2 * Copyright (C) 2007-2010 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/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <asm/msidef.h>
35 #include <asm/proto.h>
36 #include <asm/iommu.h>
40 #include "amd_iommu_proto.h"
41 #include "amd_iommu_types.h"
43 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
45 #define LOOP_TIMEOUT 100000
47 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
49 /* A list of preallocated protection domains */
50 static LIST_HEAD(iommu_pd_list
);
51 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
53 /* List of all available dev_data structures */
54 static LIST_HEAD(dev_data_list
);
55 static DEFINE_SPINLOCK(dev_data_list_lock
);
58 * Domain for untranslated devices - only allocated
59 * if iommu=pt passed on kernel cmd line.
61 static struct protection_domain
*pt_domain
;
63 static struct iommu_ops amd_iommu_ops
;
65 static ATOMIC_NOTIFIER_HEAD(ppr_notifier
);
66 int amd_iommu_max_glx_val
= -1;
69 * general struct to manage commands send to an IOMMU
75 static void update_domain(struct protection_domain
*domain
);
76 static int __init
alloc_passthrough_domain(void);
78 /****************************************************************************
82 ****************************************************************************/
84 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
86 struct iommu_dev_data
*dev_data
;
89 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
93 dev_data
->devid
= devid
;
94 atomic_set(&dev_data
->bind
, 0);
96 spin_lock_irqsave(&dev_data_list_lock
, flags
);
97 list_add_tail(&dev_data
->dev_data_list
, &dev_data_list
);
98 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
103 static void free_dev_data(struct iommu_dev_data
*dev_data
)
107 spin_lock_irqsave(&dev_data_list_lock
, flags
);
108 list_del(&dev_data
->dev_data_list
);
109 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
114 static struct iommu_dev_data
*search_dev_data(u16 devid
)
116 struct iommu_dev_data
*dev_data
;
119 spin_lock_irqsave(&dev_data_list_lock
, flags
);
120 list_for_each_entry(dev_data
, &dev_data_list
, dev_data_list
) {
121 if (dev_data
->devid
== devid
)
128 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
133 static struct iommu_dev_data
*find_dev_data(u16 devid
)
135 struct iommu_dev_data
*dev_data
;
137 dev_data
= search_dev_data(devid
);
139 if (dev_data
== NULL
)
140 dev_data
= alloc_dev_data(devid
);
145 static inline u16
get_device_id(struct device
*dev
)
147 struct pci_dev
*pdev
= to_pci_dev(dev
);
149 return calc_devid(pdev
->bus
->number
, pdev
->devfn
);
152 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
154 return dev
->archdata
.iommu
;
157 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
159 static const int caps
[] = {
166 for (i
= 0; i
< 3; ++i
) {
167 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
176 * In this function the list of preallocated protection domains is traversed to
177 * find the domain for a specific device
179 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
181 struct dma_ops_domain
*entry
, *ret
= NULL
;
183 u16 alias
= amd_iommu_alias_table
[devid
];
185 if (list_empty(&iommu_pd_list
))
188 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
190 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
191 if (entry
->target_dev
== devid
||
192 entry
->target_dev
== alias
) {
198 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
204 * This function checks if the driver got a valid device from the caller to
205 * avoid dereferencing invalid pointers.
207 static bool check_device(struct device
*dev
)
211 if (!dev
|| !dev
->dma_mask
)
214 /* No device or no PCI device */
215 if (dev
->bus
!= &pci_bus_type
)
218 devid
= get_device_id(dev
);
220 /* Out of our scope? */
221 if (devid
> amd_iommu_last_bdf
)
224 if (amd_iommu_rlookup_table
[devid
] == NULL
)
230 static int iommu_init_device(struct device
*dev
)
232 struct pci_dev
*pdev
= to_pci_dev(dev
);
233 struct iommu_dev_data
*dev_data
;
236 if (dev
->archdata
.iommu
)
239 dev_data
= find_dev_data(get_device_id(dev
));
243 alias
= amd_iommu_alias_table
[dev_data
->devid
];
244 if (alias
!= dev_data
->devid
) {
245 struct iommu_dev_data
*alias_data
;
247 alias_data
= find_dev_data(alias
);
248 if (alias_data
== NULL
) {
249 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
251 free_dev_data(dev_data
);
254 dev_data
->alias_data
= alias_data
;
257 if (pci_iommuv2_capable(pdev
)) {
258 struct amd_iommu
*iommu
;
260 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
261 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
264 dev
->archdata
.iommu
= dev_data
;
269 static void iommu_ignore_device(struct device
*dev
)
273 devid
= get_device_id(dev
);
274 alias
= amd_iommu_alias_table
[devid
];
276 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
277 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
279 amd_iommu_rlookup_table
[devid
] = NULL
;
280 amd_iommu_rlookup_table
[alias
] = NULL
;
283 static void iommu_uninit_device(struct device
*dev
)
286 * Nothing to do here - we keep dev_data around for unplugged devices
287 * and reuse it when the device is re-plugged - not doing so would
288 * introduce a ton of races.
292 void __init
amd_iommu_uninit_devices(void)
294 struct iommu_dev_data
*dev_data
, *n
;
295 struct pci_dev
*pdev
= NULL
;
297 for_each_pci_dev(pdev
) {
299 if (!check_device(&pdev
->dev
))
302 iommu_uninit_device(&pdev
->dev
);
305 /* Free all of our dev_data structures */
306 list_for_each_entry_safe(dev_data
, n
, &dev_data_list
, dev_data_list
)
307 free_dev_data(dev_data
);
310 int __init
amd_iommu_init_devices(void)
312 struct pci_dev
*pdev
= NULL
;
315 for_each_pci_dev(pdev
) {
317 if (!check_device(&pdev
->dev
))
320 ret
= iommu_init_device(&pdev
->dev
);
321 if (ret
== -ENOTSUPP
)
322 iommu_ignore_device(&pdev
->dev
);
331 amd_iommu_uninit_devices();
335 #ifdef CONFIG_AMD_IOMMU_STATS
338 * Initialization code for statistics collection
341 DECLARE_STATS_COUNTER(compl_wait
);
342 DECLARE_STATS_COUNTER(cnt_map_single
);
343 DECLARE_STATS_COUNTER(cnt_unmap_single
);
344 DECLARE_STATS_COUNTER(cnt_map_sg
);
345 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
346 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
347 DECLARE_STATS_COUNTER(cnt_free_coherent
);
348 DECLARE_STATS_COUNTER(cross_page
);
349 DECLARE_STATS_COUNTER(domain_flush_single
);
350 DECLARE_STATS_COUNTER(domain_flush_all
);
351 DECLARE_STATS_COUNTER(alloced_io_mem
);
352 DECLARE_STATS_COUNTER(total_map_requests
);
354 static struct dentry
*stats_dir
;
355 static struct dentry
*de_fflush
;
357 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
359 if (stats_dir
== NULL
)
362 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
366 static void amd_iommu_stats_init(void)
368 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
369 if (stats_dir
== NULL
)
372 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
373 (u32
*)&amd_iommu_unmap_flush
);
375 amd_iommu_stats_add(&compl_wait
);
376 amd_iommu_stats_add(&cnt_map_single
);
377 amd_iommu_stats_add(&cnt_unmap_single
);
378 amd_iommu_stats_add(&cnt_map_sg
);
379 amd_iommu_stats_add(&cnt_unmap_sg
);
380 amd_iommu_stats_add(&cnt_alloc_coherent
);
381 amd_iommu_stats_add(&cnt_free_coherent
);
382 amd_iommu_stats_add(&cross_page
);
383 amd_iommu_stats_add(&domain_flush_single
);
384 amd_iommu_stats_add(&domain_flush_all
);
385 amd_iommu_stats_add(&alloced_io_mem
);
386 amd_iommu_stats_add(&total_map_requests
);
391 /****************************************************************************
393 * Interrupt handling functions
395 ****************************************************************************/
397 static void dump_dte_entry(u16 devid
)
401 for (i
= 0; i
< 4; ++i
)
402 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
403 amd_iommu_dev_table
[devid
].data
[i
]);
406 static void dump_command(unsigned long phys_addr
)
408 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
411 for (i
= 0; i
< 4; ++i
)
412 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
415 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
418 int type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
419 int devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
420 int domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
421 int flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
422 u64 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
424 printk(KERN_ERR
"AMD-Vi: Event logged [");
427 case EVENT_TYPE_ILL_DEV
:
428 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
429 "address=0x%016llx flags=0x%04x]\n",
430 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
432 dump_dte_entry(devid
);
434 case EVENT_TYPE_IO_FAULT
:
435 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
436 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
437 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
438 domid
, address
, flags
);
440 case EVENT_TYPE_DEV_TAB_ERR
:
441 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
442 "address=0x%016llx flags=0x%04x]\n",
443 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
446 case EVENT_TYPE_PAGE_TAB_ERR
:
447 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
448 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
449 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
450 domid
, address
, flags
);
452 case EVENT_TYPE_ILL_CMD
:
453 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
454 dump_command(address
);
456 case EVENT_TYPE_CMD_HARD_ERR
:
457 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
458 "flags=0x%04x]\n", address
, flags
);
460 case EVENT_TYPE_IOTLB_INV_TO
:
461 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
462 "address=0x%016llx]\n",
463 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
466 case EVENT_TYPE_INV_DEV_REQ
:
467 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
468 "address=0x%016llx flags=0x%04x]\n",
469 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
473 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
477 static void iommu_poll_events(struct amd_iommu
*iommu
)
482 spin_lock_irqsave(&iommu
->lock
, flags
);
484 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
485 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
487 while (head
!= tail
) {
488 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
489 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
492 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
494 spin_unlock_irqrestore(&iommu
->lock
, flags
);
497 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u32 head
)
499 struct amd_iommu_fault fault
;
503 raw
= (u64
*)(iommu
->ppr_log
+ head
);
506 * Hardware bug: Interrupt may arrive before the entry is written to
507 * memory. If this happens we need to wait for the entry to arrive.
