2 * Copyright (C) 2007-2009 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_proto.h>
32 #include <asm/amd_iommu_types.h>
33 #include <asm/amd_iommu.h>
35 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
37 #define EXIT_LOOP_COUNT 10000000
39 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
41 /* A list of preallocated protection domains */
42 static LIST_HEAD(iommu_pd_list
);
43 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
46 * Domain for untranslated devices - only allocated
47 * if iommu=pt passed on kernel cmd line.
49 static struct protection_domain
*pt_domain
;
51 static struct iommu_ops amd_iommu_ops
;
54 * general struct to manage commands send to an IOMMU
60 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
61 struct unity_map_entry
*e
);
62 static struct dma_ops_domain
*find_protection_domain(u16 devid
);
63 static u64
*alloc_pte(struct protection_domain
*domain
,
64 unsigned long address
, int end_lvl
,
65 u64
**pte_page
, gfp_t gfp
);
66 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
67 unsigned long start_page
,
69 static void reset_iommu_command_buffer(struct amd_iommu
*iommu
);
70 static u64
*fetch_pte(struct protection_domain
*domain
,
71 unsigned long address
, int map_size
);
72 static void update_domain(struct protection_domain
*domain
);
74 #ifdef CONFIG_AMD_IOMMU_STATS
77 * Initialization code for statistics collection
80 DECLARE_STATS_COUNTER(compl_wait
);
81 DECLARE_STATS_COUNTER(cnt_map_single
);
82 DECLARE_STATS_COUNTER(cnt_unmap_single
);
83 DECLARE_STATS_COUNTER(cnt_map_sg
);
84 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
85 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
86 DECLARE_STATS_COUNTER(cnt_free_coherent
);
87 DECLARE_STATS_COUNTER(cross_page
);
88 DECLARE_STATS_COUNTER(domain_flush_single
);
89 DECLARE_STATS_COUNTER(domain_flush_all
);
90 DECLARE_STATS_COUNTER(alloced_io_mem
);
91 DECLARE_STATS_COUNTER(total_map_requests
);
93 static struct dentry
*stats_dir
;
94 static struct dentry
*de_isolate
;
95 static struct dentry
*de_fflush
;
97 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
99 if (stats_dir
== NULL
)
102 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
106 static void amd_iommu_stats_init(void)
108 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
109 if (stats_dir
== NULL
)
112 de_isolate
= debugfs_create_bool("isolation", 0444, stats_dir
,
113 (u32
*)&amd_iommu_isolate
);
115 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
116 (u32
*)&amd_iommu_unmap_flush
);
118 amd_iommu_stats_add(&compl_wait
);
119 amd_iommu_stats_add(&cnt_map_single
);
120 amd_iommu_stats_add(&cnt_unmap_single
);
121 amd_iommu_stats_add(&cnt_map_sg
);
122 amd_iommu_stats_add(&cnt_unmap_sg
);
123 amd_iommu_stats_add(&cnt_alloc_coherent
);
124 amd_iommu_stats_add(&cnt_free_coherent
);
125 amd_iommu_stats_add(&cross_page
);
126 amd_iommu_stats_add(&domain_flush_single
);
127 amd_iommu_stats_add(&domain_flush_all
);
128 amd_iommu_stats_add(&alloced_io_mem
);
129 amd_iommu_stats_add(&total_map_requests
);
134 /****************************************************************************
136 * Interrupt handling functions
138 ****************************************************************************/
140 static void dump_dte_entry(u16 devid
)
144 for (i
= 0; i
< 8; ++i
)
145 pr_err("AMD-Vi: DTE[%d]: %08x\n", i
,
146 amd_iommu_dev_table
[devid
].data
[i
]);
149 static void dump_command(unsigned long phys_addr
)
151 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
154 for (i
= 0; i
< 4; ++i
)
155 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
158 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
161 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
162 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
163 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
164 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
165 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
167 printk(KERN_ERR
"AMD-Vi: Event logged [");
170 case EVENT_TYPE_ILL_DEV
:
171 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
172 "address=0x%016llx flags=0x%04x]\n",
173 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
175 dump_dte_entry(devid
);
177 case EVENT_TYPE_IO_FAULT
:
178 printk("IO_PAGE_FAULT 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_DEV_TAB_ERR
:
184 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
185 "address=0x%016llx flags=0x%04x]\n",
186 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
189 case EVENT_TYPE_PAGE_TAB_ERR
:
190 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
191 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
192 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
193 domid
, address
, flags
);
195 case EVENT_TYPE_ILL_CMD
:
196 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
197 reset_iommu_command_buffer(iommu
);
198 dump_command(address
);
200 case EVENT_TYPE_CMD_HARD_ERR
:
201 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
202 "flags=0x%04x]\n", address
, flags
);
204 case EVENT_TYPE_IOTLB_INV_TO
:
205 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
206 "address=0x%016llx]\n",
207 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
210 case EVENT_TYPE_INV_DEV_REQ
:
211 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
212 "address=0x%016llx flags=0x%04x]\n",
213 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
217 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
221 static void iommu_poll_events(struct amd_iommu
*iommu
)
226 spin_lock_irqsave(&iommu
->lock
, flags
);
228 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
229 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
231 while (head
!= tail
) {
232 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
233 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
236 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
238 spin_unlock_irqrestore(&iommu
->lock
, flags
);
241 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
243 struct amd_iommu
*iommu
;
245 for_each_iommu(iommu
)
246 iommu_poll_events(iommu
);
251 /****************************************************************************
253 * IOMMU command queuing functions
255 ****************************************************************************/
258 * Writes the command to the IOMMUs command buffer and informs the
259 * hardware about the new command. Must be called with iommu->lock held.
261 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
266 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
267 target
= iommu
->cmd_buf
+ tail
;
268 memcpy_toio(target
, cmd
, sizeof(*cmd
));
269 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
270 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
273 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
279 * General queuing function for commands. Takes iommu->lock and calls
280 * __iommu_queue_command().
