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_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
702 u64 address
, bool size
)
704 memset(cmd
, 0, sizeof(*cmd
));
706 address
&= ~(0xfffULL
);
708 cmd
->data
[0] = pasid
& PASID_MASK
;
709 cmd
->data
[1] = domid
;
710 cmd
->data
[2] = lower_32_bits(address
);
711 cmd
->data
[3] = upper_32_bits(address
);
712 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
713 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
715 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
716 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
719 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
720 int qdep
, u64 address
, bool size
)
722 memset(cmd
, 0, sizeof(*cmd
));
724 address
&= ~(0xfffULL
);
726 cmd
->data
[0] = devid
;
727 cmd
->data
[0] |= (pasid
& 0xff) << 16;
728 cmd
->data
[0] |= (qdep
& 0xff) << 24;
729 cmd
->data
[1] = devid
;
730 cmd
->data
[1] |= ((pasid
>> 8) & 0xfff) << 16;
731 cmd
->data
[2] = lower_32_bits(address
);
732 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
733 cmd
->data
[3] = upper_32_bits(address
);
735 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
736 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
739 static void build_inv_all(struct iommu_cmd
*cmd
)
741 memset(cmd
, 0, sizeof(*cmd
));
742 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
746 * Writes the command to the IOMMUs command buffer and informs the
747 * hardware about the new command.
749 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
750 struct iommu_cmd
*cmd
,
753 u32 left
, tail
, head
, next_tail
;
756 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
759 spin_lock_irqsave(&iommu
->lock
, flags
);
761 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
762 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
763 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
764 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
767 struct iommu_cmd sync_cmd
;
768 volatile u64 sem
= 0;
771 build_completion_wait(&sync_cmd
, (u64
)&sem
);
772 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
774 spin_unlock_irqrestore(&iommu
->lock
, flags
);
776 if ((ret
= wait_on_sem(&sem
)) != 0)
782 copy_cmd_to_buffer(iommu
, cmd
, tail
);
784 /* We need to sync now to make sure all commands are processed */
785 iommu
->need_sync
= sync
;
787 spin_unlock_irqrestore(&iommu
->lock
, flags
);
792 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
794 return iommu_queue_command_sync(iommu
, cmd
, true);
798 * This function queues a completion wait command into the command
801 static int iommu_completion_wait(struct amd_iommu
*iommu
)
803 struct iommu_cmd cmd
;
804 volatile u64 sem
= 0;
807 if (!iommu
->need_sync
)
810 build_completion_wait(&cmd
, (u64
)&sem
);
812 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
816 return wait_on_sem(&sem
);
819 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
821 struct iommu_cmd cmd
;
823 build_inv_dte(&cmd
, devid
);
825 return iommu_queue_command(iommu
, &cmd
);
828 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
832 for (devid
= 0; devid
<= 0xffff; ++devid
)
833 iommu_flush_dte(iommu
, devid
);
835 iommu_completion_wait(iommu
);
839 * This function uses heavy locking and may disable irqs for some time. But
840 * this is no issue because it is only called during resume.
842 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
846 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
847 struct iommu_cmd cmd
;
848 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
850 iommu_queue_command(iommu
, &cmd
);
853 iommu_completion_wait(iommu
);
856 static void iommu_flush_all(struct amd_iommu
*iommu
)
858 struct iommu_cmd cmd
;
862 iommu_queue_command(iommu
, &cmd
);
863 iommu_completion_wait(iommu
);
866 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
868 if (iommu_feature(iommu
, FEATURE_IA
)) {
869 iommu_flush_all(iommu
);
871 iommu_flush_dte_all(iommu
);
872 iommu_flush_tlb_all(iommu
);
877 * Command send function for flushing on-device TLB
879 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
880 u64 address
, size_t size
)
882 struct amd_iommu
*iommu
;
883 struct iommu_cmd cmd
;
886 qdep
= dev_data
->ats
.qdep
;
887 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
889 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
891 return iommu_queue_command(iommu
, &cmd
);
895 * Command send function for invalidating a device table entry
897 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
899 struct amd_iommu
*iommu
;
902 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
904 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
908 if (dev_data
->ats
.enabled
)
909 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
915 * TLB invalidation function which is called from the mapping functions.
916 * It invalidates a single PTE if the range to flush is within a single
917 * page. Otherwise it flushes the whole TLB of the IOMMU.
919 static void __domain_flush_pages(struct protection_domain
*domain
,
920 u64 address
, size_t size
, int pde
)
922 struct iommu_dev_data
*dev_data
;
923 struct iommu_cmd cmd
;
926 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
928 for (i
= 0; i
< amd_iommus_present
; ++i
) {
929 if (!domain
->dev_iommu
[i
])
933 * Devices of this domain are behind this IOMMU
934 * We need a TLB flush
936 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
939 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
941 if (!dev_data
->ats
.enabled
)
944 ret
|= device_flush_iotlb(dev_data
, address
, size
);
950 static void domain_flush_pages(struct protection_domain
*domain
,
951 u64 address
, size_t size
)
953 __domain_flush_pages(domain
, address
, size
, 0);
956 /* Flush the whole IO/TLB for a given protection domain */
957 static void domain_flush_tlb(struct protection_domain
*domain
)
959 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
962 /* Flush the whole IO/TLB for a given protection domain - including PDE */
963 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
965 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
968 static void domain_flush_complete(struct protection_domain
*domain
)
972 for (i
= 0; i
< amd_iommus_present
; ++i
) {
973 if (!domain
->dev_iommu
[i
])
977 * Devices of this domain are behind this IOMMU
978 * We need to wait for completion of all commands.
980 iommu_completion_wait(amd_iommus
[i
]);
986 * This function flushes the DTEs for all devices in domain
988 static void domain_flush_devices(struct protection_domain
*domain
)
990 struct iommu_dev_data
*dev_data
;
992 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
993 device_flush_dte(dev_data
);
996 /****************************************************************************
998 * The functions below are used the create the page table mappings for
999 * unity mapped regions.
