2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/pci.h>
21 #include <linux/pci-ats.h>
22 #include <linux/bitmap.h>
23 #include <linux/slab.h>
24 #include <linux/debugfs.h>
25 #include <linux/scatterlist.h>
26 #include <linux/dma-mapping.h>
27 #include <linux/iommu-helper.h>
28 #include <linux/iommu.h>
29 #include <linux/delay.h>
30 #include <linux/amd-iommu.h>
31 #include <asm/msidef.h>
32 #include <asm/proto.h>
33 #include <asm/iommu.h>
37 #include "amd_iommu_proto.h"
38 #include "amd_iommu_types.h"
40 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
42 #define LOOP_TIMEOUT 100000
44 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
46 /* A list of preallocated protection domains */
47 static LIST_HEAD(iommu_pd_list
);
48 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
50 /* List of all available dev_data structures */
51 static LIST_HEAD(dev_data_list
);
52 static DEFINE_SPINLOCK(dev_data_list_lock
);
55 * Domain for untranslated devices - only allocated
56 * if iommu=pt passed on kernel cmd line.
58 static struct protection_domain
*pt_domain
;
60 static struct iommu_ops amd_iommu_ops
;
63 * general struct to manage commands send to an IOMMU
69 static void update_domain(struct protection_domain
*domain
);
70 static int __init
alloc_passthrough_domain(void);
72 /****************************************************************************
76 ****************************************************************************/
78 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
80 struct iommu_dev_data
*dev_data
;
83 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
87 dev_data
->devid
= devid
;
88 atomic_set(&dev_data
->bind
, 0);
90 spin_lock_irqsave(&dev_data_list_lock
, flags
);
91 list_add_tail(&dev_data
->dev_data_list
, &dev_data_list
);
92 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
97 static void free_dev_data(struct iommu_dev_data
*dev_data
)
101 spin_lock_irqsave(&dev_data_list_lock
, flags
);
102 list_del(&dev_data
->dev_data_list
);
103 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
108 static struct iommu_dev_data
*search_dev_data(u16 devid
)
110 struct iommu_dev_data
*dev_data
;
113 spin_lock_irqsave(&dev_data_list_lock
, flags
);
114 list_for_each_entry(dev_data
, &dev_data_list
, dev_data_list
) {
115 if (dev_data
->devid
== devid
)
122 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
127 static struct iommu_dev_data
*find_dev_data(u16 devid
)
129 struct iommu_dev_data
*dev_data
;
131 dev_data
= search_dev_data(devid
);
133 if (dev_data
== NULL
)
134 dev_data
= alloc_dev_data(devid
);
139 static inline u16
get_device_id(struct device
*dev
)
141 struct pci_dev
*pdev
= to_pci_dev(dev
);
143 return calc_devid(pdev
->bus
->number
, pdev
->devfn
);
146 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
148 return dev
->archdata
.iommu
;
151 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
153 static const int caps
[] = {
160 for (i
= 0; i
< 3; ++i
) {
161 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
170 * In this function the list of preallocated protection domains is traversed to
171 * find the domain for a specific device
173 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
175 struct dma_ops_domain
*entry
, *ret
= NULL
;
177 u16 alias
= amd_iommu_alias_table
[devid
];
179 if (list_empty(&iommu_pd_list
))
182 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
184 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
185 if (entry
->target_dev
== devid
||
186 entry
->target_dev
== alias
) {
192 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
198 * This function checks if the driver got a valid device from the caller to
199 * avoid dereferencing invalid pointers.
201 static bool check_device(struct device
*dev
)
205 if (!dev
|| !dev
->dma_mask
)
208 /* No device or no PCI device */
209 if (dev
->bus
!= &pci_bus_type
)
212 devid
= get_device_id(dev
);
214 /* Out of our scope? */
215 if (devid
> amd_iommu_last_bdf
)
218 if (amd_iommu_rlookup_table
[devid
] == NULL
)
224 static int iommu_init_device(struct device
*dev
)
226 struct pci_dev
*pdev
= to_pci_dev(dev
);
227 struct iommu_dev_data
*dev_data
;
230 if (dev
->archdata
.iommu
)
233 dev_data
= find_dev_data(get_device_id(dev
));
237 alias
= amd_iommu_alias_table
[dev_data
->devid
];
238 if (alias
!= dev_data
->devid
) {
239 struct iommu_dev_data
*alias_data
;
241 alias_data
= find_dev_data(alias
);
242 if (alias_data
== NULL
) {
243 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
245 free_dev_data(dev_data
);
248 dev_data
->alias_data
= alias_data
;
251 if (pci_iommuv2_capable(pdev
)) {
252 struct amd_iommu
*iommu
;
254 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
255 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
258 dev
->archdata
.iommu
= dev_data
;
263 static void iommu_ignore_device(struct device
*dev
)
267 devid
= get_device_id(dev
);
268 alias
= amd_iommu_alias_table
[devid
];
270 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
271 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
273 amd_iommu_rlookup_table
[devid
] = NULL
;
274 amd_iommu_rlookup_table
[alias
] = NULL
;
277 static void iommu_uninit_device(struct device
*dev
)
280 * Nothing to do here - we keep dev_data around for unplugged devices
281 * and reuse it when the device is re-plugged - not doing so would
282 * introduce a ton of races.