282 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
287 spin_lock_irqsave(&iommu
->lock
, flags
);
288 ret
= __iommu_queue_command(iommu
, cmd
);
290 iommu
->need_sync
= true;
291 spin_unlock_irqrestore(&iommu
->lock
, flags
);
297 * This function waits until an IOMMU has completed a completion
300 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
306 INC_STATS_COUNTER(compl_wait
);
308 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
310 /* wait for the bit to become one */
311 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
312 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
315 /* set bit back to zero */
316 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
317 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
319 if (unlikely(i
== EXIT_LOOP_COUNT
)) {
320 spin_unlock(&iommu
->lock
);
321 reset_iommu_command_buffer(iommu
);
322 spin_lock(&iommu
->lock
);
327 * This function queues a completion wait command into the command
330 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
332 struct iommu_cmd cmd
;
334 memset(&cmd
, 0, sizeof(cmd
));
335 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
336 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
338 return __iommu_queue_command(iommu
, &cmd
);
342 * This function is called whenever we need to ensure that the IOMMU has
343 * completed execution of all commands we sent. It sends a
344 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
345 * us about that by writing a value to a physical address we pass with
348 static int iommu_completion_wait(struct amd_iommu
*iommu
)
353 spin_lock_irqsave(&iommu
->lock
, flags
);
355 if (!iommu
->need_sync
)
358 ret
= __iommu_completion_wait(iommu
);
360 iommu
->need_sync
= false;
365 __iommu_wait_for_completion(iommu
);
368 spin_unlock_irqrestore(&iommu
->lock
, flags
);
373 static void iommu_flush_complete(struct protection_domain
*domain
)
377 for (i
= 0; i
< amd_iommus_present
; ++i
) {
378 if (!domain
->dev_iommu
[i
])
382 * Devices of this domain are behind this IOMMU
383 * We need to wait for completion of all commands.
385 iommu_completion_wait(amd_iommus
[i
]);
390 * Command send function for invalidating a device table entry
392 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
394 struct iommu_cmd cmd
;
397 BUG_ON(iommu
== NULL
);
399 memset(&cmd
, 0, sizeof(cmd
));
400 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
403 ret
= iommu_queue_command(iommu
, &cmd
);
408 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
409 u16 domid
, int pde
, int s
)
411 memset(cmd
, 0, sizeof(*cmd
));
412 address
&= PAGE_MASK
;
413 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
414 cmd
->data
[1] |= domid
;
415 cmd
->data
[2] = lower_32_bits(address
);
416 cmd
->data
[3] = upper_32_bits(address
);
417 if (s
) /* size bit - we flush more than one 4kb page */
418 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
419 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
420 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
424 * Generic command send function for invalidaing TLB entries
426 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
427 u64 address
, u16 domid
, int pde
, int s
)
429 struct iommu_cmd cmd
;
432 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
434 ret
= iommu_queue_command(iommu
, &cmd
);
440 * TLB invalidation function which is called from the mapping functions.
441 * It invalidates a single PTE if the range to flush is within a single
442 * page. Otherwise it flushes the whole TLB of the IOMMU.
444 static void __iommu_flush_pages(struct protection_domain
*domain
,
445 u64 address
, size_t size
, int pde
)
448 unsigned long pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
450 address
&= PAGE_MASK
;
454 * If we have to flush more than one page, flush all
455 * TLB entries for this domain
457 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
462 for (i
= 0; i
< amd_iommus_present
; ++i
) {
463 if (!domain
->dev_iommu
[i
])
467 * Devices of this domain are behind this IOMMU
468 * We need a TLB flush
470 iommu_queue_inv_iommu_pages(amd_iommus
[i
], address
,
477 static void iommu_flush_pages(struct protection_domain
*domain
,
478 u64 address
, size_t size
)
480 __iommu_flush_pages(domain
, address
, size
, 0);
483 /* Flush the whole IO/TLB for a given protection domain */
484 static void iommu_flush_tlb(struct protection_domain
*domain
)
486 __iommu_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
489 /* Flush the whole IO/TLB for a given protection domain - including PDE */
490 static void iommu_flush_tlb_pde(struct protection_domain
*domain
)
492 __iommu_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
496 * This function flushes all domains that have devices on the given IOMMU
498 static void flush_all_domains_on_iommu(struct amd_iommu
*iommu
)
500 u64 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
501 struct protection_domain
*domain
;
504 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
506 list_for_each_entry(domain
, &amd_iommu_pd_list
, list
) {
507 if (domain
->dev_iommu
[iommu
->index
] == 0)
510 spin_lock(&domain
->lock
);
511 iommu_queue_inv_iommu_pages(iommu
, address
, domain
->id
, 1, 1);
512 iommu_flush_complete(domain
);
513 spin_unlock(&domain
->lock
);
516 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
520 * This function uses heavy locking and may disable irqs for some time. But
521 * this is no issue because it is only called during resume.
523 void amd_iommu_flush_all_domains(void)
525 struct protection_domain
*domain
;
528 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
530 list_for_each_entry(domain
, &amd_iommu_pd_list
, list
) {
531 spin_lock(&domain
->lock
);
532 iommu_flush_tlb_pde(domain
);
533 iommu_flush_complete(domain
);
534 spin_unlock(&domain
->lock
);
537 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
540 static void flush_all_devices_for_iommu(struct amd_iommu
*iommu
)
544 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
545 if (iommu
!= amd_iommu_rlookup_table
[i
])
548 iommu_queue_inv_dev_entry(iommu
, i
);
549 iommu_completion_wait(iommu
);
553 static void flush_devices_by_domain(struct protection_domain
*domain
)
555 struct amd_iommu
*iommu
;
558 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
559 if ((domain
== NULL
&& amd_iommu_pd_table
[i
] == NULL
) ||
560 (amd_iommu_pd_table
[i
] != domain
))
563 iommu
= amd_iommu_rlookup_table
[i
];
567 iommu_queue_inv_dev_entry(iommu
, i
);
568 iommu_completion_wait(iommu
);
572 static void reset_iommu_command_buffer(struct amd_iommu
*iommu
)
574 pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
576 if (iommu
->reset_in_progress
)
577 panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
579 iommu
->reset_in_progress
= true;
581 amd_iommu_reset_cmd_buffer(iommu
);
582 flush_all_devices_for_iommu(iommu
);
583 flush_all_domains_on_iommu(iommu
);
585 iommu
->reset_in_progress
= false;
588 void amd_iommu_flush_all_devices(void)
590 flush_devices_by_domain(NULL
);
593 /****************************************************************************
595 * The functions below are used the create the page table mappings for
596 * unity mapped regions.
