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/scatterlist.h>
24 #include <linux/iommu-helper.h>
25 #ifdef CONFIG_IOMMU_API
26 #include <linux/iommu.h>
28 #include <asm/proto.h>
29 #include <asm/iommu.h>
31 #include <asm/amd_iommu_types.h>
32 #include <asm/amd_iommu.h>
34 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
36 #define EXIT_LOOP_COUNT 10000000
38 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
40 /* A list of preallocated protection domains */
41 static LIST_HEAD(iommu_pd_list
);
42 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
45 * general struct to manage commands send to an IOMMU
51 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
52 struct unity_map_entry
*e
);
53 static struct dma_ops_domain
*find_protection_domain(u16 devid
);
56 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
57 static int iommu_has_npcache(struct amd_iommu
*iommu
)
59 return iommu
->cap
& (1UL << IOMMU_CAP_NPCACHE
);
62 /****************************************************************************
64 * Interrupt handling functions
66 ****************************************************************************/
68 static void iommu_print_event(void *__evt
)
71 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
72 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
73 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
74 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
75 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
77 printk(KERN_ERR
"AMD IOMMU: Event logged [");
80 case EVENT_TYPE_ILL_DEV
:
81 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
82 "address=0x%016llx flags=0x%04x]\n",
83 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
86 case EVENT_TYPE_IO_FAULT
:
87 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
88 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
89 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
90 domid
, address
, flags
);
92 case EVENT_TYPE_DEV_TAB_ERR
:
93 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
94 "address=0x%016llx flags=0x%04x]\n",
95 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
98 case EVENT_TYPE_PAGE_TAB_ERR
:
99 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
100 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
101 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
102 domid
, address
, flags
);
104 case EVENT_TYPE_ILL_CMD
:
105 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
107 case EVENT_TYPE_CMD_HARD_ERR
:
108 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
109 "flags=0x%04x]\n", address
, flags
);
111 case EVENT_TYPE_IOTLB_INV_TO
:
112 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
113 "address=0x%016llx]\n",
114 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
117 case EVENT_TYPE_INV_DEV_REQ
:
118 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
119 "address=0x%016llx flags=0x%04x]\n",
120 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
124 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
128 static void iommu_poll_events(struct amd_iommu
*iommu
)
133 spin_lock_irqsave(&iommu
->lock
, flags
);
135 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
136 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
138 while (head
!= tail
) {
139 iommu_print_event(iommu
->evt_buf
+ head
);
140 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
143 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
145 spin_unlock_irqrestore(&iommu
->lock
, flags
);
148 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
150 struct amd_iommu
*iommu
;
152 list_for_each_entry(iommu
, &amd_iommu_list
, list
)
153 iommu_poll_events(iommu
);
158 /****************************************************************************
160 * IOMMU command queuing functions
162 ****************************************************************************/
165 * Writes the command to the IOMMUs command buffer and informs the
166 * hardware about the new command. Must be called with iommu->lock held.
168 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
173 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
174 target
= iommu
->cmd_buf
+ tail
;
175 memcpy_toio(target
, cmd
, sizeof(*cmd
));
176 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
177 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
180 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
186 * General queuing function for commands. Takes iommu->lock and calls
187 * __iommu_queue_command().
189 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
194 spin_lock_irqsave(&iommu
->lock
, flags
);
195 ret
= __iommu_queue_command(iommu
, cmd
);
197 iommu
->need_sync
= 1;
198 spin_unlock_irqrestore(&iommu
->lock
, flags
);
204 * This function waits until an IOMMU has completed a completion
207 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
213 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
215 /* wait for the bit to become one */
216 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
217 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
220 /* set bit back to zero */
221 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
222 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
224 if (unlikely(i
== EXIT_LOOP_COUNT
))
225 panic("AMD IOMMU: Completion wait loop failed\n");
229 * This function queues a completion wait command into the command
232 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
234 struct iommu_cmd cmd
;
236 memset(&cmd
, 0, sizeof(cmd
));
237 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
238 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
240 return __iommu_queue_command(iommu
, &cmd
);
244 * This function is called whenever we need to ensure that the IOMMU has
245 * completed execution of all commands we sent. It sends a
246 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
247 * us about that by writing a value to a physical address we pass with
250 static int iommu_completion_wait(struct amd_iommu
*iommu
)
255 spin_lock_irqsave(&iommu
->lock
, flags
);
257 if (!iommu
->need_sync
)
260 ret
= __iommu_completion_wait(iommu
);
262 iommu
->need_sync
= 0;
267 __iommu_wait_for_completion(iommu
);
270 spin_unlock_irqrestore(&iommu
->lock
, flags
);
276 * Command send function for invalidating a device table entry
278 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
280 struct iommu_cmd cmd
;
283 BUG_ON(iommu
== NULL
);
285 memset(&cmd
, 0, sizeof(cmd
));
286 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
289 ret
= iommu_queue_command(iommu
, &cmd
);
294 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
295 u16 domid
, int pde
, int s
)
297 memset(cmd
, 0, sizeof(*cmd
));
298 address
&= PAGE_MASK
;
299 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
300 cmd
->data
[1] |= domid
;
301 cmd
->data
[2] = lower_32_bits(address
);
302 cmd
->data
[3] = upper_32_bits(address
);
303 if (s
) /* size bit - we flush more than one 4kb page */
304 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
305 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
306 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
310 * Generic command send function for invalidaing TLB entries
312 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
313 u64 address
, u16 domid
, int pde
, int s
)
315 struct iommu_cmd cmd
;
318 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
320 ret
= iommu_queue_command(iommu
, &cmd
);
326 * TLB invalidation function which is called from the mapping functions.
