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/dma-mapping.h>
26 #include <linux/iommu-helper.h>
27 #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 * Domain for untranslated devices - only allocated
46 * if iommu=pt passed on kernel cmd line.
48 static struct protection_domain
*pt_domain
;
50 #ifdef CONFIG_IOMMU_API
51 static struct iommu_ops amd_iommu_ops
;
55 * general struct to manage commands send to an IOMMU
61 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
62 struct unity_map_entry
*e
);
63 static struct dma_ops_domain
*find_protection_domain(u16 devid
);
64 static u64
* alloc_pte(struct protection_domain
*dom
,
65 unsigned long address
, u64
66 **pte_page
, gfp_t gfp
);
67 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
68 unsigned long start_page
,
71 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
72 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
75 #ifdef CONFIG_AMD_IOMMU_STATS
78 * Initialization code for statistics collection
81 DECLARE_STATS_COUNTER(compl_wait
);
82 DECLARE_STATS_COUNTER(cnt_map_single
);
83 DECLARE_STATS_COUNTER(cnt_unmap_single
);
84 DECLARE_STATS_COUNTER(cnt_map_sg
);
85 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
86 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
87 DECLARE_STATS_COUNTER(cnt_free_coherent
);
88 DECLARE_STATS_COUNTER(cross_page
);
89 DECLARE_STATS_COUNTER(domain_flush_single
);
90 DECLARE_STATS_COUNTER(domain_flush_all
);
91 DECLARE_STATS_COUNTER(alloced_io_mem
);
92 DECLARE_STATS_COUNTER(total_map_requests
);
94 static struct dentry
*stats_dir
;
95 static struct dentry
*de_isolate
;
96 static struct dentry
*de_fflush
;
98 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
100 if (stats_dir
== NULL
)
103 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
107 static void amd_iommu_stats_init(void)
109 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
110 if (stats_dir
== NULL
)
113 de_isolate
= debugfs_create_bool("isolation", 0444, stats_dir
,
114 (u32
*)&amd_iommu_isolate
);
116 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
117 (u32
*)&amd_iommu_unmap_flush
);
119 amd_iommu_stats_add(&compl_wait
);
120 amd_iommu_stats_add(&cnt_map_single
);
121 amd_iommu_stats_add(&cnt_unmap_single
);
122 amd_iommu_stats_add(&cnt_map_sg
);
123 amd_iommu_stats_add(&cnt_unmap_sg
);
124 amd_iommu_stats_add(&cnt_alloc_coherent
);
125 amd_iommu_stats_add(&cnt_free_coherent
);
126 amd_iommu_stats_add(&cross_page
);
127 amd_iommu_stats_add(&domain_flush_single
);
128 amd_iommu_stats_add(&domain_flush_all
);
129 amd_iommu_stats_add(&alloced_io_mem
);
130 amd_iommu_stats_add(&total_map_requests
);
135 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
136 static int iommu_has_npcache(struct amd_iommu
*iommu
)
138 return iommu
->cap
& (1UL << IOMMU_CAP_NPCACHE
);
141 /****************************************************************************
143 * Interrupt handling functions
145 ****************************************************************************/
147 static void iommu_print_event(void *__evt
)
150 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
151 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
152 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
153 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
154 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
156 printk(KERN_ERR
"AMD IOMMU: Event logged [");
159 case EVENT_TYPE_ILL_DEV
:
160 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
161 "address=0x%016llx flags=0x%04x]\n",
162 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
165 case EVENT_TYPE_IO_FAULT
:
166 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
167 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
168 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
169 domid
, address
, flags
);
171 case EVENT_TYPE_DEV_TAB_ERR
:
172 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
173 "address=0x%016llx flags=0x%04x]\n",
174 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
177 case EVENT_TYPE_PAGE_TAB_ERR
:
178 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
179 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
180 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
181 domid
, address
, flags
);
183 case EVENT_TYPE_ILL_CMD
:
184 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
186 case EVENT_TYPE_CMD_HARD_ERR
:
187 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
188 "flags=0x%04x]\n", address
, flags
);
190 case EVENT_TYPE_IOTLB_INV_TO
:
191 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
192 "address=0x%016llx]\n",
193 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
196 case EVENT_TYPE_INV_DEV_REQ
:
197 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
198 "address=0x%016llx flags=0x%04x]\n",
199 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
203 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
207 static void iommu_poll_events(struct amd_iommu
*iommu
)
212 spin_lock_irqsave(&iommu
->lock
, flags
);
214 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
215 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
217 while (head
!= tail
) {
218 iommu_print_event(iommu
->evt_buf
+ head
);
219 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
222 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
224 spin_unlock_irqrestore(&iommu
->lock
, flags
);
227 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
229 struct amd_iommu
*iommu
;
231 for_each_iommu(iommu
)
232 iommu_poll_events(iommu
);
237 /****************************************************************************
239 * IOMMU command queuing functions
241 ****************************************************************************/
244 * Writes the command to the IOMMUs command buffer and informs the
245 * hardware about the new command. Must be called with iommu->lock held.
247 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
252 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
253 target
= iommu
->cmd_buf
+ tail
;
254 memcpy_toio(target
, cmd
, sizeof(*cmd
));
255 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
256 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
259 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
265 * General queuing function for commands. Takes iommu->lock and calls
266 * __iommu_queue_command().
