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 #define pr_fmt(fmt) "AMD-Vi: " fmt
21 #define dev_fmt(fmt) pr_fmt(fmt)
23 #include <linux/ratelimit.h>
24 #include <linux/pci.h>
25 #include <linux/acpi.h>
26 #include <linux/amba/bus.h>
27 #include <linux/platform_device.h>
28 #include <linux/pci-ats.h>
29 #include <linux/bitmap.h>
30 #include <linux/slab.h>
31 #include <linux/debugfs.h>
32 #include <linux/scatterlist.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/dma-direct.h>
35 #include <linux/iommu-helper.h>
36 #include <linux/iommu.h>
37 #include <linux/delay.h>
38 #include <linux/amd-iommu.h>
39 #include <linux/notifier.h>
40 #include <linux/export.h>
41 #include <linux/irq.h>
42 #include <linux/msi.h>
43 #include <linux/dma-contiguous.h>
44 #include <linux/irqdomain.h>
45 #include <linux/percpu.h>
46 #include <linux/iova.h>
47 #include <asm/irq_remapping.h>
48 #include <asm/io_apic.h>
50 #include <asm/hw_irq.h>
51 #include <asm/msidef.h>
52 #include <asm/proto.h>
53 #include <asm/iommu.h>
57 #include "amd_iommu_proto.h"
58 #include "amd_iommu_types.h"
59 #include "irq_remapping.h"
61 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
63 #define LOOP_TIMEOUT 100000
65 /* IO virtual address start page frame number */
66 #define IOVA_START_PFN (1)
67 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
69 /* Reserved IOVA ranges */
70 #define MSI_RANGE_START (0xfee00000)
71 #define MSI_RANGE_END (0xfeefffff)
72 #define HT_RANGE_START (0xfd00000000ULL)
73 #define HT_RANGE_END (0xffffffffffULL)
76 * This bitmap is used to advertise the page sizes our hardware support
77 * to the IOMMU core, which will then use this information to split
78 * physically contiguous memory regions it is mapping into page sizes
81 * 512GB Pages are not supported due to a hardware bug
83 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
85 static DEFINE_SPINLOCK(amd_iommu_devtable_lock
);
86 static DEFINE_SPINLOCK(pd_bitmap_lock
);
88 /* List of all available dev_data structures */
89 static LLIST_HEAD(dev_data_list
);
91 LIST_HEAD(ioapic_map
);
93 LIST_HEAD(acpihid_map
);
96 * Domain for untranslated devices - only allocated
97 * if iommu=pt passed on kernel cmd line.
99 const struct iommu_ops amd_iommu_ops
;
101 static ATOMIC_NOTIFIER_HEAD(ppr_notifier
);
102 int amd_iommu_max_glx_val
= -1;
104 static const struct dma_map_ops amd_iommu_dma_ops
;
107 * general struct to manage commands send to an IOMMU
113 struct kmem_cache
*amd_iommu_irq_cache
;
115 static void update_domain(struct protection_domain
*domain
);
116 static int protection_domain_init(struct protection_domain
*domain
);
117 static void detach_device(struct device
*dev
);
118 static void iova_domain_flush_tlb(struct iova_domain
*iovad
);
121 * Data container for a dma_ops specific protection domain
123 struct dma_ops_domain
{
124 /* generic protection domain information */
125 struct protection_domain domain
;
128 struct iova_domain iovad
;
131 static struct iova_domain reserved_iova_ranges
;
132 static struct lock_class_key reserved_rbtree_key
;
134 /****************************************************************************
138 ****************************************************************************/
140 static inline int match_hid_uid(struct device
*dev
,
141 struct acpihid_map_entry
*entry
)
143 struct acpi_device
*adev
= ACPI_COMPANION(dev
);
144 const char *hid
, *uid
;
149 hid
= acpi_device_hid(adev
);
150 uid
= acpi_device_uid(adev
);
156 return strcmp(hid
, entry
->hid
);
159 return strcmp(hid
, entry
->hid
);
161 return (strcmp(hid
, entry
->hid
) || strcmp(uid
, entry
->uid
));
164 static inline u16
get_pci_device_id(struct device
*dev
)
166 struct pci_dev
*pdev
= to_pci_dev(dev
);
168 return pci_dev_id(pdev
);
171 static inline int get_acpihid_device_id(struct device
*dev
,
172 struct acpihid_map_entry
**entry
)
174 struct acpihid_map_entry
*p
;
176 list_for_each_entry(p
, &acpihid_map
, list
) {
177 if (!match_hid_uid(dev
, p
)) {
186 static inline int get_device_id(struct device
*dev
)
191 devid
= get_pci_device_id(dev
);
193 devid
= get_acpihid_device_id(dev
, NULL
);
198 static struct protection_domain
*to_pdomain(struct iommu_domain
*dom
)
200 return container_of(dom
, struct protection_domain
, domain
);
203 static struct dma_ops_domain
* to_dma_ops_domain(struct protection_domain
*domain
)
205 BUG_ON(domain
->flags
!= PD_DMA_OPS_MASK
);
206 return container_of(domain
, struct dma_ops_domain
, domain
);
209 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
211 struct iommu_dev_data
*dev_data
;
213 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
217 dev_data
->devid
= devid
;
218 ratelimit_default_init(&dev_data
->rs
);
220 llist_add(&dev_data
->dev_data_list
, &dev_data_list
);
224 static struct iommu_dev_data
*search_dev_data(u16 devid
)
226 struct iommu_dev_data
*dev_data
;
227 struct llist_node
*node
;
229 if (llist_empty(&dev_data_list
))
232 node
= dev_data_list
.first
;
233 llist_for_each_entry(dev_data
, node
, dev_data_list
) {
234 if (dev_data
->devid
== devid
)
241 static int __last_alias(struct pci_dev
*pdev
, u16 alias
, void *data
)
243 *(u16
*)data
= alias
;
247 static u16
get_alias(struct device
*dev
)
249 struct pci_dev
*pdev
= to_pci_dev(dev
);
250 u16 devid
, ivrs_alias
, pci_alias
;
252 /* The callers make sure that get_device_id() does not fail here */
253 devid
= get_device_id(dev
);
255 /* For ACPI HID devices, we simply return the devid as such */
256 if (!dev_is_pci(dev
))
259 ivrs_alias
= amd_iommu_alias_table
[devid
];
261 pci_for_each_dma_alias(pdev
, __last_alias
, &pci_alias
);
263 if (ivrs_alias
== pci_alias
)
269 * The IVRS is fairly reliable in telling us about aliases, but it
270 * can't know about every screwy device. If we don't have an IVRS
271 * reported alias, use the PCI reported alias. In that case we may
272 * still need to initialize the rlookup and dev_table entries if the
273 * alias is to a non-existent device.
275 if (ivrs_alias
== devid
) {
276 if (!amd_iommu_rlookup_table
[pci_alias
]) {
277 amd_iommu_rlookup_table
[pci_alias
] =
278 amd_iommu_rlookup_table
[devid
];
279 memcpy(amd_iommu_dev_table
[pci_alias
].data
,
280 amd_iommu_dev_table
[devid
].data
,
281 sizeof(amd_iommu_dev_table
[pci_alias
].data
));
287 pci_info(pdev
, "Using IVRS reported alias %02x:%02x.%d "
288 "for device [%04x:%04x], kernel reported alias "
289 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias
), PCI_SLOT(ivrs_alias
),
290 PCI_FUNC(ivrs_alias
), pdev
->vendor
, pdev
->device
,
291 PCI_BUS_NUM(pci_alias
), PCI_SLOT(pci_alias
),
292 PCI_FUNC(pci_alias
));
295 * If we don't have a PCI DMA alias and the IVRS alias is on the same
296 * bus, then the IVRS table may know about a quirk that we don't.
298 if (pci_alias
== devid
&&
299 PCI_BUS_NUM(ivrs_alias
) == pdev
->bus
->number
) {
300 pci_add_dma_alias(pdev
, ivrs_alias
& 0xff);
301 pci_info(pdev
, "Added PCI DMA alias %02x.%d\n",
302 PCI_SLOT(ivrs_alias
), PCI_FUNC(ivrs_alias
));
308 static struct iommu_dev_data
*find_dev_data(u16 devid
)
310 struct iommu_dev_data
*dev_data
;
311 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
313 dev_data
= search_dev_data(devid
);
315 if (dev_data
== NULL
) {
316 dev_data
= alloc_dev_data(devid
);
320 if (translation_pre_enabled(iommu
))
321 dev_data
->defer_attach
= true;
327 struct iommu_dev_data
*get_dev_data(struct device
*dev
)
329 return dev
->archdata
.iommu
;
331 EXPORT_SYMBOL(get_dev_data
);
334 * Find or create an IOMMU group for a acpihid device.
336 static struct iommu_group
*acpihid_device_group(struct device
*dev
)
338 struct acpihid_map_entry
*p
, *entry
= NULL
;
341 devid
= get_acpihid_device_id(dev
, &entry
);
343 return ERR_PTR(devid
);
345 list_for_each_entry(p
, &acpihid_map
, list
) {
346 if ((devid
== p
->devid
) && p
->group
)
347 entry
->group
= p
->group
;
351 entry
->group
= generic_device_group(dev
);
353 iommu_group_ref_get(entry
->group
);
358 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
360 static const int caps
[] = {
363 PCI_EXT_CAP_ID_PASID
,
367 if (pci_ats_disabled())
370 for (i
= 0; i
< 3; ++i
) {
371 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
379 static bool pdev_pri_erratum(struct pci_dev
*pdev
, u32 erratum
)
381 struct iommu_dev_data
*dev_data
;
383 dev_data
= get_dev_data(&pdev
->dev
);
385 return dev_data
->errata
& (1 << erratum
) ? true : false;
389 * This function checks if the driver got a valid device from the caller to
390 * avoid dereferencing invalid pointers.
392 static bool check_device(struct device
*dev
)
396 if (!dev
|| !dev
->dma_mask
)
399 devid
= get_device_id(dev
);
403 /* Out of our scope? */
404 if (devid
> amd_iommu_last_bdf
)
407 if (amd_iommu_rlookup_table
[devid
] == NULL
)
413 static void init_iommu_group(struct device
*dev
)
415 struct iommu_group
*group
;
417 group
= iommu_group_get_for_dev(dev
);
421 iommu_group_put(group
);
424 static int iommu_init_device(struct device
*dev
)
426 struct iommu_dev_data
*dev_data
;
427 struct amd_iommu
*iommu
;
430 if (dev
->archdata
.iommu
)
433 devid
= get_device_id(dev
);
437 iommu
= amd_iommu_rlookup_table
[devid
];
439 dev_data
= find_dev_data(devid
);
443 dev_data
->alias
= get_alias(dev
);
446 * By default we use passthrough mode for IOMMUv2 capable device.
447 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
448 * invalid address), we ignore the capability for the device so
449 * it'll be forced to go into translation mode.
451 if ((iommu_pass_through
|| !amd_iommu_force_isolation
) &&
452 dev_is_pci(dev
) && pci_iommuv2_capable(to_pci_dev(dev
))) {
453 struct amd_iommu
*iommu
;
455 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
456 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
459 dev
->archdata
.iommu
= dev_data
;
461 iommu_device_link(&iommu
->iommu
, dev
);
466 static void iommu_ignore_device(struct device
*dev
)
471 devid
= get_device_id(dev
);
475 alias
= get_alias(dev
);
477 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
478 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
480 amd_iommu_rlookup_table
[devid
] = NULL
;
481 amd_iommu_rlookup_table
[alias
] = NULL
;
484 static void iommu_uninit_device(struct device
*dev
)
486 struct iommu_dev_data
*dev_data
;
487 struct amd_iommu
*iommu
;
490 devid
= get_device_id(dev
);
494 iommu
= amd_iommu_rlookup_table
[devid
];
496 dev_data
= search_dev_data(devid
);
500 if (dev_data
->domain
)
503 iommu_device_unlink(&iommu
->iommu
, dev
);
505 iommu_group_remove_device(dev
);
511 * We keep dev_data around for unplugged devices and reuse it when the
512 * device is re-plugged - not doing so would introduce a ton of races.