509 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
510 if (PPR_REQ_TYPE(raw
[0]) != 0)
515 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
516 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
520 fault
.address
= raw
[1];
521 fault
.pasid
= PPR_PASID(raw
[0]);
522 fault
.device_id
= PPR_DEVID(raw
[0]);
523 fault
.tag
= PPR_TAG(raw
[0]);
524 fault
.flags
= PPR_FLAGS(raw
[0]);
527 * To detect the hardware bug we need to clear the entry
532 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
535 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
540 if (iommu
->ppr_log
== NULL
)
543 spin_lock_irqsave(&iommu
->lock
, flags
);
545 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
546 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
548 while (head
!= tail
) {
550 /* Handle PPR entry */
551 iommu_handle_ppr_entry(iommu
, head
);
553 /* Update and refresh ring-buffer state*/
554 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
555 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
556 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
559 /* enable ppr interrupts again */
560 writel(MMIO_STATUS_PPR_INT_MASK
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
562 spin_unlock_irqrestore(&iommu
->lock
, flags
);
565 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
567 struct amd_iommu
*iommu
;
569 for_each_iommu(iommu
) {
570 iommu_poll_events(iommu
);
571 iommu_poll_ppr_log(iommu
);
577 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
579 return IRQ_WAKE_THREAD
;
582 /****************************************************************************
584 * IOMMU command queuing functions
586 ****************************************************************************/
588 static int wait_on_sem(volatile u64
*sem
)
592 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
597 if (i
== LOOP_TIMEOUT
) {
598 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
605 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
606 struct iommu_cmd
*cmd
,
611 target
= iommu
->cmd_buf
+ tail
;
612 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
614 /* Copy command to buffer */
615 memcpy(target
, cmd
, sizeof(*cmd
));
617 /* Tell the IOMMU about it */
618 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
621 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
623 WARN_ON(address
& 0x7ULL
);
625 memset(cmd
, 0, sizeof(*cmd
));
626 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
627 cmd
->data
[1] = upper_32_bits(__pa(address
));
629 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
632 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
634 memset(cmd
, 0, sizeof(*cmd
));
635 cmd
->data
[0] = devid
;
636 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
639 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
640 size_t size
, u16 domid
, int pde
)
645 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
650 * If we have to flush more than one page, flush all
651 * TLB entries for this domain
653 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
657 address
&= PAGE_MASK
;
659 memset(cmd
, 0, sizeof(*cmd
));
660 cmd
->data
[1] |= domid
;
661 cmd
->data
[2] = lower_32_bits(address
);
662 cmd
->data
[3] = upper_32_bits(address
);
663 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
664 if (s
) /* size bit - we flush more than one 4kb page */
665 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
666 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
667 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
670 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
671 u64 address
, size_t size
)
676 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
681 * If we have to flush more than one page, flush all
682 * TLB entries for this domain
684 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
688 address
&= PAGE_MASK
;
690 memset(cmd
, 0, sizeof(*cmd
));
691 cmd
->data
[0] = devid
;
692 cmd
->data
[0] |= (qdep
& 0xff) << 24;
693 cmd
->data
[1] = devid
;
694 cmd
->data
[2] = lower_32_bits(address
);
695 cmd
->data
[3] = upper_32_bits(address
);
696 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
698 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
701 static void build_inv_all(struct iommu_cmd
*cmd
)
703 memset(cmd
, 0, sizeof(*cmd
));
704 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
708 * Writes the command to the IOMMUs command buffer and informs the
709 * hardware about the new command.
711 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
712 struct iommu_cmd
*cmd
,
715 u32 left
, tail
, head
, next_tail
;
718 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
721 spin_lock_irqsave(&iommu
->lock
, flags
);
723 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
724 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
725 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
726 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
729 struct iommu_cmd sync_cmd
;
730 volatile u64 sem
= 0;
733 build_completion_wait(&sync_cmd
, (u64
)&sem
);
734 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
736 spin_unlock_irqrestore(&iommu
->lock
, flags
);
738 if ((ret
= wait_on_sem(&sem
)) != 0)
744 copy_cmd_to_buffer(iommu
, cmd
, tail
);
746 /* We need to sync now to make sure all commands are processed */
747 iommu
->need_sync
= sync
;
749 spin_unlock_irqrestore(&iommu
->lock
, flags
);
754 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
756 return iommu_queue_command_sync(iommu
, cmd
, true);
760 * This function queues a completion wait command into the command
763 static int iommu_completion_wait(struct amd_iommu
*iommu
)
765 struct iommu_cmd cmd
;
766 volatile u64 sem
= 0;
769 if (!iommu
->need_sync
)
772 build_completion_wait(&cmd
, (u64
)&sem
);
774 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
778 return wait_on_sem(&sem
);
781 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
783 struct iommu_cmd cmd
;
785 build_inv_dte(&cmd
, devid
);
787 return iommu_queue_command(iommu
, &cmd
);
790 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
794 for (devid
= 0; devid
<= 0xffff; ++devid
)
795 iommu_flush_dte(iommu
, devid
);
797 iommu_completion_wait(iommu
);
801 * This function uses heavy locking and may disable irqs for some time. But
802 * this is no issue because it is only called during resume.