598 ****************************************************************************/
601 * Generic mapping functions. It maps a physical address into a DMA
602 * address space. It allocates the page table pages if necessary.
603 * In the future it can be extended to a generic mapping function
604 * supporting all features of AMD IOMMU page tables like level skipping
605 * and full 64 bit address spaces.
607 static int iommu_map_page(struct protection_domain
*dom
,
608 unsigned long bus_addr
,
609 unsigned long phys_addr
,
615 bus_addr
= PAGE_ALIGN(bus_addr
);
616 phys_addr
= PAGE_ALIGN(phys_addr
);
618 BUG_ON(!PM_ALIGNED(map_size
, bus_addr
));
619 BUG_ON(!PM_ALIGNED(map_size
, phys_addr
));
621 if (!(prot
& IOMMU_PROT_MASK
))
624 pte
= alloc_pte(dom
, bus_addr
, map_size
, NULL
, GFP_KERNEL
);
626 if (IOMMU_PTE_PRESENT(*pte
))
629 __pte
= phys_addr
| IOMMU_PTE_P
;
630 if (prot
& IOMMU_PROT_IR
)
631 __pte
|= IOMMU_PTE_IR
;
632 if (prot
& IOMMU_PROT_IW
)
633 __pte
|= IOMMU_PTE_IW
;
642 static void iommu_unmap_page(struct protection_domain
*dom
,
643 unsigned long bus_addr
, int map_size
)
645 u64
*pte
= fetch_pte(dom
, bus_addr
, map_size
);
652 * This function checks if a specific unity mapping entry is needed for
653 * this specific IOMMU.
655 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
656 struct unity_map_entry
*entry
)
660 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
661 bdf
= amd_iommu_alias_table
[i
];
662 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
670 * Init the unity mappings for a specific IOMMU in the system
672 * Basically iterates over all unity mapping entries and applies them to
673 * the default domain DMA of that IOMMU if necessary.
675 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
677 struct unity_map_entry
*entry
;
680 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
681 if (!iommu_for_unity_map(iommu
, entry
))
683 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
692 * This function actually applies the mapping to the page table of the
695 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
696 struct unity_map_entry
*e
)
701 for (addr
= e
->address_start
; addr
< e
->address_end
;
703 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
708 * if unity mapping is in aperture range mark the page
709 * as allocated in the aperture
711 if (addr
< dma_dom
->aperture_size
)
712 __set_bit(addr
>> PAGE_SHIFT
,
713 dma_dom
->aperture
[0]->bitmap
);
720 * Inits the unity mappings required for a specific device
722 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
725 struct unity_map_entry
*e
;
728 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
729 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
731 ret
= dma_ops_unity_map(dma_dom
, e
);
739 /****************************************************************************
741 * The next functions belong to the address allocator for the dma_ops
742 * interface functions. They work like the allocators in the other IOMMU
743 * drivers. Its basically a bitmap which marks the allocated pages in
744 * the aperture. Maybe it could be enhanced in the future to a more
745 * efficient allocator.
747 ****************************************************************************/
750 * The address allocator core functions.
752 * called with domain->lock held
756 * This function checks if there is a PTE for a given dma address. If
757 * there is one, it returns the pointer to it.
759 static u64
*fetch_pte(struct protection_domain
*domain
,
760 unsigned long address
, int map_size
)
765 level
= domain
->mode
- 1;
766 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
768 while (level
> map_size
) {
769 if (!IOMMU_PTE_PRESENT(*pte
))
774 pte
= IOMMU_PTE_PAGE(*pte
);
775 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
777 if ((PM_PTE_LEVEL(*pte
) == 0) && level
!= map_size
) {
787 * This function is used to add a new aperture range to an existing
788 * aperture in case of dma_ops domain allocation or address allocation
791 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
792 bool populate
, gfp_t gfp
)
794 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
795 struct amd_iommu
*iommu
;
798 #ifdef CONFIG_IOMMU_STRESS
802 if (index
>= APERTURE_MAX_RANGES
)
805 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
806 if (!dma_dom
->aperture
[index
])
809 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
810 if (!dma_dom
->aperture
[index
]->bitmap
)
813 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
816 unsigned long address
= dma_dom
->aperture_size
;
817 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
820 for (i
= 0; i
< num_ptes
; ++i
) {
821 pte
= alloc_pte(&dma_dom
->domain
, address
, PM_MAP_4k
,
826 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
828 address
+= APERTURE_RANGE_SIZE
/ 64;
832 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
834 /* Intialize the exclusion range if necessary */
835 for_each_iommu(iommu
) {
836 if (iommu
->exclusion_start
&&
837 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
838 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
839 unsigned long startpage
;
840 int pages
= iommu_num_pages(iommu
->exclusion_start
,
841 iommu
->exclusion_length
,
843 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
844 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
849 * Check for areas already mapped as present in the new aperture
850 * range and mark those pages as reserved in the allocator. Such
851 * mappings may already exist as a result of requested unity
852 * mappings for devices.
854 for (i
= dma_dom
->aperture
[index
]->offset
;
855 i
< dma_dom
->aperture_size
;
857 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
, PM_MAP_4k
);
858 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
861 dma_ops_reserve_addresses(dma_dom
, i
<< PAGE_SHIFT
, 1);
864 update_domain(&dma_dom
->domain
);
869 update_domain(&dma_dom
->domain
);
871 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
873 kfree(dma_dom
->aperture
[index
]);
874 dma_dom
->aperture
[index
] = NULL
;
879 static unsigned long dma_ops_area_alloc(struct device
*dev
,
880 struct dma_ops_domain
*dom
,
882 unsigned long align_mask
,
886 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
887 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
888 int i
= start
>> APERTURE_RANGE_SHIFT
;
889 unsigned long boundary_size
;
890 unsigned long address
= -1;
893 next_bit
>>= PAGE_SHIFT
;
895 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
896 PAGE_SIZE
) >> PAGE_SHIFT
;
898 for (;i
< max_index
; ++i
) {
899 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
901 if (dom
->aperture
[i
]->offset
>= dma_mask
)
904 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
905 dma_mask
>> PAGE_SHIFT
);
907 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
908 limit
, next_bit
, pages
, 0,
909 boundary_size
, align_mask
);
911 address
= dom
->aperture
[i
]->offset
+
912 (address
<< PAGE_SHIFT
);
913 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
923 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
924 struct dma_ops_domain
*dom
,
926 unsigned long align_mask
,
929 unsigned long address
;
931 #ifdef CONFIG_IOMMU_STRESS
932 dom
->next_address
= 0;
933 dom
->need_flush
= true;
936 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
937 dma_mask
, dom
->next_address
);
940 dom
->next_address
= 0;
941 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
943 dom
->need_flush
= true;
946 if (unlikely(address
== -1))
947 address
= DMA_ERROR_CODE
;
949 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
955 * The address free function.