327 * It invalidates a single PTE if the range to flush is within a single
328 * page. Otherwise it flushes the whole TLB of the IOMMU.
330 static int iommu_flush_pages(struct amd_iommu
*iommu
, u16 domid
,
331 u64 address
, size_t size
)
334 unsigned pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
336 address
&= PAGE_MASK
;
340 * If we have to flush more than one page, flush all
341 * TLB entries for this domain
343 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
347 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, s
);
352 /* Flush the whole IO/TLB for a given protection domain */
353 static void iommu_flush_tlb(struct amd_iommu
*iommu
, u16 domid
)
355 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
357 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, 1);
360 #ifdef CONFIG_IOMMU_API
362 * This function is used to flush the IO/TLB for a given protection domain
363 * on every IOMMU in the system
365 static void iommu_flush_domain(u16 domid
)
368 struct amd_iommu
*iommu
;
369 struct iommu_cmd cmd
;
371 __iommu_build_inv_iommu_pages(&cmd
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
374 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
375 spin_lock_irqsave(&iommu
->lock
, flags
);
376 __iommu_queue_command(iommu
, &cmd
);
377 __iommu_completion_wait(iommu
);
378 __iommu_wait_for_completion(iommu
);
379 spin_unlock_irqrestore(&iommu
->lock
, flags
);
384 /****************************************************************************
386 * The functions below are used the create the page table mappings for
387 * unity mapped regions.
389 ****************************************************************************/
392 * Generic mapping functions. It maps a physical address into a DMA
393 * address space. It allocates the page table pages if necessary.
394 * In the future it can be extended to a generic mapping function
395 * supporting all features of AMD IOMMU page tables like level skipping
396 * and full 64 bit address spaces.
398 static int iommu_map_page(struct protection_domain
*dom
,
399 unsigned long bus_addr
,
400 unsigned long phys_addr
,
403 u64 __pte
, *pte
, *page
;
405 bus_addr
= PAGE_ALIGN(bus_addr
);
406 phys_addr
= PAGE_ALIGN(phys_addr
);
408 /* only support 512GB address spaces for now */
409 if (bus_addr
> IOMMU_MAP_SIZE_L3
|| !(prot
& IOMMU_PROT_MASK
))
412 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(bus_addr
)];
414 if (!IOMMU_PTE_PRESENT(*pte
)) {
415 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
418 *pte
= IOMMU_L2_PDE(virt_to_phys(page
));
421 pte
= IOMMU_PTE_PAGE(*pte
);
422 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
424 if (!IOMMU_PTE_PRESENT(*pte
)) {
425 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
428 *pte
= IOMMU_L1_PDE(virt_to_phys(page
));
431 pte
= IOMMU_PTE_PAGE(*pte
);
432 pte
= &pte
[IOMMU_PTE_L0_INDEX(bus_addr
)];
434 if (IOMMU_PTE_PRESENT(*pte
))
437 __pte
= phys_addr
| IOMMU_PTE_P
;
438 if (prot
& IOMMU_PROT_IR
)
439 __pte
|= IOMMU_PTE_IR
;
440 if (prot
& IOMMU_PROT_IW
)
441 __pte
|= IOMMU_PTE_IW
;
449 * This function checks if a specific unity mapping entry is needed for
450 * this specific IOMMU.
