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 #include <asm/proto.h>
26 #include <asm/iommu.h>
28 #include <asm/amd_iommu_types.h>
29 #include <asm/amd_iommu.h>
31 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
33 #define EXIT_LOOP_COUNT 10000000
35 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
37 /* A list of preallocated protection domains */
38 static LIST_HEAD(iommu_pd_list
);
39 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
42 * general struct to manage commands send to an IOMMU
48 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
49 struct unity_map_entry
*e
);
50 static struct dma_ops_domain
*find_protection_domain(u16 devid
);
53 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
54 static int iommu_has_npcache(struct amd_iommu
*iommu
)
56 return iommu
->cap
& (1UL << IOMMU_CAP_NPCACHE
);
59 /****************************************************************************
61 * Interrupt handling functions
63 ****************************************************************************/
65 static void iommu_print_event(void *__evt
)
68 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
69 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
70 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
71 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
72 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
74 printk(KERN_ERR
"AMD IOMMU: Event logged [");
77 case EVENT_TYPE_ILL_DEV
:
78 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
79 "address=0x%016llx flags=0x%04x]\n",
80 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
83 case EVENT_TYPE_IO_FAULT
:
84 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
85 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
86 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
87 domid
, address
, flags
);
89 case EVENT_TYPE_DEV_TAB_ERR
:
90 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
91 "address=0x%016llx flags=0x%04x]\n",
92 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
95 case EVENT_TYPE_PAGE_TAB_ERR
:
96 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
97 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
98 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
99 domid
, address
, flags
);
101 case EVENT_TYPE_ILL_CMD
:
102 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
104 case EVENT_TYPE_CMD_HARD_ERR
:
105 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
106 "flags=0x%04x]\n", address
, flags
);
108 case EVENT_TYPE_IOTLB_INV_TO
:
109 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
110 "address=0x%016llx]\n",
111 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
114 case EVENT_TYPE_INV_DEV_REQ
:
115 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
116 "address=0x%016llx flags=0x%04x]\n",
117 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
121 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
125 static void iommu_poll_events(struct amd_iommu
*iommu
)
130 spin_lock_irqsave(&iommu
->lock
, flags
);
132 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
133 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
135 while (head
!= tail
) {
136 iommu_print_event(iommu
->evt_buf
+ head
);
137 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
140 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
142 spin_unlock_irqrestore(&iommu
->lock
, flags
);
145 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
147 struct amd_iommu
*iommu
;
149 list_for_each_entry(iommu
, &amd_iommu_list
, list
)
150 iommu_poll_events(iommu
);
155 /****************************************************************************
157 * IOMMU command queuing functions
159 ****************************************************************************/
162 * Writes the command to the IOMMUs command buffer and informs the
163 * hardware about the new command. Must be called with iommu->lock held.
165 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
170 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
171 target
= iommu
->cmd_buf
+ tail
;
172 memcpy_toio(target
, cmd
, sizeof(*cmd
));
173 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
174 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
177 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
183 * General queuing function for commands. Takes iommu->lock and calls
184 * __iommu_queue_command().
186 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
191 spin_lock_irqsave(&iommu
->lock
, flags
);
192 ret
= __iommu_queue_command(iommu
, cmd
);
194 iommu
->need_sync
= 1;
195 spin_unlock_irqrestore(&iommu
->lock
, flags
);
201 * This function waits until an IOMMU has completed a completion
204 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
210 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
212 /* wait for the bit to become one */
213 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
214 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
217 /* set bit back to zero */
218 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
219 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
221 if (unlikely(i
== EXIT_LOOP_COUNT
))
222 panic("AMD IOMMU: Completion wait loop failed\n");
226 * This function queues a completion wait command into the command
229 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
231 struct iommu_cmd cmd
;
233 memset(&cmd
, 0, sizeof(cmd
));
234 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
235 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
237 return __iommu_queue_command(iommu
, &cmd
);
241 * This function is called whenever we need to ensure that the IOMMU has
242 * completed execution of all commands we sent. It sends a
243 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
244 * us about that by writing a value to a physical address we pass with
247 static int iommu_completion_wait(struct amd_iommu
*iommu
)
252 spin_lock_irqsave(&iommu
->lock
, flags
);
254 if (!iommu
->need_sync
)
257 ret
= __iommu_completion_wait(iommu
);
259 iommu
->need_sync
= 0;
264 __iommu_wait_for_completion(iommu
);
267 spin_unlock_irqrestore(&iommu
->lock
, flags
);
273 * Command send function for invalidating a device table entry
275 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
277 struct iommu_cmd cmd
;
280 BUG_ON(iommu
== NULL
);
282 memset(&cmd
, 0, sizeof(cmd
));
283 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
286 ret
= iommu_queue_command(iommu
, &cmd
);
291 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
292 u16 domid
, int pde
, int s
)
294 memset(cmd
, 0, sizeof(*cmd
));
295 address
&= PAGE_MASK
;
296 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
297 cmd
->data
[1] |= domid
;
298 cmd
->data
[2] = lower_32_bits(address
);
299 cmd
->data
[3] = upper_32_bits(address
);
300 if (s
) /* size bit - we flush more than one 4kb page */
301 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
302 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
303 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
307 * Generic command send function for invalidaing TLB entries
309 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
310 u64 address
, u16 domid
, int pde
, int s
)
312 struct iommu_cmd cmd
;
315 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
317 ret
= iommu_queue_command(iommu
, &cmd
);
323 * TLB invalidation function which is called from the mapping functions.