957 * called with domain->lock held
959 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
960 unsigned long address
,
963 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
964 struct aperture_range
*range
= dom
->aperture
[i
];
966 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
968 #ifdef CONFIG_IOMMU_STRESS
973 if (address
>= dom
->next_address
)
974 dom
->need_flush
= true;
976 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
978 iommu_area_free(range
->bitmap
, address
, pages
);
982 /****************************************************************************
984 * The next functions belong to the domain allocation. A domain is
985 * allocated for every IOMMU as the default domain. If device isolation
986 * is enabled, every device get its own domain. The most important thing
987 * about domains is the page table mapping the DMA address space they
990 ****************************************************************************/
993 * This function adds a protection domain to the global protection domain list
995 static void add_domain_to_list(struct protection_domain
*domain
)
999 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1000 list_add(&domain
->list
, &amd_iommu_pd_list
);
1001 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1005 * This function removes a protection domain to the global
1006 * protection domain list
1008 static void del_domain_from_list(struct protection_domain
*domain
)
1010 unsigned long flags
;
1012 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1013 list_del(&domain
->list
);
1014 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1017 static u16
domain_id_alloc(void)
1019 unsigned long flags
;
1022 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1023 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1025 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1026 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1029 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1034 static void domain_id_free(int id
)
1036 unsigned long flags
;
1038 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1039 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1040 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1041 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1045 * Used to reserve address ranges in the aperture (e.g. for exclusion
1048 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1049 unsigned long start_page
,
1052 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1054 if (start_page
+ pages
> last_page
)
1055 pages
= last_page
- start_page
;
1057 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1058 int index
= i
/ APERTURE_RANGE_PAGES
;
1059 int page
= i
% APERTURE_RANGE_PAGES
;
1060 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1064 static void free_pagetable(struct protection_domain
*domain
)
1069 p1
= domain
->pt_root
;
1074 for (i
= 0; i
< 512; ++i
) {
1075 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1078 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1079 for (j
= 0; j
< 512; ++j
) {
1080 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1082 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1083 free_page((unsigned long)p3
);
1086 free_page((unsigned long)p2
);
1089 free_page((unsigned long)p1
);
1091 domain
->pt_root
= NULL
;
1095 * Free a domain, only used if something went wrong in the
1096 * allocation path and we need to free an already allocated page table
1098 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1105 del_domain_from_list(&dom
->domain
);
1107 free_pagetable(&dom
->domain
);
1109 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1110 if (!dom
->aperture
[i
])
1112 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1113 kfree(dom
->aperture
[i
]);
1120 * Allocates a new protection domain usable for the dma_ops functions.
1121 * It also intializes the page table and the address allocator data
1122 * structures required for the dma_ops interface
1124 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
)
1126 struct dma_ops_domain
*dma_dom
;
1128 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1132 spin_lock_init(&dma_dom
->domain
.lock
);
1134 dma_dom
->domain
.id
= domain_id_alloc();
1135 if (dma_dom
->domain
.id
== 0)
1137 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1138 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1139 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1140 dma_dom
->domain
.priv
= dma_dom
;
1141 if (!dma_dom
->domain
.pt_root
)
1144 dma_dom
->need_flush
= false;
1145 dma_dom
->target_dev
= 0xffff;
1147 add_domain_to_list(&dma_dom
->domain
);
1149 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1153 * mark the first page as allocated so we never return 0 as
1154 * a valid dma-address. So we can use 0 as error value
1156 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1157 dma_dom
->next_address
= 0;
1163 dma_ops_domain_free(dma_dom
);
1169 * little helper function to check whether a given protection domain is a
1172 static bool dma_ops_domain(struct protection_domain
*domain
)
1174 return domain
->flags
& PD_DMA_OPS_MASK
;
1178 * Find out the protection domain structure for a given PCI device. This
1179 * will give us the pointer to the page table root for example.
1181 static struct protection_domain
*domain_for_device(u16 devid
)
1183 struct protection_domain
*dom
;
1184 unsigned long flags
;
1186 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1187 dom
= amd_iommu_pd_table
[devid
];
1188 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1193 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
)
1195 u64 pte_root
= virt_to_phys(domain
->pt_root
);
1197 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1198 << DEV_ENTRY_MODE_SHIFT
;
1199 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1201 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
1202 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
1203 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
1205 amd_iommu_pd_table
[devid
] = domain
;
1209 * If a device is not yet associated with a domain, this function does
1210 * assigns it visible for the hardware
1212 static void __attach_device(struct amd_iommu
*iommu
,
1213 struct protection_domain
*domain
,
1217 spin_lock(&domain
->lock
);
1219 /* update DTE entry */
1220 set_dte_entry(devid
, domain
);
1222 /* Do reference counting */
1223 domain
->dev_iommu
[iommu
->index
] += 1;
1224 domain
->dev_cnt
+= 1;
1227 spin_unlock(&domain
->lock
);
1231 * If a device is not yet associated with a domain, this function does
1232 * assigns it visible for the hardware
1234 static void attach_device(struct amd_iommu
*iommu
,
1235 struct protection_domain
*domain
,
1238 unsigned long flags
;
1240 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1241 __attach_device(iommu
, domain
, devid
);
1242 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1245 * We might boot into a crash-kernel here. The crashed kernel
1246 * left the caches in the IOMMU dirty. So we have to flush
1247 * here to evict all dirty stuff.
