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
48 * This bitmap is used to advertise the page sizes our hardware support
49 * to the IOMMU core, which will then use this information to split
50 * physically contiguous memory regions it is mapping into page sizes
53 * Traditionally the IOMMU core just handed us the mappings directly,
54 * after making sure the size is an order of a 4KiB page and that the
55 * mapping has natural alignment.
57 * To retain this behavior, we currently advertise that we support
58 * all page sizes that are an order of 4KiB.
60 * If at some point we'd like to utilize the IOMMU core's new behavior,
61 * we could change this to advertise the real page sizes we support.
63 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
65 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
67 /* A list of preallocated protection domains */
68 static LIST_HEAD(iommu_pd_list
);
69 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
71 /* List of all available dev_data structures */
72 static LIST_HEAD(dev_data_list
);
73 static DEFINE_SPINLOCK(dev_data_list_lock
);
76 * Domain for untranslated devices - only allocated
77 * if iommu=pt passed on kernel cmd line.
79 static struct protection_domain
*pt_domain
;
81 static struct iommu_ops amd_iommu_ops
;
83 static ATOMIC_NOTIFIER_HEAD(ppr_notifier
);
84 int amd_iommu_max_glx_val
= -1;
86 static struct dma_map_ops amd_iommu_dma_ops
;
89 * general struct to manage commands send to an IOMMU
95 static void update_domain(struct protection_domain
*domain
);
96 static int __init
alloc_passthrough_domain(void);
98 /****************************************************************************
102 ****************************************************************************/
104 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
106 struct iommu_dev_data
*dev_data
;
109 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
113 dev_data
->devid
= devid
;
114 atomic_set(&dev_data
->bind
, 0);
116 spin_lock_irqsave(&dev_data_list_lock
, flags
);
117 list_add_tail(&dev_data
->dev_data_list
, &dev_data_list
);
118 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
123 static void free_dev_data(struct iommu_dev_data
*dev_data
)
127 spin_lock_irqsave(&dev_data_list_lock
, flags
);
128 list_del(&dev_data
->dev_data_list
);
129 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
134 static struct iommu_dev_data
*search_dev_data(u16 devid
)
136 struct iommu_dev_data
*dev_data
;
139 spin_lock_irqsave(&dev_data_list_lock
, flags
);
140 list_for_each_entry(dev_data
, &dev_data_list
, dev_data_list
) {
141 if (dev_data
->devid
== devid
)
148 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
153 static struct iommu_dev_data
*find_dev_data(u16 devid
)
155 struct iommu_dev_data
*dev_data
;
157 dev_data
= search_dev_data(devid
);
159 if (dev_data
== NULL
)
160 dev_data
= alloc_dev_data(devid
);
165 static inline u16
get_device_id(struct device
*dev
)
167 struct pci_dev
*pdev
= to_pci_dev(dev
);
169 return calc_devid(pdev
->bus
->number
, pdev
->devfn
);
172 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
174 return dev
->archdata
.iommu
;
177 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
179 static const int caps
[] = {
182 PCI_EXT_CAP_ID_PASID
,
186 for (i
= 0; i
< 3; ++i
) {
187 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
195 static bool pdev_pri_erratum(struct pci_dev
*pdev
, u32 erratum
)
197 struct iommu_dev_data
*dev_data
;
199 dev_data
= get_dev_data(&pdev
->dev
);
201 return dev_data
->errata
& (1 << erratum
) ? true : false;
205 * In this function the list of preallocated protection domains is traversed to
206 * find the domain for a specific device
208 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
210 struct dma_ops_domain
*entry
, *ret
= NULL
;
212 u16 alias
= amd_iommu_alias_table
[devid
];
214 if (list_empty(&iommu_pd_list
))
217 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
219 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
220 if (entry
->target_dev
== devid
||
221 entry
->target_dev
== alias
) {
227 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
233 * This function checks if the driver got a valid device from the caller to
234 * avoid dereferencing invalid pointers.
236 static bool check_device(struct device
*dev
)
240 if (!dev
|| !dev
->dma_mask
)
243 /* No device or no PCI device */
244 if (dev
->bus
!= &pci_bus_type
)
247 devid
= get_device_id(dev
);
249 /* Out of our scope? */
250 if (devid
> amd_iommu_last_bdf
)
253 if (amd_iommu_rlookup_table
[devid
] == NULL
)
259 static void swap_pci_ref(struct pci_dev
**from
, struct pci_dev
*to
)
265 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
267 static int iommu_init_device(struct device
*dev
)
269 struct pci_dev
*dma_pdev
, *pdev
= to_pci_dev(dev
);
270 struct iommu_dev_data
*dev_data
;
271 struct iommu_group
*group
;
275 if (dev
->archdata
.iommu
)
278 dev_data
= find_dev_data(get_device_id(dev
));
282 alias
= amd_iommu_alias_table
[dev_data
->devid
];
283 if (alias
!= dev_data
->devid
) {
284 struct iommu_dev_data
*alias_data
;
286 alias_data
= find_dev_data(alias
);
287 if (alias_data
== NULL
) {
288 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
290 free_dev_data(dev_data
);
293 dev_data
->alias_data
= alias_data
;
295 dma_pdev
= pci_get_bus_and_slot(alias
>> 8, alias
& 0xff);
297 dma_pdev
= pci_dev_get(pdev
);
299 swap_pci_ref(&dma_pdev
, pci_get_dma_source(dma_pdev
));
301 if (dma_pdev
->multifunction
&&
302 !pci_acs_enabled(dma_pdev
, REQ_ACS_FLAGS
))
303 swap_pci_ref(&dma_pdev
,
304 pci_get_slot(dma_pdev
->bus
,
305 PCI_DEVFN(PCI_SLOT(dma_pdev
->devfn
),
308 while (!pci_is_root_bus(dma_pdev
->bus
)) {
309 if (pci_acs_path_enabled(dma_pdev
->bus
->self
,
310 NULL
, REQ_ACS_FLAGS
))
313 swap_pci_ref(&dma_pdev
, pci_dev_get(dma_pdev
->bus
->self
));
316 group
= iommu_group_get(&dma_pdev
->dev
);
317 pci_dev_put(dma_pdev
);
319 group
= iommu_group_alloc();
321 return PTR_ERR(group
);
324 ret
= iommu_group_add_device(group
, dev
);
326 iommu_group_put(group
);
331 if (pci_iommuv2_capable(pdev
)) {
332 struct amd_iommu
*iommu
;
334 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
335 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
338 dev
->archdata
.iommu
= dev_data
;
343 static void iommu_ignore_device(struct device
*dev
)
347 devid
= get_device_id(dev
);
348 alias
= amd_iommu_alias_table
[devid
];
350 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
351 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
353 amd_iommu_rlookup_table
[devid
] = NULL
;
354 amd_iommu_rlookup_table
[alias
] = NULL
;
357 static void iommu_uninit_device(struct device
*dev
)
359 iommu_group_remove_device(dev
);
362 * Nothing to do here - we keep dev_data around for unplugged devices
363 * and reuse it when the device is re-plugged - not doing so would
364 * introduce a ton of races.