516 /****************************************************************************
518 * Interrupt handling functions
520 ****************************************************************************/
522 static void dump_dte_entry(u16 devid
)
526 for (i
= 0; i
< 4; ++i
)
527 pr_err("DTE[%d]: %016llx\n", i
,
528 amd_iommu_dev_table
[devid
].data
[i
]);
531 static void dump_command(unsigned long phys_addr
)
533 struct iommu_cmd
*cmd
= iommu_phys_to_virt(phys_addr
);
536 for (i
= 0; i
< 4; ++i
)
537 pr_err("CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
540 static void amd_iommu_report_page_fault(u16 devid
, u16 domain_id
,
541 u64 address
, int flags
)
543 struct iommu_dev_data
*dev_data
= NULL
;
544 struct pci_dev
*pdev
;
546 pdev
= pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid
),
549 dev_data
= get_dev_data(&pdev
->dev
);
551 if (dev_data
&& __ratelimit(&dev_data
->rs
)) {
552 pci_err(pdev
, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
553 domain_id
, address
, flags
);
554 } else if (printk_ratelimit()) {
555 pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
556 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
557 domain_id
, address
, flags
);
564 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
566 struct device
*dev
= iommu
->iommu
.dev
;
567 int type
, devid
, pasid
, flags
, tag
;
568 volatile u32
*event
= __evt
;
573 type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
574 devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
575 pasid
= PPR_PASID(*(u64
*)&event
[0]);
576 flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
577 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
580 /* Did we hit the erratum? */
581 if (++count
== LOOP_TIMEOUT
) {
582 pr_err("No event written to event log\n");
589 if (type
== EVENT_TYPE_IO_FAULT
) {
590 amd_iommu_report_page_fault(devid
, pasid
, address
, flags
);
595 case EVENT_TYPE_ILL_DEV
:
596 dev_err(dev
, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
598 pasid
, address
, flags
);
599 dump_dte_entry(devid
);
601 case EVENT_TYPE_DEV_TAB_ERR
:
602 dev_err(dev
, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
603 "address=0x%llx flags=0x%04x]\n",
604 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
607 case EVENT_TYPE_PAGE_TAB_ERR
:
608 dev_err(dev
, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
609 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
610 pasid
, address
, flags
);
612 case EVENT_TYPE_ILL_CMD
:
613 dev_err(dev
, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address
);
614 dump_command(address
);
616 case EVENT_TYPE_CMD_HARD_ERR
:
617 dev_err(dev
, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
620 case EVENT_TYPE_IOTLB_INV_TO
:
621 dev_err(dev
, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
622 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
625 case EVENT_TYPE_INV_DEV_REQ
:
626 dev_err(dev
, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
627 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
628 pasid
, address
, flags
);
630 case EVENT_TYPE_INV_PPR_REQ
:
631 pasid
= ((event
[0] >> 16) & 0xFFFF)
632 | ((event
[1] << 6) & 0xF0000);
633 tag
= event
[1] & 0x03FF;
634 dev_err(dev
, "Event logged [INVALID_PPR_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
635 PCI_BUS_NUM(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
636 pasid
, address
, flags
);
639 dev_err(dev
, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
640 event
[0], event
[1], event
[2], event
[3]);
643 memset(__evt
, 0, 4 * sizeof(u32
));
646 static void iommu_poll_events(struct amd_iommu
*iommu
)
650 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
651 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
653 while (head
!= tail
) {
654 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
655 head
= (head
+ EVENT_ENTRY_SIZE
) % EVT_BUFFER_SIZE
;
658 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
661 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u64
*raw
)
663 struct amd_iommu_fault fault
;
665 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
666 pr_err_ratelimited("Unknown PPR request received\n");
670 fault
.address
= raw
[1];
671 fault
.pasid
= PPR_PASID(raw
[0]);
672 fault
.device_id
= PPR_DEVID(raw
[0]);
673 fault
.tag
= PPR_TAG(raw
[0]);
674 fault
.flags
= PPR_FLAGS(raw
[0]);
676 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
679 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
683 if (iommu
->ppr_log
== NULL
)
686 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
687 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
689 while (head
!= tail
) {
694 raw
= (u64
*)(iommu
->ppr_log
+ head
);
697 * Hardware bug: Interrupt may arrive before the entry is
698 * written to memory. If this happens we need to wait for the
701 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
702 if (PPR_REQ_TYPE(raw
[0]) != 0)
707 /* Avoid memcpy function-call overhead */
712 * To detect the hardware bug we need to clear the entry
715 raw
[0] = raw
[1] = 0UL;
717 /* Update head pointer of hardware ring-buffer */
718 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
719 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
721 /* Handle PPR entry */
722 iommu_handle_ppr_entry(iommu
, entry
);
724 /* Refresh ring-buffer information */
725 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
726 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
730 #ifdef CONFIG_IRQ_REMAP
731 static int (*iommu_ga_log_notifier
)(u32
);
733 int amd_iommu_register_ga_log_notifier(int (*notifier
)(u32
))
735 iommu_ga_log_notifier
= notifier
;
739 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier
);
741 static void iommu_poll_ga_log(struct amd_iommu
*iommu
)
743 u32 head
, tail
, cnt
= 0;
745 if (iommu
->ga_log
== NULL
)
748 head
= readl(iommu
->mmio_base
+ MMIO_GA_HEAD_OFFSET
);
749 tail
= readl(iommu
->mmio_base
+ MMIO_GA_TAIL_OFFSET
);
751 while (head
!= tail
) {
755 raw
= (u64
*)(iommu
->ga_log
+ head
);
758 /* Avoid memcpy function-call overhead */
761 /* Update head pointer of hardware ring-buffer */
762 head
= (head
+ GA_ENTRY_SIZE
) % GA_LOG_SIZE
;
763 writel(head
, iommu
->mmio_base
+ MMIO_GA_HEAD_OFFSET
);
765 /* Handle GA entry */
766 switch (GA_REQ_TYPE(log_entry
)) {
768 if (!iommu_ga_log_notifier
)
771 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
772 __func__
, GA_DEVID(log_entry
),
775 if (iommu_ga_log_notifier(GA_TAG(log_entry
)) != 0)
776 pr_err("GA log notifier failed.\n");
783 #endif /* CONFIG_IRQ_REMAP */
785 #define AMD_IOMMU_INT_MASK \
786 (MMIO_STATUS_EVT_INT_MASK | \
787 MMIO_STATUS_PPR_INT_MASK | \
788 MMIO_STATUS_GALOG_INT_MASK)
790 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
792 struct amd_iommu
*iommu
= (struct amd_iommu
*) data
;
793 u32 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
795 while (status
& AMD_IOMMU_INT_MASK
) {
796 /* Enable EVT and PPR and GA interrupts again */
797 writel(AMD_IOMMU_INT_MASK
,
798 iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
800 if (status
& MMIO_STATUS_EVT_INT_MASK
) {
801 pr_devel("Processing IOMMU Event Log\n");
802 iommu_poll_events(iommu
);
805 if (status
& MMIO_STATUS_PPR_INT_MASK
) {
806 pr_devel("Processing IOMMU PPR Log\n");
807 iommu_poll_ppr_log(iommu
);
810 #ifdef CONFIG_IRQ_REMAP
811 if (status
& MMIO_STATUS_GALOG_INT_MASK
) {
812 pr_devel("Processing IOMMU GA Log\n");
813 iommu_poll_ga_log(iommu
);
818 * Hardware bug: ERBT1312
819 * When re-enabling interrupt (by writing 1
820 * to clear the bit), the hardware might also try to set
821 * the interrupt bit in the event status register.
822 * In this scenario, the bit will be set, and disable
823 * subsequent interrupts.
825 * Workaround: The IOMMU driver should read back the
826 * status register and check if the interrupt bits are cleared.
827 * If not, driver will need to go through the interrupt handler
828 * again and re-clear the bits
830 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
835 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
837 return IRQ_WAKE_THREAD
;
840 /****************************************************************************
842 * IOMMU command queuing functions
844 ****************************************************************************/
846 static int wait_on_sem(volatile u64
*sem
)
850 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
855 if (i
== LOOP_TIMEOUT
) {
856 pr_alert("Completion-Wait loop timed out\n");
863 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
864 struct iommu_cmd
*cmd
)
868 target
= iommu
->cmd_buf
+ iommu
->cmd_buf_tail
;
870 iommu
->cmd_buf_tail
+= sizeof(*cmd
);
871 iommu
->cmd_buf_tail
%= CMD_BUFFER_SIZE
;
873 /* Copy command to buffer */
874 memcpy(target
, cmd
, sizeof(*cmd
));
876 /* Tell the IOMMU about it */
877 writel(iommu
->cmd_buf_tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
880 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
882 u64 paddr
= iommu_virt_to_phys((void *)address
);
884 WARN_ON(address
& 0x7ULL
);
886 memset(cmd
, 0, sizeof(*cmd
));
887 cmd
->data
[0] = lower_32_bits(paddr
) | CMD_COMPL_WAIT_STORE_MASK
;
888 cmd
->data
[1] = upper_32_bits(paddr
);
890 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
893 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
895 memset(cmd
, 0, sizeof(*cmd
));
896 cmd
->data
[0] = devid
;
897 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
900 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
901 size_t size
, u16 domid
, int pde
)
906 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
911 * If we have to flush more than one page, flush all
912 * TLB entries for this domain
914 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
918 address
&= PAGE_MASK
;
920 memset(cmd
, 0, sizeof(*cmd
));
921 cmd
->data
[1] |= domid
;
922 cmd
->data
[2] = lower_32_bits(address
);
923 cmd
->data
[3] = upper_32_bits(address
);
924 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
925 if (s
) /* size bit - we flush more than one 4kb page */
926 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
927 if (pde
) /* PDE bit - we want to flush everything, not only the PTEs */
928 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
931 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
932 u64 address
, size_t size
)
937 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
942 * If we have to flush more than one page, flush all
943 * TLB entries for this domain
945 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
949 address
&= PAGE_MASK
;
951 memset(cmd
, 0, sizeof(*cmd
));
952 cmd
->data
[0] = devid
;
953 cmd
->data
[0] |= (qdep
& 0xff) << 24;
954 cmd
->data
[1] = devid
;
955 cmd
->data
[2] = lower_32_bits(address
);
956 cmd
->data
[3] = upper_32_bits(address
);
957 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
959 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
962 static void build_inv_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
963 u64 address
, bool size
)
965 memset(cmd
, 0, sizeof(*cmd
));
967 address
&= ~(0xfffULL
);
969 cmd
->data
[0] = pasid
;
970 cmd
->data
[1] = domid
;
971 cmd
->data
[2] = lower_32_bits(address
);
972 cmd
->data
[3] = upper_32_bits(address
);
973 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
974 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
976 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
977 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
980 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
981 int qdep
, u64 address
, bool size
)
983 memset(cmd
, 0, sizeof(*cmd
));
985 address
&= ~(0xfffULL
);
987 cmd
->data
[0] = devid
;
988 cmd
->data
[0] |= ((pasid
>> 8) & 0xff) << 16;
989 cmd
->data
[0] |= (qdep
& 0xff) << 24;
990 cmd
->data
[1] = devid
;
991 cmd
->data
[1] |= (pasid
& 0xff) << 16;
992 cmd
->data
[2] = lower_32_bits(address
);
993 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
994 cmd
->data
[3] = upper_32_bits(address
);
996 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
997 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
1000 static void build_complete_ppr(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
1001 int status
, int tag
, bool gn
)
1003 memset(cmd
, 0, sizeof(*cmd
));
1005 cmd
->data
[0] = devid
;
1007 cmd
->data
[1] = pasid
;
1008 cmd
->data
[2] = CMD_INV_IOMMU_PAGES_GN_MASK
;
1010 cmd
->data
[3] = tag
& 0x1ff;
1011 cmd
->data
[3] |= (status
& PPR_STATUS_MASK
) << PPR_STATUS_SHIFT
;
1013 CMD_SET_TYPE(cmd
, CMD_COMPLETE_PPR
);
1016 static void build_inv_all(struct iommu_cmd
*cmd
)
1018 memset(cmd
, 0, sizeof(*cmd
));
1019 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
1022 static void build_inv_irt(struct iommu_cmd
*cmd
, u16 devid
)
1024 memset(cmd
, 0, sizeof(*cmd
));
1025 cmd
->data
[0] = devid
;
1026 CMD_SET_TYPE(cmd
, CMD_INV_IRT
);
1030 * Writes the command to the IOMMUs command buffer and informs the
1031 * hardware about the new command.
1033 static int __iommu_queue_command_sync(struct amd_iommu
*iommu
,
1034 struct iommu_cmd
*cmd
,
1037 unsigned int count
= 0;
1038 u32 left
, next_tail
;
1040 next_tail
= (iommu
->cmd_buf_tail
+ sizeof(*cmd
)) % CMD_BUFFER_SIZE
;
1042 left
= (iommu
->cmd_buf_head
- next_tail
) % CMD_BUFFER_SIZE
;
1045 /* Skip udelay() the first time around */
1047 if (count
== LOOP_TIMEOUT
) {
1048 pr_err("Command buffer timeout\n");
1055 /* Update head and recheck remaining space */
1056 iommu
->cmd_buf_head
= readl(iommu
->mmio_base
+
1057 MMIO_CMD_HEAD_OFFSET
);
1062 copy_cmd_to_buffer(iommu
, cmd
);
1064 /* Do we need to make sure all commands are processed? */
1065 iommu
->need_sync
= sync
;
1070 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
1071 struct iommu_cmd
*cmd
,
1074 unsigned long flags
;
1077 raw_spin_lock_irqsave(&iommu
->lock
, flags
);
1078 ret
= __iommu_queue_command_sync(iommu
, cmd
, sync
);
1079 raw_spin_unlock_irqrestore(&iommu
->lock
, flags
);
1084 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
1086 return iommu_queue_command_sync(iommu
, cmd
, true);
1090 * This function queues a completion wait command into the command
1091 * buffer of an IOMMU
1093 static int iommu_completion_wait(struct amd_iommu
*iommu
)
1095 struct iommu_cmd cmd
;
1096 unsigned long flags
;
1099 if (!iommu
->need_sync
)
1103 build_completion_wait(&cmd
, (u64
)&iommu
->cmd_sem
);
1105 raw_spin_lock_irqsave(&iommu
->lock
, flags
);
1109 ret
= __iommu_queue_command_sync(iommu
, &cmd
, false);
1113 ret
= wait_on_sem(&iommu
->cmd_sem
);
1116 raw_spin_unlock_irqrestore(&iommu
->lock
, flags
);
1121 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
1123 struct iommu_cmd cmd
;
1125 build_inv_dte(&cmd
, devid
);
1127 return iommu_queue_command(iommu
, &cmd
);
1130 static void amd_iommu_flush_dte_all(struct amd_iommu
*iommu
)
1134 for (devid
= 0; devid
<= 0xffff; ++devid
)
1135 iommu_flush_dte(iommu
, devid
);
1137 iommu_completion_wait(iommu
);
1141 * This function uses heavy locking and may disable irqs for some time. But
1142 * this is no issue because it is only called during resume.
1144 static void amd_iommu_flush_tlb_all(struct amd_iommu
*iommu
)
1148 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
1149 struct iommu_cmd cmd
;
1150 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
1152 iommu_queue_command(iommu
, &cmd
);
1155 iommu_completion_wait(iommu
);
1158 static void amd_iommu_flush_all(struct amd_iommu
*iommu
)
1160 struct iommu_cmd cmd
;
1162 build_inv_all(&cmd
);
1164 iommu_queue_command(iommu
, &cmd
);
1165 iommu_completion_wait(iommu
);
1168 static void iommu_flush_irt(struct amd_iommu
*iommu
, u16 devid
)
1170 struct iommu_cmd cmd
;
1172 build_inv_irt(&cmd
, devid
);
1174 iommu_queue_command(iommu
, &cmd
);
1177 static void amd_iommu_flush_irt_all(struct amd_iommu
*iommu
)
1181 for (devid
= 0; devid
<= MAX_DEV_TABLE_ENTRIES
; devid
++)
1182 iommu_flush_irt(iommu
, devid
);
1184 iommu_completion_wait(iommu
);
1187 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
1189 if (iommu_feature(iommu
, FEATURE_IA
)) {
1190 amd_iommu_flush_all(iommu
);
1192 amd_iommu_flush_dte_all(iommu
);
1193 amd_iommu_flush_irt_all(iommu
);
1194 amd_iommu_flush_tlb_all(iommu
);
1199 * Command send function for flushing on-device TLB
1201 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
1202 u64 address
, size_t size
)
1204 struct amd_iommu
*iommu
;
1205 struct iommu_cmd cmd
;
1208 qdep
= dev_data
->ats
.qdep
;
1209 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1211 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
1213 return iommu_queue_command(iommu
, &cmd
);
1217 * Command send function for invalidating a device table entry
1219 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
1221 struct amd_iommu
*iommu
;
1225 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1226 alias
= dev_data
->alias
;
1228 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
1229 if (!ret
&& alias
!= dev_data
->devid
)
1230 ret
= iommu_flush_dte(iommu
, alias
);
1234 if (dev_data
->ats
.enabled
)
1235 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
1241 * TLB invalidation function which is called from the mapping functions.