452 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
453 struct unity_map_entry
*entry
)
457 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
458 bdf
= amd_iommu_alias_table
[i
];
459 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
467 * Init the unity mappings for a specific IOMMU in the system
469 * Basically iterates over all unity mapping entries and applies them to
470 * the default domain DMA of that IOMMU if necessary.
472 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
474 struct unity_map_entry
*entry
;
477 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
478 if (!iommu_for_unity_map(iommu
, entry
))
480 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
489 * This function actually applies the mapping to the page table of the
492 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
493 struct unity_map_entry
*e
)
498 for (addr
= e
->address_start
; addr
< e
->address_end
;
500 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
);
504 * if unity mapping is in aperture range mark the page
505 * as allocated in the aperture
507 if (addr
< dma_dom
->aperture_size
)
508 __set_bit(addr
>> PAGE_SHIFT
, dma_dom
->bitmap
);
515 * Inits the unity mappings required for a specific device
517 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
520 struct unity_map_entry
*e
;
523 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
524 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
526 ret
= dma_ops_unity_map(dma_dom
, e
);
534 /****************************************************************************
536 * The next functions belong to the address allocator for the dma_ops
537 * interface functions. They work like the allocators in the other IOMMU
538 * drivers. Its basically a bitmap which marks the allocated pages in
539 * the aperture. Maybe it could be enhanced in the future to a more
540 * efficient allocator.
542 ****************************************************************************/
545 * The address allocator core function.
547 * called with domain->lock held
549 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
550 struct dma_ops_domain
*dom
,
552 unsigned long align_mask
,
556 unsigned long address
;
557 unsigned long boundary_size
;
559 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
560 PAGE_SIZE
) >> PAGE_SHIFT
;
561 limit
= iommu_device_max_index(dom
->aperture_size
>> PAGE_SHIFT
, 0,
562 dma_mask
>> PAGE_SHIFT
);
564 if (dom
->next_bit
>= limit
) {
566 dom
->need_flush
= true;
569 address
= iommu_area_alloc(dom
->bitmap
, limit
, dom
->next_bit
, pages
,
570 0 , boundary_size
, align_mask
);
572 address
= iommu_area_alloc(dom
->bitmap
, limit
, 0, pages
,
573 0, boundary_size
, align_mask
);
574 dom
->need_flush
= true;
577 if (likely(address
!= -1)) {
578 dom
->next_bit
= address
+ pages
;
579 address
<<= PAGE_SHIFT
;
581 address
= bad_dma_address
;
583 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
589 * The address free function.
591 * called with domain->lock held
593 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
594 unsigned long address
,
597 address
>>= PAGE_SHIFT
;
598 iommu_area_free(dom
->bitmap
, address
, pages
);
600 if (address
>= dom
->next_bit
)
601 dom
->need_flush
= true;
604 /****************************************************************************
606 * The next functions belong to the domain allocation. A domain is
607 * allocated for every IOMMU as the default domain. If device isolation
608 * is enabled, every device get its own domain. The most important thing
609 * about domains is the page table mapping the DMA address space they
612 ****************************************************************************/
614 static u16
domain_id_alloc(void)
619 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
620 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
622 if (id
> 0 && id
< MAX_DOMAIN_ID
)
623 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
626 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
631 #ifdef CONFIG_IOMMU_API
632 static void domain_id_free(int id
)
636 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
637 if (id
> 0 && id
< MAX_DOMAIN_ID
)
638 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
639 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
644 * Used to reserve address ranges in the aperture (e.g. for exclusion
647 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
648 unsigned long start_page
,
651 unsigned int last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
653 if (start_page
+ pages
> last_page
)
654 pages
= last_page
- start_page
;
656 iommu_area_reserve(dom
->bitmap
, start_page
, pages
);
659 static void free_pagetable(struct protection_domain
*domain
)
664 p1
= domain
->pt_root
;
669 for (i
= 0; i
< 512; ++i
) {
670 if (!IOMMU_PTE_PRESENT(p1
[i
]))
673 p2
= IOMMU_PTE_PAGE(p1
[i
]);
674 for (j
= 0; j
< 512; ++j
) {
675 if (!IOMMU_PTE_PRESENT(p2
[j
]))
677 p3
= IOMMU_PTE_PAGE(p2
[j
]);
678 free_page((unsigned long)p3
);
681 free_page((unsigned long)p2
);
684 free_page((unsigned long)p1
);
686 domain
->pt_root
= NULL
;
690 * Free a domain, only used if something went wrong in the
691 * allocation path and we need to free an already allocated page table
693 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
698 free_pagetable(&dom
->domain
);
700 kfree(dom
->pte_pages
);
708 * Allocates a new protection domain usable for the dma_ops functions.
