2 * Copyright (C) 2007-2008 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/gfp.h>
22 #include <linux/bitops.h>
23 #include <linux/debugfs.h>
24 #include <linux/scatterlist.h>
25 #include <linux/iommu-helper.h>
26 #ifdef CONFIG_IOMMU_API
27 #include <linux/iommu.h>
29 #include <asm/proto.h>
30 #include <asm/iommu.h>
32 #include <asm/amd_iommu_types.h>
33 #include <asm/amd_iommu.h>
35 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
37 #define EXIT_LOOP_COUNT 10000000
39 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
41 /* A list of preallocated protection domains */
42 static LIST_HEAD(iommu_pd_list
);
43 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
45 #ifdef CONFIG_IOMMU_API
46 static struct iommu_ops amd_iommu_ops
;
50 * general struct to manage commands send to an IOMMU
56 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
57 struct unity_map_entry
*e
);
58 static struct dma_ops_domain
*find_protection_domain(u16 devid
);
61 #ifdef CONFIG_AMD_IOMMU_STATS
64 * Initialization code for statistics collection
67 DECLARE_STATS_COUNTER(compl_wait
);
68 DECLARE_STATS_COUNTER(cnt_map_single
);
69 DECLARE_STATS_COUNTER(cnt_unmap_single
);
70 DECLARE_STATS_COUNTER(cnt_map_sg
);
71 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
72 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
73 DECLARE_STATS_COUNTER(cnt_free_coherent
);
74 DECLARE_STATS_COUNTER(cross_page
);
75 DECLARE_STATS_COUNTER(domain_flush_single
);
76 DECLARE_STATS_COUNTER(domain_flush_all
);
77 DECLARE_STATS_COUNTER(alloced_io_mem
);
78 DECLARE_STATS_COUNTER(total_map_requests
);
80 static struct dentry
*stats_dir
;
81 static struct dentry
*de_isolate
;
82 static struct dentry
*de_fflush
;
84 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
86 if (stats_dir
== NULL
)
89 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
93 static void amd_iommu_stats_init(void)
95 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
96 if (stats_dir
== NULL
)
99 de_isolate
= debugfs_create_bool("isolation", 0444, stats_dir
,
100 (u32
*)&amd_iommu_isolate
);
102 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
103 (u32
*)&amd_iommu_unmap_flush
);
105 amd_iommu_stats_add(&compl_wait
);
106 amd_iommu_stats_add(&cnt_map_single
);
107 amd_iommu_stats_add(&cnt_unmap_single
);
108 amd_iommu_stats_add(&cnt_map_sg
);
109 amd_iommu_stats_add(&cnt_unmap_sg
);
110 amd_iommu_stats_add(&cnt_alloc_coherent
);
111 amd_iommu_stats_add(&cnt_free_coherent
);
112 amd_iommu_stats_add(&cross_page
);
113 amd_iommu_stats_add(&domain_flush_single
);
114 amd_iommu_stats_add(&domain_flush_all
);
115 amd_iommu_stats_add(&alloced_io_mem
);
116 amd_iommu_stats_add(&total_map_requests
);
121 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
122 static int iommu_has_npcache(struct amd_iommu
*iommu
)
124 return iommu
->cap
& (1UL << IOMMU_CAP_NPCACHE
);
127 /****************************************************************************
129 * Interrupt handling functions
131 ****************************************************************************/
133 static void iommu_print_event(void *__evt
)
136 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
137 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
138 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
139 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
140 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
142 printk(KERN_ERR
"AMD IOMMU: Event logged [");
145 case EVENT_TYPE_ILL_DEV
:
146 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
147 "address=0x%016llx flags=0x%04x]\n",
148 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
151 case EVENT_TYPE_IO_FAULT
:
152 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
153 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
154 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
155 domid
, address
, flags
);
157 case EVENT_TYPE_DEV_TAB_ERR
:
158 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
159 "address=0x%016llx flags=0x%04x]\n",
160 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
163 case EVENT_TYPE_PAGE_TAB_ERR
:
164 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
165 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
166 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
167 domid
, address
, flags
);
169 case EVENT_TYPE_ILL_CMD
:
170 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
172 case EVENT_TYPE_CMD_HARD_ERR
:
173 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
174 "flags=0x%04x]\n", address
, flags
);
176 case EVENT_TYPE_IOTLB_INV_TO
:
177 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
178 "address=0x%016llx]\n",
179 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
182 case EVENT_TYPE_INV_DEV_REQ
:
183 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
184 "address=0x%016llx flags=0x%04x]\n",
185 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
189 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
193 static void iommu_poll_events(struct amd_iommu
*iommu
)
198 spin_lock_irqsave(&iommu
->lock
, flags
);
200 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
201 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
203 while (head
!= tail
) {
204 iommu_print_event(iommu
->evt_buf
+ head
);
205 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
208 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
210 spin_unlock_irqrestore(&iommu
->lock
, flags
);
213 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
215 struct amd_iommu
*iommu
;
217 list_for_each_entry(iommu
, &amd_iommu_list
, list
)
218 iommu_poll_events(iommu
);
223 /****************************************************************************
225 * IOMMU command queuing functions
227 ****************************************************************************/
230 * Writes the command to the IOMMUs command buffer and informs the
231 * hardware about the new command. Must be called with iommu->lock held.