1249 iommu_queue_inv_dev_entry(iommu
, devid
);
1250 iommu_flush_tlb_pde(domain
);
1254 * Removes a device from a protection domain (unlocked)
1256 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
1258 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
1263 spin_lock(&domain
->lock
);
1265 /* remove domain from the lookup table */
1266 amd_iommu_pd_table
[devid
] = NULL
;
1268 /* remove entry from the device table seen by the hardware */
1269 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1270 amd_iommu_dev_table
[devid
].data
[1] = 0;
1271 amd_iommu_dev_table
[devid
].data
[2] = 0;
1273 amd_iommu_apply_erratum_63(devid
);
1275 /* decrease reference counters */
1276 domain
->dev_iommu
[iommu
->index
] -= 1;
1277 domain
->dev_cnt
-= 1;
1280 spin_unlock(&domain
->lock
);
1283 * If we run in passthrough mode the device must be assigned to the
1284 * passthrough domain if it is detached from any other domain
1286 if (iommu_pass_through
) {
1287 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
1288 __attach_device(iommu
, pt_domain
, devid
);
1293 * Removes a device from a protection domain (with devtable_lock held)
1295 static void detach_device(struct protection_domain
*domain
, u16 devid
)
1297 unsigned long flags
;
1299 /* lock device table */
1300 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1301 __detach_device(domain
, devid
);
1302 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1305 static int device_change_notifier(struct notifier_block
*nb
,
1306 unsigned long action
, void *data
)
1308 struct device
*dev
= data
;
1309 struct pci_dev
*pdev
= to_pci_dev(dev
);
1310 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1311 struct protection_domain
*domain
;
1312 struct dma_ops_domain
*dma_domain
;
1313 struct amd_iommu
*iommu
;
1314 unsigned long flags
;
1316 if (devid
> amd_iommu_last_bdf
)
1319 devid
= amd_iommu_alias_table
[devid
];
1321 iommu
= amd_iommu_rlookup_table
[devid
];
1325 domain
= domain_for_device(devid
);
1327 if (domain
&& !dma_ops_domain(domain
))
1328 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
1329 "to a non-dma-ops domain\n", dev_name(dev
));
1332 case BUS_NOTIFY_UNBOUND_DRIVER
:
1335 if (iommu_pass_through
)
1337 detach_device(domain
, devid
);
1339 case BUS_NOTIFY_ADD_DEVICE
:
1340 /* allocate a protection domain if a device is added */
1341 dma_domain
= find_protection_domain(devid
);
1344 dma_domain
= dma_ops_domain_alloc(iommu
);
1347 dma_domain
->target_dev
= devid
;
1349 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1350 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1351 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1358 iommu_queue_inv_dev_entry(iommu
, devid
);
1359 iommu_completion_wait(iommu
);
1365 static struct notifier_block device_nb
= {
1366 .notifier_call
= device_change_notifier
,
1369 /*****************************************************************************
1371 * The next functions belong to the dma_ops mapping/unmapping code.
1373 *****************************************************************************/
1376 * This function checks if the driver got a valid device from the caller to
1377 * avoid dereferencing invalid pointers.
1379 static bool check_device(struct device
*dev
)
1382 struct pci_dev
*pcidev
;
1384 if (!dev
|| !dev
->dma_mask
)
1387 /* No device or no PCI device */
1388 if (!dev
|| dev
->bus
!= &pci_bus_type
)
1391 pcidev
= to_pci_dev(dev
);
1393 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1395 /* Out of our scope? */
1396 if (bdf
> amd_iommu_last_bdf
)
1399 if (amd_iommu_rlookup_table
[bdf
] == NULL
)
1406 * In this function the list of preallocated protection domains is traversed to
1407 * find the domain for a specific device
1409 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
1411 struct dma_ops_domain
*entry
, *ret
= NULL
;
1412 unsigned long flags
;
1414 if (list_empty(&iommu_pd_list
))
1417 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1419 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
1420 if (entry
->target_dev
== devid
) {
1426 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1432 * In the dma_ops path we only have the struct device. This function
1433 * finds the corresponding IOMMU, the protection domain and the
1434 * requestor id for a given device.
1435 * If the device is not yet associated with a domain this is also done
1438 static bool get_device_resources(struct device
*dev
,
1439 struct protection_domain
**domain
,
1442 struct dma_ops_domain
*dma_dom
;
1443 struct amd_iommu
*iommu
;
1444 struct pci_dev
*pcidev
;
1447 if (!check_device(dev
))
1450 pcidev
= to_pci_dev(dev
);
1451 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1452 *bdf
= amd_iommu_alias_table
[_bdf
];
1453 iommu
= amd_iommu_rlookup_table
[*bdf
];
1454 *domain
= domain_for_device(*bdf
);
1456 if (*domain
== NULL
) {
1457 dma_dom
= find_protection_domain(*bdf
);
1459 dma_dom
= iommu
->default_dom
;
1460 *domain
= &dma_dom
->domain
;
1461 attach_device(iommu
, *domain
, *bdf
);
1462 DUMP_printk("Using protection domain %d for device %s\n",
1463 (*domain
)->id
, dev_name(dev
));
1466 if (domain_for_device(_bdf
) == NULL
)
1467 attach_device(iommu
, *domain
, _bdf
);
1472 static void update_device_table(struct protection_domain
*domain
)
1474 unsigned long flags
;
1477 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
1478 if (amd_iommu_pd_table
[i
] != domain
)
1480 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1481 set_dte_entry(i
, domain
);
1482 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1486 static void update_domain(struct protection_domain
*domain
)
1488 if (!domain
->updated
)
1491 update_device_table(domain
);
1492 flush_devices_by_domain(domain
);
1493 iommu_flush_tlb_pde(domain
);
1495 domain
->updated
= false;
1499 * This function is used to add another level to an IO page table. Adding
1500 * another level increases the size of the address space by 9 bits to a size up
1503 static bool increase_address_space(struct protection_domain
*domain
,
1508 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1509 /* address space already 64 bit large */
1512 pte
= (void *)get_zeroed_page(gfp
);
1516 *pte
= PM_LEVEL_PDE(domain
->mode
,
1517 virt_to_phys(domain
->pt_root
));
1518 domain
->pt_root
= pte
;
1520 domain
->updated
= true;
1525 static u64
*alloc_pte(struct protection_domain
*domain
,
1526 unsigned long address
,
1534 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1535 increase_address_space(domain
, gfp
);
1537 level
= domain
->mode
- 1;
1538 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1540 while (level
> end_lvl
) {
1541 if (!IOMMU_PTE_PRESENT(*pte
)) {
1542 page
= (u64
*)get_zeroed_page(gfp
);
1545 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1550 pte
= IOMMU_PTE_PAGE(*pte
);
1552 if (pte_page
&& level
== end_lvl
)
1555 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1562 * This function fetches the PTE for a given address in the aperture
1564 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
1565 unsigned long address
)
1567 struct aperture_range
*aperture
;
1568 u64
*pte
, *pte_page
;
1570 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1574 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1576 pte
= alloc_pte(&dom
->domain
, address
, PM_MAP_4k
, &pte_page
,
1578 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
1580 pte
+= PM_LEVEL_INDEX(0, address
);
1582 update_domain(&dom
->domain
);
1588 * This is the generic map function. It maps one 4kb page at paddr to
1589 * the given address in the DMA address space for the domain.
