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 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
135 static int iommu_has_npcache(struct amd_iommu
*iommu
)
137 return iommu
->cap
& (1UL << IOMMU_CAP_NPCACHE
);
140 /****************************************************************************
142 * Interrupt handling functions
144 ****************************************************************************/
146 static void dump_dte_entry(u16 devid
)
150 for (i
= 0; i
< 8; ++i
)
151 pr_err("AMD-Vi: DTE[%d]: %08x\n", i
,
152 amd_iommu_dev_table
[devid
].data
[i
]);
155 static void dump_command(unsigned long phys_addr
)
157 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
160 for (i
= 0; i
< 4; ++i
)
161 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
164 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
167 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
168 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
169 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
170 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
171 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
173 printk(KERN_ERR
"AMD-Vi: Event logged [");
176 case EVENT_TYPE_ILL_DEV
:
177 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
178 "address=0x%016llx flags=0x%04x]\n",
179 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
181 dump_dte_entry(devid
);
183 case EVENT_TYPE_IO_FAULT
:
184 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
185 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
186 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
187 domid
, address
, flags
);
189 case EVENT_TYPE_DEV_TAB_ERR
:
190 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
191 "address=0x%016llx flags=0x%04x]\n",
192 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
195 case EVENT_TYPE_PAGE_TAB_ERR
:
196 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
197 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
198 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
199 domid
, address
, flags
);
201 case EVENT_TYPE_ILL_CMD
:
202 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
203 reset_iommu_command_buffer(iommu
);
204 dump_command(address
);
206 case EVENT_TYPE_CMD_HARD_ERR
:
207 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
208 "flags=0x%04x]\n", address
, flags
);
210 case EVENT_TYPE_IOTLB_INV_TO
:
211 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
212 "address=0x%016llx]\n",
213 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
216 case EVENT_TYPE_INV_DEV_REQ
:
217 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
218 "address=0x%016llx flags=0x%04x]\n",
219 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
223 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
227 static void iommu_poll_events(struct amd_iommu
*iommu
)
232 spin_lock_irqsave(&iommu
->lock
, flags
);
234 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
235 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
237 while (head
!= tail
) {
238 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
239 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
242 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
244 spin_unlock_irqrestore(&iommu
->lock
, flags
);
247 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
249 struct amd_iommu
*iommu
;
251 for_each_iommu(iommu
)
252 iommu_poll_events(iommu
);
257 /****************************************************************************
259 * IOMMU command queuing functions
261 ****************************************************************************/
264 * Writes the command to the IOMMUs command buffer and informs the
265 * hardware about the new command. Must be called with iommu->lock held.
267 static int __iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
272 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
273 target
= iommu
->cmd_buf
+ tail
;
274 memcpy_toio(target
, cmd
, sizeof(*cmd
));
275 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
276 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
279 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
285 * General queuing function for commands. Takes iommu->lock and calls
286 * __iommu_queue_command().
288 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
293 spin_lock_irqsave(&iommu
->lock
, flags
);
294 ret
= __iommu_queue_command(iommu
, cmd
);
296 iommu
->need_sync
= true;
297 spin_unlock_irqrestore(&iommu
->lock
, flags
);
303 * This function waits until an IOMMU has completed a completion
306 static void __iommu_wait_for_completion(struct amd_iommu
*iommu
)
312 INC_STATS_COUNTER(compl_wait
);
314 while (!ready
&& (i
< EXIT_LOOP_COUNT
)) {
316 /* wait for the bit to become one */
317 status
= readl(iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
318 ready
= status
& MMIO_STATUS_COM_WAIT_INT_MASK
;
321 /* set bit back to zero */
322 status
&= ~MMIO_STATUS_COM_WAIT_INT_MASK
;
323 writel(status
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
325 if (unlikely(i
== EXIT_LOOP_COUNT
)) {
326 spin_unlock(&iommu
->lock
);
327 reset_iommu_command_buffer(iommu
);
328 spin_lock(&iommu
->lock
);
333 * This function queues a completion wait command into the command
336 static int __iommu_completion_wait(struct amd_iommu
*iommu
)
338 struct iommu_cmd cmd
;
340 memset(&cmd
, 0, sizeof(cmd
));
341 cmd
.data
[0] = CMD_COMPL_WAIT_INT_MASK
;
342 CMD_SET_TYPE(&cmd
, CMD_COMPL_WAIT
);
344 return __iommu_queue_command(iommu
, &cmd
);
348 * This function is called whenever we need to ensure that the IOMMU has
349 * completed execution of all commands we sent. It sends a
350 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
351 * us about that by writing a value to a physical address we pass with
354 static int iommu_completion_wait(struct amd_iommu
*iommu
)
359 spin_lock_irqsave(&iommu
->lock
, flags
);
361 if (!iommu
->need_sync
)
364 ret
= __iommu_completion_wait(iommu
);
366 iommu
->need_sync
= false;
371 __iommu_wait_for_completion(iommu
);
374 spin_unlock_irqrestore(&iommu
->lock
, flags
);
379 static void iommu_flush_complete(struct protection_domain
*domain
)
383 for (i
= 0; i
< amd_iommus_present
; ++i
) {
384 if (!domain
->dev_iommu
[i
])
388 * Devices of this domain are behind this IOMMU
389 * We need to wait for completion of all commands.
391 iommu_completion_wait(amd_iommus
[i
]);
396 * Command send function for invalidating a device table entry
398 static int iommu_queue_inv_dev_entry(struct amd_iommu
*iommu
, u16 devid
)
400 struct iommu_cmd cmd
;
403 BUG_ON(iommu
== NULL
);
405 memset(&cmd
, 0, sizeof(cmd
));
406 CMD_SET_TYPE(&cmd
, CMD_INV_DEV_ENTRY
);
409 ret
= iommu_queue_command(iommu
, &cmd
);
414 static void __iommu_build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
415 u16 domid
, int pde
, int s
)
417 memset(cmd
, 0, sizeof(*cmd
));
418 address
&= PAGE_MASK
;
419 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
420 cmd
->data
[1] |= domid
;
421 cmd
->data
[2] = lower_32_bits(address
);
422 cmd
->data
[3] = upper_32_bits(address
);
423 if (s
) /* size bit - we flush more than one 4kb page */
424 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
425 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
426 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
430 * Generic command send function for invalidaing TLB entries
432 static int iommu_queue_inv_iommu_pages(struct amd_iommu
*iommu
,
433 u64 address
, u16 domid
, int pde
, int s
)
435 struct iommu_cmd cmd
;
438 __iommu_build_inv_iommu_pages(&cmd
, address
, domid
, pde
, s
);
440 ret
= iommu_queue_command(iommu
, &cmd
);
446 * TLB invalidation function which is called from the mapping functions.
