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 <asm/proto.h>
31 #include <asm/iommu.h>
33 #include <asm/amd_iommu_proto.h>
34 #include <asm/amd_iommu_types.h>
35 #include <asm/amd_iommu.h>
37 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
39 #define LOOP_TIMEOUT 100000
41 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
43 /* A list of preallocated protection domains */
44 static LIST_HEAD(iommu_pd_list
);
45 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
48 * Domain for untranslated devices - only allocated
49 * if iommu=pt passed on kernel cmd line.
51 static struct protection_domain
*pt_domain
;
53 static struct iommu_ops amd_iommu_ops
;
56 * general struct to manage commands send to an IOMMU
62 static void update_domain(struct protection_domain
*domain
);
64 /****************************************************************************
68 ****************************************************************************/
70 static inline u16
get_device_id(struct device
*dev
)
72 struct pci_dev
*pdev
= to_pci_dev(dev
);
74 return calc_devid(pdev
->bus
->number
, pdev
->devfn
);
77 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
79 return dev
->archdata
.iommu
;
83 * In this function the list of preallocated protection domains is traversed to
84 * find the domain for a specific device
86 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
88 struct dma_ops_domain
*entry
, *ret
= NULL
;
90 u16 alias
= amd_iommu_alias_table
[devid
];
92 if (list_empty(&iommu_pd_list
))
95 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
97 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
98 if (entry
->target_dev
== devid
||
99 entry
->target_dev
== alias
) {
105 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
111 * This function checks if the driver got a valid device from the caller to
112 * avoid dereferencing invalid pointers.
114 static bool check_device(struct device
*dev
)
118 if (!dev
|| !dev
->dma_mask
)
121 /* No device or no PCI device */
122 if (dev
->bus
!= &pci_bus_type
)
125 devid
= get_device_id(dev
);
127 /* Out of our scope? */
128 if (devid
> amd_iommu_last_bdf
)
131 if (amd_iommu_rlookup_table
[devid
] == NULL
)
137 static int iommu_init_device(struct device
*dev
)
139 struct iommu_dev_data
*dev_data
;
140 struct pci_dev
*pdev
;
143 if (dev
->archdata
.iommu
)
146 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
152 devid
= get_device_id(dev
);
153 alias
= amd_iommu_alias_table
[devid
];
154 pdev
= pci_get_bus_and_slot(PCI_BUS(alias
), alias
& 0xff);
156 dev_data
->alias
= &pdev
->dev
;
158 atomic_set(&dev_data
->bind
, 0);
160 dev
->archdata
.iommu
= dev_data
;
166 static void iommu_uninit_device(struct device
*dev
)
168 kfree(dev
->archdata
.iommu
);
171 void __init
amd_iommu_uninit_devices(void)
173 struct pci_dev
*pdev
= NULL
;
175 for_each_pci_dev(pdev
) {
177 if (!check_device(&pdev
->dev
))
180 iommu_uninit_device(&pdev
->dev
);
184 int __init
amd_iommu_init_devices(void)
186 struct pci_dev
*pdev
= NULL
;
189 for_each_pci_dev(pdev
) {
191 if (!check_device(&pdev
->dev
))
194 ret
= iommu_init_device(&pdev
->dev
);
203 amd_iommu_uninit_devices();
207 #ifdef CONFIG_AMD_IOMMU_STATS
210 * Initialization code for statistics collection
213 DECLARE_STATS_COUNTER(compl_wait
);
214 DECLARE_STATS_COUNTER(cnt_map_single
);
215 DECLARE_STATS_COUNTER(cnt_unmap_single
);
216 DECLARE_STATS_COUNTER(cnt_map_sg
);
217 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
218 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
219 DECLARE_STATS_COUNTER(cnt_free_coherent
);
220 DECLARE_STATS_COUNTER(cross_page
);
221 DECLARE_STATS_COUNTER(domain_flush_single
);
222 DECLARE_STATS_COUNTER(domain_flush_all
);
223 DECLARE_STATS_COUNTER(alloced_io_mem
);
224 DECLARE_STATS_COUNTER(total_map_requests
);
226 static struct dentry
*stats_dir
;
227 static struct dentry
*de_fflush
;
229 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
231 if (stats_dir
== NULL
)
234 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
238 static void amd_iommu_stats_init(void)
240 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
241 if (stats_dir
== NULL
)
244 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
245 (u32
*)&amd_iommu_unmap_flush
);
247 amd_iommu_stats_add(&compl_wait
);
248 amd_iommu_stats_add(&cnt_map_single
);
249 amd_iommu_stats_add(&cnt_unmap_single
);
250 amd_iommu_stats_add(&cnt_map_sg
);
251 amd_iommu_stats_add(&cnt_unmap_sg
);
252 amd_iommu_stats_add(&cnt_alloc_coherent
);
253 amd_iommu_stats_add(&cnt_free_coherent
);
254 amd_iommu_stats_add(&cross_page
);
255 amd_iommu_stats_add(&domain_flush_single
);
256 amd_iommu_stats_add(&domain_flush_all
);
257 amd_iommu_stats_add(&alloced_io_mem
);
258 amd_iommu_stats_add(&total_map_requests
);
263 /****************************************************************************
265 * Interrupt handling functions
267 ****************************************************************************/
269 static void dump_dte_entry(u16 devid
)
273 for (i
= 0; i
< 8; ++i
)
274 pr_err("AMD-Vi: DTE[%d]: %08x\n", i
,
275 amd_iommu_dev_table
[devid
].data
[i
]);
278 static void dump_command(unsigned long phys_addr
)
280 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
283 for (i
= 0; i
< 4; ++i
)
284 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
287 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
290 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
291 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
292 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
293 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
294 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
296 printk(KERN_ERR
"AMD-Vi: Event logged [");
299 case EVENT_TYPE_ILL_DEV
:
300 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
301 "address=0x%016llx flags=0x%04x]\n",
302 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
304 dump_dte_entry(devid
);
306 case EVENT_TYPE_IO_FAULT
:
307 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
308 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
309 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
310 domid
, address
, flags
);
312 case EVENT_TYPE_DEV_TAB_ERR
:
313 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
314 "address=0x%016llx flags=0x%04x]\n",
315 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
318 case EVENT_TYPE_PAGE_TAB_ERR
:
319 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
320 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
321 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
322 domid
, address
, flags
);
324 case EVENT_TYPE_ILL_CMD
:
325 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
326 dump_command(address
);
328 case EVENT_TYPE_CMD_HARD_ERR
:
329 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
330 "flags=0x%04x]\n", address
, flags
);
332 case EVENT_TYPE_IOTLB_INV_TO
:
333 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
334 "address=0x%016llx]\n",
335 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
338 case EVENT_TYPE_INV_DEV_REQ
:
339 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
340 "address=0x%016llx flags=0x%04x]\n",
341 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
345 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
349 static void iommu_poll_events(struct amd_iommu
*iommu
)
354 spin_lock_irqsave(&iommu
->lock
, flags
);
356 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
357 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
359 while (head
!= tail
) {
360 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
361 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
364 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
366 spin_unlock_irqrestore(&iommu
->lock
, flags
);
369 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
371 struct amd_iommu
*iommu
;
373 for_each_iommu(iommu
