlb
);
3426 static u64
*__get_gcr3_pte(u64
*root
, int level
, int pasid
, bool alloc
)
3433 index
= (pasid
>> (9 * level
)) & 0x1ff;
3439 if (!(*pte
& GCR3_VALID
)) {
3443 root
= (void *)get_zeroed_page(GFP_ATOMIC
);
3447 *pte
= __pa(root
) | GCR3_VALID
;
3450 root
= __va(*pte
& PAGE_MASK
);
3458 static int __set_gcr3(struct protection_domain
*domain
, int pasid
,
3463 if (domain
->mode
!= PAGE_MODE_NONE
)
3466 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, true);
3470 *pte
= (cr3
& PAGE_MASK
) | GCR3_VALID
;
3472 return __amd_iommu_flush_tlb(domain
, pasid
);
3475 static int __clear_gcr3(struct protection_domain
*domain
, int pasid
)
3479 if (domain
->mode
!= PAGE_MODE_NONE
)
3482 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, false);
3488 return __amd_iommu_flush_tlb(domain
, pasid
);
3491 int amd_iommu_domain_set_gcr3(struct iommu_domain
*dom
, int pasid
,
3494 struct protection_domain
*domain
= to_pdomain(dom
);
3495 unsigned long flags
;
3498 spin_lock_irqsave(&domain
->lock
, flags
);
3499 ret
= __set_gcr3(domain
, pasid
, cr3
);
3500 spin_unlock_irqrestore(&domain
->lock
, flags
);
3504 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3
);
3506 int amd_iommu_domain_clear_gcr3(struct iommu_domain
*dom
, int pasid
)
3508 struct protection_domain
*domain
= to_pdomain(dom
);
3509 unsigned long flags
;
3512 spin_lock_irqsave(&domain
->lock
, flags
);
3513 ret
= __clear_gcr3(domain
, pasid
);
3514 spin_unlock_irqrestore(&domain
->lock
, flags
);
3518 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3
);
3520 int amd_iommu_complete_ppr(struct pci_dev
*pdev
, int pasid
,
3521 int status
, int tag
)
3523 struct iommu_dev_data
*dev_data
;
3524 struct amd_iommu
*iommu
;
3525 struct iommu_cmd cmd
;
3527 dev_data
= get_dev_data(&pdev
->dev
);
3528 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3530 build_complete_ppr(&cmd
, dev_data
->devid
, pasid
, status
,
3531 tag
, dev_data
->pri_tlp
);
3533 return iommu_queue_command(iommu
, &cmd
);
3535 EXPORT_SYMBOL(amd_iommu_complete_ppr
);
3537 struct iommu_domain
*amd_iommu_get_v2_domain(struct pci_dev
*pdev
)
3539 struct protection_domain
*pdomain
;
3541 pdomain
= get_domain(&pdev
->dev
);
3542 if (IS_ERR(pdomain
))
3545 /* Only return IOMMUv2 domains */
3546 if (!(pdomain
->flags
& PD_IOMMUV2_MASK
))
3549 return &pdomain
->domain
;
3551 EXPORT_SYMBOL(amd_iommu_get_v2_domain
);
3553 void amd_iommu_enable_device_erratum(struct pci_dev
*pdev
, u32 erratum
)
3555 struct iommu_dev_data
*dev_data
;
3557 if (!amd_iommu_v2_supported())
3560 dev_data
= get_dev_data(&pdev
->dev
);
3561 dev_data
->errata
|= (1 << erratum
);
3563 EXPORT_SYMBOL(amd_iommu_enable_device_erratum
);
3565 int amd_iommu_device_info(struct pci_dev
*pdev
,
3566 struct amd_iommu_device_info
*info
)
3571 if (pdev
== NULL
|| info
== NULL
)
3574 if (!amd_iommu_v2_supported())
3577 memset(info
, 0, sizeof(*info
));
3579 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ATS
);
3581 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_ATS_SUP
;
3583 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
3585 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRI_SUP
;
3587 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PASID