804 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
808 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
809 struct iommu_cmd cmd
;
810 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
812 iommu_queue_command(iommu
, &cmd
);
815 iommu_completion_wait(iommu
);
818 static void iommu_flush_all(struct amd_iommu
*iommu
)
820 struct iommu_cmd cmd
;
824 iommu_queue_command(iommu
, &cmd
);
825 iommu_completion_wait(iommu
);
828 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
830 if (iommu_feature(iommu
, FEATURE_IA
)) {
831 iommu_flush_all(iommu
);
833 iommu_flush_dte_all(iommu
);
834 iommu_flush_tlb_all(iommu
);
839 * Command send function for flushing on-device TLB
841 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
842 u64 address
, size_t size
)
844 struct amd_iommu
*iommu
;
845 struct iommu_cmd cmd
;
848 qdep
= dev_data
->ats
.qdep
;
849 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
851 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
853 return iommu_queue_command(iommu
, &cmd
);
857 * Command send function for invalidating a device table entry
859 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
861 struct amd_iommu
*iommu
;
864 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
866 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
870 if (dev_data
->ats
.enabled
)
871 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
877 * TLB invalidation function which is called from the mapping functions.
878 * It invalidates a single PTE if the range to flush is within a single
879 * page. Otherwise it flushes the whole TLB of the IOMMU.
881 static void __domain_flush_pages(struct protection_domain
*domain
,
882 u64 address
, size_t size
, int pde
)
884 struct iommu_dev_data
*dev_data
;
885 struct iommu_cmd cmd
;
888 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
890 for (i
= 0; i
< amd_iommus_present
; ++i
) {
891 if (!domain
->dev_iommu
[i
])
895 * Devices of this domain are behind this IOMMU
896 * We need a TLB flush
898 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
901 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
903 if (!dev_data
->ats
.enabled
)
906 ret
|= device_flush_iotlb(dev_data
, address
, size
);
912 static void domain_flush_pages(struct protection_domain
*domain
,
913 u64 address
, size_t size
)
915 __domain_flush_pages(domain
, address
, size
, 0);
918 /* Flush the whole IO/TLB for a given protection domain */
919 static void domain_flush_tlb(struct protection_domain
*domain
)
921 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
924 /* Flush the whole IO/TLB for a given protection domain - including PDE */
925 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
927 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
930 static void domain_flush_complete(struct protection_domain
*domain
)
934 for (i
= 0; i
< amd_iommus_present
; ++i
) {
935 if (!domain
->dev_iommu
[i
])
939 * Devices of this domain are behind this IOMMU
940 * We need to wait for completion of all commands.
942 iommu_completion_wait(amd_iommus
[i
]);
948 * This function flushes the DTEs for all devices in domain
950 static void domain_flush_devices(struct protection_domain
*domain
)
952 struct iommu_dev_data
*dev_data
;
954 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
955 device_flush_dte(dev_data
);
958 /****************************************************************************
960 * The functions below are used the create the page table mappings for
961 * unity mapped regions.
963 ****************************************************************************/
966 * This function is used to add another level to an IO page table. Adding
967 * another level increases the size of the address space by 9 bits to a size up
970 static bool increase_address_space(struct protection_domain
*domain
,
975 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
976 /* address space already 64 bit large */
979 pte
= (void *)get_zeroed_page(gfp
);
983 *pte
= PM_LEVEL_PDE(domain
->mode
,
984 virt_to_phys(domain
->pt_root
));
985 domain
->pt_root
= pte
;
987 domain
->updated
= true;
992 static u64
*alloc_pte(struct protection_domain
*domain
,
993 unsigned long address
,
994 unsigned long page_size
,
1001 BUG_ON(!is_power_of_2(page_size
));
1003 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1004 increase_address_space(domain
, gfp
);
1006 level
= domain
->mode
- 1;
1007 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1008 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1009 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1011 while (level
> end_lvl
) {
1012 if (!IOMMU_PTE_PRESENT(*pte
)) {
1013 page
= (u64
*)get_zeroed_page(gfp
);
1016 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1019 /* No level skipping support yet */
1020 if (PM_PTE_LEVEL(*pte
) != level
)
1025 pte
= IOMMU_PTE_PAGE(*pte
);
1027 if (pte_page
&& level
== end_lvl
)
1030 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1037 * This function checks if there is a PTE for a given dma address. If
1038 * there is one, it returns the pointer to it.
1040 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
1045 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1048 level
= domain
->mode
- 1;
1049 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1054 if (!IOMMU_PTE_PRESENT(*pte
))
1058 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1059 unsigned long pte_mask
, __pte
;
1062 * If we have a series of large PTEs, make
1063 * sure to return a pointer to the first one.
1065 pte_mask
= PTE_PAGE_SIZE(*pte
);
1066 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1067 __pte
= ((unsigned long)pte
) & pte_mask
;
1069 return (u64
*)__pte
;
1072 /* No level skipping support yet */
1073 if (PM_PTE_LEVEL(*pte
) != level
)
1078 /* Walk to the next level */
1079 pte
= IOMMU_PTE_PAGE(*pte
);
1080 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1087 * Generic mapping functions. It maps a physical address into a DMA
1088 * address space. It allocates the page table pages if necessary.
1089 * In the future it can be extended to a generic mapping function
1090 * supporting all features of AMD IOMMU page tables like level skipping
1091 * and full 64 bit address spaces.
1093 static int iommu_map_page(struct protection_domain
*dom
,
1094 unsigned long bus_addr
,
1095 unsigned long phys_addr
,
1097 unsigned long page_size
)
1102 if (!(prot
& IOMMU_PROT_MASK
))
1105 bus_addr
= PAGE_ALIGN(bus_addr
);
1106 phys_addr
= PAGE_ALIGN(phys_addr
);
1107 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1108 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1110 for (i
= 0; i
< count
; ++i
)
1111 if (IOMMU_PTE_PRESENT(pte
[i
]))
1114 if (page_size
> PAGE_SIZE
) {
1115 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1116 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1118 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1120 if (prot
& IOMMU_PROT_IR
)
1121 __pte
|= IOMMU_PTE_IR
;
1122 if (prot
& IOMMU_PROT_IW
)
1123 __pte
|= IOMMU_PTE_IW
;
1125 for (i
= 0; i
< count
; ++i
)
1133 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1134 unsigned long bus_addr
,
1135 unsigned long page_size
)
1137 unsigned long long unmap_size
, unmapped
;
1140 BUG_ON(!is_power_of_2(page_size
));
1144 while (unmapped
< page_size
) {
1146 pte
= fetch_pte(dom
, bus_addr
);
1150 * No PTE for this address
1151 * move forward in 4kb steps
1153 unmap_size
= PAGE_SIZE
;
1154 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1155 /* 4kb PTE found for this address */
1156 unmap_size
= PAGE_SIZE
;
1161 /* Large PTE found which maps this address */
1162 unmap_size
= PTE_PAGE_SIZE(*pte
);
1163 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1164 for (i
= 0; i
< count
; i
++)
1168 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1169 unmapped
+= unmap_size
;
1172 BUG_ON(!is_power_of_2(unmapped
));
1178 * This function checks if a specific unity mapping entry is needed for
1179 * this specific IOMMU.
1181 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1182 struct unity_map_entry
*entry
)
1186 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1187 bdf
= amd_iommu_alias_table
[i
];
1188 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1196 * This function actually applies the mapping to the page table of the
1199 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1200 struct unity_map_entry
*e
)
1205 for (addr
= e
->address_start
; addr
< e
->address_end
;
1206 addr
+= PAGE_SIZE
) {
1207 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1212 * if unity mapping is in aperture range mark the page
1213 * as allocated in the aperture
1215 if (addr
< dma_dom
->aperture_size
)
1216 __set_bit(addr
>> PAGE_SHIFT
,
1217 dma_dom
->aperture
[0]->bitmap
);
1224 * Init the unity mappings for a specific IOMMU in the system
1226 * Basically iterates over all unity mapping entries and applies them to
1227 * the default domain DMA of that IOMMU if necessary.