447 * It invalidates a single PTE if the range to flush is within a single
448 * page. Otherwise it flushes the whole TLB of the IOMMU.
450 static void __iommu_flush_pages(struct protection_domain
*domain
,
451 u64 address
, size_t size
, int pde
)
454 unsigned long pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
456 address
&= PAGE_MASK
;
460 * If we have to flush more than one page, flush all
461 * TLB entries for this domain
463 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
468 for (i
= 0; i
< amd_iommus_present
; ++i
) {
469 if (!domain
->dev_iommu
[i
])
473 * Devices of this domain are behind this IOMMU
474 * We need a TLB flush
476 iommu_queue_inv_iommu_pages(amd_iommus
[i
], address
,
483 static void iommu_flush_pages(struct protection_domain
*domain
,
484 u64 address
, size_t size
)
486 __iommu_flush_pages(domain
, address
, size
, 0);
489 /* Flush the whole IO/TLB for a given protection domain */
490 static void iommu_flush_tlb(struct protection_domain
*domain
)
492 __iommu_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
495 /* Flush the whole IO/TLB for a given protection domain - including PDE */
496 static void iommu_flush_tlb_pde(struct protection_domain
*domain
)
498 __iommu_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
502 * This function flushes one domain on one IOMMU
504 static void flush_domain_on_iommu(struct amd_iommu
*iommu
, u16 domid
)
506 struct iommu_cmd cmd
;
509 __iommu_build_inv_iommu_pages(&cmd
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
512 spin_lock_irqsave(&iommu
->lock
, flags
);
513 __iommu_queue_command(iommu
, &cmd
);
514 __iommu_completion_wait(iommu
);
515 __iommu_wait_for_completion(iommu
);
516 spin_unlock_irqrestore(&iommu
->lock
, flags
);
519 static void flush_all_domains_on_iommu(struct amd_iommu
*iommu
)
523 for (i
= 1; i
< MAX_DOMAIN_ID
; ++i
) {
524 if (!test_bit(i
, amd_iommu_pd_alloc_bitmap
))
526 flush_domain_on_iommu(iommu
, i
);
531 void amd_iommu_flush_all_domains(void)
533 struct protection_domain
*domain
;
535 list_for_each_entry(domain
, &amd_iommu_pd_list
, list
) {
536 iommu_flush_tlb_pde(domain
);
537 iommu_flush_complete(domain
);
541 static void flush_all_devices_for_iommu(struct amd_iommu
*iommu
)
545 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
546 if (iommu
!= amd_iommu_rlookup_table
[i
])
549 iommu_queue_inv_dev_entry(iommu
, i
);
550 iommu_completion_wait(iommu
);
554 static void flush_devices_by_domain(struct protection_domain
*domain
)
556 struct amd_iommu
*iommu
;
559 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
560 if ((domain
== NULL
&& amd_iommu_pd_table
[i
] == NULL
) ||
561 (amd_iommu_pd_table
[i
] != domain
))
564 iommu
= amd_iommu_rlookup_table
[i
];
568 iommu_queue_inv_dev_entry(iommu
, i
);
569 iommu_completion_wait(iommu
);
573 static void reset_iommu_command_buffer(struct amd_iommu
*iommu
)
575 pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
577 if (iommu
->reset_in_progress
)
578 panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
580 iommu
->reset_in_progress
= true;
582 amd_iommu_reset_cmd_buffer(iommu
);
583 flush_all_devices_for_iommu(iommu
);
584 flush_all_domains_on_iommu(iommu
);
586 iommu
->reset_in_progress
= false;
589 void amd_iommu_flush_all_devices(void)
591 flush_devices_by_domain(NULL
);
594 /****************************************************************************
596 * The functions below are used the create the page table mappings for
597 * unity mapped regions.
599 ****************************************************************************/
602 * Generic mapping functions. It maps a physical address into a DMA
603 * address space. It allocates the page table pages if necessary.
604 * In the future it can be extended to a generic mapping function
605 * supporting all features of AMD IOMMU page tables like level skipping
606 * and full 64 bit address spaces.
608 static int iommu_map_page(struct protection_domain
*dom
,
609 unsigned long bus_addr
,
610 unsigned long phys_addr
,
616 bus_addr
= PAGE_ALIGN(bus_addr
);
617 phys_addr
= PAGE_ALIGN(phys_addr
);
619 BUG_ON(!PM_ALIGNED(map_size
, bus_addr
));
620 BUG_ON(!PM_ALIGNED(map_size
, phys_addr
));
622 if (!(prot
& IOMMU_PROT_MASK
))
625 pte
= alloc_pte(dom
, bus_addr
, map_size
, NULL
, GFP_KERNEL
);
627 if (IOMMU_PTE_PRESENT(*pte
))
630 __pte
= phys_addr
| IOMMU_PTE_P
;
631 if (prot
& IOMMU_PROT_IR
)
632 __pte
|= IOMMU_PTE_IR
;
633 if (prot
& IOMMU_PROT_IW
)
634 __pte
|= IOMMU_PTE_IW
;
643 static void iommu_unmap_page(struct protection_domain
*dom
,
644 unsigned long bus_addr
, int map_size
)
646 u64
*pte
= fetch_pte(dom
, bus_addr
, map_size
);
653 * This function checks if a specific unity mapping entry is needed for
654 * this specific IOMMU.
656 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
657 struct unity_map_entry
*entry
)
661 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
662 bdf
= amd_iommu_alias_table
[i
];
663 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
671 * Init the unity mappings for a specific IOMMU in the system
673 * Basically iterates over all unity mapping entries and applies them to
674 * the default domain DMA of that IOMMU if necessary.