)
374 iommu_poll_events(iommu
);
379 /****************************************************************************
381 * IOMMU command queuing functions
383 ****************************************************************************/
385 static int wait_on_sem(volatile u64
*sem
)
389 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
394 if (i
== LOOP_TIMEOUT
) {
395 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
402 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
403 struct iommu_cmd
*cmd
,
408 target
= iommu
->cmd_buf
+ tail
;
409 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
411 /* Copy command to buffer */
412 memcpy(target
, cmd
, sizeof(*cmd
));
414 /* Tell the IOMMU about it */
415 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
418 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
420 WARN_ON(address
& 0x7ULL
);
422 memset(cmd
, 0, sizeof(*cmd
));
423 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
424 cmd
->data
[1] = upper_32_bits(__pa(address
));
426 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
429 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
431 memset(cmd
, 0, sizeof(*cmd
));
432 cmd
->data
[0] = devid
;
433 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
436 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
437 size_t size
, u16 domid
, int pde
)
442 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
447 * If we have to flush more than one page, flush all
448 * TLB entries for this domain
450 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
454 address
&= PAGE_MASK
;
456 memset(cmd
, 0, sizeof(*cmd
));
457 cmd
->data
[1] |= domid
;
458 cmd
->data
[2] = lower_32_bits(address
);
459 cmd
->data
[3] = upper_32_bits(address
);
460 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
461 if (s
) /* size bit - we flush more than one 4kb page */
462 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
463 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
464 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
467 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
468 u64 address
, size_t size
)
473 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
478 * If we have to flush more than one page, flush all
479 * TLB entries for this domain
481 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
485 address
&= PAGE_MASK
;
487 memset(cmd
, 0, sizeof(*cmd
));
488 cmd
->data
[0] = devid
;
489 cmd
->data
[0] |= (qdep
& 0xff) << 24;
490 cmd
->data
[1] = devid
;
491 cmd
->data
[2] = lower_32_bits(address
);
492 cmd
->data
[3] = upper_32_bits(address
);
493 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
495 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
498 static void build_inv_all(struct iommu_cmd
*cmd
)
500 memset(cmd
, 0, sizeof(*cmd
));
501 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
505 * Writes the command to the IOMMUs command buffer and informs the
506 * hardware about the new command.
508 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
510 u32 left
, tail
, head
, next_tail
;
513 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
516 spin_lock_irqsave(&iommu
->lock
, flags
);
518 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
519 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
520 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
521 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
524 struct iommu_cmd sync_cmd
;
525 volatile u64 sem
= 0;
528 build_completion_wait(&sync_cmd
, (u64
)&sem
);
529 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
531 spin_unlock_irqrestore(&iommu
->lock
, flags
);
533 if ((ret
= wait_on_sem(&sem
)) != 0)
539 copy_cmd_to_buffer(iommu
, cmd
, tail
);
541 /* We need to sync now to make sure all commands are processed */
542 iommu
->need_sync
= true;
544 spin_unlock_irqrestore(&iommu
->lock
, flags
);
550 * This function queues a completion wait command into the command
553 static int iommu_completion_wait(struct amd_iommu
*iommu
)
555 struct iommu_cmd cmd
;
556 volatile u64 sem
= 0;
559 if (!iommu
->need_sync
)
562 build_completion_wait(&cmd
, (u64
)&sem
);
564 ret
= iommu_queue_command(iommu
, &cmd
);
568 return wait_on_sem(&sem
);
571 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
573 struct iommu_cmd cmd
;
575 build_inv_dte(&cmd
, devid
);
577 return iommu_queue_command(iommu
, &cmd
);
580 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
584 for (devid
= 0; devid
<= 0xffff; ++devid
)
585 iommu_flush_dte(iommu
, devid
);
587 iommu_completion_wait(iommu
);
591 * This function uses heavy locking and may disable irqs for some time. But
592 * this is no issue because it is only called during resume.
594 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
598 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
599 struct iommu_cmd cmd
;
600 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
602 iommu_queue_command(iommu
, &cmd
);
605 iommu_completion_wait(iommu
);
608 static void iommu_flush_all(struct amd_iommu
*iommu
)
610 struct iommu_cmd cmd
;
614 iommu_queue_command(iommu
, &cmd
);
615 iommu_completion_wait(iommu
);
618 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
620 if (iommu_feature(iommu
, FEATURE_IA
)) {
621 iommu_flush_all(iommu
);
623 iommu_flush_dte_all(iommu
);
624 iommu_flush_tlb_all(iommu
);
629 * Command send function for flushing on-device TLB
631 static int device_flush_iotlb(struct device
*dev
, u64 address
, size_t size
)
633 struct pci_dev
*pdev
= to_pci_dev(dev
);
634 struct amd_iommu
*iommu
;
635 struct iommu_cmd cmd
;
639 qdep
= pci_ats_queue_depth(pdev
);
640 devid
= get_device_id(dev
);
641 iommu
= amd_iommu_rlookup_table
[devid
];
643 build_inv_iotlb_pages(&cmd
, devid
, qdep
, address
, size
);
645 return iommu_queue_command(iommu
, &cmd
);
649 * Command send function for invalidating a device table entry
651 static int device_flush_dte(struct device
*dev
)
653 struct amd_iommu
*iommu
;
654 struct pci_dev
*pdev
;
658 pdev
= to_pci_dev(dev
);
659 devid
= get_device_id(dev
);
660 iommu
= amd_iommu_rlookup_table
[devid
];
662 ret
= iommu_flush_dte(iommu
, devid
);
666 if (pci_ats_enabled(pdev
))
667 ret
= device_flush_iotlb(dev
, 0, ~0UL);
673 * TLB invalidation function which is called from the mapping functions.
674 * It invalidates a single PTE if the range to flush is within a single
675 * page. Otherwise it flushes the whole TLB of the IOMMU.
677 static void __domain_flush_pages(struct protection_domain
*domain
,
678 u64 address
, size_t size
, int pde
)
680 struct iommu_dev_data
*dev_data
;
681 struct iommu_cmd cmd
;
684 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
686 for (i
= 0; i
< amd_iommus_present
; ++i
) {
687 if (!domain
->dev_iommu
[i
])
691 * Devices of this domain are behind this IOMMU
692 * We need a TLB flush
694 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
697 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
698 struct pci_dev
*pdev
= to_pci_dev(dev_data
->dev
);
700 if (!pci_ats_enabled(pdev
))
703 ret
|= device_flush_iotlb(dev_data
->dev
, address
, size
);
709 static void domain_flush_pages(struct protection_domain
*domain
,
710 u64 address
, size_t size
)
712 __domain_flush_pages(domain
, address
, size
, 0);
715 /* Flush the whole IO/TLB for a given protection domain */
716 static void domain_flush_tlb(struct protection_domain
*domain
)
718 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
721 /* Flush the whole IO/TLB for a given protection domain - including PDE */
722 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
724 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
727 static void domain_flush_complete(struct protection_domain
*domain
)
731 for (i
= 0; i
< amd_iommus_present
; ++i
) {
732 if (!domain
->dev_iommu
[i
])
736 * Devices of this domain are behind this IOMMU
737 * We need to wait for completion of all commands.