1591 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
1592 unsigned long address
,
1598 WARN_ON(address
> dom
->aperture_size
);
1602 pte
= dma_ops_get_pte(dom
, address
);
1604 return DMA_ERROR_CODE
;
1606 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1608 if (direction
== DMA_TO_DEVICE
)
1609 __pte
|= IOMMU_PTE_IR
;
1610 else if (direction
== DMA_FROM_DEVICE
)
1611 __pte
|= IOMMU_PTE_IW
;
1612 else if (direction
== DMA_BIDIRECTIONAL
)
1613 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1619 return (dma_addr_t
)address
;
1623 * The generic unmapping function for on page in the DMA address space.
1625 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
1626 unsigned long address
)
1628 struct aperture_range
*aperture
;
1631 if (address
>= dom
->aperture_size
)
1634 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1638 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1642 pte
+= PM_LEVEL_INDEX(0, address
);
1650 * This function contains common code for mapping of a physically
1651 * contiguous memory region into DMA address space. It is used by all
1652 * mapping functions provided with this IOMMU driver.
1653 * Must be called with the domain lock held.
1655 static dma_addr_t
__map_single(struct device
*dev
,
1656 struct dma_ops_domain
*dma_dom
,
1663 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1664 dma_addr_t address
, start
, ret
;
1666 unsigned long align_mask
= 0;
1669 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1672 INC_STATS_COUNTER(total_map_requests
);
1675 INC_STATS_COUNTER(cross_page
);
1678 align_mask
= (1UL << get_order(size
)) - 1;
1681 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1683 if (unlikely(address
== DMA_ERROR_CODE
)) {
1685 * setting next_address here will let the address
1686 * allocator only scan the new allocated range in the
1687 * first run. This is a small optimization.
1689 dma_dom
->next_address
= dma_dom
->aperture_size
;
1691 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
1695 * aperture was sucessfully enlarged by 128 MB, try
1702 for (i
= 0; i
< pages
; ++i
) {
1703 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
1704 if (ret
== DMA_ERROR_CODE
)
1712 ADD_STATS_COUNTER(alloced_io_mem
, size
);
1714 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1715 iommu_flush_tlb(&dma_dom
->domain
);
1716 dma_dom
->need_flush
= false;
1717 } else if (unlikely(amd_iommu_np_cache
))
1718 iommu_flush_pages(&dma_dom
->domain
, address
, size
);
1725 for (--i
; i
>= 0; --i
) {
1727 dma_ops_domain_unmap(dma_dom
, start
);
1730 dma_ops_free_addresses(dma_dom
, address
, pages
);
1732 return DMA_ERROR_CODE
;
1736 * Does the reverse of the __map_single function. Must be called with
1737 * the domain lock held too
1739 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
1740 dma_addr_t dma_addr
,
1744 dma_addr_t i
, start
;
1747 if ((dma_addr
== DMA_ERROR_CODE
) ||
1748 (dma_addr
+ size
> dma_dom
->aperture_size
))
1751 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1752 dma_addr
&= PAGE_MASK
;
1755 for (i
= 0; i
< pages
; ++i
) {
1756 dma_ops_domain_unmap(dma_dom
, start
);
1760 SUB_STATS_COUNTER(alloced_io_mem
, size
);
1762 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1764 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1765 iommu_flush_pages(&dma_dom
->domain
, dma_addr
, size
);
1766 dma_dom
->need_flush
= false;
1771 * The exported map_single function for dma_ops.
1773 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
1774 unsigned long offset
, size_t size
,
1775 enum dma_data_direction dir
,
1776 struct dma_attrs
*attrs
)
1778 unsigned long flags
;
1779 struct protection_domain
*domain
;
1783 phys_addr_t paddr
= page_to_phys(page
) + offset
;
1785 INC_STATS_COUNTER(cnt_map_single
);
1787 if (!get_device_resources(dev
, &domain
, &devid
))
1788 /* device not handled by any AMD IOMMU */
1789 return (dma_addr_t
)paddr
;
1791 dma_mask
= *dev
->dma_mask
;
1793 if (!dma_ops_domain(domain
))
1794 return DMA_ERROR_CODE
;
1796 spin_lock_irqsave(&domain
->lock
, flags
);
1797 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
1799 if (addr
== DMA_ERROR_CODE
)
1802 iommu_flush_complete(domain
);
1805 spin_unlock_irqrestore(&domain
->lock
, flags
);
1811 * The exported unmap_single function for dma_ops.
1813 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
1814 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
1816 unsigned long flags
;
1817 struct protection_domain
*domain
;
1820 INC_STATS_COUNTER(cnt_unmap_single
);
1822 if (!get_device_resources(dev
, &domain
, &devid
))
1823 /* device not handled by any AMD IOMMU */
1826 if (!dma_ops_domain(domain
))
1829 spin_lock_irqsave(&domain
->lock
, flags
);
1831 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
1833 iommu_flush_complete(domain
);
1835 spin_unlock_irqrestore(&domain
->lock
, flags
);
1839 * This is a special map_sg function which is used if we should map a
1840 * device which is not handled by an AMD IOMMU in the system.