233 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
238 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
239 target
= iommu
->cmd_buf
+ tail
;
240 memcpy_toio(target
, cmd
, sizeof(*cmd
));
241 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
242 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
245 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
251 * General queuing function for commands. Takes iommu->lock and calls
252 * __iommu_queue_command().
254 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
259 spin_lock_irqsave(&iommu
->lock
, flags
);
260 ret
= __iommu_queue_command(iommu
, cmd
);
262 iommu
->need_sync
= true;
263 spin_unlock_irqrestore(&iommu
->lock
, flags
);
269 * This function waits until an IOMMU has completed a completion
272 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
278 INC_STATS_COUNTER(compl_wait
);
280 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
282 /* wait for the bit to become one */
283 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
284 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
287 /* set bit back to zero */
288 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
289 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
291 if (unlikely(i
== EXIT_LOOP_COUNT
))
292 panic("AMD IOMMU: Completion wait loop failed\n");
296 * This function queues a completion wait command into the command
299 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
301 struct iommu_cmd cmd
;
303 memset(&cmd
, 0, sizeof(cmd
));
304 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
305 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
307 return __iommu_queue_command(iommu
, &cmd
);
311 * This function is called whenever we need to ensure that the IOMMU has
312 * completed execution of all commands we sent. It sends a
313 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
314 * us about that by writing a value to a physical address we pass with
317 static int iommu_completion_wait(struct amd_iommu
*iommu
)
322 spin_lock_irqsave(&iommu
->lock
, flags
);
324 if (!iommu
->need_sync
)
327 ret
= __iommu_completion_wait(iommu
);
329 iommu
->need_sync
= false;
334 __iommu_wait_for_completion(iommu
);
337 spin_unlock_irqrestore(&iommu
->lock
, flags
);
343 * Command send function for invalidating a device table entry
345 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
347 struct iommu_cmd cmd
;
350 BUG_ON(iommu
== NULL
);
352 memset(&cmd
, 0, sizeof(cmd
));
353 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
356 ret
= iommu_queue_command(iommu
, &cmd
);
361 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
362 u16 domid
, int pde
, int s
)
364 memset(cmd
, 0, sizeof(*cmd
));
365 address
&= PAGE_MASK
;
366 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
367 cmd
->data
[1] |= domid
;
368 cmd
->data
[2] = lower_32_bits(address
);
369 cmd
->data
[3] = upper_32_bits(address
);
370 if (s
) /* size bit - we flush more than one 4kb page */
371 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
372 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
373 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
377 * Generic command send function for invalidaing TLB entries
379 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
380 u64 address
, u16 domid
, int pde
, int s
)
382 struct iommu_cmd cmd
;
385 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
387 ret
= iommu_queue_command(iommu
, &cmd
);
393 * TLB invalidation function which is called from the mapping functions.
394 * It invalidates a single PTE if the range to flush is within a single
395 * page. Otherwise it flushes the whole TLB of the IOMMU.
397 static int iommu_flush_pages(struct amd_iommu
*iommu
, u16 domid
,
398 u64 address
, size_t size
)
401 unsigned pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
403 address
&= PAGE_MASK
;
407 * If we have to flush more than one page, flush all
408 * TLB entries for this domain
410 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
414 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, s
);
419 /* Flush the whole IO/TLB for a given protection domain */
420 static void iommu_flush_tlb(struct amd_iommu
*iommu
, u16 domid
)
422 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
424 INC_STATS_COUNTER(domain_flush_single
);
426 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, 1);
429 #ifdef CONFIG_IOMMU_API
431 * This function is used to flush the IO/TLB for a given protection domain
432 * on every IOMMU in the system
434 static void iommu_flush_domain(u16 domid
)
437 struct amd_iommu
*iommu
;
438 struct iommu_cmd cmd
;
440 INC_STATS_COUNTER(domain_flush_all
);
442 __iommu_build_inv_iommu_pages(&cmd
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
445 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
446 spin_lock_irqsave(&iommu
->lock
, flags
);
447 __iommu_queue_command(iommu
, &cmd
);
448 __iommu_completion_wait(iommu
);
449 __iommu_wait_for_completion(iommu
);
450 spin_unlock_irqrestore(&iommu
->lock
, flags
);
455 /****************************************************************************
457 * The functions below are used the create the page table mappings for
458 * unity mapped regions.
460 ****************************************************************************/
463 * Generic mapping functions. It maps a physical address into a DMA
464 * address space. It allocates the page table pages if necessary.
465 * In the future it can be extended to a generic mapping function
466 * supporting all features of AMD IOMMU page tables like level skipping
467 * and full 64 bit address spaces.