1242 * It invalidates a single PTE if the range to flush is within a single
1243 * page. Otherwise it flushes the whole TLB of the IOMMU.
1245 static void __domain_flush_pages(struct protection_domain
*domain
,
1246 u64 address
, size_t size
, int pde
)
1248 struct iommu_dev_data
*dev_data
;
1249 struct iommu_cmd cmd
;
1252 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
1254 for (i
= 0; i
< amd_iommu_get_num_iommus(); ++i
) {
1255 if (!domain
->dev_iommu
[i
])
1259 * Devices of this domain are behind this IOMMU
1260 * We need a TLB flush
1262 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
1265 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1267 if (!dev_data
->ats
.enabled
)
1270 ret
|= device_flush_iotlb(dev_data
, address
, size
);
1276 static void domain_flush_pages(struct protection_domain
*domain
,
1277 u64 address
, size_t size
)
1279 __domain_flush_pages(domain
, address
, size
, 0);
1282 /* Flush the whole IO/TLB for a given protection domain */
1283 static void domain_flush_tlb(struct protection_domain
*domain
)
1285 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
1288 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1289 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
1291 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
1294 static void domain_flush_complete(struct protection_domain
*domain
)
1298 for (i
= 0; i
< amd_iommu_get_num_iommus(); ++i
) {
1299 if (domain
&& !domain
->dev_iommu
[i
])
1303 * Devices of this domain are behind this IOMMU
1304 * We need to wait for completion of all commands.
1306 iommu_completion_wait(amd_iommus
[i
]);
1312 * This function flushes the DTEs for all devices in domain
1314 static void domain_flush_devices(struct protection_domain
*domain
)
1316 struct iommu_dev_data
*dev_data
;
1318 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1319 device_flush_dte(dev_data
);
1322 /****************************************************************************
1324 * The functions below are used the create the page table mappings for
1325 * unity mapped regions.
1327 ****************************************************************************/
1329 static void free_page_list(struct page
*freelist
)
1331 while (freelist
!= NULL
) {
1332 unsigned long p
= (unsigned long)page_address(freelist
);
1333 freelist
= freelist
->freelist
;
1338 static struct page
*free_pt_page(unsigned long pt
, struct page
*freelist
)
1340 struct page
*p
= virt_to_page((void *)pt
);
1342 p
->freelist
= freelist
;
1347 #define DEFINE_FREE_PT_FN(LVL, FN) \
1348 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist) \
1356 for (i = 0; i < 512; ++i) { \
1357 /* PTE present? */ \
1358 if (!IOMMU_PTE_PRESENT(pt[i])) \
1362 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1363 PM_PTE_LEVEL(pt[i]) == 7) \
1366 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1367 freelist = FN(p, freelist); \
1370 return free_pt_page((unsigned long)pt, freelist); \
1373 DEFINE_FREE_PT_FN(l2
, free_pt_page
)
1374 DEFINE_FREE_PT_FN(l3
, free_pt_l2
)
1375 DEFINE_FREE_PT_FN(l4
, free_pt_l3
)
1376 DEFINE_FREE_PT_FN(l5
, free_pt_l4
)
1377 DEFINE_FREE_PT_FN(l6
, free_pt_l5
)
1379 static struct page
*free_sub_pt(unsigned long root
, int mode
,
1380 struct page
*freelist
)
1383 case PAGE_MODE_NONE
:
1384 case PAGE_MODE_7_LEVEL
:
1386 case PAGE_MODE_1_LEVEL
:
1387 freelist
= free_pt_page(root
, freelist
);
1389 case PAGE_MODE_2_LEVEL
:
1390 freelist
= free_pt_l2(root
, freelist
);
1392 case PAGE_MODE_3_LEVEL
:
1393 freelist
= free_pt_l3(root
, freelist
);
1395 case PAGE_MODE_4_LEVEL
:
1396 freelist
= free_pt_l4(root
, freelist
);
1398 case PAGE_MODE_5_LEVEL
:
1399 freelist
= free_pt_l5(root
, freelist
);
1401 case PAGE_MODE_6_LEVEL
:
1402 freelist
= free_pt_l6(root
, freelist
);
1411 static void free_pagetable(struct protection_domain
*domain
)
1413 unsigned long root
= (unsigned long)domain
->pt_root
;
1414 struct page
*freelist
= NULL
;
1416 BUG_ON(domain
->mode
< PAGE_MODE_NONE
||
1417 domain
->mode
> PAGE_MODE_6_LEVEL
);
1419 free_sub_pt(root
, domain
->mode
, freelist
);
1421 free_page_list(freelist
);
1425 * This function is used to add another level to an IO page table. Adding
1426 * another level increases the size of the address space by 9 bits to a size up
1429 static bool increase_address_space(struct protection_domain
*domain
,
1434 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1435 /* address space already 64 bit large */
1438 pte
= (void *)get_zeroed_page(gfp
);
1442 *pte
= PM_LEVEL_PDE(domain
->mode
,
1443 iommu_virt_to_phys(domain
->pt_root
));
1444 domain
->pt_root
= pte
;
1446 domain
->updated
= true;
1451 static u64
*alloc_pte(struct protection_domain
*domain
,
1452 unsigned long address
,
1453 unsigned long page_size
,
1460 BUG_ON(!is_power_of_2(page_size
));
1462 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1463 increase_address_space(domain
, gfp
);
1465 level
= domain
->mode
- 1;
1466 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1467 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1468 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1470 while (level
> end_lvl
) {
1475 pte_level
= PM_PTE_LEVEL(__pte
);
1477 if (!IOMMU_PTE_PRESENT(__pte
) ||
1478 pte_level
== PAGE_MODE_7_LEVEL
) {
1479 page
= (u64
*)get_zeroed_page(gfp
);
1483 __npte
= PM_LEVEL_PDE(level
, iommu_virt_to_phys(page
));
1485 /* pte could have been changed somewhere. */
1486 if (cmpxchg64(pte
, __pte
, __npte
) != __pte
)
1487 free_page((unsigned long)page
);
1488 else if (pte_level
== PAGE_MODE_7_LEVEL
)
1489 domain
->updated
= true;
1494 /* No level skipping support yet */
1495 if (pte_level
!= level
)
1500 pte
= IOMMU_PTE_PAGE(__pte
);
1502 if (pte_page
&& level
== end_lvl
)
1505 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1512 * This function checks if there is a PTE for a given dma address. If
1513 * there is one, it returns the pointer to it.
1515 static u64
*fetch_pte(struct protection_domain
*domain
,
1516 unsigned long address
,
1517 unsigned long *page_size
)
1524 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1527 level
= domain
->mode
- 1;
1528 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1529 *page_size
= PTE_LEVEL_PAGE_SIZE(level
);
1534 if (!IOMMU_PTE_PRESENT(*pte
))
1538 if (PM_PTE_LEVEL(*pte
) == 7 ||
1539 PM_PTE_LEVEL(*pte
) == 0)
1542 /* No level skipping support yet */
1543 if (PM_PTE_LEVEL(*pte
) != level
)
1548 /* Walk to the next level */
1549 pte
= IOMMU_PTE_PAGE(*pte
);
1550 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1551 *page_size
= PTE_LEVEL_PAGE_SIZE(level
);
1554 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1555 unsigned long pte_mask
;
1558 * If we have a series of large PTEs, make
1559 * sure to return a pointer to the first one.
1561 *page_size
= pte_mask
= PTE_PAGE_SIZE(*pte
);
1562 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1563 pte
= (u64
*)(((unsigned long)pte
) & pte_mask
);
1569 static struct page
*free_clear_pte(u64
*pte
, u64 pteval
, struct page
*freelist
)
1574 while (cmpxchg64(pte
, pteval
, 0) != pteval
) {
1575 pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1579 if (!IOMMU_PTE_PRESENT(pteval
))
1582 pt
= (unsigned long)IOMMU_PTE_PAGE(pteval
);
1583 mode
= IOMMU_PTE_MODE(pteval
);
1585 return free_sub_pt(pt
, mode
, freelist
);
1589 * Generic mapping functions. It maps a physical address into a DMA
1590 * address space. It allocates the page table pages if necessary.
1591 * In the future it can be extended to a generic mapping function
1592 * supporting all features of AMD IOMMU page tables like level skipping
1593 * and full 64 bit address spaces.
1595 static int iommu_map_page(struct protection_domain
*dom
,
1596 unsigned long bus_addr
,
1597 unsigned long phys_addr
,
1598 unsigned long page_size
,
1602 struct page
*freelist
= NULL
;
1606 BUG_ON(!IS_ALIGNED(bus_addr
, page_size
));
1607 BUG_ON(!IS_ALIGNED(phys_addr
, page_size
));
1609 if (!(prot
& IOMMU_PROT_MASK
))
1612 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1613 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, gfp
);
1618 for (i
= 0; i
< count
; ++i
)
1619 freelist
= free_clear_pte(&pte
[i
], pte
[i
], freelist
);
1621 if (freelist
!= NULL
)
1622 dom
->updated
= true;
1625 __pte
= PAGE_SIZE_PTE(__sme_set(phys_addr
), page_size
);
1626 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_PR
| IOMMU_PTE_FC
;
1628 __pte
= __sme_set(phys_addr
) | IOMMU_PTE_PR
| IOMMU_PTE_FC
;
1630 if (prot
& IOMMU_PROT_IR
)
1631 __pte
|= IOMMU_PTE_IR
;
1632 if (prot
& IOMMU_PROT_IW
)
1633 __pte
|= IOMMU_PTE_IW
;
1635 for (i
= 0; i
< count
; ++i
)
1640 /* Everything flushed out, free pages now */
1641 free_page_list(freelist
);
1646 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1647 unsigned long bus_addr
,
1648 unsigned long page_size
)
1650 unsigned long long unmapped
;
1651 unsigned long unmap_size
;
1654 BUG_ON(!is_power_of_2(page_size
));
1658 while (unmapped
< page_size
) {
1660 pte
= fetch_pte(dom
, bus_addr
, &unmap_size
);
1665 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1666 for (i
= 0; i
< count
; i
++)
1670 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1671 unmapped
+= unmap_size
;
1674 BUG_ON(unmapped
&& !is_power_of_2(unmapped
));
1679 /****************************************************************************
1681 * The next functions belong to the address allocator for the dma_ops
1682 * interface functions.
1684 ****************************************************************************/
1687 static unsigned long dma_ops_alloc_iova(struct device
*dev
,
1688 struct dma_ops_domain
*dma_dom
,
1689 unsigned int pages
, u64 dma_mask
)
1691 unsigned long pfn
= 0;
1693 pages
= __roundup_pow_of_two(pages
);
1695 if (dma_mask
> DMA_BIT_MASK(32))
1696 pfn
= alloc_iova_fast(&dma_dom
->iovad
, pages
,
1697 IOVA_PFN(DMA_BIT_MASK(32)), false);
1700 pfn
= alloc_iova_fast(&dma_dom
->iovad
, pages
,
1701 IOVA_PFN(dma_mask
), true);
1703 return (pfn
<< PAGE_SHIFT
);
1706 static void dma_ops_free_iova(struct dma_ops_domain
*dma_dom
,
1707 unsigned long address
,
1710 pages
= __roundup_pow_of_two(pages
);
1711 address
>>= PAGE_SHIFT
;
1713 free_iova_fast(&dma_dom
->iovad
, address
, pages
);
1716 /****************************************************************************
1718 * The next functions belong to the domain allocation. A domain is
1719 * allocated for every IOMMU as the default domain. If device isolation
1720 * is enabled, every device get its own domain. The most important thing
1721 * about domains is the page table mapping the DMA address space they
1724 ****************************************************************************/
1726 static u16
domain_id_alloc(void)
1730 spin_lock(&pd_bitmap_lock
);
1731 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1733 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1734 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1737 spin_unlock(&pd_bitmap_lock
);
1742 static void domain_id_free(int id
)
1744 spin_lock(&pd_bitmap_lock
);
1745 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1746 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1747 spin_unlock(&pd_bitmap_lock
);
1750 static void free_gcr3_tbl_level1(u64
*tbl
)
1755 for (i
= 0; i
< 512; ++i
) {
1756 if (!(tbl
[i
] & GCR3_VALID
))
1759 ptr
= iommu_phys_to_virt(tbl
[i
] & PAGE_MASK
);
1761 free_page((unsigned long)ptr
);
1765 static void free_gcr3_tbl_level2(u64
*tbl
)
1770 for (i
= 0; i
< 512; ++i
) {
1771 if (!(tbl
[i
] & GCR3_VALID
))
1774 ptr
= iommu_phys_to_virt(tbl
[i
] & PAGE_MASK
);
1776 free_gcr3_tbl_level1(ptr
);
1780 static void free_gcr3_table(struct protection_domain
*domain
)
1782 if (domain
->glx
== 2)
1783 free_gcr3_tbl_level2(domain
->gcr3_tbl
);
1784 else if (domain
->glx
== 1)
1785 free_gcr3_tbl_level1(domain
->gcr3_tbl
);
1787 BUG_ON(domain
->glx
!= 0);
1789 free_page((unsigned long)domain
->gcr3_tbl
);
1792 static void dma_ops_domain_flush_tlb(struct dma_ops_domain
*dom
)
1794 domain_flush_tlb(&dom
->domain
);
1795 domain_flush_complete(&dom
->domain
);
1798 static void iova_domain_flush_tlb(struct iova_domain
*iovad
)
1800 struct dma_ops_domain
*dom
;
1802 dom
= container_of(iovad
, struct dma_ops_domain
, iovad
);
1804 dma_ops_domain_flush_tlb(dom
);
1808 * Free a domain, only used if something went wrong in the
1809 * allocation path and we need to free an already allocated page table
1811 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1816 put_iova_domain(&dom
->iovad
);
1818 free_pagetable(&dom
->domain
);
1821 domain_id_free(dom
->domain
.id
);
1827 * Allocates a new protection domain usable for the dma_ops functions.