286 void __init
amd_iommu_uninit_devices(void)
288 struct iommu_dev_data
*dev_data
, *n
;
289 struct pci_dev
*pdev
= NULL
;
291 for_each_pci_dev(pdev
) {
293 if (!check_device(&pdev
->dev
))
296 iommu_uninit_device(&pdev
->dev
);
299 /* Free all of our dev_data structures */
300 list_for_each_entry_safe(dev_data
, n
, &dev_data_list
, dev_data_list
)
301 free_dev_data(dev_data
);
304 int __init
amd_iommu_init_devices(void)
306 struct pci_dev
*pdev
= NULL
;
309 for_each_pci_dev(pdev
) {
311 if (!check_device(&pdev
->dev
))
314 ret
= iommu_init_device(&pdev
->dev
);
315 if (ret
== -ENOTSUPP
)
316 iommu_ignore_device(&pdev
->dev
);
325 amd_iommu_uninit_devices();
329 #ifdef CONFIG_AMD_IOMMU_STATS
332 * Initialization code for statistics collection
335 DECLARE_STATS_COUNTER(compl_wait
);
336 DECLARE_STATS_COUNTER(cnt_map_single
);
337 DECLARE_STATS_COUNTER(cnt_unmap_single
);
338 DECLARE_STATS_COUNTER(cnt_map_sg
);
339 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
340 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
341 DECLARE_STATS_COUNTER(cnt_free_coherent
);
342 DECLARE_STATS_COUNTER(cross_page
);
343 DECLARE_STATS_COUNTER(domain_flush_single
);
344 DECLARE_STATS_COUNTER(domain_flush_all
);
345 DECLARE_STATS_COUNTER(alloced_io_mem
);
346 DECLARE_STATS_COUNTER(total_map_requests
);
348 static struct dentry
*stats_dir
;
349 static struct dentry
*de_fflush
;
351 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
353 if (stats_dir
== NULL
)
356 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
360 static void amd_iommu_stats_init(void)
362 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
363 if (stats_dir
== NULL
)
366 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
367 (u32
*)&amd_iommu_unmap_flush
);
369 amd_iommu_stats_add(&compl_wait
);
370 amd_iommu_stats_add(&cnt_map_single
);
371 amd_iommu_stats_add(&cnt_unmap_single
);
372 amd_iommu_stats_add(&cnt_map_sg
);
373 amd_iommu_stats_add(&cnt_unmap_sg
);
374 amd_iommu_stats_add(&cnt_alloc_coherent
);
375 amd_iommu_stats_add(&cnt_free_coherent
);
376 amd_iommu_stats_add(&cross_page
);
377 amd_iommu_stats_add(&domain_flush_single
);
378 amd_iommu_stats_add(&domain_flush_all
);
379 amd_iommu_stats_add(&alloced_io_mem
);
380 amd_iommu_stats_add(&total_map_requests
);
385 /****************************************************************************
387 * Interrupt handling functions
389 ****************************************************************************/
391 static void dump_dte_entry(u16 devid
)
395 for (i
= 0; i
< 4; ++i
)
396 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
397 amd_iommu_dev_table
[devid
].data
[i
]);
400 static void dump_command(unsigned long phys_addr
)
402 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
405 for (i
= 0; i
< 4; ++i
)
406 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
409 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
412 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
413 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
414 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
415 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
416 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
418 printk(KERN_ERR
"AMD-Vi: Event logged [");
421 case EVENT_TYPE_ILL_DEV
:
422 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
423 "address=0x%016llx flags=0x%04x]\n",
424 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
426 dump_dte_entry(devid
);
428 case EVENT_TYPE_IO_FAULT
:
429 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
430 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
431 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
432 domid
, address
, flags
);
434 case EVENT_TYPE_DEV_TAB_ERR
:
435 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
436 "address=0x%016llx flags=0x%04x]\n",
437 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
440 case EVENT_TYPE_PAGE_TAB_ERR
:
441 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
442 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
443 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
444 domid
, address
, flags
);
446 case EVENT_TYPE_ILL_CMD
:
447 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
448 dump_command(address
);
450 case EVENT_TYPE_CMD_HARD_ERR
:
451 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
452 "flags=0x%04x]\n", address
, flags
);
454 case EVENT_TYPE_IOTLB_INV_TO
:
455 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
456 "address=0x%016llx]\n",
457 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
460 case EVENT_TYPE_INV_DEV_REQ
:
461 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
462 "address=0x%016llx flags=0x%04x]\n",
463 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
467 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
471 static void iommu_poll_events(struct amd_iommu
*iommu
)
476 spin_lock_irqsave(&iommu
->lock
, flags
);
478 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
479 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
481 while (head
!= tail
) {
482 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
483 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
486 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
488 spin_unlock_irqrestore(&iommu
->lock
, flags
);
491 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
493 struct amd_iommu
*iommu
;
495 for_each_iommu(iommu
)
496 iommu_poll_events(iommu
);
501 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
503 return IRQ_WAKE_THREAD
;
506 /****************************************************************************
508 * IOMMU command queuing functions
510 ****************************************************************************/
512 static int wait_on_sem(volatile u64
*sem
)
516 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
521 if (i
== LOOP_TIMEOUT
) {
522 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
529 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
530 struct iommu_cmd
*cmd
,
535 target
= iommu
->cmd_buf
+ tail
;
536 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
538 /* Copy command to buffer */
539 memcpy(target
, cmd
, sizeof(*cmd
));
541 /* Tell the IOMMU about it */
542 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
545 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
547 WARN_ON(address
& 0x7ULL
);
549 memset(cmd
, 0, sizeof(*cmd
));
550 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
551 cmd
->data
[1] = upper_32_bits(__pa(address
));
553 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
556 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
558 memset(cmd
, 0, sizeof(*cmd
));
559 cmd
->data
[0] = devid
;
560 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
563 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
564 size_t size
, u16 domid
, int pde
)
569 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
574 * If we have to flush more than one page, flush all
575 * TLB entries for this domain
577 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
581 address
&= PAGE_MASK
;
583 memset(cmd
, 0, sizeof(*cmd
));
584 cmd
->data
[1] |= domid
;
585 cmd
->data
[2] = lower_32_bits(address
);
586 cmd
->data
[3] = upper_32_bits(address
);
587 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
588 if (s
) /* size bit - we flush more than one 4kb page */
589 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
590 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
591 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
594 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
595 u64 address
, size_t size
)
600 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
605 * If we have to flush more than one page, flush all
606 * TLB entries for this domain
608 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
612 address
&= PAGE_MASK
;
614 memset(cmd
, 0, sizeof(*cmd
));
615 cmd
->data
[0] = devid
;
616 cmd
->data
[0] |= (qdep
& 0xff) << 24;
617 cmd
->data
[1] = devid
;
618 cmd
->data
[2] = lower_32_bits(address
);
619 cmd
->data
[3] = upper_32_bits(address
);
620 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
622 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
625 static void build_inv_all(struct iommu_cmd
*cmd
)
627 memset(cmd
, 0, sizeof(*cmd
));
628 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
632 * Writes the command to the IOMMUs command buffer and informs the
633 * hardware about the new command.
635 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
636 struct iommu_cmd
*cmd
,
639 u32 left
, tail
, head
, next_tail
;
642 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
645 spin_lock_irqsave(&iommu
->lock
, flags
);
647 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
648 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
649 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
650 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
653 struct iommu_cmd sync_cmd
;
654 volatile u64 sem
= 0;
657 build_completion_wait(&sync_cmd
, (u64
)&sem
);
658 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
660 spin_unlock_irqrestore(&iommu
->lock
, flags
);
662 if ((ret
= wait_on_sem(&sem
)) != 0)
668 copy_cmd_to_buffer(iommu
, cmd
, tail
);
670 /* We need to sync now to make sure all commands are processed */
671 iommu
->need_sync
= sync
;
673 spin_unlock_irqrestore(&iommu
->lock
, flags
);
678 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
680 return iommu_queue_command_sync(iommu
, cmd
, true);
684 * This function queues a completion wait command into the command
687 static int iommu_completion_wait(struct amd_iommu
*iommu
)
689 struct iommu_cmd cmd
;
690 volatile u64 sem
= 0;
693 if (!iommu
->need_sync
)
696 build_completion_wait(&cmd
, (u64
)&sem
);
698 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
702 return wait_on_sem(&sem
);
705 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
707 struct iommu_cmd cmd
;
709 build_inv_dte(&cmd
, devid
);
711 return iommu_queue_command(iommu
, &cmd
);
714 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
718 for (devid
= 0; devid
<= 0xffff; ++devid
)
719 iommu_flush_dte(iommu
, devid
);
721 iommu_completion_wait(iommu
);
725 * This function uses heavy locking and may disable irqs for some time. But
726 * this is no issue because it is only called during resume.