324 * It invalidates a single PTE if the range to flush is within a single
325 * page. Otherwise it flushes the whole TLB of the IOMMU.
327 static int iommu_flush_pages(struct amd_iommu
*iommu
, u16 domid
,
328 u64 address
, size_t size
)
331 unsigned pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
333 address
&= PAGE_MASK
;
337 * If we have to flush more than one page, flush all
338 * TLB entries for this domain
340 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
344 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, s
);
349 /* Flush the whole IO/TLB for a given protection domain */
350 static void iommu_flush_tlb(struct amd_iommu
*iommu
, u16 domid
)
352 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
354 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, 1);
357 #ifdef CONFIG_IOMMU_API
359 * This function is used to flush the IO/TLB for a given protection domain
360 * on every IOMMU in the system
362 static void iommu_flush_domain(u16 domid
)
365 struct amd_iommu
*iommu
;
366 struct iommu_cmd cmd
;
368 __iommu_build_inv_iommu_pages(&cmd
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
371 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
372 spin_lock_irqsave(&iommu
->lock
, flags
);
373 __iommu_queue_command(iommu
, &cmd
);
374 __iommu_completion_wait(iommu
);
375 __iommu_wait_for_completion(iommu
);
376 spin_unlock_irqrestore(&iommu
->lock
, flags
);
381 /****************************************************************************
383 * The functions below are used the create the page table mappings for
384 * unity mapped regions.
386 ****************************************************************************/
389 * Generic mapping functions. It maps a physical address into a DMA
390 * address space. It allocates the page table pages if necessary.
391 * In the future it can be extended to a generic mapping function
392 * supporting all features of AMD IOMMU page tables like level skipping
393 * and full 64 bit address spaces.
395 static int iommu_map_page(struct protection_domain
*dom
,
396 unsigned long bus_addr
,
397 unsigned long phys_addr
,
400 u64 __pte
, *pte
, *page
;
402 bus_addr
= PAGE_ALIGN(bus_addr
);
403 phys_addr
= PAGE_ALIGN(phys_addr
);
405 /* only support 512GB address spaces for now */
406 if (bus_addr
> IOMMU_MAP_SIZE_L3
|| !(prot
& IOMMU_PROT_MASK
))
409 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(bus_addr
)];
411 if (!IOMMU_PTE_PRESENT(*pte
)) {
412 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
415 *pte
= IOMMU_L2_PDE(virt_to_phys(page
));
418 pte
= IOMMU_PTE_PAGE(*pte
);
419 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
421 if (!IOMMU_PTE_PRESENT(*pte
)) {
422 page
= (u64
*)get_zeroed_page(GFP_KERNEL
);
425 *pte
= IOMMU_L1_PDE(virt_to_phys(page
));
428 pte
= IOMMU_PTE_PAGE(*pte
);
429 pte
= &pte
[IOMMU_PTE_L0_INDEX(bus_addr
)];
431 if (IOMMU_PTE_PRESENT(*pte
))
434 __pte
= phys_addr
| IOMMU_PTE_P
;
435 if (prot
& IOMMU_PROT_IR
)
436 __pte
|= IOMMU_PTE_IR
;
437 if (prot
& IOMMU_PROT_IW
)
438 __pte
|= IOMMU_PTE_IW
;
446 * This function checks if a specific unity mapping entry is needed for
447 * this specific IOMMU.