676 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
678 struct unity_map_entry
*entry
;
681 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
682 if (!iommu_for_unity_map(iommu
, entry
))
684 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
693 * This function actually applies the mapping to the page table of the
696 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
697 struct unity_map_entry
*e
)
702 for (addr
= e
->address_start
; addr
< e
->address_end
;
704 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
709 * if unity mapping is in aperture range mark the page
710 * as allocated in the aperture
712 if (addr
< dma_dom
->aperture_size
)
713 __set_bit(addr
>> PAGE_SHIFT
,
714 dma_dom
->aperture
[0]->bitmap
);
721 * Inits the unity mappings required for a specific device
723 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
726 struct unity_map_entry
*e
;
729 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
730 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
732 ret
= dma_ops_unity_map(dma_dom
, e
);
740 /****************************************************************************
742 * The next functions belong to the address allocator for the dma_ops
743 * interface functions. They work like the allocators in the other IOMMU
744 * drivers. Its basically a bitmap which marks the allocated pages in
745 * the aperture. Maybe it could be enhanced in the future to a more
746 * efficient allocator.
748 ****************************************************************************/
751 * The address allocator core functions.
753 * called with domain->lock held
757 * This function checks if there is a PTE for a given dma address. If
758 * there is one, it returns the pointer to it.
760 static u64
*fetch_pte(struct protection_domain
*domain
,
761 unsigned long address
, int map_size
)
766 level
= domain
->mode
- 1;
767 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
769 while (level
> map_size
) {
770 if (!IOMMU_PTE_PRESENT(*pte
))
775 pte
= IOMMU_PTE_PAGE(*pte
);
776 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
778 if ((PM_PTE_LEVEL(*pte
) == 0) && level
!= map_size
) {
788 * This function is used to add a new aperture range to an existing
789 * aperture in case of dma_ops domain allocation or address allocation
792 static int alloc_new_range(struct amd_iommu
*iommu
,
793 struct dma_ops_domain
*dma_dom
,
794 bool populate
, gfp_t gfp
)
796 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
799 #ifdef CONFIG_IOMMU_STRESS
803 if (index
>= APERTURE_MAX_RANGES
)
806 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
807 if (!dma_dom
->aperture
[index
])
810 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
811 if (!dma_dom
->aperture
[index
]->bitmap
)
814 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
817 unsigned long address
= dma_dom
->aperture_size
;
818 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
821 for (i
= 0; i
< num_ptes
; ++i
) {
822 pte
= alloc_pte(&dma_dom
->domain
, address
, PM_MAP_4k
,
827 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
829 address
+= APERTURE_RANGE_SIZE
/ 64;
833 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
835 /* Intialize the exclusion range if necessary */
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
= iommu
->exclusion_start
>> PAGE_SHIFT
;
840 int pages
= iommu_num_pages(iommu
->exclusion_start
,
841 iommu
->exclusion_length
,
843 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
847 * Check for areas already mapped as present in the new aperture
848 * range and mark those pages as reserved in the allocator. Such
849 * mappings may already exist as a result of requested unity
850 * mappings for devices.
852 for (i
= dma_dom
->aperture
[index
]->offset
;
853 i
< dma_dom
->aperture_size
;
855 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
, PM_MAP_4k
);
856 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
859 dma_ops_reserve_addresses(dma_dom
, i
<< PAGE_SHIFT
, 1);
862 update_domain(&dma_dom
->domain
);
867 update_domain(&dma_dom
->domain
);
869 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
871 kfree(dma_dom
->aperture
[index
]);
872 dma_dom
->aperture
[index
] = NULL
;
877 static unsigned long dma_ops_area_alloc(struct device
*dev
,
878 struct dma_ops_domain
*dom
,
880 unsigned long align_mask
,
884 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
885 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
886 int i
= start
>> APERTURE_RANGE_SHIFT
;
887 unsigned long boundary_size
;
888 unsigned long address
= -1;
891 next_bit
>>= PAGE_SHIFT
;
893 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
894 PAGE_SIZE
) >> PAGE_SHIFT
;
896 for (;i
< max_index
; ++i
) {
897 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
899 if (dom
->aperture
[i
]->offset
>= dma_mask
)
902 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
903 dma_mask
>> PAGE_SHIFT
);
905 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
906 limit
, next_bit
, pages
, 0,
907 boundary_size
, align_mask
);
909 address
= dom
->aperture
[i
]->offset
+
910 (address
<< PAGE_SHIFT
);
911 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
921 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
922 struct dma_ops_domain
*dom
,
924 unsigned long align_mask
,
927 unsigned long address
;
929 #ifdef CONFIG_IOMMU_STRESS
930 dom
->next_address
= 0;
931 dom
->need_flush
= true;
934 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
935 dma_mask
, dom
->next_address
);
938 dom
->next_address
= 0;
939 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
941 dom
->need_flush
= true;
944 if (unlikely(address
== -1))
945 address
= DMA_ERROR_CODE
;
947 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
953 * The address free function.
955 * called with domain->lock held
957 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
958 unsigned long address
,
961 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
962 struct aperture_range
*range
= dom
->aperture
[i
];
964 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
966 #ifdef CONFIG_IOMMU_STRESS
971 if (address
>= dom
->next_address
)
972 dom
->need_flush
= true;
974 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
976 iommu_area_free(range
->bitmap
, address
, pages
);
980 /****************************************************************************
982 * The next functions belong to the domain allocation. A domain is
983 * allocated for every IOMMU as the default domain. If device isolation
984 * is enabled, every device get its own domain. The most important thing
985 * about domains is the page table mapping the DMA address space they
988 ****************************************************************************/
991 * This function adds a protection domain to the global protection domain list
993 static void add_domain_to_list(struct protection_domain
*domain
)
997 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
998 list_add(&domain
->list
, &amd_iommu_pd_list
);
999 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1003 * This function removes a protection domain to the global
1004 * protection domain list
1006 static void del_domain_from_list(struct protection_domain
*domain
)
1008 unsigned long flags
;
1010 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1011 list_del(&domain
->list
);
1012 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1015 static u16
domain_id_alloc(void)
1017 unsigned long flags
;
1020 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1021 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1023 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1024 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1027 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1032 static void domain_id_free(int id
)
1034 unsigned long flags
;
1036 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1037 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1038 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1039 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1043 * Used to reserve address ranges in the aperture (e.g. for exclusion
1046 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1047 unsigned long start_page
,
1050 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1052 if (start_page
+ pages
> last_page
)
1053 pages
= last_page
- start_page
;
1055 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1056 int index
= i
/ APERTURE_RANGE_PAGES
;
1057 int page
= i
% APERTURE_RANGE_PAGES
;
1058 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1062 static void free_pagetable(struct protection_domain
*domain
)
1067 p1
= domain
->pt_root
;
1072 for (i
= 0; i
< 512; ++i
) {
1073 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1076 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1077 for (j
= 0; j
< 512; ++j
) {
1078 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1080 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1081 free_page((unsigned long)p3
);
1084 free_page((unsigned long)p2
);
1087 free_page((unsigned long)p1
);
1089 domain
->pt_root
= NULL
;
1093 * Free a domain, only used if something went wrong in the
1094 * allocation path and we need to free an already allocated page table
1096 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1103 del_domain_from_list(&dom
->domain
);
1105 free_pagetable(&dom
->domain
);
1107 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1108 if (!dom
->aperture
[i
])
1110 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1111 kfree(dom
->aperture
[i
]);
1118 * Allocates a new protection domain usable for the dma_ops functions.