469 static int iommu_map_page(struct protection_domain
*dom
,
470 unsigned long bus_addr
,
471 unsigned long phys_addr
,
474 u64 __pte
, *pte
, *page
;
476 bus_addr
= PAGE_ALIGN(bus_addr
);
477 phys_addr
= PAGE_ALIGN(phys_addr
);
479 /* only support 512GB address spaces for now */
480 if (bus_addr
> IOMMU_MAP_SIZE_L3
|| !(prot
& IOMMU_PROT_MASK
))
483 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(bus_addr
)];
485 if (!IOMMU_PTE_PRESENT(*pte
)) {
486 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
489 *pte
= IOMMU_L2_PDE(virt_to_phys(page
));
492 pte
= IOMMU_PTE_PAGE(*pte
);
493 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
495 if (!IOMMU_PTE_PRESENT(*pte
)) {
496 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
499 *pte
= IOMMU_L1_PDE(virt_to_phys(page
));
502 pte
= IOMMU_PTE_PAGE(*pte
);
503 pte
= &pte
[IOMMU_PTE_L0_INDEX(bus_addr
)];
505 if (IOMMU_PTE_PRESENT(*pte
))
508 __pte
= phys_addr
| IOMMU_PTE_P
;
509 if (prot
& IOMMU_PROT_IR
)
510 __pte
|= IOMMU_PTE_IR
;
511 if (prot
& IOMMU_PROT_IW
)
512 __pte
|= IOMMU_PTE_IW
;
519 #ifdef CONFIG_IOMMU_API
520 static void iommu_unmap_page(struct protection_domain
*dom
,
521 unsigned long bus_addr
)
525 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(bus_addr
)];
527 if (!IOMMU_PTE_PRESENT(*pte
))
530 pte
= IOMMU_PTE_PAGE(*pte
);
531 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
533 if (!IOMMU_PTE_PRESENT(*pte
))
536 pte
= IOMMU_PTE_PAGE(*pte
);
537 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
544 * This function checks if a specific unity mapping entry is needed for
545 * this specific IOMMU.
547 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
548 struct unity_map_entry
*entry
)
552 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
553 bdf
= amd_iommu_alias_table
[i
];
554 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
562 * Init the unity mappings for a specific IOMMU in the system
564 * Basically iterates over all unity mapping entries and applies them to
565 * the default domain DMA of that IOMMU if necessary.
567 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
569 struct unity_map_entry
*entry
;
572 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
573 if (!iommu_for_unity_map(iommu
, entry
))
575 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
584 * This function actually applies the mapping to the page table of the
587 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
588 struct unity_map_entry
*e
)
593 for (addr
= e
->address_start
; addr
< e
->address_end
;
595 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
);
599 * if unity mapping is in aperture range mark the page
600 * as allocated in the aperture
602 if (addr
< dma_dom
->aperture_size
)
603 __set_bit(addr
>> PAGE_SHIFT
, dma_dom
->bitmap
);
610 * Inits the unity mappings required for a specific device
612 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
615 struct unity_map_entry
*e
;
618 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
619 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
621 ret
= dma_ops_unity_map(dma_dom
, e
);
629 /****************************************************************************
631 * The next functions belong to the address allocator for the dma_ops
632 * interface functions. They work like the allocators in the other IOMMU
633 * drivers. Its basically a bitmap which marks the allocated pages in
634 * the aperture. Maybe it could be enhanced in the future to a more
635 * efficient allocator.
637 ****************************************************************************/
640 * The address allocator core function.
642 * called with domain->lock held
644 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
645 struct dma_ops_domain
*dom
,
647 unsigned long align_mask
,
651 unsigned long address
;
652 unsigned long boundary_size
;
654 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
655 PAGE_SIZE
) >> PAGE_SHIFT
;
656 limit
= iommu_device_max_index(dom
->aperture_size
>> PAGE_SHIFT
, 0,
657 dma_mask
>> PAGE_SHIFT
);
659 if (dom
->next_bit
>= limit
) {
661 dom
->need_flush
= true;
664 address
= iommu_area_alloc(dom
->bitmap
, limit
, dom
->next_bit
, pages
,
665 0 , boundary_size
, align_mask
);
667 address
= iommu_area_alloc(dom
->bitmap
, limit
, 0, pages
,
668 0, boundary_size
, align_mask
);
669 dom
->need_flush
= true;
672 if (likely(address
!= -1)) {
673 dom
->next_bit
= address
+ pages
;
674 address
<<= PAGE_SHIFT
;
676 address
= bad_dma_address
;
678 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
684 * The address free function.
686 * called with domain->lock held
688 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
689 unsigned long address
,
692 address
>>= PAGE_SHIFT
;
693 iommu_area_free(dom
->bitmap
, address
, pages
);
695 if (address
>= dom
->next_bit
)
696 dom
->need_flush
= true;
699 /****************************************************************************
701 * The next functions belong to the domain allocation. A domain is
702 * allocated for every IOMMU as the default domain. If device isolation
703 * is enabled, every device get its own domain. The most important thing
704 * about domains is the page table mapping the DMA address space they
707 ****************************************************************************/
709 static u16
domain_id_alloc(void)
714 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
715 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
717 if (id
> 0 && id
< MAX_DOMAIN_ID
)
718 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
721 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
726 #ifdef CONFIG_IOMMU_API
727 static void domain_id_free(int id
)
731 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
732 if (id
> 0 && id
< MAX_DOMAIN_ID
)
733 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
734 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
739 * Used to reserve address ranges in the aperture (e.g. for exclusion
742 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
743 unsigned long start_page
,
746 unsigned int last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
748 if (start_page
+ pages
> last_page
)
749 pages
= last_page
- start_page
;
751 iommu_area_reserve(dom
->bitmap
, start_page
, pages
);
754 static void free_pagetable(struct protection_domain
*domain
)
759 p1
= domain
->pt_root
;
764 for (i
= 0; i
< 512; ++i
) {
765 if (!IOMMU_PTE_PRESENT(p1
[i
]))
768 p2
= IOMMU_PTE_PAGE(p1
[i
]);
769 for (j
= 0; j
< 512; ++j
) {
770 if (!IOMMU_PTE_PRESENT(p2
[j
]))
772 p3
= IOMMU_PTE_PAGE(p2
[j
]);
773 free_page((unsigned long)p3
);
776 free_page((unsigned long)p2
);
779 free_page((unsigned long)p1
);
781 domain
->pt_root
= NULL
;
785 * Free a domain, only used if something went wrong in the
786 * allocation path and we need to free an already allocated page table
788 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
793 free_pagetable(&dom
->domain
);
795 kfree(dom
->pte_pages
);
803 * Allocates a new protection domain usable for the dma_ops functions.