1229 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1231 struct unity_map_entry
*entry
;
1234 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1235 if (!iommu_for_unity_map(iommu
, entry
))
1237 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1246 * Inits the unity mappings required for a specific device
1248 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1251 struct unity_map_entry
*e
;
1254 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1255 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1257 ret
= dma_ops_unity_map(dma_dom
, e
);
1265 /****************************************************************************
1267 * The next functions belong to the address allocator for the dma_ops
1268 * interface functions. They work like the allocators in the other IOMMU
1269 * drivers. Its basically a bitmap which marks the allocated pages in
1270 * the aperture. Maybe it could be enhanced in the future to a more
1271 * efficient allocator.
1273 ****************************************************************************/
1276 * The address allocator core functions.
1278 * called with domain->lock held
1282 * Used to reserve address ranges in the aperture (e.g. for exclusion
1285 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1286 unsigned long start_page
,
1289 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1291 if (start_page
+ pages
> last_page
)
1292 pages
= last_page
- start_page
;
1294 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1295 int index
= i
/ APERTURE_RANGE_PAGES
;
1296 int page
= i
% APERTURE_RANGE_PAGES
;
1297 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1302 * This function is used to add a new aperture range to an existing
1303 * aperture in case of dma_ops domain allocation or address allocation
1306 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1307 bool populate
, gfp_t gfp
)
1309 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1310 struct amd_iommu
*iommu
;
1311 unsigned long i
, old_size
;
1313 #ifdef CONFIG_IOMMU_STRESS
1317 if (index
>= APERTURE_MAX_RANGES
)
1320 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1321 if (!dma_dom
->aperture
[index
])
1324 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1325 if (!dma_dom
->aperture
[index
]->bitmap
)
1328 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1331 unsigned long address
= dma_dom
->aperture_size
;
1332 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1333 u64
*pte
, *pte_page
;
1335 for (i
= 0; i
< num_ptes
; ++i
) {
1336 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1341 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1343 address
+= APERTURE_RANGE_SIZE
/ 64;
1347 old_size
= dma_dom
->aperture_size
;
1348 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1350 /* Reserve address range used for MSI messages */
1351 if (old_size
< MSI_ADDR_BASE_LO
&&
1352 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1353 unsigned long spage
;
1356 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1357 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1359 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1362 /* Initialize the exclusion range if necessary */
1363 for_each_iommu(iommu
) {
1364 if (iommu
->exclusion_start
&&
1365 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1366 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1367 unsigned long startpage
;
1368 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1369 iommu
->exclusion_length
,
1371 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1372 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1377 * Check for areas already mapped as present in the new aperture
1378 * range and mark those pages as reserved in the allocator. Such
1379 * mappings may already exist as a result of requested unity
1380 * mappings for devices.
1382 for (i
= dma_dom
->aperture
[index
]->offset
;
1383 i
< dma_dom
->aperture_size
;
1385 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1386 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1389 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1392 update_domain(&dma_dom
->domain
);
1397 update_domain(&dma_dom
->domain
);
1399 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1401 kfree(dma_dom
->aperture
[index
]);
1402 dma_dom
->aperture
[index
] = NULL
;
1407 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1408 struct dma_ops_domain
*dom
,
1410 unsigned long align_mask
,
1412 unsigned long start
)
1414 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1415 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1416 int i
= start
>> APERTURE_RANGE_SHIFT
;
1417 unsigned long boundary_size
;
1418 unsigned long address
= -1;
1419 unsigned long limit
;
1421 next_bit
>>= PAGE_SHIFT
;
1423 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1424 PAGE_SIZE
) >> PAGE_SHIFT
;
1426 for (;i
< max_index
; ++i
) {
1427 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1429 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1432 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1433 dma_mask
>> PAGE_SHIFT
);
1435 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1436 limit
, next_bit
, pages
, 0,
1437 boundary_size
, align_mask
);
1438 if (address
!= -1) {
1439 address
= dom
->aperture
[i
]->offset
+
1440 (address
<< PAGE_SHIFT
);
1441 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1451 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1452 struct dma_ops_domain
*dom
,
1454 unsigned long align_mask
,
1457 unsigned long address
;
1459 #ifdef CONFIG_IOMMU_STRESS
1460 dom
->next_address
= 0;
1461 dom
->need_flush
= true;
1464 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1465 dma_mask
, dom
->next_address
);
1467 if (address
== -1) {
1468 dom
->next_address
= 0;
1469 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1471 dom
->need_flush
= true;
1474 if (unlikely(address
== -1))
1475 address
= DMA_ERROR_CODE
;
1477 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1483 * The address free function.
1485 * called with domain->lock held
1487 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1488 unsigned long address
,
1491 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1492 struct aperture_range
*range
= dom
->aperture
[i
];
1494 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1496 #ifdef CONFIG_IOMMU_STRESS
1501 if (address
>= dom
->next_address
)
1502 dom
->need_flush
= true;
1504 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1506 bitmap_clear(range
->bitmap
, address
, pages
);
1510 /****************************************************************************
1512 * The next functions belong to the domain allocation. A domain is
1513 * allocated for every IOMMU as the default domain. If device isolation
1514 * is enabled, every device get its own domain. The most important thing
1515 * about domains is the page table mapping the DMA address space they
1518 ****************************************************************************/
1521 * This function adds a protection domain to the global protection domain list
1523 static void add_domain_to_list(struct protection_domain
*domain
)
1525 unsigned long flags
;
1527 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1528 list_add(&domain
->list
, &amd_iommu_pd_list
);
1529 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1533 * This function removes a protection domain to the global
1534 * protection domain list
1536 static void del_domain_from_list(struct protection_domain
*domain
)
1538 unsigned long flags
;
1540 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1541 list_del(&domain
->list
);
1542 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1545 static u16
domain_id_alloc(void)
1547 unsigned long flags
;
1550 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1551 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1553 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1554 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1557 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1562 static void domain_id_free(int id
)
1564 unsigned long flags
;
1566 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1567 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1568 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1569 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1572 static void free_pagetable(struct protection_domain
*domain
)
1577 p1
= domain
->pt_root
;
1582 for (i
= 0; i
< 512; ++i
) {
1583 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1586 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1587 for (j
= 0; j
< 512; ++j
) {
1588 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1590 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1591 free_page((unsigned long)p3
);
1594 free_page((unsigned long)p2
);
1597 free_page((unsigned long)p1
);
1599 domain
->pt_root
= NULL
;
1602 static void free_gcr3_table(struct protection_domain
*domain
)
1604 free_page((unsigned long)domain
->gcr3_tbl
);
1608 * Free a domain, only used if something went wrong in the
1609 * allocation path and we need to free an already allocated page table
1611 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1618 del_domain_from_list(&dom
->domain
);
1620 free_pagetable(&dom
->domain
);
1622 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1623 if (!dom
->aperture
[i
])
1625 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1626 kfree(dom
->aperture
[i
]);
1633 * Allocates a new protection domain usable for the dma_ops functions.