709 * It also intializes the page table and the address allocator data
710 * structures required for the dma_ops interface
712 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
,
715 struct dma_ops_domain
*dma_dom
;
716 unsigned i
, num_pte_pages
;
721 * Currently the DMA aperture must be between 32 MB and 1GB in size
723 if ((order
< 25) || (order
> 30))
726 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
730 spin_lock_init(&dma_dom
->domain
.lock
);
732 dma_dom
->domain
.id
= domain_id_alloc();
733 if (dma_dom
->domain
.id
== 0)
735 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
736 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
737 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
738 dma_dom
->domain
.priv
= dma_dom
;
739 if (!dma_dom
->domain
.pt_root
)
741 dma_dom
->aperture_size
= (1ULL << order
);
742 dma_dom
->bitmap
= kzalloc(dma_dom
->aperture_size
/ (PAGE_SIZE
* 8),
744 if (!dma_dom
->bitmap
)
747 * mark the first page as allocated so we never return 0 as
748 * a valid dma-address. So we can use 0 as error value
750 dma_dom
->bitmap
[0] = 1;
751 dma_dom
->next_bit
= 0;
753 dma_dom
->need_flush
= false;
754 dma_dom
->target_dev
= 0xffff;
756 /* Intialize the exclusion range if necessary */
757 if (iommu
->exclusion_start
&&
758 iommu
->exclusion_start
< dma_dom
->aperture_size
) {
759 unsigned long startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
760 int pages
= iommu_num_pages(iommu
->exclusion_start
,
761 iommu
->exclusion_length
,
763 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
767 * At the last step, build the page tables so we don't need to
768 * allocate page table pages in the dma_ops mapping/unmapping
771 num_pte_pages
= dma_dom
->aperture_size
/ (PAGE_SIZE
* 512);
772 dma_dom
->pte_pages
= kzalloc(num_pte_pages
* sizeof(void *),
774 if (!dma_dom
->pte_pages
)
777 l2_pde
= (u64
*)get_zeroed_page(GFP_KERNEL
);
781 dma_dom
->domain
.pt_root
[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde
));
783 for (i
= 0; i
< num_pte_pages
; ++i
) {
784 dma_dom
->pte_pages
[i
] = (u64
*)get_zeroed_page(GFP_KERNEL
);
785 if (!dma_dom
->pte_pages
[i
])
787 address
= virt_to_phys(dma_dom
->pte_pages
[i
]);
788 l2_pde
[i
] = IOMMU_L1_PDE(address
);
794 dma_ops_domain_free(dma_dom
);
800 * little helper function to check whether a given protection domain is a
803 static bool dma_ops_domain(struct protection_domain
*domain
)
805 return domain
->flags
& PD_DMA_OPS_MASK
;
809 * Find out the protection domain structure for a given PCI device. This
810 * will give us the pointer to the page table root for example.
812 static struct protection_domain
*domain_for_device(u16 devid
)
814 struct protection_domain
*dom
;
817 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
818 dom
= amd_iommu_pd_table
[devid
];
819 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
825 * If a device is not yet associated with a domain, this function does
826 * assigns it visible for the hardware
828 static void attach_device(struct amd_iommu
*iommu
,
829 struct protection_domain
*domain
,
833 u64 pte_root
= virt_to_phys(domain
->pt_root
);
835 domain
->dev_cnt
+= 1;
837 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
838 << DEV_ENTRY_MODE_SHIFT
;
839 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
841 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
842 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
843 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
844 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
846 amd_iommu_pd_table
[devid
] = domain
;
847 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
849 iommu_queue_inv_dev_entry(iommu
, devid
);
853 * Removes a device from a protection domain (unlocked)
855 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
859 spin_lock(&domain
->lock
);
861 /* remove domain from the lookup table */
862 amd_iommu_pd_table
[devid
] = NULL
;
864 /* remove entry from the device table seen by the hardware */
865 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
866 amd_iommu_dev_table
[devid
].data
[1] = 0;
867 amd_iommu_dev_table
[devid
].data
[2] = 0;
869 /* decrease reference counter */
870 domain
->dev_cnt
-= 1;
873 spin_unlock(&domain
->lock
);
877 * Removes a device from a protection domain (with devtable_lock held)
879 static void detach_device(struct protection_domain
*domain