1001 ****************************************************************************/
1004 * This function is used to add another level to an IO page table. Adding
1005 * another level increases the size of the address space by 9 bits to a size up
1008 static bool increase_address_space(struct protection_domain
*domain
,
1013 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1014 /* address space already 64 bit large */
1017 pte
= (void *)get_zeroed_page(gfp
);
1021 *pte
= PM_LEVEL_PDE(domain
->mode
,
1022 virt_to_phys(domain
->pt_root
));
1023 domain
->pt_root
= pte
;
1025 domain
->updated
= true;
1030 static u64
*alloc_pte(struct protection_domain
*domain
,
1031 unsigned long address
,
1032 unsigned long page_size
,
1039 BUG_ON(!is_power_of_2(page_size
));
1041 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1042 increase_address_space(domain
, gfp
);
1044 level
= domain
->mode
- 1;
1045 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1046 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1047 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1049 while (level
> end_lvl
) {
1050 if (!IOMMU_PTE_PRESENT(*pte
)) {
1051 page
= (u64
*)get_zeroed_page(gfp
);
1054 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1057 /* No level skipping support yet */
1058 if (PM_PTE_LEVEL(*pte
) != level
)
1063 pte
= IOMMU_PTE_PAGE(*pte
);
1065 if (pte_page
&& level
== end_lvl
)
1068 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1075 * This function checks if there is a PTE for a given dma address. If
1076 * there is one, it returns the pointer to it.
1078 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
1083 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1086 level
= domain
->mode
- 1;
1087 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1092 if (!IOMMU_PTE_PRESENT(*pte
))
1096 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1097 unsigned long pte_mask
, __pte
;
1100 * If we have a series of large PTEs, make
1101 * sure to return a pointer to the first one.
1103 pte_mask
= PTE_PAGE_SIZE(*pte
);
1104 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1105 __pte
= ((unsigned long)pte
) & pte_mask
;
1107 return (u64
*)__pte
;
1110 /* No level skipping support yet */
1111 if (PM_PTE_LEVEL(*pte
) != level
)
1116 /* Walk to the next level */
1117 pte
= IOMMU_PTE_PAGE(*pte
);
1118 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1125 * Generic mapping functions. It maps a physical address into a DMA
1126 * address space. It allocates the page table pages if necessary.
1127 * In the future it can be extended to a generic mapping function
1128 * supporting all features of AMD IOMMU page tables like level skipping
1129 * and full 64 bit address spaces.
1131 static int iommu_map_page(struct protection_domain
*dom
,
1132 unsigned long bus_addr
,
1133 unsigned long phys_addr
,
1135 unsigned long page_size
)
1140 if (!(prot
& IOMMU_PROT_MASK
))
1143 bus_addr
= PAGE_ALIGN(bus_addr
);
1144 phys_addr
= PAGE_ALIGN(phys_addr
);
1145 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1146 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1148 for (i
= 0; i
< count
; ++i
)
1149 if (IOMMU_PTE_PRESENT(pte
[i
]))
1152 if (page_size
> PAGE_SIZE
) {
1153 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1154 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1156 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1158 if (prot
& IOMMU_PROT_IR
)
1159 __pte
|= IOMMU_PTE_IR
;
1160 if (prot
& IOMMU_PROT_IW
)
1161 __pte
|= IOMMU_PTE_IW
;
1163 for (i
= 0; i
< count
; ++i
)
1171 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1172 unsigned long bus_addr
,
1173 unsigned long page_size
)
1175 unsigned long long unmap_size
, unmapped
;
1178 BUG_ON(!is_power_of_2(page_size
));
1182 while (unmapped
< page_size
) {
1184 pte
= fetch_pte(dom
, bus_addr
);
1188 * No PTE for this address
1189 * move forward in 4kb steps
1191 unmap_size
= PAGE_SIZE
;
1192 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1193 /* 4kb PTE found for this address */
1194 unmap_size
= PAGE_SIZE
;
1199 /* Large PTE found which maps this address */
1200 unmap_size
= PTE_PAGE_SIZE(*pte
);
1201 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1202 for (i
= 0; i
< count
; i
++)
1206 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1207 unmapped
+= unmap_size
;
1210 BUG_ON(!is_power_of_2(unmapped
));
1216 * This function checks if a specific unity mapping entry is needed for
1217 * this specific IOMMU.
1219 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1220 struct unity_map_entry
*entry
)
1224 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1225 bdf
= amd_iommu_alias_table
[i
];
1226 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1234 * This function actually applies the mapping to the page table of the
1237 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1238 struct unity_map_entry
*e
)
1243 for (addr
= e
->address_start
; addr
< e
->address_end
;
1244 addr
+= PAGE_SIZE
) {
1245 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1250 * if unity mapping is in aperture range mark the page
1251 * as allocated in the aperture
1253 if (addr
< dma_dom
->aperture_size
)
1254 __set_bit(addr
>> PAGE_SHIFT
,
1255 dma_dom
->aperture
[0]->bitmap
);
1262 * Init the unity mappings for a specific IOMMU in the system
1264 * Basically iterates over all unity mapping entries and applies them to
1265 * the default domain DMA of that IOMMU if necessary.
1267 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1269 struct unity_map_entry
*entry
;
1272 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1273 if (!iommu_for_unity_map(iommu
, entry
))
1275 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1284 * Inits the unity mappings required for a specific device
1286 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1289 struct unity_map_entry
*e
;
1292 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1293 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1295 ret
= dma_ops_unity_map(dma_dom
, e
);
1303 /****************************************************************************
1305 * The next functions belong to the address allocator for the dma_ops
1306 * interface functions. They work like the allocators in the other IOMMU
1307 * drivers. Its basically a bitmap which marks the allocated pages in
1308 * the aperture. Maybe it could be enhanced in the future to a more
1309 * efficient allocator.
1311 ****************************************************************************/
1314 * The address allocator core functions.
1316 * called with domain->lock held
1320 * Used to reserve address ranges in the aperture (e.g. for exclusion
1323 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1324 unsigned long start_page
,
1327 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1329 if (start_page
+ pages
> last_page
)
1330 pages
= last_page
- start_page
;
1332 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1333 int index
= i
/ APERTURE_RANGE_PAGES
;
1334 int page
= i
% APERTURE_RANGE_PAGES
;
1335 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1340 * This function is used to add a new aperture range to an existing
1341 * aperture in case of dma_ops domain allocation or address allocation
1344 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1345 bool populate
, gfp_t gfp
)
1347 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1348 struct amd_iommu
*iommu
;
1349 unsigned long i
, old_size
;
1351 #ifdef CONFIG_IOMMU_STRESS
1355 if (index
>= APERTURE_MAX_RANGES
)
1358 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1359 if (!dma_dom
->aperture
[index
])
1362 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1363 if (!dma_dom
->aperture
[index
]->bitmap
)
1366 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1369 unsigned long address
= dma_dom
->aperture_size
;
1370 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1371 u64
*pte
, *pte_page
;
1373 for (i
= 0; i
< num_ptes
; ++i
) {
1374 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1379 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1381 address
+= APERTURE_RANGE_SIZE
/ 64;
1385 old_size
= dma_dom
->aperture_size
;
1386 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1388 /* Reserve address range used for MSI messages */
1389 if (old_size
< MSI_ADDR_BASE_LO
&&
1390 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1391 unsigned long spage
;
1394 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1395 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1397 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1400 /* Initialize the exclusion range if necessary */
1401 for_each_iommu(iommu
) {
1402 if (iommu
->exclusion_start
&&
1403 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1404 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1405 unsigned long startpage
;
1406 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1407 iommu
->exclusion_length
,
1409 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1410 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1415 * Check for areas already mapped as present in the new aperture
1416 * range and mark those pages as reserved in the allocator. Such
1417 * mappings may already exist as a result of requested unity
1418 * mappings for devices.