1634 * It also initializes the page table and the address allocator data
1635 * structures required for the dma_ops interface
1637 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1639 struct dma_ops_domain
*dma_dom
;
1641 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1645 spin_lock_init(&dma_dom
->domain
.lock
);
1647 dma_dom
->domain
.id
= domain_id_alloc();
1648 if (dma_dom
->domain
.id
== 0)
1650 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1651 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1652 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1653 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1654 dma_dom
->domain
.priv
= dma_dom
;
1655 if (!dma_dom
->domain
.pt_root
)
1658 dma_dom
->need_flush
= false;
1659 dma_dom
->target_dev
= 0xffff;
1661 add_domain_to_list(&dma_dom
->domain
);
1663 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1667 * mark the first page as allocated so we never return 0 as
1668 * a valid dma-address. So we can use 0 as error value
1670 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1671 dma_dom
->next_address
= 0;
1677 dma_ops_domain_free(dma_dom
);
1683 * little helper function to check whether a given protection domain is a
1686 static bool dma_ops_domain(struct protection_domain
*domain
)
1688 return domain
->flags
& PD_DMA_OPS_MASK
;
1691 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1696 if (domain
->mode
!= PAGE_MODE_NONE
)
1697 pte_root
= virt_to_phys(domain
->pt_root
);
1699 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1700 << DEV_ENTRY_MODE_SHIFT
;
1701 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1703 flags
= amd_iommu_dev_table
[devid
].data
[1];
1706 flags
|= DTE_FLAG_IOTLB
;
1708 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1709 u64 gcr3
= __pa(domain
->gcr3_tbl
);
1710 u64 glx
= domain
->glx
;
1713 pte_root
|= DTE_FLAG_GV
;
1714 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1716 /* First mask out possible old values for GCR3 table */
1717 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1720 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1723 /* Encode GCR3 table into DTE */
1724 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1727 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1730 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1734 flags
&= ~(0xffffUL
);
1735 flags
|= domain
->id
;
1737 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1738 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1741 static void clear_dte_entry(u16 devid
)
1743 /* remove entry from the device table seen by the hardware */
1744 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1745 amd_iommu_dev_table
[devid
].data
[1] = 0;
1747 amd_iommu_apply_erratum_63(devid
);
1750 static void do_attach(struct iommu_dev_data
*dev_data
,
1751 struct protection_domain
*domain
)
1753 struct amd_iommu
*iommu
;
1756 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1757 ats
= dev_data
->ats
.enabled
;
1759 /* Update data structures */
1760 dev_data
->domain
= domain
;
1761 list_add(&dev_data
->list
, &domain
->dev_list
);
1762 set_dte_entry(dev_data
->devid
, domain
, ats
);
1764 /* Do reference counting */
1765 domain
->dev_iommu
[iommu
->index
] += 1;
1766 domain
->dev_cnt
+= 1;
1768 /* Flush the DTE entry */
1769 device_flush_dte(dev_data
);
1772 static void do_detach(struct iommu_dev_data
*dev_data
)
1774 struct amd_iommu
*iommu
;
1776 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1778 /* decrease reference counters */
1779 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1780 dev_data
->domain
->dev_cnt
-= 1;
1782 /* Update data structures */
1783 dev_data
->domain
= NULL
;
1784 list_del(&dev_data
->list
);
1785 clear_dte_entry(dev_data
->devid
);
1787 /* Flush the DTE entry */
1788 device_flush_dte(dev_data
);
1792 * If a device is not yet associated with a domain, this function does
1793 * assigns it visible for the hardware
1795 static int __attach_device(struct iommu_dev_data
*dev_data
,
1796 struct protection_domain
*domain
)
1801 spin_lock(&domain
->lock
);
1803 if (dev_data
->alias_data
!= NULL
) {
1804 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1806 /* Some sanity checks */
1808 if (alias_data
->domain
!= NULL
&&
1809 alias_data
->domain
!= domain
)
1812 if (dev_data
->domain
!= NULL
&&
1813 dev_data
->domain
!= domain
)
1816 /* Do real assignment */
1817 if (alias_data
->domain
== NULL
)
1818 do_attach(alias_data
, domain
);
1820 atomic_inc(&alias_data
->bind
);
1823 if (dev_data
->domain
== NULL
)
1824 do_attach(dev_data
, domain
);
1826 atomic_inc(&dev_data
->bind
);
1833 spin_unlock(&domain
->lock
);
1839 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
1841 pci_disable_ats(pdev
);
1842 pci_disable_pri(pdev
);
1843 pci_disable_pasid(pdev
);
1846 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
1850 /* Only allow access to user-accessible pages */
1851 ret
= pci_enable_pasid(pdev
, 0);
1855 /* First reset the PRI state of the device */
1856 ret
= pci_reset_pri(pdev
);
1860 /* FIXME: Hardcode number of outstanding requests for now */
1861 ret
= pci_enable_pri(pdev
, 32);
1865 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
1872 pci_disable_pri(pdev
);
1873 pci_disable_pasid(pdev
);
1879 * If a device is not yet associated with a domain, this function does
1880 * assigns it visible for the hardware
1882 static int attach_device(struct device
*dev
,
1883 struct protection_domain
*domain
)
1885 struct pci_dev
*pdev
= to_pci_dev(dev
);
1886 struct iommu_dev_data
*dev_data
;
1887 unsigned long flags
;
1890 dev_data
= get_dev_data(dev
);
1892 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1893 if (!dev_data
->iommu_v2
|| !dev_data
->passthrough
)
1896 if (pdev_iommuv2_enable(pdev
) != 0)
1899 dev_data
->ats
.enabled
= true;
1900 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
1901 } else if (amd_iommu_iotlb_sup
&&
1902 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
1903 dev_data
->ats
.enabled
= true;
1904 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
1907 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1908 ret
= __attach_device(dev_data
, domain
);
1909 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1912 * We might boot into a crash-kernel here. The crashed kernel
1913 * left the caches in the IOMMU dirty. So we have to flush
1914 * here to evict all dirty stuff.
1916 domain_flush_tlb_pde(domain
);
1922 * Removes a device from a protection domain (unlocked)
1924 static void __detach_device(struct iommu_dev_data
*dev_data
)
1926 struct protection_domain
*domain
;
1927 unsigned long flags
;
1929 BUG_ON(!dev_data
->domain
);
1931 domain
= dev_data
->domain
;
1933 spin_lock_irqsave(&domain
->lock
, flags
);
1935 if (dev_data
->alias_data
!= NULL
) {
1936 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1938 if (atomic_dec_and_test(&alias_data
->bind
))
1939 do_detach(alias_data
);
1942 if (atomic_dec_and_test(&dev_data
->bind
))
1943 do_detach(dev_data
);
1945 spin_unlock_irqrestore(&domain
->lock
, flags
);
1948 * If we run in passthrough mode the device must be assigned to the
1949 * passthrough domain if it is detached from any other domain.
1950 * Make sure we can deassign from the pt_domain itself.
1952 if (dev_data
->passthrough
&&
1953 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
1954 __attach_device(dev_data
, pt_domain
);
1958 * Removes a device from a protection domain (with devtable_lock held)
1960 static void detach_device(struct device
*dev
)
1962 struct protection_domain
*domain
;
1963 struct iommu_dev_data
*dev_data
;
1964 unsigned long flags
;
1966 dev_data
= get_dev_data(dev
);
1967 domain
= dev_data
->domain
;
1969 /* lock device table */
1970 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1971 __detach_device(dev_data
);
1972 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1974 if (domain
->flags
& PD_IOMMUV2_MASK
)
1975 pdev_iommuv2_disable(to_pci_dev(dev
));
1976 else if (dev_data
->ats
.enabled
)
1977 pci_disable_ats(to_pci_dev(dev
));
1979 dev_data
->ats
.enabled
= false;
1983 * Find out the protection domain structure for a given PCI device. This
1984 * will give us the pointer to the page table root for example.