449 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
450 struct unity_map_entry
*entry
)
454 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
455 bdf
= amd_iommu_alias_table
[i
];
456 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
464 * Init the unity mappings for a specific IOMMU in the system
466 * Basically iterates over all unity mapping entries and applies them to
467 * the default domain DMA of that IOMMU if necessary.
469 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
471 struct unity_map_entry
*entry
;
474 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
475 if (!iommu_for_unity_map(iommu
, entry
))
477 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
486 * This function actually applies the mapping to the page table of the
489 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
490 struct unity_map_entry
*e
)
495 for (addr
= e
->address_start
; addr
< e
->address_end
;
497 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
);
501 * if unity mapping is in aperture range mark the page
502 * as allocated in the aperture
504 if (addr
< dma_dom
->aperture_size
)
505 __set_bit(addr
>> PAGE_SHIFT
, dma_dom
->bitmap
);
512 * Inits the unity mappings required for a specific device
514 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
517 struct unity_map_entry
*e
;
520 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
521 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
523 ret
= dma_ops_unity_map(dma_dom
, e
);
531 /****************************************************************************
533 * The next functions belong to the address allocator for the dma_ops
534 * interface functions. They work like the allocators in the other IOMMU
535 * drivers. Its basically a bitmap which marks the allocated pages in
536 * the aperture. Maybe it could be enhanced in the future to a more
537 * efficient allocator.
539 ****************************************************************************/
542 * The address allocator core function.
544 * called with domain->lock held
546 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
547 struct dma_ops_domain
*dom
,
549 unsigned long align_mask
,
553 unsigned long address
;
554 unsigned long boundary_size
;
556 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
557 PAGE_SIZE
) >> PAGE_SHIFT
;
558 limit
= iommu_device_max_index(dom
->aperture_size
>> PAGE_SHIFT
, 0,
559 dma_mask
>> PAGE_SHIFT
);
561 if (dom
->next_bit
>= limit
) {
563 dom
->need_flush
= true;
566 address
= iommu_area_alloc(dom
->bitmap
, limit
, dom
->next_bit
, pages
,
567 0 , boundary_size
, align_mask
);
569 address
= iommu_area_alloc(dom
->bitmap
, limit
, 0, pages
,
570 0, boundary_size
, align_mask
);
571 dom
->need_flush
= true;
574 if (likely(address
!= -1)) {
575 dom
->next_bit
= address
+ pages
;
576 address
<<= PAGE_SHIFT
;
578 address
= bad_dma_address
;
580 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
586 * The address free function.
588 * called with domain->lock held
590 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
591 unsigned long address
,
594 address
>>= PAGE_SHIFT
;
595 iommu_area_free(dom
->bitmap
, address
, pages
);
597 if (address
>= dom
->next_bit
)
598 dom
->need_flush
= true;
601 /****************************************************************************
603 * The next functions belong to the domain allocation. A domain is
604 * allocated for every IOMMU as the default domain. If device isolation
605 * is enabled, every device get its own domain. The most important thing
606 * about domains is the page table mapping the DMA address space they
609 ****************************************************************************/
611 static u16
domain_id_alloc(void)
616 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
617 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
619 if (id
> 0 && id
< MAX_DOMAIN_ID
)
620 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
623 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
628 #ifdef CONFIG_IOMMU_API
629 static void domain_id_free(int id
)
633 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
634 if (id
> 0 && id
< MAX_DOMAIN_ID
)
635 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
636 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
641 * Used to reserve address ranges in the aperture (e.g. for exclusion
644 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
645 unsigned long start_page
,
648 unsigned int last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
650 if (start_page
+ pages
> last_page
)
651 pages
= last_page
- start_page
;
653 iommu_area_reserve(dom
->bitmap
, start_page
, pages
);
656 static void free_pagetable(struct protection_domain
*domain
)
661 p1
= domain
->pt_root
;
666 for (i
= 0; i
< 512; ++i
) {
667 if (!IOMMU_PTE_PRESENT(p1
[i
]))
670 p2
= IOMMU_PTE_PAGE(p1
[i
]);
671 for (j
= 0; j
< 512; ++j
) {
672 if (!IOMMU_PTE_PRESENT(p2
[j
]))
674 p3
= IOMMU_PTE_PAGE(p2
[j
]);
675 free_page((unsigned long)p3
);
678 free_page((unsigned long)p2
);
681 free_page((unsigned long)p1
);
683 domain
->pt_root
= NULL
;
687 * Free a domain, only used if something went wrong in the
688 * allocation path and we need to free an already allocated page table
690 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
695 free_pagetable(&dom
->domain
);
697 kfree(dom
->pte_pages
);
705 * Allocates a new protection domain usable for the dma_ops functions.
