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
= NULL
, *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);
298 if (dma_pdev
== NULL
)
299 dma_pdev
= pci_dev_get(pdev
);
301 /* Account for quirked devices */
302 swap_pci_ref(&dma_pdev
, pci_get_dma_source(dma_pdev
));
305 * If it's a multifunction device that does not support our
306 * required ACS flags, add to the same group as function 0.
308 if (dma_pdev
->multifunction
&&
309 !pci_acs_enabled(dma_pdev
, REQ_ACS_FLAGS
))
310 swap_pci_ref(&dma_pdev
,
311 pci_get_slot(dma_pdev
->bus
,
312 PCI_DEVFN(PCI_SLOT(dma_pdev
->devfn
),
316 * Devices on the root bus go through the iommu. If that's not us,
317 * find the next upstream device and test ACS up to the root bus.
318 * Finding the next device may require skipping virtual buses.
320 while (!pci_is_root_bus(dma_pdev
->bus
)) {
321 struct pci_bus
*bus
= dma_pdev
->bus
;
324 if (!pci_is_root_bus(bus
))
330 if (pci_acs_path_enabled(bus
->self
, NULL
, REQ_ACS_FLAGS
))
333 swap_pci_ref(&dma_pdev
, pci_dev_get(bus
->self
));
337 group
= iommu_group_get(&dma_pdev
->dev
);
338 pci_dev_put(dma_pdev
);
340 group
= iommu_group_alloc();
342 return PTR_ERR(group
);
345 ret
= iommu_group_add_device(group
, dev
);
347 iommu_group_put(group
);
352 if (pci_iommuv2_capable(pdev
)) {
353 struct amd_iommu
*iommu
;
355 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
356 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
359 dev
->archdata
.iommu
= dev_data
;
364 static void iommu_ignore_device(struct device
*dev
)
368 devid
= get_device_id(dev
);
369 alias
= amd_iommu_alias_table
[devid
];
371 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
372 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
374 amd_iommu_rlookup_table
[devid
] = NULL
;
375 amd_iommu_rlookup_table
[alias
] = NULL
;
378 static void iommu_uninit_device(struct device
*dev
)
380 iommu_group_remove_device(dev
);
383 * Nothing to do here - we keep dev_data around for unplugged devices
384 * and reuse it when the device is re-plugged - not doing so would
385 * introduce a ton of races.
389 void __init
amd_iommu_uninit_devices(void)
391 struct iommu_dev_data
*dev_data
, *n
;
392 struct pci_dev
*pdev
= NULL
;
394 for_each_pci_dev(pdev
) {
396 if (!check_device(&pdev
->dev
))
399 iommu_uninit_device(&pdev
->dev
);
402 /* Free all of our dev_data structures */
403 list_for_each_entry_safe(dev_data
, n
, &dev_data_list
, dev_data_list
)
404 free_dev_data(dev_data
);
407 int __init
amd_iommu_init_devices(void)
409 struct pci_dev
*pdev
= NULL
;
412 for_each_pci_dev(pdev
) {
414 if (!check_device(&pdev
->dev
))
417 ret
= iommu_init_device(&pdev
->dev
);
418 if (ret
== -ENOTSUPP
)
419 iommu_ignore_device(&pdev
->dev
);
428 amd_iommu_uninit_devices();
432 #ifdef CONFIG_AMD_IOMMU_STATS
435 * Initialization code for statistics collection
438 DECLARE_STATS_COUNTER(compl_wait
);
439 DECLARE_STATS_COUNTER(cnt_map_single
);
440 DECLARE_STATS_COUNTER(cnt_unmap_single
);
441 DECLARE_STATS_COUNTER(cnt_map_sg
);
442 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
443 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
444 DECLARE_STATS_COUNTER(cnt_free_coherent
);
445 DECLARE_STATS_COUNTER(cross_page
);
446 DECLARE_STATS_COUNTER(domain_flush_single
);
447 DECLARE_STATS_COUNTER(domain_flush_all
);
448 DECLARE_STATS_COUNTER(alloced_io_mem
);
449 DECLARE_STATS_COUNTER(total_map_requests
);
450 DECLARE_STATS_COUNTER(complete_ppr
);
451 DECLARE_STATS_COUNTER(invalidate_iotlb
);
452 DECLARE_STATS_COUNTER(invalidate_iotlb_all
);
453 DECLARE_STATS_COUNTER(pri_requests
);
455 static struct dentry
*stats_dir
;
456 static struct dentry
*de_fflush
;
458 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
460 if (stats_dir
== NULL
)
463 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
467 static void amd_iommu_stats_init(void)
469 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
470 if (stats_dir
== NULL
)
473 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
474 &amd_iommu_unmap_flush
);
476 amd_iommu_stats_add(&compl_wait
);
477 amd_iommu_stats_add(&cnt_map_single
);
478 amd_iommu_stats_add(&cnt_unmap_single
);
479 amd_iommu_stats_add(&cnt_map_sg
);
480 amd_iommu_stats_add(&cnt_unmap_sg
);
481 amd_iommu_stats_add(&cnt_alloc_coherent
);
482 amd_iommu_stats_add(&cnt_free_coherent
);
483 amd_iommu_stats_add(&cross_page
);
484 amd_iommu_stats_add(&domain_flush_single
);
485 amd_iommu_stats_add(&domain_flush_all
);
486 amd_iommu_stats_add(&alloced_io_mem
);
487 amd_iommu_stats_add(&total_map_requests
);
488 amd_iommu_stats_add(&complete_ppr
);
489 amd_iommu_stats_add(&invalidate_iotlb
);
490 amd_iommu_stats_add(&invalidate_iotlb_all
);
491 amd_iommu_stats_add(&pri_requests
);
496 /****************************************************************************
498 * Interrupt handling functions
500 ****************************************************************************/
502 static void dump_dte_entry(u16 devid
)
506 for (i
= 0; i
< 4; ++i
)
507 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
508 amd_iommu_dev_table
[devid
].data
[i
]);
511 static void dump_command(unsigned long phys_addr
)
513 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
516 for (i
= 0; i
< 4; ++i
)
517 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
520 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
522 int type
, devid
, domid
, flags
;
523 volatile u32
*event
= __evt
;
528 type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
529 devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
530 domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
531 flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
532 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
535 /* Did we hit the erratum? */
536 if (++count
== LOOP_TIMEOUT
) {
537 pr_err("AMD-Vi: No event written to event log\n");
544 printk(KERN_ERR
"AMD-Vi: Event logged [");
547 case EVENT_TYPE_ILL_DEV
:
548 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
549 "address=0x%016llx flags=0x%04x]\n",
550 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
552 dump_dte_entry(devid
);
554 case EVENT_TYPE_IO_FAULT
:
555 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
556 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
557 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
558 domid
, address
, flags
);
560 case EVENT_TYPE_DEV_TAB_ERR
:
561 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
562 "address=0x%016llx flags=0x%04x]\n",
563 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
566 case EVENT_TYPE_PAGE_TAB_ERR
:
567 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
568 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
569 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
570 domid
, address
, flags
);
572 case EVENT_TYPE_ILL_CMD
:
573 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
574 dump_command(address
);
576 case EVENT_TYPE_CMD_HARD_ERR
:
577 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
578 "flags=0x%04x]\n", address
, flags
);
580 case EVENT_TYPE_IOTLB_INV_TO
:
581 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
582 "address=0x%016llx]\n",
583 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
586 case EVENT_TYPE_INV_DEV_REQ
:
587 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
588 "address=0x%016llx flags=0x%04x]\n",
589 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
593 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
596 memset(__evt
, 0, 4 * sizeof(u32
));
599 static void iommu_poll_events(struct amd_iommu
*iommu
)
604 spin_lock_irqsave(&iommu
->lock
, flags
);
606 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
607 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
609 while (head
!= tail
) {
610 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
611 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
614 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
616 spin_unlock_irqrestore(&iommu
->lock
, flags
);
619 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u64
*raw
)
621 struct amd_iommu_fault fault
;
623 INC_STATS_COUNTER(pri_requests
);
625 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
626 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
630 fault
.address
= raw
[1];
631 fault
.pasid
= PPR_PASID(raw
[0]);
632 fault
.device_id
= PPR_DEVID(raw
[0]);
633 fault
.tag
= PPR_TAG(raw
[0]);
634 fault
.flags
= PPR_FLAGS(raw
[0]);
636 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
639 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
644 if (iommu
->ppr_log
== NULL
)
647 /* enable ppr interrupts again */
648 writel(MMIO_STATUS_PPR_INT_MASK
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
650 spin_lock_irqsave(&iommu
->lock
, flags
);
652 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
653 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
655 while (head
!= tail
) {
660 raw
= (u64
*)(iommu
->ppr_log
+ head
);
663 * Hardware bug: Interrupt may arrive before the entry is
664 * written to memory. If this happens we need to wait for the
667 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
668 if (PPR_REQ_TYPE(raw
[0]) != 0)
673 /* Avoid memcpy function-call overhead */
678 * To detect the hardware bug we need to clear the entry
681 raw
[0] = raw
[1] = 0UL;
683 /* Update head pointer of hardware ring-buffer */
684 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
685 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
688 * Release iommu->lock because ppr-handling might need to
691 spin_unlock_irqrestore(&iommu
->lock
, flags
);
693 /* Handle PPR entry */
694 iommu_handle_ppr_entry(iommu
, entry
);
696 spin_lock_irqsave(&iommu
->lock
, flags
);
698 /* Refresh ring-buffer information */
699 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
700 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