1986 static struct protection_domain
*domain_for_device(struct device
*dev
)
1988 struct iommu_dev_data
*dev_data
;
1989 struct protection_domain
*dom
= NULL
;
1990 unsigned long flags
;
1992 dev_data
= get_dev_data(dev
);
1994 if (dev_data
->domain
)
1995 return dev_data
->domain
;
1997 if (dev_data
->alias_data
!= NULL
) {
1998 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2000 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2001 if (alias_data
->domain
!= NULL
) {
2002 __attach_device(dev_data
, alias_data
->domain
);
2003 dom
= alias_data
->domain
;
2005 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2011 static int device_change_notifier(struct notifier_block
*nb
,
2012 unsigned long action
, void *data
)
2014 struct dma_ops_domain
*dma_domain
;
2015 struct protection_domain
*domain
;
2016 struct iommu_dev_data
*dev_data
;
2017 struct device
*dev
= data
;
2018 struct amd_iommu
*iommu
;
2019 unsigned long flags
;
2022 if (!check_device(dev
))
2025 devid
= get_device_id(dev
);
2026 iommu
= amd_iommu_rlookup_table
[devid
];
2027 dev_data
= get_dev_data(dev
);
2030 case BUS_NOTIFY_UNBOUND_DRIVER
:
2032 domain
= domain_for_device(dev
);
2036 if (dev_data
->passthrough
)
2040 case BUS_NOTIFY_ADD_DEVICE
:
2042 iommu_init_device(dev
);
2044 domain
= domain_for_device(dev
);
2046 /* allocate a protection domain if a device is added */
2047 dma_domain
= find_protection_domain(devid
);
2050 dma_domain
= dma_ops_domain_alloc();
2053 dma_domain
->target_dev
= devid
;
2055 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
2056 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
2057 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
2060 case BUS_NOTIFY_DEL_DEVICE
:
2062 iommu_uninit_device(dev
);
2068 iommu_completion_wait(iommu
);
2074 static struct notifier_block device_nb
= {
2075 .notifier_call
= device_change_notifier
,
2078 void amd_iommu_init_notifier(void)
2080 bus_register_notifier(&pci_bus_type
, &device_nb
);
2083 /*****************************************************************************
2085 * The next functions belong to the dma_ops mapping/unmapping code.
2087 *****************************************************************************/
2090 * In the dma_ops path we only have the struct device. This function
2091 * finds the corresponding IOMMU, the protection domain and the
2092 * requestor id for a given device.
2093 * If the device is not yet associated with a domain this is also done
2096 static struct protection_domain
*get_domain(struct device
*dev
)
2098 struct protection_domain
*domain
;
2099 struct dma_ops_domain
*dma_dom
;
2100 u16 devid
= get_device_id(dev
);
2102 if (!check_device(dev
))
2103 return ERR_PTR(-EINVAL
);
2105 domain
= domain_for_device(dev
);
2106 if (domain
!= NULL
&& !dma_ops_domain(domain
))
2107 return ERR_PTR(-EBUSY
);
2112 /* Device not bount yet - bind it */
2113 dma_dom
= find_protection_domain(devid
);
2115 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
2116 attach_device(dev
, &dma_dom
->domain
);
2117 DUMP_printk("Using protection domain %d for device %s\n",
2118 dma_dom
->domain
.id
, dev_name(dev
));
2120 return &dma_dom
->domain
;
2123 static void update_device_table(struct protection_domain
*domain
)
2125 struct iommu_dev_data
*dev_data
;
2127 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
2128 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2131 static void update_domain(struct protection_domain
*domain
)
2133 if (!domain
->updated
)
2136 update_device_table(domain
);
2138 domain_flush_devices(domain
);
2139 domain_flush_tlb_pde(domain
);
2141 domain
->updated
= false;
2145 * This function fetches the PTE for a given address in the aperture
2147 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
2148 unsigned long address
)
2150 struct aperture_range
*aperture
;
2151 u64
*pte
, *pte_page
;
2153 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2157 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2159 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
2161 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
2163 pte
+= PM_LEVEL_INDEX(0, address
);
2165 update_domain(&dom
->domain
);
2171 * This is the generic map function. It maps one 4kb page at paddr to
2172 * the given address in the DMA address space for the domain.
2174 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2175 unsigned long address
,
2181 WARN_ON(address
> dom
->aperture_size
);
2185 pte
= dma_ops_get_pte(dom
, address
);
2187 return DMA_ERROR_CODE
;
2189 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2191 if (direction
== DMA_TO_DEVICE
)
2192 __pte
|= IOMMU_PTE_IR
;
2193 else if (direction
== DMA_FROM_DEVICE
)
2194 __pte
|= IOMMU_PTE_IW
;
2195 else if (direction
== DMA_BIDIRECTIONAL
)
2196 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2202 return (dma_addr_t
)address
;
2206 * The generic unmapping function for on page in the DMA address space.
2208 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2209 unsigned long address
)
2211 struct aperture_range
*aperture
;
2214 if (address
>= dom
->aperture_size
)
2217 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2221 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2225 pte
+= PM_LEVEL_INDEX(0, address
);
2233 * This function contains common code for mapping of a physically
2234 * contiguous memory region into DMA address space. It is used by all
2235 * mapping functions provided with this IOMMU driver.
2236 * Must be called with the domain lock held.
2238 static dma_addr_t
__map_single(struct device
*dev
,
2239 struct dma_ops_domain
*dma_dom
,
2246 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2247 dma_addr_t address
, start
, ret
;
2249 unsigned long align_mask
= 0;
2252 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2255 INC_STATS_COUNTER(total_map_requests
);
2258 INC_STATS_COUNTER(cross_page
);
2261 align_mask
= (1UL << get_order(size
)) - 1;
2264 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2266 if (unlikely(address
== DMA_ERROR_CODE
)) {
2268 * setting next_address here will let the address
2269 * allocator only scan the new allocated range in the
2270 * first run. This is a small optimization.
2272 dma_dom
->next_address
= dma_dom
->aperture_size
;
2274 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2278 * aperture was successfully enlarged by 128 MB, try
2285 for (i
= 0; i
< pages
; ++i
) {
2286 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2287 if (ret
== DMA_ERROR_CODE
)
2295 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2297 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2298 domain_flush_tlb(&dma_dom
->domain
);
2299 dma_dom
->need_flush
= false;
2300 } else if (unlikely(amd_iommu_np_cache
))
2301 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2308 for (--i
; i
>= 0; --i
) {
2310 dma_ops_domain_unmap(dma_dom
, start
);
2313 dma_ops_free_addresses(dma_dom
, address
, pages
);
2315 return DMA_ERROR_CODE
;
2319 * Does the reverse of the __map_single function. Must be called with
2320 * the domain lock held too
2322 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2323 dma_addr_t dma_addr
,
2327 dma_addr_t flush_addr
;
2328 dma_addr_t i
, start
;
2331 if ((dma_addr
== DMA_ERROR_CODE
) ||
2332 (dma_addr
+ size
> dma_dom
->aperture_size
))
2335 flush_addr
= dma_addr
;
2336 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2337 dma_addr
&= PAGE_MASK
;
2340 for (i
= 0; i
< pages
; ++i
) {
2341 dma_ops_domain_unmap(dma_dom
, start
);
2345 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2347 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2349 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2350 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2351 dma_dom
->need_flush
= false;
2356 * The exported map_single function for dma_ops.
2358 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2359 unsigned long offset
, size_t size
,
2360 enum dma_data_direction dir
,
2361 struct dma_attrs
*attrs
)
2363 unsigned long flags
;
2364 struct protection_domain
*domain
;
2367 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2369 INC_STATS_COUNTER(cnt_map_single
);
2371 domain
= get_domain(dev
);
2372 if (PTR_ERR(domain
) == -EINVAL
)
2373 return (dma_addr_t
)paddr
;
2374 else if (IS_ERR(domain
))
2375 return DMA_ERROR_CODE
;
2377 dma_mask
= *dev
->dma_mask
;
2379 spin_lock_irqsave(&domain
->lock
, flags
);
2381 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2383 if (addr
== DMA_ERROR_CODE
)
2386 domain_flush_complete(domain
);
2389 spin_unlock_irqrestore(&domain
->lock
, flags
);
2395 * The exported unmap_single function for dma_ops.