706 * It also intializes the page table and the address allocator data
707 * structures required for the dma_ops interface
709 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
,
712 struct dma_ops_domain
*dma_dom
;
713 unsigned i
, num_pte_pages
;
718 * Currently the DMA aperture must be between 32 MB and 1GB in size
720 if ((order
< 25) || (order
> 30))
723 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
727 spin_lock_init(&dma_dom
->domain
.lock
);
729 dma_dom
->domain
.id
= domain_id_alloc();
730 if (dma_dom
->domain
.id
== 0)
732 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
733 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
734 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
735 dma_dom
->domain
.priv
= dma_dom
;
736 if (!dma_dom
->domain
.pt_root
)
738 dma_dom
->aperture_size
= (1ULL << order
);
739 dma_dom
->bitmap
= kzalloc(dma_dom
->aperture_size
/ (PAGE_SIZE
* 8),
741 if (!dma_dom
->bitmap
)
744 * mark the first page as allocated so we never return 0 as
745 * a valid dma-address. So we can use 0 as error value
747 dma_dom
->bitmap
[0] = 1;
748 dma_dom
->next_bit
= 0;
750 dma_dom
->need_flush
= false;
751 dma_dom
->target_dev
= 0xffff;
753 /* Intialize the exclusion range if necessary */
754 if (iommu
->exclusion_start
&&
755 iommu
->exclusion_start
< dma_dom
->aperture_size
) {
756 unsigned long startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
757 int pages
= iommu_num_pages(iommu
->exclusion_start
,
758 iommu
->exclusion_length
,
760 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
764 * At the last step, build the page tables so we don't need to
765 * allocate page table pages in the dma_ops mapping/unmapping
768 num_pte_pages
= dma_dom
->aperture_size
/ (PAGE_SIZE
* 512);
769 dma_dom
->pte_pages
= kzalloc(num_pte_pages
* sizeof(void *),
771 if (!dma_dom
->pte_pages
)
774 l2_pde
= (u64
*)get_zeroed_page(GFP_KERNEL
);
778 dma_dom
->domain
.pt_root
[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde
));
780 for (i
= 0; i
< num_pte_pages
; ++i
) {
781 dma_dom
->pte_pages
[i
] = (u64
*)get_zeroed_page(GFP_KERNEL
);
782 if (!dma_dom
->pte_pages
[i
])
784 address
= virt_to_phys(dma_dom
->pte_pages
[i
]);
785 l2_pde
[i
] = IOMMU_L1_PDE(address
);
791 dma_ops_domain_free(dma_dom
);
797 * little helper function to check whether a given protection domain is a
800 static bool dma_ops_domain(struct protection_domain
*domain
)
802 return domain
->flags
& PD_DMA_OPS_MASK
;
806 * Find out the protection domain structure for a given PCI device. This
807 * will give us the pointer to the page table root for example.