268 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
273 spin_lock_irqsave(&iommu
->lock
, flags
);
274 ret
= __iommu_queue_command(iommu
, cmd
);
276 iommu
->need_sync
= true;
277 spin_unlock_irqrestore(&iommu
->lock
, flags
);
283 * This function waits until an IOMMU has completed a completion
286 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
292 INC_STATS_COUNTER(compl_wait
);
294 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
296 /* wait for the bit to become one */
297 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
298 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
301 /* set bit back to zero */
302 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
303 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
305 if (unlikely(i
== EXIT_LOOP_COUNT
))
306 panic("AMD IOMMU: Completion wait loop failed\n");
310 * This function queues a completion wait command into the command
313 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
315 struct iommu_cmd cmd
;
317 memset(&cmd
, 0, sizeof(cmd
));
318 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
319 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
321 return __iommu_queue_command(iommu
, &cmd
);
325 * This function is called whenever we need to ensure that the IOMMU has
326 * completed execution of all commands we sent. It sends a
327 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
328 * us about that by writing a value to a physical address we pass with
331 static int iommu_completion_wait(struct amd_iommu
*iommu
)
336 spin_lock_irqsave(&iommu
->lock
, flags
);
338 if (!iommu
->need_sync
)
341 ret
= __iommu_completion_wait(iommu
);
343 iommu
->need_sync
= false;
348 __iommu_wait_for_completion(iommu
);
351 spin_unlock_irqrestore(&iommu
->lock
, flags
);
357 * Command send function for invalidating a device table entry
359 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
361 struct iommu_cmd cmd
;
364 BUG_ON(iommu
== NULL
);
366 memset(&cmd
, 0, sizeof(cmd
));
367 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
370 ret
= iommu_queue_command(iommu
, &cmd
);
375 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
376 u16 domid
, int pde
, int s
)
378 memset(cmd
, 0, sizeof(*cmd
));
379 address
&= PAGE_MASK
;
380 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
381 cmd
->data
[1] |= domid
;
382 cmd
->data
[2] = lower_32_bits(address
);
383 cmd
->data
[3] = upper_32_bits(address
);
384 if (s
) /* size bit - we flush more than one 4kb page */
385 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
386 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
387 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
391 * Generic command send function for invalidaing TLB entries
393 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
394 u64 address
, u16 domid
, int pde
, int s
)
396 struct iommu_cmd cmd
;
399 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
401 ret
= iommu_queue_command(iommu
, &cmd
);
407 * TLB invalidation function which is called from the mapping functions.
408 * It invalidates a single PTE if the range to flush is within a single
409 * page. Otherwise it flushes the whole TLB of the IOMMU.
411 static int iommu_flush_pages(struct amd_iommu
*iommu
, u16 domid
,
412 u64 address
, size_t size
)
415 unsigned pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
417 address
&= PAGE_MASK
;
421 * If we have to flush more than one page, flush all
422 * TLB entries for this domain
424 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
428 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, s
);
433 /* Flush the whole IO/TLB for a given protection domain */
434 static void iommu_flush_tlb(struct amd_iommu
*iommu
, u16 domid
)
436 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
438 INC_STATS_COUNTER(domain_flush_single
);
440 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 0, 1);
443 /* Flush the whole IO/TLB for a given protection domain - including PDE */
444 static void iommu_flush_tlb_pde(struct amd_iommu
*iommu
, u16 domid
)
446 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
448 INC_STATS_COUNTER(domain_flush_single
);
450 iommu_queue_inv_iommu_pages(iommu
, address
, domid
, 1, 1);
454 * This function is used to flush the IO/TLB for a given protection domain
455 * on every IOMMU in the system
457 static void iommu_flush_domain(u16 domid
)
460 struct amd_iommu
*iommu
;
461 struct iommu_cmd cmd
;
463 INC_STATS_COUNTER(domain_flush_all
);
465 __iommu_build_inv_iommu_pages(&cmd
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
468 for_each_iommu(iommu
) {
469 spin_lock_irqsave(&iommu
->lock
, flags
);
470 __iommu_queue_command(iommu
, &cmd
);
471 __iommu_completion_wait(iommu
);
472 __iommu_wait_for_completion(iommu
);
473 spin_unlock_irqrestore(&iommu
->lock
, flags
);
477 void amd_iommu_flush_all_domains(void)
481 for (i
= 1; i
< MAX_DOMAIN_ID
; ++i
) {
482 if (!test_bit(i
, amd_iommu_pd_alloc_bitmap
))
484 iommu_flush_domain(i
);
488 void amd_iommu_flush_all_devices(void)
490 struct amd_iommu
*iommu
;
493 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
494 if (amd_iommu_pd_table
[i
] == NULL
)
497 iommu
= amd_iommu_rlookup_table
[i
];
501 iommu_queue_inv_dev_entry(iommu
, i
);
502 iommu_completion_wait(iommu
);
506 /****************************************************************************
508 * The functions below are used the create the page table mappings for
509 * unity mapped regions.
511 ****************************************************************************/
514 * Generic mapping functions. It maps a physical address into a DMA
515 * address space. It allocates the page table pages if necessary.
516 * In the future it can be extended to a generic mapping function
517 * supporting all features of AMD IOMMU page tables like level skipping
518 * and full 64 bit address spaces.
520 static int iommu_map_page(struct protection_domain
*dom
,
521 unsigned long bus_addr
,
522 unsigned long phys_addr
,
527 bus_addr
= PAGE_ALIGN(bus_addr
);
528 phys_addr
= PAGE_ALIGN(phys_addr
);
530 /* only support 512GB address spaces for now */
531 if (bus_addr
> IOMMU_MAP_SIZE_L3
|| !(prot
& IOMMU_PROT_MASK
))
534 pte
= alloc_pte(dom
, bus_addr
, NULL
, GFP_KERNEL
);
536 if (IOMMU_PTE_PRESENT(*pte
))
539 __pte
= phys_addr
| IOMMU_PTE_P
;
540 if (prot
& IOMMU_PROT_IR
)
541 __pte
|= IOMMU_PTE_IR
;
542 if (prot
& IOMMU_PROT_IW
)
543 __pte
|= IOMMU_PTE_IW
;
550 static void iommu_unmap_page(struct protection_domain
*dom
,
551 unsigned long bus_addr
)
555 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(bus_addr
)];
557 if (!IOMMU_PTE_PRESENT(*pte
))
560 pte
= IOMMU_PTE_PAGE(*pte
);
561 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
563 if (!IOMMU_PTE_PRESENT(*pte
))
566 pte
= IOMMU_PTE_PAGE(*pte
);
567 pte
= &pte
[IOMMU_PTE_L1_INDEX(bus_addr
)];
573 * This function checks if a specific unity mapping entry is needed for
574 * this specific IOMMU.