);
3591 max_pasids
= 1 << (9 * (amd_iommu_max_glx_val
+ 1));
3592 max_pasids
= min(max_pasids
, (1 << 20));
3594 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PASID_SUP
;
3595 info
->max_pasids
= min(pci_max_pasids(pdev
), max_pasids
);
3597 features
= pci_pasid_features(pdev
);
3598 if (features
& PCI_PASID_CAP_EXEC
)
3599 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP
;
3600 if (features
& PCI_PASID_CAP_PRIV
)
3601 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP
;
3606 EXPORT_SYMBOL(amd_iommu_device_info
);
3608 #ifdef CONFIG_IRQ_REMAP
3610 /*****************************************************************************
3612 * Interrupt Remapping Implementation
3614 *****************************************************************************/
3616 static struct irq_chip amd_ir_chip
;
3618 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3619 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3620 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3621 #define DTE_IRQ_REMAP_ENABLE 1ULL
3623 static void set_dte_irq_entry(u16 devid
, struct irq_remap_table
*table
)
3627 dte
= amd_iommu_dev_table
[devid
].data
[2];
3628 dte
&= ~DTE_IRQ_PHYS_ADDR_MASK
;
3629 dte
|= virt_to_phys(table
->table
);
3630 dte
|= DTE_IRQ_REMAP_INTCTL
;
3631 dte
|= DTE_IRQ_TABLE_LEN
;
3632 dte
|= DTE_IRQ_REMAP_ENABLE
;
3634 amd_iommu_dev_table
[devid
].data
[2] = dte
;
3637 static struct irq_remap_table
*get_irq_table(u16 devid
, bool ioapic
)
3639 struct irq_remap_table
*table
= NULL
;
3640 struct amd_iommu
*iommu
;
3641 unsigned long flags
;
3644 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
3646 iommu
= amd_iommu_rlookup_table
[devid
];
3650 table
= irq_lookup_table
[devid
];
3654 alias
= amd_iommu_alias_table
[devid
];
3655 table
= irq_lookup_table
[alias
];
3657 irq_lookup_table
[devid
] = table
;
3658 set_dte_irq_entry(devid
, table
);
3659 iommu_flush_dte(iommu
, devid
);
3663 /* Nothing there yet, allocate new irq remapping table */
3664 table
= kzalloc(sizeof(*table
), GFP_ATOMIC
);
3668 /* Initialize table spin-lock */
3669 spin_lock_init(&table
->lock
);
3672 /* Keep the first 32 indexes free for IOAPIC interrupts */
3673 table
->min_index
= 32;
3675 table
->table
= kmem_cache_alloc(amd_iommu_irq_cache
, GFP_ATOMIC
);
3676 if (!table
->table
) {
3682 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir
))
3683 memset(table
->table
, 0,
3684 MAX_IRQS_PER_TABLE
* sizeof(u32
));
3686 memset(table
->table
, 0,
3687 (MAX_IRQS_PER_TABLE
* (sizeof(u64
) * 2)));
3692 for (i
= 0; i
< 32; ++i
)
3693 iommu
->irte_ops
->set_allocated(table
, i
);
3696 irq_lookup_table
[devid
] = table
;
3697 set_dte_irq_entry(devid
, table
);
3698 iommu_flush_dte(iommu
, devid
);
3699 if (devid
!= alias
) {
3700 irq_lookup_table
[alias
] = table
;
3701 set_dte_irq_entry(alias
, table
);
3702 iommu_flush_dte(iommu
, alias
);
3706 iommu_completion_wait(iommu
);
3709 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
3714 static int alloc_irq_index(u16 devid
, int count
)
3716 struct irq_remap_table
*table
;
3717 unsigned long flags
;
3719 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
3724 table
= get_irq_table(devid