809 static struct protection_domain
*domain_for_device(u16 devid
)
811 struct protection_domain
*dom
;
814 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
815 dom
= amd_iommu_pd_table
[devid
];
816 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
822 * If a device is not yet associated with a domain, this function does
823 * assigns it visible for the hardware
825 static void attach_device(struct amd_iommu
*iommu
,
826 struct protection_domain
*domain
,
830 u64 pte_root
= virt_to_phys(domain
->pt_root
);
832 domain
->dev_cnt
+= 1;
834 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
835 << DEV_ENTRY_MODE_SHIFT
;
836 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
838 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
839 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
840 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
841 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
843 amd_iommu_pd_table
[devid
] = domain
;
844 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
846 iommu_queue_inv_dev_entry(iommu
, devid
);
850 * Removes a device from a protection domain (unlocked)
852 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
856 spin_lock(&domain
->lock
);
858 /* remove domain from the lookup table */
859 amd_iommu_pd_table
[devid
] = NULL
;
861 /* remove entry from the device table seen by the hardware */
862 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
863 amd_iommu_dev_table
[devid
].data
[1] = 0;
864 amd_iommu_dev_table
[devid
].data
[2] = 0;
866 /* decrease reference counter */
867 domain
->dev_cnt
-= 1;
870 spin_unlock(&domain
->lock
);
874 * Removes a device from a protection domain (with devtable_lock held)
876 static void detach_device(struct protection_domain
*domain
, u16 devid
)
880 /* lock device table */
881 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
882 __detach_device(domain
, devid
);
883 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
886 static int device_change_notifier(struct notifier_block
*nb
,
887 unsigned long action
, void *data
)
889 struct device
*dev
= data
;
890 struct pci_dev
*pdev
= to_pci_dev(dev
);
891 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
892 struct protection_domain
*domain
;
893 struct dma_ops_domain
*dma_domain
;
894 struct amd_iommu
*iommu
;
896 if (devid
> amd_iommu_last_bdf
)
899 devid
= amd_iommu_alias_table
[devid
];
901 iommu
= amd_iommu_rlookup_table
[devid
];
905 domain
= domain_for_device(devid
);
907 if (domain
&& !dma_ops_domain(domain
))
908 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
909 "to a non-dma-ops domain\n", dev_name(dev
));
912 case BUS_NOTIFY_BOUND_DRIVER
:
915 dma_domain
= find_protection_domain(devid
);
917 dma_domain
= iommu
->default_dom
;
918 attach_device(iommu
, &dma_domain
->domain
, devid
);
919 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
920 "device %s\n", dma_domain
->domain
.id
, dev_name(dev
));
922 case BUS_NOTIFY_UNBIND_DRIVER
:
925 detach_device(domain
, devid
);
931 iommu_queue_inv_dev_entry(iommu
, devid
);
932 iommu_completion_wait(iommu
);
938 struct notifier_block device_nb
= {
939 .notifier_call
= device_change_notifier
,
942 /*****************************************************************************
944 * The next functions belong to the dma_ops mapping/unmapping code.
946 *****************************************************************************/
949 * This function checks if the driver got a valid device from the caller to
950 * avoid dereferencing invalid pointers.
952 static bool check_device(struct device
*dev
)
954 if (!dev
|| !dev
->dma_mask
)
961 * In this function the list of preallocated protection domains is traversed to
962 * find the domain for a specific device
964 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
966 struct dma_ops_domain
*entry
, *ret
= NULL
;
969 if (list_empty(&iommu_pd_list
))
972 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
974 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
975 if (entry
->target_dev
== devid
) {
981 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
987 * In the dma_ops path we only have the struct device. This function
988 * finds the corresponding IOMMU, the protection domain and the
989 * requestor id for a given device.
990 * If the device is not yet associated with a domain this is also done
993 static int get_device_resources(struct device
*dev
,
994 struct amd_iommu
**iommu
,
995 struct protection_domain
**domain
,
998 struct dma_ops_domain
*dma_dom
;
999 struct pci_dev
*pcidev
;
1006 if (dev
->bus
!= &pci_bus_type
)
1009 pcidev
= to_pci_dev(dev
);
1010 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1012 /* device not translated by any IOMMU in the system? */
1013 if (_bdf
> amd_iommu_last_bdf
)
1016 *bdf
= amd_iommu_alias_table
[_bdf
];
1018 *iommu
= amd_iommu_rlookup_table
[*bdf
];
1021 *domain
= domain_for_device(*bdf
);
1022 if (*domain
== NULL
) {
1023 dma_dom
= find_protection_domain(*bdf
);
1025 dma_dom
= (*iommu
)->default_dom
;
1026 *domain
= &dma_dom
->domain
;
1027 attach_device(*iommu
, *domain
, *bdf
);
1028 printk(KERN_INFO
"AMD IOMMU: Using protection domain %d for "
1029 "device ", (*domain
)->id
);
1030 print_devid(_bdf
, 1);
1033 if (domain_for_device(_bdf
) == NULL
)
1034 attach_device(*iommu
, *domain
, _bdf
);
1040 * This is the generic map function. It maps one 4kb page at paddr to
1041 * the given address in the DMA address space for the domain.