368 void __init
amd_iommu_uninit_devices(void)
370 struct iommu_dev_data
*dev_data
, *n
;
371 struct pci_dev
*pdev
= NULL
;
373 for_each_pci_dev(pdev
) {
375 if (!check_device(&pdev
->dev
))
378 iommu_uninit_device(&pdev
->dev
);
381 /* Free all of our dev_data structures */
382 list_for_each_entry_safe(dev_data
, n
, &dev_data_list
, dev_data_list
)
383 free_dev_data(dev_data
);
386 int __init
amd_iommu_init_devices(void)
388 struct pci_dev
*pdev
= NULL
;
391 for_each_pci_dev(pdev
) {
393 if (!check_device(&pdev
->dev
))
396 ret
= iommu_init_device(&pdev
->dev
);
397 if (ret
== -ENOTSUPP
)
398 iommu_ignore_device(&pdev
->dev
);
407 amd_iommu_uninit_devices();
411 #ifdef CONFIG_AMD_IOMMU_STATS
414 * Initialization code for statistics collection
417 DECLARE_STATS_COUNTER(compl_wait
);
418 DECLARE_STATS_COUNTER(cnt_map_single
);
419 DECLARE_STATS_COUNTER(cnt_unmap_single
);
420 DECLARE_STATS_COUNTER(cnt_map_sg
);
421 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
422 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
423 DECLARE_STATS_COUNTER(cnt_free_coherent
);
424 DECLARE_STATS_COUNTER(cross_page
);
425 DECLARE_STATS_COUNTER(domain_flush_single
);
426 DECLARE_STATS_COUNTER(domain_flush_all
);
427 DECLARE_STATS_COUNTER(alloced_io_mem
);
428 DECLARE_STATS_COUNTER(total_map_requests
);
429 DECLARE_STATS_COUNTER(complete_ppr
);
430 DECLARE_STATS_COUNTER(invalidate_iotlb
);
431 DECLARE_STATS_COUNTER(invalidate_iotlb_all
);
432 DECLARE_STATS_COUNTER(pri_requests
);
434 static struct dentry
*stats_dir
;
435 static struct dentry
*de_fflush
;
437 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
439 if (stats_dir
== NULL
)
442 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
446 static void amd_iommu_stats_init(void)
448 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
449 if (stats_dir
== NULL
)
452 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
453 &amd_iommu_unmap_flush
);
455 amd_iommu_stats_add(&compl_wait
);
456 amd_iommu_stats_add(&cnt_map_single
);
457 amd_iommu_stats_add(&cnt_unmap_single
);
458 amd_iommu_stats_add(&cnt_map_sg
);
459 amd_iommu_stats_add(&cnt_unmap_sg
);
460 amd_iommu_stats_add(&cnt_alloc_coherent
);
461 amd_iommu_stats_add(&cnt_free_coherent
);
462 amd_iommu_stats_add(&cross_page
);
463 amd_iommu_stats_add(&domain_flush_single
);
464 amd_iommu_stats_add(&domain_flush_all
);
465 amd_iommu_stats_add(&alloced_io_mem
);
466 amd_iommu_stats_add(&total_map_requests
);
467 amd_iommu_stats_add(&complete_ppr
);
468 amd_iommu_stats_add(&invalidate_iotlb
);
469 amd_iommu_stats_add(&invalidate_iotlb_all
);
470 amd_iommu_stats_add(&pri_requests
);
475 /****************************************************************************
477 * Interrupt handling functions
479 ****************************************************************************/
481 static void dump_dte_entry(u16 devid
)
485 for (i
= 0; i
< 4; ++i
)
486 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
487 amd_iommu_dev_table
[devid
].data
[i
]);
490 static void dump_command(unsigned long phys_addr
)
492 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
495 for (i
= 0; i
< 4; ++i
)
496 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
499 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
501 int type
, devid
, domid
, flags
;
502 volatile u32
*event
= __evt
;
507 type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
508 devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
509 domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
510 flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
511 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
514 /* Did we hit the erratum? */
515 if (++count
== LOOP_TIMEOUT
) {
516 pr_err("AMD-Vi: No event written to event log\n");
523 printk(KERN_ERR
"AMD-Vi: Event logged [");
526 case EVENT_TYPE_ILL_DEV
:
527 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
528 "address=0x%016llx flags=0x%04x]\n",
529 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
531 dump_dte_entry(devid
);
533 case EVENT_TYPE_IO_FAULT
:
534 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
535 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
536 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
537 domid
, address
, flags
);
539 case EVENT_TYPE_DEV_TAB_ERR
:
540 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
541 "address=0x%016llx flags=0x%04x]\n",
542 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
545 case EVENT_TYPE_PAGE_TAB_ERR
:
546 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
547 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
548 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
549 domid
, address
, flags
);
551 case EVENT_TYPE_ILL_CMD
:
552 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
553 dump_command(address
);
555 case EVENT_TYPE_CMD_HARD_ERR
:
556 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
557 "flags=0x%04x]\n", address
, flags
);
559 case EVENT_TYPE_IOTLB_INV_TO
:
560 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
561 "address=0x%016llx]\n",
562 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
565 case EVENT_TYPE_INV_DEV_REQ
:
566 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
567 "address=0x%016llx flags=0x%04x]\n",
568 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
572 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
575 memset(__evt
, 0, 4 * sizeof(u32
));
578 static void iommu_poll_events(struct amd_iommu
*iommu
)
583 spin_lock_irqsave(&iommu
->lock
, flags
);
585 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
586 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
588 while (head
!= tail
) {
589 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
590 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
593 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
595 spin_unlock_irqrestore(&iommu
->lock
, flags
);
598 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u64
*raw
)
600 struct amd_iommu_fault fault
;
602 INC_STATS_COUNTER(pri_requests
);
604 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
605 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
609 fault
.address
= raw
[1];
610 fault
.pasid
= PPR_PASID(raw
[0]);
611 fault
.device_id
= PPR_DEVID(raw
[0]);
612 fault
.tag
= PPR_TAG(raw
[0]);
613 fault
.flags
= PPR_FLAGS(raw
[0]);
615 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
618 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
623 if (iommu
->ppr_log
== NULL
)
626 /* enable ppr interrupts again */
627 writel(MMIO_STATUS_PPR_INT_MASK
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
629 spin_lock_irqsave(&iommu
->lock
, flags
);
631 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
632 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
634 while (head
!= tail
) {
639 raw
= (u64
*)(iommu
->ppr_log
+ head
);
642 * Hardware bug: Interrupt may arrive before the entry is
643 * written to memory. If this happens we need to wait for the
646 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
647 if (PPR_REQ_TYPE(raw
[0]) != 0)
652 /* Avoid memcpy function-call overhead */
657 * To detect the hardware bug we need to clear the entry
660 raw
[0] = raw
[1] = 0UL;
662 /* Update head pointer of hardware ring-buffer */
663 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
664 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
667 * Release iommu->lock because ppr-handling might need to
670 spin_unlock_irqrestore(&iommu
->lock
, flags
);
672 /* Handle PPR entry */
673 iommu_handle_ppr_entry(iommu
, entry
);
675 spin_lock_irqsave(&iommu
->lock
, flags
);
677 /* Refresh ring-buffer information */
678 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
679 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
682 spin_unlock_irqrestore(&iommu
->lock
, flags
);
685 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
687 struct amd_iommu
*iommu
;
689 for_each_iommu(iommu
) {
690 iommu_poll_events(iommu
);
691 iommu_poll_ppr_log(iommu
);
697 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
699 return IRQ_WAKE_THREAD
;
702 /****************************************************************************
704 * IOMMU command queuing functions
706 ****************************************************************************/
708 static int wait_on_sem(volatile u64
*sem
)
712 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
717 if (i
== LOOP_TIMEOUT
) {
718 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
725 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
726 struct iommu_cmd
*cmd
,
731 target
= iommu
->cmd_buf
+ tail
;
732 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
734 /* Copy command to buffer */
735 memcpy(target
, cmd
, sizeof(*cmd
));
737 /* Tell the IOMMU about it */
738 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
741 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
743 WARN_ON(address
& 0x7ULL
);
745 memset(cmd
, 0, sizeof(*cmd
));
746 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
747 cmd
->data
[1] = upper_32_bits(__pa(address
));
749 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
752 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
754 memset(cmd
, 0, sizeof(*cmd
));
755 cmd
->data
[0] = devid
;
756 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
759 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
760 size_t size
, u16 domid
, int pde
)
765 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
770 * If we have to flush more than one page, flush all
771 * TLB entries for this domain
773 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
777 address
&= PAGE_MASK
;
779 memset(cmd
, 0, sizeof(*cmd
));
780 cmd
->data
[1] |= domid
;
781 cmd
->data
[2] = lower_32_bits(address
);
782 cmd
->data
[3] = upper_32_bits(address
);
783 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
784 if (s
) /* size bit - we flush more than one 4kb page */
785 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
786 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
787 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
790 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
791 u64 address
, size_t size
)
796 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
801 * If we have to flush more than one page, flush all
802 * TLB entries for this domain
804 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
808 address
&= PAGE_MASK
;
810 memset(cmd
, 0, sizeof(*cmd
));
811 cmd
->data
[0] = devid
;
812 cmd
->data
[0] |= (qdep
& 0xff) << 24;
813 cmd
->data
[1] = devid
;
814 cmd
->data
[2] = lower_32_bits(address
);
815 cmd
->data
[3] = upper_32_bits(address
);
816 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
818 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
821 static void build_inv_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
822 u64 address
, bool size
)
824 memset(cmd
, 0, sizeof(*cmd
));
826 address
&= ~(0xfffULL
);
828 cmd
->data
[0] = pasid
& PASID_MASK
;
829 cmd
->data
[1] = domid
;
830 cmd
->data
[2] = lower_32_bits(address
);
831 cmd
->data
[3] = upper_32_bits(address
);
832 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
833 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
835 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
836 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
839 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
840 int qdep
, u64 address
, bool size
)
842 memset(cmd
, 0, sizeof(*cmd
));
844 address
&= ~(0xfffULL
);
846 cmd
->data
[0] = devid
;
847 cmd
->data
[0] |= (pasid
& 0xff) << 16;
848 cmd
->data
[0] |= (qdep
& 0xff) << 24;
849 cmd
->data
[1] = devid
;
850 cmd
->data
[1] |= ((pasid
>> 8) & 0xfff) << 16;
851 cmd
->data
[2] = lower_32_bits(address
);
852 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
853 cmd
->data
[3] = upper_32_bits(address
);
855 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
856 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
859 static void build_complete_ppr(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
860 int status
, int tag
, bool gn
)
862 memset(cmd
, 0, sizeof(*cmd
));
864 cmd
->data
[0] = devid
;
866 cmd
->data
[1] = pasid
& PASID_MASK
;
867 cmd
->data
[2] = CMD_INV_IOMMU_PAGES_GN_MASK
;
869 cmd
->data
[3] = tag
& 0x1ff;
870 cmd
->data
[3] |= (status
& PPR_STATUS_MASK
) << PPR_STATUS_SHIFT
;
872 CMD_SET_TYPE(cmd
, CMD_COMPLETE_PPR
);
875 static void build_inv_all(struct iommu_cmd
*cmd
)
877 memset(cmd
, 0, sizeof(*cmd
));
878 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
882 * Writes the command to the IOMMUs command buffer and informs the
883 * hardware about the new command.