576 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
577 struct unity_map_entry
*entry
)
581 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
582 bdf
= amd_iommu_alias_table
[i
];
583 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
591 * Init the unity mappings for a specific IOMMU in the system
593 * Basically iterates over all unity mapping entries and applies them to
594 * the default domain DMA of that IOMMU if necessary.
596 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
598 struct unity_map_entry
*entry
;
601 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
602 if (!iommu_for_unity_map(iommu
, entry
))
604 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
613 * This function actually applies the mapping to the page table of the
616 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
617 struct unity_map_entry
*e
)
622 for (addr
= e
->address_start
; addr
< e
->address_end
;
624 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
);
628 * if unity mapping is in aperture range mark the page
629 * as allocated in the aperture
631 if (addr
< dma_dom
->aperture_size
)
632 __set_bit(addr
>> PAGE_SHIFT
,
633 dma_dom
->aperture
[0]->bitmap
);
640 * Inits the unity mappings required for a specific device
642 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
645 struct unity_map_entry
*e
;
648 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
649 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
651 ret
= dma_ops_unity_map(dma_dom
, e
);
659 /****************************************************************************
661 * The next functions belong to the address allocator for the dma_ops
662 * interface functions. They work like the allocators in the other IOMMU
663 * drivers. Its basically a bitmap which marks the allocated pages in
664 * the aperture. Maybe it could be enhanced in the future to a more
665 * efficient allocator.
667 ****************************************************************************/
670 * The address allocator core functions.
672 * called with domain->lock held
676 * This function checks if there is a PTE for a given dma address. If
677 * there is one, it returns the pointer to it.
679 static u64
* fetch_pte(struct protection_domain
*domain
,
680 unsigned long address
)
684 pte
= &domain
->pt_root
[IOMMU_PTE_L2_INDEX(address
)];
686 if (!IOMMU_PTE_PRESENT(*pte
))
689 pte
= IOMMU_PTE_PAGE(*pte
);
690 pte
= &pte
[IOMMU_PTE_L1_INDEX(address
)];
692 if (!IOMMU_PTE_PRESENT(*pte
))
695 pte
= IOMMU_PTE_PAGE(*pte
);
696 pte
= &pte
[IOMMU_PTE_L0_INDEX(address
)];
702 * This function is used to add a new aperture range to an existing
703 * aperture in case of dma_ops domain allocation or address allocation
706 static int alloc_new_range(struct amd_iommu
*iommu
,
707 struct dma_ops_domain
*dma_dom
,
708 bool populate
, gfp_t gfp
)
710 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
713 #ifdef CONFIG_IOMMU_STRESS
717 if (index
>= APERTURE_MAX_RANGES
)
720 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
721 if (!dma_dom
->aperture
[index
])
724 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
725 if (!dma_dom
->aperture
[index
]->bitmap
)
728 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
731 unsigned long address
= dma_dom
->aperture_size
;
732 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
735 for (i
= 0; i
< num_ptes
; ++i
) {
736 pte
= alloc_pte(&dma_dom
->domain
, address
,
741 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
743 address
+= APERTURE_RANGE_SIZE
/ 64;
747 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
749 /* Intialize the exclusion range if necessary */
750 if (iommu
->exclusion_start
&&
751 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
&&
752 iommu
->exclusion_start
< dma_dom
->aperture_size
) {
753 unsigned long startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
754 int pages
= iommu_num_pages(iommu
->exclusion_start
,
755 iommu
->exclusion_length
,
757 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
761 * Check for areas already mapped as present in the new aperture
762 * range and mark those pages as reserved in the allocator. Such
763 * mappings may already exist as a result of requested unity
764 * mappings for devices.
766 for (i
= dma_dom
->aperture
[index
]->offset
;
767 i
< dma_dom
->aperture_size
;
769 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
770 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
773 dma_ops_reserve_addresses(dma_dom
, i
<< PAGE_SHIFT
, 1);
779 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
781 kfree(dma_dom
->aperture
[index
]);
782 dma_dom
->aperture
[index
] = NULL
;
787 static unsigned long dma_ops_area_alloc(struct device
*dev
,
788 struct dma_ops_domain
*dom
,
790 unsigned long align_mask
,
794 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
795 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
796 int i
= start
>> APERTURE_RANGE_SHIFT
;
797 unsigned long boundary_size
;
798 unsigned long address
= -1;
801 next_bit
>>= PAGE_SHIFT
;
803 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
804 PAGE_SIZE
) >> PAGE_SHIFT
;
806 for (;i
< max_index
; ++i
) {
807 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
809 if (dom
->aperture
[i
]->offset
>= dma_mask
)
812 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
813 dma_mask
>> PAGE_SHIFT
);
815 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
816 limit
, next_bit
, pages
, 0,
817 boundary_size
, align_mask
);
819 address
= dom
->aperture
[i
]->offset
+
820 (address
<< PAGE_SHIFT
);
821 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
831 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
832 struct dma_ops_domain
*dom
,
834 unsigned long align_mask
,
837 unsigned long address
;
839 #ifdef CONFIG_IOMMU_STRESS
840 dom
->next_address
= 0;
841 dom
->need_flush
= true;
844 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
845 dma_mask
, dom
->next_address
);
848 dom
->next_address
= 0;
849 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
851 dom
->need_flush
= true;
854 if (unlikely(address
== -1))
855 address
= bad_dma_address
;
857 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
863 * The address free function.