1043 static dma_addr_t
dma_ops_domain_map(struct amd_iommu
*iommu
,
1044 struct dma_ops_domain
*dom
,
1045 unsigned long address
,
1051 WARN_ON(address
> dom
->aperture_size
);
1055 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1056 pte
+= IOMMU_PTE_L0_INDEX(address
);
1058 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1060 if (direction
== DMA_TO_DEVICE
)
1061 __pte
|= IOMMU_PTE_IR
;
1062 else if (direction
== DMA_FROM_DEVICE
)
1063 __pte
|= IOMMU_PTE_IW
;
1064 else if (direction
== DMA_BIDIRECTIONAL
)
1065 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1071 return (dma_addr_t
)address
;
1075 * The generic unmapping function for on page in the DMA address space.
1077 static void dma_ops_domain_unmap(struct amd_iommu
*iommu
,
1078 struct dma_ops_domain
*dom
,
1079 unsigned long address
)
1083 if (address
>= dom
->aperture_size
)
1086 WARN_ON(address
& ~PAGE_MASK
|| address
>= dom
->aperture_size
);
1088 pte
= dom
->pte_pages
[IOMMU_PTE_L1_INDEX(address
)];
1089 pte
+= IOMMU_PTE_L0_INDEX(address
);
1097 * This function contains common code for mapping of a physically
1098 * contiguous memory region into DMA address space. It is used by all
1099 * mapping functions provided with this IOMMU driver.
1100 * Must be called with the domain lock held.
1102 static dma_addr_t
__map_single(struct device
*dev
,
1103 struct amd_iommu
*iommu
,
1104 struct dma_ops_domain
*dma_dom
,
1111 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1112 dma_addr_t address
, start
;
1114 unsigned long align_mask
= 0;
1117 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1121 align_mask
= (1UL << get_order(size
)) - 1;
1123 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1125 if (unlikely(address
== bad_dma_address
))
1129 for (i
= 0; i
< pages
; ++i
) {
1130 dma_ops_domain_map(iommu
, dma_dom
, start
, paddr
, dir
);
1136 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1137 iommu_flush_tlb(iommu
, dma_dom
->domain
.id
);
1138 dma_dom
->need_flush
= false;
1139 } else if (unlikely(iommu_has_npcache(iommu
)))
1140 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, address
, size
);
1147 * Does the reverse of the __map_single function. Must be called with
1148 * the domain lock held too
1150 static void __unmap_single(struct amd_iommu
*iommu
,
1151 struct dma_ops_domain
*dma_dom
,
1152 dma_addr_t dma_addr
,
1156 dma_addr_t i
, start
;
1159 if ((dma_addr
== bad_dma_address
) ||
1160 (dma_addr
+ size
> dma_dom
->aperture_size
))
1163 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1164 dma_addr
&= PAGE_MASK
;
1167 for (i
= 0; i
< pages
; ++i
) {
1168 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1172 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1174 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1175 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, dma_addr
, size
);
1176 dma_dom
->need_flush
= false;
1181 * The exported map_single function for dma_ops.
1183 static dma_addr_t
map_single(struct device
*dev
, phys_addr_t paddr
,
1184 size_t size
, int dir
)
1186 unsigned long flags
;
1187 struct amd_iommu
*iommu
;
1188 struct protection_domain
*domain
;
1193 if (!check_device(dev
))
1194 return bad_dma_address
;
1196 dma_mask
= *dev
->dma_mask
;
1198 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1200 if (iommu
== NULL
|| domain
== NULL
)
1201 /* device not handled by any AMD IOMMU */
1202 return (dma_addr_t
)paddr
;
1204 if (!dma_ops_domain(domain
))
1205 return bad_dma_address
;
1207 spin_lock_irqsave(&domain
->lock
, flags
);
1208 addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
, size
, dir
, false,
1210 if (addr
== bad_dma_address
)
1213 iommu_completion_wait(iommu
);
1216 spin_unlock_irqrestore(&domain
->lock
, flags
);
1222 * The exported unmap_single function for dma_ops.