, false);
3728 spin_lock_irqsave(&table
->lock
, flags
);
3730 /* Scan table for free entries */
3731 for (c
= 0, index
= table
->min_index
;
3732 index
< MAX_IRQS_PER_TABLE
;
3734 if (!iommu
->irte_ops
->is_allocated(table
, index
))
3741 iommu
->irte_ops
->set_allocated(table
, index
- c
+ 1);
3751 spin_unlock_irqrestore(&table
->lock
, flags
);
3756 static int modify_irte_ga(u16 devid
, int index
, struct irte_ga
*irte
,
3757 struct amd_ir_data
*data
)
3759 struct irq_remap_table
*table
;
3760 struct amd_iommu
*iommu
;
3761 unsigned long flags
;
3762 struct irte_ga
*entry
;
3764 iommu
= amd_iommu_rlookup_table
[devid
];
3768 table
= get_irq_table(devid
, false);
3772 spin_lock_irqsave(&table
->lock
, flags
);
3774 entry
= (struct irte_ga
*)table
->table
;
3775 entry
= &entry
[index
];
3776 entry
->lo
.fields_remap
.valid
= 0;
3777 entry
->hi
.val
= irte
->hi
.val
;
3778 entry
->lo
.val
= irte
->lo
.val
;
3779 entry
->lo
.fields_remap
.valid
= 1;
3783 spin_unlock_irqrestore(&table
->lock
, flags
);
3785 iommu_flush_irt(iommu
, devid
);
3786 iommu_completion_wait(iommu
);
3791 static int modify_irte(u16 devid
, int index
, union irte
*irte
)
3793 struct irq_remap_table
*table
;
3794 struct amd_iommu
*iommu
;
3795 unsigned long flags
;
3797 iommu
= amd_iommu_rlookup_table
[devid
];
3801 table
= get_irq_table(devid
, false);
3805 spin_lock_irqsave(&table
->lock
, flags
);
3806 table
->table
[index
] = irte
->val
;
3807 spin_unlock_irqrestore(&table
->lock
, flags
);
3809 iommu_flush_irt(iommu
, devid
);
3810 iommu_completion_wait(iommu
);
3815 static void free_irte(u16 devid
, int index
)
3817 struct irq_remap_table
*table
;
3818 struct amd_iommu
*iommu
;
3819 unsigned long flags
;
3821 iommu
= amd_iommu_rlookup_table
[devid
];
3825 table
= get_irq_table(devid
, false);
3829 spin_lock_irqsave(&table
->lock
, flags
);
3830 iommu
->irte_ops
->clear_allocated(table
, index
);
3831 spin_unlock_irqrestore(&table
->lock
, flags
);
3833 iommu_flush_irt(iommu
, devid
);
3834 iommu_completion_wait(iommu
);
3837 static void irte_prepare(void *entry
,
3838 u32 delivery_mode
, u32 dest_mode
,
3839 u8 vector
, u32 dest_apicid
, int devid
)
3841 union irte
*irte
= (union irte
*) entry
;
3844 irte
->fields
.vector
= vector
;
3845 irte
->fields
.int_type
= delivery_mode
;
3846 irte
->fields
.destination
= dest_apicid
;
3847 irte
->fields
.dm
= dest_mode
;
3848 irte
->fields
.valid
= 1;
3851 static void irte_ga_prepare(void *entry
,
3852 u32 delivery_mode
, u32 dest_mode
,
3853 u8 vector
, u32 dest_apicid
, int devid
)
3855 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3856 struct iommu_dev_data
*dev_data
= search_dev_data(devid
);
3860 irte
->lo
.fields_remap
.guest_mode
= dev_data
? dev_data
->use_vapic
: 0;
3861 irte
->lo
.fields_remap
.int_type
= delivery_mode
;
3862 irte
->lo
.fields_remap
.dm
= dest_mode
;
3863 irte
->hi
.fields
.vector
= vector
;
3864 irte
->lo
.fields_remap
.destination
= dest_apicid
;
3865 irte
->lo
.fields_remap
.valid
= 1;
3868 static void irte_activate(void *entry
, u16 devid
, u16 index
)
3870 union irte
*irte
= (union irte
*) entry
;
3872 irte
->fields
.valid
= 1;
3873 modify_irte(devid
, index
, irte
);