804 * It also intializes the page table and the address allocator data
805 * structures required for the dma_ops interface
807 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
,
810 struct dma_ops_domain
*dma_dom
;
811 unsigned i
, num_pte_pages
;
816 * Currently the DMA aperture must be between 32 MB and 1GB in size
818 if ((order
< 25) || (order
> 30))
821 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
825 spin_lock_init(&dma_dom
->domain
.lock
);
827 dma_dom
->domain
.id
= domain_id_alloc();
828 if (dma_dom
->domain
.id
== 0)
830 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
831 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
832 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
833 dma_dom
->domain
.priv
= dma_dom
;
834 if (!dma_dom
->domain
.pt_root
)
836 dma_dom
->aperture_size
= (1ULL << order
);
837 dma_dom
->bitmap
= kzalloc(dma_dom
->aperture_size
/ (PAGE_SIZE
* 8),
839 if (!dma_dom
->bitmap
)
842 * mark the first page as allocated so we never return 0 as
843 * a valid dma-address. So we can use 0 as error value
845 dma_dom
->bitmap
[0] = 1;
846 dma_dom
->next_bit
= 0;
848 dma_dom
->need_flush
= false;
849 dma_dom
->target_dev
= 0xffff;
851 /* Intialize the exclusion range if necessary */
852 if (iommu
->exclusion_start
&&
853 iommu
->exclusion_start
< dma_dom
->aperture_size
) {
854 unsigned long startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
855 int pages
= iommu_num_pages(iommu
->exclusion_start
,
856 iommu
->exclusion_length
,
858 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
862 * At the last step, build the page tables so we don't need to
863 * allocate page table pages in the dma_ops mapping/unmapping
866 num_pte_pages
= dma_dom
->aperture_size
/ (PAGE_SIZE
* 512);
867 dma_dom
->pte_pages
= kzalloc(num_pte_pages
* sizeof(void *),
869 if (!dma_dom
->pte_pages
)
872 l2_pde
= (u64
*)get_zeroed_page(GFP_KERNEL
);
876 dma_dom
->domain
.pt_root
[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde
));
878 for (i
= 0; i
< num_pte_pages
; ++i
) {
879 dma_dom
->pte_pages
[i
] = (u64
*)get_zeroed_page(GFP_KERNEL
);
880 if (!dma_dom
->pte_pages
[i
])
882 address
= virt_to_phys(dma_dom
->pte_pages
[i
]);
883 l2_pde
[i
] = IOMMU_L1_PDE(address
);
889 dma_ops_domain_free(dma_dom
);
895 * little helper function to check whether a given protection domain is a
898 static bool dma_ops_domain(struct protection_domain
*domain
)
900 return domain
->flags
& PD_DMA_OPS_MASK
;
904 * Find out the protection domain structure for a given PCI device. This
905 * will give us the pointer to the page table root for example.
907 static struct protection_domain
*domain_for_device(u16 devid
)
909 struct protection_domain
*dom
;
912 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
913 dom
= amd_iommu_pd_table
[devid
];
914 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
920 * If a device is not yet associated with a domain, this function does
921 * assigns it visible for the hardware
923 static void attach_device(struct amd_iommu
*iommu
,
924 struct protection_domain
*domain
,
928 u64 pte_root
= virt_to_phys(domain
->pt_root
);
930 domain
->dev_cnt
+= 1;
932 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
933 << DEV_ENTRY_MODE_SHIFT
;
934 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
936 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
937 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
938 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
939 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
941 amd_iommu_pd_table
[devid
] = domain
;
942 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
944 iommu_queue_inv_dev_entry(iommu
, devid
);
948 * Removes a device from a protection domain (unlocked)
950 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
954 spin_lock(&domain
->lock
);
956 /* remove domain from the lookup table */
957 amd_iommu_pd_table
[devid
] = NULL
;
959 /* remove entry from the device table seen by the hardware */
960 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
961 amd_iommu_dev_table
[devid
].data
[1] = 0;
962 amd_iommu_dev_table
[devid
].data
[2] = 0;
964 /* decrease reference counter */
965 domain
->dev_cnt
-= 1;
968 spin_unlock(&domain
->lock
);
972 * Removes a device from a protection domain (with devtable_lock held)
974 static void detach_device(struct protection_domain
*domain
, u16 devid
)
978 /* lock device table */
979 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
980 __detach_device(domain
, devid
);
981 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
984 static int device_change_notifier(struct notifier_block
*nb
,
985 unsigned long action
, void *data
)
987 struct device
*dev
= data
;
988 struct pci_dev
*pdev
= to_pci_dev(dev
);
989 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
990 struct protection_domain
*domain
;
991 struct dma_ops_domain
*dma_domain
;
992 struct amd_iommu
*iommu
;
993 int order