703 spin_unlock_irqrestore(&iommu
->lock
, flags
);
706 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
708 struct amd_iommu
*iommu
;
710 for_each_iommu(iommu
) {
711 iommu_poll_events(iommu
);
712 iommu_poll_ppr_log(iommu
);
718 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
720 return IRQ_WAKE_THREAD
;
723 /****************************************************************************
725 * IOMMU command queuing functions
727 ****************************************************************************/
729 static int wait_on_sem(volatile u64
*sem
)
733 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
738 if (i
== LOOP_TIMEOUT
) {
739 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
746 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
747 struct iommu_cmd
*cmd
,
752 target
= iommu
->cmd_buf
+ tail
;
753 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
755 /* Copy command to buffer */
756 memcpy(target
, cmd
, sizeof(*cmd
));
758 /* Tell the IOMMU about it */
759 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
762 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
764 WARN_ON(address
& 0x7ULL
);
766 memset(cmd
, 0, sizeof(*cmd
));
767 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
768 cmd
->data
[1] = upper_32_bits(__pa(address
));
770 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
773 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
775 memset(cmd
, 0, sizeof(*cmd
));
776 cmd
->data
[0] = devid
;
777 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
780 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
781 size_t size
, u16 domid
, int pde
)
786 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
791 * If we have to flush more than one page, flush all
792 * TLB entries for this domain
794 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
798 address
&= PAGE_MASK
;
800 memset(cmd
, 0, sizeof(*cmd
));
801 cmd
->data
[1] |= domid
;
802 cmd
->data
[2] = lower_32_bits(address
);
803 cmd
->data
[3] = upper_32_bits(address
);
804 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
805 if (s
) /* size bit - we flush more than one 4kb page */
806 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
807 if (pde
) /* PDE bit - we wan't flush everything not only the PTEs */
808 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
811 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
812 u64 address
, size_t size
)
817 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
822 * If we have to flush more than one page, flush all
823 * TLB entries for this domain
825 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
829 address
&= PAGE_MASK
;
831 memset(cmd
, 0, sizeof(*cmd
));
832 cmd
->data
[0] = devid
;
833 cmd
->data
[0] |= (qdep
& 0xff) << 24;
834 cmd
->data
[1] = devid
;
835 cmd
->data
[2] = lower_32_bits(address
);
836 cmd
->data
[3] = upper_32_bits(address
);
837 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
839 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
842 static void build_inv_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
843 u64 address
, bool size
)
845 memset(cmd
, 0, sizeof(*cmd
));
847 address
&= ~(0xfffULL
);
849 cmd
->data
[0] = pasid
& PASID_MASK
;
850 cmd
->data
[1] = domid
;
851 cmd
->data
[2] = lower_32_bits(address
);
852 cmd
->data
[3] = upper_32_bits(address
);
853 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
854 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
856 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
857 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
860 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
861 int qdep
, u64 address
, bool size
)
863 memset(cmd
, 0, sizeof(*cmd
));
865 address
&= ~(0xfffULL
);
867 cmd
->data
[0] = devid
;
868 cmd
->data
[0] |= (pasid
& 0xff) << 16;
869 cmd
->data
[0] |= (qdep
& 0xff) << 24;
870 cmd
->data
[1] = devid
;
871 cmd
->data
[1] |= ((pasid
>> 8) & 0xfff) << 16;
872 cmd
->data
[2] = lower_32_bits(address
);
873 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
874 cmd
->data
[3] = upper_32_bits(address
);
876 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
877 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
880 static void build_complete_ppr(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
881 int status
, int tag
, bool gn
)
883 memset(cmd
, 0, sizeof(*cmd
));
885 cmd
->data
[0] = devid
;
887 cmd
->data
[1] = pasid
& PASID_MASK
;
888 cmd
->data
[2] = CMD_INV_IOMMU_PAGES_GN_MASK
;
890 cmd
->data
[3] = tag
& 0x1ff;
891 cmd
->data
[3] |= (status
& PPR_STATUS_MASK
) << PPR_STATUS_SHIFT
;
893 CMD_SET_TYPE(cmd
, CMD_COMPLETE_PPR
);
896 static void build_inv_all(struct iommu_cmd
*cmd
)
898 memset(cmd
, 0, sizeof(*cmd
));
899 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
903 * Writes the command to the IOMMUs command buffer and informs the
904 * hardware about the new command.
906 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
907 struct iommu_cmd
*cmd
,
910 u32 left
, tail
, head
, next_tail
;
913 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
916 spin_lock_irqsave(&iommu
->lock
, flags
);
918 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
919 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
920 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
921 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
924 struct iommu_cmd sync_cmd
;
925 volatile u64 sem
= 0;
928 build_completion_wait(&sync_cmd
, (u64
)&sem
);
929 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
931 spin_unlock_irqrestore(&iommu
->lock
, flags
);
933 if ((ret
= wait_on_sem(&sem
)) != 0)
939 copy_cmd_to_buffer(iommu
, cmd
, tail
);
941 /* We need to sync now to make sure all commands are processed */
942 iommu
->need_sync
= sync
;
944 spin_unlock_irqrestore(&iommu
->lock
, flags
);
949 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
951 return iommu_queue_command_sync(iommu
, cmd
, true);
955 * This function queues a completion wait command into the command
958 static int iommu_completion_wait(struct amd_iommu
*iommu
)
960 struct iommu_cmd cmd
;
961 volatile u64 sem
= 0;
964 if (!iommu
->need_sync
)
967 build_completion_wait(&cmd
, (u64
)&sem
);
969 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
973 return wait_on_sem(&sem
);
976 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
978 struct iommu_cmd cmd
;
980 build_inv_dte(&cmd
, devid
);
982 return iommu_queue_command(iommu
, &cmd
);
985 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
989 for (devid
= 0; devid
<= 0xffff; ++devid
)
990 iommu_flush_dte(iommu
, devid
);
992 iommu_completion_wait(iommu
);
996 * This function uses heavy locking and may disable irqs for some time. But
997 * this is no issue because it is only called during resume.
999 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
1003 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
1004 struct iommu_cmd cmd
;
1005 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
1007 iommu_queue_command(iommu
, &cmd
);
1010 iommu_completion_wait(iommu
);
1013 static void iommu_flush_all(struct amd_iommu
*iommu
)
1015 struct iommu_cmd cmd
;
1017 build_inv_all(&cmd
);
1019 iommu_queue_command(iommu
, &cmd
);
1020 iommu_completion_wait(iommu
);
1023 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
1025 if (iommu_feature(iommu
, FEATURE_IA
)) {
1026 iommu_flush_all(iommu
);
1028 iommu_flush_dte_all(iommu
);
1029 iommu_flush_tlb_all(iommu
);
1034 * Command send function for flushing on-device TLB
1036 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
1037 u64 address
, size_t size
)
1039 struct amd_iommu
*iommu
;
1040 struct iommu_cmd cmd
;
1043 qdep
= dev_data
->ats
.qdep
;
1044 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1046 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
1048 return iommu_queue_command(iommu
, &cmd
);
1052 * Command send function for invalidating a device table entry
1054 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
1056 struct amd_iommu
*iommu
;
1059 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1061 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
1065 if (dev_data
->ats
.enabled
)
1066 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
1072 * TLB invalidation function which is called from the mapping functions.
1073 * It invalidates a single PTE if the range to flush is within a single
1074 * page. Otherwise it flushes the whole TLB of the IOMMU.
1076 static void __domain_flush_pages(struct protection_domain
*domain
,
1077 u64 address
, size_t size
, int pde
)
1079 struct iommu_dev_data
*dev_data
;
1080 struct iommu_cmd cmd
;
1083 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
1085 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1086 if (!domain
->dev_iommu
[i
])
1090 * Devices of this domain are behind this IOMMU
1091 * We need a TLB flush
1093 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
1096 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1098 if (!dev_data
->ats
.enabled
)
1101 ret
|= device_flush_iotlb(dev_data
, address
, size
);
1107 static void domain_flush_pages(struct protection_domain
*domain
,
1108 u64 address
, size_t size
)
1110 __domain_flush_pages(domain
, address
, size
, 0);
1113 /* Flush the whole IO/TLB for a given protection domain */
1114 static void domain_flush_tlb(struct protection_domain
*domain
)
1116 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
1119 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1120 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
1122 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
1125 static void domain_flush_complete(struct protection_domain
*domain
)
1129 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1130 if (!domain
->dev_iommu
[i
])
1134 * Devices of this domain are behind this IOMMU
1135 * We need to wait for completion of all commands.