885 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
886 struct iommu_cmd
*cmd
,
889 u32 left
, tail
, head
, next_tail
;
892 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
895 spin_lock_irqsave(&iommu
->lock
, flags
);
897 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
898 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
899 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
900 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
903 struct iommu_cmd sync_cmd
;
904 volatile u64 sem
= 0;
907 build_completion_wait(&sync_cmd
, (u64
)&sem
);
908 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
910 spin_unlock_irqrestore(&iommu
->lock
, flags
);
912 if ((ret
= wait_on_sem(&sem
)) != 0)
918 copy_cmd_to_buffer(iommu
, cmd
, tail
);
920 /* We need to sync now to make sure all commands are processed */
921 iommu
->need_sync
= sync
;
923 spin_unlock_irqrestore(&iommu
->lock
, flags
);
928 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
930 return iommu_queue_command_sync(iommu
, cmd
, true);
934 * This function queues a completion wait command into the command
937 static int iommu_completion_wait(struct amd_iommu
*iommu
)
939 struct iommu_cmd cmd
;
940 volatile u64 sem
= 0;
943 if (!iommu
->need_sync
)
946 build_completion_wait(&cmd
, (u64
)&sem
);
948 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
952 return wait_on_sem(&sem
);
955 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
957 struct iommu_cmd cmd
;
959 build_inv_dte(&cmd
, devid
);
961 return iommu_queue_command(iommu
, &cmd
);
964 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
968 for (devid
= 0; devid
<= 0xffff; ++devid
)
969 iommu_flush_dte(iommu
, devid
);
971 iommu_completion_wait(iommu
);
975 * This function uses heavy locking and may disable irqs for some time. But
976 * this is no issue because it is only called during resume.
978 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
982 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
983 struct iommu_cmd cmd
;
984 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
986 iommu_queue_command(iommu
, &cmd
);
989 iommu_completion_wait(iommu
);
992 static void iommu_flush_all(struct amd_iommu
*iommu
)
994 struct iommu_cmd cmd
;
998 iommu_queue_command(iommu
, &cmd
);
999 iommu_completion_wait(iommu
);
1002 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
1004 if (iommu_feature(iommu
, FEATURE_IA
)) {
1005 iommu_flush_all(iommu
);
1007 iommu_flush_dte_all(iommu
);
1008 iommu_flush_tlb_all(iommu
);
1013 * Command send function for flushing on-device TLB
1015 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
1016 u64 address
, size_t size
)
1018 struct amd_iommu
*iommu
;
1019 struct iommu_cmd cmd
;
1022 qdep
= dev_data
->ats
.qdep
;
1023 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1025 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
1027 return iommu_queue_command(iommu
, &cmd
);
1031 * Command send function for invalidating a device table entry
1033 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
1035 struct amd_iommu
*iommu
;
1038 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1040 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
1044 if (dev_data
->ats
.enabled
)
1045 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
1051 * TLB invalidation function which is called from the mapping functions.
1052 * It invalidates a single PTE if the range to flush is within a single
1053 * page. Otherwise it flushes the whole TLB of the IOMMU.
1055 static void __domain_flush_pages(struct protection_domain
*domain
,
1056 u64 address
, size_t size
, int pde
)
1058 struct iommu_dev_data
*dev_data
;
1059 struct iommu_cmd cmd
;
1062 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
1064 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1065 if (!domain
->dev_iommu
[i
])
1069 * Devices of this domain are behind this IOMMU
1070 * We need a TLB flush
1072 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
1075 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1077 if (!dev_data
->ats
.enabled
)
1080 ret
|= device_flush_iotlb(dev_data
, address
, size
);
1086 static void domain_flush_pages(struct protection_domain
*domain
,
1087 u64 address
, size_t size
)
1089 __domain_flush_pages(domain
, address
, size
, 0);
1092 /* Flush the whole IO/TLB for a given protection domain */
1093 static void domain_flush_tlb(struct protection_domain
*domain
)
1095 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
1098 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1099 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
1101 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
1104 static void domain_flush_complete(struct protection_domain
*domain
)
1108 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1109 if (!domain
->dev_iommu
[i
])
1113 * Devices of this domain are behind this IOMMU
1114 * We need to wait for completion of all commands.
1116 iommu_completion_wait(amd_iommus
[i
]);
1122 * This function flushes the DTEs for all devices in domain
1124 static void domain_flush_devices(struct protection_domain
*domain
)
1126 struct iommu_dev_data
*dev_data
;
1128 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1129 device_flush_dte(dev_data
);
1132 /****************************************************************************
1134 * The functions below are used the create the page table mappings for
1135 * unity mapped regions.
1137 ****************************************************************************/
1140 * This function is used to add another level to an IO page table. Adding
1141 * another level increases the size of the address space by 9 bits to a size up
1144 static bool increase_address_space(struct protection_domain
*domain
,
1149 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1150 /* address space already 64 bit large */
1153 pte
= (void *)get_zeroed_page(gfp
);
1157 *pte
= PM_LEVEL_PDE(domain
->mode
,
1158 virt_to_phys(domain
->pt_root
));
1159 domain
->pt_root
= pte
;
1161 domain
->updated
= true;
1166 static u64
*alloc_pte(struct protection_domain
*domain
,
1167 unsigned long address
,
1168 unsigned long page_size
,
1175 BUG_ON(!is_power_of_2(page_size
));
1177 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1178 increase_address_space(domain
, gfp
);
1180 level
= domain
->mode
- 1;
1181 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1182 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1183 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1185 while (level
> end_lvl
) {
1186 if (!IOMMU_PTE_PRESENT(*pte
)) {
1187 page
= (u64
*)get_zeroed_page(gfp
);
1190 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1193 /* No level skipping support yet */
1194 if (PM_PTE_LEVEL(*pte
) != level
)
1199 pte
= IOMMU_PTE_PAGE(*pte
);
1201 if (pte_page
&& level
== end_lvl
)
1204 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1211 * This function checks if there is a PTE for a given dma address. If
1212 * there is one, it returns the pointer to it.
1214 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
1219 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1222 level
= domain
->mode
- 1;
1223 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1228 if (!IOMMU_PTE_PRESENT(*pte
))
1232 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1233 unsigned long pte_mask
, __pte
;
1236 * If we have a series of large PTEs, make
1237 * sure to return a pointer to the first one.
1239 pte_mask
= PTE_PAGE_SIZE(*pte
);
1240 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1241 __pte
= ((unsigned long)pte
) & pte_mask
;
1243 return (u64
*)__pte
;
1246 /* No level skipping support yet */
1247 if (PM_PTE_LEVEL(*pte
) != level
)
1252 /* Walk to the next level */
1253 pte
= IOMMU_PTE_PAGE(*pte
);
1254 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1261 * Generic mapping functions. It maps a physical address into a DMA
1262 * address space. It allocates the page table pages if necessary.
1263 * In the future it can be extended to a generic mapping function
1264 * supporting all features of AMD IOMMU page tables like level skipping
1265 * and full 64 bit address spaces.
1267 static int iommu_map_page(struct protection_domain
*dom
,
1268 unsigned long bus_addr
,
1269 unsigned long phys_addr
,
1271 unsigned long page_size
)
1276 if (!(prot
& IOMMU_PROT_MASK
))
1279 bus_addr
= PAGE_ALIGN(bus_addr
);
1280 phys_addr
= PAGE_ALIGN(phys_addr
);
1281 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1282 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1284 for (i
= 0; i
< count
; ++i
)
1285 if (IOMMU_PTE_PRESENT(pte
[i
]))
1288 if (page_size
> PAGE_SIZE
) {
1289 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1290 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1292 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1294 if (prot
& IOMMU_PROT_IR
)
1295 __pte
|= IOMMU_PTE_IR
;
1296 if (prot
& IOMMU_PROT_IW
)
1297 __pte
|= IOMMU_PTE_IW
;
1299 for (i
= 0; i
< count
; ++i
)
1307 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1308 unsigned long bus_addr
,
1309 unsigned long page_size
)
1311 unsigned long long unmap_size
, unmapped
;
1314 BUG_ON(!is_power_of_2(page_size
));
1318 while (unmapped
< page_size
) {
1320 pte
= fetch_pte(dom
, bus_addr
);
1324 * No PTE for this address
1325 * move forward in 4kb steps
1327 unmap_size
= PAGE_SIZE
;
1328 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1329 /* 4kb PTE found for this address */
1330 unmap_size
= PAGE_SIZE
;
1335 /* Large PTE found which maps this address */
1336 unmap_size
= PTE_PAGE_SIZE(*pte
);
1337 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1338 for (i
= 0; i
< count
; i
++)
1342 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1343 unmapped
+= unmap_size
;
1346 BUG_ON(!is_power_of_2(unmapped
));
1352 * This function checks if a specific unity mapping entry is needed for
1353 * this specific IOMMU.
1355 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1356 struct unity_map_entry
*entry
)
1360 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1361 bdf
= amd_iommu_alias_table
[i
];
1362 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1370 * This function actually applies the mapping to the page table of the
1373 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1374 struct unity_map_entry
*e
)
1379 for (addr
= e
->address_start
; addr
< e
->address_end
;
1380 addr
+= PAGE_SIZE
) {
1381 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1386 * if unity mapping is in aperture range mark the page
1387 * as allocated in the aperture
1389 if (addr
< dma_dom
->aperture_size
)
1390 __set_bit(addr
>> PAGE_SHIFT
,
1391 dma_dom
->aperture
[0]->bitmap
);
1398 * Init the unity mappings for a specific IOMMU in the system
1400 * Basically iterates over all unity mapping entries and applies them to
1401 * the default domain DMA of that IOMMU if necessary.