1828 * It also initializes the page table and the address allocator data
1829 * structures required for the dma_ops interface
1831 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1833 struct dma_ops_domain
*dma_dom
;
1835 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1839 if (protection_domain_init(&dma_dom
->domain
))
1842 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
1843 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1844 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1845 if (!dma_dom
->domain
.pt_root
)
1848 init_iova_domain(&dma_dom
->iovad
, PAGE_SIZE
, IOVA_START_PFN
);
1850 if (init_iova_flush_queue(&dma_dom
->iovad
, iova_domain_flush_tlb
, NULL
))
1853 /* Initialize reserved ranges */
1854 copy_reserved_iova(&reserved_iova_ranges
, &dma_dom
->iovad
);
1859 dma_ops_domain_free(dma_dom
);
1865 * little helper function to check whether a given protection domain is a
1868 static bool dma_ops_domain(struct protection_domain
*domain
)
1870 return domain
->flags
& PD_DMA_OPS_MASK
;
1873 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
,
1879 if (domain
->mode
!= PAGE_MODE_NONE
)
1880 pte_root
= iommu_virt_to_phys(domain
->pt_root
);
1882 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1883 << DEV_ENTRY_MODE_SHIFT
;
1884 pte_root
|= DTE_FLAG_IR
| DTE_FLAG_IW
| DTE_FLAG_V
| DTE_FLAG_TV
;
1886 flags
= amd_iommu_dev_table
[devid
].data
[1];
1889 flags
|= DTE_FLAG_IOTLB
;
1892 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
1894 if (iommu_feature(iommu
, FEATURE_EPHSUP
))
1895 pte_root
|= 1ULL << DEV_ENTRY_PPR
;
1898 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1899 u64 gcr3
= iommu_virt_to_phys(domain
->gcr3_tbl
);
1900 u64 glx
= domain
->glx
;
1903 pte_root
|= DTE_FLAG_GV
;
1904 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1906 /* First mask out possible old values for GCR3 table */
1907 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1910 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1913 /* Encode GCR3 table into DTE */
1914 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1917 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1920 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1924 flags
&= ~DEV_DOMID_MASK
;
1925 flags
|= domain
->id
;
1927 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1928 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1931 static void clear_dte_entry(u16 devid
)
1933 /* remove entry from the device table seen by the hardware */
1934 amd_iommu_dev_table
[devid
].data
[0] = DTE_FLAG_V
| DTE_FLAG_TV
;
1935 amd_iommu_dev_table
[devid
].data
[1] &= DTE_FLAG_MASK
;
1937 amd_iommu_apply_erratum_63(devid
);
1940 static void do_attach(struct iommu_dev_data
*dev_data
,
1941 struct protection_domain
*domain
)
1943 struct amd_iommu
*iommu
;
1947 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1948 alias
= dev_data
->alias
;
1949 ats
= dev_data
->ats
.enabled
;
1951 /* Update data structures */
1952 dev_data
->domain
= domain
;
1953 list_add(&dev_data
->list
, &domain
->dev_list
);
1955 /* Do reference counting */
1956 domain
->dev_iommu
[iommu
->index
] += 1;
1957 domain
->dev_cnt
+= 1;
1959 /* Update device table */
1960 set_dte_entry(dev_data
->devid
, domain
, ats
, dev_data
->iommu_v2
);
1961 if (alias
!= dev_data
->devid
)
1962 set_dte_entry(alias
, domain
, ats
, dev_data
->iommu_v2
);
1964 device_flush_dte(dev_data
);
1967 static void do_detach(struct iommu_dev_data
*dev_data
)
1969 struct protection_domain
*domain
= dev_data
->domain
;
1970 struct amd_iommu
*iommu
;
1973 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1974 alias
= dev_data
->alias
;
1976 /* Update data structures */
1977 dev_data
->domain
= NULL
;
1978 list_del(&dev_data
->list
);
1979 clear_dte_entry(dev_data
->devid
);
1980 if (alias
!= dev_data
->devid
)
1981 clear_dte_entry(alias
);
1983 /* Flush the DTE entry */
1984 device_flush_dte(dev_data
);
1987 domain_flush_tlb_pde(domain
);
1989 /* Wait for the flushes to finish */
1990 domain_flush_complete(domain
);
1992 /* decrease reference counters - needs to happen after the flushes */
1993 domain
->dev_iommu
[iommu
->index
] -= 1;
1994 domain
->dev_cnt
-= 1;
1998 * If a device is not yet associated with a domain, this function makes the
1999 * device visible in the domain
2001 static int __attach_device(struct iommu_dev_data
*dev_data
,
2002 struct protection_domain
*domain
)
2007 spin_lock(&domain
->lock
);
2010 if (dev_data
->domain
!= NULL
)
2013 /* Attach alias group root */
2014 do_attach(dev_data
, domain
);
2021 spin_unlock(&domain
->lock
);
2027 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
2029 pci_disable_ats(pdev
);
2030 pci_disable_pri(pdev
);
2031 pci_disable_pasid(pdev
);
2034 /* FIXME: Change generic reset-function to do the same */
2035 static int pri_reset_while_enabled(struct pci_dev
*pdev
)
2040 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2044 pci_read_config_word(pdev
, pos
+ PCI_PRI_CTRL
, &control
);
2045 control
|= PCI_PRI_CTRL_RESET
;
2046 pci_write_config_word(pdev
, pos
+ PCI_PRI_CTRL
, control
);
2051 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
2056 /* FIXME: Hardcode number of outstanding requests for now */
2058 if (pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE
))
2060 reset_enable
= pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_ENABLE_RESET
);
2062 /* Only allow access to user-accessible pages */
2063 ret
= pci_enable_pasid(pdev
, 0);
2067 /* First reset the PRI state of the device */
2068 ret
= pci_reset_pri(pdev
);
2073 ret
= pci_enable_pri(pdev
, reqs
);
2078 ret
= pri_reset_while_enabled(pdev
);
2083 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
2090 pci_disable_pri(pdev
);
2091 pci_disable_pasid(pdev
);
2097 * If a device is not yet associated with a domain, this function makes the
2098 * device visible in the domain
2100 static int attach_device(struct device
*dev
,
2101 struct protection_domain
*domain
)
2103 struct pci_dev
*pdev
;
2104 struct iommu_dev_data
*dev_data
;
2105 unsigned long flags
;
2108 dev_data
= get_dev_data(dev
);
2110 if (!dev_is_pci(dev
))
2111 goto skip_ats_check
;
2113 pdev
= to_pci_dev(dev
);
2114 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2115 if (!dev_data
->passthrough
)
2118 if (dev_data
->iommu_v2
) {
2119 if (pdev_iommuv2_enable(pdev
) != 0)
2122 dev_data
->ats
.enabled
= true;
2123 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2124 dev_data
->pri_tlp
= pci_prg_resp_pasid_required(pdev
);
2126 } else if (amd_iommu_iotlb_sup
&&
2127 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
2128 dev_data
->ats
.enabled
= true;
2129 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2133 spin_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2134 ret
= __attach_device(dev_data
, domain
);
2135 spin_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2138 * We might boot into a crash-kernel here. The crashed kernel
2139 * left the caches in the IOMMU dirty. So we have to flush
2140 * here to evict all dirty stuff.
2142 domain_flush_tlb_pde(domain
);
2148 * Removes a device from a protection domain (unlocked)
2150 static void __detach_device(struct iommu_dev_data
*dev_data
)
2152 struct protection_domain
*domain
;
2154 domain
= dev_data
->domain
;
2156 spin_lock(&domain
->lock
);
2158 do_detach(dev_data
);
2160 spin_unlock(&domain
->lock
);
2164 * Removes a device from a protection domain (with devtable_lock held)
2166 static void detach_device(struct device
*dev
)
2168 struct protection_domain
*domain
;
2169 struct iommu_dev_data
*dev_data
;
2170 unsigned long flags
;
2172 dev_data
= get_dev_data(dev
);
2173 domain
= dev_data
->domain
;
2176 * First check if the device is still attached. It might already
2177 * be detached from its domain because the generic
2178 * iommu_detach_group code detached it and we try again here in
2179 * our alias handling.
2181 if (WARN_ON(!dev_data
->domain
))
2184 /* lock device table */
2185 spin_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2186 __detach_device(dev_data
);
2187 spin_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2189 if (!dev_is_pci(dev
))
2192 if (domain
->flags
& PD_IOMMUV2_MASK
&& dev_data
->iommu_v2
)
2193 pdev_iommuv2_disable(to_pci_dev(dev
));
2194 else if (dev_data
->ats
.enabled
)
2195 pci_disable_ats(to_pci_dev(dev
));
2197 dev_data
->ats
.enabled
= false;
2200 static int amd_iommu_add_device(struct device
*dev
)
2202 struct iommu_dev_data
*dev_data
;
2203 struct iommu_domain
*domain
;
2204 struct amd_iommu
*iommu
;
2207 if (!check_device(dev
) || get_dev_data(dev
))
2210 devid
= get_device_id(dev
);
2214 iommu
= amd_iommu_rlookup_table
[devid
];
2216 ret
= iommu_init_device(dev
);
2218 if (ret
!= -ENOTSUPP
)
2219 dev_err(dev
, "Failed to initialize - trying to proceed anyway\n");
2221 iommu_ignore_device(dev
);
2222 dev
->dma_ops
= NULL
;
2225 init_iommu_group(dev
);
2227 dev_data
= get_dev_data(dev
);
2231 if (iommu_pass_through
|| dev_data
->iommu_v2
)
2232 iommu_request_dm_for_dev(dev
);
2234 /* Domains are initialized for this device - have a look what we ended up with */
2235 domain
= iommu_get_domain_for_dev(dev
);
2236 if (domain
->type
== IOMMU_DOMAIN_IDENTITY
)
2237 dev_data
->passthrough
= true;
2239 dev
->dma_ops
= &amd_iommu_dma_ops
;
2242 iommu_completion_wait(iommu
);
2247 static void amd_iommu_remove_device(struct device
*dev
)
2249 struct amd_iommu
*iommu
;
2252 if (!check_device(dev
))
2255 devid
= get_device_id(dev
);
2259 iommu
= amd_iommu_rlookup_table
[devid
];
2261 iommu_uninit_device(dev
);
2262 iommu_completion_wait(iommu
);
2265 static struct iommu_group
*amd_iommu_device_group(struct device
*dev
)
2267 if (dev_is_pci(dev
))
2268 return pci_device_group(dev
);
2270 return acpihid_device_group(dev
);
2273 /*****************************************************************************
2275 * The next functions belong to the dma_ops mapping/unmapping code.
2277 *****************************************************************************/
2280 * In the dma_ops path we only have the struct device. This function
2281 * finds the corresponding IOMMU, the protection domain and the
2282 * requestor id for a given device.
2283 * If the device is not yet associated with a domain this is also done
2286 static struct protection_domain
*get_domain(struct device
*dev
)
2288 struct protection_domain
*domain
;
2289 struct iommu_domain
*io_domain
;
2291 if (!check_device(dev
))
2292 return ERR_PTR(-EINVAL
);
2294 domain
= get_dev_data(dev
)->domain
;
2295 if (domain
== NULL
&& get_dev_data(dev
)->defer_attach
) {
2296 get_dev_data(dev
)->defer_attach
= false;
2297 io_domain
= iommu_get_domain_for_dev(dev
);
2298 domain
= to_pdomain(io_domain
);
2299 attach_device(dev
, domain
);
2302 return ERR_PTR(-EBUSY
);
2304 if (!dma_ops_domain(domain
))
2305 return ERR_PTR(-EBUSY
);
2310 static void update_device_table(struct protection_domain
*domain
)
2312 struct iommu_dev_data
*dev_data
;
2314 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
2315 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
,
2316 dev_data
->iommu_v2
);
2318 if (dev_data
->devid
== dev_data
->alias
)
2321 /* There is an alias, update device table entry for it */
2322 set_dte_entry(dev_data
->alias
, domain
, dev_data
->ats
.enabled
,
2323 dev_data
->iommu_v2
);
2327 static void update_domain(struct protection_domain
*domain
)
2329 if (!domain
->updated
)
2332 update_device_table(domain
);
2334 domain_flush_devices(domain
);
2335 domain_flush_tlb_pde(domain
);
2337 domain
->updated
= false;
2340 static int dir2prot(enum dma_data_direction direction
)
2342 if (direction
== DMA_TO_DEVICE
)
2343 return IOMMU_PROT_IR
;
2344 else if (direction
== DMA_FROM_DEVICE
)
2345 return IOMMU_PROT_IW
;
2346 else if (direction
== DMA_BIDIRECTIONAL
)
2347 return IOMMU_PROT_IW
| IOMMU_PROT_IR
;
2353 * This function contains common code for mapping of a physically
2354 * contiguous memory region into DMA address space. It is used by all
2355 * mapping functions provided with this IOMMU driver.
2356 * Must be called with the domain lock held.
2358 static dma_addr_t
__map_single(struct device
*dev
,
2359 struct dma_ops_domain
*dma_dom
,
2362 enum dma_data_direction direction
,
2365 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2366 dma_addr_t address
, start
, ret
;
2371 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2374 address
= dma_ops_alloc_iova(dev
, dma_dom
, pages
, dma_mask
);
2378 prot
= dir2prot(direction
);
2381 for (i
= 0; i
< pages
; ++i
) {
2382 ret
= iommu_map_page(&dma_dom
->domain
, start
, paddr
,
2383 PAGE_SIZE
, prot
, GFP_ATOMIC
);
2392 if (unlikely(amd_iommu_np_cache
)) {
2393 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2394 domain_flush_complete(&dma_dom
->domain
);
2402 for (--i
; i
>= 0; --i
) {
2404 iommu_unmap_page(&dma_dom
->domain
, start
, PAGE_SIZE
);
2407 domain_flush_tlb(&dma_dom
->domain
);
2408 domain_flush_complete(&dma_dom
->domain
);
2410 dma_ops_free_iova(dma_dom
, address
, pages
);
2412 return DMA_MAPPING_ERROR
;
2416 * Does the reverse of the __map_single function. Must be called with
2417 * the domain lock held too
2419 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2420 dma_addr_t dma_addr
,
2424 dma_addr_t i
, start
;
2427 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2428 dma_addr
&= PAGE_MASK
;
2431 for (i
= 0; i
< pages
; ++i
) {
2432 iommu_unmap_page(&dma_dom
->domain
, start
, PAGE_SIZE
);
2436 if (amd_iommu_unmap_flush
) {
2437 domain_flush_tlb(&dma_dom
->domain
);
2438 domain_flush_complete(&dma_dom
->domain
);
2439 dma_ops_free_iova(dma_dom
, dma_addr
, pages
);
2441 pages
= __roundup_pow_of_two(pages
);
2442 queue_iova(&dma_dom
->iovad
, dma_addr
>> PAGE_SHIFT
, pages
, 0);
2447 * The exported map_single function for dma_ops.