1137 iommu_completion_wait(amd_iommus
[i
]);
1143 * This function flushes the DTEs for all devices in domain
1145 static void domain_flush_devices(struct protection_domain
*domain
)
1147 struct iommu_dev_data
*dev_data
;
1149 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1150 device_flush_dte(dev_data
);
1153 /****************************************************************************
1155 * The functions below are used the create the page table mappings for
1156 * unity mapped regions.
1158 ****************************************************************************/
1161 * This function is used to add another level to an IO page table. Adding
1162 * another level increases the size of the address space by 9 bits to a size up
1165 static bool increase_address_space(struct protection_domain
*domain
,
1170 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1171 /* address space already 64 bit large */
1174 pte
= (void *)get_zeroed_page(gfp
);
1178 *pte
= PM_LEVEL_PDE(domain
->mode
,
1179 virt_to_phys(domain
->pt_root
));
1180 domain
->pt_root
= pte
;
1182 domain
->updated
= true;
1187 static u64
*alloc_pte(struct protection_domain
*domain
,
1188 unsigned long address
,
1189 unsigned long page_size
,
1196 BUG_ON(!is_power_of_2(page_size
));
1198 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1199 increase_address_space(domain
, gfp
);
1201 level
= domain
->mode
- 1;
1202 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1203 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1204 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1206 while (level
> end_lvl
) {
1207 if (!IOMMU_PTE_PRESENT(*pte
)) {
1208 page
= (u64
*)get_zeroed_page(gfp
);
1211 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1214 /* No level skipping support yet */
1215 if (PM_PTE_LEVEL(*pte
) != level
)
1220 pte
= IOMMU_PTE_PAGE(*pte
);
1222 if (pte_page
&& level
== end_lvl
)
1225 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1232 * This function checks if there is a PTE for a given dma address. If
1233 * there is one, it returns the pointer to it.
1235 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
1240 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1243 level
= domain
->mode
- 1;
1244 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1249 if (!IOMMU_PTE_PRESENT(*pte
))
1253 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1254 unsigned long pte_mask
, __pte
;
1257 * If we have a series of large PTEs, make
1258 * sure to return a pointer to the first one.
1260 pte_mask
= PTE_PAGE_SIZE(*pte
);
1261 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1262 __pte
= ((unsigned long)pte
) & pte_mask
;
1264 return (u64
*)__pte
;
1267 /* No level skipping support yet */
1268 if (PM_PTE_LEVEL(*pte
) != level
)
1273 /* Walk to the next level */
1274 pte
= IOMMU_PTE_PAGE(*pte
);
1275 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1282 * Generic mapping functions. It maps a physical address into a DMA
1283 * address space. It allocates the page table pages if necessary.
1284 * In the future it can be extended to a generic mapping function
1285 * supporting all features of AMD IOMMU page tables like level skipping
1286 * and full 64 bit address spaces.
1288 static int iommu_map_page(struct protection_domain
*dom
,
1289 unsigned long bus_addr
,
1290 unsigned long phys_addr
,
1292 unsigned long page_size
)
1297 if (!(prot
& IOMMU_PROT_MASK
))
1300 bus_addr
= PAGE_ALIGN(bus_addr
);
1301 phys_addr
= PAGE_ALIGN(phys_addr
);
1302 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1303 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1305 for (i
= 0; i
< count
; ++i
)
1306 if (IOMMU_PTE_PRESENT(pte
[i
]))
1309 if (page_size
> PAGE_SIZE
) {
1310 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1311 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1313 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1315 if (prot
& IOMMU_PROT_IR
)
1316 __pte
|= IOMMU_PTE_IR
;
1317 if (prot
& IOMMU_PROT_IW
)
1318 __pte
|= IOMMU_PTE_IW
;
1320 for (i
= 0; i
< count
; ++i
)
1328 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1329 unsigned long bus_addr
,
1330 unsigned long page_size
)
1332 unsigned long long unmap_size
, unmapped
;
1335 BUG_ON(!is_power_of_2(page_size
));
1339 while (unmapped
< page_size
) {
1341 pte
= fetch_pte(dom
, bus_addr
);
1345 * No PTE for this address
1346 * move forward in 4kb steps
1348 unmap_size
= PAGE_SIZE
;
1349 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1350 /* 4kb PTE found for this address */
1351 unmap_size
= PAGE_SIZE
;
1356 /* Large PTE found which maps this address */
1357 unmap_size
= PTE_PAGE_SIZE(*pte
);
1358 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1359 for (i
= 0; i
< count
; i
++)
1363 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1364 unmapped
+= unmap_size
;
1367 BUG_ON(!is_power_of_2(unmapped
));
1373 * This function checks if a specific unity mapping entry is needed for
1374 * this specific IOMMU.
1376 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1377 struct unity_map_entry
*entry
)
1381 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1382 bdf
= amd_iommu_alias_table
[i
];
1383 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1391 * This function actually applies the mapping to the page table of the
1394 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1395 struct unity_map_entry
*e
)
1400 for (addr
= e
->address_start
; addr
< e
->address_end
;
1401 addr
+= PAGE_SIZE
) {
1402 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1407 * if unity mapping is in aperture range mark the page
1408 * as allocated in the aperture
1410 if (addr
< dma_dom
->aperture_size
)
1411 __set_bit(addr
>> PAGE_SHIFT
,
1412 dma_dom
->aperture
[0]->bitmap
);
1419 * Init the unity mappings for a specific IOMMU in the system
1421 * Basically iterates over all unity mapping entries and applies them to
1422 * the default domain DMA of that IOMMU if necessary.
1424 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1426 struct unity_map_entry
*entry
;
1429 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1430 if (!iommu_for_unity_map(iommu
, entry
))
1432 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1441 * Inits the unity mappings required for a specific device
1443 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1446 struct unity_map_entry
*e
;
1449 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1450 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1452 ret
= dma_ops_unity_map(dma_dom
, e
);
1460 /****************************************************************************
1462 * The next functions belong to the address allocator for the dma_ops
1463 * interface functions. They work like the allocators in the other IOMMU
1464 * drivers. Its basically a bitmap which marks the allocated pages in
1465 * the aperture. Maybe it could be enhanced in the future to a more
1466 * efficient allocator.
1468 ****************************************************************************/
1471 * The address allocator core functions.
1473 * called with domain->lock held
1477 * Used to reserve address ranges in the aperture (e.g. for exclusion
1480 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1481 unsigned long start_page
,
1484 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1486 if (start_page
+ pages
> last_page
)
1487 pages
= last_page
- start_page
;
1489 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1490 int index
= i
/ APERTURE_RANGE_PAGES
;
1491 int page
= i
% APERTURE_RANGE_PAGES
;
1492 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1497 * This function is used to add a new aperture range to an existing
1498 * aperture in case of dma_ops domain allocation or address allocation
1501 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1502 bool populate
, gfp_t gfp
)
1504 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1505 struct amd_iommu
*iommu
;
1506 unsigned long i
, old_size
;
1508 #ifdef CONFIG_IOMMU_STRESS
1512 if (index
>= APERTURE_MAX_RANGES
)
1515 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1516 if (!dma_dom
->aperture
[index
])
1519 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1520 if (!dma_dom
->aperture
[index
]->bitmap
)
1523 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1526 unsigned long address
= dma_dom
->aperture_size
;
1527 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1528 u64
*pte
, *pte_page
;
1530 for (i
= 0; i
< num_ptes
; ++i
) {
1531 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1536 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1538 address
+= APERTURE_RANGE_SIZE
/ 64;
1542 old_size
= dma_dom
->aperture_size
;
1543 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1545 /* Reserve address range used for MSI messages */
1546 if (old_size
< MSI_ADDR_BASE_LO
&&
1547 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1548 unsigned long spage
;
1551 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1552 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1554 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1557 /* Initialize the exclusion range if necessary */
1558 for_each_iommu(iommu
) {
1559 if (iommu
->exclusion_start
&&
1560 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1561 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1562 unsigned long startpage
;
1563 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1564 iommu
->exclusion_length
,
1566 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1567 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1572 * Check for areas already mapped as present in the new aperture
1573 * range and mark those pages as reserved in the allocator. Such
1574 * mappings may already exist as a result of requested unity
1575 * mappings for devices.