728 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
732 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
733 struct iommu_cmd cmd
;
734 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
736 iommu_queue_command(iommu
, &cmd
);
739 iommu_completion_wait(iommu
);
742 static void iommu_flush_all(struct amd_iommu
*iommu
)
744 struct iommu_cmd cmd
;
748 iommu_queue_command(iommu
, &cmd
);
749 iommu_completion_wait(iommu
);
752 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
754 if (iommu_feature(iommu
, FEATURE_IA
)) {
755 iommu_flush_all(iommu
);
757 iommu_flush_dte_all(iommu
);
758 iommu_flush_tlb_all(iommu
);
763 * Command send function for flushing on-device TLB
765 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
766 u64 address
, size_t size
)
768 struct amd_iommu
*iommu
;
769 struct iommu_cmd cmd
;
772 qdep
= dev_data
->ats
.qdep
;
773 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
775 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
777 return iommu_queue_command(iommu
, &cmd
);
781 * Command send function for invalidating a device table entry
783 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
785 struct amd_iommu
*iommu
;
788 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
790 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
794 if (dev_data
->ats
.enabled
)
795 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
801 * TLB invalidation function which is called from the mapping functions.
802 * It invalidates a single PTE if the range to flush is within a single
803 * page. Otherwise it flushes the whole TLB of the IOMMU.
805 static void __domain_flush_pages(struct protection_domain
*domain
,
806 u64 address
, size_t size
, int pde
)
808 struct iommu_dev_data
*dev_data
;
809 struct iommu_cmd cmd
;
812 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
814 for (i
= 0; i
< amd_iommus_present
; ++i
) {
815 if (!domain
->dev_iommu
[i
])
819 * Devices of this domain are behind this IOMMU
820 * We need a TLB flush
822 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
825 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
827 if (!dev_data
->ats
.enabled
)
830 ret
|= device_flush_iotlb(dev_data
, address
, size
);
836 static void domain_flush_pages(struct protection_domain
*domain
,
837 u64 address
, size_t size
)
839 __domain_flush_pages(domain
, address
, size
, 0);
842 /* Flush the whole IO/TLB for a given protection domain */
843 static void domain_flush_tlb(struct protection_domain
*domain
)
845 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
848 /* Flush the whole IO/TLB for a given protection domain - including PDE */
849 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
851 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
854 static void domain_flush_complete(struct protection_domain
*domain
)
858 for (i
= 0; i
< amd_iommus_present
; ++i
) {
859 if (!domain
->dev_iommu
[i
])
863 * Devices of this domain are behind this IOMMU
864 * We need to wait for completion of all commands.
866 iommu_completion_wait(amd_iommus
[i
]);
872 * This function flushes the DTEs for all devices in domain
874 static void domain_flush_devices(struct protection_domain
*domain
)
876 struct iommu_dev_data
*dev_data
;
878 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
879 device_flush_dte(dev_data
);
882 /****************************************************************************
884 * The functions below are used the create the page table mappings for
885 * unity mapped regions.
887 ****************************************************************************/
890 * This function is used to add another level to an IO page table. Adding
891 * another level increases the size of the address space by 9 bits to a size up
894 static bool increase_address_space(struct protection_domain
*domain
,
899 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
900 /* address space already 64 bit large */
903 pte
= (void *)get_zeroed_page(gfp
);
907 *pte
= PM_LEVEL_PDE(domain
->mode
,
908 virt_to_phys(domain
->pt_root
));
909 domain
->pt_root
= pte
;
911 domain
->updated
= true;
916 static u64
*alloc_pte(struct protection_domain
*domain
,
917 unsigned long address
,
918 unsigned long page_size
,
925 BUG_ON(!is_power_of_2(page_size
));
927 while (address
> PM_LEVEL_SIZE(domain
->mode
))
928 increase_address_space(domain
, gfp
);
930 level
= domain
->mode
- 1;
931 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
932 address
= PAGE_SIZE_ALIGN(address
, page_size
);
933 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
935 while (level
> end_lvl
) {
936 if (!IOMMU_PTE_PRESENT(*pte
)) {
937 page
= (u64
*)get_zeroed_page(gfp
);
940 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
943 /* No level skipping support yet */
944 if (PM_PTE_LEVEL(*pte
) != level
)
949 pte
= IOMMU_PTE_PAGE(*pte
);
951 if (pte_page
&& level
== end_lvl
)
954 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
961 * This function checks if there is a PTE for a given dma address. If
962 * there is one, it returns the pointer to it.
964 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
969 if (address
> PM_LEVEL_SIZE(domain
->mode
))
972 level
= domain
->mode
- 1;
973 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
978 if (!IOMMU_PTE_PRESENT(*pte
))
982 if (PM_PTE_LEVEL(*pte
) == 0x07) {
983 unsigned long pte_mask
, __pte
;
986 * If we have a series of large PTEs, make
987 * sure to return a pointer to the first one.
989 pte_mask
= PTE_PAGE_SIZE(*pte
);
990 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
991 __pte
= ((unsigned long)pte
) & pte_mask
;
996 /* No level skipping support yet */
997 if (PM_PTE_LEVEL(*pte
) != level
)
1002 /* Walk to the next level */
1003 pte
= IOMMU_PTE_PAGE(*pte
);
1004 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1011 * Generic mapping functions. It maps a physical address into a DMA
1012 * address space. It allocates the page table pages if necessary.
1013 * In the future it can be extended to a generic mapping function
1014 * supporting all features of AMD IOMMU page tables like level skipping
1015 * and full 64 bit address spaces.
1017 static int iommu_map_page(struct protection_domain
*dom
,
1018 unsigned long bus_addr
,
1019 unsigned long phys_addr
,
1021 unsigned long page_size
)
1026 if (!(prot
& IOMMU_PROT_MASK
))
1029 bus_addr
= PAGE_ALIGN(bus_addr
);
1030 phys_addr
= PAGE_ALIGN(phys_addr
);
1031 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1032 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1034 for (i
= 0; i
< count
; ++i
)
1035 if (IOMMU_PTE_PRESENT(pte
[i
]))
1038 if (page_size
> PAGE_SIZE
) {
1039 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1040 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1042 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1044 if (prot
& IOMMU_PROT_IR
)
1045 __pte
|= IOMMU_PTE_IR
;
1046 if (prot
& IOMMU_PROT_IW
)
1047 __pte
|= IOMMU_PTE_IW
;
1049 for (i
= 0; i
< count
; ++i
)
1057 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1058 unsigned long bus_addr
,
1059 unsigned long page_size
)
1061 unsigned long long unmap_size
, unmapped
;
1064 BUG_ON(!is_power_of_2(page_size
));
1068 while (unmapped
< page_size
) {
1070 pte
= fetch_pte(dom
, bus_addr
);
1074 * No PTE for this address
1075 * move forward in 4kb steps
1077 unmap_size
= PAGE_SIZE
;
1078 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1079 /* 4kb PTE found for this address */
1080 unmap_size
= PAGE_SIZE
;
1085 /* Large PTE found which maps this address */
1086 unmap_size
= PTE_PAGE_SIZE(*pte
);
1087 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1088 for (i
= 0; i
< count
; i
++)
1092 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1093 unmapped
+= unmap_size
;
1096 BUG_ON(!is_power_of_2(unmapped
));
1102 * This function checks if a specific unity mapping entry is needed for
1103 * this specific IOMMU.