865 * called with domain->lock held
867 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
868 unsigned long address
,
871 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
872 struct aperture_range
*range
= dom
->aperture
[i
];
874 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
876 #ifdef CONFIG_IOMMU_STRESS
881 if (address
>= dom
->next_address
)
882 dom
->need_flush
= true;
884 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
886 iommu_area_free(range
->bitmap
, address
, pages
);
890 /****************************************************************************
892 * The next functions belong to the domain allocation. A domain is
893 * allocated for every IOMMU as the default domain. If device isolation
894 * is enabled, every device get its own domain. The most important thing
895 * about domains is the page table mapping the DMA address space they
898 ****************************************************************************/
900 static u16
domain_id_alloc(void)
905 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
906 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
908 if (id
> 0 && id
< MAX_DOMAIN_ID
)
909 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
912 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
917 static void domain_id_free(int id
)
921 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
922 if (id
> 0 && id
< MAX_DOMAIN_ID
)
923 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
924 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
928 * Used to reserve address ranges in the aperture (e.g. for exclusion
931 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
932 unsigned long start_page
,
935 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
937 if (start_page
+ pages
> last_page
)
938 pages
= last_page
- start_page
;
940 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
941 int index
= i
/ APERTURE_RANGE_PAGES
;
942 int page
= i
% APERTURE_RANGE_PAGES
;
943 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
947 static void free_pagetable(struct protection_domain
*domain
)
952 p1
= domain
->pt_root
;
957 for (i
= 0; i
< 512; ++i
) {
958 if (!IOMMU_PTE_PRESENT(p1
[i
]))
961 p2
= IOMMU_PTE_PAGE(p1
[i
]);
962 for (j
= 0; j
< 512; ++j
) {
963 if (!IOMMU_PTE_PRESENT(p2
[j
]))
965 p3
= IOMMU_PTE_PAGE(p2
[j
]);
966 free_page((unsigned long)p3
);
969 free_page((unsigned long)p2
);
972 free_page((unsigned long)p1
);
974 domain
->pt_root
= NULL
;
978 * Free a domain, only used if something went wrong in the
979 * allocation path and we need to free an already allocated page table
981 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
988 free_pagetable(&dom
->domain
);
990 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
991 if (!dom
->aperture
[i
])
993 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
994 kfree(dom
->aperture
[i
]);
1001 * Allocates a new protection domain usable for the dma_ops functions.
1002 * It also intializes the page table and the address allocator data
1003 * structures required for the dma_ops interface
1005 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
)
1007 struct dma_ops_domain
*dma_dom
;
1009 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1013 spin_lock_init(&dma_dom
->domain
.lock
);
1015 dma_dom
->domain
.id
= domain_id_alloc();
1016 if (dma_dom
->domain
.id
== 0)
1018 dma_dom
->domain
.mode
= PAGE_MODE_3_LEVEL
;
1019 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1020 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1021 dma_dom
->domain
.priv
= dma_dom
;
1022 if (!dma_dom
->domain
.pt_root
)
1025 dma_dom
->need_flush
= false;
1026 dma_dom
->target_dev
= 0xffff;
1028 if (alloc_new_range(iommu
, dma_dom
, true, GFP_KERNEL
))
1032 * mark the first page as allocated so we never return 0 as
1033 * a valid dma-address. So we can use 0 as error value
1035 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1036 dma_dom
->next_address
= 0;
1042 dma_ops_domain_free(dma_dom
);
1048 * little helper function to check whether a given protection domain is a
1051 static bool dma_ops_domain(struct protection_domain
*domain
)
1053 return domain
->flags
& PD_DMA_OPS_MASK
;
1057 * Find out the protection domain structure for a given PCI device. This
1058 * will give us the pointer to the page table root for example.
1060 static struct protection_domain
*domain_for_device(u16 devid
)
1062 struct protection_domain
*dom
;
1063 unsigned long flags
;
1065 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1066 dom
= amd_iommu_pd_table
[devid
];
1067 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1073 * If a device is not yet associated with a domain, this function does
1074 * assigns it visible for the hardware
1076 static void __attach_device(struct amd_iommu
*iommu
,
1077 struct protection_domain
*domain
,
1080 unsigned long flags
;
1081 u64 pte_root
= virt_to_phys(domain
->pt_root
);
1083 domain
->dev_cnt
+= 1;
1085 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1086 << DEV_ENTRY_MODE_SHIFT
;
1087 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1089 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1090 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
1091 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
1092 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
1094 amd_iommu_pd_table
[devid
] = domain
;
1095 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1098 static void attach_device(struct amd_iommu
*iommu
,
1099 struct protection_domain
*domain
,
1102 __attach_device(iommu
, domain
, devid
);
1105 * We might boot into a crash-kernel here. The crashed kernel
1106 * left the caches in the IOMMU dirty. So we have to flush
1107 * here to evict all dirty stuff.
1109 iommu_queue_inv_dev_entry(iommu
, devid
);
1110 iommu_flush_tlb_pde(iommu
, domain
->id
);
1114 * Removes a device from a protection domain (unlocked)
1116 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
1120 spin_lock(&domain
->lock
);
1122 /* remove domain from the lookup table */
1123 amd_iommu_pd_table
[devid
] = NULL
;
1125 /* remove entry from the device table seen by the hardware */
1126 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1127 amd_iommu_dev_table
[devid
].data
[1] = 0;
1128 amd_iommu_dev_table
[devid
].data
[2] = 0;
1130 /* decrease reference counter */
1131 domain
->dev_cnt
-= 1;
1134 spin_unlock(&domain
->lock
);
1138 * Removes a device from a protection domain (with devtable_lock held)
1140 static void detach_device(struct protection_domain
*domain
, u16 devid
)
1142 unsigned long flags
;
1144 /* lock device table */
1145 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1146 __detach_device(domain
, devid
);
1147 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1150 static int device_change_notifier(struct notifier_block
*nb
,
1151 unsigned long action
, void *data
)
1153 struct device
*dev
= data
;
1154 struct pci_dev
*pdev
= to_pci_dev(dev
);
1155 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1156 struct protection_domain
*domain
;
1157 struct dma_ops_domain
*dma_domain
;
1158 struct amd_iommu
*iommu
;
1159 unsigned long flags
;
1161 if (devid
> amd_iommu_last_bdf
)
1164 devid
= amd_iommu_alias_table
[devid
];
1166 iommu
= amd_iommu_rlookup_table
[devid
];
1170 domain
= domain_for_device(devid
);
1172 if (domain
&& !dma_ops_domain(domain
))
1173 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
1174 "to a non-dma-ops domain\n", dev_name(dev
));
1177 case BUS_NOTIFY_UNBOUND_DRIVER
:
1180 detach_device(domain
, devid
);
1182 case BUS_NOTIFY_ADD_DEVICE
:
1183 /* allocate a protection domain if a device is added */
1184 dma_domain
= find_protection_domain(devid
);
1187 dma_domain
= dma_ops_domain_alloc(iommu
);
1190 dma_domain
->target_dev
= devid
;
1192 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1193 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1194 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1201 iommu_queue_inv_dev_entry(iommu
, devid
);
1202 iommu_completion_wait(iommu
);
1208 static struct notifier_block device_nb
= {
1209 .notifier_call
= device_change_notifier
,
1212 /*****************************************************************************
1214 * The next functions belong to the dma_ops mapping/unmapping code.