3876 static void irte_ga_activate(void *entry
, u16 devid
, u16 index
)
3878 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3880 irte
->lo
.fields_remap
.valid
= 1;
3881 modify_irte_ga(devid
, index
, irte
, NULL
);
3884 static void irte_deactivate(void *entry
, u16 devid
, u16 index
)
3886 union irte
*irte
= (union irte
*) entry
;
3888 irte
->fields
.valid
= 0;
3889 modify_irte(devid
, index
, irte
);
3892 static void irte_ga_deactivate(void *entry
, u16 devid
, u16 index
)
3894 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3896 irte
->lo
.fields_remap
.valid
= 0;
3897 modify_irte_ga(devid
, index
, irte
, NULL
);
3900 static void irte_set_affinity(void *entry
, u16 devid
, u16 index
,
3901 u8 vector
, u32 dest_apicid
)
3903 union irte
*irte
= (union irte
*) entry
;
3905 irte
->fields
.vector
= vector
;
3906 irte
->fields
.destination
= dest_apicid
;
3907 modify_irte(devid
, index
, irte
);
3910 static void irte_ga_set_affinity(void *entry
, u16 devid
, u16 index
,
3911 u8 vector
, u32 dest_apicid
)
3913 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3914 struct iommu_dev_data
*dev_data
= search_dev_data(devid
);
3916 if (!dev_data
|| !dev_data
->use_vapic
) {
3917 irte
->hi
.fields
.vector
= vector
;
3918 irte
->lo
.fields_remap
.destination
= dest_apicid
;
3919 irte
->lo
.fields_remap
.guest_mode
= 0;
3920 modify_irte_ga(devid
, index
, irte
, NULL
);
3924 #define IRTE_ALLOCATED (~1U)
3925 static void irte_set_allocated(struct irq_remap_table
*table
, int index
)
3927 table
->table
[index
] = IRTE_ALLOCATED
;
3930 static void irte_ga_set_allocated(struct irq_remap_table
*table
, int index
)
3932 struct irte_ga
*ptr
= (struct irte_ga
*)table
->table
;
3933 struct irte_ga
*irte
= &ptr
[index
];
3935 memset(&irte
->lo
.val
, 0, sizeof(u64
));
3936 memset(&irte
->hi
.val
, 0, sizeof(u64
));
3937 irte
->hi
.fields
.vector
= 0xff;
3940 static bool irte_is_allocated(struct irq_remap_table
*table
, int index
)
3942 union irte
*ptr
= (union irte
*)table
->table
;
3943 union irte
*irte
= &ptr
[index
];
3945 return irte
->val
!= 0;
3948 static bool irte_ga_is_allocated(struct irq_remap_table
*table
, int index
)
3950 struct irte_ga
*ptr
= (struct irte_ga
*)table
->table
;
3951 struct irte_ga
*irte
= &ptr
[index
];
3953 return irte
->hi
.fields
.vector
!= 0;
3956 static void irte_clear_allocated(struct irq_remap_table
*table
, int index
)
3958 table
->table
[index
] = 0;
3961 static void irte_ga_clear_allocated(struct irq_remap_table
*table
, int index
)
3963 struct irte_ga
*ptr
= (struct irte_ga
*)table
->table
;
3964 struct irte_ga
*irte
= &ptr
[index
];
3966 memset(&irte
->lo
.val
, 0, sizeof(u64
));
3967 memset(&irte
->hi
.val
, 0, sizeof(u64
));
3970 static int get_devid(struct irq_alloc_info
*info
)
3974 switch (info
->type
) {
3975 case X86_IRQ_ALLOC_TYPE_IOAPIC
:
3976 devid
= get_ioapic_devid(info
->ioapic_id
);
3978 case X86_IRQ_ALLOC_TYPE_HPET
:
3979 devid
= get_hpet_devid(info
->hpet_id
);
3981 case X86_IRQ_ALLOC_TYPE_MSI
:
3982 case X86_IRQ_ALLOC_TYPE_MSIX
:
3983 devid
= get_device_id(&info
->msi_dev
->dev
);
3993 static struct irq_domain
*get_ir_irq_domain(struct irq_alloc_info
*info
)
3995 struct amd_iommu
*iommu