= amd_iommu_aperture_order
;
996 if (devid
> amd_iommu_last_bdf
)
999 devid
= amd_iommu_alias_table
[devid
];
1001 iommu
= amd_iommu_rlookup_table
[devid
];
1005 domain
= domain_for_device(devid
);
1007 if (domain
&& !dma_ops_domain(domain
))
1008 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
1009 "to a non-dma-ops domain\n", dev_name(dev
));
1012 case BUS_NOTIFY_BOUND_DRIVER
:
1015 dma_domain
= find_protection_domain(devid
);
1017 dma_domain
= iommu
->default_dom
;
1018 attach_device(iommu
, &dma_domain
->domain
, devid
);
1019 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
1020 "device %s\n", dma_domain
->domain
.id
, dev_name(dev
));
1022 case BUS_NOTIFY_UNBIND_DRIVER
:
1025 detach_device(domain
, devid
);
1027 case BUS_NOTIFY_ADD_DEVICE
:
1028 /* allocate a protection domain if a device is added */
1029 dma_domain
= find_protection_domain(devid
);
1032 dma_domain
= dma_ops_domain_alloc(iommu
, order
);
1035 dma_domain
->target_dev
= devid
;
1037 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1038 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1039 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1046 iommu_queue_inv_dev_entry(iommu
, devid
);
1047 iommu_completion_wait(iommu
);
1053 struct notifier_block device_nb
= {
1054 .notifier_call
= device_change_notifier
,
1057 /*****************************************************************************
1059 * The next functions belong to the dma_ops mapping/unmapping code.
1061 *****************************************************************************/
1064 * This function checks if the driver got a valid device from the caller to
1065 * avoid dereferencing invalid pointers.
1067 static bool check_device(struct device
*dev
)
1069 if (!dev
|| !dev
->dma_mask
)
1076 * In this function the list of preallocated protection domains is traversed to
1077 * find the domain for a specific device
1079 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
1081 struct dma_ops_domain
*entry
, *ret
= NULL
;
1082 unsigned long flags
;
1084 if (list_empty(&iommu_pd_list
))
1087 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1089 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
1090 if (entry
->target_dev
== devid
) {
1096 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1102 * In the dma_ops path we only have the struct device. This function
1103 * finds the corresponding IOMMU, the protection domain and the
1104 * requestor id for a given device.
1105 * If the device is not yet associated with a domain this is also done
1108 static int get_device_resources(struct device
*dev
,
1109 struct amd_iommu
**iommu
,
1110 struct protection_domain
**domain
,
1113 struct dma_ops_domain
*dma_dom
;
1114 struct pci_dev
*pcidev
;
1121 if (dev
->bus
!= &pci_bus_type
)
1124 pcidev
= to_pci_dev(dev
);
1125 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1127 /* device not translated by any IOMMU in the system? */
1128 if (_bdf
> amd_iommu_last_bdf
)
1131 *bdf
= amd_iommu_alias_table
[_bdf
];
1133 *iommu
= amd_iommu_rlookup_table
[*bdf
];
1136 *domain
= domain_for_device(*bdf
);
1137 if (*domain
== NULL
) {
1138 dma_dom
= find_protection_domain(*bdf
);
1140 dma_dom
= (*iommu
)->default_dom
;
1141 *domain
= &dma_dom
->domain
;
1142 attach_device(*iommu
, *domain
, *bdf
);
1143 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
1144 "device %s\n", (*domain
)->id
, dev_name(dev
));
1147 if (domain_for_device(_bdf
) == NULL
)
1148 attach_device(*iommu
, *domain
, _bdf
);
1154 * This is the generic map function. It maps one 4kb page at paddr to
1155 * the given address in the DMA address space for the domain.
1157 static dma_addr_t
dma_ops_domain_map(struct amd_iommu
*iommu
,
1158 struct dma_ops_domain
*dom
,
1159 unsigned long address
,
1165 WARN_ON(address
> dom
->aperture_size
);
1169 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1170 pte
+= IOMMU_PTE_L0_INDEX(address
);
1172 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1174 if (direction
== DMA_TO_DEVICE
)
1175 __pte
|= IOMMU_PTE_IR
;
1176 else if (direction
== DMA_FROM_DEVICE
)
1177 __pte
|= IOMMU_PTE_IW
;
1178 else if (direction
== DMA_BIDIRECTIONAL
)
1179 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1185 return (dma_addr_t
)address
;
1189 * The generic unmapping function for on page in the DMA address space.
1191 static void dma_ops_domain_unmap(struct amd_iommu
*iommu
,
1192 struct dma_ops_domain
*dom
,
1193 unsigned long address
)
1197 if (address
>= dom
->aperture_size
)
1200 WARN_ON(address
& ~PAGE_MASK
|| address
>= dom
->aperture_size
);
1202 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1203 pte
+= IOMMU_PTE_L0_INDEX(address
);
1211 * This function contains common code for mapping of a physically
1212 * contiguous memory region into DMA address space. It is used by all
1213 * mapping functions provided with this IOMMU driver.
1214 * Must be called with the domain lock held.