2449 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2450 unsigned long offset
, size_t size
,
2451 enum dma_data_direction dir
,
2452 unsigned long attrs
)
2454 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2455 struct protection_domain
*domain
;
2456 struct dma_ops_domain
*dma_dom
;
2459 domain
= get_domain(dev
);
2460 if (PTR_ERR(domain
) == -EINVAL
)
2461 return (dma_addr_t
)paddr
;
2462 else if (IS_ERR(domain
))
2463 return DMA_MAPPING_ERROR
;
2465 dma_mask
= *dev
->dma_mask
;
2466 dma_dom
= to_dma_ops_domain(domain
);
2468 return __map_single(dev
, dma_dom
, paddr
, size
, dir
, dma_mask
);
2472 * The exported unmap_single function for dma_ops.
2474 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2475 enum dma_data_direction dir
, unsigned long attrs
)
2477 struct protection_domain
*domain
;
2478 struct dma_ops_domain
*dma_dom
;
2480 domain
= get_domain(dev
);
2484 dma_dom
= to_dma_ops_domain(domain
);
2486 __unmap_single(dma_dom
, dma_addr
, size
, dir
);
2489 static int sg_num_pages(struct device
*dev
,
2490 struct scatterlist
*sglist
,
2493 unsigned long mask
, boundary_size
;
2494 struct scatterlist
*s
;
2497 mask
= dma_get_seg_boundary(dev
);
2498 boundary_size
= mask
+ 1 ? ALIGN(mask
+ 1, PAGE_SIZE
) >> PAGE_SHIFT
:
2499 1UL << (BITS_PER_LONG
- PAGE_SHIFT
);
2501 for_each_sg(sglist
, s
, nelems
, i
) {
2504 s
->dma_address
= npages
<< PAGE_SHIFT
;
2505 p
= npages
% boundary_size
;
2506 n
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2507 if (p
+ n
> boundary_size
)
2508 npages
+= boundary_size
- p
;
2516 * The exported map_sg function for dma_ops (handles scatter-gather
2519 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2520 int nelems
, enum dma_data_direction direction
,
2521 unsigned long attrs
)
2523 int mapped_pages
= 0, npages
= 0, prot
= 0, i
;
2524 struct protection_domain
*domain
;
2525 struct dma_ops_domain
*dma_dom
;
2526 struct scatterlist
*s
;
2527 unsigned long address
;
2531 domain
= get_domain(dev
);
2535 dma_dom
= to_dma_ops_domain(domain
);
2536 dma_mask
= *dev
->dma_mask
;
2538 npages
= sg_num_pages(dev
, sglist
, nelems
);
2540 address
= dma_ops_alloc_iova(dev
, dma_dom
, npages
, dma_mask
);
2541 if (address
== DMA_MAPPING_ERROR
)
2544 prot
= dir2prot(direction
);
2546 /* Map all sg entries */
2547 for_each_sg(sglist
, s
, nelems
, i
) {
2548 int j
, pages
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2550 for (j
= 0; j
< pages
; ++j
) {
2551 unsigned long bus_addr
, phys_addr
;
2553 bus_addr
= address
+ s
->dma_address
+ (j
<< PAGE_SHIFT
);
2554 phys_addr
= (sg_phys(s
) & PAGE_MASK
) + (j
<< PAGE_SHIFT
);
2555 ret
= iommu_map_page(domain
, bus_addr
, phys_addr
, PAGE_SIZE
, prot
, GFP_ATOMIC
);
2563 /* Everything is mapped - write the right values into s->dma_address */
2564 for_each_sg(sglist
, s
, nelems
, i
) {
2566 * Add in the remaining piece of the scatter-gather offset that
2567 * was masked out when we were determining the physical address
2568 * via (sg_phys(s) & PAGE_MASK) earlier.
2570 s
->dma_address
+= address
+ (s
->offset
& ~PAGE_MASK
);
2571 s
->dma_length
= s
->length
;
2577 dev_err(dev
, "IOMMU mapping error in map_sg (io-pages: %d reason: %d)\n",
2580 for_each_sg(sglist
, s
, nelems
, i
) {
2581 int j
, pages
= iommu_num_pages(sg_phys(s
), s
->length
, PAGE_SIZE
);
2583 for (j
= 0; j
< pages
; ++j
) {
2584 unsigned long bus_addr
;
2586 bus_addr
= address
+ s
->dma_address
+ (j
<< PAGE_SHIFT
);
2587 iommu_unmap_page(domain
, bus_addr
, PAGE_SIZE
);
2589 if (--mapped_pages
== 0)
2595 free_iova_fast(&dma_dom
->iovad
, address
>> PAGE_SHIFT
, npages
);
2602 * The exported map_sg function for dma_ops (handles scatter-gather
2605 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2606 int nelems
, enum dma_data_direction dir
,
2607 unsigned long attrs
)
2609 struct protection_domain
*domain
;
2610 struct dma_ops_domain
*dma_dom
;
2611 unsigned long startaddr
;
2614 domain
= get_domain(dev
);
2618 startaddr
= sg_dma_address(sglist
) & PAGE_MASK
;
2619 dma_dom
= to_dma_ops_domain(domain
);
2620 npages
= sg_num_pages(dev
, sglist
, nelems
);
2622 __unmap_single(dma_dom
, startaddr
, npages
<< PAGE_SHIFT
, dir
);
2626 * The exported alloc_coherent function for dma_ops.
2628 static void *alloc_coherent(struct device
*dev
, size_t size
,
2629 dma_addr_t
*dma_addr
, gfp_t flag
,
2630 unsigned long attrs
)
2632 u64 dma_mask
= dev
->coherent_dma_mask
;
2633 struct protection_domain
*domain
;
2634 struct dma_ops_domain
*dma_dom
;
2637 domain
= get_domain(dev
);
2638 if (PTR_ERR(domain
) == -EINVAL
) {
2639 page
= alloc_pages(flag
, get_order(size
));
2640 *dma_addr
= page_to_phys(page
);
2641 return page_address(page
);
2642 } else if (IS_ERR(domain
))
2645 dma_dom
= to_dma_ops_domain(domain
);
2646 size
= PAGE_ALIGN(size
);
2647 dma_mask
= dev
->coherent_dma_mask
;
2648 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2651 page
= alloc_pages(flag
| __GFP_NOWARN
, get_order(size
));
2653 if (!gfpflags_allow_blocking(flag
))
2656 page
= dma_alloc_from_contiguous(dev
, size
>> PAGE_SHIFT
,
2657 get_order(size
), flag
& __GFP_NOWARN
);
2663 dma_mask
= *dev
->dma_mask
;
2665 *dma_addr
= __map_single(dev
, dma_dom
, page_to_phys(page
),
2666 size
, DMA_BIDIRECTIONAL
, dma_mask
);
2668 if (*dma_addr
== DMA_MAPPING_ERROR
)
2671 return page_address(page
);
2675 if (!dma_release_from_contiguous(dev
, page
, size
>> PAGE_SHIFT
))
2676 __free_pages(page
, get_order(size
));
2682 * The exported free_coherent function for dma_ops.
2684 static void free_coherent(struct device
*dev
, size_t size
,
2685 void *virt_addr
, dma_addr_t dma_addr
,
2686 unsigned long attrs
)
2688 struct protection_domain
*domain
;
2689 struct dma_ops_domain
*dma_dom
;
2692 page
= virt_to_page(virt_addr
);
2693 size
= PAGE_ALIGN(size
);
2695 domain
= get_domain(dev
);
2699 dma_dom
= to_dma_ops_domain(domain
);
2701 __unmap_single(dma_dom
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2704 if (!dma_release_from_contiguous(dev
, page
, size
>> PAGE_SHIFT
))
2705 __free_pages(page
, get_order(size
));
2709 * This function is called by the DMA layer to find out if we can handle a
2710 * particular device. It is part of the dma_ops.
2712 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2714 if (!dma_direct_supported(dev
, mask
))
2716 return check_device(dev
);
2719 static const struct dma_map_ops amd_iommu_dma_ops
= {
2720 .alloc
= alloc_coherent
,
2721 .free
= free_coherent
,
2722 .map_page
= map_page
,
2723 .unmap_page
= unmap_page
,
2725 .unmap_sg
= unmap_sg
,
2726 .dma_supported
= amd_iommu_dma_supported
,
2729 static int init_reserved_iova_ranges(void)
2731 struct pci_dev
*pdev
= NULL
;
2734 init_iova_domain(&reserved_iova_ranges
, PAGE_SIZE
, IOVA_START_PFN
);
2736 lockdep_set_class(&reserved_iova_ranges
.iova_rbtree_lock
,
2737 &reserved_rbtree_key
);
2739 /* MSI memory range */
2740 val
= reserve_iova(&reserved_iova_ranges
,
2741 IOVA_PFN(MSI_RANGE_START
), IOVA_PFN(MSI_RANGE_END
));
2743 pr_err("Reserving MSI range failed\n");
2747 /* HT memory range */
2748 val
= reserve_iova(&reserved_iova_ranges
,
2749 IOVA_PFN(HT_RANGE_START
), IOVA_PFN(HT_RANGE_END
));
2751 pr_err("Reserving HT range failed\n");
2756 * Memory used for PCI resources
2757 * FIXME: Check whether we can reserve the PCI-hole completly
2759 for_each_pci_dev(pdev
) {
2762 for (i
= 0; i
< PCI_NUM_RESOURCES
; ++i
) {
2763 struct resource
*r
= &pdev
->resource
[i
];
2765 if (!(r
->flags
& IORESOURCE_MEM
))
2768 val
= reserve_iova(&reserved_iova_ranges
,
2772 pci_err(pdev
, "Reserve pci-resource range %pR failed\n", r
);
2781 int __init
amd_iommu_init_api(void)
2785 ret
= iova_cache_get();
2789 ret
= init_reserved_iova_ranges();
2793 err
= bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2796 #ifdef CONFIG_ARM_AMBA
2797 err
= bus_set_iommu(&amba_bustype
, &amd_iommu_ops
);
2801 err
= bus_set_iommu(&platform_bus_type
, &amd_iommu_ops
);
2808 int __init
amd_iommu_init_dma_ops(void)
2810 swiotlb
= (iommu_pass_through
|| sme_me_mask
) ? 1 : 0;
2813 if (amd_iommu_unmap_flush
)
2814 pr_info("IO/TLB flush on unmap enabled\n");
2816 pr_info("Lazy IO/TLB flushing enabled\n");
2822 /*****************************************************************************
2824 * The following functions belong to the exported interface of AMD IOMMU
2826 * This interface allows access to lower level functions of the IOMMU
2827 * like protection domain handling and assignement of devices to domains
2828 * which is not possible with the dma_ops interface.