1842 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1843 int nelems
, int dir
)
1845 struct scatterlist
*s
;
1848 for_each_sg(sglist
, s
, nelems
, i
) {
1849 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1850 s
->dma_length
= s
->length
;
1857 * The exported map_sg function for dma_ops (handles scatter-gather
1860 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1861 int nelems
, enum dma_data_direction dir
,
1862 struct dma_attrs
*attrs
)
1864 unsigned long flags
;
1865 struct protection_domain
*domain
;
1868 struct scatterlist
*s
;
1870 int mapped_elems
= 0;
1873 INC_STATS_COUNTER(cnt_map_sg
);
1875 if (!get_device_resources(dev
, &domain
, &devid
))
1876 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1878 dma_mask
= *dev
->dma_mask
;
1880 if (!dma_ops_domain(domain
))
1883 spin_lock_irqsave(&domain
->lock
, flags
);
1885 for_each_sg(sglist
, s
, nelems
, i
) {
1888 s
->dma_address
= __map_single(dev
, domain
->priv
,
1889 paddr
, s
->length
, dir
, false,
1892 if (s
->dma_address
) {
1893 s
->dma_length
= s
->length
;
1899 iommu_flush_complete(domain
);
1902 spin_unlock_irqrestore(&domain
->lock
, flags
);
1904 return mapped_elems
;
1906 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1908 __unmap_single(domain
->priv
, s
->dma_address
,
1909 s
->dma_length
, dir
);
1910 s
->dma_address
= s
->dma_length
= 0;
1919 * The exported map_sg function for dma_ops (handles scatter-gather
1922 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1923 int nelems
, enum dma_data_direction dir
,
1924 struct dma_attrs
*attrs
)
1926 unsigned long flags
;
1927 struct protection_domain
*domain
;
1928 struct scatterlist
*s
;
1932 INC_STATS_COUNTER(cnt_unmap_sg
);
1934 if (!get_device_resources(dev
, &domain
, &devid
))
1937 if (!dma_ops_domain(domain
))
1940 spin_lock_irqsave(&domain
->lock
, flags
);
1942 for_each_sg(sglist
, s
, nelems
, i
) {
1943 __unmap_single(domain
->priv
, s
->dma_address
,
1944 s
->dma_length
, dir
);
1945 s
->dma_address
= s
->dma_length
= 0;
1948 iommu_flush_complete(domain
);
1950 spin_unlock_irqrestore(&domain
->lock
, flags
);
1954 * The exported alloc_coherent function for dma_ops.
1956 static void *alloc_coherent(struct device
*dev
, size_t size
,
1957 dma_addr_t
*dma_addr
, gfp_t flag
)
1959 unsigned long flags
;
1961 struct protection_domain
*domain
;
1964 u64 dma_mask
= dev
->coherent_dma_mask
;
1966 INC_STATS_COUNTER(cnt_alloc_coherent
);
1968 if (!get_device_resources(dev
, &domain
, &devid
)) {
1969 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1970 *dma_addr
= __pa(virt_addr
);
1974 dma_mask
= dev
->coherent_dma_mask
;
1975 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1978 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1982 paddr
= virt_to_phys(virt_addr
);
1984 if (!dma_ops_domain(domain
))
1988 dma_mask
= *dev
->dma_mask
;
1990 spin_lock_irqsave(&domain
->lock
, flags
);
1992 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
1993 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
1995 if (*dma_addr
== DMA_ERROR_CODE
) {
1996 spin_unlock_irqrestore(&domain
->lock
, flags
);
2000 iommu_flush_complete(domain
);
2002 spin_unlock_irqrestore(&domain
->lock
, flags
);
2008 free_pages((unsigned long)virt_addr
, get_order(size
));
2014 * The exported free_coherent function for dma_ops.
2016 static void free_coherent(struct device
*dev
, size_t size
,
2017 void *virt_addr
, dma_addr_t dma_addr
)
2019 unsigned long flags
;
2020 struct protection_domain
*domain
;
2023 INC_STATS_COUNTER(cnt_free_coherent
);
2025 if (!get_device_resources(dev
, &domain
, &devid
))
2028 if (!dma_ops_domain(domain
))
2031 spin_lock_irqsave(&domain
->lock
, flags
);
2033 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2035 iommu_flush_complete(domain
);
2037 spin_unlock_irqrestore(&domain
->lock
, flags
);
2040 free_pages((unsigned long)virt_addr
, get_order(size
));
2044 * This function is called by the DMA layer to find out if we can handle a
2045 * particular device. It is part of the dma_ops.
2047 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2049 return check_device(dev
);
2053 * The function for pre-allocating protection domains.
2055 * If the driver core informs the DMA layer if a driver grabs a device
2056 * we don't need to preallocate the protection domains anymore.
2057 * For now we have to.
2059 static void prealloc_protection_domains(void)
2061 struct pci_dev
*dev
= NULL
;
2062 struct dma_ops_domain
*dma_dom
;
2063 struct amd_iommu
*iommu
;
2066 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
2067 __devid
= devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
2068 if (devid
> amd_iommu_last_bdf
)
2070 devid
= amd_iommu_alias_table
[devid
];
2071 if (domain_for_device(devid
))
2073 iommu
= amd_iommu_rlookup_table
[devid
];
2076 dma_dom
= dma_ops_domain_alloc(iommu
);
2079 init_unity_mappings_for_device(dma_dom
, devid
);
2080 dma_dom
->target_dev
= devid
;
2082 attach_device(iommu
, &dma_dom
->domain
, devid
);
2083 if (__devid
!= devid
)
2084 attach_device(iommu
, &dma_dom
->domain
, __devid
);
2086 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2090 static struct dma_map_ops amd_iommu_dma_ops
= {
2091 .alloc_coherent
= alloc_coherent
,
2092 .free_coherent
= free_coherent
,
2093 .map_page
= map_page
,
2094 .unmap_page
= unmap_page
,
2096 .unmap_sg
= unmap_sg
,
2097 .dma_supported
= amd_iommu_dma_supported
,
2101 * The function which clues the AMD IOMMU driver into dma_ops.
2103 int __init
amd_iommu_init_dma_ops(void)
2105 struct amd_iommu
*iommu
;
2109 * first allocate a default protection domain for every IOMMU we
2110 * found in the system. Devices not assigned to any other
2111 * protection domain will be assigned to the default one.
2113 for_each_iommu(iommu
) {
2114 iommu
->default_dom
= dma_ops_domain_alloc(iommu
);
2115 if (iommu
->default_dom
== NULL
)
2117 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2118 ret
= iommu_init_unity_mappings(iommu
);
2124 * If device isolation is enabled, pre-allocate the protection
2125 * domains for each device.