1105 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1106 struct unity_map_entry
*entry
)
1110 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1111 bdf
= amd_iommu_alias_table
[i
];
1112 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1120 * This function actually applies the mapping to the page table of the
1123 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1124 struct unity_map_entry
*e
)
1129 for (addr
= e
->address_start
; addr
< e
->address_end
;
1130 addr
+= PAGE_SIZE
) {
1131 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1136 * if unity mapping is in aperture range mark the page
1137 * as allocated in the aperture
1139 if (addr
< dma_dom
->aperture_size
)
1140 __set_bit(addr
>> PAGE_SHIFT
,
1141 dma_dom
->aperture
[0]->bitmap
);
1148 * Init the unity mappings for a specific IOMMU in the system
1150 * Basically iterates over all unity mapping entries and applies them to
1151 * the default domain DMA of that IOMMU if necessary.
1153 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1155 struct unity_map_entry
*entry
;
1158 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1159 if (!iommu_for_unity_map(iommu
, entry
))
1161 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1170 * Inits the unity mappings required for a specific device
1172 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1175 struct unity_map_entry
*e
;
1178 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1179 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1181 ret
= dma_ops_unity_map(dma_dom
, e
);
1189 /****************************************************************************
1191 * The next functions belong to the address allocator for the dma_ops
1192 * interface functions. They work like the allocators in the other IOMMU
1193 * drivers. Its basically a bitmap which marks the allocated pages in
1194 * the aperture. Maybe it could be enhanced in the future to a more
1195 * efficient allocator.
1197 ****************************************************************************/
1200 * The address allocator core functions.
1202 * called with domain->lock held
1206 * Used to reserve address ranges in the aperture (e.g. for exclusion
1209 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1210 unsigned long start_page
,
1213 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1215 if (start_page
+ pages
> last_page
)
1216 pages
= last_page
- start_page
;
1218 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1219 int index
= i
/ APERTURE_RANGE_PAGES
;
1220 int page
= i
% APERTURE_RANGE_PAGES
;
1221 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1226 * This function is used to add a new aperture range to an existing
1227 * aperture in case of dma_ops domain allocation or address allocation
1230 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1231 bool populate
, gfp_t gfp
)
1233 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1234 struct amd_iommu
*iommu
;
1235 unsigned long i
, old_size
;
1237 #ifdef CONFIG_IOMMU_STRESS
1241 if (index
>= APERTURE_MAX_RANGES
)
1244 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1245 if (!dma_dom
->aperture
[index
])
1248 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1249 if (!dma_dom
->aperture
[index
]->bitmap
)
1252 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1255 unsigned long address
= dma_dom
->aperture_size
;
1256 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1257 u64
*pte
, *pte_page
;
1259 for (i
= 0; i
< num_ptes
; ++i
) {
1260 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1265 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1267 address
+= APERTURE_RANGE_SIZE
/ 64;
1271 old_size
= dma_dom
->aperture_size
;
1272 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1274 /* Reserve address range used for MSI messages */
1275 if (old_size
< MSI_ADDR_BASE_LO
&&
1276 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1277 unsigned long spage
;
1280 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1281 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1283 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1286 /* Initialize the exclusion range if necessary */
1287 for_each_iommu(iommu
) {
1288 if (iommu
->exclusion_start
&&
1289 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1290 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1291 unsigned long startpage
;
1292 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1293 iommu
->exclusion_length
,
1295 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1296 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1301 * Check for areas already mapped as present in the new aperture
1302 * range and mark those pages as reserved in the allocator. Such
1303 * mappings may already exist as a result of requested unity
1304 * mappings for devices.
1306 for (i
= dma_dom
->aperture
[index
]->offset
;
1307 i
< dma_dom
->aperture_size
;
1309 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1310 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1313 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1316 update_domain(&dma_dom
->domain
);
1321 update_domain(&dma_dom
->domain
);
1323 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1325 kfree(dma_dom
->aperture
[index
]);
1326 dma_dom
->aperture
[index
] = NULL
;
1331 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1332 struct dma_ops_domain
*dom
,
1334 unsigned long align_mask
,
1336 unsigned long start
)
1338 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1339 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1340 int i
= start
>> APERTURE_RANGE_SHIFT
;
1341 unsigned long boundary_size
;
1342 unsigned long address
= -1;
1343 unsigned long limit
;
1345 next_bit
>>= PAGE_SHIFT
;
1347 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1348 PAGE_SIZE
) >> PAGE_SHIFT
;
1350 for (;i
< max_index
; ++i
) {
1351 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1353 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1356 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1357 dma_mask
>> PAGE_SHIFT
);
1359 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1360 limit
, next_bit
, pages
, 0,
1361 boundary_size
, align_mask
);
1362 if (address
!= -1) {
1363 address
= dom
->aperture
[i
]->offset
+
1364 (address
<< PAGE_SHIFT
);
1365 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1375 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1376 struct dma_ops_domain
*dom
,
1378 unsigned long align_mask
,
1381 unsigned long address
;
1383 #ifdef CONFIG_IOMMU_STRESS
1384 dom
->next_address
= 0;
1385 dom
->need_flush
= true;
1388 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1389 dma_mask
, dom
->next_address
);
1391 if (address
== -1) {
1392 dom
->next_address
= 0;
1393 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1395 dom
->need_flush
= true;
1398 if (unlikely(address
== -1))
1399 address
= DMA_ERROR_CODE
;
1401 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1407 * The address free function.
1409 * called with domain->lock held
1411 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1412 unsigned long address
,
1415 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1416 struct aperture_range
*range
= dom
->aperture
[i
];
1418 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1420 #ifdef CONFIG_IOMMU_STRESS
1425 if (address
>= dom
->next_address
)
1426 dom
->need_flush
= true;
1428 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1430 bitmap_clear(range
->bitmap
, address
, pages
);
1434 /****************************************************************************
1436 * The next functions belong to the domain allocation. A domain is
1437 * allocated for every IOMMU as the default domain. If device isolation
1438 * is enabled, every device get its own domain. The most important thing
1439 * about domains is the page table mapping the DMA address space they
1442 ****************************************************************************/
1445 * This function adds a protection domain to the global protection domain list
1447 static void add_domain_to_list(struct protection_domain
*domain
)
1449 unsigned long flags
;
1451 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1452 list_add(&domain
->list
, &amd_iommu_pd_list
);
1453 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1457 * This function removes a protection domain to the global
1458 * protection domain list
1460 static void del_domain_from_list(struct protection_domain
*domain
)
1462 unsigned long flags
;
1464 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1465 list_del(&domain
->list
);
1466 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1469 static u16
domain_id_alloc(void)
1471 unsigned long flags
;
1474 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1475 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1477 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1478 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1481 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1486 static void domain_id_free(int id
)
1488 unsigned long flags
;
1490 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1491 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1492 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1493 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1496 static void free_pagetable(struct protection_domain
*domain
)
1501 p1
= domain
->pt_root
;
1506 for (i
= 0; i
< 512; ++i
) {
1507 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1510 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1511 for (j
= 0; j
< 512; ++j
) {
1512 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1514 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1515 free_page((unsigned long)p3
);
1518 free_page((unsigned long)p2
);
1521 free_page((unsigned long)p1
);
1523 domain
->pt_root
= NULL
;
1527 * Free a domain, only used if something went wrong in the
1528 * allocation path and we need to free an already allocated page table
1530 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1537 del_domain_from_list(&dom
->domain
);
1539 free_pagetable(&dom
->domain
);
1541 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1542 if (!dom
->aperture
[i
])
1544 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1545 kfree(dom
->aperture
[i
]);
1552 * Allocates a new protection domain usable for the dma_ops functions.