2830 *****************************************************************************/
2832 static void cleanup_domain(struct protection_domain
*domain
)
2834 struct iommu_dev_data
*entry
;
2835 unsigned long flags
;
2837 spin_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2839 while (!list_empty(&domain
->dev_list
)) {
2840 entry
= list_first_entry(&domain
->dev_list
,
2841 struct iommu_dev_data
, list
);
2842 BUG_ON(!entry
->domain
);
2843 __detach_device(entry
);
2846 spin_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2849 static void protection_domain_free(struct protection_domain
*domain
)
2855 domain_id_free(domain
->id
);
2860 static int protection_domain_init(struct protection_domain
*domain
)
2862 spin_lock_init(&domain
->lock
);
2863 mutex_init(&domain
->api_lock
);
2864 domain
->id
= domain_id_alloc();
2867 INIT_LIST_HEAD(&domain
->dev_list
);
2872 static struct protection_domain
*protection_domain_alloc(void)
2874 struct protection_domain
*domain
;
2876 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2880 if (protection_domain_init(domain
))
2891 static struct iommu_domain
*amd_iommu_domain_alloc(unsigned type
)
2893 struct protection_domain
*pdomain
;
2894 struct dma_ops_domain
*dma_domain
;
2897 case IOMMU_DOMAIN_UNMANAGED
:
2898 pdomain
= protection_domain_alloc();
2902 pdomain
->mode
= PAGE_MODE_3_LEVEL
;
2903 pdomain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2904 if (!pdomain
->pt_root
) {
2905 protection_domain_free(pdomain
);
2909 pdomain
->domain
.geometry
.aperture_start
= 0;
2910 pdomain
->domain
.geometry
.aperture_end
= ~0ULL;
2911 pdomain
->domain
.geometry
.force_aperture
= true;
2914 case IOMMU_DOMAIN_DMA
:
2915 dma_domain
= dma_ops_domain_alloc();
2917 pr_err("Failed to allocate\n");
2920 pdomain
= &dma_domain
->domain
;
2922 case IOMMU_DOMAIN_IDENTITY
:
2923 pdomain
= protection_domain_alloc();
2927 pdomain
->mode
= PAGE_MODE_NONE
;
2933 return &pdomain
->domain
;
2936 static void amd_iommu_domain_free(struct iommu_domain
*dom
)
2938 struct protection_domain
*domain
;
2939 struct dma_ops_domain
*dma_dom
;
2941 domain
= to_pdomain(dom
);
2943 if (domain
->dev_cnt
> 0)
2944 cleanup_domain(domain
);
2946 BUG_ON(domain
->dev_cnt
!= 0);
2951 switch (dom
->type
) {
2952 case IOMMU_DOMAIN_DMA
:
2953 /* Now release the domain */
2954 dma_dom
= to_dma_ops_domain(domain
);
2955 dma_ops_domain_free(dma_dom
);
2958 if (domain
->mode
!= PAGE_MODE_NONE
)
2959 free_pagetable(domain
);
2961 if (domain
->flags
& PD_IOMMUV2_MASK
)
2962 free_gcr3_table(domain
);
2964 protection_domain_free(domain
);
2969 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2972 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
2973 struct amd_iommu
*iommu
;
2976 if (!check_device(dev
))
2979 devid
= get_device_id(dev
);
2983 if (dev_data
->domain
!= NULL
)
2986 iommu
= amd_iommu_rlookup_table
[devid
];
2990 #ifdef CONFIG_IRQ_REMAP
2991 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
) &&
2992 (dom
->type
== IOMMU_DOMAIN_UNMANAGED
))
2993 dev_data
->use_vapic
= 0;
2996 iommu_completion_wait(iommu
);
2999 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
3002 struct protection_domain
*domain
= to_pdomain(dom
);
3003 struct iommu_dev_data
*dev_data
;
3004 struct amd_iommu
*iommu
;
3007 if (!check_device(dev
))
3010 dev_data
= dev
->archdata
.iommu
;
3012 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3016 if (dev_data
->domain
)
3019 ret
= attach_device(dev
, domain
);
3021 #ifdef CONFIG_IRQ_REMAP
3022 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
)) {
3023 if (dom
->type
== IOMMU_DOMAIN_UNMANAGED
)
3024 dev_data
->use_vapic
= 1;
3026 dev_data
->use_vapic
= 0;
3030 iommu_completion_wait(iommu
);
3035 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
3036 phys_addr_t paddr
, size_t page_size
, int iommu_prot
)
3038 struct protection_domain
*domain
= to_pdomain(dom
);
3042 if (domain
->mode
== PAGE_MODE_NONE
)
3045 if (iommu_prot
& IOMMU_READ
)
3046 prot
|= IOMMU_PROT_IR
;
3047 if (iommu_prot
& IOMMU_WRITE
)
3048 prot
|= IOMMU_PROT_IW
;
3050 mutex_lock(&domain
->api_lock
);
3051 ret
= iommu_map_page(domain
, iova
, paddr
, page_size
, prot
, GFP_KERNEL
);
3052 mutex_unlock(&domain
->api_lock
);
3057 static size_t amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
3060 struct protection_domain
*domain
= to_pdomain(dom
);
3063 if (domain
->mode
== PAGE_MODE_NONE
)
3066 mutex_lock(&domain
->api_lock
);
3067 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3068 mutex_unlock(&domain
->api_lock
);
3073 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3076 struct protection_domain
*domain
= to_pdomain(dom
);
3077 unsigned long offset_mask
, pte_pgsize
;
3080 if (domain
->mode
== PAGE_MODE_NONE
)
3083 pte
= fetch_pte(domain
, iova
, &pte_pgsize
);
3085 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3088 offset_mask
= pte_pgsize
- 1;
3089 __pte
= __sme_clr(*pte
& PM_ADDR_MASK
);
3091 return (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3094 static bool amd_iommu_capable(enum iommu_cap cap
)
3097 case IOMMU_CAP_CACHE_COHERENCY
:
3099 case IOMMU_CAP_INTR_REMAP
:
3100 return (irq_remapping_enabled
== 1);
3101 case IOMMU_CAP_NOEXEC
:
3110 static void amd_iommu_get_resv_regions(struct device
*dev
,
3111 struct list_head
*head
)
3113 struct iommu_resv_region
*region
;
3114 struct unity_map_entry
*entry
;
3117 devid
= get_device_id(dev
);
3121 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
3125 if (devid
< entry
->devid_start
|| devid
> entry
->devid_end
)
3128 type
= IOMMU_RESV_DIRECT
;
3129 length
= entry
->address_end
- entry
->address_start
;
3130 if (entry
->prot
& IOMMU_PROT_IR
)
3132 if (entry
->prot
& IOMMU_PROT_IW
)
3133 prot
|= IOMMU_WRITE
;
3134 if (entry
->prot
& IOMMU_UNITY_MAP_FLAG_EXCL_RANGE
)
3135 /* Exclusion range */
3136 type
= IOMMU_RESV_RESERVED
;
3138 region
= iommu_alloc_resv_region(entry
->address_start
,
3139 length
, prot
, type
);
3141 dev_err(dev
, "Out of memory allocating dm-regions\n");
3144 list_add_tail(®ion
->list
, head
);
3147 region
= iommu_alloc_resv_region(MSI_RANGE_START
,
3148 MSI_RANGE_END
- MSI_RANGE_START
+ 1,
3152 list_add_tail(®ion
->list
, head
);
3154 region
= iommu_alloc_resv_region(HT_RANGE_START
,
3155 HT_RANGE_END
- HT_RANGE_START
+ 1,
3156 0, IOMMU_RESV_RESERVED
);
3159 list_add_tail(®ion
->list
, head
);
3162 static void amd_iommu_put_resv_regions(struct device
*dev
,
3163 struct list_head
*head
)
3165 struct iommu_resv_region
*entry
, *next
;
3167 list_for_each_entry_safe(entry
, next
, head
, list
)
3171 static void amd_iommu_apply_resv_region(struct device
*dev
,
3172 struct iommu_domain
*domain
,
3173 struct iommu_resv_region
*region
)
3175 struct dma_ops_domain
*dma_dom
= to_dma_ops_domain(to_pdomain(domain
));
3176 unsigned long start
, end
;
3178 start
= IOVA_PFN(region
->start
);
3179 end
= IOVA_PFN(region
->start
+ region
->length
- 1);
3181 WARN_ON_ONCE(reserve_iova(&dma_dom
->iovad
, start
, end
) == NULL
);
3184 static bool amd_iommu_is_attach_deferred(struct iommu_domain
*domain
,
3187 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
3188 return dev_data
->defer_attach
;
3191 static void amd_iommu_flush_iotlb_all(struct iommu_domain
*domain
)
3193 struct protection_domain
*dom
= to_pdomain(domain
);
3195 domain_flush_tlb_pde(dom
);
3196 domain_flush_complete(dom
);
3199 static void amd_iommu_iotlb_range_add(struct iommu_domain
*domain
,
3200 unsigned long iova
, size_t size
)
3204 const struct iommu_ops amd_iommu_ops
= {
3205 .capable
= amd_iommu_capable
,
3206 .domain_alloc
= amd_iommu_domain_alloc
,
3207 .domain_free
= amd_iommu_domain_free
,
3208 .attach_dev
= amd_iommu_attach_device
,
3209 .detach_dev
= amd_iommu_detach_device
,
3210 .map
= amd_iommu_map
,
3211 .unmap
= amd_iommu_unmap
,
3212 .iova_to_phys
= amd_iommu_iova_to_phys
,
3213 .add_device
= amd_iommu_add_device
,
3214 .remove_device
= amd_iommu_remove_device
,
3215 .device_group
= amd_iommu_device_group
,
3216 .get_resv_regions
= amd_iommu_get_resv_regions
,
3217 .put_resv_regions
= amd_iommu_put_resv_regions
,
3218 .apply_resv_region
= amd_iommu_apply_resv_region
,
3219 .is_attach_deferred
= amd_iommu_is_attach_deferred
,
3220 .pgsize_bitmap
= AMD_IOMMU_PGSIZES
,
3221 .flush_iotlb_all
= amd_iommu_flush_iotlb_all
,
3222 .iotlb_range_add
= amd_iommu_iotlb_range_add
,
3223 .iotlb_sync
= amd_iommu_flush_iotlb_all
,
3226 /*****************************************************************************
3228 * The next functions do a basic initialization of IOMMU for pass through
3231 * In passthrough mode the IOMMU is initialized and enabled but not used for
3232 * DMA-API translation.
3234 *****************************************************************************/
3236 /* IOMMUv2 specific functions */
3237 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3239 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3241 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3243 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3245 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3247 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3249 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3251 struct protection_domain
*domain
= to_pdomain(dom
);
3252 unsigned long flags
;
3254 spin_lock_irqsave(&domain
->lock
, flags
);
3256 /* Update data structure */
3257 domain
->mode
= PAGE_MODE_NONE
;
3258 domain
->updated
= true;
3260 /* Make changes visible to IOMMUs */
3261 update_domain(domain
);
3263 /* Page-table is not visible to IOMMU anymore, so free it */
3264 free_pagetable(domain
);
3266 spin_unlock_irqrestore(&domain
->lock
, flags
);
3268 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3270 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3272 struct protection_domain
*domain
= to_pdomain(dom
);
3273 unsigned long flags
;
3276 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3279 /* Number of GCR3 table levels required */
3280 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3283 if (levels
> amd_iommu_max_glx_val
)
3286 spin_lock_irqsave(&domain
->lock
, flags
);
3289 * Save us all sanity checks whether devices already in the
3290 * domain support IOMMUv2. Just force that the domain has no
3291 * devices attached when it is switched into IOMMUv2 mode.
3294 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3298 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3299 if (domain
->gcr3_tbl
== NULL
)
3302 domain
->glx
= levels
;
3303 domain
->flags
|= PD_IOMMUV2_MASK
;
3304 domain
->updated
= true;
3306 update_domain(domain
);
3311 spin_unlock_irqrestore(&domain
->lock
, flags
);
3315 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3317 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3318 u64 address
, bool size
)
3320 struct iommu_dev_data
*dev_data
;
3321 struct iommu_cmd cmd
;
3324 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3327 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3330 * IOMMU TLB needs to be flushed before Device TLB to
3331 * prevent device TLB refill from IOMMU TLB
3333 for (i
= 0; i
< amd_iommu_get_num_iommus(); ++i
) {
3334 if (domain
->dev_iommu
[i
] == 0)
3337 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3342 /* Wait until IOMMU TLB flushes are complete */
3343 domain_flush_complete(domain
);
3345 /* Now flush device TLBs */
3346 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3347 struct amd_iommu
*iommu
;
3351 There might be non-IOMMUv2 capable devices in an IOMMUv2
3354 if (!dev_data
->ats
.enabled
)
3357 qdep
= dev_data
->ats
.qdep
;
3358 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3360 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3361 qdep
, address
, size
);
3363 ret
= iommu_queue_command(iommu
, &cmd
);
3368 /* Wait until all device TLBs are flushed */
3369 domain_flush_complete(domain
);
3378 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3381 return __flush_pasid(domain
, pasid
, address
, false);
3384 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3387 struct protection_domain
*domain
= to_pdomain(dom
);
3388 unsigned long flags
;
3391 spin_lock_irqsave(&domain
->lock
, flags
);
3392 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3393 spin_unlock_irqrestore(&domain
->lock
, flags
);
3397 EXPORT_SYMBOL(amd_iommu_flush_page
);
3399 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3401 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3405 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3407 struct protection_domain
*domain
= to_pdomain(dom
);
3408 unsigned long flags
;
3411 spin_lock_irqsave(&domain
->lock
, flags
);
3412 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3413 spin_unlock_irqrestore(&domain
->lock
, flags
);
3417 EXPORT_SYMBOL(amd_iommu_flush_tlb
);
3419 static u64
*__get_gcr3_pte(u64
*root
, int level
, int pasid
, bool alloc
)
3426 index
= (pasid
>> (9 * level
)) & 0x1ff;
3432 if (!(*pte
& GCR3_VALID
)) {
3436 root
= (void *)get_zeroed_page(GFP_ATOMIC
);
3440 *pte
= iommu_virt_to_phys(root
) | GCR3_VALID
;
3443 root
= iommu_phys_to_virt(*pte
& PAGE_MASK
);
3451 static int __set_gcr3(struct protection_domain
*domain
, int pasid
,
3456 if (domain
->mode
!= PAGE_MODE_NONE
)
3459 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, true);
3463 *pte
= (cr3
& PAGE_MASK
) | GCR3_VALID
;
3465 return __amd_iommu_flush_tlb(domain
, pasid
);
3468 static int __clear_gcr3(struct protection_domain
*domain
, int pasid
)
3472 if (domain
->mode
!= PAGE_MODE_NONE
)
3475 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, false);
3481 return __amd_iommu_flush_tlb(domain
, pasid
);
3484 int amd_iommu_domain_set_gcr3(struct iommu_domain
*dom
, int pasid
,
3487 struct protection_domain
*domain
= to_pdomain(dom
);
3488 unsigned long flags
;
3491 spin_lock_irqsave(&domain
->lock
, flags
);
3492 ret
= __set_gcr3(domain
, pasid
, cr3
);
3493 spin_unlock_irqrestore(&domain
->lock
, flags
);
3497 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3
);
3499 int amd_iommu_domain_clear_gcr3(struct iommu_domain
*dom
, int pasid
)
3501 struct protection_domain
*domain
= to_pdomain(dom
);
3502 unsigned long flags
;
3505 spin_lock_irqsave(&domain
->lock
, flags
);
3506 ret
= __clear_gcr3(domain
, pasid
);
3507 spin_unlock_irqrestore(&domain
->lock
, flags
);
3511 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3
);
3513 int amd_iommu_complete_ppr(struct pci_dev
*pdev
, int pasid
,
3514 int status
, int tag
)
3516 struct iommu_dev_data
*dev_data
;
3517 struct amd_iommu
*iommu
;
3518 struct iommu_cmd cmd
;
3520 dev_data
= get_dev_data(&pdev
->dev
);