1216 static dma_addr_t
__map_single(struct device
*dev
,
1217 struct amd_iommu
*iommu
,
1218 struct dma_ops_domain
*dma_dom
,
1225 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1226 dma_addr_t address
, start
;
1228 unsigned long align_mask
= 0;
1231 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1234 INC_STATS_COUNTER(total_map_requests
);
1237 INC_STATS_COUNTER(cross_page
);
1240 align_mask
= (1UL << get_order(size
)) - 1;
1242 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1244 if (unlikely(address
== bad_dma_address
))
1248 for (i
= 0; i
< pages
; ++i
) {
1249 dma_ops_domain_map(iommu
, dma_dom
, start
, paddr
, dir
);
1255 ADD_STATS_COUNTER(alloced_io_mem
, size
);
1257 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1258 iommu_flush_tlb(iommu
, dma_dom
->domain
.id
);
1259 dma_dom
->need_flush
= false;
1260 } else if (unlikely(iommu_has_npcache(iommu
)))
1261 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, address
, size
);
1268 * Does the reverse of the __map_single function. Must be called with
1269 * the domain lock held too
1271 static void __unmap_single(struct amd_iommu
*iommu
,
1272 struct dma_ops_domain
*dma_dom
,
1273 dma_addr_t dma_addr
,
1277 dma_addr_t i
, start
;
1280 if ((dma_addr
== bad_dma_address
) ||
1281 (dma_addr
+ size
> dma_dom
->aperture_size
))
1284 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1285 dma_addr
&= PAGE_MASK
;
1288 for (i
= 0; i
< pages
; ++i
) {
1289 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1293 SUB_STATS_COUNTER(alloced_io_mem
, size
);
1295 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1297 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1298 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, dma_addr
, size
);
1299 dma_dom
->need_flush
= false;
1304 * The exported map_single function for dma_ops.
1306 static dma_addr_t
map_single(struct device
*dev
, phys_addr_t paddr
,
1307 size_t size
, int dir
)
1309 unsigned long flags
;
1310 struct amd_iommu
*iommu
;
1311 struct protection_domain
*domain
;
1316 INC_STATS_COUNTER(cnt_map_single
);
1318 if (!check_device(dev
))
1319 return bad_dma_address
;
1321 dma_mask
= *dev
->dma_mask
;
1323 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1325 if (iommu
== NULL
|| domain
== NULL
)
1326 /* device not handled by any AMD IOMMU */
1327 return (dma_addr_t
)paddr
;
1329 if (!dma_ops_domain(domain
))
1330 return bad_dma_address
;
1332 spin_lock_irqsave(&domain
->lock
, flags
);
1333 addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
, size
, dir
, false,
1335 if (addr
== bad_dma_address
)
1338 iommu_completion_wait(iommu
);
1341 spin_unlock_irqrestore(&domain
->lock
, flags
);
1347 * The exported unmap_single function for dma_ops.
1349 static void unmap_single(struct device
*dev
, dma_addr_t dma_addr
,
1350 size_t size
, int dir
)
1352 unsigned long flags
;
1353 struct amd_iommu
*iommu
;
1354 struct protection_domain
*domain
;
1357 INC_STATS_COUNTER(cnt_unmap_single
);
1359 if (!check_device(dev
) ||
1360 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1361 /* device not handled by any AMD IOMMU */
1364 if (!dma_ops_domain(domain
))
1367 spin_lock_irqsave(&domain
->lock
, flags
);
1369 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, dir
);
1371 iommu_completion_wait(iommu
);
1373 spin_unlock_irqrestore(&domain
->lock
, flags
);
1377 * This is a special map_sg function which is used if we should map a
1378 * device which is not handled by an AMD IOMMU in the system.
1380 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1381 int nelems
, int dir
)
1383 struct scatterlist
*s
;
1386 for_each_sg(sglist
, s
, nelems
, i
) {
1387 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1388 s
->dma_length
= s
->length
;
1395 * The exported map_sg function for dma_ops (handles scatter-gather
1398 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1399 int nelems
, int dir
)
1401 unsigned long flags
;
1402 struct amd_iommu
*iommu
;
1403 struct protection_domain
*domain
;
1406 struct scatterlist
*s
;
1408 int mapped_elems
= 0;
1411 INC_STATS_COUNTER(cnt_map_sg
);
1413 if (!check_device(dev
))
1416 dma_mask
= *dev
->dma_mask
;
1418 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1420 if (!iommu
|| !domain
)
1421 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1423 if (!dma_ops_domain(domain
))
1426 spin_lock_irqsave(&domain
->lock
, flags
);
1428 for_each_sg(sglist
, s
, nelems
, i
) {
1431 s
->dma_address
= __map_single(dev
, iommu
, domain
->priv
,
1432 paddr
, s
->length
, dir
, false,
1435 if (s
->dma_address
) {
1436 s
->dma_length
= s
->length
;
1442 iommu_completion_wait(iommu
);
1445 spin_unlock_irqrestore(&domain
->lock
, flags
);
1447 return mapped_elems
;
1449 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1451 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1452 s
->dma_length
, dir
);
1453 s
->dma_address
= s
->dma_length
= 0;
1462 * The exported map_sg function for dma_ops (handles scatter-gather
1465 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1466 int nelems
, int dir
)
1468 unsigned long flags
;
1469 struct amd_iommu
*iommu
;
1470 struct protection_domain
*domain
;
1471 struct scatterlist
*s
;
1475 INC_STATS_COUNTER(cnt_unmap_sg
);
1477 if (!check_device(dev
) ||
1478 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1481 if (!dma_ops_domain(domain
))
1484 spin_lock_irqsave(&domain
->lock
, flags
);
1486 for_each_sg(sglist
, s
, nelems
, i
) {
1487 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1488 s
->dma_length
, dir
);
1489 s
->dma_address
= s
->dma_length
= 0;
1492 iommu_completion_wait(iommu
);
1494 spin_unlock_irqrestore(&domain
->lock
, flags
);
1498 * The exported alloc_coherent function for dma_ops.