2397 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2398 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2400 unsigned long flags
;
2401 struct protection_domain
*domain
;
2403 INC_STATS_COUNTER(cnt_unmap_single
);
2405 domain
= get_domain(dev
);
2409 spin_lock_irqsave(&domain
->lock
, flags
);
2411 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2413 domain_flush_complete(domain
);
2415 spin_unlock_irqrestore(&domain
->lock
, flags
);
2419 * This is a special map_sg function which is used if we should map a
2420 * device which is not handled by an AMD IOMMU in the system.
2422 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2423 int nelems
, int dir
)
2425 struct scatterlist
*s
;
2428 for_each_sg(sglist
, s
, nelems
, i
) {
2429 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2430 s
->dma_length
= s
->length
;
2437 * The exported map_sg function for dma_ops (handles scatter-gather
2440 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2441 int nelems
, enum dma_data_direction dir
,
2442 struct dma_attrs
*attrs
)
2444 unsigned long flags
;
2445 struct protection_domain
*domain
;
2447 struct scatterlist
*s
;
2449 int mapped_elems
= 0;
2452 INC_STATS_COUNTER(cnt_map_sg
);
2454 domain
= get_domain(dev
);
2455 if (PTR_ERR(domain
) == -EINVAL
)
2456 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2457 else if (IS_ERR(domain
))
2460 dma_mask
= *dev
->dma_mask
;
2462 spin_lock_irqsave(&domain
->lock
, flags
);
2464 for_each_sg(sglist
, s
, nelems
, i
) {
2467 s
->dma_address
= __map_single(dev
, domain
->priv
,
2468 paddr
, s
->length
, dir
, false,
2471 if (s
->dma_address
) {
2472 s
->dma_length
= s
->length
;
2478 domain_flush_complete(domain
);
2481 spin_unlock_irqrestore(&domain
->lock
, flags
);
2483 return mapped_elems
;
2485 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2487 __unmap_single(domain
->priv
, s
->dma_address
,
2488 s
->dma_length
, dir
);
2489 s
->dma_address
= s
->dma_length
= 0;
2498 * The exported map_sg function for dma_ops (handles scatter-gather
2501 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2502 int nelems
, enum dma_data_direction dir
,
2503 struct dma_attrs
*attrs
)
2505 unsigned long flags
;
2506 struct protection_domain
*domain
;
2507 struct scatterlist
*s
;
2510 INC_STATS_COUNTER(cnt_unmap_sg
);
2512 domain
= get_domain(dev
);
2516 spin_lock_irqsave(&domain
->lock
, flags
);
2518 for_each_sg(sglist
, s
, nelems
, i
) {
2519 __unmap_single(domain
->priv
, s
->dma_address
,
2520 s
->dma_length
, dir
);
2521 s
->dma_address
= s
->dma_length
= 0;
2524 domain_flush_complete(domain
);
2526 spin_unlock_irqrestore(&domain
->lock
, flags
);
2530 * The exported alloc_coherent function for dma_ops.
2532 static void *alloc_coherent(struct device
*dev
, size_t size
,
2533 dma_addr_t
*dma_addr
, gfp_t flag
)
2535 unsigned long flags
;
2537 struct protection_domain
*domain
;
2539 u64 dma_mask
= dev
->coherent_dma_mask
;
2541 INC_STATS_COUNTER(cnt_alloc_coherent
);
2543 domain
= get_domain(dev
);
2544 if (PTR_ERR(domain
) == -EINVAL
) {
2545 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2546 *dma_addr
= __pa(virt_addr
);
2548 } else if (IS_ERR(domain
))
2551 dma_mask
= dev
->coherent_dma_mask
;
2552 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2555 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2559 paddr
= virt_to_phys(virt_addr
);
2562 dma_mask
= *dev
->dma_mask
;
2564 spin_lock_irqsave(&domain
->lock
, flags
);
2566 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2567 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2569 if (*dma_addr
== DMA_ERROR_CODE
) {
2570 spin_unlock_irqrestore(&domain
->lock
, flags
);
2574 domain_flush_complete(domain
);
2576 spin_unlock_irqrestore(&domain
->lock
, flags
);
2582 free_pages((unsigned long)virt_addr
, get_order(size
));
2588 * The exported free_coherent function for dma_ops.
2590 static void free_coherent(struct device
*dev
, size_t size
,
2591 void *virt_addr
, dma_addr_t dma_addr
)
2593 unsigned long flags
;
2594 struct protection_domain
*domain
;
2596 INC_STATS_COUNTER(cnt_free_coherent
);
2598 domain
= get_domain(dev
);
2602 spin_lock_irqsave(&domain
->lock
, flags
);
2604 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2606 domain_flush_complete(domain
);
2608 spin_unlock_irqrestore(&domain
->lock
, flags
);
2611 free_pages((unsigned long)virt_addr
, get_order(size
));
2615 * This function is called by the DMA layer to find out if we can handle a
2616 * particular device. It is part of the dma_ops.
2618 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2620 return check_device(dev
);
2624 * The function for pre-allocating protection domains.
2626 * If the driver core informs the DMA layer if a driver grabs a device
2627 * we don't need to preallocate the protection domains anymore.
2628 * For now we have to.
2630 static void prealloc_protection_domains(void)
2632 struct iommu_dev_data
*dev_data
;
2633 struct dma_ops_domain
*dma_dom
;
2634 struct pci_dev
*dev
= NULL
;
2637 for_each_pci_dev(dev
) {
2639 /* Do we handle this device? */
2640 if (!check_device(&dev
->dev
))
2643 dev_data
= get_dev_data(&dev
->dev
);
2644 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
2645 /* Make sure passthrough domain is allocated */
2646 alloc_passthrough_domain();
2647 dev_data
->passthrough
= true;
2648 attach_device(&dev
->dev
, pt_domain
);
2649 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2650 dev_name(&dev
->dev
));
2653 /* Is there already any domain for it? */
2654 if (domain_for_device(&dev
->dev
))
2657 devid
= get_device_id(&dev
->dev
);
2659 dma_dom
= dma_ops_domain_alloc();
2662 init_unity_mappings_for_device(dma_dom
, devid
);
2663 dma_dom
->target_dev
= devid
;
2665 attach_device(&dev
->dev
, &dma_dom
->domain
);
2667 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2671 static struct dma_map_ops amd_iommu_dma_ops
= {
2672 .alloc_coherent
= alloc_coherent
,
2673 .free_coherent
= free_coherent
,
2674 .map_page
= map_page
,
2675 .unmap_page
= unmap_page
,
2677 .unmap_sg
= unmap_sg
,
2678 .dma_supported
= amd_iommu_dma_supported
,
2681 static unsigned device_dma_ops_init(void)
2683 struct iommu_dev_data
*dev_data
;
2684 struct pci_dev
*pdev
= NULL
;
2685 unsigned unhandled
= 0;
2687 for_each_pci_dev(pdev
) {
2688 if (!check_device(&pdev
->dev
)) {
2693 dev_data
= get_dev_data(&pdev
->dev
);
2695 if (!dev_data
->passthrough
)
2696 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
2698 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
2705 * The function which clues the AMD IOMMU driver into dma_ops.
2708 void __init
amd_iommu_init_api(void)
2710 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2713 int __init
amd_iommu_init_dma_ops(void)
2715 struct amd_iommu
*iommu
;
2719 * first allocate a default protection domain for every IOMMU we
2720 * found in the system. Devices not assigned to any other
2721 * protection domain will be assigned to the default one.
2723 for_each_iommu(iommu
) {
2724 iommu
->default_dom
= dma_ops_domain_alloc();
2725 if (iommu
->default_dom
== NULL
)
2727 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2728 ret
= iommu_init_unity_mappings(iommu
);
2734 * Pre-allocate the protection domains for each device.