;
4001 devid
= get_devid(info
);
4003 iommu
= amd_iommu_rlookup_table
[devid
];
4005 return iommu
->ir_domain
;
4011 static struct irq_domain
*get_irq_domain(struct irq_alloc_info
*info
)
4013 struct amd_iommu
*iommu
;
4019 switch (info
->type
) {
4020 case X86_IRQ_ALLOC_TYPE_MSI
:
4021 case X86_IRQ_ALLOC_TYPE_MSIX
:
4022 devid
= get_device_id(&info
->msi_dev
->dev
);
4026 iommu
= amd_iommu_rlookup_table
[devid
];
4028 return iommu
->msi_domain
;
4037 struct irq_remap_ops amd_iommu_irq_ops
= {
4038 .prepare
= amd_iommu_prepare
,
4039 .enable
= amd_iommu_enable
,
4040 .disable
= amd_iommu_disable
,
4041 .reenable
= amd_iommu_reenable
,
4042 .enable_faulting
= amd_iommu_enable_faulting
,
4043 .get_ir_irq_domain
= get_ir_irq_domain
,
4044 .get_irq_domain
= get_irq_domain
,
4047 static void irq_remapping_prepare_irte(struct amd_ir_data
*data
,
4048 struct irq_cfg
*irq_cfg
,
4049 struct irq_alloc_info
*info
,
4050 int devid
, int index
, int sub_handle
)
4052 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4053 struct msi_msg
*msg
= &data
->msi_entry
;
4054 struct IO_APIC_route_entry
*entry
;
4055 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
4060 data
->irq_2_irte
.devid
= devid
;
4061 data
->irq_2_irte
.index
= index
+ sub_handle
;
4062 iommu
->irte_ops
->prepare(data
->entry
, apic
->irq_delivery_mode
,
4063 apic
->irq_dest_mode
, irq_cfg
->vector
,
4064 irq_cfg
->dest_apicid
, devid
);
4066 switch (info
->type
) {
4067 case X86_IRQ_ALLOC_TYPE_IOAPIC
:
4068 /* Setup IOAPIC entry */
4069 entry
= info
->ioapic_entry
;
4070 info
->ioapic_entry
= NULL
;
4071 memset(entry
, 0, sizeof(*entry
));
4072 entry
->vector
= index
;
4074 entry
->trigger
= info
->ioapic_trigger
;
4075 entry
->polarity
= info
->ioapic_polarity
;
4076 /* Mask level triggered irqs. */
4077 if (info
->ioapic_trigger
)
4081 case X86_IRQ_ALLOC_TYPE_HPET
:
4082 case X86_IRQ_ALLOC_TYPE_MSI
:
4083 case X86_IRQ_ALLOC_TYPE_MSIX
:
4084 msg
->address_hi
= MSI_ADDR_BASE_HI
;
4085 msg
->address_lo
= MSI_ADDR_BASE_LO
;
4086 msg
->data
= irte_info
->index
;
4095 struct amd_irte_ops irte_32_ops
= {
4096 .prepare
= irte_prepare
,
4097 .activate
= irte_activate
,
4098 .deactivate
= irte_deactivate
,
4099 .set_affinity
= irte_set_affinity
,
4100 .set_allocated
= irte_set_allocated
,
4101 .is_allocated
= irte_is_allocated
,
4102 .clear_allocated
= irte_clear_allocated
,
4105 struct amd_irte_ops irte_128_ops
= {
4106 .prepare
= irte_ga_prepare
,
4107 .activate
= irte_ga_activate
,
4108 .deactivate
= irte_ga_deactivate
,
4109 .set_affinity
= irte_ga_set_affinity
,
4110 .set_allocated
= irte_ga_set_allocated
,
4111 .is_allocated
= irte_ga_is_allocated
,
4112 .clear_allocated
= irte_ga_clear_allocated
,
4115 static int irq_remapping_alloc(struct irq_domain
*domain
, unsigned int virq
,
4116 unsigned int nr_irqs
, void *arg
)
4118 struct irq_alloc_info
*info
= arg
;
4119 struct irq_data
*irq_data
;
4120 struct amd_ir_data
*data
= NULL
;
4121 struct irq_cfg
*cfg
;
4127 if (nr_irqs
> 1 && info
->type
!= X86_IRQ_ALLOC_TYPE_MSI
&&
4128 info
->type
!= X86_IRQ_ALLOC_TYPE_MSIX
)
4132 * With IRQ remapping enabled, don't need contiguous CPU vectors
4133 * to support multiple MSI interrupts.