1420 for (i
= dma_dom
->aperture
[index
]->offset
;
1421 i
< dma_dom
->aperture_size
;
1423 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1424 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1427 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1430 update_domain(&dma_dom
->domain
);
1435 update_domain(&dma_dom
->domain
);
1437 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1439 kfree(dma_dom
->aperture
[index
]);
1440 dma_dom
->aperture
[index
] = NULL
;
1445 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1446 struct dma_ops_domain
*dom
,
1448 unsigned long align_mask
,
1450 unsigned long start
)
1452 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1453 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1454 int i
= start
>> APERTURE_RANGE_SHIFT
;
1455 unsigned long boundary_size
;
1456 unsigned long address
= -1;
1457 unsigned long limit
;
1459 next_bit
>>= PAGE_SHIFT
;
1461 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1462 PAGE_SIZE
) >> PAGE_SHIFT
;
1464 for (;i
< max_index
; ++i
) {
1465 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1467 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1470 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1471 dma_mask
>> PAGE_SHIFT
);
1473 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1474 limit
, next_bit
, pages
, 0,
1475 boundary_size
, align_mask
);
1476 if (address
!= -1) {
1477 address
= dom
->aperture
[i
]->offset
+
1478 (address
<< PAGE_SHIFT
);
1479 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1489 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1490 struct dma_ops_domain
*dom
,
1492 unsigned long align_mask
,
1495 unsigned long address
;
1497 #ifdef CONFIG_IOMMU_STRESS
1498 dom
->next_address
= 0;
1499 dom
->need_flush
= true;
1502 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1503 dma_mask
, dom
->next_address
);
1505 if (address
== -1) {
1506 dom
->next_address
= 0;
1507 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1509 dom
->need_flush
= true;
1512 if (unlikely(address
== -1))
1513 address
= DMA_ERROR_CODE
;
1515 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1521 * The address free function.
1523 * called with domain->lock held
1525 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1526 unsigned long address
,
1529 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1530 struct aperture_range
*range
= dom
->aperture
[i
];
1532 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1534 #ifdef CONFIG_IOMMU_STRESS
1539 if (address
>= dom
->next_address
)
1540 dom
->need_flush
= true;
1542 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1544 bitmap_clear(range
->bitmap
, address
, pages
);
1548 /****************************************************************************
1550 * The next functions belong to the domain allocation. A domain is
1551 * allocated for every IOMMU as the default domain. If device isolation
1552 * is enabled, every device get its own domain. The most important thing
1553 * about domains is the page table mapping the DMA address space they
1556 ****************************************************************************/
1559 * This function adds a protection domain to the global protection domain list
1561 static void add_domain_to_list(struct protection_domain
*domain
)
1563 unsigned long flags
;
1565 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1566 list_add(&domain
->list
, &amd_iommu_pd_list
);
1567 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1571 * This function removes a protection domain to the global
1572 * protection domain list
1574 static void del_domain_from_list(struct protection_domain
*domain
)
1576 unsigned long flags
;
1578 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1579 list_del(&domain
->list
);
1580 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1583 static u16
domain_id_alloc(void)
1585 unsigned long flags
;
1588 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1589 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1591 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1592 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1595 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1600 static void domain_id_free(int id
)
1602 unsigned long flags
;
1604 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1605 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1606 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1607 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1610 static void free_pagetable(struct protection_domain
*domain
)
1615 p1
= domain
->pt_root
;
1620 for (i
= 0; i
< 512; ++i
) {
1621 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1624 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1625 for (j
= 0; j
< 512; ++j
) {
1626 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1628 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1629 free_page((unsigned long)p3
);
1632 free_page((unsigned long)p2
);
1635 free_page((unsigned long)p1
);
1637 domain
->pt_root
= NULL
;
1640 static void free_gcr3_table(struct protection_domain
*domain
)
1642 free_page((unsigned long)domain
->gcr3_tbl
);
1646 * Free a domain, only used if something went wrong in the
1647 * allocation path and we need to free an already allocated page table
1649 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1656 del_domain_from_list(&dom
->domain
);
1658 free_pagetable(&dom
->domain
);
1660 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1661 if (!dom
->aperture
[i
])
1663 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1664 kfree(dom
->aperture
[i
]);
1671 * Allocates a new protection domain usable for the dma_ops functions.