, u16 devid
)
883 /* lock device table */
884 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
885 __detach_device(domain
, devid
);
886 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
889 static int device_change_notifier(struct notifier_block
*nb
,
890 unsigned long action
, void *data
)
892 struct device
*dev
= data
;
893 struct pci_dev
*pdev
= to_pci_dev(dev
);
894 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
895 struct protection_domain
*domain
;
896 struct dma_ops_domain
*dma_domain
;
897 struct amd_iommu
*iommu
;
899 if (devid
> amd_iommu_last_bdf
)
902 devid
= amd_iommu_alias_table
[devid
];
904 iommu
= amd_iommu_rlookup_table
[devid
];
908 domain
= domain_for_device(devid
);
910 if (domain
&& !dma_ops_domain(domain
))
911 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
912 "to a non-dma-ops domain\n", dev_name(dev
));
915 case BUS_NOTIFY_BOUND_DRIVER
:
918 dma_domain
= find_protection_domain(devid
);
920 dma_domain
= iommu
->default_dom
;
921 attach_device(iommu
, &dma_domain
->domain
, devid
);
922 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
923 "device %s\n", dma_domain
->domain
.id
, dev_name(dev
));
925 case BUS_NOTIFY_UNBIND_DRIVER
:
928 detach_device(domain
, devid
);
934 iommu_queue_inv_dev_entry(iommu
, devid
);
935 iommu_completion_wait(iommu
);
941 struct notifier_block device_nb
= {
942 .notifier_call
= device_change_notifier
,
945 /*****************************************************************************
947 * The next functions belong to the dma_ops mapping/unmapping code.
949 *****************************************************************************/
952 * This function checks if the driver got a valid device from the caller to
953 * avoid dereferencing invalid pointers.
955 static bool check_device(struct device
*dev
)
957 if (!dev
|| !dev
->dma_mask
)
964 * In this function the list of preallocated protection domains is traversed to
965 * find the domain for a specific device
967 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
969 struct dma_ops_domain
*entry
, *ret
= NULL
;
972 if (list_empty(&iommu_pd_list
))
975 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
977 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
978 if (entry
->target_dev
== devid
) {
984 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
990 * In the dma_ops path we only have the struct device. This function
991 * finds the corresponding IOMMU, the protection domain and the
992 * requestor id for a given device.
993 * If the device is not yet associated with a domain this is also done
996 static int get_device_resources(struct device
*dev
,
997 struct amd_iommu
**iommu
,
998 struct protection_domain
**domain
,
1001 struct dma_ops_domain
*dma_dom
;
1002 struct pci_dev
*pcidev
;
1009 if (dev
->bus
!= &pci_bus_type
)
1012 pcidev
= to_pci_dev(dev
);
1013 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1015 /* device not translated by any IOMMU in the system? */
1016 if (_bdf
> amd_iommu_last_bdf
)
1019 *bdf
= amd_iommu_alias_table
[_bdf
];
1021 *iommu
= amd_iommu_rlookup_table
[*bdf
];
1024 *domain
= domain_for_device(*bdf
);
1025 if (*domain
== NULL
) {
1026 dma_dom
= find_protection_domain(*bdf
);
1028 dma_dom
= (*iommu
)->default_dom
;
1029 *domain
= &dma_dom
->domain
;
1030 attach_device(*iommu
, *domain
, *bdf
);
1031 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
1032 "device ", (*domain
)->id
);
1033 print_devid(_bdf
, 1);
1036 if (domain_for_device(_bdf
) == NULL
)
1037 attach_device(*iommu
, *domain
, _bdf
);
1043 * This is the generic map function. It maps one 4kb page at paddr to
1044 * the given address in the DMA address space for the domain.
1046 static dma_addr_t
dma_ops_domain_map(struct amd_iommu
*iommu
,
1047 struct dma_ops_domain
*dom
,
1048 unsigned long address
,
1054 WARN_ON(address
> dom
->aperture_size
);
1058 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1059 pte
+= IOMMU_PTE_L0_INDEX(address
);
1061 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1063 if (direction
== DMA_TO_DEVICE
)
1064 __pte
|= IOMMU_PTE_IR
;
1065 else if (direction
== DMA_FROM_DEVICE
)
1066 __pte
|= IOMMU_PTE_IW
;
1067 else if (direction
== DMA_BIDIRECTIONAL
)
1068 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1074 return (dma_addr_t
)address
;
1078 * The generic unmapping function for on page in the DMA address space.