1216 *****************************************************************************/
1219 * This function checks if the driver got a valid device from the caller to
1220 * avoid dereferencing invalid pointers.
1222 static bool check_device(struct device
*dev
)
1224 if (!dev
|| !dev
->dma_mask
)
1231 * In this function the list of preallocated protection domains is traversed to
1232 * find the domain for a specific device
1234 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
1236 struct dma_ops_domain
*entry
, *ret
= NULL
;
1237 unsigned long flags
;
1239 if (list_empty(&iommu_pd_list
))
1242 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1244 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
1245 if (entry
->target_dev
== devid
) {
1251 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1257 * In the dma_ops path we only have the struct device. This function
1258 * finds the corresponding IOMMU, the protection domain and the
1259 * requestor id for a given device.
1260 * If the device is not yet associated with a domain this is also done
1263 static int get_device_resources(struct device
*dev
,
1264 struct amd_iommu
**iommu
,
1265 struct protection_domain
**domain
,
1268 struct dma_ops_domain
*dma_dom
;
1269 struct pci_dev
*pcidev
;
1276 if (dev
->bus
!= &pci_bus_type
)
1279 pcidev
= to_pci_dev(dev
);
1280 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1282 /* device not translated by any IOMMU in the system? */
1283 if (_bdf
> amd_iommu_last_bdf
)
1286 *bdf
= amd_iommu_alias_table
[_bdf
];
1288 *iommu
= amd_iommu_rlookup_table
[*bdf
];
1291 *domain
= domain_for_device(*bdf
);
1292 if (*domain
== NULL
) {
1293 dma_dom
= find_protection_domain(*bdf
);
1295 dma_dom
= (*iommu
)->default_dom
;
1296 *domain
= &dma_dom
->domain
;
1297 attach_device(*iommu
, *domain
, *bdf
);
1298 DUMP_printk("Using protection domain %d for device %s\n",
1299 (*domain
)->id
, dev_name(dev
));
1302 if (domain_for_device(_bdf
) == NULL
)
1303 attach_device(*iommu
, *domain
, _bdf
);
1309 * If the pte_page is not yet allocated this function is called
1311 static u64
* alloc_pte(struct protection_domain
*dom
,
1312 unsigned long address
, u64
**pte_page
, gfp_t gfp
)
1316 pte
= &dom
->pt_root
[IOMMU_PTE_L2_INDEX(address
)];
1318 if (!IOMMU_PTE_PRESENT(*pte
)) {
1319 page
= (u64
*)get_zeroed_page(gfp
);
1322 *pte
= IOMMU_L2_PDE(virt_to_phys(page
));
1325 pte
= IOMMU_PTE_PAGE(*pte
);
1326 pte
= &pte
[IOMMU_PTE_L1_INDEX(address
)];
1328 if (!IOMMU_PTE_PRESENT(*pte
)) {
1329 page
= (u64
*)get_zeroed_page(gfp
);
1332 *pte
= IOMMU_L1_PDE(virt_to_phys(page
));
1335 pte
= IOMMU_PTE_PAGE(*pte
);
1340 pte
= &pte
[IOMMU_PTE_L0_INDEX(address
)];
1346 * This function fetches the PTE for a given address in the aperture
1348 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
1349 unsigned long address
)
1351 struct aperture_range
*aperture
;
1352 u64
*pte
, *pte_page
;
1354 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1358 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1360 pte
= alloc_pte(&dom
->domain
, address
, &pte_page
, GFP_ATOMIC
);
1361 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
1363 pte
+= IOMMU_PTE_L0_INDEX(address
);
1369 * This is the generic map function. It maps one 4kb page at paddr to
1370 * the given address in the DMA address space for the domain.
1372 static dma_addr_t
dma_ops_domain_map(struct amd_iommu
*iommu
,
1373 struct dma_ops_domain
*dom
,
1374 unsigned long address
,
1380 WARN_ON(address
> dom
->aperture_size
);
1384 pte
= dma_ops_get_pte(dom
, address
);
1386 return bad_dma_address
;
1388 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1390 if (direction
== DMA_TO_DEVICE
)
1391 __pte
|= IOMMU_PTE_IR
;
1392 else if (direction
== DMA_FROM_DEVICE
)
1393 __pte
|= IOMMU_PTE_IW
;
1394 else if (direction
== DMA_BIDIRECTIONAL
)
1395 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1401 return (dma_addr_t
)address
;
1405 * The generic unmapping function for on page in the DMA address space.
1407 static void dma_ops_domain_unmap(struct amd_iommu
*iommu
,
1408 struct dma_ops_domain
*dom
,
1409 unsigned long address
)
1411 struct aperture_range
*aperture
;
1414 if (address
>= dom
->aperture_size
)
1417 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1421 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1425 pte
+= IOMMU_PTE_L0_INDEX(address
);
1433 * This function contains common code for mapping of a physically
1434 * contiguous memory region into DMA address space. It is used by all
1435 * mapping functions provided with this IOMMU driver.
1436 * Must be called with the domain lock held.