739 iommu_completion_wait(amd_iommus
[i
]);
745 * This function flushes the DTEs for all devices in domain
747 static void domain_flush_devices(struct protection_domain
*domain
)
749 struct iommu_dev_data
*dev_data
;
752 spin_lock_irqsave(&domain
->lock
, flags
);
754 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
755 device_flush_dte(dev_data
->dev
);
757 spin_unlock_irqrestore(&domain
->lock
, flags
);
760 /****************************************************************************
762 * The functions below are used the create the page table mappings for
763 * unity mapped regions.
765 ****************************************************************************/
768 * This function is used to add another level to an IO page table. Adding
769 * another level increases the size of the address space by 9 bits to a size up
772 static bool increase_address_space(struct protection_domain
*domain
,
777 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
778 /* address space already 64 bit large */
781 pte
= (void *)get_zeroed_page(gfp
);
785 *pte
= PM_LEVEL_PDE(domain
->mode
,
786 virt_to_phys(domain
->pt_root
));
787 domain
->pt_root
= pte
;
789 domain
->updated
= true;
794 static u64
*alloc_pte(struct protection_domain
*domain
,
795 unsigned long address
,
796 unsigned long page_size
,
803 BUG_ON(!is_power_of_2(page_size
));
805 while (address
> PM_LEVEL_SIZE(domain
->mode
))
806 increase_address_space(domain
, gfp
);
808 level
= domain
->mode
- 1;
809 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
810 address
= PAGE_SIZE_ALIGN(address
, page_size
);
811 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
813 while (level
> end_lvl
) {
814 if (!IOMMU_PTE_PRESENT(*pte
)) {
815 page
= (u64
*)get_zeroed_page(gfp
);
818 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
821 /* No level skipping support yet */
822 if (PM_PTE_LEVEL(*pte
) != level
)
827 pte
= IOMMU_PTE_PAGE(*pte
);
829 if (pte_page
&& level
== end_lvl
)
832 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
839 * This function checks if there is a PTE for a given dma address. If
840 * there is one, it returns the pointer to it.
842 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
847 if (address
> PM_LEVEL_SIZE(domain
->mode
))
850 level
= domain
->mode
- 1;
851 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
856 if (!IOMMU_PTE_PRESENT(*pte
))
860 if (PM_PTE_LEVEL(*pte
) == 0x07) {
861 unsigned long pte_mask
, __pte
;
864 * If we have a series of large PTEs, make
865 * sure to return a pointer to the first one.
867 pte_mask
= PTE_PAGE_SIZE(*pte
);
868 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
869 __pte
= ((unsigned long)pte
) & pte_mask
;
874 /* No level skipping support yet */
875 if (PM_PTE_LEVEL(*pte
) != level
)
880 /* Walk to the next level */
881 pte
= IOMMU_PTE_PAGE(*pte
);
882 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
889 * Generic mapping functions. It maps a physical address into a DMA
890 * address space. It allocates the page table pages if necessary.
891 * In the future it can be extended to a generic mapping function
892 * supporting all features of AMD IOMMU page tables like level skipping
893 * and full 64 bit address spaces.
895 static int iommu_map_page(struct protection_domain
*dom
,
896 unsigned long bus_addr
,
897 unsigned long phys_addr
,
899 unsigned long page_size
)
904 if (!(prot
& IOMMU_PROT_MASK
))
907 bus_addr
= PAGE_ALIGN(bus_addr
);
908 phys_addr
= PAGE_ALIGN(phys_addr
);
909 count
= PAGE_SIZE_PTE_COUNT(page_size
);
910 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
912 for (i
= 0; i
< count
; ++i
)
913 if (IOMMU_PTE_PRESENT(pte
[i
]))
916 if (page_size
> PAGE_SIZE
) {
917 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
918 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
920 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
922 if (prot
& IOMMU_PROT_IR
)
923 __pte
|= IOMMU_PTE_IR
;
924 if (prot
& IOMMU_PROT_IW
)
925 __pte
|= IOMMU_PTE_IW
;
927 for (i
= 0; i
< count
; ++i
)
935 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
936 unsigned long bus_addr
,
937 unsigned long page_size
)
939 unsigned long long unmap_size
, unmapped
;
942 BUG_ON(!is_power_of_2(page_size
));
946 while (unmapped
< page_size
) {
948 pte
= fetch_pte(dom
, bus_addr
);
952 * No PTE for this address
953 * move forward in 4kb steps
955 unmap_size
= PAGE_SIZE
;
956 } else if (PM_PTE_LEVEL(*pte
) == 0) {
957 /* 4kb PTE found for this address */
958 unmap_size
= PAGE_SIZE
;
963 /* Large PTE found which maps this address */
964 unmap_size
= PTE_PAGE_SIZE(*pte
);
965 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
966 for (i
= 0; i
< count
; i
++)
970 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
971 unmapped
+= unmap_size
;
974 BUG_ON(!is_power_of_2(unmapped
));
980 * This function checks if a specific unity mapping entry is needed for
981 * this specific IOMMU.
983 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
984 struct unity_map_entry
*entry
)
988 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
989 bdf
= amd_iommu_alias_table
[i
];
990 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
998 * This function actually applies the mapping to the page table of the
1001 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1002 struct unity_map_entry
*e
)
1007 for (addr
= e
->address_start
; addr
< e
->address_end
;
1008 addr
+= PAGE_SIZE
) {
1009 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1014 * if unity mapping is in aperture range mark the page
1015 * as allocated in the aperture
1017 if (addr
< dma_dom
->aperture_size
)
1018 __set_bit(addr
>> PAGE_SHIFT
,
1019 dma_dom
->aperture
[0]->bitmap
);
1026 * Init the unity mappings for a specific IOMMU in the system
1028 * Basically iterates over all unity mapping entries and applies them to
1029 * the default domain DMA of that IOMMU if necessary.
1031 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1033 struct unity_map_entry
*entry
;
1036 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1037 if (!iommu_for_unity_map(iommu
, entry
))
1039 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1048 * Inits the unity mappings required for a specific device
1050 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1053 struct unity_map_entry
*e
;
1056 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1057 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1059 ret
= dma_ops_unity_map(dma_dom
, e
);
1067 /****************************************************************************
1069 * The next functions belong to the address allocator for the dma_ops
1070 * interface functions. They work like the allocators in the other IOMMU
1071 * drivers. Its basically a bitmap which marks the allocated pages in
1072 * the aperture. Maybe it could be enhanced in the future to a more
1073 * efficient allocator.