1119 * It also intializes the page table and the address allocator data
1120 * structures required for the dma_ops interface
1122 static struct dma_ops_domain
*dma_ops_domain_alloc(struct amd_iommu
*iommu
)
1124 struct dma_ops_domain
*dma_dom
;
1126 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1130 spin_lock_init(&dma_dom
->domain
.lock
);
1132 dma_dom
->domain
.id
= domain_id_alloc();
1133 if (dma_dom
->domain
.id
== 0)
1135 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1136 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1137 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1138 dma_dom
->domain
.priv
= dma_dom
;
1139 if (!dma_dom
->domain
.pt_root
)
1142 dma_dom
->need_flush
= false;
1143 dma_dom
->target_dev
= 0xffff;
1145 add_domain_to_list(&dma_dom
->domain
);
1147 if (alloc_new_range(iommu
, dma_dom
, true, GFP_KERNEL
))
1151 * mark the first page as allocated so we never return 0 as
1152 * a valid dma-address. So we can use 0 as error value
1154 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1155 dma_dom
->next_address
= 0;
1161 dma_ops_domain_free(dma_dom
);
1167 * little helper function to check whether a given protection domain is a
1170 static bool dma_ops_domain(struct protection_domain
*domain
)
1172 return domain
->flags
& PD_DMA_OPS_MASK
;
1176 * Find out the protection domain structure for a given PCI device. This
1177 * will give us the pointer to the page table root for example.
1179 static struct protection_domain
*domain_for_device(u16 devid
)
1181 struct protection_domain
*dom
;
1182 unsigned long flags
;
1184 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1185 dom
= amd_iommu_pd_table
[devid
];
1186 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1191 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
)
1193 u64 pte_root
= virt_to_phys(domain
->pt_root
);
1195 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1196 << DEV_ENTRY_MODE_SHIFT
;
1197 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1199 amd_iommu_dev_table
[devid
].data
[2] = domain
->id
;
1200 amd_iommu_dev_table
[devid
].data
[1] = upper_32_bits(pte_root
);
1201 amd_iommu_dev_table
[devid
].data
[0] = lower_32_bits(pte_root
);
1203 amd_iommu_pd_table
[devid
] = domain
;
1207 * If a device is not yet associated with a domain, this function does
1208 * assigns it visible for the hardware
1210 static void __attach_device(struct amd_iommu
*iommu
,
1211 struct protection_domain
*domain
,
1215 spin_lock(&domain
->lock
);
1217 /* update DTE entry */
1218 set_dte_entry(devid
, domain
);
1220 /* Do reference counting */
1221 domain
->dev_iommu
[iommu
->index
] += 1;
1222 domain
->dev_cnt
+= 1;
1225 spin_unlock(&domain
->lock
);
1229 * If a device is not yet associated with a domain, this function does
1230 * assigns it visible for the hardware
1232 static void attach_device(struct amd_iommu
*iommu
,
1233 struct protection_domain
*domain
,
1236 unsigned long flags
;
1238 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1239 __attach_device(iommu
, domain
, devid
);
1240 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1243 * We might boot into a crash-kernel here. The crashed kernel
1244 * left the caches in the IOMMU dirty. So we have to flush
1245 * here to evict all dirty stuff.
1247 iommu_queue_inv_dev_entry(iommu
, devid
);
1248 iommu_flush_tlb_pde(domain
);
1252 * Removes a device from a protection domain (unlocked)
1254 static void __detach_device(struct protection_domain
*domain
, u16 devid
)
1256 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
1261 spin_lock(&domain
->lock
);
1263 /* remove domain from the lookup table */
1264 amd_iommu_pd_table
[devid
] = NULL
;
1266 /* remove entry from the device table seen by the hardware */
1267 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1268 amd_iommu_dev_table
[devid
].data
[1] = 0;
1269 amd_iommu_dev_table
[devid
].data
[2] = 0;
1271 amd_iommu_apply_erratum_63(devid
);
1273 /* decrease reference counters */
1274 domain
->dev_iommu
[iommu
->index
] -= 1;
1275 domain
->dev_cnt
-= 1;
1278 spin_unlock(&domain
->lock
);
1281 * If we run in passthrough mode the device must be assigned to the
1282 * passthrough domain if it is detached from any other domain
1284 if (iommu_pass_through
) {
1285 struct amd_iommu
*iommu
= amd_iommu_rlookup_table
[devid
];
1286 __attach_device(iommu
, pt_domain
, devid
);
1291 * Removes a device from a protection domain (with devtable_lock held)
1293 static void detach_device(struct protection_domain
*domain
, u16 devid
)
1295 unsigned long flags
;
1297 /* lock device table */
1298 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1299 __detach_device(domain
, devid
);
1300 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1303 static int device_change_notifier(struct notifier_block
*nb
,
1304 unsigned long action
, void *data
)
1306 struct device
*dev
= data
;
1307 struct pci_dev
*pdev
= to_pci_dev(dev
);
1308 u16 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
1309 struct protection_domain
*domain
;
1310 struct dma_ops_domain
*dma_domain
;
1311 struct amd_iommu
*iommu
;
1312 unsigned long flags
;
1314 if (devid
> amd_iommu_last_bdf
)
1317 devid
= amd_iommu_alias_table
[devid
];
1319 iommu
= amd_iommu_rlookup_table
[devid
];
1323 domain
= domain_for_device(devid
);
1325 if (domain
&& !dma_ops_domain(domain
))
1326 WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
1327 "to a non-dma-ops domain\n", dev_name(dev
));
1330 case BUS_NOTIFY_UNBOUND_DRIVER
:
1333 if (iommu_pass_through
)
1335 detach_device(domain
, devid
);
1337 case BUS_NOTIFY_ADD_DEVICE
:
1338 /* allocate a protection domain if a device is added */
1339 dma_domain
= find_protection_domain(devid
);
1342 dma_domain
= dma_ops_domain_alloc(iommu
);
1345 dma_domain
->target_dev
= devid
;
1347 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1348 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
1349 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1356 iommu_queue_inv_dev_entry(iommu
, devid
);
1357 iommu_completion_wait(iommu
);
1363 static struct notifier_block device_nb
= {
1364 .notifier_call
= device_change_notifier
,
1367 /*****************************************************************************
1369 * The next functions belong to the dma_ops mapping/unmapping code.