1438 static dma_addr_t
__map_single(struct device
*dev
,
1439 struct amd_iommu
*iommu
,
1440 struct dma_ops_domain
*dma_dom
,
1447 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1448 dma_addr_t address
, start
, ret
;
1450 unsigned long align_mask
= 0;
1453 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1456 INC_STATS_COUNTER(total_map_requests
);
1459 INC_STATS_COUNTER(cross_page
);
1462 align_mask
= (1UL << get_order(size
)) - 1;
1465 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1467 if (unlikely(address
== bad_dma_address
)) {
1469 * setting next_address here will let the address
1470 * allocator only scan the new allocated range in the
1471 * first run. This is a small optimization.
1473 dma_dom
->next_address
= dma_dom
->aperture_size
;
1475 if (alloc_new_range(iommu
, dma_dom
, false, GFP_ATOMIC
))
1479 * aperture was sucessfully enlarged by 128 MB, try
1486 for (i
= 0; i
< pages
; ++i
) {
1487 ret
= dma_ops_domain_map(iommu
, dma_dom
, start
, paddr
, dir
);
1488 if (ret
== bad_dma_address
)
1496 ADD_STATS_COUNTER(alloced_io_mem
, size
);
1498 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1499 iommu_flush_tlb(iommu
, dma_dom
->domain
.id
);
1500 dma_dom
->need_flush
= false;
1501 } else if (unlikely(iommu_has_npcache(iommu
)))
1502 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, address
, size
);
1509 for (--i
; i
>= 0; --i
) {
1511 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1514 dma_ops_free_addresses(dma_dom
, address
, pages
);
1516 return bad_dma_address
;
1520 * Does the reverse of the __map_single function. Must be called with
1521 * the domain lock held too
1523 static void __unmap_single(struct amd_iommu
*iommu
,
1524 struct dma_ops_domain
*dma_dom
,
1525 dma_addr_t dma_addr
,
1529 dma_addr_t i
, start
;
1532 if ((dma_addr
== bad_dma_address
) ||
1533 (dma_addr
+ size
> dma_dom
->aperture_size
))
1536 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1537 dma_addr
&= PAGE_MASK
;
1540 for (i
= 0; i
< pages
; ++i
) {
1541 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1545 SUB_STATS_COUNTER(alloced_io_mem
, size
);
1547 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1549 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1550 iommu_flush_pages(iommu
, dma_dom
->domain
.id
, dma_addr
, size
);
1551 dma_dom
->need_flush
= false;
1556 * The exported map_single function for dma_ops.
1558 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
1559 unsigned long offset
, size_t size
,
1560 enum dma_data_direction dir
,
1561 struct dma_attrs
*attrs
)
1563 unsigned long flags
;
1564 struct amd_iommu
*iommu
;
1565 struct protection_domain
*domain
;
1569 phys_addr_t paddr
= page_to_phys(page
) + offset
;
1571 INC_STATS_COUNTER(cnt_map_single
);
1573 if (!check_device(dev
))
1574 return bad_dma_address
;
1576 dma_mask
= *dev
->dma_mask
;
1578 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1580 if (iommu
== NULL
|| domain
== NULL
)
1581 /* device not handled by any AMD IOMMU */
1582 return (dma_addr_t
)paddr
;
1584 if (!dma_ops_domain(domain
))
1585 return bad_dma_address
;
1587 spin_lock_irqsave(&domain
->lock
, flags
);
1588 addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
, size
, dir
, false,
1590 if (addr
== bad_dma_address
)
1593 iommu_completion_wait(iommu
);
1596 spin_unlock_irqrestore(&domain
->lock
, flags
);
1602 * The exported unmap_single function for dma_ops.
1604 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
1605 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
1607 unsigned long flags
;
1608 struct amd_iommu
*iommu
;
1609 struct protection_domain
*domain
;
1612 INC_STATS_COUNTER(cnt_unmap_single
);
1614 if (!check_device(dev
) ||
1615 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1616 /* device not handled by any AMD IOMMU */
1619 if (!dma_ops_domain(domain
))
1622 spin_lock_irqsave(&domain
->lock
, flags
);
1624 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, dir
);
1626 iommu_completion_wait(iommu
);
1628 spin_unlock_irqrestore(&domain
->lock
, flags
);
1632 * This is a special map_sg function which is used if we should map a
1633 * device which is not handled by an AMD IOMMU in the system.
1635 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1636 int nelems
, int dir
)
1638 struct scatterlist
*s
;
1641 for_each_sg(sglist
, s
, nelems
, i
) {
1642 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1643 s
->dma_length
= s
->length
;
1650 * The exported map_sg function for dma_ops (handles scatter-gather
1653 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1654 int nelems
, enum dma_data_direction dir
,
1655 struct dma_attrs
*attrs
)
1657 unsigned long flags
;
1658 struct amd_iommu
*iommu
;
1659 struct protection_domain
*domain
;
1662 struct scatterlist
*s
;
1664 int mapped_elems
= 0;
1667 INC_STATS_COUNTER(cnt_map_sg
);
1669 if (!check_device(dev
))
1672 dma_mask
= *dev
->dma_mask
;
1674 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1676 if (!iommu
|| !domain
)
1677 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1679 if (!dma_ops_domain(domain
))
1682 spin_lock_irqsave(&domain
->lock
, flags
);
1684 for_each_sg(sglist
, s
, nelems
, i
) {
1687 s
->dma_address
= __map_single(dev
, iommu
, domain
->priv
,
1688 paddr
, s
->length
, dir
, false,
1691 if (s
->dma_address
) {
1692 s
->dma_length
= s
->length
;
1698 iommu_completion_wait(iommu
);
1701 spin_unlock_irqrestore(&domain
->lock
, flags
);
1703 return mapped_elems
;
1705 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1707 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1708 s
->dma_length
, dir
);
1709 s
->dma_address
= s
->dma_length
= 0;
1718 * The exported map_sg function for dma_ops (handles scatter-gather
1721 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1722 int nelems
, enum dma_data_direction dir
,
1723 struct dma_attrs
*attrs
)
1725 unsigned long flags
;
1726 struct amd_iommu
*iommu
;
1727 struct protection_domain
*domain
;
1728 struct scatterlist
*s
;
1732 INC_STATS_COUNTER(cnt_unmap_sg
);
1734 if (!check_device(dev
) ||
1735 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1738 if (!dma_ops_domain(domain
))
1741 spin_lock_irqsave(&domain
->lock
, flags
);
1743 for_each_sg(sglist
, s
, nelems
, i
) {
1744 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1745 s
->dma_length
, dir
);
1746 s
->dma_address
= s
->dma_length
= 0;
1749 iommu_completion_wait(iommu
);
1751 spin_unlock_irqrestore(&domain
->lock
, flags
);
1755 * The exported alloc_coherent function for dma_ops.