1403 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1405 struct unity_map_entry
*entry
;
1408 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1409 if (!iommu_for_unity_map(iommu
, entry
))
1411 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1420 * Inits the unity mappings required for a specific device
1422 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1425 struct unity_map_entry
*e
;
1428 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1429 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1431 ret
= dma_ops_unity_map(dma_dom
, e
);
1439 /****************************************************************************
1441 * The next functions belong to the address allocator for the dma_ops
1442 * interface functions. They work like the allocators in the other IOMMU
1443 * drivers. Its basically a bitmap which marks the allocated pages in
1444 * the aperture. Maybe it could be enhanced in the future to a more
1445 * efficient allocator.
1447 ****************************************************************************/
1450 * The address allocator core functions.
1452 * called with domain->lock held
1456 * Used to reserve address ranges in the aperture (e.g. for exclusion
1459 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1460 unsigned long start_page
,
1463 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1465 if (start_page
+ pages
> last_page
)
1466 pages
= last_page
- start_page
;
1468 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1469 int index
= i
/ APERTURE_RANGE_PAGES
;
1470 int page
= i
% APERTURE_RANGE_PAGES
;
1471 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1476 * This function is used to add a new aperture range to an existing
1477 * aperture in case of dma_ops domain allocation or address allocation
1480 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1481 bool populate
, gfp_t gfp
)
1483 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1484 struct amd_iommu
*iommu
;
1485 unsigned long i
, old_size
;
1487 #ifdef CONFIG_IOMMU_STRESS
1491 if (index
>= APERTURE_MAX_RANGES
)
1494 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1495 if (!dma_dom
->aperture
[index
])
1498 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1499 if (!dma_dom
->aperture
[index
]->bitmap
)
1502 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1505 unsigned long address
= dma_dom
->aperture_size
;
1506 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1507 u64
*pte
, *pte_page
;
1509 for (i
= 0; i
< num_ptes
; ++i
) {
1510 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1515 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1517 address
+= APERTURE_RANGE_SIZE
/ 64;
1521 old_size
= dma_dom
->aperture_size
;
1522 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1524 /* Reserve address range used for MSI messages */
1525 if (old_size
< MSI_ADDR_BASE_LO
&&
1526 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1527 unsigned long spage
;
1530 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1531 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1533 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1536 /* Initialize the exclusion range if necessary */
1537 for_each_iommu(iommu
) {
1538 if (iommu
->exclusion_start
&&
1539 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1540 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1541 unsigned long startpage
;
1542 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1543 iommu
->exclusion_length
,
1545 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1546 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1551 * Check for areas already mapped as present in the new aperture
1552 * range and mark those pages as reserved in the allocator. Such
1553 * mappings may already exist as a result of requested unity
1554 * mappings for devices.
1556 for (i
= dma_dom
->aperture
[index
]->offset
;
1557 i
< dma_dom
->aperture_size
;
1559 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1560 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1563 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1566 update_domain(&dma_dom
->domain
);
1571 update_domain(&dma_dom
->domain
);
1573 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1575 kfree(dma_dom
->aperture
[index
]);
1576 dma_dom
->aperture
[index
] = NULL
;
1581 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1582 struct dma_ops_domain
*dom
,
1584 unsigned long align_mask
,
1586 unsigned long start
)
1588 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1589 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1590 int i
= start
>> APERTURE_RANGE_SHIFT
;
1591 unsigned long boundary_size
;
1592 unsigned long address
= -1;
1593 unsigned long limit
;
1595 next_bit
>>= PAGE_SHIFT
;
1597 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1598 PAGE_SIZE
) >> PAGE_SHIFT
;
1600 for (;i
< max_index
; ++i
) {
1601 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1603 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1606 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1607 dma_mask
>> PAGE_SHIFT
);
1609 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1610 limit
, next_bit
, pages
, 0,
1611 boundary_size
, align_mask
);
1612 if (address
!= -1) {
1613 address
= dom
->aperture
[i
]->offset
+
1614 (address
<< PAGE_SHIFT
);
1615 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1625 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1626 struct dma_ops_domain
*dom
,
1628 unsigned long align_mask
,
1631 unsigned long address
;
1633 #ifdef CONFIG_IOMMU_STRESS
1634 dom
->next_address
= 0;
1635 dom
->need_flush
= true;
1638 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1639 dma_mask
, dom
->next_address
);
1641 if (address
== -1) {
1642 dom
->next_address
= 0;
1643 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1645 dom
->need_flush
= true;
1648 if (unlikely(address
== -1))
1649 address
= DMA_ERROR_CODE
;
1651 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1657 * The address free function.
1659 * called with domain->lock held
1661 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1662 unsigned long address
,
1665 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1666 struct aperture_range
*range
= dom
->aperture
[i
];
1668 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1670 #ifdef CONFIG_IOMMU_STRESS
1675 if (address
>= dom
->next_address
)
1676 dom
->need_flush
= true;
1678 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1680 bitmap_clear(range
->bitmap
, address
, pages
);
1684 /****************************************************************************
1686 * The next functions belong to the domain allocation. A domain is
1687 * allocated for every IOMMU as the default domain. If device isolation
1688 * is enabled, every device get its own domain. The most important thing
1689 * about domains is the page table mapping the DMA address space they
1692 ****************************************************************************/
1695 * This function adds a protection domain to the global protection domain list
1697 static void add_domain_to_list(struct protection_domain
*domain
)
1699 unsigned long flags
;
1701 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1702 list_add(&domain
->list
, &amd_iommu_pd_list
);
1703 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1707 * This function removes a protection domain to the global
1708 * protection domain list
1710 static void del_domain_from_list(struct protection_domain
*domain
)
1712 unsigned long flags
;
1714 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1715 list_del(&domain
->list
);
1716 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1719 static u16
domain_id_alloc(void)
1721 unsigned long flags
;
1724 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1725 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1727 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1728 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1731 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1736 static void domain_id_free(int id
)
1738 unsigned long flags
;
1740 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1741 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1742 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1743 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1746 static void free_pagetable(struct protection_domain
*domain
)
1751 p1
= domain
->pt_root
;
1756 for (i
= 0; i
< 512; ++i
) {
1757 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1760 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1761 for (j
= 0; j
< 512; ++j
) {
1762 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1764 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1765 free_page((unsigned long)p3
);
1768 free_page((unsigned long)p2
);
1771 free_page((unsigned long)p1
);
1773 domain
->pt_root
= NULL
;
1776 static void free_gcr3_tbl_level1(u64
*tbl
)
1781 for (i
= 0; i
< 512; ++i
) {
1782 if (!(tbl
[i
] & GCR3_VALID
))
1785 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1787 free_page((unsigned long)ptr
);
1791 static void free_gcr3_tbl_level2(u64
*tbl
)
1796 for (i
= 0; i
< 512; ++i
) {
1797 if (!(tbl
[i
] & GCR3_VALID
))
1800 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1802 free_gcr3_tbl_level1(ptr
);
1806 static void free_gcr3_table(struct protection_domain
*domain
)
1808 if (domain
->glx
== 2)
1809 free_gcr3_tbl_level2(domain
->gcr3_tbl
);
1810 else if (domain
->glx
== 1)
1811 free_gcr3_tbl_level1(domain
->gcr3_tbl
);
1812 else if (domain
->glx
!= 0)
1815 free_page((unsigned long)domain
->gcr3_tbl
);
1819 * Free a domain, only used if something went wrong in the
1820 * allocation path and we need to free an already allocated page table
1822 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1829 del_domain_from_list(&dom
->domain
);
1831 free_pagetable(&dom
->domain
);
1833 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1834 if (!dom
->aperture
[i
])
1836 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1837 kfree(dom
->aperture
[i
]);
1844 * Allocates a new protection domain usable for the dma_ops functions.