1075 ****************************************************************************/
1078 * The address allocator core functions.
1080 * called with domain->lock held
1084 * Used to reserve address ranges in the aperture (e.g. for exclusion
1087 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1088 unsigned long start_page
,
1091 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1093 if (start_page
+ pages
> last_page
)
1094 pages
= last_page
- start_page
;
1096 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1097 int index
= i
/ APERTURE_RANGE_PAGES
;
1098 int page
= i
% APERTURE_RANGE_PAGES
;
1099 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1104 * This function is used to add a new aperture range to an existing
1105 * aperture in case of dma_ops domain allocation or address allocation
1108 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1109 bool populate
, gfp_t gfp
)
1111 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1112 struct amd_iommu
*iommu
;
1115 #ifdef CONFIG_IOMMU_STRESS
1119 if (index
>= APERTURE_MAX_RANGES
)
1122 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1123 if (!dma_dom
->aperture
[index
])
1126 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1127 if (!dma_dom
->aperture
[index
]->bitmap
)
1130 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1133 unsigned long address
= dma_dom
->aperture_size
;
1134 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1135 u64
*pte
, *pte_page
;
1137 for (i
= 0; i
< num_ptes
; ++i
) {
1138 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1143 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1145 address
+= APERTURE_RANGE_SIZE
/ 64;
1149 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1151 /* Initialize the exclusion range if necessary */
1152 for_each_iommu(iommu
) {
1153 if (iommu
->exclusion_start
&&
1154 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1155 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1156 unsigned long startpage
;
1157 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1158 iommu
->exclusion_length
,
1160 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1161 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1166 * Check for areas already mapped as present in the new aperture
1167 * range and mark those pages as reserved in the allocator. Such
1168 * mappings may already exist as a result of requested unity
1169 * mappings for devices.
1171 for (i
= dma_dom
->aperture
[index
]->offset
;
1172 i
< dma_dom
->aperture_size
;
1174 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1175 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1178 dma_ops_reserve_addresses(dma_dom
, i
<< PAGE_SHIFT
, 1);
1181 update_domain(&dma_dom
->domain
);
1186 update_domain(&dma_dom
->domain
);
1188 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1190 kfree(dma_dom
->aperture
[index
]);
1191 dma_dom
->aperture
[index
] = NULL
;
1196 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1197 struct dma_ops_domain
*dom
,
1199 unsigned long align_mask
,
1201 unsigned long start
)
1203 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1204 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1205 int i
= start
>> APERTURE_RANGE_SHIFT
;
1206 unsigned long boundary_size
;
1207 unsigned long address
= -1;
1208 unsigned long limit
;
1210 next_bit
>>= PAGE_SHIFT
;
1212 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1213 PAGE_SIZE
) >> PAGE_SHIFT
;
1215 for (;i
< max_index
; ++i
) {
1216 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1218 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1221 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1222 dma_mask
>> PAGE_SHIFT
);
1224 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1225 limit
, next_bit
, pages
, 0,
1226 boundary_size
, align_mask
);
1227 if (address
!= -1) {
1228 address
= dom
->aperture
[i
]->offset
+
1229 (address
<< PAGE_SHIFT
);
1230 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1240 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1241 struct dma_ops_domain
*dom
,
1243 unsigned long align_mask
,
1246 unsigned long address
;
1248 #ifdef CONFIG_IOMMU_STRESS
1249 dom
->next_address
= 0;
1250 dom
->need_flush
= true;
1253 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1254 dma_mask
, dom
->next_address
);
1256 if (address
== -1) {
1257 dom
->next_address
= 0;
1258 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1260 dom
->need_flush
= true;
1263 if (unlikely(address
== -1))
1264 address
= DMA_ERROR_CODE
;
1266 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1272 * The address free function.
1274 * called with domain->lock held
1276 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1277 unsigned long address
,
1280 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1281 struct aperture_range
*range
= dom
->aperture
[i
];
1283 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1285 #ifdef CONFIG_IOMMU_STRESS
1290 if (address
>= dom
->next_address
)
1291 dom
->need_flush
= true;
1293 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1295 bitmap_clear(range
->bitmap
, address
, pages
);
1299 /****************************************************************************
1301 * The next functions belong to the domain allocation. A domain is
1302 * allocated for every IOMMU as the default domain. If device isolation
1303 * is enabled, every device get its own domain. The most important thing
1304 * about domains is the page table mapping the DMA address space they
1307 ****************************************************************************/
1310 * This function adds a protection domain to the global protection domain list
1312 static void add_domain_to_list(struct protection_domain
*domain
)
1314 unsigned long flags
;
1316 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1317 list_add(&domain
->list
, &amd_iommu_pd_list
);
1318 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1322 * This function removes a protection domain to the global
1323 * protection domain list
1325 static void del_domain_from_list(struct protection_domain
*domain
)
1327 unsigned long flags
;
1329 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1330 list_del(&domain
->list
);
1331 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1334 static u16
domain_id_alloc(void)
1336 unsigned long flags
;
1339 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1340 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1342 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1343 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1346 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1351 static void domain_id_free(int id
)
1353 unsigned long flags
;
1355 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1356 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1357 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1358 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1361 static void free_pagetable(struct protection_domain
*domain
)
1366 p1
= domain
->pt_root
;
1371 for (i
= 0; i
< 512; ++i
) {
1372 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1375 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1376 for (j
= 0; j
< 512; ++j
) {
1377 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1379 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1380 free_page((unsigned long)p3
);
1383 free_page((unsigned long)p2
);
1386 free_page((unsigned long)p1
);
1388 domain
->pt_root
= NULL
;
1392 * Free a domain, only used if something went wrong in the
1393 * allocation path and we need to free an already allocated page table
1395 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1402 del_domain_from_list(&dom
->domain
);
1404 free_pagetable(&dom
->domain
);
1406 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1407 if (!dom
->aperture
[i
])
1409 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1410 kfree(dom
->aperture
[i
]);
1417 * Allocates a new protection domain usable for the dma_ops functions.
