2 * VFIO: IOMMU DMA mapping support for TCE on POWER
4 * Copyright (C) 2013 IBM Corp. All rights reserved.
5 * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio_iommu_type1.c:
12 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
13 * Author: Alex Williamson <alex.williamson@redhat.com>
16 #include <linux/module.h>
17 #include <linux/pci.h>
18 #include <linux/slab.h>
19 #include <linux/uaccess.h>
20 #include <linux/err.h>
21 #include <linux/vfio.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sched/mm.h>
24 #include <linux/sched/signal.h>
26 #include <asm/iommu.h>
28 #include <asm/mmu_context.h>
30 #define DRIVER_VERSION "0.1"
31 #define DRIVER_AUTHOR "aik@ozlabs.ru"
32 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
34 static void tce_iommu_detach_group(void *iommu_data
,
35 struct iommu_group
*iommu_group
);
37 static long try_increment_locked_vm(struct mm_struct
*mm
, long npages
)
39 long ret
= 0, locked
, lock_limit
;
41 if (WARN_ON_ONCE(!mm
))
47 down_write(&mm
->mmap_sem
);
48 locked
= mm
->locked_vm
+ npages
;
49 lock_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
50 if (locked
> lock_limit
&& !capable(CAP_IPC_LOCK
))
53 mm
->locked_vm
+= npages
;
55 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current
->pid
,
57 mm
->locked_vm
<< PAGE_SHIFT
,
58 rlimit(RLIMIT_MEMLOCK
),
59 ret
? " - exceeded" : "");
61 up_write(&mm
->mmap_sem
);
66 static void decrement_locked_vm(struct mm_struct
*mm
, long npages
)
71 down_write(&mm
->mmap_sem
);
72 if (WARN_ON_ONCE(npages
> mm
->locked_vm
))
73 npages
= mm
->locked_vm
;
74 mm
->locked_vm
-= npages
;
75 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current
->pid
,
77 mm
->locked_vm
<< PAGE_SHIFT
,
78 rlimit(RLIMIT_MEMLOCK
));
79 up_write(&mm
->mmap_sem
);
83 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
85 * This code handles mapping and unmapping of user data buffers
86 * into DMA'ble space using the IOMMU
89 struct tce_iommu_group
{
90 struct list_head next
;
91 struct iommu_group
*grp
;
95 * A container needs to remember which preregistered region it has
96 * referenced to do proper cleanup at the userspace process exit.
98 struct tce_iommu_prereg
{
99 struct list_head next
;
100 struct mm_iommu_table_group_mem_t
*mem
;
104 * The container descriptor supports only a single group per container.
105 * Required by the API as the container is not supplied with the IOMMU group
106 * at the moment of initialization.
108 struct tce_container
{
112 bool def_window_pending
;
113 unsigned long locked_pages
;
114 struct mm_struct
*mm
;
115 struct iommu_table
*tables
[IOMMU_TABLE_GROUP_MAX_TABLES
];
116 struct list_head group_list
;
117 struct list_head prereg_list
;
120 static long tce_iommu_mm_set(struct tce_container
*container
)
123 if (container
->mm
== current
->mm
)
127 BUG_ON(!current
->mm
);
128 container
->mm
= current
->mm
;
129 atomic_inc(&container
->mm
->mm_count
);
134 static long tce_iommu_prereg_free(struct tce_container
*container
,
135 struct tce_iommu_prereg
*tcemem
)
139 ret
= mm_iommu_put(container
->mm
, tcemem
->mem
);
143 list_del(&tcemem
->next
);
149 static long tce_iommu_unregister_pages(struct tce_container
*container
,
150 __u64 vaddr
, __u64 size
)
152 struct mm_iommu_table_group_mem_t
*mem
;
153 struct tce_iommu_prereg
*tcemem
;
157 if ((vaddr
& ~PAGE_MASK
) || (size
& ~PAGE_MASK
))
160 mem
= mm_iommu_get(container
->mm
, vaddr
, size
>> PAGE_SHIFT
);
164 list_for_each_entry(tcemem
, &container
->prereg_list
, next
) {
165 if (tcemem
->mem
== mem
) {
174 ret
= tce_iommu_prereg_free(container
, tcemem
);
176 mm_iommu_put(container
->mm
, mem
);
181 static long tce_iommu_register_pages(struct tce_container