1500 static void *alloc_coherent(struct device
*dev
, size_t size
,
1501 dma_addr_t
*dma_addr
, gfp_t flag
)
1503 unsigned long flags
;
1505 struct amd_iommu
*iommu
;
1506 struct protection_domain
*domain
;
1509 u64 dma_mask
= dev
->coherent_dma_mask
;
1511 INC_STATS_COUNTER(cnt_alloc_coherent
);
1513 if (!check_device(dev
))
1516 if (!get_device_resources(dev
, &iommu
, &domain
, &devid
))
1517 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1520 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1524 paddr
= virt_to_phys(virt_addr
);
1526 if (!iommu
|| !domain
) {
1527 *dma_addr
= (dma_addr_t
)paddr
;
1531 if (!dma_ops_domain(domain
))
1535 dma_mask
= *dev
->dma_mask
;
1537 spin_lock_irqsave(&domain
->lock
, flags
);
1539 *dma_addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
,
1540 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
1542 if (*dma_addr
== bad_dma_address
)
1545 iommu_completion_wait(iommu
);
1547 spin_unlock_irqrestore(&domain
->lock
, flags
);
1553 free_pages((unsigned long)virt_addr
, get_order(size
));
1559 * The exported free_coherent function for dma_ops.
1561 static void free_coherent(struct device
*dev
, size_t size
,
1562 void *virt_addr
, dma_addr_t dma_addr
)
1564 unsigned long flags
;
1565 struct amd_iommu
*iommu
;
1566 struct protection_domain
*domain
;
1569 INC_STATS_COUNTER(cnt_free_coherent
);
1571 if (!check_device(dev
))
1574 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1576 if (!iommu
|| !domain
)
1579 if (!dma_ops_domain(domain
))
1582 spin_lock_irqsave(&domain
->lock
, flags
);
1584 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
1586 iommu_completion_wait(iommu
);
1588 spin_unlock_irqrestore(&domain
->lock
, flags
);
1591 free_pages((unsigned long)virt_addr
, get_order(size
));
1595 * This function is called by the DMA layer to find out if we can handle a
1596 * particular device. It is part of the dma_ops.
1598 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
1601 struct pci_dev
*pcidev
;
1603 /* No device or no PCI device */
1604 if (!dev
|| dev
->bus
!= &pci_bus_type
)
1607 pcidev
= to_pci_dev(dev
);
1609 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1611 /* Out of our scope? */
1612 if (bdf
> amd_iommu_last_bdf
)
1619 * The function for pre-allocating protection domains.
1621 * If the driver core informs the DMA layer if a driver grabs a device
1622 * we don't need to preallocate the protection domains anymore.
1623 * For now we have to.
1625 void prealloc_protection_domains(void)
1627 struct pci_dev
*dev
= NULL
;
1628 struct dma_ops_domain
*dma_dom
;
1629 struct amd_iommu
*iommu
;
1630 int order
= amd_iommu_aperture_order
;
1633 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
1634 devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
1635 if (devid
> amd_iommu_last_bdf
)
1637 devid
= amd_iommu_alias_table
[devid
];
1638 if (domain_for_device(devid
))
1640 iommu
= amd_iommu_rlookup_table
[devid
];
1643 dma_dom
= dma_ops_domain_alloc(iommu
, order
);
1646 init_unity_mappings_for_device(dma_dom
, devid
);
1647 dma_dom
->target_dev
= devid
;
1649 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
1653 static struct dma_mapping_ops amd_iommu_dma_ops
= {
1654 .alloc_coherent
= alloc_coherent
,
1655 .free_coherent
= free_coherent
,
1656 .map_single
= map_single
,
1657 .unmap_single
= unmap_single
,
1659 .unmap_sg
= unmap_sg
,
1660 .dma_supported
= amd_iommu_dma_supported
,
1664 * The function which clues the AMD IOMMU driver into dma_ops.
1666 int __init
amd_iommu_init_dma_ops(void)
1668 struct amd_iommu
*iommu
;
1669 int order
= amd_iommu_aperture_order
;
1673 * first allocate a default protection domain for every IOMMU we
1674 * found in the system. Devices not assigned to any other
1675 * protection domain will be assigned to the default one.
1677 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1678 iommu
->default_dom
= dma_ops_domain_alloc(iommu
, order
);
1679 if (iommu
->default_dom
== NULL
)
1681 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
1682 ret
= iommu_init_unity_mappings(iommu
);
1688 * If device isolation is enabled, pre-allocate the protection
1689 * domains for each device.