1672 * It also initializes the page table and the address allocator data
1673 * structures required for the dma_ops interface
1675 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1677 struct dma_ops_domain
*dma_dom
;
1679 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1683 spin_lock_init(&dma_dom
->domain
.lock
);
1685 dma_dom
->domain
.id
= domain_id_alloc();
1686 if (dma_dom
->domain
.id
== 0)
1688 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1689 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1690 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1691 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1692 dma_dom
->domain
.priv
= dma_dom
;
1693 if (!dma_dom
->domain
.pt_root
)
1696 dma_dom
->need_flush
= false;
1697 dma_dom
->target_dev
= 0xffff;
1699 add_domain_to_list(&dma_dom
->domain
);
1701 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1705 * mark the first page as allocated so we never return 0 as
1706 * a valid dma-address. So we can use 0 as error value
1708 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1709 dma_dom
->next_address
= 0;
1715 dma_ops_domain_free(dma_dom
);
1721 * little helper function to check whether a given protection domain is a
1724 static bool dma_ops_domain(struct protection_domain
*domain
)
1726 return domain
->flags
& PD_DMA_OPS_MASK
;
1729 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1734 if (domain
->mode
!= PAGE_MODE_NONE
)
1735 pte_root
= virt_to_phys(domain
->pt_root
);
1737 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1738 << DEV_ENTRY_MODE_SHIFT
;
1739 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1741 flags
= amd_iommu_dev_table
[devid
].data
[1];
1744 flags
|= DTE_FLAG_IOTLB
;
1746 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1747 u64 gcr3
= __pa(domain
->gcr3_tbl
);
1748 u64 glx
= domain
->glx
;
1751 pte_root
|= DTE_FLAG_GV
;
1752 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1754 /* First mask out possible old values for GCR3 table */
1755 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1758 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1761 /* Encode GCR3 table into DTE */
1762 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1765 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1768 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1772 flags
&= ~(0xffffUL
);
1773 flags
|= domain
->id
;
1775 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1776 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1779 static void clear_dte_entry(u16 devid
)
1781 /* remove entry from the device table seen by the hardware */
1782 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1783 amd_iommu_dev_table
[devid
].data
[1] = 0;
1785 amd_iommu_apply_erratum_63(devid
);
1788 static void do_attach(struct iommu_dev_data
*dev_data
,
1789 struct protection_domain
*domain
)
1791 struct amd_iommu
*iommu
;
1794 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1795 ats
= dev_data
->ats
.enabled
;
1797 /* Update data structures */
1798 dev_data
->domain
= domain
;
1799 list_add(&dev_data
->list
, &domain
->dev_list
);
1800 set_dte_entry(dev_data
->devid
, domain
, ats
);
1802 /* Do reference counting */
1803 domain
->dev_iommu
[iommu
->index
] += 1;
1804 domain
->dev_cnt
+= 1;
1806 /* Flush the DTE entry */
1807 device_flush_dte(dev_data
);
1810 static void do_detach(struct iommu_dev_data
*dev_data
)
1812 struct amd_iommu
*iommu
;
1814 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1816 /* decrease reference counters */
1817 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1818 dev_data
->domain
->dev_cnt
-= 1;
1820 /* Update data structures */
1821 dev_data
->domain
= NULL
;
1822 list_del(&dev_data
->list
);
1823 clear_dte_entry(dev_data
->devid
);
1825 /* Flush the DTE entry */
1826 device_flush_dte(dev_data
);
1830 * If a device is not yet associated with a domain, this function does
1831 * assigns it visible for the hardware
1833 static int __attach_device(struct iommu_dev_data
*dev_data
,
1834 struct protection_domain
*domain
)
1839 spin_lock(&domain
->lock
);
1841 if (dev_data
->alias_data
!= NULL
) {
1842 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1844 /* Some sanity checks */
1846 if (alias_data
->domain
!= NULL
&&
1847 alias_data
->domain
!= domain
)
1850 if (dev_data
->domain
!= NULL
&&
1851 dev_data
->domain
!= domain
)
1854 /* Do real assignment */
1855 if (alias_data
->domain
== NULL
)
1856 do_attach(alias_data
, domain
);
1858 atomic_inc(&alias_data
->bind
);
1861 if (dev_data
->domain
== NULL
)
1862 do_attach(dev_data
, domain
);
1864 atomic_inc(&dev_data
->bind
);
1871 spin_unlock(&domain
->lock
);
1877 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
1879 pci_disable_ats(pdev
);
1880 pci_disable_pri(pdev
);
1881 pci_disable_pasid(pdev
);
1884 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
1888 /* Only allow access to user-accessible pages */
1889 ret
= pci_enable_pasid(pdev
, 0);
1893 /* First reset the PRI state of the device */
1894 ret
= pci_reset_pri(pdev
);
1898 /* FIXME: Hardcode number of outstanding requests for now */
1899 ret
= pci_enable_pri(pdev
, 32);
1903 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
1910 pci_disable_pri(pdev
);
1911 pci_disable_pasid(pdev
);
1917 * If a device is not yet associated with a domain, this function does
1918 * assigns it visible for the hardware
1920 static int attach_device(struct device
*dev
,
1921 struct protection_domain
*domain
)
1923 struct pci_dev
*pdev
= to_pci_dev(dev
);
1924 struct iommu_dev_data
*dev_data
;
1925 unsigned long flags
;
1928 dev_data
= get_dev_data(dev
);
1930 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1931 if (!dev_data
->iommu_v2
|| !dev_data
->passthrough
)
1934 if (pdev_iommuv2_enable(pdev
) != 0)
1937 dev_data
->ats
.enabled
= true;
1938 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
1939 } else if (amd_iommu_iotlb_sup
&&
1940 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
1941 dev_data
->ats
.enabled
= true;
1942 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
1945 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1946 ret
= __attach_device(dev_data
, domain
);
1947 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1950 * We might boot into a crash-kernel here. The crashed kernel
1951 * left the caches in the IOMMU dirty. So we have to flush
1952 * here to evict all dirty stuff.
1954 domain_flush_tlb_pde(domain
);
1960 * Removes a device from a protection domain (unlocked)
1962 static void __detach_device(struct iommu_dev_data
*dev_data
)
1964 struct protection_domain
*domain
;
1965 unsigned long flags
;
1967 BUG_ON(!dev_data
->domain
);
1969 domain
= dev_data
->domain
;
1971 spin_lock_irqsave(&domain
->lock
, flags
);
1973 if (dev_data
->alias_data
!= NULL
) {
1974 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
1976 if (atomic_dec_and_test(&alias_data
->bind
))
1977 do_detach(alias_data
);
1980 if (atomic_dec_and_test(&dev_data
->bind
))
1981 do_detach(dev_data
);
1983 spin_unlock_irqrestore(&domain
->lock
, flags
);
1986 * If we run in passthrough mode the device must be assigned to the
1987 * passthrough domain if it is detached from any other domain.
1988 * Make sure we can deassign from the pt_domain itself.
1990 if (dev_data
->passthrough
&&
1991 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
1992 __attach_device(dev_data
, pt_domain
);
1996 * Removes a device from a protection domain (with devtable_lock held)
1998 static void detach_device(struct device
*dev
)
2000 struct protection_domain
*domain
;
2001 struct iommu_dev_data
*dev_data
;
2002 unsigned long flags
;
2004 dev_data
= get_dev_data(dev
);
2005 domain
= dev_data
->domain
;
2007 /* lock device table */
2008 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2009 __detach_device(dev_data
);
2010 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2012 if (domain
->flags
& PD_IOMMUV2_MASK
)
2013 pdev_iommuv2_disable(to_pci_dev(dev
));
2014 else if (dev_data
->ats
.enabled
)
2015 pci_disable_ats(to_pci_dev(dev
));
2017 dev_data
->ats
.enabled
= false;
2021 * Find out the protection domain structure for a given PCI device. This
2022 * will give us the pointer to the page table root for example.