1577 for (i
= dma_dom
->aperture
[index
]->offset
;
1578 i
< dma_dom
->aperture_size
;
1580 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1581 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1584 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1587 update_domain(&dma_dom
->domain
);
1592 update_domain(&dma_dom
->domain
);
1594 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1596 kfree(dma_dom
->aperture
[index
]);
1597 dma_dom
->aperture
[index
] = NULL
;
1602 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1603 struct dma_ops_domain
*dom
,
1605 unsigned long align_mask
,
1607 unsigned long start
)
1609 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1610 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1611 int i
= start
>> APERTURE_RANGE_SHIFT
;
1612 unsigned long boundary_size
;
1613 unsigned long address
= -1;
1614 unsigned long limit
;
1616 next_bit
>>= PAGE_SHIFT
;
1618 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1619 PAGE_SIZE
) >> PAGE_SHIFT
;
1621 for (;i
< max_index
; ++i
) {
1622 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1624 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1627 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1628 dma_mask
>> PAGE_SHIFT
);
1630 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1631 limit
, next_bit
, pages
, 0,
1632 boundary_size
, align_mask
);
1633 if (address
!= -1) {
1634 address
= dom
->aperture
[i
]->offset
+
1635 (address
<< PAGE_SHIFT
);
1636 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1646 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1647 struct dma_ops_domain
*dom
,
1649 unsigned long align_mask
,
1652 unsigned long address
;
1654 #ifdef CONFIG_IOMMU_STRESS
1655 dom
->next_address
= 0;
1656 dom
->need_flush
= true;
1659 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1660 dma_mask
, dom
->next_address
);
1662 if (address
== -1) {
1663 dom
->next_address
= 0;
1664 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1666 dom
->need_flush
= true;
1669 if (unlikely(address
== -1))
1670 address
= DMA_ERROR_CODE
;
1672 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1678 * The address free function.
1680 * called with domain->lock held
1682 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1683 unsigned long address
,
1686 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1687 struct aperture_range
*range
= dom
->aperture
[i
];
1689 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1691 #ifdef CONFIG_IOMMU_STRESS
1696 if (address
>= dom
->next_address
)
1697 dom
->need_flush
= true;
1699 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1701 bitmap_clear(range
->bitmap
, address
, pages
);
1705 /****************************************************************************
1707 * The next functions belong to the domain allocation. A domain is
1708 * allocated for every IOMMU as the default domain. If device isolation
1709 * is enabled, every device get its own domain. The most important thing
1710 * about domains is the page table mapping the DMA address space they
1713 ****************************************************************************/
1716 * This function adds a protection domain to the global protection domain list
1718 static void add_domain_to_list(struct protection_domain
*domain
)
1720 unsigned long flags
;
1722 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1723 list_add(&domain
->list
, &amd_iommu_pd_list
);
1724 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1728 * This function removes a protection domain to the global
1729 * protection domain list
1731 static void del_domain_from_list(struct protection_domain
*domain
)
1733 unsigned long flags
;
1735 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1736 list_del(&domain
->list
);
1737 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1740 static u16
domain_id_alloc(void)
1742 unsigned long flags
;
1745 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1746 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1748 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1749 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1752 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1757 static void domain_id_free(int id
)
1759 unsigned long flags
;
1761 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1762 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1763 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1764 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1767 static void free_pagetable(struct protection_domain
*domain
)
1772 p1
= domain
->pt_root
;
1777 for (i
= 0; i
< 512; ++i
) {
1778 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1781 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1782 for (j
= 0; j
< 512; ++j
) {
1783 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1785 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1786 free_page((unsigned long)p3
);
1789 free_page((unsigned long)p2
);
1792 free_page((unsigned long)p1
);
1794 domain
->pt_root
= NULL
;
1797 static void free_gcr3_tbl_level1(u64
*tbl
)
1802 for (i
= 0; i
< 512; ++i
) {
1803 if (!(tbl
[i
] & GCR3_VALID
))
1806 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1808 free_page((unsigned long)ptr
);
1812 static void free_gcr3_tbl_level2(u64
*tbl
)
1817 for (i
= 0; i
< 512; ++i
) {
1818 if (!(tbl
[i
] & GCR3_VALID
))
1821 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1823 free_gcr3_tbl_level1(ptr
);
1827 static void free_gcr3_table(struct protection_domain
*domain
)
1829 if (domain
->glx
== 2)
1830 free_gcr3_tbl_level2(domain
->gcr3_tbl
);
1831 else if (domain
->glx
== 1)
1832 free_gcr3_tbl_level1(domain
->gcr3_tbl
);
1833 else if (domain
->glx
!= 0)
1836 free_page((unsigned long)domain
->gcr3_tbl
);
1840 * Free a domain, only used if something went wrong in the
1841 * allocation path and we need to free an already allocated page table
1843 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1850 del_domain_from_list(&dom
->domain
);
1852 free_pagetable(&dom
->domain
);
1854 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1855 if (!dom
->aperture
[i
])
1857 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1858 kfree(dom
->aperture
[i
]);
1865 * Allocates a new protection domain usable for the dma_ops functions.
1866 * It also initializes the page table and the address allocator data
1867 * structures required for the dma_ops interface
1869 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1871 struct dma_ops_domain
*dma_dom
;
1873 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1877 spin_lock_init(&dma_dom
->domain
.lock
);
1879 dma_dom
->domain
.id
= domain_id_alloc();
1880 if (dma_dom
->domain
.id
== 0)
1882 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1883 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1884 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1885 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1886 dma_dom
->domain
.priv
= dma_dom
;
1887 if (!dma_dom
->domain
.pt_root
)
1890 dma_dom
->need_flush
= false;
1891 dma_dom
->target_dev
= 0xffff;
1893 add_domain_to_list(&dma_dom
->domain
);
1895 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1899 * mark the first page as allocated so we never return 0 as
1900 * a valid dma-address. So we can use 0 as error value
1902 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1903 dma_dom
->next_address
= 0;
1909 dma_ops_domain_free(dma_dom
);
1915 * little helper function to check whether a given protection domain is a
1918 static bool dma_ops_domain(struct protection_domain
*domain
)
1920 return domain
->flags
& PD_DMA_OPS_MASK
;
1923 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
1928 if (domain
->mode
!= PAGE_MODE_NONE
)
1929 pte_root
= virt_to_phys(domain
->pt_root
);
1931 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
1932 << DEV_ENTRY_MODE_SHIFT
;
1933 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
1935 flags
= amd_iommu_dev_table
[devid
].data
[1];
1938 flags
|= DTE_FLAG_IOTLB
;
1940 if (domain
->flags
& PD_IOMMUV2_MASK
) {
1941 u64 gcr3
= __pa(domain
->gcr3_tbl
);
1942 u64 glx
= domain
->glx
;
1945 pte_root
|= DTE_FLAG_GV
;
1946 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
1948 /* First mask out possible old values for GCR3 table */
1949 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
1952 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
1955 /* Encode GCR3 table into DTE */
1956 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
1959 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
1962 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
1966 flags
&= ~(0xffffUL
);
1967 flags
|= domain
->id
;
1969 amd_iommu_dev_table
[devid
].data
[1] = flags
;
1970 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
1973 static void clear_dte_entry(u16 devid
)
1975 /* remove entry from the device table seen by the hardware */
1976 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
1977 amd_iommu_dev_table
[devid
].data
[1] = 0;
1979 amd_iommu_apply_erratum_63(devid
);
1982 static void do_attach(struct iommu_dev_data
*dev_data
,
1983 struct protection_domain
*domain
)
1985 struct amd_iommu
*iommu
;
1988 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1989 ats
= dev_data
->ats
.enabled
;
1991 /* Update data structures */
1992 dev_data
->domain
= domain
;
1993 list_add(&dev_data
->list
, &domain
->dev_list
);
1994 set_dte_entry(dev_data
->devid
, domain
, ats
);
1996 /* Do reference counting */
1997 domain
->dev_iommu
[iommu
->index
] += 1;
1998 domain
->dev_cnt
+= 1;
2000 /* Flush the DTE entry */
2001 device_flush_dte(dev_data
);
2004 static void do_detach(struct iommu_dev_data
*dev_data
)
2006 struct amd_iommu
*iommu
;
2008 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2010 /* decrease reference counters */
2011 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
2012 dev_data
->domain
->dev_cnt
-= 1;
2014 /* Update data structures */
2015 dev_data
->domain
= NULL
;
2016 list_del(&dev_data
->list
);
2017 clear_dte_entry(dev_data
->devid