1757 static void *alloc_coherent(struct device
*dev
, size_t size
,
1758 dma_addr_t
*dma_addr
, gfp_t flag
)
1760 unsigned long flags
;
1762 struct amd_iommu
*iommu
;
1763 struct protection_domain
*domain
;
1766 u64 dma_mask
= dev
->coherent_dma_mask
;
1768 INC_STATS_COUNTER(cnt_alloc_coherent
);
1770 if (!check_device(dev
))
1773 if (!get_device_resources(dev
, &iommu
, &domain
, &devid
))
1774 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1777 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1781 paddr
= virt_to_phys(virt_addr
);
1783 if (!iommu
|| !domain
) {
1784 *dma_addr
= (dma_addr_t
)paddr
;
1788 if (!dma_ops_domain(domain
))
1792 dma_mask
= *dev
->dma_mask
;
1794 spin_lock_irqsave(&domain
->lock
, flags
);
1796 *dma_addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
,
1797 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
1799 if (*dma_addr
== bad_dma_address
) {
1800 spin_unlock_irqrestore(&domain
->lock
, flags
);
1804 iommu_completion_wait(iommu
);
1806 spin_unlock_irqrestore(&domain
->lock
, flags
);
1812 free_pages((unsigned long)virt_addr
, get_order(size
));
1818 * The exported free_coherent function for dma_ops.
1820 static void free_coherent(struct device
*dev
, size_t size
,
1821 void *virt_addr
, dma_addr_t dma_addr
)
1823 unsigned long flags
;
1824 struct amd_iommu
*iommu
;
1825 struct protection_domain
*domain
;
1828 INC_STATS_COUNTER(cnt_free_coherent
);
1830 if (!check_device(dev
))
1833 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1835 if (!iommu
|| !domain
)
1838 if (!dma_ops_domain(domain
))
1841 spin_lock_irqsave(&domain
->lock
, flags
);
1843 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
1845 iommu_completion_wait(iommu
);
1847 spin_unlock_irqrestore(&domain
->lock
, flags
);
1850 free_pages((unsigned long)virt_addr
, get_order(size
));
1854 * This function is called by the DMA layer to find out if we can handle a
1855 * particular device. It is part of the dma_ops.
1857 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
1860 struct pci_dev
*pcidev
;
1862 /* No device or no PCI device */
1863 if (!dev
|| dev
->bus
!= &pci_bus_type
)
1866 pcidev
= to_pci_dev(dev
);
1868 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1870 /* Out of our scope? */
1871 if (bdf
> amd_iommu_last_bdf
)
1878 * The function for pre-allocating protection domains.
1880 * If the driver core informs the DMA layer if a driver grabs a device
1881 * we don't need to preallocate the protection domains anymore.
1882 * For now we have to.
1884 static void prealloc_protection_domains(void)
1886 struct pci_dev
*dev
= NULL
;
1887 struct dma_ops_domain
*dma_dom
;
1888 struct amd_iommu
*iommu
;
1891 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
1892 devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
1893 if (devid
> amd_iommu_last_bdf
)
1895 devid
= amd_iommu_alias_table
[devid
];
1896 if (domain_for_device(devid
))
1898 iommu
= amd_iommu_rlookup_table
[devid
];
1901 dma_dom
= dma_ops_domain_alloc(iommu
);
1904 init_unity_mappings_for_device(dma_dom
, devid
);
1905 dma_dom
->target_dev
= devid
;
1907 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
1911 static struct dma_map_ops amd_iommu_dma_ops
= {
1912 .alloc_coherent
= alloc_coherent
,
1913 .free_coherent
= free_coherent
,
1914 .map_page
= map_page
,
1915 .unmap_page
= unmap_page
,
1917 .unmap_sg
= unmap_sg
,
1918 .dma_supported
= amd_iommu_dma_supported
,
1922 * The function which clues the AMD IOMMU driver into dma_ops.
1924 int __init
amd_iommu_init_dma_ops(void)
1926 struct amd_iommu
*iommu
;
1930 * first allocate a default protection domain for every IOMMU we
1931 * found in the system. Devices not assigned to any other
1932 * protection domain will be assigned to the default one.
1934 for_each_iommu(iommu
) {
1935 iommu
->default_dom
= dma_ops_domain_alloc(iommu
);
1936 if (iommu
->default_dom
== NULL
)
1938 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
1939 ret
= iommu_init_unity_mappings(iommu
);
1945 * If device isolation is enabled, pre-allocate the protection
1946 * domains for each device.
1948 if (amd_iommu_isolate
)
1949 prealloc_protection_domains();
1953 bad_dma_address
= 0;
1954 #ifdef CONFIG_GART_IOMMU
1955 gart_iommu_aperture_disabled
= 1;
1956 gart_iommu_aperture
= 0;
1959 /* Make the driver finally visible to the drivers */
1960 dma_ops
= &amd_iommu_dma_ops
;
1962 register_iommu(&amd_iommu_ops
);
1964 bus_register_notifier(&pci_bus_type
, &device_nb
);
1966 amd_iommu_stats_init();
1972 for_each_iommu(iommu
) {
1973 if (iommu
->default_dom
)
1974 dma_ops_domain_free(iommu
->default_dom
);
1980 /*****************************************************************************
1982 * The following functions belong to the exported interface of AMD IOMMU
1984 * This interface allows access to lower level functions of the IOMMU
1985 * like protection domain handling and assignement of devices to domains
1986 * which is not possible with the dma_ops interface.