3521 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3523 build_complete_ppr(&cmd
, dev_data
->devid
, pasid
, status
,
3524 tag
, dev_data
->pri_tlp
);
3526 return iommu_queue_command(iommu
, &cmd
);
3528 EXPORT_SYMBOL(amd_iommu_complete_ppr
);
3530 struct iommu_domain
*amd_iommu_get_v2_domain(struct pci_dev
*pdev
)
3532 struct protection_domain
*pdomain
;
3534 pdomain
= get_domain(&pdev
->dev
);
3535 if (IS_ERR(pdomain
))
3538 /* Only return IOMMUv2 domains */
3539 if (!(pdomain
->flags
& PD_IOMMUV2_MASK
))
3542 return &pdomain
->domain
;
3544 EXPORT_SYMBOL(amd_iommu_get_v2_domain
);
3546 void amd_iommu_enable_device_erratum(struct pci_dev
*pdev
, u32 erratum
)
3548 struct iommu_dev_data
*dev_data
;
3550 if (!amd_iommu_v2_supported())
3553 dev_data
= get_dev_data(&pdev
->dev
);
3554 dev_data
->errata
|= (1 << erratum
);
3556 EXPORT_SYMBOL(amd_iommu_enable_device_erratum
);
3558 int amd_iommu_device_info(struct pci_dev
*pdev
,
3559 struct amd_iommu_device_info
*info
)
3564 if (pdev
== NULL
|| info
== NULL
)
3567 if (!amd_iommu_v2_supported())
3570 memset(info
, 0, sizeof(*info
));
3572 if (!pci_ats_disabled()) {
3573 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ATS
);
3575 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_ATS_SUP
;
3578 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
3580 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRI_SUP
;
3582 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PASID
);
3586 max_pasids
= 1 << (9 * (amd_iommu_max_glx_val
+ 1));
3587 max_pasids
= min(max_pasids
, (1 << 20));
3589 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PASID_SUP
;
3590 info
->max_pasids
= min(pci_max_pasids(pdev
), max_pasids
);
3592 features
= pci_pasid_features(pdev
);
3593 if (features
& PCI_PASID_CAP_EXEC
)
3594 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP
;
3595 if (features
& PCI_PASID_CAP_PRIV
)
3596 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP
;
3601 EXPORT_SYMBOL(amd_iommu_device_info
);
3603 #ifdef CONFIG_IRQ_REMAP
3605 /*****************************************************************************
3607 * Interrupt Remapping Implementation
3609 *****************************************************************************/
3611 static struct irq_chip amd_ir_chip
;
3612 static DEFINE_SPINLOCK(iommu_table_lock
);
3614 static void set_dte_irq_entry(u16 devid
, struct irq_remap_table
*table
)
3618 dte
= amd_iommu_dev_table
[devid
].data
[2];
3619 dte
&= ~DTE_IRQ_PHYS_ADDR_MASK
;
3620 dte
|= iommu_virt_to_phys(table
->table
);
3621 dte
|= DTE_IRQ_REMAP_INTCTL
;
3622 dte
|= DTE_IRQ_TABLE_LEN
;
3623 dte
|= DTE_IRQ_REMAP_ENABLE
;
3625 amd_iommu_dev_table
[devid
].data
[2] = dte
;
3628 static struct irq_remap_table
*get_irq_table(u16 devid
)
3630 struct irq_remap_table
*table
;
3632 if (WARN_ONCE(!amd_iommu_rlookup_table
[devid
],
3633 "%s: no iommu for devid %x\n", __func__
, devid
))
3636 table
= irq_lookup_table
[devid
];
3637 if (WARN_ONCE(!table
, "%s: no table for devid %x\n", __func__
, devid
))
3643 static struct irq_remap_table
*__alloc_irq_table(void)
3645 struct irq_remap_table
*table
;
3647 table
= kzalloc(sizeof(*table
), GFP_KERNEL
);
3651 table
->table
= kmem_cache_alloc(amd_iommu_irq_cache
, GFP_KERNEL
);
3652 if (!table
->table
) {
3656 raw_spin_lock_init(&table
->lock
);
3658 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir
))
3659 memset(table
->table
, 0,
3660 MAX_IRQS_PER_TABLE
* sizeof(u32
));
3662 memset(table
->table
, 0,
3663 (MAX_IRQS_PER_TABLE
* (sizeof(u64
) * 2)));
3667 static void set_remap_table_entry(struct amd_iommu
*iommu
, u16 devid
,
3668 struct irq_remap_table
*table
)
3670 irq_lookup_table
[devid
] = table
;
3671 set_dte_irq_entry(devid
, table
);
3672 iommu_flush_dte(iommu
, devid
);
3675 static struct irq_remap_table
*alloc_irq_table(u16 devid
)
3677 struct irq_remap_table
*table
= NULL
;
3678 struct irq_remap_table
*new_table
= NULL
;
3679 struct amd_iommu
*iommu
;
3680 unsigned long flags
;
3683 spin_lock_irqsave(&iommu_table_lock
, flags
);
3685 iommu
= amd_iommu_rlookup_table
[devid
];
3689 table
= irq_lookup_table
[devid
];
3693 alias
= amd_iommu_alias_table
[devid
];
3694 table
= irq_lookup_table
[alias
];
3696 set_remap_table_entry(iommu
, devid
, table
);
3699 spin_unlock_irqrestore(&iommu_table_lock
, flags
);
3701 /* Nothing there yet, allocate new irq remapping table */
3702 new_table
= __alloc_irq_table();
3706 spin_lock_irqsave(&iommu_table_lock
, flags
);
3708 table
= irq_lookup_table
[devid
];
3712 table
= irq_lookup_table
[alias
];
3714 set_remap_table_entry(iommu
, devid
, table
);
3721 set_remap_table_entry(iommu
, devid
, table
);
3723 set_remap_table_entry(iommu
, alias
, table
);
3726 iommu_completion_wait(iommu
);
3729 spin_unlock_irqrestore(&iommu_table_lock
, flags
);
3732 kmem_cache_free(amd_iommu_irq_cache
, new_table
->table
);
3738 static int alloc_irq_index(u16 devid
, int count
, bool align
)
3740 struct irq_remap_table
*table
;
3741 int index
, c
, alignment
= 1;
3742 unsigned long flags
;
3743 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
3748 table
= alloc_irq_table(devid
);
3753 alignment
= roundup_pow_of_two(count
);
3755 raw_spin_lock_irqsave(&table
->lock
, flags
);
3757 /* Scan table for free entries */
3758 for (index
= ALIGN(table
->min_index
, alignment
), c
= 0;
3759 index
< MAX_IRQS_PER_TABLE
;) {
3760 if (!iommu
->irte_ops
->is_allocated(table
, index
)) {
3764 index
= ALIGN(index
+ 1, alignment
);
3770 iommu
->irte_ops
->set_allocated(table
, index
- c
+ 1);
3782 raw_spin_unlock_irqrestore(&table
->lock
, flags
);
3787 static int modify_irte_ga(u16 devid
, int index
, struct irte_ga
*irte
,
3788 struct amd_ir_data
*data
)
3790 struct irq_remap_table
*table
;
3791 struct amd_iommu
*iommu
;
3792 unsigned long flags
;
3793 struct irte_ga
*entry
;
3795 iommu
= amd_iommu_rlookup_table
[devid
];
3799 table
= get_irq_table(devid
);
3803 raw_spin_lock_irqsave(&table
->lock
, flags
);
3805 entry
= (struct irte_ga
*)table
->table
;
3806 entry
= &entry
[index
];
3807 entry
->lo
.fields_remap
.valid
= 0;
3808 entry
->hi
.val
= irte
->hi
.val
;
3809 entry
->lo
.val
= irte
->lo
.val
;
3810 entry
->lo
.fields_remap
.valid
= 1;
3814 raw_spin_unlock_irqrestore(&table
->lock
, flags
);
3816 iommu_flush_irt(iommu
, devid
);
3817 iommu_completion_wait(iommu
);
3822 static int modify_irte(u16 devid
, int index
, union irte
*irte
)
3824 struct irq_remap_table
*table
;
3825 struct amd_iommu
*iommu
;
3826 unsigned long flags
;
3828 iommu
= amd_iommu_rlookup_table
[devid
];
3832 table
= get_irq_table(devid
);
3836 raw_spin_lock_irqsave(&table
->lock
, flags
);
3837 table
->table
[index
] = irte
->val
;
3838 raw_spin_unlock_irqrestore(&table
->lock
, flags
);
3840 iommu_flush_irt(iommu
, devid
);
3841 iommu_completion_wait(iommu
);
3846 static void free_irte(u16 devid
, int index
)
3848 struct irq_remap_table
*table
;
3849 struct amd_iommu
*iommu
;
3850 unsigned long flags
;
3852 iommu
= amd_iommu_rlookup_table
[devid
];
3856 table
= get_irq_table(devid
);
3860 raw_spin_lock_irqsave(&table
->lock
, flags
);
3861 iommu
->irte_ops
->clear_allocated(table
, index
);
3862 raw_spin_unlock_irqrestore(&table
->lock
, flags
);
3864 iommu_flush_irt(iommu
, devid
);
3865 iommu_completion_wait(iommu
);
3868 static void irte_prepare(void *entry
,
3869 u32 delivery_mode
, u32 dest_mode
,
3870 u8 vector
, u32 dest_apicid
, int devid
)
3872 union irte
*irte
= (union irte
*) entry
;
3875 irte
->fields
.vector
= vector
;
3876 irte
->fields
.int_type
= delivery_mode
;
3877 irte
->fields
.destination
= dest_apicid
;
3878 irte
->fields
.dm
= dest_mode
;
3879 irte
->fields
.valid
= 1;
3882 static void irte_ga_prepare(void *entry
,
3883 u32 delivery_mode
, u32 dest_mode
,
3884 u8 vector
, u32 dest_apicid
, int devid
)
3886 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3890 irte
->lo
.fields_remap
.int_type
= delivery_mode
;
3891 irte
->lo
.fields_remap
.dm
= dest_mode
;
3892 irte
->hi
.fields
.vector
= vector
;
3893 irte
->lo
.fields_remap
.destination
= APICID_TO_IRTE_DEST_LO(dest_apicid
);
3894 irte
->hi
.fields
.destination
= APICID_TO_IRTE_DEST_HI(dest_apicid
);
3895 irte
->lo
.fields_remap
.valid
= 1;
3898 static void irte_activate(void *entry
, u16 devid
, u16 index
)
3900 union irte
*irte
= (union irte
*) entry
;
3902 irte
->fields
.valid
= 1;
3903 modify_irte(devid
, index
, irte
);
3906 static void irte_ga_activate(void *entry
, u16 devid
, u16 index
)
3908 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3910 irte
->lo
.fields_remap
.valid
= 1;
3911 modify_irte_ga(devid
, index
, irte
, NULL
);
3914 static void irte_deactivate(void *entry
, u16 devid
, u16 index
)
3916 union irte
*irte
= (union irte
*) entry
;
3918 irte
->fields
.valid
= 0;
3919 modify_irte(devid
, index
, irte
);
3922 static void irte_ga_deactivate(void *entry
, u16 devid
, u16 index
)
3924 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3926 irte
->lo
.fields_remap
.valid
= 0;
3927 modify_irte_ga(devid
, index
, irte
, NULL
);
3930 static void irte_set_affinity(void *entry
, u16 devid
, u16 index
,
3931 u8 vector
, u32 dest_apicid
)
3933 union irte
*irte
= (union irte
*) entry
;
3935 irte
->fields
.vector
= vector
;
3936 irte
->fields
.destination
= dest_apicid
;
3937 modify_irte(devid
, index
, irte
);
3940 static void irte_ga_set_affinity(void *entry
, u16 devid
, u16 index
,
3941 u8 vector
, u32 dest_apicid
)
3943 struct irte_ga
*irte
= (struct irte_ga
*) entry
;
3945 if (!irte
->lo
.fields_remap
.guest_mode
) {
3946 irte
->hi
.fields
.vector
= vector
;
3947 irte
->lo
.fields_remap
.destination
=
3948 APICID_TO_IRTE_DEST_LO(dest_apicid
);
3949 irte
->hi
.fields
.destination
=
3950 APICID_TO_IRTE_DEST_HI(dest_apicid
);
3951 modify_irte_ga(devid
, index
, irte
, NULL
);
3955 #define IRTE_ALLOCATED (~1U)
3956 static void irte_set_allocated(struct irq_remap_table
*table
, int index
)
3958 table
->table
[index
] = IRTE_ALLOCATED
;
3961 static void irte_ga_set_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
));
3968 irte
->hi
.fields
.vector
= 0xff;
3971 static bool irte_is_allocated(struct irq_remap_table
*table
, int index
)
3973 union irte
*ptr
= (union irte
*)table
->table
;
3974 union irte
*irte
= &ptr
[index
];
3976 return irte
->val
!= 0;
3979 static bool irte_ga_is_allocated(struct irq_remap_table
*table
, int index
)
3981 struct irte_ga
*ptr
= (struct irte_ga
*)table
->table
;
3982 struct irte_ga
*irte
= &ptr
[index
];
3984 return irte
->hi
.fields
.vector
!= 0;
3987 static void irte_clear_allocated(struct irq_remap_table
*table
, int index
)
3989 table
->table
[index
] = 0;
3992 static void irte_ga_clear_allocated(struct irq_remap_table
*table
, int index
)
3994 struct irte_ga
*ptr
= (struct irte_ga
*)table
->table
;
3995 struct irte_ga
*irte
= &ptr
[index
];
3997 memset(&irte
->lo
.val
, 0, sizeof(u64
));
3998 memset(&irte
->hi
.val
, 0, sizeof(u64
));
4001 static int get_devid(struct irq_alloc_info
*info
)
4005 switch (info
->type
) {
4006 case X86_IRQ_ALLOC_TYPE_IOAPIC
:
4007 devid
= get_ioapic_devid(info
->ioapic_id
);
4009 case X86_IRQ_ALLOC_TYPE_HPET
:
4010 devid
= get_hpet_devid(info
->hpet_id
);
4012 case X86_IRQ_ALLOC_TYPE_MSI
:
4013 case X86_IRQ_ALLOC_TYPE_MSIX
:
4014 devid
= get_device_id(&info
->msi_dev
->dev
);
4024 static struct irq_domain
*get_ir_irq_domain(struct irq_alloc_info
*info
)
4026 struct amd_iommu
*iommu
;
4032 devid
= get_devid(info
);
4034 iommu
= amd_iommu_rlookup_table
[devid
];
4036 return iommu
->ir_domain
;
4042 static struct irq_domain
*get_irq_domain(struct irq_alloc_info
*info
)
4044 struct amd_iommu
*iommu
;
4050 switch (info
->type
) {
4051 case X86_IRQ_ALLOC_TYPE_MSI
:
4052 case X86_IRQ_ALLOC_TYPE_MSIX
:
4053 devid
= get_device_id(&info
->msi_dev
->dev
);
4057 iommu
= amd_iommu_rlookup_table
[devid
];
4059 return iommu
->msi_domain
;
4068 struct irq_remap_ops amd_iommu_irq_ops
= {
4069 .prepare
= amd_iommu_prepare
,
4070 .enable
= amd_iommu_enable
,
4071 .disable
= amd_iommu_disable
,
4072 .reenable
= amd_iommu_reenable
,
4073 .enable_faulting
= amd_iommu_enable_faulting
,
4074 .get_ir_irq_domain
= get_ir_irq_domain
,
4075 .get_irq_domain
= get_irq_domain
,
4078 static void irq_remapping_prepare_irte(struct amd_ir_data
*data
,
4079 struct irq_cfg
*irq_cfg
,
4080 struct irq_alloc_info
*info
,
4081 int devid
, int index
, int sub_handle
)
4083 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4084 struct msi_msg
*msg
= &data
->msi_entry
;
4085 struct IO_APIC_route_entry
*entry
;
4086 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
4091 data
->irq_2_irte
.devid
= devid
;
4092 data
->irq_2_irte
.index
= index
+ sub_handle
;
4093 iommu
->irte_ops
->prepare(data
->entry
, apic
->irq_delivery_mode
,
4094 apic
->irq_dest_mode
, irq_cfg
->vector
,
4095 irq_cfg
->dest_apicid
, devid
);
4097 switch (info
->type
) {
4098 case X86_IRQ_ALLOC_TYPE_IOAPIC
:
4099 /* Setup IOAPIC entry */
4100 entry
= info
->ioapic_entry
;
4101 info
->ioapic_entry
= NULL
;
4102 memset(entry
, 0, sizeof(*entry
));
4103 entry
->vector
= index
;
4105 entry
->trigger
= info
->ioapic_trigger
;
4106 entry
->polarity
= info
->ioapic_polarity
;
4107 /* Mask level triggered irqs. */
4108 if (info
->ioapic_trigger
)
4112 case X86_IRQ_ALLOC_TYPE_HPET
:
4113 case X86_IRQ_ALLOC_TYPE_MSI
:
4114 case X86_IRQ_ALLOC_TYPE_MSIX
:
4115 msg
->address_hi
= MSI_ADDR_BASE_HI
;
4116 msg
->address_lo
= MSI_ADDR_BASE_LO
;
4117 msg
->data
= irte_info
->index
;
4126 struct amd_irte_ops irte_32_ops
= {
4127 .prepare
= irte_prepare
,
4128 .activate
= irte_activate
,
4129 .deactivate
= irte_deactivate
,
4130 .set_affinity
= irte_set_affinity
,
4131 .set_allocated
= irte_set_allocated
,
4132 .is_allocated
= irte_is_allocated
,
4133 .clear_allocated
= irte_clear_allocated
,
4136 struct amd_irte_ops irte_128_ops
= {
4137 .prepare
= irte_ga_prepare
,
4138 .activate
= irte_ga_activate
,
4139 .deactivate
= irte_ga_deactivate
,
4140 .set_affinity
= irte_ga_set_affinity
,
4141 .set_allocated
= irte_ga_set_allocated
,
4142 .is_allocated
= irte_ga_is_allocated
,
4143 .clear_allocated
= irte_ga_clear_allocated
,
4146 static int irq_remapping_alloc(struct irq_domain
*domain
, unsigned int virq
,
4147 unsigned int nr_irqs
, void *arg
)
4149 struct irq_alloc_info
*info
= arg
;
4150 struct irq_data
*irq_data
;
4151 struct amd_ir_data
*data
= NULL
;
4152 struct irq_cfg
*cfg
;
4158 if (nr_irqs
> 1 && info
->type
!= X86_IRQ_ALLOC_TYPE_MSI
&&
4159 info
->type
!= X86_IRQ_ALLOC_TYPE_MSIX
)
4163 * With IRQ remapping enabled, don't need contiguous CPU vectors
4164 * to support multiple MSI interrupts.