);
2019 /* Flush the DTE entry */
2020 device_flush_dte(dev_data
);
2024 * If a device is not yet associated with a domain, this function does
2025 * assigns it visible for the hardware
2027 static int __attach_device(struct iommu_dev_data
*dev_data
,
2028 struct protection_domain
*domain
)
2033 spin_lock(&domain
->lock
);
2035 if (dev_data
->alias_data
!= NULL
) {
2036 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2038 /* Some sanity checks */
2040 if (alias_data
->domain
!= NULL
&&
2041 alias_data
->domain
!= domain
)
2044 if (dev_data
->domain
!= NULL
&&
2045 dev_data
->domain
!= domain
)
2048 /* Do real assignment */
2049 if (alias_data
->domain
== NULL
)
2050 do_attach(alias_data
, domain
);
2052 atomic_inc(&alias_data
->bind
);
2055 if (dev_data
->domain
== NULL
)
2056 do_attach(dev_data
, domain
);
2058 atomic_inc(&dev_data
->bind
);
2065 spin_unlock(&domain
->lock
);
2071 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
2073 pci_disable_ats(pdev
);
2074 pci_disable_pri(pdev
);
2075 pci_disable_pasid(pdev
);
2078 /* FIXME: Change generic reset-function to do the same */
2079 static int pri_reset_while_enabled(struct pci_dev
*pdev
)
2084 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2088 pci_read_config_word(pdev
, pos
+ PCI_PRI_CTRL
, &control
);
2089 control
|= PCI_PRI_CTRL_RESET
;
2090 pci_write_config_word(pdev
, pos
+ PCI_PRI_CTRL
, control
);
2095 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
2100 /* FIXME: Hardcode number of outstanding requests for now */
2102 if (pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE
))
2104 reset_enable
= pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_ENABLE_RESET
);
2106 /* Only allow access to user-accessible pages */
2107 ret
= pci_enable_pasid(pdev
, 0);
2111 /* First reset the PRI state of the device */
2112 ret
= pci_reset_pri(pdev
);
2117 ret
= pci_enable_pri(pdev
, reqs
);
2122 ret
= pri_reset_while_enabled(pdev
);
2127 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
2134 pci_disable_pri(pdev
);
2135 pci_disable_pasid(pdev
);
2140 /* FIXME: Move this to PCI code */
2141 #define PCI_PRI_TLP_OFF (1 << 15)
2143 static bool pci_pri_tlp_required(struct pci_dev
*pdev
)
2148 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2152 pci_read_config_word(pdev
, pos
+ PCI_PRI_STATUS
, &status
);
2154 return (status
& PCI_PRI_TLP_OFF
) ? true : false;
2158 * If a device is not yet associated with a domain, this function does
2159 * assigns it visible for the hardware
2161 static int attach_device(struct device
*dev
,
2162 struct protection_domain
*domain
)
2164 struct pci_dev
*pdev
= to_pci_dev(dev
);
2165 struct iommu_dev_data
*dev_data
;
2166 unsigned long flags
;
2169 dev_data
= get_dev_data(dev
);
2171 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2172 if (!dev_data
->iommu_v2
|| !dev_data
->passthrough
)
2175 if (pdev_iommuv2_enable(pdev
) != 0)
2178 dev_data
->ats
.enabled
= true;
2179 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2180 dev_data
->pri_tlp
= pci_pri_tlp_required(pdev
);
2181 } else if (amd_iommu_iotlb_sup
&&
2182 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
2183 dev_data
->ats
.enabled
= true;
2184 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2187 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2188 ret
= __attach_device(dev_data
, domain
);
2189 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2192 * We might boot into a crash-kernel here. The crashed kernel
2193 * left the caches in the IOMMU dirty. So we have to flush
2194 * here to evict all dirty stuff.
2196 domain_flush_tlb_pde(domain
);
2202 * Removes a device from a protection domain (unlocked)
2204 static void __detach_device(struct iommu_dev_data
*dev_data
)
2206 struct protection_domain
*domain
;
2207 unsigned long flags
;
2209 BUG_ON(!dev_data
->domain
);
2211 domain
= dev_data
->domain
;
2213 spin_lock_irqsave(&domain
->lock
, flags
);
2215 if (dev_data
->alias_data
!= NULL
) {
2216 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2218 if (atomic_dec_and_test(&alias_data
->bind
))
2219 do_detach(alias_data
);
2222 if (atomic_dec_and_test(&dev_data
->bind
))
2223 do_detach(dev_data
);
2225 spin_unlock_irqrestore(&domain
->lock
, flags
);
2228 * If we run in passthrough mode the device must be assigned to the
2229 * passthrough domain if it is detached from any other domain.
2230 * Make sure we can deassign from the pt_domain itself.
2232 if (dev_data
->passthrough
&&
2233 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
2234 __attach_device(dev_data
, pt_domain
);
2238 * Removes a device from a protection domain (with devtable_lock held)
2240 static void detach_device(struct device
*dev
)
2242 struct protection_domain
*domain
;
2243 struct iommu_dev_data
*dev_data
;
2244 unsigned long flags
;
2246 dev_data
= get_dev_data(dev
);
2247 domain
= dev_data
->domain
;
2249 /* lock device table */
2250 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2251 __detach_device(dev_data
);
2252 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2254 if (domain
->flags
& PD_IOMMUV2_MASK
)
2255 pdev_iommuv2_disable(to_pci_dev(dev
));
2256 else if (dev_data
->ats
.enabled
)
2257 pci_disable_ats(to_pci_dev(dev
));
2259 dev_data
->ats
.enabled
= false;
2263 * Find out the protection domain structure for a given PCI device. This
2264 * will give us the pointer to the page table root for example.
2266 static struct protection_domain
*domain_for_device(struct device
*dev
)
2268 struct iommu_dev_data
*dev_data
;
2269 struct protection_domain
*dom
= NULL
;
2270 unsigned long flags
;
2272 dev_data
= get_dev_data(dev
);
2274 if (dev_data
->domain
)
2275 return dev_data
->domain
;
2277 if (dev_data
->alias_data
!= NULL
) {
2278 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2280 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2281 if (alias_data
->domain
!= NULL
) {
2282 __attach_device(dev_data
, alias_data
->domain
);
2283 dom
= alias_data
->domain
;
2285 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2291 static int device_change_notifier(struct notifier_block
*nb
,
2292 unsigned long action
, void *data
)
2294 struct dma_ops_domain
*dma_domain
;
2295 struct protection_domain
*domain
;
2296 struct iommu_dev_data
*dev_data
;
2297 struct device
*dev
= data
;
2298 struct amd_iommu
*iommu
;
2299 unsigned long flags
;
2302 if (!check_device(dev
))
2305 devid
= get_device_id(dev
);
2306 iommu
= amd_iommu_rlookup_table
[devid
];
2307 dev_data
= get_dev_data(dev
);
2310 case BUS_NOTIFY_UNBOUND_DRIVER
:
2312 domain
= domain_for_device(dev
);
2316 if (dev_data
->passthrough
)
2320 case BUS_NOTIFY_ADD_DEVICE
:
2322 iommu_init_device(dev
);
2325 * dev_data is still NULL and
2326 * got initialized in iommu_init_device
2328 dev_data
= get_dev_data(dev
);
2330 if (iommu_pass_through
|| dev_data
->iommu_v2
) {
2331 dev_data
->passthrough
= true;
2332 attach_device(dev
, pt_domain
);
2336 domain
= domain_for_device(dev
);
2338 /* allocate a protection domain if a device is added */
2339 dma_domain
= find_protection_domain(devid
);
2342 dma_domain
= dma_ops_domain_alloc();
2345 dma_domain
->target_dev
= devid
;
2347 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
2348 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
2349 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
2351 dev_data
= get_dev_data(dev
);
2353 dev
->archdata
.dma_ops
= &amd_iommu_dma_ops
;
2356 case BUS_NOTIFY_DEL_DEVICE
:
2358 iommu_uninit_device(dev
);
2364 iommu_completion_wait(iommu
);
2370 static struct notifier_block device_nb
= {
2371 .notifier_call
= device_change_notifier
,
2374 void amd_iommu_init_notifier(void)
2376 bus_register_notifier(&pci_bus_type
, &device_nb
);
2379 /*****************************************************************************
2381 * The next functions belong to the dma_ops mapping/unmapping code.
2383 *****************************************************************************/
2386 * In the dma_ops path we only have the struct device. This function
2387 * finds the corresponding IOMMU, the protection domain and the
2388 * requestor id for a given device.
2389 * If the device is not yet associated with a domain this is also done
2392 static struct protection_domain
*get_domain(struct device
*dev
)
2394 struct protection_domain
*domain
;
2395 struct dma_ops_domain
*dma_dom
;
2396 u16 devid
= get_device_id(dev
);
2398 if (!check_device(dev
))
2399 return ERR_PTR(-EINVAL
);
2401 domain
= domain_for_device(dev
);
2402 if (domain
!= NULL
&& !dma_ops_domain(domain
))
2403 return ERR_PTR(-EBUSY
);
2408 /* Device not bount yet - bind it */
2409 dma_dom
= find_protection_domain(devid
);
2411 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
2412 attach_device(dev
, &dma_dom
->domain
);
2413 DUMP_printk("Using protection domain %d for device %s\n",
2414 dma_dom
->domain
.id
, dev_name(dev
));
2416 return &dma_dom
->domain
;
2419 static void update_device_table(struct protection_domain
*domain
)
2421 struct iommu_dev_data
*dev_data
;
2423 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
2424 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2427 static void update_domain(struct protection_domain
*domain
)
2429 if (!domain
->updated
)
2432 update_device_table(domain
);
2434 domain_flush_devices(domain
);
2435 domain_flush_tlb_pde(domain
);
2437 domain
->updated
= false;
2441 * This function fetches the PTE for a given address in the aperture
2443 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
2444 unsigned long address
)
2446 struct aperture_range
*aperture
;
2447 u64
*pte
, *pte_page
;
2449 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2453 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2455 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
2457 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
2459 pte
+= PM_LEVEL_INDEX(0, address
);
2461 update_domain(&dom
->domain
);
2467 * This is the generic map function. It maps one 4kb page at paddr to
2468 * the given address in the DMA address space for the domain.