1371 *****************************************************************************/
1374 * This function checks if the driver got a valid device from the caller to
1375 * avoid dereferencing invalid pointers.
1377 static bool check_device(struct device
*dev
)
1379 if (!dev
|| !dev
->dma_mask
)
1386 * In this function the list of preallocated protection domains is traversed to
1387 * find the domain for a specific device
1389 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
1391 struct dma_ops_domain
*entry
, *ret
= NULL
;
1392 unsigned long flags
;
1394 if (list_empty(&iommu_pd_list
))
1397 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
1399 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
1400 if (entry
->target_dev
== devid
) {
1406 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
1412 * In the dma_ops path we only have the struct device. This function
1413 * finds the corresponding IOMMU, the protection domain and the
1414 * requestor id for a given device.
1415 * If the device is not yet associated with a domain this is also done
1418 static int get_device_resources(struct device
*dev
,
1419 struct amd_iommu
**iommu
,
1420 struct protection_domain
**domain
,
1423 struct dma_ops_domain
*dma_dom
;
1424 struct pci_dev
*pcidev
;
1431 if (dev
->bus
!= &pci_bus_type
)
1434 pcidev
= to_pci_dev(dev
);
1435 _bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
1437 /* device not translated by any IOMMU in the system? */
1438 if (_bdf
> amd_iommu_last_bdf
)
1441 *bdf
= amd_iommu_alias_table
[_bdf
];
1443 *iommu
= amd_iommu_rlookup_table
[*bdf
];
1446 *domain
= domain_for_device(*bdf
);
1447 if (*domain
== NULL
) {
1448 dma_dom
= find_protection_domain(*bdf
);
1450 dma_dom
= (*iommu
)->default_dom
;
1451 *domain
= &dma_dom
->domain
;
1452 attach_device(*iommu
, *domain
, *bdf
);
1453 DUMP_printk("Using protection domain %d for device %s\n",
1454 (*domain
)->id
, dev_name(dev
));
1457 if (domain_for_device(_bdf
) == NULL
)
1458 attach_device(*iommu
, *domain
, _bdf
);
1463 static void update_device_table(struct protection_domain
*domain
)
1465 unsigned long flags
;
1468 for (i
= 0; i
<= amd_iommu_last_bdf
; ++i
) {
1469 if (amd_iommu_pd_table
[i
] != domain
)
1471 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1472 set_dte_entry(i
, domain
);
1473 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1477 static void update_domain(struct protection_domain
*domain
)
1479 if (!domain
->updated
)
1482 update_device_table(domain
);
1483 flush_devices_by_domain(domain
);
1484 iommu_flush_tlb_pde(domain
);
1486 domain
->updated
= false;
1490 * This function is used to add another level to an IO page table. Adding
1491 * another level increases the size of the address space by 9 bits to a size up
1494 static bool increase_address_space(struct protection_domain
*domain
,
1499 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1500 /* address space already 64 bit large */
1503 pte
= (void *)get_zeroed_page(gfp
);
1507 *pte
= PM_LEVEL_PDE(domain
->mode
,
1508 virt_to_phys(domain
->pt_root
));
1509 domain
->pt_root
= pte
;
1511 domain
->updated
= true;
1516 static u64
*alloc_pte(struct protection_domain
*domain
,
1517 unsigned long address
,
1525 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1526 increase_address_space(domain
, gfp
);
1528 level
= domain
->mode
- 1;
1529 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1531 while (level
> end_lvl
) {
1532 if (!IOMMU_PTE_PRESENT(*pte
)) {
1533 page
= (u64
*)get_zeroed_page(gfp
);
1536 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1541 pte
= IOMMU_PTE_PAGE(*pte
);
1543 if (pte_page
&& level
== end_lvl
)
1546 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1553 * This function fetches the PTE for a given address in the aperture
1555 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
1556 unsigned long address
)
1558 struct aperture_range
*aperture
;
1559 u64
*pte
, *pte_page
;
1561 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1565 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1567 pte
= alloc_pte(&dom
->domain
, address
, PM_MAP_4k
, &pte_page
,
1569 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
1571 pte
+= PM_LEVEL_INDEX(0, address
);
1573 update_domain(&dom
->domain
);
1579 * This is the generic map function. It maps one 4kb page at paddr to
1580 * the given address in the DMA address space for the domain.
1582 static dma_addr_t
dma_ops_domain_map(struct amd_iommu
*iommu
,
1583 struct dma_ops_domain
*dom
,
1584 unsigned long address
,
1590 WARN_ON(address
> dom
->aperture_size
);
1594 pte
= dma_ops_get_pte(dom
, address
);
1596 return DMA_ERROR_CODE
;
1598 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1600 if (direction
== DMA_TO_DEVICE
)
1601 __pte
|= IOMMU_PTE_IR
;
1602 else if (direction
== DMA_FROM_DEVICE
)
1603 __pte
|= IOMMU_PTE_IW
;
1604 else if (direction
== DMA_BIDIRECTIONAL
)
1605 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
1611 return (dma_addr_t
)address
;
1615 * The generic unmapping function for on page in the DMA address space.
1617 static void dma_ops_domain_unmap(struct amd_iommu
*iommu
,
1618 struct dma_ops_domain
*dom
,
1619 unsigned long address
)
1621 struct aperture_range
*aperture
;
1624 if (address
>= dom
->aperture_size
)
1627 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
1631 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
1635 pte
+= PM_LEVEL_INDEX(0, address
);
1643 * This function contains common code for mapping of a physically
1644 * contiguous memory region into DMA address space. It is used by all
1645 * mapping functions provided with this IOMMU driver.