2024 static struct protection_domain
*domain_for_device(struct device
*dev
)
2026 struct iommu_dev_data
*dev_data
;
2027 struct protection_domain
*dom
= NULL
;
2028 unsigned long flags
;
2030 dev_data
= get_dev_data(dev
);
2032 if (dev_data
->domain
)
2033 return dev_data
->domain
;
2035 if (dev_data
->alias_data
!= NULL
) {
2036 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2038 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2039 if (alias_data
->domain
!= NULL
) {
2040 __attach_device(dev_data
, alias_data
->domain
);
2041 dom
= alias_data
->domain
;
2043 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2049 static int device_change_notifier(struct notifier_block
*nb
,
2050 unsigned long action
, void *data
)
2052 struct dma_ops_domain
*dma_domain
;
2053 struct protection_domain
*domain
;
2054 struct iommu_dev_data
*dev_data
;
2055 struct device
*dev
= data
;
2056 struct amd_iommu
*iommu
;
2057 unsigned long flags
;
2060 if (!check_device(dev
))
2063 devid
= get_device_id(dev
);
2064 iommu
= amd_iommu_rlookup_table
[devid
];
2065 dev_data
= get_dev_data(dev
);
2068 case BUS_NOTIFY_UNBOUND_DRIVER
:
2070 domain
= domain_for_device(dev
);
2074 if (dev_data
->passthrough
)
2078 case BUS_NOTIFY_ADD_DEVICE
:
2080 iommu_init_device(dev
);
2082 domain
= domain_for_device(dev
);
2084 /* allocate a protection domain if a device is added */
2085 dma_domain
= find_protection_domain(devid
);
2088 dma_domain
= dma_ops_domain_alloc();
2091 dma_domain
->target_dev
= devid
;
2093 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
2094 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
2095 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
2098 case BUS_NOTIFY_DEL_DEVICE
:
2100 iommu_uninit_device(dev
);
2106 iommu_completion_wait(iommu
);
2112 static struct notifier_block device_nb
= {
2113 .notifier_call
= device_change_notifier
,
2116 void amd_iommu_init_notifier(void)
2118 bus_register_notifier(&pci_bus_type
, &device_nb
);
2121 /*****************************************************************************
2123 * The next functions belong to the dma_ops mapping/unmapping code.
2125 *****************************************************************************/
2128 * In the dma_ops path we only have the struct device. This function
2129 * finds the corresponding IOMMU, the protection domain and the
2130 * requestor id for a given device.
2131 * If the device is not yet associated with a domain this is also done
2134 static struct protection_domain
*get_domain(struct device
*dev
)
2136 struct protection_domain
*domain
;
2137 struct dma_ops_domain
*dma_dom
;
2138 u16 devid
= get_device_id(dev
);
2140 if (!check_device(dev
))
2141 return ERR_PTR(-EINVAL
);
2143 domain
= domain_for_device(dev
);
2144 if (domain
!= NULL
&& !dma_ops_domain(domain
))
2145 return ERR_PTR(-EBUSY
);
2150 /* Device not bount yet - bind it */
2151 dma_dom
= find_protection_domain(devid
);
2153 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
2154 attach_device(dev
, &dma_dom
->domain
);
2155 DUMP_printk("Using protection domain %d for device %s\n",
2156 dma_dom
->domain
.id
, dev_name(dev
));
2158 return &dma_dom
->domain
;
2161 static void update_device_table(struct protection_domain
*domain
)
2163 struct iommu_dev_data
*dev_data
;
2165 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
2166 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2169 static void update_domain(struct protection_domain
*domain
)
2171 if (!domain
->updated
)
2174 update_device_table(domain
);
2176 domain_flush_devices(domain
);
2177 domain_flush_tlb_pde(domain
);
2179 domain
->updated
= false;
2183 * This function fetches the PTE for a given address in the aperture
2185 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
2186 unsigned long address
)
2188 struct aperture_range
*aperture
;
2189 u64
*pte
, *pte_page
;
2191 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2195 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2197 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
2199 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
2201 pte
+= PM_LEVEL_INDEX(0, address
);
2203 update_domain(&dom
->domain
);
2209 * This is the generic map function. It maps one 4kb page at paddr to
2210 * the given address in the DMA address space for the domain.
2212 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2213 unsigned long address
,
2219 WARN_ON(address
> dom
->aperture_size
);
2223 pte
= dma_ops_get_pte(dom
, address
);
2225 return DMA_ERROR_CODE
;
2227 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2229 if (direction
== DMA_TO_DEVICE
)
2230 __pte
|= IOMMU_PTE_IR
;
2231 else if (direction
== DMA_FROM_DEVICE
)
2232 __pte
|= IOMMU_PTE_IW
;
2233 else if (direction
== DMA_BIDIRECTIONAL
)
2234 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2240 return (dma_addr_t
)address
;
2244 * The generic unmapping function for on page in the DMA address space.
2246 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2247 unsigned long address
)
2249 struct aperture_range
*aperture
;
2252 if (address
>= dom
->aperture_size
)
2255 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2259 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2263 pte
+= PM_LEVEL_INDEX(0, address
);
2271 * This function contains common code for mapping of a physically
2272 * contiguous memory region into DMA address space. It is used by all
2273 * mapping functions provided with this IOMMU driver.
2274 * Must be called with the domain lock held.
2276 static dma_addr_t
__map_single(struct device
*dev
,
2277 struct dma_ops_domain
*dma_dom
,
2284 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2285 dma_addr_t address
, start
, ret
;
2287 unsigned long align_mask
= 0;
2290 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2293 INC_STATS_COUNTER(total_map_requests
);
2296 INC_STATS_COUNTER(cross_page
);
2299 align_mask
= (1UL << get_order(size
)) - 1;
2302 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2304 if (unlikely(address
== DMA_ERROR_CODE
)) {
2306 * setting next_address here will let the address
2307 * allocator only scan the new allocated range in the
2308 * first run. This is a small optimization.