1845 * It also initializes the page table and the address allocator data
1846 * structures required for the dma_ops interface
1848 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1850 struct dma_ops_domain
*dma_dom
;
1852 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1856 spin_lock_init(&dma_dom
->domain
.lock
);
1858 dma_dom
->domain
.id
= domain_id_alloc();
1859 if (dma_dom
->domain
.id
== 0)
1861 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1862 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1863 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1864 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1865 dma_dom
->domain
.priv
= dma_dom
;
1866 if (!dma_dom
->domain
.pt_root
)
1869 dma_dom
->need_flush
= false;
1870 dma_dom
->target_dev
= 0xffff;
1872 add_domain_to_list(&dma_dom
->domain
);
1874 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1878 * mark the first page as allocated so we never return 0 as
1879 * a valid dma-address. So we can use 0 as error value
1881 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1882 dma_dom
->next_address
= 0;
1888 dma_ops_domain_free(dma_dom
);
1894 * little helper function to check whether a given protection domain is a
1897 static bool dma_ops_domain(struct protection_domain
*domain
)
1899 return domain
->flags
& PD_DMA_OPS_MASK
;
1902 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1907 if (domain
->mode
!= PAGE_MODE_NONE
)
1908 pte_root
= virt_to_phys(domain
->pt_root
);
1910 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1911 << DEV_ENTRY_MODE_SHIFT
;
1912 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1914 flags
= amd_iommu_dev_table
[devid
].data
[1];
1917 flags
|= DTE_FLAG_IOTLB
;
1919 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1920 u64 gcr3
= __pa(domain
->gcr3_tbl
);
1921 u64 glx
= domain
->glx
;
1924 pte_root
|= DTE_FLAG_GV
;
1925 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1927 /* First mask out possible old values for GCR3 table */
1928 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1931 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1934 /* Encode GCR3 table into DTE */
1935 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1938 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1941 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1945 flags
&= ~(0xffffUL
);
1946 flags
|= domain
->id
;
1948 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1949 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1952 static void clear_dte_entry(u16 devid
)
1954 /* remove entry from the device table seen by the hardware */
1955 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1956 amd_iommu_dev_table
[devid
].data
[1] = 0;
1958 amd_iommu_apply_erratum_63(devid
);
1961 static void do_attach(struct iommu_dev_data
*dev_data
,
1962 struct protection_domain
*domain
)
1964 struct amd_iommu
*iommu
;
1967 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1968 ats
= dev_data
->ats
.enabled
;
1970 /* Update data structures */
1971 dev_data
->domain
= domain
;
1972 list_add(&dev_data
->list
, &domain
->dev_list
);
1973 set_dte_entry(dev_data
->devid
, domain
, ats
);
1975 /* Do reference counting */
1976 domain
->dev_iommu
[iommu
->index
] += 1;
1977 domain
->dev_cnt
+= 1;
1979 /* Flush the DTE entry */
1980 device_flush_dte(dev_data
);
1983 static void do_detach(struct iommu_dev_data
*dev_data
)
1985 struct amd_iommu
*iommu
;
1987 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1989 /* decrease reference counters */
1990 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
1991 dev_data
->domain
->dev_cnt
-= 1;
1993 /* Update data structures */
1994 dev_data
->domain
= NULL
;
1995 list_del(&dev_data
->list
);
1996 clear_dte_entry(dev_data
->devid
);
1998 /* Flush the DTE entry */
1999 device_flush_dte(dev_data
);
2003 * If a device is not yet associated with a domain, this function does
2004 * assigns it visible for the hardware
2006 static int __attach_device(struct iommu_dev_data
*dev_data
,
2007 struct protection_domain
*domain
)
2012 spin_lock(&domain
->lock
);
2014 if (dev_data
->alias_data
!= NULL
) {
2015 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2017 /* Some sanity checks */
2019 if (alias_data
->domain
!= NULL
&&
2020 alias_data
->domain
!= domain
)
2023 if (dev_data
->domain
!= NULL
&&
2024 dev_data
->domain
!= domain
)
2027 /* Do real assignment */
2028 if (alias_data
->domain
== NULL
)
2029 do_attach(alias_data
, domain
);
2031 atomic_inc(&alias_data
->bind
);
2034 if (dev_data
->domain
== NULL
)
2035 do_attach(dev_data
, domain
);
2037 atomic_inc(&dev_data
->bind
);
2044 spin_unlock(&domain
->lock
);
2050 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
2052 pci_disable_ats(pdev
);
2053 pci_disable_pri(pdev
);
2054 pci_disable_pasid(pdev
);
2057 /* FIXME: Change generic reset-function to do the same */
2058 static int pri_reset_while_enabled(struct pci_dev
*pdev
)
2063 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2067 pci_read_config_word(pdev
, pos
+ PCI_PRI_CTRL
, &control
);
2068 control
|= PCI_PRI_CTRL_RESET
;
2069 pci_write_config_word(pdev
, pos
+ PCI_PRI_CTRL
, control
);
2074 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
2079 /* FIXME: Hardcode number of outstanding requests for now */
2081 if (pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE
))
2083 reset_enable
= pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_ENABLE_RESET
);
2085 /* Only allow access to user-accessible pages */
2086 ret
= pci_enable_pasid(pdev
, 0);
2090 /* First reset the PRI state of the device */
2091 ret
= pci_reset_pri(pdev
);
2096 ret
= pci_enable_pri(pdev
, reqs
);
2101 ret
= pri_reset_while_enabled(pdev
);
2106 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
2113 pci_disable_pri(pdev
);
2114 pci_disable_pasid(pdev
);
2119 /* FIXME: Move this to PCI code */
2120 #define PCI_PRI_TLP_OFF (1 << 15)
2122 static bool pci_pri_tlp_required(struct pci_dev
*pdev
)
2127 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2131 pci_read_config_word(pdev
, pos
+ PCI_PRI_STATUS
, &status
);
2133 return (status
& PCI_PRI_TLP_OFF
) ? true : false;
2137 * If a device is not yet associated with a domain, this function does
2138 * assigns it visible for the hardware
2140 static int attach_device(struct device
*dev
,
2141 struct protection_domain
*domain
)
2143 struct pci_dev
*pdev
= to_pci_dev(dev
);
2144 struct iommu_dev_data
*dev_data
;
2145 unsigned long flags
;
2148 dev_data
= get_dev_data(dev
);
2150 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2151 if (!dev_data
->iommu_v2
|| !dev_data
->passthrough
)
2154 if (pdev_iommuv2_enable(pdev
) != 0)
2157 dev_data
->ats
.enabled
= true;
2158 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2159 dev_data
->pri_tlp
= pci_pri_tlp_required(pdev
);
2160 } else if (amd_iommu_iotlb_sup
&&
2161 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
2162 dev_data
->ats
.enabled
= true;
2163 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2166 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2167 ret
= __attach_device(dev_data
, domain
);
2168 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2171 * We might boot into a crash-kernel here. The crashed kernel
2172 * left the caches in the IOMMU dirty. So we have to flush
2173 * here to evict all dirty stuff.
2175 domain_flush_tlb_pde(domain
);
2181 * Removes a device from a protection domain (unlocked)
2183 static void __detach_device(struct iommu_dev_data
*dev_data
)
2185 struct protection_domain
*domain
;
2186 unsigned long flags
;
2188 BUG_ON(!dev_data
->domain
);
2190 domain
= dev_data
->domain
;
2192 spin_lock_irqsave(&domain
->lock
, flags
);
2194 if (dev_data
->alias_data
!= NULL
) {
2195 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2197 if (atomic_dec_and_test(&alias_data
->bind
))
2198 do_detach(alias_data
);
2201 if (atomic_dec_and_test(&dev_data
->bind
))
2202 do_detach(dev_data
);
2204 spin_unlock_irqrestore(&domain
->lock
, flags
);
2207 * If we run in passthrough mode the device must be assigned to the
2208 * passthrough domain if it is detached from any other domain.
2209 * Make sure we can deassign from the pt_domain itself.
2211 if (dev_data
->passthrough
&&
2212 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
2213 __attach_device(dev_data
, pt_domain
);
2217 * Removes a device from a protection domain (with devtable_lock held)
2219 static void detach_device(struct device
*dev
)
2221 struct protection_domain
*domain
;
2222 struct iommu_dev_data
*dev_data
;
2223 unsigned long flags
;
2225 dev_data
= get_dev_data(dev
);
2226 domain
= dev_data
->domain
;
2228 /* lock device table */
2229 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2230 __detach_device(dev_data
);
2231 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2233 if (domain
->flags
& PD_IOMMUV2_MASK
)
2234 pdev_iommuv2_disable(to_pci_dev(dev
));
2235 else if (dev_data
->ats
.enabled
)
2236 pci_disable_ats(to_pci_dev(dev
));
2238 dev_data
->ats
.enabled
= false;
2242 * Find out the protection domain structure for a given PCI device. This
2243 * will give us the pointer to the page table root for example.
2245 static struct protection_domain
*domain_for_device(struct device
*dev
)
2247 struct iommu_dev_data
*dev_data
;
2248 struct protection_domain
*dom
= NULL
;
2249 unsigned long flags
;
2251 dev_data
= get_dev_data(dev
);
2253 if (dev_data
->domain
)
2254 return dev_data
->domain
;
2256 if (dev_data
->alias_data
!= NULL
) {
2257 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2259 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2260 if (alias_data
->domain
!= NULL
) {
2261 __attach_device(dev_data
, alias_data
->domain
);
2262 dom
= alias_data
->domain
;
2264 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2270 static int device_change_notifier(struct notifier_block
*nb
,
2271 unsigned long action
, void *data
)
2273 struct dma_ops_domain
*dma_domain
;
2274 struct protection_domain
*domain
;
2275 struct iommu_dev_data
*dev_data
;
2276 struct device
*dev
= data
;
2277 struct amd_iommu
*iommu
;
2278 unsigned long flags
;
2281 if (!check_device(dev
))
2284 devid
= get_device_id(dev
);
2285 iommu
= amd_iommu_rlookup_table
[devid
];
2286 dev_data
= get_dev_data(dev
);
2289 case BUS_NOTIFY_UNBOUND_DRIVER
:
2291 domain
= domain_for_device(dev
);
2295 if (dev_data
->passthrough
)
2299 case BUS_NOTIFY_ADD_DEVICE
:
2301 iommu_init_device(dev
);
2304 * dev_data is still NULL and
2305 * got initialized in iommu_init_device
2307 dev_data
= get_dev_data(dev
);
2309 if (iommu_pass_through
|| dev_data
->iommu_v2
) {
2310 dev_data
->passthrough
= true;
2311 attach_device(dev
, pt_domain
);
2315 domain
= domain_for_device(dev
);
2317 /* allocate a protection domain if a device is added */
2318 dma_domain
= find_protection_domain(devid
);
2321 dma_domain
= dma_ops_domain_alloc();
2324 dma_domain
->target_dev
= devid
;
2326 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
2327 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
2328 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
2330 dev_data
= get_dev_data(dev
);
2332 dev
->archdata
.dma_ops
= &amd_iommu_dma_ops
;
2335 case BUS_NOTIFY_DEL_DEVICE
:
2337 iommu_uninit_device(dev
);
2343 iommu_completion_wait(iommu
);
2349 static struct notifier_block device_nb
= {
2350 .notifier_call
= device_change_notifier
,
2353 void amd_iommu_init_notifier(void)
2355 bus_register_notifier(&pci_bus_type
, &device_nb
);
2358 /*****************************************************************************
2360 * The next functions belong to the dma_ops mapping/unmapping code.