1646 * Must be called with the domain lock held.
1648 static dma_addr_t
__map_single(struct device
*dev
,
1649 struct amd_iommu
*iommu
,
1650 struct dma_ops_domain
*dma_dom
,
1657 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
1658 dma_addr_t address
, start
, ret
;
1660 unsigned long align_mask
= 0;
1663 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
1666 INC_STATS_COUNTER(total_map_requests
);
1669 INC_STATS_COUNTER(cross_page
);
1672 align_mask
= (1UL << get_order(size
)) - 1;
1675 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
1677 if (unlikely(address
== DMA_ERROR_CODE
)) {
1679 * setting next_address here will let the address
1680 * allocator only scan the new allocated range in the
1681 * first run. This is a small optimization.
1683 dma_dom
->next_address
= dma_dom
->aperture_size
;
1685 if (alloc_new_range(iommu
, dma_dom
, false, GFP_ATOMIC
))
1689 * aperture was sucessfully enlarged by 128 MB, try
1696 for (i
= 0; i
< pages
; ++i
) {
1697 ret
= dma_ops_domain_map(iommu
, dma_dom
, start
, paddr
, dir
);
1698 if (ret
== DMA_ERROR_CODE
)
1706 ADD_STATS_COUNTER(alloced_io_mem
, size
);
1708 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
1709 iommu_flush_tlb(&dma_dom
->domain
);
1710 dma_dom
->need_flush
= false;
1711 } else if (unlikely(iommu_has_npcache(iommu
)))
1712 iommu_flush_pages(&dma_dom
->domain
, address
, size
);
1719 for (--i
; i
>= 0; --i
) {
1721 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1724 dma_ops_free_addresses(dma_dom
, address
, pages
);
1726 return DMA_ERROR_CODE
;
1730 * Does the reverse of the __map_single function. Must be called with
1731 * the domain lock held too
1733 static void __unmap_single(struct amd_iommu
*iommu
,
1734 struct dma_ops_domain
*dma_dom
,
1735 dma_addr_t dma_addr
,
1739 dma_addr_t i
, start
;
1742 if ((dma_addr
== DMA_ERROR_CODE
) ||
1743 (dma_addr
+ size
> dma_dom
->aperture_size
))
1746 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
1747 dma_addr
&= PAGE_MASK
;
1750 for (i
= 0; i
< pages
; ++i
) {
1751 dma_ops_domain_unmap(iommu
, dma_dom
, start
);
1755 SUB_STATS_COUNTER(alloced_io_mem
, size
);
1757 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
1759 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
1760 iommu_flush_pages(&dma_dom
->domain
, dma_addr
, size
);
1761 dma_dom
->need_flush
= false;
1766 * The exported map_single function for dma_ops.
1768 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
1769 unsigned long offset
, size_t size
,
1770 enum dma_data_direction dir
,
1771 struct dma_attrs
*attrs
)
1773 unsigned long flags
;
1774 struct amd_iommu
*iommu
;
1775 struct protection_domain
*domain
;
1779 phys_addr_t paddr
= page_to_phys(page
) + offset
;
1781 INC_STATS_COUNTER(cnt_map_single
);
1783 if (!check_device(dev
))
1784 return DMA_ERROR_CODE
;
1786 dma_mask
= *dev
->dma_mask
;
1788 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1790 if (iommu
== NULL
|| domain
== NULL
)
1791 /* device not handled by any AMD IOMMU */
1792 return (dma_addr_t
)paddr
;
1794 if (!dma_ops_domain(domain
))
1795 return DMA_ERROR_CODE
;
1797 spin_lock_irqsave(&domain
->lock
, flags
);
1798 addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
, size
, dir
, false,
1800 if (addr
== DMA_ERROR_CODE
)
1803 iommu_flush_complete(domain
);
1806 spin_unlock_irqrestore(&domain
->lock
, flags
);
1812 * The exported unmap_single function for dma_ops.
1814 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
1815 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
1817 unsigned long flags
;
1818 struct amd_iommu
*iommu
;
1819 struct protection_domain
*domain
;
1822 INC_STATS_COUNTER(cnt_unmap_single
);
1824 if (!check_device(dev
) ||
1825 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1826 /* device not handled by any AMD IOMMU */
1829 if (!dma_ops_domain(domain
))
1832 spin_lock_irqsave(&domain
->lock
, flags
);
1834 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, dir
);
1836 iommu_flush_complete(domain
);
1838 spin_unlock_irqrestore(&domain
->lock
, flags
);
1842 * This is a special map_sg function which is used if we should map a
1843 * device which is not handled by an AMD IOMMU in the system.
1845 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
1846 int nelems
, int dir
)
1848 struct scatterlist
*s
;
1851 for_each_sg(sglist
, s
, nelems
, i
) {
1852 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
1853 s
->dma_length
= s
->length
;
1860 * The exported map_sg function for dma_ops (handles scatter-gather
1863 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
1864 int nelems
, enum dma_data_direction dir
,
1865 struct dma_attrs
*attrs
)
1867 unsigned long flags
;
1868 struct amd_iommu
*iommu
;
1869 struct protection_domain
*domain
;
1872 struct scatterlist
*s
;
1874 int mapped_elems
= 0;
1877 INC_STATS_COUNTER(cnt_map_sg
);
1879 if (!check_device(dev
))
1882 dma_mask
= *dev
->dma_mask
;
1884 get_device_resources(dev
, &iommu
, &domain
, &devid
);
1886 if (!iommu
|| !domain
)
1887 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
1889 if (!dma_ops_domain(domain
))
1892 spin_lock_irqsave(&domain
->lock
, flags
);
1894 for_each_sg(sglist
, s
, nelems
, i
) {
1897 s
->dma_address
= __map_single(dev
, iommu
, domain
->priv
,
1898 paddr
, s
->length
, dir
, false,
1901 if (s
->dma_address
) {
1902 s
->dma_length
= s
->length
;
1908 iommu_flush_complete(domain
);
1911 spin_unlock_irqrestore(&domain
->lock
, flags
);
1913 return mapped_elems
;
1915 for_each_sg(sglist
, s
, mapped_elems
, i
) {
1917 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1918 s
->dma_length
, dir
);
1919 s
->dma_address
= s
->dma_length
= 0;
1928 * The exported map_sg function for dma_ops (handles scatter-gather
1931 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1932 int nelems
, enum dma_data_direction dir
,
1933 struct dma_attrs
*attrs
)
1935 unsigned long flags
;
1936 struct amd_iommu
*iommu
;
1937 struct protection_domain
*domain
;
1938 struct scatterlist
*s
;
1942 INC_STATS_COUNTER(cnt_unmap_sg
);
1944 if (!check_device(dev
) ||
1945 !get_device_resources(dev
, &iommu
, &domain
, &devid
))
1948 if (!dma_ops_domain(domain
))
1951 spin_lock_irqsave(&domain
->lock
, flags
);
1953 for_each_sg(sglist
, s
, nelems
, i
) {
1954 __unmap_single(iommu
, domain
->priv
, s
->dma_address
,
1955 s
->dma_length
, dir
);
1956 s
->dma_address
= s
->dma_length
= 0;
1959 iommu_flush_complete(domain
);
1961 spin_unlock_irqrestore(&domain
->lock
, flags
);
1965 * The exported alloc_coherent function for dma_ops.