4166 if (info
->type
== X86_IRQ_ALLOC_TYPE_MSI
)
4167 info
->flags
&= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS
;
4169 devid
= get_devid(info
);
4173 ret
= irq_domain_alloc_irqs_parent(domain
, virq
, nr_irqs
, arg
);
4177 if (info
->type
== X86_IRQ_ALLOC_TYPE_IOAPIC
) {
4178 struct irq_remap_table
*table
;
4179 struct amd_iommu
*iommu
;
4181 table
= alloc_irq_table(devid
);
4183 if (!table
->min_index
) {
4185 * Keep the first 32 indexes free for IOAPIC
4188 table
->min_index
= 32;
4189 iommu
= amd_iommu_rlookup_table
[devid
];
4190 for (i
= 0; i
< 32; ++i
)
4191 iommu
->irte_ops
->set_allocated(table
, i
);
4193 WARN_ON(table
->min_index
!= 32);
4194 index
= info
->ioapic_pin
;
4199 bool align
= (info
->type
== X86_IRQ_ALLOC_TYPE_MSI
);
4201 index
= alloc_irq_index(devid
, nr_irqs
, align
);
4204 pr_warn("Failed to allocate IRTE\n");
4206 goto out_free_parent
;
4209 for (i
= 0; i
< nr_irqs
; i
++) {
4210 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4211 cfg
= irqd_cfg(irq_data
);
4212 if (!irq_data
|| !cfg
) {
4218 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
4222 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir
))
4223 data
->entry
= kzalloc(sizeof(union irte
), GFP_KERNEL
);
4225 data
->entry
= kzalloc(sizeof(struct irte_ga
),
4232 irq_data
->hwirq
= (devid
<< 16) + i
;
4233 irq_data
->chip_data
= data
;
4234 irq_data
->chip
= &amd_ir_chip
;
4235 irq_remapping_prepare_irte(data
, cfg
, info
, devid
, index
, i
);
4236 irq_set_status_flags(virq
+ i
, IRQ_MOVE_PCNTXT
);
4242 for (i
--; i
>= 0; i
--) {
4243 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4245 kfree(irq_data
->chip_data
);
4247 for (i
= 0; i
< nr_irqs
; i
++)
4248 free_irte(devid
, index
+ i
);
4250 irq_domain_free_irqs_common(domain
, virq
, nr_irqs
);
4254 static void irq_remapping_free(struct irq_domain
*domain
, unsigned int virq
,
4255 unsigned int nr_irqs
)
4257 struct irq_2_irte
*irte_info
;
4258 struct irq_data
*irq_data
;
4259 struct amd_ir_data
*data
;
4262 for (i
= 0; i
< nr_irqs
; i
++) {
4263 irq_data
= irq_domain_get_irq_data(domain
, virq
+ i
);
4264 if (irq_data
&& irq_data
->chip_data
) {
4265 data
= irq_data
->chip_data
;
4266 irte_info
= &data
->irq_2_irte
;
4267 free_irte(irte_info
->devid
, irte_info
->index
);
4272 irq_domain_free_irqs_common(domain
, virq
, nr_irqs
);
4275 static void amd_ir_update_irte(struct irq_data
*irqd
, struct amd_iommu
*iommu
,
4276 struct amd_ir_data
*ir_data
,
4277 struct irq_2_irte
*irte_info
,
4278 struct irq_cfg
*cfg
);
4280 static int irq_remapping_activate(struct irq_domain
*domain
,
4281 struct irq_data
*irq_data
, bool reserve
)
4283 struct amd_ir_data
*data
= irq_data
->chip_data
;
4284 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4285 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4286 struct irq_cfg
*cfg
= irqd_cfg(irq_data
);
4291 iommu
->irte_ops
->activate(data
->entry
, irte_info
->devid
,
4293 amd_ir_update_irte(irq_data
, iommu
, data
, irte_info
, cfg
);
4297 static void irq_remapping_deactivate(struct irq_domain
*domain
,
4298 struct irq_data
*irq_data
)
4300 struct amd_ir_data
*data
= irq_data
->chip_data
;
4301 struct irq_2_irte
*irte_info
= &data
->irq_2_irte
;
4302 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4305 iommu
->irte_ops
->deactivate(data
->entry
, irte_info
->devid
,
4309 static const struct irq_domain_ops amd_ir_domain_ops
= {
4310 .alloc
= irq_remapping_alloc
,
4311 .free
= irq_remapping_free
,
4312 .activate
= irq_remapping_activate
,
4313 .deactivate
= irq_remapping_deactivate
,
4316 static int amd_ir_set_vcpu_affinity(struct irq_data
*data
, void *vcpu_info
)
4318 struct amd_iommu
*iommu
;
4319 struct amd_iommu_pi_data
*pi_data
= vcpu_info
;
4320 struct vcpu_data
*vcpu_pi_info
= pi_data
->vcpu_data
;
4321 struct amd_ir_data
*ir_data
= data
->chip_data
;
4322 struct irte_ga
*irte
= (struct irte_ga
*) ir_data
->entry
;
4323 struct irq_2_irte
*irte_info
= &ir_data
->irq_2_irte
;
4324 struct iommu_dev_data
*dev_data
= search_dev_data(irte_info
->devid
);
4327 * This device has never been set up for guest mode.
4328 * we should not modify the IRTE
4330 if (!dev_data
|| !dev_data
->use_vapic
)
4333 pi_data
->ir_data
= ir_data
;
4336 * SVM tries to set up for VAPIC mode, but we are in
4337 * legacy mode. So, we force legacy mode instead.
4339 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
)) {
4340 pr_debug("%s: Fall back to using intr legacy remap\n",
4342 pi_data
->is_guest_mode
= false;
4345 iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4349 pi_data
->prev_ga_tag
= ir_data
->cached_ga_tag
;
4350 if (pi_data
->is_guest_mode
) {
4352 irte
->hi
.fields
.ga_root_ptr
= (pi_data
->base
>> 12);
4353 irte
->hi
.fields
.vector
= vcpu_pi_info
->vector
;
4354 irte
->lo
.fields_vapic
.ga_log_intr
= 1;
4355 irte
->lo
.fields_vapic
.guest_mode
= 1;
4356 irte
->lo
.fields_vapic
.ga_tag
= pi_data
->ga_tag
;
4358 ir_data
->cached_ga_tag
= pi_data
->ga_tag
;
4361 struct irq_cfg
*cfg
= irqd_cfg(data
);
4365 irte
->hi
.fields
.vector
= cfg
->vector
;
4366 irte
->lo
.fields_remap
.guest_mode
= 0;
4367 irte
->lo
.fields_remap
.destination
=
4368 APICID_TO_IRTE_DEST_LO(cfg
->dest_apicid
);
4369 irte
->hi
.fields
.destination
=
4370 APICID_TO_IRTE_DEST_HI(cfg
->dest_apicid
);
4371 irte
->lo
.fields_remap
.int_type
= apic
->irq_delivery_mode
;
4372 irte
->lo
.fields_remap
.dm
= apic
->irq_dest_mode
;
4375 * This communicates the ga_tag back to the caller
4376 * so that it can do all the necessary clean up.
4378 ir_data
->cached_ga_tag
= 0;
4381 return modify_irte_ga(irte_info
->devid
, irte_info
->index
, irte
, ir_data
);
4385 static void amd_ir_update_irte(struct irq_data
*irqd
, struct amd_iommu
*iommu
,
4386 struct amd_ir_data
*ir_data
,
4387 struct irq_2_irte
*irte_info
,
4388 struct irq_cfg
*cfg
)
4392 * Atomically updates the IRTE with the new destination, vector
4393 * and flushes the interrupt entry cache.
4395 iommu
->irte_ops
->set_affinity(ir_data
->entry
, irte_info
->devid
,
4396 irte_info
->index
, cfg
->vector
,
4400 static int amd_ir_set_affinity(struct irq_data
*data
,
4401 const struct cpumask
*mask
, bool force
)
4403 struct amd_ir_data
*ir_data
= data
->chip_data
;
4404 struct irq_2_irte
*irte_info
= &ir_data
->irq_2_irte
;
4405 struct irq_cfg
*cfg
= irqd_cfg(data
);
4406 struct irq_data
*parent
= data
->parent_data
;
4407 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[irte_info
->devid
];
4413 ret
= parent
->chip
->irq_set_affinity(parent
, mask
, force
);
4414 if (ret
< 0 || ret
== IRQ_SET_MASK_OK_DONE
)
4417 amd_ir_update_irte(data
, iommu
, ir_data
, irte_info
, cfg
);
4419 * After this point, all the interrupts will start arriving
4420 * at the new destination. So, time to cleanup the previous
4421 * vector allocation.
4423 send_cleanup_vector(cfg
);
4425 return IRQ_SET_MASK_OK_DONE
;
4428 static void ir_compose_msi_msg(struct irq_data
*irq_data
, struct msi_msg
*msg
)
4430 struct amd_ir_data
*ir_data
= irq_data
->chip_data
;
4432 *msg
= ir_data
->msi_entry
;
4435 static struct irq_chip amd_ir_chip
= {
4437 .irq_ack
= apic_ack_irq
,
4438 .irq_set_affinity
= amd_ir_set_affinity
,
4439 .irq_set_vcpu_affinity
= amd_ir_set_vcpu_affinity
,
4440 .irq_compose_msi_msg
= ir_compose_msi_msg
,
4443 int amd_iommu_create_irq_domain(struct amd_iommu
*iommu
)
4445 struct fwnode_handle
*fn
;
4447 fn
= irq_domain_alloc_named_id_fwnode("AMD-IR", iommu
->index
);
4450 iommu
->ir_domain
= irq_domain_create_tree(fn
, &amd_ir_domain_ops
, iommu
);
4451 irq_domain_free_fwnode(fn
);
4452 if (!iommu
->ir_domain
)
4455 iommu
->ir_domain
->parent
= arch_get_ir_parent_domain();
4456 iommu
->msi_domain
= arch_create_remap_msi_irq_domain(iommu
->ir_domain
,
4462 int amd_iommu_update_ga(int cpu
, bool is_run
, void *data
)
4464 unsigned long flags
;
4465 struct amd_iommu
*iommu
;
4466 struct irq_remap_table
*table
;
4467 struct amd_ir_data
*ir_data
= (struct amd_ir_data
*)data
;
4468 int devid
= ir_data
->irq_2_irte
.devid
;
4469 struct irte_ga
*entry
= (struct irte_ga
*) ir_data
->entry
;
4470 struct irte_ga
*ref
= (struct irte_ga
*) ir_data
->ref
;
4472 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir
) ||
4473 !ref
|| !entry
|| !entry
->lo
.fields_vapic
.guest_mode
)
4476 iommu
= amd_iommu_rlookup_table
[devid
];
4480 table
= get_irq_table(devid
);
4484 raw_spin_lock_irqsave(&table
->lock
, flags
);
4486 if (ref
->lo
.fields_vapic
.guest_mode
) {
4488 ref
->lo
.fields_vapic
.destination
=
4489 APICID_TO_IRTE_DEST_LO(cpu
);
4490 ref
->hi
.fields
.destination
=
4491 APICID_TO_IRTE_DEST_HI(cpu
);
4493 ref
->lo
.fields_vapic
.is_run
= is_run
;
4497 raw_spin_unlock_irqrestore(&table
->lock
, flags
);
4499 iommu_flush_irt(iommu
, devid
);
4500 iommu_completion_wait(iommu
);
4503 EXPORT_SYMBOL(amd_iommu_update_ga
);