1691 if (amd_iommu_isolate
)
1692 prealloc_protection_domains();
1696 bad_dma_address
= 0;
1697 #ifdef CONFIG_GART_IOMMU
1698 gart_iommu_aperture_disabled
= 1;
1699 gart_iommu_aperture
= 0;
1702 /* Make the driver finally visible to the drivers */
1703 dma_ops
= &amd_iommu_dma_ops
;
1705 #ifdef CONFIG_IOMMU_API
1706 register_iommu(&amd_iommu_ops
);
1709 bus_register_notifier(&pci_bus_type
, &device_nb
);
1711 amd_iommu_stats_init();
1717 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1718 if (iommu
->default_dom
)
1719 dma_ops_domain_free(iommu
->default_dom
);
1725 /*****************************************************************************
1727 * The following functions belong to the exported interface of AMD IOMMU
1729 * This interface allows access to lower level functions of the IOMMU
1730 * like protection domain handling and assignement of devices to domains
1731 * which is not possible with the dma_ops interface.
1733 *****************************************************************************/
1735 #ifdef CONFIG_IOMMU_API
1737 static void cleanup_domain(struct protection_domain
*domain
)
1739 unsigned long flags
;
1742 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1744 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
1745 if (amd_iommu_pd_table
[devid
] == domain
)
1746 __detach_device(domain
, devid
);
1748 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1751 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
1753 struct protection_domain
*domain
;
1755 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
1759 spin_lock_init(&domain
->lock
);
1760 domain
->mode
= PAGE_MODE_3_LEVEL
;
1761 domain
->id
= domain_id_alloc();
1764 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1765 if (!domain
->pt_root
)
1778 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
1780 struct protection_domain
*domain
= dom
->priv
;
1785 if (domain
->dev_cnt
> 0)
1786 cleanup_domain(domain
);
1788 BUG_ON(domain
->dev_cnt
!= 0);
1790 free_pagetable(domain
);
1792 domain_id_free(domain
->id
);
1799 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
1802 struct protection_domain
*domain
= dom
->priv
;
1803 struct amd_iommu
*iommu
;
1804 struct pci_dev
*pdev
;
1807 if (dev
->bus
!= &pci_bus_type
)
1810 pdev
= to_pci_dev(dev
);
1812 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1815 detach_device(domain
, devid
);
1817 iommu
= amd_iommu_rlookup_table
[devid
];
1821 iommu_queue_inv_dev_entry(iommu
, devid
);
1822 iommu_completion_wait(iommu
);
1825 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
1828 struct protection_domain
*domain
= dom
->priv
;
1829 struct protection_domain
*old_domain
;
1830 struct amd_iommu
*iommu
;
1831 struct pci_dev
*pdev
;
1834 if (dev
->bus
!= &pci_bus_type
)
1837 pdev
= to_pci_dev(dev
);
1839 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1841 if (devid
>= amd_iommu_last_bdf
||
1842 devid
!= amd_iommu_alias_table
[devid
])
1845 iommu
= amd_iommu_rlookup_table
[devid
];
1849 old_domain
= domain_for_device(devid
);
1853 attach_device(iommu
, domain
, devid
);
1855 iommu_completion_wait(iommu
);
1860 static int amd_iommu_map_range(struct iommu_domain
*dom
,
1861 unsigned long iova
, phys_addr_t paddr
,
1862 size_t size
, int iommu_prot
)
1864 struct protection_domain
*domain
= dom
->priv
;
1865 unsigned long i
, npages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1869 if (iommu_prot
& IOMMU_READ
)
1870 prot
|= IOMMU_PROT_IR
;
1871 if (iommu_prot
& IOMMU_WRITE
)
1872 prot
|= IOMMU_PROT_IW
;
1877 for (i
= 0; i
< npages
; ++i
) {
1878 ret
= iommu_map_page(domain
, iova
, paddr
, prot
);
1889 static void amd_iommu_unmap_range(struct iommu_domain
*dom
,
1890 unsigned long iova
, size_t size
)
1893 struct protection_domain
*domain
= dom
->priv
;
1894 unsigned long i
, npages
= iommu_num_pages(iova
, size
, PAGE_SIZE
);
1898 for (i
= 0; i
< npages
; ++i
) {
1899 iommu_unmap_page(domain
, iova
);
1903 iommu_flush_domain(domain
->id
);
1906 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
1909 struct protection_domain
*domain
= dom
->priv
;
1910 unsigned long offset
= iova
& ~PAGE_MASK
;
1914 pte
= &domain
->pt_root
[IOMMU_PTE_L2_INDEX(iova
)];
1916 if (!IOMMU_PTE_PRESENT(*pte
))
1919 pte
= IOMMU_PTE_PAGE(*pte
);
1920 pte
= &pte
[IOMMU_PTE_L1_INDEX(iova
)];
1922 if (!IOMMU_PTE_PRESENT(*pte
))
1925 pte
= IOMMU_PTE_PAGE(*pte
);
1926 pte
= &pte
[IOMMU_PTE_L0_INDEX(iova
)];
1928 if (!IOMMU_PTE_PRESENT(*pte
))
1931 paddr
= *pte
& IOMMU_PAGE_MASK
;
1937 static struct iommu_ops amd_iommu_ops
= {
1938 .domain_init
= amd_iommu_domain_init
,
1939 .domain_destroy
= amd_iommu_domain_destroy
,
1940 .attach_dev
= amd_iommu_attach_device
,
1941 .detach_dev
= amd_iommu_detach_device
,
1942 .map
= amd_iommu_map_range
,
1943 .unmap
= amd_iommu_unmap_range
,
1944 .iova_to_phys
= amd_iommu_iova_to_phys
,