2310 dma_dom
->next_address
= dma_dom
->aperture_size
;
2312 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2316 * aperture was successfully enlarged by 128 MB, try
2323 for (i
= 0; i
< pages
; ++i
) {
2324 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2325 if (ret
== DMA_ERROR_CODE
)
2333 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2335 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2336 domain_flush_tlb(&dma_dom
->domain
);
2337 dma_dom
->need_flush
= false;
2338 } else if (unlikely(amd_iommu_np_cache
))
2339 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2346 for (--i
; i
>= 0; --i
) {
2348 dma_ops_domain_unmap(dma_dom
, start
);
2351 dma_ops_free_addresses(dma_dom
, address
, pages
);
2353 return DMA_ERROR_CODE
;
2357 * Does the reverse of the __map_single function. Must be called with
2358 * the domain lock held too
2360 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2361 dma_addr_t dma_addr
,
2365 dma_addr_t flush_addr
;
2366 dma_addr_t i
, start
;
2369 if ((dma_addr
== DMA_ERROR_CODE
) ||
2370 (dma_addr
+ size
> dma_dom
->aperture_size
))
2373 flush_addr
= dma_addr
;
2374 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2375 dma_addr
&= PAGE_MASK
;
2378 for (i
= 0; i
< pages
; ++i
) {
2379 dma_ops_domain_unmap(dma_dom
, start
);
2383 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2385 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2387 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2388 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2389 dma_dom
->need_flush
= false;
2394 * The exported map_single function for dma_ops.
2396 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2397 unsigned long offset
, size_t size
,
2398 enum dma_data_direction dir
,
2399 struct dma_attrs
*attrs
)
2401 unsigned long flags
;
2402 struct protection_domain
*domain
;
2405 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2407 INC_STATS_COUNTER(cnt_map_single
);
2409 domain
= get_domain(dev
);
2410 if (PTR_ERR(domain
) == -EINVAL
)
2411 return (dma_addr_t
)paddr
;
2412 else if (IS_ERR(domain
))
2413 return DMA_ERROR_CODE
;
2415 dma_mask
= *dev
->dma_mask
;
2417 spin_lock_irqsave(&domain
->lock
, flags
);
2419 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2421 if (addr
== DMA_ERROR_CODE
)
2424 domain_flush_complete(domain
);
2427 spin_unlock_irqrestore(&domain
->lock
, flags
);
2433 * The exported unmap_single function for dma_ops.
2435 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2436 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2438 unsigned long flags
;
2439 struct protection_domain
*domain
;
2441 INC_STATS_COUNTER(cnt_unmap_single
);
2443 domain
= get_domain(dev
);
2447 spin_lock_irqsave(&domain
->lock
, flags
);
2449 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2451 domain_flush_complete(domain
);
2453 spin_unlock_irqrestore(&domain
->lock
, flags
);
2457 * This is a special map_sg function which is used if we should map a
2458 * device which is not handled by an AMD IOMMU in the system.
2460 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2461 int nelems
, int dir
)
2463 struct scatterlist
*s
;
2466 for_each_sg(sglist
, s
, nelems
, i
) {
2467 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2468 s
->dma_length
= s
->length
;
2475 * The exported map_sg function for dma_ops (handles scatter-gather
2478 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2479 int nelems
, enum dma_data_direction dir
,
2480 struct dma_attrs
*attrs
)
2482 unsigned long flags
;
2483 struct protection_domain
*domain
;
2485 struct scatterlist
*s
;
2487 int mapped_elems
= 0;
2490 INC_STATS_COUNTER(cnt_map_sg
);
2492 domain
= get_domain(dev
);
2493 if (PTR_ERR(domain
) == -EINVAL
)
2494 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2495 else if (IS_ERR(domain
))
2498 dma_mask
= *dev
->dma_mask
;
2500 spin_lock_irqsave(&domain
->lock
, flags
);
2502 for_each_sg(sglist
, s
, nelems
, i
) {
2505 s
->dma_address
= __map_single(dev
, domain
->priv
,
2506 paddr
, s
->length
, dir
, false,
2509 if (s
->dma_address
) {
2510 s
->dma_length
= s
->length
;
2516 domain_flush_complete(domain
);
2519 spin_unlock_irqrestore(&domain
->lock
, flags
);
2521 return mapped_elems
;
2523 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2525 __unmap_single(domain
->priv
, s
->dma_address
,
2526 s
->dma_length
, dir
);
2527 s
->dma_address
= s
->dma_length
= 0;
2536 * The exported map_sg function for dma_ops (handles scatter-gather
2539 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2540 int nelems
, enum dma_data_direction dir
,
2541 struct dma_attrs
*attrs
)
2543 unsigned long flags
;
2544 struct protection_domain
*domain
;
2545 struct scatterlist
*s
;
2548 INC_STATS_COUNTER(cnt_unmap_sg
);
2550 domain
= get_domain(dev
);
2554 spin_lock_irqsave(&domain
->lock
, flags
);
2556 for_each_sg(sglist
, s
, nelems
, i
) {
2557 __unmap_single(domain
->priv
, s
->dma_address
,
2558 s
->dma_length
, dir
);
2559 s
->dma_address
= s
->dma_length
= 0;
2562 domain_flush_complete(domain
);
2564 spin_unlock_irqrestore(&domain
->lock
, flags
);
2568 * The exported alloc_coherent function for dma_ops.
2570 static void *alloc_coherent(struct device
*dev
, size_t size
,
2571 dma_addr_t
*dma_addr
, gfp_t flag
)
2573 unsigned long flags
;
2575 struct protection_domain
*domain
;
2577 u64 dma_mask
= dev
->coherent_dma_mask
;
2579 INC_STATS_COUNTER(cnt_alloc_coherent
);
2581 domain
= get_domain(dev
);
2582 if (PTR_ERR(domain
) == -EINVAL
) {
2583 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2584 *dma_addr
= __pa(virt_addr
);
2586 } else if (IS_ERR(domain
))
2589 dma_mask
= dev
->coherent_dma_mask
;
2590 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2593 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2597 paddr
= virt_to_phys(virt_addr
);
2600 dma_mask
= *dev
->dma_mask
;
2602 spin_lock_irqsave(&domain
->lock
, flags
);
2604 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2605 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2607 if (*dma_addr
== DMA_ERROR_CODE
) {
2608 spin_unlock_irqrestore(&domain
->lock
, flags
);
2612 domain_flush_complete(domain
);
2614 spin_unlock_irqrestore(&domain
->lock
, flags
);
2620 free_pages((unsigned long)virt_addr
, get_order(size
));
2626 * The exported free_coherent function for dma_ops.
2628 static void free_coherent(struct device
*dev
, size_t size
,
2629 void *virt_addr
, dma_addr_t dma_addr
)
2631 unsigned long flags
;
2632 struct protection_domain
*domain
;
2634 INC_STATS_COUNTER(cnt_free_coherent
);
2636 domain
= get_domain(dev
);
2640 spin_lock_irqsave(&domain
->lock
, flags
);
2642 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2644 domain_flush_complete(domain
);
2646 spin_unlock_irqrestore(&domain
->lock
, flags
);
2649 free_pages((unsigned long)virt_addr
, get_order(size
));
2653 * This function is called by the DMA layer to find out if we can handle a
2654 * particular device. It is part of the dma_ops.