1553 * It also initializes the page table and the address allocator data
1554 * structures required for the dma_ops interface
1556 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1558 struct dma_ops_domain
*dma_dom
;
1560 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1564 spin_lock_init(&dma_dom
->domain
.lock
);
1566 dma_dom
->domain
.id
= domain_id_alloc();
1567 if (dma_dom
->domain
.id
== 0)
1569 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1570 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1571 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1572 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1573 dma_dom
->domain
.priv
= dma_dom
;
1574 if (!dma_dom
->domain
.pt_root
)
1577 dma_dom
->need_flush
= false;
1578 dma_dom
->target_dev
= 0xffff;
1580 add_domain_to_list(&dma_dom
->domain
);
1582 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1586 * mark the first page as allocated so we never return 0 as
1587 * a valid dma-address. So we can use 0 as error value
1589 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1590 dma_dom
->next_address
= 0;
1596 dma_ops_domain_free(dma_dom
);
1602 * little helper function to check whether a given protection domain is a
1605 static bool dma_ops_domain(struct protection_domain
*domain
)
1607 return domain
->flags
& PD_DMA_OPS_MASK
;
1610 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1612 u64 pte_root
= virt_to_phys(domain
->pt_root
);
1615 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1616 << DEV_ENTRY_MODE_SHIFT
;
1617 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1619 flags
= amd_iommu_dev_table
[devid
].data
[1];
1622 flags
|= DTE_FLAG_IOTLB
;
1624 flags
&= ~(0xffffUL
);
1625 flags
|= domain
->id
;
1627 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1628 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1631 static void clear_dte_entry(u16 devid
)
1633 /* remove entry from the device table seen by the hardware */
1634 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1635 amd_iommu_dev_table
[devid
].data
[1] = 0;
1637 amd_iommu_apply_erratum_63(devid
);
1640 static void do_attach(struct iommu_dev_data
*dev_data
,
1641 struct protection_domain
*domain
)
1643 struct amd_iommu
*iommu
;
1646 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1647 ats
= dev_data
->ats
.enabled
;
1649 /* Update data structures */
1650 dev_data
->domain
= domain
;
1651 list_add(&dev_data
->list
, &domain
->dev_list
);
1652 set_dte_entry(dev_data
->devid
, domain
, ats
);
1654 /* Do reference counting */
1655 domain
->dev_iommu
[iommu
->index
] += 1;
1656 domain
->dev_cnt
+= 1;
1658 /* Flush the DTE entry */
1659 device_flush_dte(dev_data
);
1662 static void do_detach(struct iommu_dev_data
*dev_data
)
1664 struct amd_iommu
*iommu
;
1666 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1668 /* decrease reference counters */
1669 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1670 dev_data
->domain
->dev_cnt
-= 1;
1672 /* Update data structures */
1673 dev_data
->domain
= NULL
;
1674 list_del(&dev_data
->list
);
1675 clear_dte_entry(dev_data
->devid
);
1677 /* Flush the DTE entry */
1678 device_flush_dte(dev_data
);
1682 * If a device is not yet associated with a domain, this function does
1683 * assigns it visible for the hardware
1685 static int __attach_device(struct iommu_dev_data
*dev_data
,
1686 struct protection_domain
*domain
)
1691 spin_lock(&domain
->lock
);
1693 if (dev_data
->alias_data
!= NULL
) {
1694 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1696 /* Some sanity checks */
1698 if (alias_data
->domain
!= NULL
&&
1699 alias_data
->domain
!= domain
)
1702 if (dev_data
->domain
!= NULL
&&
1703 dev_data
->domain
!= domain
)
1706 /* Do real assignment */
1707 if (alias_data
->domain
== NULL
)
1708 do_attach(alias_data
, domain
);
1710 atomic_inc(&alias_data
->bind
);
1713 if (dev_data
->domain
== NULL
)
1714 do_attach(dev_data
, domain
);
1716 atomic_inc(&dev_data
->bind
);
1723 spin_unlock(&domain
->lock
);
1729 * If a device is not yet associated with a domain, this function does
1730 * assigns it visible for the hardware
1732 static int attach_device(struct device
*dev
,
1733 struct protection_domain
*domain
)
1735 struct pci_dev
*pdev
= to_pci_dev(dev
);
1736 struct iommu_dev_data
*dev_data
;
1737 unsigned long flags
;
1740 dev_data
= get_dev_data(dev
);
1742 if (amd_iommu_iotlb_sup
&& pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
1743 dev_data
->ats
.enabled
= true;
1744 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
1747 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1748 ret
= __attach_device(dev_data
, domain
);
1749 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1752 * We might boot into a crash-kernel here. The crashed kernel
1753 * left the caches in the IOMMU dirty. So we have to flush
1754 * here to evict all dirty stuff.
1756 domain_flush_tlb_pde(domain
);
1762 * Removes a device from a protection domain (unlocked)
1764 static void __detach_device(struct iommu_dev_data
*dev_data
)
1766 struct protection_domain
*domain
;
1767 unsigned long flags
;
1769 BUG_ON(!dev_data
->domain
);
1771 domain
= dev_data
->domain
;
1773 spin_lock_irqsave(&domain
->lock
, flags
);
1775 if (dev_data
->alias_data
!= NULL
) {
1776 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1778 if (atomic_dec_and_test(&alias_data
->bind
))
1779 do_detach(alias_data
);
1782 if (atomic_dec_and_test(&dev_data
->bind
))
1783 do_detach(dev_data
);
1785 spin_unlock_irqrestore(&domain
->lock
, flags
);
1788 * If we run in passthrough mode the device must be assigned to the
1789 * passthrough domain if it is detached from any other domain.
1790 * Make sure we can deassign from the pt_domain itself.
1792 if (dev_data
->passthrough
&&
1793 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
1794 __attach_device(dev_data
, pt_domain
);
1798 * Removes a device from a protection domain (with devtable_lock held)
1800 static void detach_device(struct device
*dev
)
1802 struct iommu_dev_data
*dev_data
;
1803 unsigned long flags
;
1805 dev_data
= get_dev_data(dev
);
1807 /* lock device table */
1808 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1809 __detach_device(dev_data
);
1810 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1812 if (dev_data
->ats
.enabled
) {
1813 pci_disable_ats(to_pci_dev(dev
));
1814 dev_data
->ats
.enabled
= false;
1819 * Find out the protection domain structure for a given PCI device. This
1820 * will give us the pointer to the page table root for example.