2470 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2471 unsigned long address
,
2477 WARN_ON(address
> dom
->aperture_size
);
2481 pte
= dma_ops_get_pte(dom
, address
);
2483 return DMA_ERROR_CODE
;
2485 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2487 if (direction
== DMA_TO_DEVICE
)
2488 __pte
|= IOMMU_PTE_IR
;
2489 else if (direction
== DMA_FROM_DEVICE
)
2490 __pte
|= IOMMU_PTE_IW
;
2491 else if (direction
== DMA_BIDIRECTIONAL
)
2492 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2498 return (dma_addr_t
)address
;
2502 * The generic unmapping function for on page in the DMA address space.
2504 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2505 unsigned long address
)
2507 struct aperture_range
*aperture
;
2510 if (address
>= dom
->aperture_size
)
2513 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2517 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2521 pte
+= PM_LEVEL_INDEX(0, address
);
2529 * This function contains common code for mapping of a physically
2530 * contiguous memory region into DMA address space. It is used by all
2531 * mapping functions provided with this IOMMU driver.
2532 * Must be called with the domain lock held.
2534 static dma_addr_t
__map_single(struct device
*dev
,
2535 struct dma_ops_domain
*dma_dom
,
2542 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2543 dma_addr_t address
, start
, ret
;
2545 unsigned long align_mask
= 0;
2548 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2551 INC_STATS_COUNTER(total_map_requests
);
2554 INC_STATS_COUNTER(cross_page
);
2557 align_mask
= (1UL << get_order(size
)) - 1;
2560 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2562 if (unlikely(address
== DMA_ERROR_CODE
)) {
2564 * setting next_address here will let the address
2565 * allocator only scan the new allocated range in the
2566 * first run. This is a small optimization.
2568 dma_dom
->next_address
= dma_dom
->aperture_size
;
2570 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2574 * aperture was successfully enlarged by 128 MB, try
2581 for (i
= 0; i
< pages
; ++i
) {
2582 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2583 if (ret
== DMA_ERROR_CODE
)
2591 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2593 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2594 domain_flush_tlb(&dma_dom
->domain
);
2595 dma_dom
->need_flush
= false;
2596 } else if (unlikely(amd_iommu_np_cache
))
2597 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2604 for (--i
; i
>= 0; --i
) {
2606 dma_ops_domain_unmap(dma_dom
, start
);
2609 dma_ops_free_addresses(dma_dom
, address
, pages
);
2611 return DMA_ERROR_CODE
;
2615 * Does the reverse of the __map_single function. Must be called with
2616 * the domain lock held too
2618 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2619 dma_addr_t dma_addr
,
2623 dma_addr_t flush_addr
;
2624 dma_addr_t i
, start
;
2627 if ((dma_addr
== DMA_ERROR_CODE
) ||
2628 (dma_addr
+ size
> dma_dom
->aperture_size
))
2631 flush_addr
= dma_addr
;
2632 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2633 dma_addr
&= PAGE_MASK
;
2636 for (i
= 0; i
< pages
; ++i
) {
2637 dma_ops_domain_unmap(dma_dom
, start
);
2641 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2643 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2645 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2646 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2647 dma_dom
->need_flush
= false;
2652 * The exported map_single function for dma_ops.
2654 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2655 unsigned long offset
, size_t size
,
2656 enum dma_data_direction dir
,
2657 struct dma_attrs
*attrs
)
2659 unsigned long flags
;
2660 struct protection_domain
*domain
;
2663 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2665 INC_STATS_COUNTER(cnt_map_single
);
2667 domain
= get_domain(dev
);
2668 if (PTR_ERR(domain
) == -EINVAL
)
2669 return (dma_addr_t
)paddr
;
2670 else if (IS_ERR(domain
))
2671 return DMA_ERROR_CODE
;
2673 dma_mask
= *dev
->dma_mask
;
2675 spin_lock_irqsave(&domain
->lock
, flags
);
2677 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2679 if (addr
== DMA_ERROR_CODE
)
2682 domain_flush_complete(domain
);
2685 spin_unlock_irqrestore(&domain
->lock
, flags
);
2691 * The exported unmap_single function for dma_ops.
2693 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2694 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2696 unsigned long flags
;
2697 struct protection_domain
*domain
;
2699 INC_STATS_COUNTER(cnt_unmap_single
);
2701 domain
= get_domain(dev
);
2705 spin_lock_irqsave(&domain
->lock
, flags
);
2707 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2709 domain_flush_complete(domain
);
2711 spin_unlock_irqrestore(&domain
->lock
, flags
);
2715 * This is a special map_sg function which is used if we should map a
2716 * device which is not handled by an AMD IOMMU in the system.
2718 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2719 int nelems
, int dir
)
2721 struct scatterlist
*s
;
2724 for_each_sg(sglist
, s
, nelems
, i
) {
2725 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2726 s
->dma_length
= s
->length
;
2733 * The exported map_sg function for dma_ops (handles scatter-gather
2736 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2737 int nelems
, enum dma_data_direction dir
,
2738 struct dma_attrs
*attrs
)
2740 unsigned long flags
;
2741 struct protection_domain
*domain
;
2743 struct scatterlist
*s
;
2745 int mapped_elems
= 0;
2748 INC_STATS_COUNTER(cnt_map_sg
);
2750 domain
= get_domain(dev
);
2751 if (PTR_ERR(domain
) == -EINVAL
)
2752 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2753 else if (IS_ERR(domain
))
2756 dma_mask
= *dev
->dma_mask
;
2758 spin_lock_irqsave(&domain
->lock
, flags
);
2760 for_each_sg(sglist
, s
, nelems
, i
) {
2763 s
->dma_address
= __map_single(dev
, domain
->priv
,
2764 paddr
, s
->length
, dir
, false,
2767 if (s
->dma_address
) {
2768 s
->dma_length
= s
->length
;
2774 domain_flush_complete(domain
);
2777 spin_unlock_irqrestore(&domain
->lock
, flags
);
2779 return mapped_elems
;
2781 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2783 __unmap_single(domain
->priv
, s
->dma_address
,
2784 s
->dma_length
, dir
);
2785 s
->dma_address
= s
->dma_length
= 0;
2794 * The exported map_sg function for dma_ops (handles scatter-gather
2797 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2798 int nelems
, enum dma_data_direction dir
,
2799 struct dma_attrs
*attrs
)
2801 unsigned long flags
;
2802 struct protection_domain
*domain
;
2803 struct scatterlist
*s
;
2806 INC_STATS_COUNTER(cnt_unmap_sg
);
2808 domain
= get_domain(dev
);
2812 spin_lock_irqsave(&domain
->lock
, flags
);
2814 for_each_sg(sglist
, s
, nelems
, i
) {
2815 __unmap_single(domain
->priv
, s
->dma_address
,
2816 s
->dma_length
, dir
);
2817 s
->dma_address
= s
->dma_length
= 0;
2820 domain_flush_complete(domain
);
2822 spin_unlock_irqrestore(&domain
->lock
, flags
);
2826 * The exported alloc_coherent function for dma_ops.
2828 static void *alloc_coherent(struct device
*dev
, size_t size
,
2829 dma_addr_t
*dma_addr
, gfp_t flag
,
2830 struct dma_attrs
*attrs
)
2832 unsigned long flags
;
2834 struct protection_domain
*domain
;
2836 u64 dma_mask
= dev
->coherent_dma_mask
;
2838 INC_STATS_COUNTER(cnt_alloc_coherent
);
2840 domain
= get_domain(dev
);
2841 if (PTR_ERR(domain
) == -EINVAL
) {
2842 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2843 *dma_addr
= __pa(virt_addr
);
2845 } else if (IS_ERR(domain
))
2848 dma_mask
= dev
->coherent_dma_mask
;
2849 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2852 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2856 paddr
= virt_to_phys(virt_addr
);
2859 dma_mask
= *dev
->dma_mask
;
2861 spin_lock_irqsave(&domain
->lock
, flags
);
2863 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2864 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2866 if (*dma_addr
== DMA_ERROR_CODE
) {
2867 spin_unlock_irqrestore(&domain
->lock
, flags
);
2871 domain_flush_complete(domain
);
2873 spin_unlock_irqrestore(&domain
->lock
, flags
);
2879 free_pages((unsigned long)virt_addr
, get_order(size
));
2885 * The exported free_coherent function for dma_ops.