1418 * It also initializes the page table and the address allocator data
1419 * structures required for the dma_ops interface
1421 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1423 struct dma_ops_domain
*dma_dom
;
1425 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1429 spin_lock_init(&dma_dom
->domain
.lock
);
1431 dma_dom
->domain
.id
= domain_id_alloc();
1432 if (dma_dom
->domain
.id
== 0)
1434 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1435 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1436 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1437 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1438 dma_dom
->domain
.priv
= dma_dom
;
1439 if (!dma_dom
->domain
.pt_root
)
1442 dma_dom
->need_flush
= false;
1443 dma_dom
->target_dev
= 0xffff;
1445 add_domain_to_list(&dma_dom
->domain
);
1447 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1451 * mark the first page as allocated so we never return 0 as
1452 * a valid dma-address. So we can use 0 as error value
1454 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1455 dma_dom
->next_address
= 0;
1461 dma_ops_domain_free(dma_dom
);
1467 * little helper function to check whether a given protection domain is a
1470 static bool dma_ops_domain(struct protection_domain
*domain
)
1472 return domain
->flags
& PD_DMA_OPS_MASK
;
1475 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1477 u64 pte_root
= virt_to_phys(domain
->pt_root
);
1480 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1481 << DEV_ENTRY_MODE_SHIFT
;
1482 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1485 flags
|= DTE_FLAG_IOTLB
;
1487 amd_iommu_dev_table
[devid
].data
[3] |= flags
;
1488 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
1489 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
1490 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
1493 static void clear_dte_entry(u16 devid
)
1495 /* remove entry from the device table seen by the hardware */
1496 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1497 amd_iommu_dev_table
[devid
].data
[1] = 0;
1498 amd_iommu_dev_table
[devid
].data
[2] = 0;
1500 amd_iommu_apply_erratum_63(devid
);
1503 static void do_attach(struct device
*dev
, struct protection_domain
*domain
)
1505 struct iommu_dev_data
*dev_data
;
1506 struct amd_iommu
*iommu
;
1507 struct pci_dev
*pdev
;
1511 devid
= get_device_id(dev
);
1512 iommu
= amd_iommu_rlookup_table
[devid
];
1513 dev_data
= get_dev_data(dev
);
1514 pdev
= to_pci_dev(dev
);
1516 if (amd_iommu_iotlb_sup
)
1517 ats
= pci_ats_enabled(pdev
);
1519 /* Update data structures */
1520 dev_data
->domain
= domain
;
1521 list_add(&dev_data
->list
, &domain
->dev_list
);
1522 set_dte_entry(devid
, domain
, ats
);
1524 /* Do reference counting */
1525 domain
->dev_iommu
[iommu
->index
] += 1;
1526 domain
->dev_cnt
+= 1;
1528 /* Flush the DTE entry */
1529 device_flush_dte(dev
);
1532 static void do_detach(struct device
*dev
)
1534 struct iommu_dev_data
*dev_data
;
1535 struct amd_iommu
*iommu
;
1536 struct pci_dev
*pdev
;
1539 devid
= get_device_id(dev
);
1540 iommu
= amd_iommu_rlookup_table
[devid
];
1541 dev_data
= get_dev_data(dev
);
1542 pdev
= to_pci_dev(dev
);
1544 /* decrease reference counters */
1545 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1546 dev_data
->domain
->dev_cnt
-= 1;
1548 /* Update data structures */
1549 dev_data
->domain
= NULL
;
1550 list_del(&dev_data
->list
);
1551 clear_dte_entry(devid
);
1553 /* Flush the DTE entry */
1554 device_flush_dte(dev
);
1558 * If a device is not yet associated with a domain, this function does
1559 * assigns it visible for the hardware
1561 static int __attach_device(struct device
*dev
,
1562 struct protection_domain
*domain
)
1564 struct iommu_dev_data
*dev_data
, *alias_data
;
1567 dev_data
= get_dev_data(dev
);
1568 alias_data
= get_dev_data(dev_data
->alias
);
1574 spin_lock(&domain
->lock
);
1576 /* Some sanity checks */
1578 if (alias_data
->domain
!= NULL
&&
1579 alias_data
->domain
!= domain
)
1582 if (dev_data
->domain
!= NULL
&&
1583 dev_data
->domain
!= domain
)
1586 /* Do real assignment */
1587 if (dev_data
->alias
!= dev
) {
1588 alias_data
= get_dev_data(dev_data
->alias
);
1589 if (alias_data
->domain
== NULL
)
1590 do_attach(dev_data
->alias
, domain
);
1592 atomic_inc(&alias_data
->bind
);
1595 if (dev_data
->domain
== NULL
)
1596 do_attach(dev
, domain
);
1598 atomic_inc(&dev_data
->bind
);
1605 spin_unlock(&domain
->lock
);
1611 * If a device is not yet associated with a domain, this function does
1612 * assigns it visible for the hardware
1614 static int attach_device(struct device
*dev
,
1615 struct protection_domain
*domain
)
1617 struct pci_dev
*pdev
= to_pci_dev(dev
);
1618 unsigned long flags
;
1621 if (amd_iommu_iotlb_sup
)
1622 pci_enable_ats(pdev
, PAGE_SHIFT
);
1624 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1625 ret
= __attach_device(dev
, domain
);
1626 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1629 * We might boot into a crash-kernel here. The crashed kernel
1630 * left the caches in the IOMMU dirty. So we have to flush
1631 * here to evict all dirty stuff.
1633 domain_flush_tlb_pde(domain
);
1639 * Removes a device from a protection domain (unlocked)
1641 static void __detach_device(struct device
*dev
)
1643 struct iommu_dev_data
*dev_data
= get_dev_data(dev
);
1644 struct iommu_dev_data
*alias_data
;
1645 struct protection_domain
*domain
;
1646 unsigned long flags
;
1648 BUG_ON(!dev_data
->domain
);
1650 domain
= dev_data
->domain
;
1652 spin_lock_irqsave(&domain
->lock
, flags
);
1654 if (dev_data
->alias
!= dev
) {
1655 alias_data
= get_dev_data(dev_data
->alias
);
1656 if (atomic_dec_and_test(&alias_data
->bind
))
1657 do_detach(dev_data
->alias
);
1660 if (atomic_dec_and_test(&dev_data
->bind
))
1663 spin_unlock_irqrestore(&domain
->lock
, flags
);
1666 * If we run in passthrough mode the device must be assigned to the
1667 * passthrough domain if it is detached from any other domain.
1668 * Make sure we can deassign from the pt_domain itself.
1670 if (iommu_pass_through
&&
1671 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
1672 __attach_device(dev
, pt_domain
);
1676 * Removes a device from a protection domain (with devtable_lock held)
1678 static void detach_device(struct device
*dev
)
1680 struct pci_dev
*pdev
= to_pci_dev(dev
);
1681 unsigned long flags
;
1683 /* lock device table */
1684 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1685 __detach_device(dev
);
1686 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1688 if (amd_iommu_iotlb_sup
&& pci_ats_enabled(pdev
))
1689 pci_disable_ats(pdev
);
1693 * Find out the protection domain structure for a given PCI device. This
1694 * will give us the pointer to the page table root for example.