1967 static void *alloc_coherent(struct device
*dev
, size_t size
,
1968 dma_addr_t
*dma_addr
, gfp_t flag
)
1970 unsigned long flags
;
1972 struct amd_iommu
*iommu
;
1973 struct protection_domain
*domain
;
1976 u64 dma_mask
= dev
->coherent_dma_mask
;
1978 INC_STATS_COUNTER(cnt_alloc_coherent
);
1980 if (!check_device(dev
))
1983 if (!get_device_resources(dev
, &iommu
, &domain
, &devid
))
1984 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
1987 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
1991 paddr
= virt_to_phys(virt_addr
);
1993 if (!iommu
|| !domain
) {
1994 *dma_addr
= (dma_addr_t
)paddr
;
1998 if (!dma_ops_domain(domain
))
2002 dma_mask
= *dev
->dma_mask
;
2004 spin_lock_irqsave(&domain
->lock
, flags
);
2006 *dma_addr
= __map_single(dev
, iommu
, domain
->priv
, paddr
,
2007 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2009 if (*dma_addr
== DMA_ERROR_CODE
) {
2010 spin_unlock_irqrestore(&domain
->lock
, flags
);
2014 iommu_flush_complete(domain
);
2016 spin_unlock_irqrestore(&domain
->lock
, flags
);
2022 free_pages((unsigned long)virt_addr
, get_order(size
));
2028 * The exported free_coherent function for dma_ops.
2030 static void free_coherent(struct device
*dev
, size_t size
,
2031 void *virt_addr
, dma_addr_t dma_addr
)
2033 unsigned long flags
;
2034 struct amd_iommu
*iommu
;
2035 struct protection_domain
*domain
;
2038 INC_STATS_COUNTER(cnt_free_coherent
);
2040 if (!check_device(dev
))
2043 get_device_resources(dev
, &iommu
, &domain
, &devid
);
2045 if (!iommu
|| !domain
)
2048 if (!dma_ops_domain(domain
))
2051 spin_lock_irqsave(&domain
->lock
, flags
);
2053 __unmap_single(iommu
, domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2055 iommu_flush_complete(domain
);
2057 spin_unlock_irqrestore(&domain
->lock
, flags
);
2060 free_pages((unsigned long)virt_addr
, get_order(size
));
2064 * This function is called by the DMA layer to find out if we can handle a
2065 * particular device. It is part of the dma_ops.
2067 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2070 struct pci_dev
*pcidev
;
2072 /* No device or no PCI device */
2073 if (!dev
|| dev
->bus
!= &pci_bus_type
)
2076 pcidev
= to_pci_dev(dev
);
2078 bdf
= calc_devid(pcidev
->bus
->number
, pcidev
->devfn
);
2080 /* Out of our scope? */
2081 if (bdf
> amd_iommu_last_bdf
)
2088 * The function for pre-allocating protection domains.
2090 * If the driver core informs the DMA layer if a driver grabs a device
2091 * we don't need to preallocate the protection domains anymore.
2092 * For now we have to.
2094 static void prealloc_protection_domains(void)
2096 struct pci_dev
*dev
= NULL
;
2097 struct dma_ops_domain
*dma_dom
;
2098 struct amd_iommu
*iommu
;
2101 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
2102 __devid
= devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
2103 if (devid
> amd_iommu_last_bdf
)
2105 devid
= amd_iommu_alias_table
[devid
];
2106 if (domain_for_device(devid
))
2108 iommu
= amd_iommu_rlookup_table
[devid
];
2111 dma_dom
= dma_ops_domain_alloc(iommu
);
2114 init_unity_mappings_for_device(dma_dom
, devid
);
2115 dma_dom
->target_dev
= devid
;
2117 attach_device(iommu
, &dma_dom
->domain
, devid
);
2118 if (__devid
!= devid
)
2119 attach_device(iommu
, &dma_dom
->domain
, __devid
);
2121 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2125 static struct dma_map_ops amd_iommu_dma_ops
= {
2126 .alloc_coherent
= alloc_coherent
,
2127 .free_coherent
= free_coherent
,
2128 .map_page
= map_page
,
2129 .unmap_page
= unmap_page
,
2131 .unmap_sg
= unmap_sg
,
2132 .dma_supported
= amd_iommu_dma_supported
,
2136 * The function which clues the AMD IOMMU driver into dma_ops.
2138 int __init
amd_iommu_init_dma_ops(void)
2140 struct amd_iommu
*iommu
;
2144 * first allocate a default protection domain for every IOMMU we
2145 * found in the system. Devices not assigned to any other
2146 * protection domain will be assigned to the default one.
2148 for_each_iommu(iommu
) {
2149 iommu
->default_dom
= dma_ops_domain_alloc(iommu
);
2150 if (iommu
->default_dom
== NULL
)
2152 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2153 ret
= iommu_init_unity_mappings(iommu
);
2159 * If device isolation is enabled, pre-allocate the protection
2160 * domains for each device.