2656 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2658 return check_device(dev
);
2662 * The function for pre-allocating protection domains.
2664 * If the driver core informs the DMA layer if a driver grabs a device
2665 * we don't need to preallocate the protection domains anymore.
2666 * For now we have to.
2668 static void prealloc_protection_domains(void)
2670 struct iommu_dev_data
*dev_data
;
2671 struct dma_ops_domain
*dma_dom
;
2672 struct pci_dev
*dev
= NULL
;
2675 for_each_pci_dev(dev
) {
2677 /* Do we handle this device? */
2678 if (!check_device(&dev
->dev
))
2681 dev_data
= get_dev_data(&dev
->dev
);
2682 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
2683 /* Make sure passthrough domain is allocated */
2684 alloc_passthrough_domain();
2685 dev_data
->passthrough
= true;
2686 attach_device(&dev
->dev
, pt_domain
);
2687 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2688 dev_name(&dev
->dev
));
2691 /* Is there already any domain for it? */
2692 if (domain_for_device(&dev
->dev
))
2695 devid
= get_device_id(&dev
->dev
);
2697 dma_dom
= dma_ops_domain_alloc();
2700 init_unity_mappings_for_device(dma_dom
, devid
);
2701 dma_dom
->target_dev
= devid
;
2703 attach_device(&dev
->dev
, &dma_dom
->domain
);
2705 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2709 static struct dma_map_ops amd_iommu_dma_ops
= {
2710 .alloc_coherent
= alloc_coherent
,
2711 .free_coherent
= free_coherent
,
2712 .map_page
= map_page
,
2713 .unmap_page
= unmap_page
,
2715 .unmap_sg
= unmap_sg
,
2716 .dma_supported
= amd_iommu_dma_supported
,
2719 static unsigned device_dma_ops_init(void)
2721 struct iommu_dev_data
*dev_data
;
2722 struct pci_dev
*pdev
= NULL
;
2723 unsigned unhandled
= 0;
2725 for_each_pci_dev(pdev
) {
2726 if (!check_device(&pdev
->dev
)) {
2731 dev_data
= get_dev_data(&pdev
->dev
);
2733 if (!dev_data
->passthrough
)
2734 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
2736 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
2743 * The function which clues the AMD IOMMU driver into dma_ops.
2746 void __init
amd_iommu_init_api(void)
2748 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2751 int __init
amd_iommu_init_dma_ops(void)
2753 struct amd_iommu
*iommu
;
2757 * first allocate a default protection domain for every IOMMU we
2758 * found in the system. Devices not assigned to any other
2759 * protection domain will be assigned to the default one.
2761 for_each_iommu(iommu
) {
2762 iommu
->default_dom
= dma_ops_domain_alloc();
2763 if (iommu
->default_dom
== NULL
)
2765 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
2766 ret
= iommu_init_unity_mappings(iommu
);
2772 * Pre-allocate the protection domains for each device.
2774 prealloc_protection_domains();
2779 /* Make the driver finally visible to the drivers */
2780 unhandled
= device_dma_ops_init();
2781 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
2782 /* There are unhandled devices - initialize swiotlb for them */
2786 amd_iommu_stats_init();
2792 for_each_iommu(iommu
) {
2793 if (iommu
->default_dom
)
2794 dma_ops_domain_free(iommu
->default_dom
);
2800 /*****************************************************************************
2802 * The following functions belong to the exported interface of AMD IOMMU
2804 * This interface allows access to lower level functions of the IOMMU
2805 * like protection domain handling and assignement of devices to domains
2806 * which is not possible with the dma_ops interface.
2808 *****************************************************************************/
2810 static void cleanup_domain(struct protection_domain
*domain
)
2812 struct iommu_dev_data
*dev_data
, *next
;
2813 unsigned long flags
;
2815 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2817 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
2818 __detach_device(dev_data
);
2819 atomic_set(&dev_data
->bind
, 0);
2822 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2825 static void protection_domain_free(struct protection_domain
*domain
)
2830 del_domain_from_list(domain
);
2833 domain_id_free(domain
->id
);
2838 static struct protection_domain
*protection_domain_alloc(void)
2840 struct protection_domain
*domain
;
2842 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
2846 spin_lock_init(&domain
->lock
);
2847 mutex_init(&domain
->api_lock
);
2848 domain
->id
= domain_id_alloc();
2851 INIT_LIST_HEAD(&domain
->dev_list
);
2853 add_domain_to_list(domain
);
2863 static int __init
alloc_passthrough_domain(void)
2865 if (pt_domain
!= NULL
)
2868 /* allocate passthrough domain */
2869 pt_domain
= protection_domain_alloc();
2873 pt_domain
->mode
= PAGE_MODE_NONE
;
2877 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
2879 struct protection_domain
*domain
;
2881 domain
= protection_domain_alloc();
2885 domain
->mode
= PAGE_MODE_3_LEVEL
;
2886 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
2887 if (!domain
->pt_root
)
2895 protection_domain_free(domain
);
2900 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
2902 struct protection_domain
*domain
= dom
->priv
;
2907 if (domain
->dev_cnt
> 0)
2908 cleanup_domain(domain
);
2910 BUG_ON(domain
->dev_cnt
!= 0);
2912 if (domain
->mode
!= PAGE_MODE_NONE
)
2913 free_pagetable(domain
);
2915 if (domain
->flags
& PD_IOMMUV2_MASK
)
2916 free_gcr3_table(domain
);
2918 protection_domain_free(domain
);
2923 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
2926 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
2927 struct amd_iommu
*iommu
;
2930 if (!check_device(dev
))
2933 devid
= get_device_id(dev
);
2935 if (dev_data
->domain
!= NULL
)
2938 iommu
= amd_iommu_rlookup_table
[devid
];
2942 iommu_completion_wait(iommu
);
2945 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
2948 struct protection_domain
*domain
= dom
->priv
;
2949 struct iommu_dev_data
*dev_data
;
2950 struct amd_iommu
*iommu
;
2953 if (!check_device(dev
))
2956 dev_data
= dev
->archdata
.iommu
;
2958 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2962 if (dev_data
->domain
)
2965 ret
= attach_device(dev
, domain
);
2967 iommu_completion_wait(iommu
);
2972 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
2973 phys_addr_t paddr
, int gfp_order
, int iommu_prot
)
2975 unsigned long page_size
= 0x1000UL
<< gfp_order
;
2976 struct protection_domain
*domain
= dom
->priv
;
2980 if (domain
->mode
== PAGE_MODE_NONE
)
2983 if (iommu_prot
& IOMMU_READ
)
2984 prot
|= IOMMU_PROT_IR
;
2985 if (iommu_prot