1696 static struct protection_domain
*domain_for_device(struct device
*dev
)
1698 struct protection_domain
*dom
;
1699 struct iommu_dev_data
*dev_data
, *alias_data
;
1700 unsigned long flags
;
1703 devid
= get_device_id(dev
);
1704 alias
= amd_iommu_alias_table
[devid
];
1705 dev_data
= get_dev_data(dev
);
1706 alias_data
= get_dev_data(dev_data
->alias
);
1710 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1711 dom
= dev_data
->domain
;
1713 alias_data
->domain
!= NULL
) {
1714 __attach_device(dev
, alias_data
->domain
);
1715 dom
= alias_data
->domain
;
1718 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1723 static int device_change_notifier(struct notifier_block
*nb
,
1724 unsigned long action
, void *data
)
1726 struct device
*dev
= data
;
1728 struct protection_domain
*domain
;
1729 struct dma_ops_domain
*dma_domain
;
1730 struct amd_iommu
*iommu
;
1731 unsigned long flags
;
1733 if (!check_device(dev
))
1736 devid
= get_device_id(dev
);
1737 iommu
= amd_iommu_rlookup_table
[devid
];
1740 case BUS_NOTIFY_UNBOUND_DRIVER
:
1742 domain
= domain_for_device(dev
);
1746 if (iommu_pass_through
)
1750 case BUS_NOTIFY_ADD_DEVICE
:
1752 iommu_init_device(dev
);
1754 domain
= domain_for_device(dev
);
1756 /* allocate a protection domain if a device is added */
1757 dma_domain
= find_protection_domain(devid
);
1760 dma_domain
= dma_ops_domain_alloc();
1763 dma_domain
->target_dev
= devid
;
1765 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1766 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1767 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1770 case BUS_NOTIFY_DEL_DEVICE
:
1772 iommu_uninit_device(dev
);
1778 device_flush_dte(dev
);
1779 iommu_completion_wait(iommu
);
1785 static struct notifier_block device_nb
= {
1786 .notifier_call
= device_change_notifier
,
1789 void amd_iommu_init_notifier(void)
1791 bus_register_notifier(&pci_bus_type
, &device_nb
);
1794 /*****************************************************************************
1796 * The next functions belong to the dma_ops mapping/unmapping code.
1798 *****************************************************************************/
1801 * In the dma_ops path we only have the struct device. This function
1802 * finds the corresponding IOMMU, the protection domain and the
1803 * requestor id for a given device.
1804 * If the device is not yet associated with a domain this is also done
1807 static struct protection_domain
*get_domain(struct device
*dev
)
1809 struct protection_domain
*domain
;
1810 struct dma_ops_domain
*dma_dom
;
1811 u16 devid
= get_device_id(dev
);
1813 if (!check_device(dev
))
1814 return ERR_PTR(-EINVAL
);
1816 domain
= domain_for_device(dev
);
1817 if (domain
!= NULL
&& !dma_ops_domain(domain
))
1818 return ERR_PTR(-EBUSY
);
1823 /* Device not bount yet - bind it */
1824 dma_dom
= find_protection_domain(devid
);
1826 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
1827 attach_device(dev
, &dma_dom
->domain
);
1828 DUMP_printk("Using protection domain %d for device %s\n",
1829 dma_dom
->domain
.id
, dev_name(dev
));
1831 return &dma_dom
->domain
;
1834 static void update_device_table(struct protection_domain
*domain
)
1836 struct iommu_dev_data
*dev_data
;
1838 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1839 struct pci_dev
*pdev
= to_pci_dev(dev_data
->dev
);
1840 u16 devid
= get_device_id(dev_data
->dev
);
1841 set_dte_entry(devid
, domain
, pci_ats_enabled(pdev
));
1845 static void update_domain(struct protection_domain
*domain
)
1847 if (!domain
->updated
)
1850 update_device_table(domain
);
1852 domain_flush_devices(domain
);
1853 domain_flush_tlb_pde(domain
);
1855 domain
->updated
= false;
1859 * This function fetches the PTE for a given address in the aperture
1861 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
1862 unsigned long address
)
1864 struct aperture_range
*aperture
;
1865 u64
*pte
, *pte_page
;
1867 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1871 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1873 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
1875 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
1877 pte
+= PM_LEVEL_INDEX(0, address
);
1879 update_domain(&dom
->domain
);
1885 * This is the generic map function. It maps one 4kb page at paddr to
1886 * the given address in the DMA address space for the domain.
1888 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
1889 unsigned long address
,
1895 WARN_ON(address
> dom
->aperture_size
);
1899 pte
= dma_ops_get_pte(dom
, address
);
1901 return DMA_ERROR_CODE
;
1903 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1905 if (direction
== DMA_TO_DEVICE
)
1906 __pte
|= IOMMU_PTE_IR
;
1907 else if (direction
== DMA_FROM_DEVICE
)
1908 __pte
|= IOMMU_PTE_IW
;
1909 else if (direction
== DMA_BIDIRECTIONAL
)
1910 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1916 return (dma_addr_t
)address
;
1920 * The generic unmapping function for on page in the DMA address space.
1922 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
1923 unsigned long address
)
1925 struct aperture_range
*aperture
;
1928 if (address
>= dom
->aperture_size
)
1931 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1935 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1939 pte
+= PM_LEVEL_INDEX(0, address
);
1947 * This function contains common code for mapping of a physically
1948 * contiguous memory region into DMA address space. It is used by all
1949 * mapping functions provided with this IOMMU driver.
1950 * Must be called with the domain lock held.
1952 static dma_addr_t
__map_single(struct device
*dev
,
1953 struct dma_ops_domain
*dma_dom
,
1960 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1961 dma_addr_t address
, start
, ret
;
1963 unsigned long align_mask
= 0;
1966 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1969 INC_STATS_COUNTER(total_map_requests
);
1972 INC_STATS_COUNTER(cross_page
);
1975 align_mask
= (1UL << get_order(size
)) - 1;
1978 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1980 if (unlikely(address
== DMA_ERROR_CODE
)) {
1982 * setting next_address here will let the address
1983 * allocator only scan the new allocated range in the
1984 * first run. This is a small optimization.
1986 dma_dom
->next_address
= dma_dom
->aperture_size
;
1988 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
1992 * aperture was successfully enlarged by 128 MB, try
1999 for (i
= 0; i
< pages
; ++i
) {
2000 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2001 if (ret
== DMA_ERROR_CODE
)
2009 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2011 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2012 domain_flush_tlb(&dma_dom
->domain
);
2013 dma_dom
->need_flush
= false;
2014 } else if (unlikely(amd_iommu_np_cache
))
2015 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2022 for (--i
; i
>= 0; --i
) {
2024 dma_ops_domain_unmap(dma_dom
, start
);
2027 dma_ops_free_addresses(dma_dom
, address
, pages
);
2029 return DMA_ERROR_CODE
;
2033 * Does the reverse of the __map_single function. Must be called with
2034 * the domain lock held too
2036 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2037 dma_addr_t dma_addr
,
2041 dma_addr_t flush_addr
;
2042 dma_addr_t i
, start
;
2045 if ((dma_addr
== DMA_ERROR_CODE
) ||
2046 (dma_addr
+ size
> dma_dom
->aperture_size
))
2049 flush_addr
= dma_addr
;
2050 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2051 dma_addr
&= PAGE_MASK
;
2054 for (i
= 0; i
< pages
; ++i
) {
2055 dma_ops_domain_unmap(dma_dom
, start
);
2059 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2061 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2063 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2064 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2065 dma_dom
->need_flush
= false;
2070 * The exported map_single function for dma_ops.