2362 *****************************************************************************/
2365 * In the dma_ops path we only have the struct device. This function
2366 * finds the corresponding IOMMU, the protection domain and the
2367 * requestor id for a given device.
2368 * If the device is not yet associated with a domain this is also done
2371 static struct protection_domain
*get_domain(struct device
*dev
)
2373 struct protection_domain
*domain
;
2374 struct dma_ops_domain
*dma_dom
;
2375 u16 devid
= get_device_id(dev
);
2377 if (!check_device(dev
))
2378 return ERR_PTR(-EINVAL
);
2380 domain
= domain_for_device(dev
);
2381 if (domain
!= NULL
&& !dma_ops_domain(domain
))
2382 return ERR_PTR(-EBUSY
);
2387 /* Device not bount yet - bind it */
2388 dma_dom
= find_protection_domain(devid
);
2390 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
2391 attach_device(dev
, &dma_dom
->domain
);
2392 DUMP_printk("Using protection domain %d for device %s\n",
2393 dma_dom
->domain
.id
, dev_name(dev
));
2395 return &dma_dom
->domain
;
2398 static void update_device_table(struct protection_domain
*domain
)
2400 struct iommu_dev_data
*dev_data
;
2402 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
2403 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2406 static void update_domain(struct protection_domain
*domain
)
2408 if (!domain
->updated
)
2411 update_device_table(domain
);
2413 domain_flush_devices(domain
);
2414 domain_flush_tlb_pde(domain
);
2416 domain
->updated
= false;
2420 * This function fetches the PTE for a given address in the aperture
2422 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
2423 unsigned long address
)
2425 struct aperture_range
*aperture
;
2426 u64
*pte
, *pte_page
;
2428 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2432 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2434 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
2436 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
2438 pte
+= PM_LEVEL_INDEX(0, address
);
2440 update_domain(&dom
->domain
);
2446 * This is the generic map function. It maps one 4kb page at paddr to
2447 * the given address in the DMA address space for the domain.
2449 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2450 unsigned long address
,
2456 WARN_ON(address
> dom
->aperture_size
);
2460 pte
= dma_ops_get_pte(dom
, address
);
2462 return DMA_ERROR_CODE
;
2464 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2466 if (direction
== DMA_TO_DEVICE
)
2467 __pte
|= IOMMU_PTE_IR
;
2468 else if (direction
== DMA_FROM_DEVICE
)
2469 __pte
|= IOMMU_PTE_IW
;
2470 else if (direction
== DMA_BIDIRECTIONAL
)
2471 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2477 return (dma_addr_t
)address
;
2481 * The generic unmapping function for on page in the DMA address space.
2483 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2484 unsigned long address
)
2486 struct aperture_range
*aperture
;
2489 if (address
>= dom
->aperture_size
)
2492 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2496 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2500 pte
+= PM_LEVEL_INDEX(0, address
);
2508 * This function contains common code for mapping of a physically
2509 * contiguous memory region into DMA address space. It is used by all
2510 * mapping functions provided with this IOMMU driver.
2511 * Must be called with the domain lock held.
2513 static dma_addr_t
__map_single(struct device
*dev
,
2514 struct dma_ops_domain
*dma_dom
,
2521 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2522 dma_addr_t address
, start
, ret
;
2524 unsigned long align_mask
= 0;
2527 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2530 INC_STATS_COUNTER(total_map_requests
);
2533 INC_STATS_COUNTER(cross_page
);
2536 align_mask
= (1UL << get_order(size
)) - 1;
2539 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2541 if (unlikely(address
== DMA_ERROR_CODE
)) {
2543 * setting next_address here will let the address
2544 * allocator only scan the new allocated range in the
2545 * first run. This is a small optimization.
2547 dma_dom
->next_address
= dma_dom
->aperture_size
;
2549 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2553 * aperture was successfully enlarged by 128 MB, try
2560 for (i
= 0; i
< pages
; ++i
) {
2561 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2562 if (ret
== DMA_ERROR_CODE
)
2570 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2572 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2573 domain_flush_tlb(&dma_dom
->domain
);
2574 dma_dom
->need_flush
= false;
2575 } else if (unlikely(amd_iommu_np_cache
))
2576 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2583 for (--i
; i
>= 0; --i
) {
2585 dma_ops_domain_unmap(dma_dom
, start
);
2588 dma_ops_free_addresses(dma_dom
, address
, pages
);
2590 return DMA_ERROR_CODE
;
2594 * Does the reverse of the __map_single function. Must be called with
2595 * the domain lock held too
2597 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2598 dma_addr_t dma_addr
,
2602 dma_addr_t flush_addr
;
2603 dma_addr_t i
, start
;
2606 if ((dma_addr
== DMA_ERROR_CODE
) ||
2607 (dma_addr
+ size
> dma_dom
->aperture_size
))
2610 flush_addr
= dma_addr
;
2611 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2612 dma_addr
&= PAGE_MASK
;
2615 for (i
= 0; i
< pages
; ++i
) {
2616 dma_ops_domain_unmap(dma_dom
, start
);
2620 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2622 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2624 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2625 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2626 dma_dom
->need_flush
= false;
2631 * The exported map_single function for dma_ops.
2633 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2634 unsigned long offset
, size_t size
,
2635 enum dma_data_direction dir
,
2636 struct dma_attrs
*attrs
)
2638 unsigned long flags
;
2639 struct protection_domain
*domain
;
2642 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2644 INC_STATS_COUNTER(cnt_map_single
);
2646 domain
= get_domain(dev
);
2647 if (PTR_ERR(domain
) == -EINVAL
)
2648 return (dma_addr_t
)paddr
;
2649 else if (IS_ERR(domain
))
2650 return DMA_ERROR_CODE
;
2652 dma_mask
= *dev
->dma_mask
;
2654 spin_lock_irqsave(&domain
->lock
, flags
);
2656 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2658 if (addr
== DMA_ERROR_CODE
)
2661 domain_flush_complete(domain
);
2664 spin_unlock_irqrestore(&domain
->lock
, flags
);
2670 * The exported unmap_single function for dma_ops.
2672 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2673 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2675 unsigned long flags
;
2676 struct protection_domain
*domain
;
2678 INC_STATS_COUNTER(cnt_unmap_single
);
2680 domain
= get_domain(dev
);
2684 spin_lock_irqsave(&domain
->lock
, flags
);
2686 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2688 domain_flush_complete(domain
);
2690 spin_unlock_irqrestore(&domain
->lock
, flags
);
2694 * This is a special map_sg function which is used if we should map a
2695 * device which is not handled by an AMD IOMMU in the system.
2697 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2698 int nelems
, int dir
)
2700 struct scatterlist
*s
;
2703 for_each_sg(sglist
, s
, nelems
, i
) {
2704 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2705 s
->dma_length
= s
->length
;
2712 * The exported map_sg function for dma_ops (handles scatter-gather
2715 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2716 int nelems
, enum dma_data_direction dir
,
2717 struct dma_attrs
*attrs
)
2719 unsigned long flags
;
2720 struct protection_domain
*domain
;
2722 struct scatterlist
*s
;
2724 int mapped_elems
= 0;
2727 INC_STATS_COUNTER(cnt_map_sg
);
2729 domain
= get_domain(dev
);
2730 if (PTR_ERR(domain
) == -EINVAL
)
2731 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2732 else if (IS_ERR(domain
))
2735 dma_mask
= *dev
->dma_mask
;
2737 spin_lock_irqsave(&domain
->lock
, flags
);
2739 for_each_sg(sglist
, s
, nelems
, i
) {
2742 s
->dma_address
= __map_single(dev
, domain
->priv
,
2743 paddr
, s
->length
, dir
, false,
2746 if (s
->dma_address
) {
2747 s
->dma_length
= s
->length
;
2753 domain_flush_complete(domain
);
2756 spin_unlock_irqrestore(&domain
->lock
, flags
);
2758 return mapped_elems
;
2760 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2762 __unmap_single(domain
->priv
, s
->dma_address
,
2763 s
->dma_length
, dir
);
2764 s
->dma_address
= s
->dma_length
= 0;
2773 * The exported map_sg function for dma_ops (handles scatter-gather
2776 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2777 int nelems
, enum dma_data_direction dir
,
2778 struct dma_attrs
*attrs
)
2780 unsigned long flags
;
2781 struct protection_domain
*domain
;
2782 struct scatterlist
*s
;
2785 INC_STATS_COUNTER(cnt_unmap_sg
);
2787 domain
= get_domain(dev
);
2791 spin_lock_irqsave(&domain
->lock
, flags
);
2793 for_each_sg(sglist
, s
, nelems
, i
) {
2794 __unmap_single(domain
->priv
, s
->dma_address
,
2795 s
->dma_length
, dir
);
2796 s
->dma_address
= s
->dma_length
= 0;
2799 domain_flush_complete(domain
);
2801 spin_unlock_irqrestore(&domain
->lock
, flags
);
2805 * The exported alloc_coherent function for dma_ops.