& IOMMU_WRITE
)
2986 prot
|= IOMMU_PROT_IW
;
2988 mutex_lock(&domain
->api_lock
);
2989 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
2990 mutex_unlock(&domain
->api_lock
);
2995 static int amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
2998 struct protection_domain
*domain
= dom
->priv
;
2999 unsigned long page_size
, unmap_size
;
3001 if (domain
->mode
== PAGE_MODE_NONE
)
3004 page_size
= 0x1000UL
<< gfp_order
;
3006 mutex_lock(&domain
->api_lock
);
3007 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3008 mutex_unlock(&domain
->api_lock
);
3010 domain_flush_tlb_pde(domain
);
3012 return get_order(unmap_size
);
3015 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3018 struct protection_domain
*domain
= dom
->priv
;
3019 unsigned long offset_mask
;
3023 if (domain
->mode
== PAGE_MODE_NONE
)
3026 pte
= fetch_pte(domain
, iova
);
3028 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3031 if (PM_PTE_LEVEL(*pte
) == 0)
3032 offset_mask
= PAGE_SIZE
- 1;
3034 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
3036 __pte
= *pte
& PM_ADDR_MASK
;
3037 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3042 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
3046 case IOMMU_CAP_CACHE_COHERENCY
:
3053 static struct iommu_ops amd_iommu_ops
= {
3054 .domain_init
= amd_iommu_domain_init
,
3055 .domain_destroy
= amd_iommu_domain_destroy
,
3056 .attach_dev
= amd_iommu_attach_device
,
3057 .detach_dev
= amd_iommu_detach_device
,
3058 .map
= amd_iommu_map
,
3059 .unmap
= amd_iommu_unmap
,
3060 .iova_to_phys
= amd_iommu_iova_to_phys
,
3061 .domain_has_cap
= amd_iommu_domain_has_cap
,
3064 /*****************************************************************************
3066 * The next functions do a basic initialization of IOMMU for pass through
3069 * In passthrough mode the IOMMU is initialized and enabled but not used for
3070 * DMA-API translation.
3072 *****************************************************************************/
3074 int __init
amd_iommu_init_passthrough(void)
3076 struct iommu_dev_data
*dev_data
;
3077 struct pci_dev
*dev
= NULL
;
3078 struct amd_iommu
*iommu
;
3082 ret
= alloc_passthrough_domain();
3086 for_each_pci_dev(dev
) {
3087 if (!check_device(&dev
->dev
))
3090 dev_data
= get_dev_data(&dev
->dev
);
3091 dev_data
->passthrough
= true;
3093 devid
= get_device_id(&dev
->dev
);
3095 iommu
= amd_iommu_rlookup_table
[devid
];
3099 attach_device(&dev
->dev
, pt_domain
);
3102 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3107 /* IOMMUv2 specific functions */
3108 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3110 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3112 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3114 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3116 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3118 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3120 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3122 struct protection_domain
*domain
= dom
->priv
;
3123 unsigned long flags
;
3125 spin_lock_irqsave(&domain
->lock
, flags
);
3127 /* Update data structure */
3128 domain
->mode
= PAGE_MODE_NONE
;
3129 domain
->updated
= true;
3131 /* Make changes visible to IOMMUs */
3132 update_domain(domain
);
3134 /* Page-table is not visible to IOMMU anymore, so free it */
3135 free_pagetable(domain
);
3137 spin_unlock_irqrestore(&domain
->lock
, flags
);
3139 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3141 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3143 struct protection_domain
*domain
= dom
->priv
;
3144 unsigned long flags
;
3147 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3150 /* Number of GCR3 table levels required */
3151 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3154 if (levels
> amd_iommu_max_glx_val
)
3157 spin_lock_irqsave(&domain
->lock
, flags
);
3160 * Save us all sanity checks whether devices already in the
3161 * domain support IOMMUv2. Just force that the domain has no
3162 * devices attached when it is switched into IOMMUv2 mode.
3165 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3169 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3170 if (domain
->gcr3_tbl
== NULL
)
3173 domain
->glx
= levels
;
3174 domain
->flags
|= PD_IOMMUV2_MASK
;
3175 domain
->updated
= true;
3177 update_domain(domain
);
3182 spin_unlock_irqrestore(&domain
->lock
, flags
);
3186 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3188 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3189 u64 address
, bool size
)
3191 struct iommu_dev_data
*dev_data
;
3192 struct iommu_cmd cmd
;
3195 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3198 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3201 * IOMMU TLB needs to be flushed before Device TLB to
3202 * prevent device TLB refill from IOMMU TLB
3204 for (i
= 0; i
< amd_iommus_present
; ++i
) {
3205 if (domain
->dev_iommu
[i
] == 0)
3208 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3213 /* Wait until IOMMU TLB flushes are complete */
3214 domain_flush_complete(domain
);
3216 /* Now flush device TLBs */
3217 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3218 struct amd_iommu
*iommu
;
3221 BUG_ON(!dev_data
->ats
.enabled
);
3223 qdep
= dev_data
->ats
.qdep
;
3224 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3226 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3227 qdep
, address
, size
);
3229 ret
= iommu_queue_command(iommu
, &cmd
);
3234 /* Wait until all device TLBs are flushed */
3235 domain_flush_complete(domain
);
3244 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3247 return __flush_pasid(domain
, pasid
, address
, false);
3250 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3253 struct protection_domain
*domain
= dom
->priv
;
3254 unsigned long flags
;
3257 spin_lock_irqsave(&domain
->lock
, flags
);
3258 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3259 spin_unlock_irqrestore(&domain
->lock
, flags
);
3263 EXPORT_SYMBOL(amd_iommu_flush_page
);
3265 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3267 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3271 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3273 struct protection_domain
*domain
= dom
->priv
;
3274 unsigned long flags
;
3277 spin_lock_irqsave(&domain
->lock
, flags
);
3278 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3279 spin_unlock_irqrestore(&domain
->lock
, flags
);
3283 EXPORT_SYMBOL(amd_iommu_flush_tlb
);