1080 static void dma_ops_domain_unmap(struct amd_iommu
*iommu
,
1081 struct dma_ops_domain
*dom
,
1082 unsigned long address
)
1086 if (address
>= dom
->aperture_size
)
1089 WARN_ON(address
& ~PAGE_MASK
|| address
>= dom
->aperture_size
);
1091 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1092 pte
+= IOMMU_PTE_L0_INDEX(address
);
1100 * This function contains common code for mapping of a physically
1101 * contiguous memory region into DMA address space. It is used by all
1102 * mapping functions provided with this IOMMU driver.
1103 * Must be called with the domain lock held.
1105 static dma_addr_t
__map_single(struct device
*dev
,
1106 struct amd_iommu
*iommu
,
1107 struct dma_ops_domain
*dma_dom
,
1114 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1115 dma_addr_t address
, start
;
1117 unsigned long align_mask
= 0;
1120 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1124 align_mask
= (1UL << get_order(size
)) - 1;
1126 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1128 if (unlikely(address
== bad_dma_address
))
1132 for (i
= 0; i
< pages
; ++i
) {
1133 dma_ops_domain_map(iommu
, dma_dom
, start
, paddr
, dir
);
1139 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1140 iommu_flush_tlb(iommu
, dma_dom
->domain
.id
);
1141 dma_dom
->need_flush
= false;
1142 } else if (unlikely(iommu_has_npcache(iommu
)))
1143 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, address
, size
);
1150 * Does the reverse of the __map_single function. Must be called with
1151 * the domain lock held too
1153 static void __unmap_single(struct amd_iommu
*iommu
,
1154 struct dma_ops_domain
*dma_dom
,
1155 dma_addr_t dma_addr
,
1159 dma_addr_t i
, start
;
1162 if ((dma_addr
== bad_dma_address
) ||
1163 (dma_addr
+ size
> dma_dom
->aperture_size
))
1166 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1167 dma_addr
&= PAGE_MASK
;
1170 for (i
= 0; i
< pages
; ++i
) {
1171 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1175 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1177 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1178 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, dma_addr
, size
);
1179 dma_dom
->need_flush
= false;
1184 * The exported map_single function for dma_ops.
1186 static dma_addr_t
map_single(struct device
*dev
, phys_addr_t paddr
,
1187 size_t size
, int dir
)
1189 unsigned long flags
;
1190 struct amd_iommu
*iommu
;
1191 struct protection_domain
*domain
;
1196 if (!check_device(dev
))
1197 return bad_dma_address
;
1199 dma_mask
= *dev
->dma_mask
;
1201 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1203 if (iommu
== NULL
|| domain
== NULL
)
1204 /* device not handled by any AMD IOMMU */
1205 return (dma_addr_t
)paddr
;
1207 if (!dma_ops_domain(domain
))
1208 return bad_dma_address
;
1210 spin_lock_irqsave(&domain
->lock
, flags
);
1211 addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
, size
, dir
, false,
1213 if (addr
== bad_dma_address
)
1216 iommu_completion_wait(iommu
);
1219 spin_unlock_irqrestore(&domain
->lock
, flags
);
1225 * The exported unmap_single function for dma_ops.
1227 static void unmap_single(struct device
*dev
, dma_addr_t dma_addr
,
1228 size_t size
, int dir
)
1230 unsigned long flags
;
1231 struct amd_iommu
*iommu
;
1232 struct protection_domain
*domain
;
1235 if (!check_device(dev
) ||
1236 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1237 /* device not handled by any AMD IOMMU */
1240 if (!dma_ops_domain(domain
))
1243 spin_lock_irqsave(&domain
->lock
, flags
);
1245 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, dir
);
1247 iommu_completion_wait(iommu
);
1249 spin_unlock_irqrestore(&domain
->lock
, flags
);
1253 * This is a special map_sg function which is used if we should map a
1254 * device which is not handled by an AMD IOMMU in the system.