1822 static struct protection_domain
*domain_for_device(struct device
*dev
)
1824 struct iommu_dev_data
*dev_data
;
1825 struct protection_domain
*dom
= NULL
;
1826 unsigned long flags
;
1828 dev_data
= get_dev_data(dev
);
1830 if (dev_data
->domain
)
1831 return dev_data
->domain
;
1833 if (dev_data
->alias_data
!= NULL
) {
1834 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1836 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1837 if (alias_data
->domain
!= NULL
) {
1838 __attach_device(dev_data
, alias_data
->domain
);
1839 dom
= alias_data
->domain
;
1841 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1847 static int device_change_notifier(struct notifier_block
*nb
,
1848 unsigned long action
, void *data
)
1850 struct dma_ops_domain
*dma_domain
;
1851 struct protection_domain
*domain
;
1852 struct iommu_dev_data
*dev_data
;
1853 struct device
*dev
= data
;
1854 struct amd_iommu
*iommu
;
1855 unsigned long flags
;
1858 if (!check_device(dev
))
1861 devid
= get_device_id(dev
);
1862 iommu
= amd_iommu_rlookup_table
[devid
];
1863 dev_data
= get_dev_data(dev
);
1866 case BUS_NOTIFY_UNBOUND_DRIVER
:
1868 domain
= domain_for_device(dev
);
1872 if (dev_data
->passthrough
)
1876 case BUS_NOTIFY_ADD_DEVICE
:
1878 iommu_init_device(dev
);
1880 domain
= domain_for_device(dev
);
1882 /* allocate a protection domain if a device is added */
1883 dma_domain
= find_protection_domain(devid
);
1886 dma_domain
= dma_ops_domain_alloc();
1889 dma_domain
->target_dev
= devid
;
1891 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1892 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1893 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1896 case BUS_NOTIFY_DEL_DEVICE
:
1898 iommu_uninit_device(dev
);
1904 iommu_completion_wait(iommu
);
1910 static struct notifier_block device_nb
= {
1911 .notifier_call
= device_change_notifier
,
1914 void amd_iommu_init_notifier(void)
1916 bus_register_notifier(&pci_bus_type
, &device_nb
);
1919 /*****************************************************************************
1921 * The next functions belong to the dma_ops mapping/unmapping code.
1923 *****************************************************************************/
1926 * In the dma_ops path we only have the struct device. This function
1927 * finds the corresponding IOMMU, the protection domain and the
1928 * requestor id for a given device.
1929 * If the device is not yet associated with a domain this is also done
1932 static struct protection_domain
*get_domain(struct device
*dev
)
1934 struct protection_domain
*domain
;
1935 struct dma_ops_domain
*dma_dom
;
1936 u16 devid
= get_device_id(dev
);
1938 if (!check_device(dev
))
1939 return ERR_PTR(-EINVAL
);
1941 domain
= domain_for_device(dev
);
1942 if (domain
!= NULL
&& !dma_ops_domain(domain
))
1943 return ERR_PTR(-EBUSY
);
1948 /* Device not bount yet - bind it */
1949 dma_dom
= find_protection_domain(devid
);
1951 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
1952 attach_device(dev
, &dma_dom
->domain
);
1953 DUMP_printk("Using protection domain %d for device %s\n",
1954 dma_dom
->domain
.id
, dev_name(dev
));
1956 return &dma_dom
->domain
;
1959 static void update_device_table(struct protection_domain
*domain
)
1961 struct iommu_dev_data
*dev_data
;
1963 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1964 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
1967 static void update_domain(struct protection_domain
*domain
)
1969 if (!domain
->updated
)
1972 update_device_table(domain
);
1974 domain_flush_devices(domain
);
1975 domain_flush_tlb_pde(domain
);
1977 domain
->updated
= false;
1981 * This function fetches the PTE for a given address in the aperture
1983 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
1984 unsigned long address
)
1986 struct aperture_range
*aperture
;
1987 u64
*pte
, *pte_page
;
1989 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1993 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1995 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
1997 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
1999 pte
+= PM_LEVEL_INDEX(0, address
);
2001 update_domain(&dom
->domain
);
2007 * This is the generic map function. It maps one 4kb page at paddr to
2008 * the given address in the DMA address space for the domain.
2010 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2011 unsigned long address
,
2017 WARN_ON(address
> dom
->aperture_size
);
2021 pte
= dma_ops_get_pte(dom
, address
);
2023 return DMA_ERROR_CODE
;
2025 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2027 if (direction
== DMA_TO_DEVICE
)
2028 __pte
|= IOMMU_PTE_IR
;
2029 else if (direction
== DMA_FROM_DEVICE
)
2030 __pte
|= IOMMU_PTE_IW
;
2031 else if (direction
== DMA_BIDIRECTIONAL
)
2032 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2038 return (dma_addr_t
)address
;
2042 * The generic unmapping function for on page in the DMA address space.
2044 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2045 unsigned long address
)
2047 struct aperture_range
*aperture
;
2050 if (address
>= dom
->aperture_size
)
2053 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2057 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2061 pte
+= PM_LEVEL_INDEX(0, address
);
2069 * This function contains common code for mapping of a physically
2070 * contiguous memory region into DMA address space. It is used by all
2071 * mapping functions provided with this IOMMU driver.
2072 * Must be called with the domain lock held.
2074 static dma_addr_t
__map_single(struct device
*dev
,
2075 struct dma_ops_domain
*dma_dom
,
2082 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2083 dma_addr_t address
, start
, ret
;
2085 unsigned long align_mask
= 0;
2088 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2091 INC_STATS_COUNTER(total_map_requests
);
2094 INC_STATS_COUNTER(cross_page
);
2097 align_mask
= (1UL << get_order(size
)) - 1;
2100 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2102 if (unlikely(address
== DMA_ERROR_CODE
)) {
2104 * setting next_address here will let the address
2105 * allocator only scan the new allocated range in the
2106 * first run. This is a small optimization.
2108 dma_dom
->next_address
= dma_dom
->aperture_size
;
2110 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2114 * aperture was successfully enlarged by 128 MB, try
2121 for (i
= 0; i
< pages
; ++i
) {
2122 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2123 if (ret
== DMA_ERROR_CODE
)
2131 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2133 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2134 domain_flush_tlb(&dma_dom
->domain
);
2135 dma_dom
->need_flush
= false;
2136 } else if (unlikely(amd_iommu_np_cache
))
2137 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2144 for (--i
; i
>= 0; --i
) {
2146 dma_ops_domain_unmap(dma_dom
, start
);
2149 dma_ops_free_addresses(dma_dom
, address
, pages
);
2151 return DMA_ERROR_CODE
;
2155 * Does the reverse of the __map_single function. Must be called with
2156 * the domain lock held too
2158 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2159 dma_addr_t dma_addr
,
2163 dma_addr_t flush_addr
;
2164 dma_addr_t i
, start
;
2167 if ((dma_addr
== DMA_ERROR_CODE
) ||
2168 (dma_addr
+ size
> dma_dom
->aperture_size
))
2171 flush_addr
= dma_addr
;
2172 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2173 dma_addr
&= PAGE_MASK
;
2176 for (i
= 0; i
< pages
; ++i
) {
2177 dma_ops_domain_unmap(dma_dom
, start
);
2181 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2183 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2185 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2186 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2187 dma_dom
->need_flush
= false;
2192 * The exported map_single function for dma_ops.