2807 static void *alloc_coherent(struct device
*dev
, size_t size
,
2808 dma_addr_t
*dma_addr
, gfp_t flag
,
2809 struct dma_attrs
*attrs
)
2811 unsigned long flags
;
2813 struct protection_domain
*domain
;
2815 u64 dma_mask
= dev
->coherent_dma_mask
;
2817 INC_STATS_COUNTER(cnt_alloc_coherent
);
2819 domain
= get_domain(dev
);
2820 if (PTR_ERR(domain
) == -EINVAL
) {
2821 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2822 *dma_addr
= __pa(virt_addr
);
2824 } else if (IS_ERR(domain
))
2827 dma_mask
= dev
->coherent_dma_mask
;
2828 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2831 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2835 paddr
= virt_to_phys(virt_addr
);
2838 dma_mask
= *dev
->dma_mask
;
2840 spin_lock_irqsave(&domain
->lock
, flags
);
2842 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2843 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2845 if (*dma_addr
== DMA_ERROR_CODE
) {
2846 spin_unlock_irqrestore(&domain
->lock
, flags
);
2850 domain_flush_complete(domain
);
2852 spin_unlock_irqrestore(&domain
->lock
, flags
);
2858 free_pages((unsigned long)virt_addr
, get_order(size
));
2864 * The exported free_coherent function for dma_ops.
2866 static void free_coherent(struct device
*dev
, size_t size
,
2867 void *virt_addr
, dma_addr_t dma_addr
,
2868 struct dma_attrs
*attrs
)
2870 unsigned long flags
;
2871 struct protection_domain
*domain
;
2873 INC_STATS_COUNTER(cnt_free_coherent
);
2875 domain
= get_domain(dev
);
2879 spin_lock_irqsave(&domain
->lock
, flags
);
2881 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2883 domain_flush_complete(domain
);
2885 spin_unlock_irqrestore(&domain
->lock
, flags
);
2888 free_pages((unsigned long)virt_addr
, get_order(size
));
2892 * This function is called by the DMA layer to find out if we can handle a
2893 * particular device. It is part of the dma_ops.
2895 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2897 return check_device(dev
);
2901 * The function for pre-allocating protection domains.
2903 * If the driver core informs the DMA layer if a driver grabs a device
2904 * we don't need to preallocate the protection domains anymore.
2905 * For now we have to.
2907 static void __init
prealloc_protection_domains(void)
2909 struct iommu_dev_data
*dev_data
;
2910 struct dma_ops_domain
*dma_dom
;
2911 struct pci_dev
*dev
= NULL
;
2914 for_each_pci_dev(dev
) {
2916 /* Do we handle this device? */
2917 if (!check_device(&dev
->dev
))
2920 dev_data
= get_dev_data(&dev
->dev
);
2921 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
2922 /* Make sure passthrough domain is allocated */
2923 alloc_passthrough_domain();
2924 dev_data
->passthrough
= true;
2925 attach_device(&dev
->dev
, pt_domain
);
2926 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2927 dev_name(&dev
->dev
));
2930 /* Is there already any domain for it? */
2931 if (domain_for_device(&dev
->dev
))
2934 devid
= get_device_id(&dev
->dev
);
2936 dma_dom
= dma_ops_domain_alloc();
2939 init_unity_mappings_for_device(dma_dom
, devid
);
2940 dma_dom
->target_dev
= devid
;
2942 attach_device(&dev
->dev
, &dma_dom
->domain
);
2944 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2948 static struct dma_map_ops amd_iommu_dma_ops
= {
2949 .alloc
= alloc_coherent
,
2950 .free
= free_coherent
,
2951 .map_page
= map_page
,
2952 .unmap_page
= unmap_page
,
2954 .unmap_sg
= unmap_sg
,
2955 .dma_supported
= amd_iommu_dma_supported
,
2958 static unsigned device_dma_ops_init(void)
2960 struct iommu_dev_data
*dev_data
;
2961 struct pci_dev
*pdev
= NULL
;
2962 unsigned unhandled
= 0;
2964 for_each_pci_dev(pdev
) {
2965 if (!check_device(&pdev
->dev
)) {
2967 iommu_ignore_device(&pdev
->dev
);
2973 dev_data
= get_dev_data(&pdev
->dev
);
2975 if (!dev_data
->passthrough
)
2976 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
2978 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
2985 * The function which clues the AMD IOMMU driver into dma_ops.
2988 void __init
amd_iommu_init_api(void)
2990 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
2993 int __init
amd_iommu_init_dma_ops(void)
2995 struct amd_iommu
*iommu
;
2999 * first allocate a default protection domain for every IOMMU we
3000 * found in the system. Devices not assigned to any other
3001 * protection domain will be assigned to the default one.
3003 for_each_iommu(iommu
) {
3004 iommu
->default_dom
= dma_ops_domain_alloc();
3005 if (iommu
->default_dom
== NULL
)
3007 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
3008 ret
= iommu_init_unity_mappings(iommu
);
3014 * Pre-allocate the protection domains for each device.
3016 prealloc_protection_domains();
3021 /* Make the driver finally visible to the drivers */
3022 unhandled
= device_dma_ops_init();
3023 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
3024 /* There are unhandled devices - initialize swiotlb for them */
3028 amd_iommu_stats_init();
3030 if (amd_iommu_unmap_flush
)
3031 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3033 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3039 for_each_iommu(iommu
) {
3040 if (iommu
->default_dom
)
3041 dma_ops_domain_free(iommu
->default_dom
);
3047 /*****************************************************************************
3049 * The following functions belong to the exported interface of AMD IOMMU
3051 * This interface allows access to lower level functions of the IOMMU
3052 * like protection domain handling and assignement of devices to domains
3053 * which is not possible with the dma_ops interface.
3055 *****************************************************************************/
3057 static void cleanup_domain(struct protection_domain
*domain
)
3059 struct iommu_dev_data
*dev_data
, *next
;
3060 unsigned long flags
;
3062 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
3064 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
3065 __detach_device(dev_data
);
3066 atomic_set(&dev_data
->bind
, 0);
3069 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
3072 static void protection_domain_free(struct protection_domain
*domain
)
3077 del_domain_from_list(domain
);
3080 domain_id_free(domain
->id
);
3085 static struct protection_domain
*protection_domain_alloc(void)
3087 struct protection_domain
*domain
;
3089 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
3093 spin_lock_init(&domain
->lock
);
3094 mutex_init(&domain
->api_lock
);
3095 domain
->id
= domain_id_alloc();
3098 INIT_LIST_HEAD(&domain
->dev_list
);
3100 add_domain_to_list(domain
);
3110 static int __init
alloc_passthrough_domain(void)
3112 if (pt_domain
!= NULL
)
3115 /* allocate passthrough domain */
3116 pt_domain
= protection_domain_alloc();
3120 pt_domain
->mode
= PAGE_MODE_NONE
;
3124 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
3126 struct protection_domain
*domain
;
3128 domain
= protection_domain_alloc();
3132 domain
->mode
= PAGE_MODE_3_LEVEL
;
3133 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
3134 if (!domain
->pt_root
)
3137 domain
->iommu_domain
= dom
;
3141 dom
->geometry
.aperture_start
= 0;
3142 dom
->geometry
.aperture_end
= ~0ULL;
3143 dom
->geometry
.force_aperture
= true;
3148 protection_domain_free(domain
);
3153 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
3155 struct protection_domain
*domain
= dom
->priv
;
3160 if (domain
->dev_cnt
> 0)
3161 cleanup_domain(domain
);
3163 BUG_ON(domain
->dev_cnt
!= 0);
3165 if (domain
->mode
!= PAGE_MODE_NONE
)
3166 free_pagetable(domain
);
3168 if (domain
->flags
& PD_IOMMUV2_MASK
)
3169 free_gcr3_table(domain
);
3171 protection_domain_free(domain
);
3176 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
3179 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
3180 struct amd_iommu
*iommu
;
3183 if (!check_device(dev
))
3186 devid
= get_device_id(dev
);
3188 if (dev_data
->domain
!= NULL
)
3191 iommu
= amd_iommu_rlookup_table
[devid
];
3195 iommu_completion_wait(iommu
);
3198 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
3201 struct protection_domain
*domain
= dom
->priv
;
3202 struct iommu_dev_data
*dev_data
;
3203 struct amd_iommu
*iommu
;
3206 if (!check_device(dev
))
3209 dev_data
= dev
->archdata
.iommu
;
3211 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3215 if (dev_data
->domain
)
3218 ret
= attach_device(dev
, domain
);
3220 iommu_completion_wait(iommu
);
3225 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
3226 phys_addr_t paddr
, size_t page_size
, int iommu_prot
)
3228 struct protection_domain
*domain
= dom
->priv
;
3232 if (domain
->mode
== PAGE_MODE_NONE
)
3235 if (iommu_prot
& IOMMU_READ
)
3236 prot
|= IOMMU_PROT_IR
;
3237 if (iommu_prot
& IOMMU_WRITE
)
3238 prot
|= IOMMU_PROT_IW
;
3240 mutex_lock(&domain
->api_lock
);
3241 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
3242 mutex_unlock(&domain
->api_lock
);
3247 static size_t amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
3250 struct protection_domain
*domain
= dom
->priv
;
3253 if (domain
->mode
== PAGE_MODE_NONE
)
3256 mutex_lock(&domain
->api_lock
);
3257 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3258 mutex_unlock(&domain
->api_lock
);
3260 domain_flush_tlb_pde(domain
);
3265 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3268 struct protection_domain
*domain
= dom
->priv
;
3269 unsigned long offset_mask
;
3273 if (domain
->mode
== PAGE_MODE_NONE
)
3276 pte
= fetch_pte(domain
, iova
);
3278 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3281 if (PM_PTE_LEVEL(*pte
) == 0)
3282 offset_mask
= PAGE_SIZE
- 1;
3284 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
3286 __pte
= *pte
& PM_ADDR_MASK
;
3287 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3292 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
3296 case IOMMU_CAP_CACHE_COHERENCY
:
3303 static struct iommu_ops amd_iommu_ops
= {
3304 .domain_init
= amd_iommu_domain_init
,
3305 .domain_destroy
= amd_iommu_domain_destroy
,
3306 .attach_dev
= amd_iommu_attach_device
,
3307 .detach_dev
= amd_iommu_detach_device
,
3308 .map
= amd_iommu_map
,
3309 .unmap
= amd_iommu_unmap
,
3310 .iova_to_phys
= amd_iommu_iova_to_phys
,
3311 .domain_has_cap
= amd_iommu_domain_has_cap
,
3312 .pgsize_bitmap
= AMD_IOMMU_PGSIZES
,
3315 /*****************************************************************************
3317 * The next functions do a basic initialization of IOMMU for pass through
3320 * In passthrough mode the IOMMU is initialized and enabled but not used for
3321 * DMA-API translation.