2887 static void free_coherent(struct device
*dev
, size_t size
,
2888 void *virt_addr
, dma_addr_t dma_addr
,
2889 struct dma_attrs
*attrs
)
2891 unsigned long flags
;
2892 struct protection_domain
*domain
;
2894 INC_STATS_COUNTER(cnt_free_coherent
);
2896 domain
= get_domain(dev
);
2900 spin_lock_irqsave(&domain
->lock
, flags
);
2902 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2904 domain_flush_complete(domain
);
2906 spin_unlock_irqrestore(&domain
->lock
, flags
);
2909 free_pages((unsigned long)virt_addr
, get_order(size
));
2913 * This function is called by the DMA layer to find out if we can handle a
2914 * particular device. It is part of the dma_ops.
2916 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2918 return check_device(dev
);
2922 * The function for pre-allocating protection domains.
2924 * If the driver core informs the DMA layer if a driver grabs a device
2925 * we don't need to preallocate the protection domains anymore.
2926 * For now we have to.
2928 static void __init
prealloc_protection_domains(void)
2930 struct iommu_dev_data
*dev_data
;
2931 struct dma_ops_domain
*dma_dom
;
2932 struct pci_dev
*dev
= NULL
;
2935 for_each_pci_dev(dev
) {
2937 /* Do we handle this device? */
2938 if (!check_device(&dev
->dev
))
2941 dev_data
= get_dev_data(&dev
->dev
);
2942 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
2943 /* Make sure passthrough domain is allocated */
2944 alloc_passthrough_domain();
2945 dev_data
->passthrough
= true;
2946 attach_device(&dev
->dev
, pt_domain
);
2947 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2948 dev_name(&dev
->dev
));
2951 /* Is there already any domain for it? */
2952 if (domain_for_device(&dev
->dev
))
2955 devid
= get_device_id(&dev
->dev
);
2957 dma_dom
= dma_ops_domain_alloc();
2960 init_unity_mappings_for_device(dma_dom
, devid
);
2961 dma_dom
->target_dev
= devid
;
2963 attach_device(&dev
->dev
, &dma_dom
->domain
);
2965 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
2969 static struct dma_map_ops amd_iommu_dma_ops
= {
2970 .alloc
= alloc_coherent
,
2971 .free
= free_coherent
,
2972 .map_page
= map_page
,
2973 .unmap_page
= unmap_page
,
2975 .unmap_sg
= unmap_sg
,
2976 .dma_supported
= amd_iommu_dma_supported
,
2979 static unsigned device_dma_ops_init(void)
2981 struct iommu_dev_data
*dev_data
;
2982 struct pci_dev
*pdev
= NULL
;
2983 unsigned unhandled
= 0;
2985 for_each_pci_dev(pdev
) {
2986 if (!check_device(&pdev
->dev
)) {
2988 iommu_ignore_device(&pdev
->dev
);
2994 dev_data
= get_dev_data(&pdev
->dev
);
2996 if (!dev_data
->passthrough
)
2997 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
2999 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
3006 * The function which clues the AMD IOMMU driver into dma_ops.
3009 void __init
amd_iommu_init_api(void)
3011 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
3014 int __init
amd_iommu_init_dma_ops(void)
3016 struct amd_iommu
*iommu
;
3020 * first allocate a default protection domain for every IOMMU we
3021 * found in the system. Devices not assigned to any other
3022 * protection domain will be assigned to the default one.
3024 for_each_iommu(iommu
) {
3025 iommu
->default_dom
= dma_ops_domain_alloc();
3026 if (iommu
->default_dom
== NULL
)
3028 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
3029 ret
= iommu_init_unity_mappings(iommu
);
3035 * Pre-allocate the protection domains for each device.
3037 prealloc_protection_domains();
3042 /* Make the driver finally visible to the drivers */
3043 unhandled
= device_dma_ops_init();
3044 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
3045 /* There are unhandled devices - initialize swiotlb for them */
3049 amd_iommu_stats_init();
3051 if (amd_iommu_unmap_flush
)
3052 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3054 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3060 for_each_iommu(iommu
) {
3061 if (iommu
->default_dom
)
3062 dma_ops_domain_free(iommu
->default_dom
);
3068 /*****************************************************************************
3070 * The following functions belong to the exported interface of AMD IOMMU
3072 * This interface allows access to lower level functions of the IOMMU
3073 * like protection domain handling and assignement of devices to domains
3074 * which is not possible with the dma_ops interface.
3076 *****************************************************************************/
3078 static void cleanup_domain(struct protection_domain
*domain
)
3080 struct iommu_dev_data
*dev_data
, *next
;
3081 unsigned long flags
;
3083 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
3085 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
3086 __detach_device(dev_data
);
3087 atomic_set(&dev_data
->bind
, 0);
3090 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
3093 static void protection_domain_free(struct protection_domain
*domain
)
3098 del_domain_from_list(domain
);
3101 domain_id_free(domain
->id
);
3106 static struct protection_domain
*protection_domain_alloc(void)
3108 struct protection_domain
*domain
;
3110 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
3114 spin_lock_init(&domain
->lock
);
3115 mutex_init(&domain
->api_lock
);
3116 domain
->id
= domain_id_alloc();
3119 INIT_LIST_HEAD(&domain
->dev_list
);
3121 add_domain_to_list(domain
);
3131 static int __init
alloc_passthrough_domain(void)
3133 if (pt_domain
!= NULL
)
3136 /* allocate passthrough domain */
3137 pt_domain
= protection_domain_alloc();
3141 pt_domain
->mode
= PAGE_MODE_NONE
;
3145 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
3147 struct protection_domain
*domain
;
3149 domain
= protection_domain_alloc();
3153 domain
->mode
= PAGE_MODE_3_LEVEL
;
3154 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
3155 if (!domain
->pt_root
)
3158 domain
->iommu_domain
= dom
;
3162 dom
->geometry
.aperture_start
= 0;
3163 dom
->geometry
.aperture_end
= ~0ULL;
3164 dom
->geometry
.force_aperture
= true;
3169 protection_domain_free(domain
);
3174 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
3176 struct protection_domain
*domain
= dom
->priv
;
3181 if (domain
->dev_cnt
> 0)
3182 cleanup_domain(domain
);
3184 BUG_ON(domain
->dev_cnt
!= 0);
3186 if (domain
->mode
!= PAGE_MODE_NONE
)
3187 free_pagetable(domain
);
3189 if (domain
->flags
& PD_IOMMUV2_MASK
)
3190 free_gcr3_table(domain
);
3192 protection_domain_free(domain
);
3197 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
3200 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
3201 struct amd_iommu
*iommu
;
3204 if (!check_device(dev
))
3207 devid
= get_device_id(dev
);
3209 if (dev_data
->domain
!= NULL
)
3212 iommu
= amd_iommu_rlookup_table
[devid
];
3216 iommu_completion_wait(iommu
);
3219 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
3222 struct protection_domain
*domain
= dom
->priv
;
3223 struct iommu_dev_data
*dev_data
;
3224 struct amd_iommu
*iommu
;
3227 if (!check_device(dev
))
3230 dev_data
= dev
->archdata
.iommu
;
3232 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3236 if (dev_data
->domain
)
3239 ret
= attach_device(dev
, domain
);
3241 iommu_completion_wait(iommu
);
3246 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
3247 phys_addr_t paddr
, size_t page_size
, int iommu_prot
)
3249 struct protection_domain
*domain
= dom
->priv
;
3253 if (domain
->mode
== PAGE_MODE_NONE
)
3256 if (iommu_prot
& IOMMU_READ
)
3257 prot
|= IOMMU_PROT_IR
;
3258 if (iommu_prot
& IOMMU_WRITE
)
3259 prot
|= IOMMU_PROT_IW
;
3261 mutex_lock(&domain
->api_lock
);
3262 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
3263 mutex_unlock(&domain
->api_lock
);
3268 static size_t amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
3271 struct protection_domain
*domain
= dom
->priv
;
3274 if (domain
->mode
== PAGE_MODE_NONE
)
3277 mutex_lock(&domain
->api_lock
);
3278 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3279 mutex_unlock(&domain
->api_lock
);
3281 domain_flush_tlb_pde(domain
);
3286 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3289 struct protection_domain
*domain
= dom
->priv
;
3290 unsigned long offset_mask
;
3294 if (domain
->mode
== PAGE_MODE_NONE
)
3297 pte
= fetch_pte(domain
, iova
);
3299 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3302 if (PM_PTE_LEVEL(*pte
) == 0)
3303 offset_mask
= PAGE_SIZE
- 1;
3305 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
3307 __pte
= *pte
& PM_ADDR_MASK
;
3308 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3313 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
3317 case IOMMU_CAP_CACHE_COHERENCY
:
3324 static struct iommu_ops amd_iommu_ops
= {
3325 .domain_init
= amd_iommu_domain_init
,
3326 .domain_destroy
= amd_iommu_domain_destroy
,
3327 .attach_dev
= amd_iommu_attach_device
,
3328 .detach_dev
= amd_iommu_detach_device
,
3329 .map
= amd_iommu_map
,
3330 .unmap
= amd_iommu_unmap
,
3331 .iova_to_phys
= amd_iommu_iova_to_phys
,
3332 .domain_has_cap
= amd_iommu_domain_has_cap
,
3333 .pgsize_bitmap
= AMD_IOMMU_PGSIZES
,
3336 /*****************************************************************************
3338 * The next functions do a basic initialization of IOMMU for pass through
3341 * In passthrough mode the IOMMU is initialized and enabled but not used for
3342 * DMA-API translation.