4135 if (info
->type
== X86_IRQ_ALLOC_TYPE_MSI
)
4136 info
->flags
&= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS
;
4138 devid
= get_devid(info
);
4142 ret
= irq_domain_alloc_irqs_parent(domain
, virq
, nr_irqs
, arg
);
4146 if (info
->type
== X86_IRQ_ALLOC_TYPE_IOAPIC
) {
4147 if (get_irq_table(devid
, true))
4148 index
= info
->ioapic_pin
;
4152 index
= alloc_irq_index(devid
, nr_irqs
);
4155 pr_warn("Failed to allocate IRTE\n");
4156 goto out_free_parent
;
4159 for (i
= 0; i
< nr_irqs
; i
++) {
4160 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4161 cfg
= irqd_cfg(irq_data
);
4162 if (!irq_data
|| !cfg
) {
4168 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
4172 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir
))
4173 data
->entry
= kzalloc(sizeof(union irte
), GFP_KERNEL
);
4175 data
->entry
= kzalloc(sizeof(struct irte_ga
),
4182 irq_data
->hwirq
= (devid
<< 16) + i
;
4183 irq_data
->chip_data
= data
;
4184 irq_data
->chip
= &amd_ir_chip
;
4185 irq_remapping_prepare_irte(data
, cfg
, info
, devid
, index
, i
);
4186 irq_set_status_flags(virq
+ i
, IRQ_MOVE_PCNTXT
);
4192 for (i
--; i
>= 0; i
--) {
4193 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4195 kfree(irq_data
->chip_data
);
4197 for (i
= 0; i
< nr_irqs
; i
++)
4198 free_irte(devid
, index
+ i
);
4200 irq_domain_free_irqs_common(domain
, virq
, nr_irqs
);
4204 static void irq_remapping_free(struct irq_domain
*domain
, unsigned int virq
,
4205 unsigned int nr_irqs
)
4207 struct irq_2_irte
*irte_info
;
4208 struct irq_data
*irq_data
;
4209 struct amd_ir_data
*data
;
4212 for (i
= 0; i
< nr_irqs
; i
++) {
4213 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4214 if (irq_data
&& irq_data
->chip_data
) {
4215 data
= irq_data
->chip_data
;
4216 irte_info
= &data
->irq_2_irte
;
4217 free_irte(irte_info
->devid
, irte_info
->index
);
4222 irq_domain_free_irqs_common(domain
, virq
, nr_irqs
);
4225 static void irq_remapping_activate(struct irq_domain
*domain
,
4226 struct irq_data
*irq_data
)
4228 struct amd_ir_data
*data
= irq_data
->chip_data
;
4229 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4230 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4233 iommu
->irte_ops
->activate(data
->entry
, irte_info
->devid
,
4237 static void irq_remapping_deactivate(struct irq_domain
*domain
,
4238 struct irq_data
*irq_data
)
4240 struct amd_ir_data
*data
= irq_data
->chip_data
;
4241 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4242 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4245 iommu
->irte_ops
->deactivate(data
->entry
, irte_info
->devid
,
4249 static struct irq_domain_ops amd_ir_domain_ops
= {
4250 .alloc
= irq_remapping_alloc
,
4251 .free
= irq_remapping_free
,
4252 .activate
= irq_remapping_activate
,
4253 .deactivate
= irq_remapping_deactivate
,
4256 static int amd_ir_set_vcpu_affinity(struct irq_data
*data
, void *vcpu_info
)
4258 struct amd_iommu
*iommu
;
4259 struct amd_iommu_pi_data
*pi_data
= vcpu_info
;
4260 struct vcpu_data
*vcpu_pi_info
= pi_data
->vcpu_data
;
4261 struct amd_ir_data
*ir_data
= data
->chip_data
;
4262 struct irte_ga
*irte
= (struct irte_ga
*) ir_data
->entry
;
4263 struct irq_2_irte
*irte_info
= &ir_data
->irq_2_irte
;
4264 struct iommu_dev_data
*dev_data
= search_dev_data(irte_info
->devid
);
4267 * This device has never been set up for guest mode.
4268 * we should not modify the IRTE
4270 if (!dev_data
|| !dev_data
->use_vapic
)
4273 pi_data
->ir_data
= ir_data
;
4276 * SVM tries to set up for VAPIC mode, but we are in
4277 * legacy mode. So, we force legacy mode instead.
4279 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
)) {
4280 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4282 pi_data
->is_guest_mode
= false;
4285 iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4289 pi_data
->prev_ga_tag
= ir_data
->cached_ga_tag
;
4290 if (pi_data
->is_guest_mode
) {
4292 irte
->hi
.fields
.ga_root_ptr
= (pi_data
->base
>> 12);
4293 irte
->hi
.fields
.vector
= vcpu_pi_info
->vector
;
4294 irte
->lo
.fields_vapic
.guest_mode
= 1;
4295 irte
->lo
.fields_vapic
.ga_tag
= pi_data
->ga_tag
;
4297 ir_data
->cached_ga_tag
= pi_data
->ga_tag
;
4300 struct irq_cfg
*cfg
= irqd_cfg(data
);
4304 irte
->hi
.fields
.vector
= cfg
->vector
;
4305 irte
->lo
.fields_remap
.guest_mode
= 0;
4306 irte
->lo
.fields_remap
.destination
= cfg
->dest_apicid
;
4307 irte
->lo
.fields_remap
.int_type
= apic
->irq_delivery_mode
;
4308 irte
->lo
.fields_remap
.dm
= apic
->irq_dest_mode
;
4311 * This communicates the ga_tag back to the caller
4312 * so that it can do all the necessary clean up.