*container
,
182 __u64 vaddr
, __u64 size
)
185 struct mm_iommu_table_group_mem_t
*mem
= NULL
;
186 struct tce_iommu_prereg
*tcemem
;
187 unsigned long entries
= size
>> PAGE_SHIFT
;
189 if ((vaddr
& ~PAGE_MASK
) || (size
& ~PAGE_MASK
) ||
190 ((vaddr
+ size
) < vaddr
))
193 mem
= mm_iommu_get(container
->mm
, vaddr
, entries
);
195 list_for_each_entry(tcemem
, &container
->prereg_list
, next
) {
196 if (tcemem
->mem
== mem
) {
202 ret
= mm_iommu_new(container
->mm
, vaddr
, entries
, &mem
);
207 tcemem
= kzalloc(sizeof(*tcemem
), GFP_KERNEL
);
214 list_add(&tcemem
->next
, &container
->prereg_list
);
216 container
->enabled
= true;
221 mm_iommu_put(container
->mm
, mem
);
225 static bool tce_page_is_contained(struct mm_struct
*mm
, unsigned long hpa
,
226 unsigned int page_shift
)
229 unsigned long size
= 0;
231 if (mm_iommu_is_devmem(mm
, hpa
, page_shift
, &size
))
232 return size
== (1UL << page_shift
);
234 page
= pfn_to_page(hpa
>> PAGE_SHIFT
);
236 * Check that the TCE table granularity is not bigger than the size of
237 * a page we just found. Otherwise the hardware can get access to
238 * a bigger memory chunk that it should.
240 return (PAGE_SHIFT
+ compound_order(compound_head(page
))) >= page_shift
;
243 static inline bool tce_groups_attached(struct tce_container
*container
)
245 return !list_empty(&container
->group_list
);
248 static long tce_iommu_find_table(struct tce_container
*container
,
249 phys_addr_t ioba
, struct iommu_table
**ptbl
)
253 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
254 struct iommu_table
*tbl
= container
->tables
[i
];
257 unsigned long entry
= ioba
>> tbl
->it_page_shift
;
258 unsigned long start
= tbl
->it_offset
;
259 unsigned long end
= start
+ tbl
->it_size
;
261 if ((start
<= entry
) && (entry
< end
)) {
271 static int tce_iommu_find_free_table(struct tce_container
*container
)
275 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
276 if (!container
->tables
[i
])
283 static int tce_iommu_enable(struct tce_container
*container
)
286 unsigned long locked
;
287 struct iommu_table_group
*table_group
;
288 struct tce_iommu_group
*tcegrp
;
290 if (container
->enabled
)
294 * When userspace pages are mapped into the IOMMU, they are effectively
295 * locked memory, so, theoretically, we need to update the accounting
296 * of locked pages on each map and unmap. For powerpc, the map unmap
297 * paths can be very hot, though, and the accounting would kill
298 * performance, especially since it would be difficult to impossible
299 * to handle the accounting in real mode only.
301 * To address that, rather than precisely accounting every page, we
302 * instead account for a worst case on locked memory when the iommu is
303 * enabled and disabled. The worst case upper bound on locked memory
304 * is the size of the whole iommu window, which is usually relatively
305 * small (compared to total memory sizes) on POWER hardware.
307 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
308 * that would effectively kill the guest at random points, much better
309 * enforcing the limit based on the max that the guest can map.
311 * Unfortunately at the moment it counts whole tables, no matter how
312 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
313 * each with 2GB DMA window, 8GB will be counted here. The reason for
314 * this is that we cannot tell here the amount of RAM used by the guest
315 * as this information is only available from KVM and VFIO is
318 * So we do not allow enabling a container without a group attached
319 * as there is no way to know how much we should increment
320 * the locked_vm counter.