2072 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2073 unsigned long offset
, size_t size
,
2074 enum dma_data_direction dir
,
2075 struct dma_attrs
*attrs
)
2077 unsigned long flags
;
2078 struct protection_domain
*domain
;
2081 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2083 INC_STATS_COUNTER(cnt_map_single
);
2085 domain
= get_domain(dev
);
2086 if (PTR_ERR(domain
) == -EINVAL
)
2087 return (dma_addr_t
)paddr
;
2088 else if (IS_ERR(domain
))
2089 return DMA_ERROR_CODE
;
2091 dma_mask
= *dev
->dma_mask
;
2093 spin_lock_irqsave(&domain
->lock
, flags
);
2095 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2097 if (addr
== DMA_ERROR_CODE
)
2100 domain_flush_complete(domain
);
2103 spin_unlock_irqrestore(&domain
->lock
, flags
);
2109 * The exported unmap_single function for dma_ops.
2111 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2112 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2114 unsigned long flags
;
2115 struct protection_domain
*domain
;
2117 INC_STATS_COUNTER(cnt_unmap_single
);
2119 domain
= get_domain(dev
);
2123 spin_lock_irqsave(&domain
->lock
, flags
);
2125 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2127 domain_flush_complete(domain
);
2129 spin_unlock_irqrestore(&domain
->lock
, flags
);
2133 * This is a special map_sg function which is used if we should map a
2134 * device which is not handled by an AMD IOMMU in the system.
2136 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2137 int nelems
, int dir
)
2139 struct scatterlist
*s
;
2142 for_each_sg(sglist
, s
, nelems
, i
) {
2143 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2144 s
->dma_length
= s
->length
;
2151 * The exported map_sg function for dma_ops (handles scatter-gather
2154 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2155 int nelems
, enum dma_data_direction dir
,
2156 struct dma_attrs
*attrs
)
2158 unsigned long flags
;
2159 struct protection_domain
*domain
;
2161 struct scatterlist
*s
;
2163 int mapped_elems
= 0;
2166 INC_STATS_COUNTER(cnt_map_sg
);
2168 domain
= get_domain(dev
);
2169 if (PTR_ERR(domain
) == -EINVAL
)
2170 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2171 else if (IS_ERR(domain
))
2174 dma_mask
= *dev
->dma_mask
;
2176 spin_lock_irqsave(&domain
->lock
, flags
);
2178 for_each_sg(sglist
, s
, nelems
, i
) {
2181 s
->dma_address
= __map_single(dev
, domain
->priv
,
2182 paddr
, s
->length
, dir
, false,
2185 if (s
->dma_address
) {
2186 s
->dma_length
= s
->length
;
2192 domain_flush_complete(domain
);
2195 spin_unlock_irqrestore(&domain
->lock
, flags
);
2197 return mapped_elems
;
2199 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2201 __unmap_single(domain
->priv
, s
->dma_address
,
2202 s
->dma_length
, dir
);
2203 s
->dma_address
= s
->dma_length
= 0;
2212 * The exported map_sg function for dma_ops (handles scatter-gather
2215 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2216 int nelems
, enum dma_data_direction dir
,
2217 struct dma_attrs
*attrs
)
2219 unsigned long flags
;
2220 struct protection_domain
*domain
;
2221 struct scatterlist
*s
;
2224 INC_STATS_COUNTER(cnt_unmap_sg
);
2226 domain
= get_domain(dev
);
2230 spin_lock_irqsave(&domain
->lock
, flags
);
2232 for_each_sg(sglist
, s
, nelems
, i
) {
2233 __unmap_single(domain
->priv
, s
->dma_address
,
2234 s
->dma_length
, dir
);
2235 s
->dma_address
= s
->dma_length
= 0;
2238 domain_flush_complete(domain
);
2240 spin_unlock_irqrestore(&domain
->lock
, flags
);
2244 * The exported alloc_coherent function for dma_ops.
2246 static void *alloc_coherent(struct device
*dev
, size_t size
,
2247 dma_addr_t
*dma_addr
, gfp_t flag
)
2249 unsigned long flags
;
2251 struct protection_domain
*domain
;
2253 u64 dma_mask
= dev
->coherent_dma_mask
;
2255 INC_STATS_COUNTER(cnt_alloc_coherent
);
2257 domain
= get_domain(dev
);
2258 if (PTR_ERR(domain
) == -EINVAL
) {
2259 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2260 *dma_addr
= __pa(virt_addr
);
2262 } else if (IS_ERR(domain
))
2265 dma_mask
= dev
->coherent_dma_mask
;
2266 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2269 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2273 paddr
= virt_to_phys(virt_addr
);
2276 dma_mask
= *dev
->dma_mask
;
2278 spin_lock_irqsave(&domain
->lock
, flags
);
2280 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2281 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2283 if (*dma_addr
== DMA_ERROR_CODE
) {
2284 spin_unlock_irqrestore(&domain
->lock
, flags
);
2288 domain_flush_complete(domain
);
2290 spin_unlock_irqrestore(&domain
->lock
, flags
);
2296 free_pages((unsigned long)virt_addr
, get_order(size
));
2302 * The exported free_coherent function for dma_ops.
2304 static void free_coherent(struct device
*dev
, size_t size
,
2305 void *virt_addr
, dma_addr_t dma_addr
)
2307 unsigned long flags
;
2308 struct protection_domain
*domain
;
2310 INC_STATS_COUNTER(cnt_free_coherent
);
2312 domain
= get_domain(dev
);
2316 spin_lock_irqsave(&domain
->lock
, flags
);
2318 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2320 domain_flush_complete(domain
);
2322 spin_unlock_irqrestore(&domain
->lock
, flags
);
2325 free_pages((unsigned long)virt_addr
, get_order(size
));
2329 * This function is called by the DMA layer to find out if we can handle a
2330 * particular device. It is part of the dma_ops.
2332 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2334 return check_device(dev
);
2338 * The function for pre-allocating protection domains.
2340 * If the driver core informs the DMA layer if a driver grabs a device
2341 * we don't need to preallocate the protection domains anymore.
2342 * For now we have to.
2344 static void prealloc_protection_domains(void)
2346 struct pci_dev
*dev
= NULL
;
2347 struct dma_ops_domain
*dma_dom
;
2350 for_each_pci_dev(dev
) {
2352 /* Do we handle this device? */
2353 if (!check_device(&dev
->dev
))
2356 /* Is there already any domain for it? */
2357 if (domain_for_device(&dev
->dev
))
2360 devid
= get_device_id(&dev
->dev
);
2362 dma_dom
= dma_ops_domain_alloc();
2365 init_unity_mappings_for_device(dma_dom
, devid
);
2366 dma_dom
->target_dev
= devid
;
2368 attach_device(&dev
->dev
, &dma_dom
->domain
);
2370 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2374 static struct dma_map_ops amd_iommu_dma_ops
= {
2375 .alloc_coherent
= alloc_coherent
,
2376 .free_coherent
= free_coherent
,
2377 .map_page
= map_page
,
2378 .unmap_page
= unmap_page
,
2380 .unmap_sg
= unmap_sg
,
2381 .dma_supported
= amd_iommu_dma_supported
,
2385 * The function which clues the AMD IOMMU driver into dma_ops.