2127 if (amd_iommu_isolate
)
2128 prealloc_protection_domains();
2132 #ifdef CONFIG_GART_IOMMU
2133 gart_iommu_aperture_disabled
= 1;
2134 gart_iommu_aperture
= 0;
2137 /* Make the driver finally visible to the drivers */
2138 dma_ops
= &amd_iommu_dma_ops
;
2140 register_iommu(&amd_iommu_ops
);
2142 bus_register_notifier(&pci_bus_type
, &device_nb
);
2144 amd_iommu_stats_init();
2150 for_each_iommu(iommu
) {
2151 if (iommu
->default_dom
)
2152 dma_ops_domain_free(iommu
->default_dom
);
2158 /*****************************************************************************
2160 * The following functions belong to the exported interface of AMD IOMMU
2162 * This interface allows access to lower level functions of the IOMMU
2163 * like protection domain handling and assignement of devices to domains
2164 * which is not possible with the dma_ops interface.
2166 *****************************************************************************/
2168 static void cleanup_domain(struct protection_domain
*domain
)
2170 unsigned long flags
;
2173 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2175 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
2176 if (amd_iommu_pd_table
[devid
] == domain
)
2177 __detach_device(domain
, devid
);
2179 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2182 static void protection_domain_free(struct protection_domain
*domain
)
2187 del_domain_from_list(domain
);
2190 domain_id_free(domain
->id
);
2195 static struct protection_domain
*protection_domain_alloc(void)
2197 struct protection_domain
*domain
;
2199 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2203 spin_lock_init(&domain
->lock
);
2204 domain
->id
= domain_id_alloc();
2208 add_domain_to_list(domain
);
2218 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2220 struct protection_domain
*domain
;
2222 domain
= protection_domain_alloc();
2226 domain
->mode
= PAGE_MODE_3_LEVEL
;
2227 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2228 if (!domain
->pt_root
)
2236 protection_domain_free(domain
);
2241 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2243 struct protection_domain
*domain
= dom
->priv
;
2248 if (domain
->dev_cnt
> 0)
2249 cleanup_domain(domain
);
2251 BUG_ON(domain
->dev_cnt
!= 0);
2253 free_pagetable(domain
);
2255 domain_id_free(domain
->id
);
2262 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2265 struct protection_domain
*domain
= dom
->priv
;
2266 struct amd_iommu
*iommu
;
2267 struct pci_dev
*pdev
;
2270 if (dev
->bus
!= &pci_bus_type
)
2273 pdev
= to_pci_dev(dev
);
2275 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2278 detach_device(domain
, devid
);
2280 iommu
= amd_iommu_rlookup_table
[devid
];
2284 iommu_queue_inv_dev_entry(iommu
, devid
);
2285 iommu_completion_wait(iommu
);
2288 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2291 struct protection_domain
*domain
= dom
->priv
;
2292 struct protection_domain
*old_domain
;
2293 struct amd_iommu
*iommu
;
2294 struct pci_dev
*pdev
;
2297 if (dev
->bus
!= &pci_bus_type
)
2300 pdev
= to_pci_dev(dev
);
2302 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2304 if (devid
>= amd_iommu_last_bdf
||
2305 devid
!= amd_iommu_alias_table
[devid
])
2308 iommu
= amd_iommu_rlookup_table
[devid
];
2312 old_domain
= domain_for_device(devid
);
2314 detach_device(old_domain
, devid
);
2316 attach_device(iommu
, domain
, devid
);
2318 iommu_completion_wait(iommu
);
2323 static int amd_iommu_map_range(struct iommu_domain
*dom
,
2324 unsigned long iova
, phys_addr_t paddr
,
2325 size_t size
, int iommu_prot
)
2327 struct protection_domain
*domain
= dom
->priv
;
2328 unsigned long i
, npages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2332 if (iommu_prot
& IOMMU_READ
)
2333 prot
|= IOMMU_PROT_IR
;
2334 if (iommu_prot
& IOMMU_WRITE
)
2335 prot
|= IOMMU_PROT_IW
;
2340 for (i
= 0; i
< npages
; ++i
) {
2341 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, PM_MAP_4k
);
2352 static void amd_iommu_unmap_range(struct iommu_domain
*dom
,
2353 unsigned long iova
, size_t size
)
2356 struct protection_domain
*domain
= dom
->priv
;
2357 unsigned long i
, npages
= iommu_num_pages(iova
, size
, PAGE_SIZE
);
2361 for (i
= 0; i
< npages
; ++i
) {
2362 iommu_unmap_page(domain
, iova
, PM_MAP_4k
);
2366 iommu_flush_tlb_pde(domain
);
2369 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2372 struct protection_domain
*domain
= dom
->priv
;
2373 unsigned long offset
= iova
& ~PAGE_MASK
;
2377 pte
= fetch_pte(domain
, iova
, PM_MAP_4k
);
2379 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
2382 paddr
= *pte
& IOMMU_PAGE_MASK
;
2388 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
2394 static struct iommu_ops amd_iommu_ops
= {
2395 .domain_init
= amd_iommu_domain_init
,
2396 .domain_destroy
= amd_iommu_domain_destroy
,
2397 .attach_dev
= amd_iommu_attach_device
,
2398 .detach_dev
= amd_iommu_detach_device
,
2399 .map
= amd_iommu_map_range
,
2400 .unmap
= amd_iommu_unmap_range
,
2401 .iova_to_phys
= amd_iommu_iova_to_phys
,
2402 .domain_has_cap
= amd_iommu_domain_has_cap
,
2405 /*****************************************************************************
2407 * The next functions do a basic initialization of IOMMU for pass through
2410 * In passthrough mode the IOMMU is initialized and enabled but not used for
2411 * DMA-API translation.
2413 *****************************************************************************/
2415 int __init
amd_iommu_init_passthrough(void)
2417 struct pci_dev
*dev
= NULL
;
2420 /* allocate passthroug domain */
2421 pt_domain
= protection_domain_alloc();
2425 pt_domain
->mode
|= PAGE_MODE_NONE
;
2427 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
2428 struct amd_iommu
*iommu
;
2430 devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
2431 if (devid
> amd_iommu_last_bdf
)
2434 devid2
= amd_iommu_alias_table
[devid
];
2436 iommu
= amd_iommu_rlookup_table
[devid2
];
2440 __attach_device(iommu
, pt_domain
, devid
);
2441 __attach_device(iommu
, pt_domain
, devid2
);
2444 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");