2736 prealloc_protection_domains();
2741 /* Make the driver finally visible to the drivers */
2742 unhandled
= device_dma_ops_init();
2743 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
2744 /* There are unhandled devices - initialize swiotlb for them */
2748 amd_iommu_stats_init();
2754 for_each_iommu(iommu
) {
2755 if (iommu
->default_dom
)
2756 dma_ops_domain_free(iommu
->default_dom
);
2762 /*****************************************************************************
2764 * The following functions belong to the exported interface of AMD IOMMU
2766 * This interface allows access to lower level functions of the IOMMU
2767 * like protection domain handling and assignement of devices to domains
2768 * which is not possible with the dma_ops interface.
2770 *****************************************************************************/
2772 static void cleanup_domain(struct protection_domain
*domain
)
2774 struct iommu_dev_data
*dev_data
, *next
;
2775 unsigned long flags
;
2777 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2779 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
2780 __detach_device(dev_data
);
2781 atomic_set(&dev_data
->bind
, 0);
2784 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2787 static void protection_domain_free(struct protection_domain
*domain
)
2792 del_domain_from_list(domain
);
2795 domain_id_free(domain
->id
);
2800 static struct protection_domain
*protection_domain_alloc(void)
2802 struct protection_domain
*domain
;
2804 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2808 spin_lock_init(&domain
->lock
);
2809 mutex_init(&domain
->api_lock
);
2810 domain
->id
= domain_id_alloc();
2813 INIT_LIST_HEAD(&domain
->dev_list
);
2815 add_domain_to_list(domain
);
2825 static int __init
alloc_passthrough_domain(void)
2827 if (pt_domain
!= NULL
)
2830 /* allocate passthrough domain */
2831 pt_domain
= protection_domain_alloc();
2835 pt_domain
->mode
= PAGE_MODE_NONE
;
2839 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2841 struct protection_domain
*domain
;
2843 domain
= protection_domain_alloc();
2847 domain
->mode
= PAGE_MODE_3_LEVEL
;
2848 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2849 if (!domain
->pt_root
)
2857 protection_domain_free(domain
);
2862 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2864 struct protection_domain
*domain
= dom
->priv
;
2869 if (domain
->dev_cnt
> 0)
2870 cleanup_domain(domain
);
2872 BUG_ON(domain
->dev_cnt
!= 0);
2874 if (domain
->mode
!= PAGE_MODE_NONE
)
2875 free_pagetable(domain
);
2877 if (domain
->flags
& PD_IOMMUV2_MASK
)
2878 free_gcr3_table(domain
);
2880 protection_domain_free(domain
);
2885 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2888 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
2889 struct amd_iommu
*iommu
;
2892 if (!check_device(dev
))
2895 devid
= get_device_id(dev
);
2897 if (dev_data
->domain
!= NULL
)
2900 iommu
= amd_iommu_rlookup_table
[devid
];
2904 iommu_completion_wait(iommu
);
2907 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2910 struct protection_domain
*domain
= dom
->priv
;
2911 struct iommu_dev_data
*dev_data
;
2912 struct amd_iommu
*iommu
;
2915 if (!check_device(dev
))
2918 dev_data
= dev
->archdata
.iommu
;
2920 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2924 if (dev_data
->domain
)
2927 ret
= attach_device(dev
, domain
);
2929 iommu_completion_wait(iommu
);
2934 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
2935 phys_addr_t paddr
, int gfp_order
, int iommu_prot
)
2937 unsigned long page_size
= 0x1000UL
<< gfp_order
;
2938 struct protection_domain
*domain
= dom
->priv
;
2942 if (domain
->mode
== PAGE_MODE_NONE
)
2945 if (iommu_prot
& IOMMU_READ
)
2946 prot
|= IOMMU_PROT_IR
;
2947 if (iommu_prot
& IOMMU_WRITE
)
2948 prot
|= IOMMU_PROT_IW
;
2950 mutex_lock(&domain
->api_lock
);
2951 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
2952 mutex_unlock(&domain
->api_lock
);
2957 static int amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
2960 struct protection_domain
*domain
= dom
->priv
;
2961 unsigned long page_size
, unmap_size
;
2963 if (domain
->mode
== PAGE_MODE_NONE
)
2966 page_size
= 0x1000UL
<< gfp_order
;
2968 mutex_lock(&domain
->api_lock
);
2969 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
2970 mutex_unlock(&domain
->api_lock
);
2972 domain_flush_tlb_pde(domain
);
2974 return get_order(unmap_size
);
2977 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
2980 struct protection_domain
*domain
= dom
->priv
;
2981 unsigned long offset_mask
;
2985 if (domain
->mode
== PAGE_MODE_NONE
)
2988 pte
= fetch_pte(domain
, iova
);
2990 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
2993 if (PM_PTE_LEVEL(*pte
) == 0)
2994 offset_mask
= PAGE_SIZE
- 1;
2996 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
2998 __pte
= *pte
& PM_ADDR_MASK
;
2999 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3004 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
3008 case IOMMU_CAP_CACHE_COHERENCY
:
3015 static struct iommu_ops amd_iommu_ops
= {
3016 .domain_init
= amd_iommu_domain_init
,
3017 .domain_destroy
= amd_iommu_domain_destroy
,
3018 .attach_dev
= amd_iommu_attach_device
,
3019 .detach_dev
= amd_iommu_detach_device
,
3020 .map
= amd_iommu_map
,
3021 .unmap
= amd_iommu_unmap
,
3022 .iova_to_phys
= amd_iommu_iova_to_phys
,
3023 .domain_has_cap
= amd_iommu_domain_has_cap
,
3026 /*****************************************************************************
3028 * The next functions do a basic initialization of IOMMU for pass through
3031 * In passthrough mode the IOMMU is initialized and enabled but not used for
3032 * DMA-API translation.
3034 *****************************************************************************/
3036 int __init
amd_iommu_init_passthrough(void)
3038 struct iommu_dev_data
*dev_data
;
3039 struct pci_dev
*dev
= NULL
;
3040 struct amd_iommu
*iommu
;
3044 ret
= alloc_passthrough_domain();
3048 for_each_pci_dev(dev
) {
3049 if (!check_device(&dev
->dev
))
3052 dev_data
= get_dev_data(&dev
->dev
);
3053 dev_data
->passthrough
= true;
3055 devid
= get_device_id(&dev
->dev
);
3057 iommu
= amd_iommu_rlookup_table
[devid
];
3061 attach_device(&dev
->dev
, pt_domain
);
3064 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3069 /* IOMMUv2 specific functions */
3070 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3072 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3074 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3076 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3078 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3080 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3082 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3084 struct protection_domain
*domain
= dom
->priv
;
3085 unsigned long flags
;
3087 spin_lock_irqsave(&domain
->lock
, flags
);
3089 /* Update data structure */
3090 domain
->mode
= PAGE_MODE_NONE
;
3091 domain
->updated
= true;
3093 /* Make changes visible to IOMMUs */
3094 update_domain(domain
);
3096 /* Page-table is not visible to IOMMU anymore, so free it */
3097 free_pagetable(domain
);
3099 spin_unlock_irqrestore(&domain
->lock
, flags
);
3101 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3103 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3105 struct protection_domain
*domain
= dom
->priv
;
3106 unsigned long flags
;
3109 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3112 /* Number of GCR3 table levels required */
3113 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3116 if (levels
> amd_iommu_max_glx_val
)
3119 spin_lock_irqsave(&domain
->lock
, flags
);
3122 * Save us all sanity checks whether devices already in the
3123 * domain support IOMMUv2. Just force that the domain has no
3124 * devices attached when it is switched into IOMMUv2 mode.
3127 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3131 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3132 if (domain
->gcr3_tbl
== NULL
)
3135 domain
->glx
= levels
;
3136 domain
->flags
|= PD_IOMMUV2_MASK
;
3137 domain
->updated
= true;
3139 update_domain(domain
);
3144 spin_unlock_irqrestore(&domain
->lock
, flags
);
3148 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);