1224 static void unmap_single(struct device
*dev
, dma_addr_t dma_addr
,
1225 size_t size
, int dir
)
1227 unsigned long flags
;
1228 struct amd_iommu
*iommu
;
1229 struct protection_domain
*domain
;
1232 if (!check_device(dev
) ||
1233 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1234 /* device not handled by any AMD IOMMU */
1237 if (!dma_ops_domain(domain
))
1240 spin_lock_irqsave(&domain
->lock
, flags
);
1242 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, dir
);
1244 iommu_completion_wait(iommu
);
1246 spin_unlock_irqrestore(&domain
->lock
, flags
);
1250 * This is a special map_sg function which is used if we should map a
1251 * device which is not handled by an AMD IOMMU in the system.
1253 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1254 int nelems
, int dir
)
1256 struct scatterlist
*s
;
1259 for_each_sg(sglist
, s
, nelems
, i
) {
1260 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1261 s
->dma_length
= s
->length
;
1268 * The exported map_sg function for dma_ops (handles scatter-gather
1271 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1272 int nelems
, int dir
)
1274 unsigned long flags
;
1275 struct amd_iommu
*iommu
;
1276 struct protection_domain
*domain
;
1279 struct scatterlist
*s
;
1281 int mapped_elems
= 0;
1284 if (!check_device(dev
))
1287 dma_mask
= *dev
->dma_mask
;
1289 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1291 if (!iommu
|| !domain
)
1292 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1294 if (!dma_ops_domain(domain
))
1297 spin_lock_irqsave(&domain
->lock
, flags
);
1299 for_each_sg(sglist
, s
, nelems
, i
) {
1302 s
->dma_address
= __map_single(dev
, iommu
, domain
->priv
,
1303 paddr
, s
->length
, dir
, false,
1306 if (s
->dma_address
) {
1307 s
->dma_length
= s
->length
;
1313 iommu_completion_wait(iommu
);
1316 spin_unlock_irqrestore(&domain
->lock
, flags
);
1318 return mapped_elems
;
1320 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1322 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1323 s
->dma_length
, dir
);
1324 s
->dma_address
= s
->dma_length
= 0;
1333 * The exported map_sg function for dma_ops (handles scatter-gather
1336 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1337 int nelems
, int dir
)
1339 unsigned long flags
;
1340 struct amd_iommu
*iommu
;
1341 struct protection_domain
*domain
;
1342 struct scatterlist
*s
;
1346 if (!check_device(dev
) ||
1347 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1350 if (!dma_ops_domain(domain
))
1353 spin_lock_irqsave(&domain
->lock
, flags
);
1355 for_each_sg(sglist
, s
, nelems
, i
) {
1356 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1357 s
->dma_length
, dir
);
1358 s
->dma_address
= s
->dma_length
= 0;
1361 iommu_completion_wait(iommu
);
1363 spin_unlock_irqrestore(&domain
->lock
, flags
);
1367 * The exported alloc_coherent function for dma_ops.
1369 static void *alloc_coherent(struct device
*dev
, size_t size
,
1370 dma_addr_t
*dma_addr
, gfp_t flag
)
1372 unsigned long flags
;
1374 struct amd_iommu
*iommu
;
1375 struct protection_domain
*domain
;
1378 u64 dma_mask
= dev
->coherent_dma_mask
;
1380 if (!check_device(dev
))
1383 if (!get_device_resources(dev
, &iommu
, &domain
, &devid
))
1384 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1387 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1391 paddr
= virt_to_phys(virt_addr
);
1393 if (!iommu
|| !domain
) {
1394 *dma_addr
= (dma_addr_t
)paddr
;
1398 if (!dma_ops_domain(domain
))
1402 dma_mask
= *dev
->dma_mask
;
1404 spin_lock_irqsave(&domain
->lock
, flags
);
1406 *dma_addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
,
1407 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
1409 if (*dma_addr
== bad_dma_address
)
1412 iommu_completion_wait(iommu
);
1414 spin_unlock_irqrestore(&domain
->lock
, flags
);
1420 free_pages((unsigned long)virt_addr
, get_order(size
));
1426 * The exported free_coherent function for dma_ops.