2194 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2195 unsigned long offset
, size_t size
,
2196 enum dma_data_direction dir
,
2197 struct dma_attrs
*attrs
)
2199 unsigned long flags
;
2200 struct protection_domain
*domain
;
2203 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2205 INC_STATS_COUNTER(cnt_map_single
);
2207 domain
= get_domain(dev
);
2208 if (PTR_ERR(domain
) == -EINVAL
)
2209 return (dma_addr_t
)paddr
;
2210 else if (IS_ERR(domain
))
2211 return DMA_ERROR_CODE
;
2213 dma_mask
= *dev
->dma_mask
;
2215 spin_lock_irqsave(&domain
->lock
, flags
);
2217 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2219 if (addr
== DMA_ERROR_CODE
)
2222 domain_flush_complete(domain
);
2225 spin_unlock_irqrestore(&domain
->lock
, flags
);
2231 * The exported unmap_single function for dma_ops.
2233 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2234 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2236 unsigned long flags
;
2237 struct protection_domain
*domain
;
2239 INC_STATS_COUNTER(cnt_unmap_single
);
2241 domain
= get_domain(dev
);
2245 spin_lock_irqsave(&domain
->lock
, flags
);
2247 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2249 domain_flush_complete(domain
);
2251 spin_unlock_irqrestore(&domain
->lock
, flags
);
2255 * This is a special map_sg function which is used if we should map a
2256 * device which is not handled by an AMD IOMMU in the system.
2258 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2259 int nelems
, int dir
)
2261 struct scatterlist
*s
;
2264 for_each_sg(sglist
, s
, nelems
, i
) {
2265 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2266 s
->dma_length
= s
->length
;
2273 * The exported map_sg function for dma_ops (handles scatter-gather
2276 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2277 int nelems
, enum dma_data_direction dir
,
2278 struct dma_attrs
*attrs
)
2280 unsigned long flags
;
2281 struct protection_domain
*domain
;
2283 struct scatterlist
*s
;
2285 int mapped_elems
= 0;
2288 INC_STATS_COUNTER(cnt_map_sg
);
2290 domain
= get_domain(dev
);
2291 if (PTR_ERR(domain
) == -EINVAL
)
2292 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2293 else if (IS_ERR(domain
))
2296 dma_mask
= *dev
->dma_mask
;
2298 spin_lock_irqsave(&domain
->lock
, flags
);
2300 for_each_sg(sglist
, s
, nelems
, i
) {
2303 s
->dma_address
= __map_single(dev
, domain
->priv
,
2304 paddr
, s
->length
, dir
, false,
2307 if (s
->dma_address
) {
2308 s
->dma_length
= s
->length
;
2314 domain_flush_complete(domain
);
2317 spin_unlock_irqrestore(&domain
->lock
, flags
);
2319 return mapped_elems
;
2321 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2323 __unmap_single(domain
->priv
, s
->dma_address
,
2324 s
->dma_length
, dir
);
2325 s
->dma_address
= s
->dma_length
= 0;
2334 * The exported map_sg function for dma_ops (handles scatter-gather
2337 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2338 int nelems
, enum dma_data_direction dir
,
2339 struct dma_attrs
*attrs
)
2341 unsigned long flags
;
2342 struct protection_domain
*domain
;
2343 struct scatterlist
*s
;
2346 INC_STATS_COUNTER(cnt_unmap_sg
);
2348 domain
= get_domain(dev
);
2352 spin_lock_irqsave(&domain
->lock
, flags
);
2354 for_each_sg(sglist
, s
, nelems
, i
) {
2355 __unmap_single(domain
->priv
, s
->dma_address
,
2356 s
->dma_length
, dir
);
2357 s
->dma_address
= s
->dma_length
= 0;
2360 domain_flush_complete(domain
);
2362 spin_unlock_irqrestore(&domain
->lock
, flags
);
2366 * The exported alloc_coherent function for dma_ops.
2368 static void *alloc_coherent(struct device
*dev
, size_t size
,
2369 dma_addr_t
*dma_addr
, gfp_t flag
)
2371 unsigned long flags
;
2373 struct protection_domain
*domain
;
2375 u64 dma_mask
= dev
->coherent_dma_mask
;
2377 INC_STATS_COUNTER(cnt_alloc_coherent
);
2379 domain
= get_domain(dev
);
2380 if (PTR_ERR(domain
) == -EINVAL
) {
2381 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2382 *dma_addr
= __pa(virt_addr
);
2384 } else if (IS_ERR(domain
))
2387 dma_mask
= dev
->coherent_dma_mask
;
2388 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2391 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2395 paddr
= virt_to_phys(virt_addr
);
2398 dma_mask
= *dev
->dma_mask
;
2400 spin_lock_irqsave(&domain
->lock
, flags
);
2402 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2403 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2405 if (*dma_addr
== DMA_ERROR_CODE
) {
2406 spin_unlock_irqrestore(&domain
->lock
, flags
);
2410 domain_flush_complete(domain
);
2412 spin_unlock_irqrestore(&domain
->lock
, flags
);
2418 free_pages((unsigned long)virt_addr
, get_order(size
));
2424 * The exported free_coherent function for dma_ops.
2426 static void free_coherent(struct device
*dev
, size_t size
,
2427 void *virt_addr
, dma_addr_t dma_addr
)
2429 unsigned long flags
;
2430 struct protection_domain
*domain
;
2432 INC_STATS_COUNTER(cnt_free_coherent
);
2434 domain
= get_domain(dev
);
2438 spin_lock_irqsave(&domain
->lock
, flags
);
2440 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2442 domain_flush_complete(domain
);
2444 spin_unlock_irqrestore(&domain
->lock
, flags
);
2447 free_pages((unsigned long)virt_addr
, get_order(size
));
2451 * This function is called by the DMA layer to find out if we can handle a
2452 * particular device. It is part of the dma_ops.
2454 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2456 return check_device(dev
);
2460 * The function for pre-allocating protection domains.
2462 * If the driver core informs the DMA layer if a driver grabs a device
2463 * we don't need to preallocate the protection domains anymore.
2464 * For now we have to.
2466 static void prealloc_protection_domains(void)
2468 struct iommu_dev_data
*dev_data
;
2469 struct dma_ops_domain
*dma_dom
;
2470 struct pci_dev
*dev
= NULL
;
2473 for_each_pci_dev(dev
) {
2475 /* Do we handle this device? */
2476 if (!check_device(&dev
->dev
))
2479 dev_data
= get_dev_data(&dev
->dev
);
2480 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
2481 /* Make sure passthrough domain is allocated */
2482 alloc_passthrough_domain();
2483 dev_data
->passthrough
= true;
2484 attach_device(&dev
->dev
, pt_domain
);
2485 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2486 dev_name(&dev
->dev
));
2489 /* Is there already any domain for it? */
2490 if (domain_for_device(&dev
->dev
))
2493 devid
= get_device_id(&dev
->dev
);
2495 dma_dom
= dma_ops_domain_alloc();
2498 init_unity_mappings_for_device(dma_dom
, devid
);
2499 dma_dom
->target_dev
= devid
;
2501 attach_device(&dev
->dev
, &dma_dom
->domain
);
2503 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2507 static struct dma_map_ops amd_iommu_dma_ops
= {
2508 .alloc_coherent
= alloc_coherent
,
2509 .free_coherent
= free_coherent
,
2510 .map_page
= map_page
,
2511 .unmap_page
= unmap_page
,
2513 .unmap_sg
= unmap_sg
,
2514 .dma_supported
= amd_iommu_dma_supported
,
2517 static unsigned device_dma_ops_init(void)
2519 struct iommu_dev_data
*dev_data
;
2520 struct pci_dev
*pdev
= NULL
;
2521 unsigned unhandled
= 0;
2523 for_each_pci_dev(pdev
) {
2524 if (!check_device(&pdev
->dev
)) {
2529 dev_data
= get_dev_data(&pdev
->dev
);
2531 if (!dev_data
->passthrough
)
2532 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
2534 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
2541 * The function which clues the AMD IOMMU driver into dma_ops.