1256 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1257 int nelems
, int dir
)
1259 struct scatterlist
*s
;
1262 for_each_sg(sglist
, s
, nelems
, i
) {
1263 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1264 s
->dma_length
= s
->length
;
1271 * The exported map_sg function for dma_ops (handles scatter-gather
1274 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1275 int nelems
, int dir
)
1277 unsigned long flags
;
1278 struct amd_iommu
*iommu
;
1279 struct protection_domain
*domain
;
1282 struct scatterlist
*s
;
1284 int mapped_elems
= 0;
1287 if (!check_device(dev
))
1290 dma_mask
= *dev
->dma_mask
;
1292 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1294 if (!iommu
|| !domain
)
1295 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1297 if (!dma_ops_domain(domain
))
1300 spin_lock_irqsave(&domain
->lock
, flags
);
1302 for_each_sg(sglist
, s
, nelems
, i
) {
1305 s
->dma_address
= __map_single(dev
, iommu
, domain
->priv
,
1306 paddr
, s
->length
, dir
, false,
1309 if (s
->dma_address
) {
1310 s
->dma_length
= s
->length
;
1316 iommu_completion_wait(iommu
);
1319 spin_unlock_irqrestore(&domain
->lock
, flags
);
1321 return mapped_elems
;
1323 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1325 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1326 s
->dma_length
, dir
);
1327 s
->dma_address
= s
->dma_length
= 0;
1336 * The exported map_sg function for dma_ops (handles scatter-gather
1339 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1340 int nelems
, int dir
)
1342 unsigned long flags
;
1343 struct amd_iommu
*iommu
;
1344 struct protection_domain
*domain
;
1345 struct scatterlist
*s
;
1349 if (!check_device(dev
) ||
1350 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1353 if (!dma_ops_domain(domain
))
1356 spin_lock_irqsave(&domain
->lock
, flags
);
1358 for_each_sg(sglist
, s
, nelems
, i
) {
1359 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1360 s
->dma_length
, dir
);
1361 s
->dma_address
= s
->dma_length
= 0;
1364 iommu_completion_wait(iommu
);
1366 spin_unlock_irqrestore(&domain
->lock
, flags
);
1370 * The exported alloc_coherent function for dma_ops.
1372 static void *alloc_coherent(struct device
*dev
, size_t size
,
1373 dma_addr_t
*dma_addr
, gfp_t flag
)
1375 unsigned long flags
;
1377 struct amd_iommu
*iommu
;
1378 struct protection_domain
*domain
;
1381 u64 dma_mask
= dev
->coherent_dma_mask
;
1383 if (!check_device(dev
))
1386 if (!get_device_resources(dev
, &iommu
, &domain
, &devid
))
1387 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1390 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1394 paddr
= virt_to_phys(virt_addr
);
1396 if (!iommu
|| !domain
) {
1397 *dma_addr
= (dma_addr_t
)paddr
;
1401 if (!dma_ops_domain(domain
))
1405 dma_mask
= *dev
->dma_mask
;
1407 spin_lock_irqsave(&domain
->lock
, flags
);
1409 *dma_addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
,
1410 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
1412 if (*dma_addr
== bad_dma_address
)
1415 iommu_completion_wait(iommu
);
1417 spin_unlock_irqrestore(&domain
->lock
, flags
);
1423 free_pages((unsigned long)virt_addr
, get_order(size
));
1429 * The exported free_coherent function for dma_ops.
1431 static void free_coherent(struct device
*dev
, size_t size
,
1432 void *virt_addr
, dma_addr_t dma_addr
)
1434 unsigned long flags
;
1435 struct amd_iommu
*iommu
;
1436 struct protection_domain
*domain
;
1439 if (!check_device(dev
))
1442 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1444 if (!iommu
|| !domain
)
1447 if (!dma_ops_domain(domain
))
1450 spin_lock_irqsave(&domain
->lock
, flags
);
1452 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
1454 iommu_completion_wait(iommu
);
1456 spin_unlock_irqrestore(&domain
->lock
, flags
);
1459 free_pages((unsigned long)virt_addr
, get_order(size
));
1463 * This function is called by the DMA layer to find out if we can handle a
1464 * particular device. It is part of the dma_ops.
1466 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
1469 struct pci_dev
*pcidev
;
1471 /* No device or no PCI device */
1472 if (!dev
|| dev
->bus
!= &pci_bus_type
)
1475 pcidev
= to_pci_dev(dev
);
1477 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1479 /* Out of our scope? */
1480 if (bdf
> amd_iommu_last_bdf
)
1487 * The function for pre-allocating protection domains.
1489 * If the driver core informs the DMA layer if a driver grabs a device
1490 * we don't need to preallocate the protection domains anymore.
1491 * For now we have to.