2388 void __init
amd_iommu_init_api(void)
2390 register_iommu(&amd_iommu_ops
);
2393 int __init
amd_iommu_init_dma_ops(void)
2395 struct amd_iommu
*iommu
;
2399 * first allocate a default protection domain for every IOMMU we
2400 * found in the system. Devices not assigned to any other
2401 * protection domain will be assigned to the default one.
2403 for_each_iommu(iommu
) {
2404 iommu
->default_dom
= dma_ops_domain_alloc();
2405 if (iommu
->default_dom
== NULL
)
2407 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2408 ret
= iommu_init_unity_mappings(iommu
);
2414 * Pre-allocate the protection domains for each device.
2416 prealloc_protection_domains();
2421 /* Make the driver finally visible to the drivers */
2422 dma_ops
= &amd_iommu_dma_ops
;
2424 amd_iommu_stats_init();
2430 for_each_iommu(iommu
) {
2431 if (iommu
->default_dom
)
2432 dma_ops_domain_free(iommu
->default_dom
);
2438 /*****************************************************************************
2440 * The following functions belong to the exported interface of AMD IOMMU
2442 * This interface allows access to lower level functions of the IOMMU
2443 * like protection domain handling and assignement of devices to domains
2444 * which is not possible with the dma_ops interface.
2446 *****************************************************************************/
2448 static void cleanup_domain(struct protection_domain
*domain
)
2450 struct iommu_dev_data
*dev_data
, *next
;
2451 unsigned long flags
;
2453 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2455 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
2456 struct device
*dev
= dev_data
->dev
;
2458 __detach_device(dev
);
2459 atomic_set(&dev_data
->bind
, 0);
2462 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2465 static void protection_domain_free(struct protection_domain
*domain
)
2470 del_domain_from_list(domain
);
2473 domain_id_free(domain
->id
);
2478 static struct protection_domain
*protection_domain_alloc(void)
2480 struct protection_domain
*domain
;
2482 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2486 spin_lock_init(&domain
->lock
);
2487 mutex_init(&domain
->api_lock
);
2488 domain
->id
= domain_id_alloc();
2491 INIT_LIST_HEAD(&domain
->dev_list
);
2493 add_domain_to_list(domain
);
2503 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2505 struct protection_domain
*domain
;
2507 domain
= protection_domain_alloc();
2511 domain
->mode
= PAGE_MODE_3_LEVEL
;
2512 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2513 if (!domain
->pt_root
)
2521 protection_domain_free(domain
);
2526 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2528 struct protection_domain
*domain
= dom
->priv
;
2533 if (domain
->dev_cnt
> 0)
2534 cleanup_domain(domain
);
2536 BUG_ON(domain
->dev_cnt
!= 0);
2538 free_pagetable(domain
);
2540 protection_domain_free(domain
);
2545 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2548 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
2549 struct amd_iommu
*iommu
;
2552 if (!check_device(dev
))
2555 devid
= get_device_id(dev
);
2557 if (dev_data
->domain
!= NULL
)
2560 iommu
= amd_iommu_rlookup_table
[devid
];
2564 device_flush_dte(dev
);
2565 iommu_completion_wait(iommu
);
2568 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2571 struct protection_domain
*domain
= dom
->priv
;
2572 struct iommu_dev_data
*dev_data
;
2573 struct amd_iommu
*iommu
;
2577 if (!check_device(dev
))
2580 dev_data
= dev
->archdata
.iommu
;
2582 devid
= get_device_id(dev
);
2584 iommu
= amd_iommu_rlookup_table
[devid
];
2588 if (dev_data
->domain
)
2591 ret
= attach_device(dev
, domain
);
2593 iommu_completion_wait(iommu
);
2598 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
2599 phys_addr_t paddr
, int gfp_order
, int iommu_prot
)
2601 unsigned long page_size
= 0x1000UL
<< gfp_order
;
2602 struct protection_domain
*domain
= dom
->priv
;
2606 if (iommu_prot
& IOMMU_READ
)
2607 prot
|= IOMMU_PROT_IR
;
2608 if (iommu_prot
& IOMMU_WRITE
)
2609 prot
|= IOMMU_PROT_IW
;
2611 mutex_lock(&domain
->api_lock
);
2612 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
2613 mutex_unlock(&domain
->api_lock
);
2618 static int amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
2621 struct protection_domain
*domain
= dom
->priv
;
2622 unsigned long page_size
, unmap_size
;
2624 page_size
= 0x1000UL
<< gfp_order
;
2626 mutex_lock(&domain
->api_lock
);
2627 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
2628 mutex_unlock(&domain
->api_lock
);
2630 domain_flush_tlb_pde(domain
);
2632 return get_order(unmap_size
);
2635 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2638 struct protection_domain
*domain
= dom
->priv
;
2639 unsigned long offset_mask
;
2643 pte
= fetch_pte(domain
, iova
);
2645 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
2648 if (PM_PTE_LEVEL(*pte
) == 0)
2649 offset_mask
= PAGE_SIZE
- 1;
2651 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
2653 __pte
= *pte
& PM_ADDR_MASK
;
2654 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
2659 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
2663 case IOMMU_CAP_CACHE_COHERENCY
:
2670 static struct iommu_ops amd_iommu_ops
= {
2671 .domain_init
= amd_iommu_domain_init
,
2672 .domain_destroy
= amd_iommu_domain_destroy
,
2673 .attach_dev
= amd_iommu_attach_device
,
2674 .detach_dev
= amd_iommu_detach_device
,
2675 .map
= amd_iommu_map
,
2676 .unmap
= amd_iommu_unmap
,
2677 .iova_to_phys
= amd_iommu_iova_to_phys
,
2678 .domain_has_cap
= amd_iommu_domain_has_cap
,
2681 /*****************************************************************************
2683 * The next functions do a basic initialization of IOMMU for pass through
2686 * In passthrough mode the IOMMU is initialized and enabled but not used for
2687 * DMA-API translation.
2689 *****************************************************************************/
2691 int __init
amd_iommu_init_passthrough(void)
2693 struct amd_iommu
*iommu
;
2694 struct pci_dev
*dev
= NULL
;
2697 /* allocate passthrough domain */
2698 pt_domain
= protection_domain_alloc();
2702 pt_domain
->mode
|= PAGE_MODE_NONE
;
2704 for_each_pci_dev(dev
) {
2705 if (!check_device(&dev
->dev
))
2708 devid
= get_device_id(&dev
->dev
);
2710 iommu
= amd_iommu_rlookup_table
[devid
];
2714 attach_device(&dev
->dev
, pt_domain
);
2717 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");