1428 static void free_coherent(struct device
*dev
, size_t size
,
1429 void *virt_addr
, dma_addr_t dma_addr
)
1431 unsigned long flags
;
1432 struct amd_iommu
*iommu
;
1433 struct protection_domain
*domain
;
1436 if (!check_device(dev
))
1439 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1441 if (!iommu
|| !domain
)
1444 if (!dma_ops_domain(domain
))
1447 spin_lock_irqsave(&domain
->lock
, flags
);
1449 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
1451 iommu_completion_wait(iommu
);
1453 spin_unlock_irqrestore(&domain
->lock
, flags
);
1456 free_pages((unsigned long)virt_addr
, get_order(size
));
1460 * This function is called by the DMA layer to find out if we can handle a
1461 * particular device. It is part of the dma_ops.
1463 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
1466 struct pci_dev
*pcidev
;
1468 /* No device or no PCI device */
1469 if (!dev
|| dev
->bus
!= &pci_bus_type
)
1472 pcidev
= to_pci_dev(dev
);
1474 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1476 /* Out of our scope? */
1477 if (bdf
> amd_iommu_last_bdf
)
1484 * The function for pre-allocating protection domains.
1486 * If the driver core informs the DMA layer if a driver grabs a device
1487 * we don't need to preallocate the protection domains anymore.
1488 * For now we have to.
1490 void prealloc_protection_domains(void)
1492 struct pci_dev
*dev
= NULL
;
1493 struct dma_ops_domain
*dma_dom
;
1494 struct amd_iommu
*iommu
;
1495 int order
= amd_iommu_aperture_order
;
1498 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
1499 devid
= (dev
->bus
->number
<< 8) | dev
->devfn
;
1500 if (devid
> amd_iommu_last_bdf
)
1502 devid
= amd_iommu_alias_table
[devid
];
1503 if (domain_for_device(devid
))
1505 iommu
= amd_iommu_rlookup_table
[devid
];
1508 dma_dom
= dma_ops_domain_alloc(iommu
, order
);
1511 init_unity_mappings_for_device(dma_dom
, devid
);
1512 dma_dom
->target_dev
= devid
;
1514 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
1518 static struct dma_mapping_ops amd_iommu_dma_ops
= {
1519 .alloc_coherent
= alloc_coherent
,
1520 .free_coherent
= free_coherent
,
1521 .map_single
= map_single
,
1522 .unmap_single
= unmap_single
,
1524 .unmap_sg
= unmap_sg
,
1525 .dma_supported
= amd_iommu_dma_supported
,
1529 * The function which clues the AMD IOMMU driver into dma_ops.
1531 int __init
amd_iommu_init_dma_ops(void)
1533 struct amd_iommu
*iommu
;
1534 int order
= amd_iommu_aperture_order
;
1538 * first allocate a default protection domain for every IOMMU we
1539 * found in the system. Devices not assigned to any other
1540 * protection domain will be assigned to the default one.
1542 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1543 iommu
->default_dom
= dma_ops_domain_alloc(iommu
, order
);
1544 if (iommu
->default_dom
== NULL
)
1546 ret
= iommu_init_unity_mappings(iommu
);
1552 * If device isolation is enabled, pre-allocate the protection
1553 * domains for each device.
1555 if (amd_iommu_isolate
)
1556 prealloc_protection_domains();
1560 bad_dma_address
= 0;
1561 #ifdef CONFIG_GART_IOMMU
1562 gart_iommu_aperture_disabled
= 1;
1563 gart_iommu_aperture
= 0;
1566 /* Make the driver finally visible to the drivers */
1567 dma_ops
= &amd_iommu_dma_ops
;
1569 bus_register_notifier(&pci_bus_type
, &device_nb
);
1575 list_for_each_entry(iommu
, &amd_iommu_list
, list
) {
1576 if (iommu
->default_dom
)
1577 dma_ops_domain_free(iommu
->default_dom
);
1583 /*****************************************************************************
1585 * The following functions belong to the exported interface of AMD IOMMU
1587 * This interface allows access to lower level functions of the IOMMU
1588 * like protection domain handling and assignement of devices to domains
1589 * which is not possible with the dma_ops interface.
1591 *****************************************************************************/
1593 #ifdef CONFIG_IOMMU_API
1595 static void cleanup_domain(struct protection_domain
*domain
)
1597 unsigned long flags
;
1600 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1602 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
1603 if (amd_iommu_pd_table
[devid
] == domain
)
1604 __detach_device(domain
, devid
);
1606 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);