1493 void prealloc_protection_domains(void)
1495 struct pci_dev
*dev
= NULL
;
1496 struct dma_ops_domain
*dma_dom
;
1497 struct amd_iommu
*iommu
;
1498 int order
= amd_iommu_aperture_order
;
1501 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
1502 devid
= (dev
->bus
->number
<< 8) | dev
->devfn
;
1503 if (devid
> amd_iommu_last_bdf
)
1505 devid
= amd_iommu_alias_table
[devid
];
1506 if (domain_for_device(devid
))
1508 iommu
= amd_iommu_rlookup_table
[devid
];
1511 dma_dom
= dma_ops_domain_alloc(iommu
, order
);
1514 init_unity_mappings_for_device(dma_dom
, devid
);
1515 dma_dom
->target_dev
= devid
;
1517 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
1521 static struct dma_mapping_ops amd_iommu_dma_ops
= {
1522 .alloc_coherent
= alloc_coherent
,
1523 .free_coherent
= free_coherent
,
1524 .map_single
= map_single
,
1525 .unmap_single
= unmap_single
,
1527 .unmap_sg
= unmap_sg
,
1528 .dma_supported
= amd_iommu_dma_supported
,
1532 * The function which clues the AMD IOMMU driver into dma_ops.
1534 int __init
amd_iommu_init_dma_ops(void)
1536 struct amd_iommu
*iommu
;
1537 int order
= amd_iommu_aperture_order
;
1541 * first allocate a default protection domain for every IOMMU we
1542 * found in the system. Devices not assigned to any other
1543 * protection domain will be assigned to the default one.
1545 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1546 iommu
->default_dom
= dma_ops_domain_alloc(iommu
, order
);
1547 if (iommu
->default_dom
== NULL
)
1549 ret
= iommu_init_unity_mappings(iommu
);
1555 * If device isolation is enabled, pre-allocate the protection
1556 * domains for each device.
1558 if (amd_iommu_isolate
)
1559 prealloc_protection_domains();
1563 bad_dma_address
= 0;
1564 #ifdef CONFIG_GART_IOMMU
1565 gart_iommu_aperture_disabled
= 1;
1566 gart_iommu_aperture
= 0;
1569 /* Make the driver finally visible to the drivers */
1570 dma_ops
= &amd_iommu_dma_ops
;
1572 bus_register_notifier(&pci_bus_type
, &device_nb
);
1578 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1579 if (iommu
->default_dom
)
1580 dma_ops_domain_free(iommu
->default_dom
);
1586 /*****************************************************************************
1588 * The following functions belong to the exported interface of AMD IOMMU
1590 * This interface allows access to lower level functions of the IOMMU
1591 * like protection domain handling and assignement of devices to domains
1592 * which is not possible with the dma_ops interface.
1594 *****************************************************************************/
1596 #ifdef CONFIG_IOMMU_API
1598 static void cleanup_domain(struct protection_domain
*domain
)
1600 unsigned long flags
;
1603 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1605 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
1606 if (amd_iommu_pd_table
[devid
] == domain
)
1607 __detach_device(domain
, devid
);
1609 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1612 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
1614 struct protection_domain
*domain
;
1616 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
1620 spin_lock_init(&domain
->lock
);
1621 domain
->mode
= PAGE_MODE_3_LEVEL
;
1622 domain
->id
= domain_id_alloc();
1625 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1626 if (!domain
->pt_root
)
1639 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
1641 struct protection_domain
*domain
= dom
->priv
;
1646 if (domain
->dev_cnt
> 0)
1647 cleanup_domain(domain
);
1649 BUG_ON(domain
->dev_cnt
!= 0);
1651 free_pagetable(domain
);
1653 domain_id_free(domain
->id
);
1660 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
1663 struct protection_domain
*domain
= dom
->priv
;
1664 struct amd_iommu
*iommu
;
1665 struct pci_dev
*pdev
;
1668 if (dev
->bus
!= &pci_bus_type
)
1671 pdev
= to_pci_dev(dev
);
1673 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1676 detach_device(domain
, devid
);
1678 iommu
= amd_iommu_rlookup_table
[devid
];
1682 iommu_queue_inv_dev_entry(iommu
, devid
);
1683 iommu_completion_wait(iommu
);
1686 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
1689 struct protection_domain
*domain
= dom
->priv
;
1690 struct protection_domain
*old_domain
;
1691 struct amd_iommu
*iommu
;
1692 struct pci_dev
*pdev
;
1695 if (dev
->bus
!= &pci_bus_type
)
1698 pdev
= to_pci_dev(dev
);
1700 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1702 if (devid
>= amd_iommu_last_bdf
||
1703 devid
!= amd_iommu_alias_table
[devid
])
1706 iommu
= amd_iommu_rlookup_table
[devid
];
1710 old_domain
= domain_for_device(devid
);
1714 attach_device(iommu
, domain
, devid
);
1716 iommu_completion_wait(iommu
);