1988 *****************************************************************************/
1990 static void cleanup_domain(struct protection_domain
*domain
)
1992 unsigned long flags
;
1995 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1997 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
1998 if (amd_iommu_pd_table
[devid
] == domain
)
1999 __detach_device(domain
, devid
);
2001 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2004 static void protection_domain_free(struct protection_domain
*domain
)
2010 domain_id_free(domain
->id
);
2015 static struct protection_domain
*protection_domain_alloc(void)
2017 struct protection_domain
*domain
;
2019 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2023 spin_lock_init(&domain
->lock
);
2024 domain
->id
= domain_id_alloc();
2036 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2038 struct protection_domain
*domain
;
2040 domain
= protection_domain_alloc();
2044 domain
->mode
= PAGE_MODE_3_LEVEL
;
2045 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2046 if (!domain
->pt_root
)
2054 protection_domain_free(domain
);
2059 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2061 struct protection_domain
*domain
= dom
->priv
;
2066 if (domain
->dev_cnt
> 0)
2067 cleanup_domain(domain
);
2069 BUG_ON(domain
->dev_cnt
!= 0);
2071 free_pagetable(domain
);
2073 domain_id_free(domain
->id
);
2080 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2083 struct protection_domain
*domain
= dom
->priv
;
2084 struct amd_iommu
*iommu
;
2085 struct pci_dev
*pdev
;
2088 if (dev
->bus
!= &pci_bus_type
)
2091 pdev
= to_pci_dev(dev
);
2093 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2096 detach_device(domain
, devid
);
2098 iommu
= amd_iommu_rlookup_table
[devid
];
2102 iommu_queue_inv_dev_entry(iommu
, devid
);
2103 iommu_completion_wait(iommu
);
2106 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2109 struct protection_domain
*domain
= dom
->priv
;
2110 struct protection_domain
*old_domain
;
2111 struct amd_iommu
*iommu
;
2112 struct pci_dev
*pdev
;
2115 if (dev
->bus
!= &pci_bus_type
)
2118 pdev
= to_pci_dev(dev
);
2120 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2122 if (devid
>= amd_iommu_last_bdf
||
2123 devid
!= amd_iommu_alias_table
[devid
])
2126 iommu
= amd_iommu_rlookup_table
[devid
];
2130 old_domain
= domain_for_device(devid
);
2132 detach_device(old_domain
, devid
);
2134 attach_device(iommu
, domain
, devid
);
2136 iommu_completion_wait(iommu
);
2141 static int amd_iommu_map_range(struct iommu_domain
*dom
,
2142 unsigned long iova
, phys_addr_t paddr
,
2143 size_t size
, int iommu_prot
)
2145 struct protection_domain
*domain
= dom
->priv
;
2146 unsigned long i
, npages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2150 if (iommu_prot
& IOMMU_READ
)
2151 prot
|= IOMMU_PROT_IR
;
2152 if (iommu_prot
& IOMMU_WRITE
)
2153 prot
|= IOMMU_PROT_IW
;
2158 for (i
= 0; i
< npages
; ++i
) {
2159 ret
= iommu_map_page(domain
, iova
, paddr
, prot
);
2170 static void amd_iommu_unmap_range(struct iommu_domain
*dom
,
2171 unsigned long iova
, size_t size
)
2174 struct protection_domain
*domain
= dom
->priv
;
2175 unsigned long i
, npages
= iommu_num_pages(iova
, size
, PAGE_SIZE
);
2179 for (i
= 0; i
< npages
; ++i
) {
2180 iommu_unmap_page(domain
, iova
);
2184 iommu_flush_domain(domain
->id
);
2187 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2190 struct protection_domain
*domain
= dom
->priv
;
2191 unsigned long offset
= iova
& ~PAGE_MASK
;
2195 pte
= &domain
->pt_root
[IOMMU_PTE_L2_INDEX(iova
)];
2197 if (!IOMMU_PTE_PRESENT(*pte
))
2200 pte
= IOMMU_PTE_PAGE(*pte
);
2201 pte
= &pte
[IOMMU_PTE_L1_INDEX(iova
)];
2203 if (!IOMMU_PTE_PRESENT(*pte
))
2206 pte
= IOMMU_PTE_PAGE(*pte
);
2207 pte
= &pte
[IOMMU_PTE_L0_INDEX(iova
)];
2209 if (!IOMMU_PTE_PRESENT(*pte
))
2212 paddr
= *pte
& IOMMU_PAGE_MASK
;
2218 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
2224 static struct iommu_ops amd_iommu_ops
= {
2225 .domain_init
= amd_iommu_domain_init
,
2226 .domain_destroy
= amd_iommu_domain_destroy
,
2227 .attach_dev
= amd_iommu_attach_device
,
2228 .detach_dev
= amd_iommu_detach_device
,
2229 .map
= amd_iommu_map_range
,
2230 .unmap
= amd_iommu_unmap_range
,
2231 .iova_to_phys
= amd_iommu_iova_to_phys
,
2232 .domain_has_cap
= amd_iommu_domain_has_cap
,
2235 /*****************************************************************************
2237 * The next functions do a basic initialization of IOMMU for pass through
2240 * In passthrough mode the IOMMU is initialized and enabled but not used for
2241 * DMA-API translation.
2243 *****************************************************************************/
2245 int __init
amd_iommu_init_passthrough(void)
2247 struct pci_dev
*dev
= NULL
;
2250 /* allocate passthroug domain */
2251 pt_domain
= protection_domain_alloc();
2255 pt_domain
->mode
|= PAGE_MODE_NONE
;
2257 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
2258 struct amd_iommu
*iommu
;
2260 devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
2261 if (devid
> amd_iommu_last_bdf
)
2264 devid2
= amd_iommu_alias_table
[devid
];
2266 iommu
= amd_iommu_rlookup_table
[devid2
];
2270 __attach_device(iommu
, pt_domain
, devid
);
2271 __attach_device(iommu
, pt_domain
, devid2
);
2274 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");