2162 if (amd_iommu_isolate
)
2163 prealloc_protection_domains();
2167 #ifdef CONFIG_GART_IOMMU
2168 gart_iommu_aperture_disabled
= 1;
2169 gart_iommu_aperture
= 0;
2172 /* Make the driver finally visible to the drivers */
2173 dma_ops
= &amd_iommu_dma_ops
;
2175 register_iommu(&amd_iommu_ops
);
2177 bus_register_notifier(&pci_bus_type
, &device_nb
);
2179 amd_iommu_stats_init();
2185 for_each_iommu(iommu
) {
2186 if (iommu
->default_dom
)
2187 dma_ops_domain_free(iommu
->default_dom
);
2193 /*****************************************************************************
2195 * The following functions belong to the exported interface of AMD IOMMU
2197 * This interface allows access to lower level functions of the IOMMU
2198 * like protection domain handling and assignement of devices to domains
2199 * which is not possible with the dma_ops interface.
2201 *****************************************************************************/
2203 static void cleanup_domain(struct protection_domain
*domain
)
2205 unsigned long flags
;
2208 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2210 for (devid
= 0; devid
<= amd_iommu_last_bdf
; ++devid
)
2211 if (amd_iommu_pd_table
[devid
] == domain
)
2212 __detach_device(domain
, devid
);
2214 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2217 static void protection_domain_free(struct protection_domain
*domain
)
2222 del_domain_from_list(domain
);
2225 domain_id_free(domain
->id
);
2230 static struct protection_domain
*protection_domain_alloc(void)
2232 struct protection_domain
*domain
;
2234 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2238 spin_lock_init(&domain
->lock
);
2239 domain
->id
= domain_id_alloc();
2243 add_domain_to_list(domain
);
2253 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2255 struct protection_domain
*domain
;
2257 domain
= protection_domain_alloc();
2261 domain
->mode
= PAGE_MODE_3_LEVEL
;
2262 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2263 if (!domain
->pt_root
)
2271 protection_domain_free(domain
);
2276 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2278 struct protection_domain
*domain
= dom
->priv
;
2283 if (domain
->dev_cnt
> 0)
2284 cleanup_domain(domain
);
2286 BUG_ON(domain
->dev_cnt
!= 0);
2288 free_pagetable(domain
);
2290 domain_id_free(domain
->id
);
2297 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2300 struct protection_domain
*domain
= dom
->priv
;
2301 struct amd_iommu
*iommu
;
2302 struct pci_dev
*pdev
;
2305 if (dev
->bus
!= &pci_bus_type
)
2308 pdev
= to_pci_dev(dev
);
2310 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2313 detach_device(domain
, devid
);
2315 iommu
= amd_iommu_rlookup_table
[devid
];
2319 iommu_queue_inv_dev_entry(iommu
, devid
);
2320 iommu_completion_wait(iommu
);
2323 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2326 struct protection_domain
*domain
= dom
->priv
;
2327 struct protection_domain
*old_domain
;
2328 struct amd_iommu
*iommu
;
2329 struct pci_dev
*pdev
;
2332 if (dev
->bus
!= &pci_bus_type
)
2335 pdev
= to_pci_dev(dev
);
2337 devid
= calc_devid(pdev
->bus
->number
, pdev
->devfn
);
2339 if (devid
>= amd_iommu_last_bdf
||
2340 devid
!= amd_iommu_alias_table
[devid
])
2343 iommu
= amd_iommu_rlookup_table
[devid
];
2347 old_domain
= domain_for_device(devid
);
2349 detach_device(old_domain
, devid
);
2351 attach_device(iommu
, domain
, devid
);
2353 iommu_completion_wait(iommu
);
2358 static int amd_iommu_map_range(struct iommu_domain
*dom
,
2359 unsigned long iova
, phys_addr_t paddr
,
2360 size_t size
, int iommu_prot
)
2362 struct protection_domain
*domain
= dom
->priv
;
2363 unsigned long i
, npages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2367 if (iommu_prot
& IOMMU_READ
)
2368 prot
|= IOMMU_PROT_IR
;
2369 if (iommu_prot
& IOMMU_WRITE
)
2370 prot
|= IOMMU_PROT_IW
;
2375 for (i
= 0; i
< npages
; ++i
) {
2376 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, PM_MAP_4k
);
2387 static void amd_iommu_unmap_range(struct iommu_domain
*dom
,
2388 unsigned long iova
, size_t size
)
2391 struct protection_domain
*domain
= dom
->priv
;
2392 unsigned long i
, npages
= iommu_num_pages(iova
, size
, PAGE_SIZE
);
2396 for (i
= 0; i
< npages
; ++i
) {
2397 iommu_unmap_page(domain
, iova
, PM_MAP_4k
);
2401 iommu_flush_tlb_pde(domain
);
2404 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2407 struct protection_domain
*domain
= dom
->priv
;
2408 unsigned long offset
= iova
& ~PAGE_MASK
;
2412 pte
= fetch_pte(domain
, iova
, PM_MAP_4k
);
2414 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
2417 paddr
= *pte
& IOMMU_PAGE_MASK
;
2423 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
2429 static struct iommu_ops amd_iommu_ops
= {
2430 .domain_init
= amd_iommu_domain_init
,
2431 .domain_destroy
= amd_iommu_domain_destroy
,
2432 .attach_dev
= amd_iommu_attach_device
,
2433 .detach_dev
= amd_iommu_detach_device
,
2434 .map
= amd_iommu_map_range
,
2435 .unmap
= amd_iommu_unmap_range
,
2436 .iova_to_phys
= amd_iommu_iova_to_phys
,
2437 .domain_has_cap
= amd_iommu_domain_has_cap
,
2440 /*****************************************************************************
2442 * The next functions do a basic initialization of IOMMU for pass through
2445 * In passthrough mode the IOMMU is initialized and enabled but not used for
2446 * DMA-API translation.
2448 *****************************************************************************/
2450 int __init
amd_iommu_init_passthrough(void)
2452 struct pci_dev
*dev
= NULL
;
2455 /* allocate passthroug domain */
2456 pt_domain
= protection_domain_alloc();
2460 pt_domain
->mode
|= PAGE_MODE_NONE
;
2462 while ((dev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, dev
)) != NULL
) {
2463 struct amd_iommu
*iommu
;
2465 devid
= calc_devid(dev
->bus
->number
, dev
->devfn
);
2466 if (devid
> amd_iommu_last_bdf
)
2469 devid2
= amd_iommu_alias_table
[devid
];
2471 iommu
= amd_iommu_rlookup_table
[devid2
];
2475 __attach_device(iommu
, pt_domain
, devid
);
2476 __attach_device(iommu
, pt_domain
, devid2
);
2479 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");