3344 *****************************************************************************/
3346 int __init
amd_iommu_init_passthrough(void)
3348 struct iommu_dev_data
*dev_data
;
3349 struct pci_dev
*dev
= NULL
;
3350 struct amd_iommu
*iommu
;
3354 ret
= alloc_passthrough_domain();
3358 for_each_pci_dev(dev
) {
3359 if (!check_device(&dev
->dev
))
3362 dev_data
= get_dev_data(&dev
->dev
);
3363 dev_data
->passthrough
= true;
3365 devid
= get_device_id(&dev
->dev
);
3367 iommu
= amd_iommu_rlookup_table
[devid
];
3371 attach_device(&dev
->dev
, pt_domain
);
3374 amd_iommu_stats_init();
3376 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3381 /* IOMMUv2 specific functions */
3382 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3384 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3386 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3388 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3390 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3392 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3394 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3396 struct protection_domain
*domain
= dom
->priv
;
3397 unsigned long flags
;
3399 spin_lock_irqsave(&domain
->lock
, flags
);
3401 /* Update data structure */
3402 domain
->mode
= PAGE_MODE_NONE
;
3403 domain
->updated
= true;
3405 /* Make changes visible to IOMMUs */
3406 update_domain(domain
);
3408 /* Page-table is not visible to IOMMU anymore, so free it */
3409 free_pagetable(domain
);
3411 spin_unlock_irqrestore(&domain
->lock
, flags
);
3413 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3415 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3417 struct protection_domain
*domain
= dom
->priv
;
3418 unsigned long flags
;
3421 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3424 /* Number of GCR3 table levels required */
3425 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3428 if (levels
> amd_iommu_max_glx_val
)
3431 spin_lock_irqsave(&domain
->lock
, flags
);
3434 * Save us all sanity checks whether devices already in the
3435 * domain support IOMMUv2. Just force that the domain has no
3436 * devices attached when it is switched into IOMMUv2 mode.
3439 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3443 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3444 if (domain
->gcr3_tbl
== NULL
)
3447 domain
->glx
= levels
;
3448 domain
->flags
|= PD_IOMMUV2_MASK
;
3449 domain
->updated
= true;
3451 update_domain(domain
);
3456 spin_unlock_irqrestore(&domain
->lock
, flags
);
3460 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3462 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3463 u64 address
, bool size
)
3465 struct iommu_dev_data
*dev_data
;
3466 struct iommu_cmd cmd
;
3469 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3472 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3475 * IOMMU TLB needs to be flushed before Device TLB to
3476 * prevent device TLB refill from IOMMU TLB
3478 for (i
= 0; i
< amd_iommus_present
; ++i
) {
3479 if (domain
->dev_iommu
[i
] == 0)
3482 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3487 /* Wait until IOMMU TLB flushes are complete */
3488 domain_flush_complete(domain
);
3490 /* Now flush device TLBs */
3491 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3492 struct amd_iommu
*iommu
;
3495 BUG_ON(!dev_data
->ats
.enabled
);
3497 qdep
= dev_data
->ats
.qdep
;
3498 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3500 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3501 qdep
, address
, size
);
3503 ret
= iommu_queue_command(iommu
, &cmd
);
3508 /* Wait until all device TLBs are flushed */
3509 domain_flush_complete(domain
);
3518 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3521 INC_STATS_COUNTER(invalidate_iotlb
);
3523 return __flush_pasid(domain
, pasid
, address
, false);
3526 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3529 struct protection_domain
*domain
= dom
->priv
;
3530 unsigned long flags
;
3533 spin_lock_irqsave(&domain
->lock
, flags
);
3534 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3535 spin_unlock_irqrestore(&domain
->lock
, flags
);
3539 EXPORT_SYMBOL(amd_iommu_flush_page
);
3541 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3543 INC_STATS_COUNTER(invalidate_iotlb_all
);
3545 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3549 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3551 struct protection_domain
*domain
= dom
->priv
;
3552 unsigned long flags
;
3555 spin_lock_irqsave(&domain
->lock
, flags
);
3556 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3557 spin_unlock_irqrestore(&domain
->lock
, flags
);
3561 EXPORT_SYMBOL(amd_iommu_flush_tlb
);
3563 static u64
*__get_gcr3_pte(u64
*root
, int level
, int pasid
, bool alloc
)
3570 index
= (pasid
>> (9 * level
)) & 0x1ff;
3576 if (!(*pte
& GCR3_VALID
)) {
3580 root
= (void *)get_zeroed_page(GFP_ATOMIC
);
3584 *pte
= __pa(root
) | GCR3_VALID
;
3587 root
= __va(*pte
& PAGE_MASK
);
3595 static int __set_gcr3(struct protection_domain
*domain
, int pasid
,
3600 if (domain
->mode
!= PAGE_MODE_NONE
)
3603 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, true);
3607 *pte
= (cr3
& PAGE_MASK
) | GCR3_VALID
;
3609 return __amd_iommu_flush_tlb(domain
, pasid
);
3612 static int __clear_gcr3(struct protection_domain
*domain
, int pasid
)
3616 if (domain
->mode
!= PAGE_MODE_NONE
)
3619 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, false);
3625 return __amd_iommu_flush_tlb(domain
, pasid
);
3628 int amd_iommu_domain_set_gcr3(struct iommu_domain
*dom
, int pasid
,
3631 struct protection_domain
*domain
= dom
->priv
;
3632 unsigned long flags
;
3635 spin_lock_irqsave(&domain
->lock
, flags
);
3636 ret
= __set_gcr3(domain
, pasid
, cr3
);
3637 spin_unlock_irqrestore(&domain
->lock
, flags
);
3641 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3
);
3643 int amd_iommu_domain_clear_gcr3(struct iommu_domain
*dom
, int pasid
)
3645 struct protection_domain
*domain
= dom
->priv
;
3646 unsigned long flags
;
3649 spin_lock_irqsave(&domain
->lock
, flags
);
3650 ret
= __clear_gcr3(domain
, pasid
);
3651 spin_unlock_irqrestore(&domain
->lock
, flags
);
3655 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3
);
3657 int amd_iommu_complete_ppr(struct pci_dev
*pdev
, int pasid
,
3658 int status
, int tag
)
3660 struct iommu_dev_data
*dev_data
;
3661 struct amd_iommu
*iommu
;
3662 struct iommu_cmd cmd
;
3664 INC_STATS_COUNTER(complete_ppr
);
3666 dev_data
= get_dev_data(&pdev
->dev
);
3667 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3669 build_complete_ppr(&cmd
, dev_data
->devid
, pasid
, status
,
3670 tag
, dev_data
->pri_tlp
);
3672 return iommu_queue_command(iommu
, &cmd
);
3674 EXPORT_SYMBOL(amd_iommu_complete_ppr
);
3676 struct iommu_domain
*amd_iommu_get_v2_domain(struct pci_dev
*pdev
)
3678 struct protection_domain
*domain
;
3680 domain
= get_domain(&pdev
->dev
);
3684 /* Only return IOMMUv2 domains */
3685 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3688 return domain
->iommu_domain
;
3690 EXPORT_SYMBOL(amd_iommu_get_v2_domain
);
3692 void amd_iommu_enable_device_erratum(struct pci_dev
*pdev
, u32 erratum
)
3694 struct iommu_dev_data
*dev_data
;
3696 if (!amd_iommu_v2_supported())
3699 dev_data
= get_dev_data(&pdev
->dev
);
3700 dev_data
->errata
|= (1 << erratum
);
3702 EXPORT_SYMBOL(amd_iommu_enable_device_erratum
);
3704 int amd_iommu_device_info(struct pci_dev
*pdev
,
3705 struct amd_iommu_device_info
*info
)
3710 if (pdev
== NULL
|| info
== NULL
)
3713 if (!amd_iommu_v2_supported())
3716 memset(info
, 0, sizeof(*info
));
3718 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ATS
);
3720 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_ATS_SUP
;
3722 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
3724 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRI_SUP
;
3726 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PASID
);
3730 max_pasids
= 1 << (9 * (amd_iommu_max_glx_val
+ 1));
3731 max_pasids
= min(max_pasids
, (1 << 20));
3733 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PASID_SUP
;
3734 info
->max_pasids
= min(pci_max_pasids(pdev
), max_pasids
);
3736 features
= pci_pasid_features(pdev
);
3737 if (features
& PCI_PASID_CAP_EXEC
)
3738 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP
;
3739 if (features
& PCI_PASID_CAP_PRIV
)
3740 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP
;
3745 EXPORT_SYMBOL(amd_iommu_device_info
);