2544 void __init
amd_iommu_init_api(void)
2546 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2549 int __init
amd_iommu_init_dma_ops(void)
2551 struct amd_iommu
*iommu
;
2555 * first allocate a default protection domain for every IOMMU we
2556 * found in the system. Devices not assigned to any other
2557 * protection domain will be assigned to the default one.
2559 for_each_iommu(iommu
) {
2560 iommu
->default_dom
= dma_ops_domain_alloc();
2561 if (iommu
->default_dom
== NULL
)
2563 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2564 ret
= iommu_init_unity_mappings(iommu
);
2570 * Pre-allocate the protection domains for each device.
2572 prealloc_protection_domains();
2577 /* Make the driver finally visible to the drivers */
2578 unhandled
= device_dma_ops_init();
2579 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
2580 /* There are unhandled devices - initialize swiotlb for them */
2584 amd_iommu_stats_init();
2590 for_each_iommu(iommu
) {
2591 if (iommu
->default_dom
)
2592 dma_ops_domain_free(iommu
->default_dom
);
2598 /*****************************************************************************
2600 * The following functions belong to the exported interface of AMD IOMMU
2602 * This interface allows access to lower level functions of the IOMMU
2603 * like protection domain handling and assignement of devices to domains
2604 * which is not possible with the dma_ops interface.
2606 *****************************************************************************/
2608 static void cleanup_domain(struct protection_domain
*domain
)
2610 struct iommu_dev_data
*dev_data
, *next
;
2611 unsigned long flags
;
2613 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2615 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
2616 __detach_device(dev_data
);
2617 atomic_set(&dev_data
->bind
, 0);
2620 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2623 static void protection_domain_free(struct protection_domain
*domain
)
2628 del_domain_from_list(domain
);
2631 domain_id_free(domain
->id
);
2636 static struct protection_domain
*protection_domain_alloc(void)
2638 struct protection_domain
*domain
;
2640 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2644 spin_lock_init(&domain
->lock
);
2645 mutex_init(&domain
->api_lock
);
2646 domain
->id
= domain_id_alloc();
2649 INIT_LIST_HEAD(&domain
->dev_list
);
2651 add_domain_to_list(domain
);
2661 static int __init
alloc_passthrough_domain(void)
2663 if (pt_domain
!= NULL
)
2666 /* allocate passthrough domain */
2667 pt_domain
= protection_domain_alloc();
2671 pt_domain
->mode
= PAGE_MODE_NONE
;
2675 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2677 struct protection_domain
*domain
;
2679 domain
= protection_domain_alloc();
2683 domain
->mode
= PAGE_MODE_3_LEVEL
;
2684 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2685 if (!domain
->pt_root
)
2693 protection_domain_free(domain
);
2698 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2700 struct protection_domain
*domain
= dom
->priv
;
2705 if (domain
->dev_cnt
> 0)
2706 cleanup_domain(domain
);
2708 BUG_ON(domain
->dev_cnt
!= 0);
2710 free_pagetable(domain
);
2712 protection_domain_free(domain
);
2717 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2720 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
2721 struct amd_iommu
*iommu
;
2724 if (!check_device(dev
))
2727 devid
= get_device_id(dev
);
2729 if (dev_data
->domain
!= NULL
)
2732 iommu
= amd_iommu_rlookup_table
[devid
];
2736 iommu_completion_wait(iommu
);
2739 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2742 struct protection_domain
*domain
= dom
->priv
;
2743 struct iommu_dev_data
*dev_data
;
2744 struct amd_iommu
*iommu
;
2747 if (!check_device(dev
))
2750 dev_data
= dev
->archdata
.iommu
;
2752 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2756 if (dev_data
->domain
)
2759 ret
= attach_device(dev
, domain
);
2761 iommu_completion_wait(iommu
);
2766 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
2767 phys_addr_t paddr
, int gfp_order
, int iommu_prot
)
2769 unsigned long page_size
= 0x1000UL
<< gfp_order
;
2770 struct protection_domain
*domain
= dom
->priv
;
2774 if (iommu_prot
& IOMMU_READ
)
2775 prot
|= IOMMU_PROT_IR
;
2776 if (iommu_prot
& IOMMU_WRITE
)
2777 prot
|= IOMMU_PROT_IW
;
2779 mutex_lock(&domain
->api_lock
);
2780 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
2781 mutex_unlock(&domain
->api_lock
);
2786 static int amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
2789 struct protection_domain
*domain
= dom
->priv
;
2790 unsigned long page_size
, unmap_size
;
2792 page_size
= 0x1000UL
<< gfp_order
;
2794 mutex_lock(&domain
->api_lock
);
2795 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
2796 mutex_unlock(&domain
->api_lock
);
2798 domain_flush_tlb_pde(domain
);
2800 return get_order(unmap_size
);
2803 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2806 struct protection_domain
*domain
= dom
->priv
;
2807 unsigned long offset_mask
;
2811 pte
= fetch_pte(domain
, iova
);
2813 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
2816 if (PM_PTE_LEVEL(*pte
) == 0)
2817 offset_mask
= PAGE_SIZE
- 1;
2819 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
2821 __pte
= *pte
& PM_ADDR_MASK
;
2822 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
2827 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
2831 case IOMMU_CAP_CACHE_COHERENCY
:
2838 static struct iommu_ops amd_iommu_ops
= {
2839 .domain_init
= amd_iommu_domain_init
,
2840 .domain_destroy
= amd_iommu_domain_destroy
,
2841 .attach_dev
= amd_iommu_attach_device
,
2842 .detach_dev
= amd_iommu_detach_device
,
2843 .map
= amd_iommu_map
,
2844 .unmap
= amd_iommu_unmap
,
2845 .iova_to_phys
= amd_iommu_iova_to_phys
,
2846 .domain_has_cap
= amd_iommu_domain_has_cap
,
2849 /*****************************************************************************
2851 * The next functions do a basic initialization of IOMMU for pass through
2854 * In passthrough mode the IOMMU is initialized and enabled but not used for
2855 * DMA-API translation.
2857 *****************************************************************************/
2859 int __init
amd_iommu_init_passthrough(void)
2861 struct iommu_dev_data
*dev_data
;
2862 struct pci_dev
*dev
= NULL
;
2863 struct amd_iommu
*iommu
;
2867 ret
= alloc_passthrough_domain();
2871 for_each_pci_dev(dev
) {
2872 if (!check_device(&dev
->dev
))
2875 dev_data
= get_dev_data(&dev
->dev
);
2876 dev_data
->passthrough
= true;
2878 devid
= get_device_id(&dev
->dev
);
2880 iommu
= amd_iommu_rlookup_table
[devid
];
2884 attach_device(&dev
->dev
, pt_domain
);
2887 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");