4314 ir_data
->cached_ga_tag
= 0;
4317 return modify_irte_ga(irte_info
->devid
, irte_info
->index
, irte
, ir_data
);
4320 static int amd_ir_set_affinity(struct irq_data
*data
,
4321 const struct cpumask
*mask
, bool force
)
4323 struct amd_ir_data
*ir_data
= data
->chip_data
;
4324 struct irq_2_irte
*irte_info
= &ir_data
->irq_2_irte
;
4325 struct irq_cfg
*cfg
= irqd_cfg(data
);
4326 struct irq_data
*parent
= data
->parent_data
;
4327 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4333 ret
= parent
->chip
->irq_set_affinity(parent
, mask
, force
);
4334 if (ret
< 0 || ret
== IRQ_SET_MASK_OK_DONE
)
4338 * Atomically updates the IRTE with the new destination, vector
4339 * and flushes the interrupt entry cache.
4341 iommu
->irte_ops
->set_affinity(ir_data
->entry
, irte_info
->devid
,
4342 irte_info
->index
, cfg
->vector
, cfg
->dest_apicid
);
4345 * After this point, all the interrupts will start arriving
4346 * at the new destination. So, time to cleanup the previous
4347 * vector allocation.
4349 send_cleanup_vector(cfg
);
4351 return IRQ_SET_MASK_OK_DONE
;
4354 static void ir_compose_msi_msg(struct irq_data
*irq_data
, struct msi_msg
*msg
)
4356 struct amd_ir_data
*ir_data
= irq_data
->chip_data
;
4358 *msg
= ir_data
->msi_entry
;
4361 static struct irq_chip amd_ir_chip
= {
4362 .irq_ack
= ir_ack_apic_edge
,
4363 .irq_set_affinity
= amd_ir_set_affinity
,
4364 .irq_set_vcpu_affinity
= amd_ir_set_vcpu_affinity
,
4365 .irq_compose_msi_msg
= ir_compose_msi_msg
,
4368 int amd_iommu_create_irq_domain(struct amd_iommu
*iommu
)
4370 iommu
->ir_domain
= irq_domain_add_tree(NULL
, &amd_ir_domain_ops
, iommu
);
4371 if (!iommu
->ir_domain
)
4374 iommu
->ir_domain
->parent
= arch_get_ir_parent_domain();
4375 iommu
->msi_domain
= arch_create_msi_irq_domain(iommu
->ir_domain
);
4380 int amd_iommu_update_ga(int cpu
, bool is_run
, void *data
)
4382 unsigned long flags
;
4383 struct amd_iommu
*iommu
;
4384 struct irq_remap_table
*irt
;
4385 struct amd_ir_data
*ir_data
= (struct amd_ir_data
*)data
;
4386 int devid
= ir_data
->irq_2_irte
.devid
;
4387 struct irte_ga
*entry
= (struct irte_ga
*) ir_data
->entry
;
4388 struct irte_ga
*ref
= (struct irte_ga
*) ir_data
->ref
;
4390 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
) ||
4391 !ref
|| !entry
|| !entry
->lo
.fields_vapic
.guest_mode
)
4394 iommu
= amd_iommu_rlookup_table
[devid
];
4398 irt
= get_irq_table(devid
, false);
4402 spin_lock_irqsave(&irt
->lock
, flags
);
4404 if (ref
->lo
.fields_vapic
.guest_mode
) {
4406 ref
->lo
.fields_vapic
.destination
= cpu
;
4407 ref
->lo
.fields_vapic
.is_run
= is_run
;
4411 spin_unlock_irqrestore(&irt
->lock
, flags
);
4413 iommu_flush_irt(iommu
, devid
);
4414 iommu_completion_wait(iommu
);
4417 EXPORT_SYMBOL(amd_iommu_update_ga
);