3323 *****************************************************************************/
3325 int __init
amd_iommu_init_passthrough(void)
3327 struct iommu_dev_data
*dev_data
;
3328 struct pci_dev
*dev
= NULL
;
3329 struct amd_iommu
*iommu
;
3333 ret
= alloc_passthrough_domain();
3337 for_each_pci_dev(dev
) {
3338 if (!check_device(&dev
->dev
))
3341 dev_data
= get_dev_data(&dev
->dev
);
3342 dev_data
->passthrough
= true;
3344 devid
= get_device_id(&dev
->dev
);
3346 iommu
= amd_iommu_rlookup_table
[devid
];
3350 attach_device(&dev
->dev
, pt_domain
);
3353 amd_iommu_stats_init();
3355 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3360 /* IOMMUv2 specific functions */
3361 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3363 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3365 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3367 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3369 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3371 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3373 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3375 struct protection_domain
*domain
= dom
->priv
;
3376 unsigned long flags
;
3378 spin_lock_irqsave(&domain
->lock
, flags
);
3380 /* Update data structure */
3381 domain
->mode
= PAGE_MODE_NONE
;
3382 domain
->updated
= true;
3384 /* Make changes visible to IOMMUs */
3385 update_domain(domain
);
3387 /* Page-table is not visible to IOMMU anymore, so free it */
3388 free_pagetable(domain
);
3390 spin_unlock_irqrestore(&domain
->lock
, flags
);
3392 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3394 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3396 struct protection_domain
*domain
= dom
->priv
;
3397 unsigned long flags
;
3400 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3403 /* Number of GCR3 table levels required */
3404 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3407 if (levels
> amd_iommu_max_glx_val
)
3410 spin_lock_irqsave(&domain
->lock
, flags
);
3413 * Save us all sanity checks whether devices already in the
3414 * domain support IOMMUv2. Just force that the domain has no
3415 * devices attached when it is switched into IOMMUv2 mode.
3418 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3422 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3423 if (domain
->gcr3_tbl
== NULL
)
3426 domain
->glx
= levels
;
3427 domain
->flags
|= PD_IOMMUV2_MASK
;
3428 domain
->updated
= true;
3430 update_domain(domain
);
3435 spin_unlock_irqrestore(&domain
->lock
, flags
);
3439 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3441 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3442 u64 address
, bool size
)
3444 struct iommu_dev_data
*dev_data
;
3445 struct iommu_cmd cmd
;
3448 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3451 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3454 * IOMMU TLB needs to be flushed before Device TLB to
3455 * prevent device TLB refill from IOMMU TLB
3457 for (i
= 0; i
< amd_iommus_present
; ++i
) {
3458 if (domain
->dev_iommu
[i
] == 0)
3461 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3466 /* Wait until IOMMU TLB flushes are complete */
3467 domain_flush_complete(domain
);
3469 /* Now flush device TLBs */
3470 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3471 struct amd_iommu
*iommu
;
3474 BUG_ON(!dev_data
->ats
.enabled
);
3476 qdep
= dev_data
->ats
.qdep
;
3477 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3479 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3480 qdep
, address
, size
);
3482 ret
= iommu_queue_command(iommu
, &cmd
);
3487 /* Wait until all device TLBs are flushed */
3488 domain_flush_complete(domain
);
3497 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3500 INC_STATS_COUNTER(invalidate_iotlb
);
3502 return __flush_pasid(domain
, pasid
, address
, false);
3505 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3508 struct protection_domain
*domain
= dom
->priv
;
3509 unsigned long flags
;
3512 spin_lock_irqsave(&domain
->lock
, flags
);
3513 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3514 spin_unlock_irqrestore(&domain
->lock
, flags
);
3518 EXPORT_SYMBOL(amd_iommu_flush_page
);
3520 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3522 INC_STATS_COUNTER(invalidate_iotlb_all
);
3524 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3528 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3530 struct protection_domain
*domain
= dom
->priv
;
3531 unsigned long flags
;
3534 spin_lock_irqsave(&domain
->lock
, flags
);
3535 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3536 spin_unlock_irqrestore(&domain
->lock
, flags
);
3540 EXPORT_SYMBOL(amd_iommu_flush_tlb
);
3542 static u64
*__get_gcr3_pte(u64
*root
, int level
, int pasid
, bool alloc
)
3549 index
= (pasid
>> (9 * level
)) & 0x1ff;
3555 if (!(*pte
& GCR3_VALID
)) {
3559 root
= (void *)get_zeroed_page(GFP_ATOMIC
);
3563 *pte
= __pa(root
) | GCR3_VALID
;
3566 root
= __va(*pte
& PAGE_MASK
);
3574 static int __set_gcr3(struct protection_domain
*domain
, int pasid
,
3579 if (domain
->mode
!= PAGE_MODE_NONE
)
3582 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, true);
3586 *pte
= (cr3
& PAGE_MASK
) | GCR3_VALID
;
3588 return __amd_iommu_flush_tlb(domain
, pasid
);
3591 static int __clear_gcr3(struct protection_domain
*domain
, int pasid
)
3595 if (domain
->mode
!= PAGE_MODE_NONE
)
3598 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, false);
3604 return __amd_iommu_flush_tlb(domain
, pasid
);
3607 int amd_iommu_domain_set_gcr3(struct iommu_domain
*dom
, int pasid
,
3610 struct protection_domain
*domain
= dom
->priv
;
3611 unsigned long flags
;
3614 spin_lock_irqsave(&domain
->lock
, flags
);
3615 ret
= __set_gcr3(domain
, pasid
, cr3
);
3616 spin_unlock_irqrestore(&domain
->lock
, flags
);
3620 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3
);
3622 int amd_iommu_domain_clear_gcr3(struct iommu_domain
*dom
, int pasid
)
3624 struct protection_domain
*domain
= dom
->priv
;
3625 unsigned long flags
;
3628 spin_lock_irqsave(&domain
->lock
, flags
);
3629 ret
= __clear_gcr3(domain
, pasid
);
3630 spin_unlock_irqrestore(&domain
->lock
, flags
);
3634 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3
);
3636 int amd_iommu_complete_ppr(struct pci_dev
*pdev
, int pasid
,
3637 int status
, int tag
)
3639 struct iommu_dev_data
*dev_data
;
3640 struct amd_iommu
*iommu
;
3641 struct iommu_cmd cmd
;
3643 INC_STATS_COUNTER(complete_ppr
);
3645 dev_data
= get_dev_data(&pdev
->dev
);
3646 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3648 build_complete_ppr(&cmd
, dev_data
->devid
, pasid
, status
,
3649 tag
, dev_data
->pri_tlp
);
3651 return iommu_queue_command(iommu
, &cmd
);
3653 EXPORT_SYMBOL(amd_iommu_complete_ppr
);
3655 struct iommu_domain
*amd_iommu_get_v2_domain(struct pci_dev
*pdev
)
3657 struct protection_domain
*domain
;
3659 domain
= get_domain(&pdev
->dev
);
3663 /* Only return IOMMUv2 domains */
3664 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3667 return domain
->iommu_domain
;
3669 EXPORT_SYMBOL(amd_iommu_get_v2_domain
);
3671 void amd_iommu_enable_device_erratum(struct pci_dev
*pdev
, u32 erratum
)
3673 struct iommu_dev_data
*dev_data
;
3675 if (!amd_iommu_v2_supported())
3678 dev_data
= get_dev_data(&pdev
->dev
);
3679 dev_data
->errata
|= (1 << erratum
);
3681 EXPORT_SYMBOL(amd_iommu_enable_device_erratum
);
3683 int amd_iommu_device_info(struct pci_dev
*pdev
,
3684 struct amd_iommu_device_info
*info
)
3689 if (pdev
== NULL
|| info
== NULL
)
3692 if (!amd_iommu_v2_supported())
3695 memset(info
, 0, sizeof(*info
));
3697 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ATS
);
3699 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_ATS_SUP
;
3701 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
3703 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRI_SUP
;
3705 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PASID
);
3709 max_pasids
= 1 << (9 * (amd_iommu_max_glx_val
+ 1));
3710 max_pasids
= min(max_pasids
, (1 << 20));
3712 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PASID_SUP
;
3713 info
->max_pasids
= min(pci_max_pasids(pdev
), max_pasids
);
3715 features
= pci_pasid_features(pdev
);
3716 if (features
& PCI_PASID_CAP_EXEC
)
3717 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP
;
3718 if (features
& PCI_PASID_CAP_PRIV
)
3719 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP
;
3724 EXPORT_SYMBOL(amd_iommu_device_info
);