322 if (!tce_groups_attached(container
))
325 tcegrp
= list_first_entry(&container
->group_list
,
326 struct tce_iommu_group
, next
);
327 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
331 if (!table_group
->tce32_size
)
334 ret
= tce_iommu_mm_set(container
);
338 locked
= table_group
->tce32_size
>> PAGE_SHIFT
;
339 ret
= try_increment_locked_vm(container
->mm
, locked
);
343 container
->locked_pages
= locked
;
345 container
->enabled
= true;
350 static void tce_iommu_disable(struct tce_container
*container
)
352 if (!container
->enabled
)
355 container
->enabled
= false;
357 BUG_ON(!container
->mm
);
358 decrement_locked_vm(container
->mm
, container
->locked_pages
);
361 static void *tce_iommu_open(unsigned long arg
)
363 struct tce_container
*container
;
365 if ((arg
!= VFIO_SPAPR_TCE_IOMMU
) && (arg
!= VFIO_SPAPR_TCE_v2_IOMMU
)) {
366 pr_err("tce_vfio: Wrong IOMMU type\n");
367 return ERR_PTR(-EINVAL
);
370 container
= kzalloc(sizeof(*container
), GFP_KERNEL
);
372 return ERR_PTR(-ENOMEM
);
374 mutex_init(&container
->lock
);
375 INIT_LIST_HEAD_RCU(&container
->group_list
);
376 INIT_LIST_HEAD_RCU(&container
->prereg_list
);
378 container
->v2
= arg
== VFIO_SPAPR_TCE_v2_IOMMU
;
383 static int tce_iommu_clear(struct tce_container
*container
,
384 struct iommu_table
*tbl
,
385 unsigned long entry
, unsigned long pages
);
386 static void tce_iommu_free_table(struct tce_container
*container
,
387 struct iommu_table
*tbl
);
389 static void tce_iommu_release(void *iommu_data
)
391 struct tce_container
*container
= iommu_data
;
392 struct tce_iommu_group
*tcegrp
;
393 struct tce_iommu_prereg
*tcemem
, *tmtmp
;
396 while (tce_groups_attached(container
)) {
397 tcegrp
= list_first_entry(&container
->group_list
,
398 struct tce_iommu_group
, next
);
399 tce_iommu_detach_group(iommu_data
, tcegrp
->grp
);
403 * If VFIO created a table, it was not disposed
404 * by tce_iommu_detach_group() so do it now.
406 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
407 struct iommu_table
*tbl
= container
->tables
[i
];
412 tce_iommu_clear(container
, tbl
, tbl
->it_offset
, tbl
->it_size
);
413 tce_iommu_free_table(container
, tbl
);
416 list_for_each_entry_safe(tcemem
, tmtmp
, &container
->prereg_list
, next
)
417 WARN_ON(tce_iommu_prereg_free(container
, tcemem
));
419 tce_iommu_disable(container
);
421 mmdrop(container
->mm
);
422 mutex_destroy(&container
->lock
);
427 static void tce_iommu_unuse_page(struct tce_container
*container
,
432 page
= pfn_to_page(hpa
>> PAGE_SHIFT
);
436 static int tce_iommu_prereg_ua_to_hpa(struct tce_container
*container
,
437 unsigned long tce
, unsigned long shift
,
438 unsigned long *phpa
, struct mm_iommu_table_group_mem_t
**pmem
)
441 struct mm_iommu_table_group_mem_t
*mem
;
443 mem
= mm_iommu_lookup(container
->mm
, tce
, 1ULL << shift
);
447 ret
= mm_iommu_ua_to_hpa(mem
, tce
, shift
, phpa
);
456 static void tce_iommu_unuse_page_v2(struct tce_container
*container
,
457 struct iommu_table
*tbl
, unsigned long entry
)
459 struct mm_iommu_table_group_mem_t
*mem
= NULL
;
461 unsigned long hpa
= 0;
462 __be64
*pua
= IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl
, entry
);
467 ret
= tce_iommu_prereg_ua_to_hpa(container
, be64_to_cpu(*pua
),
468 tbl
->it_page_shift
, &hpa
, &mem
);
470 pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
471 __func__
, be64_to_cpu(*pua
), entry
, ret
);
473 mm_iommu_mapped_dec(mem
);
475 *pua
= cpu_to_be64(0);
478 static int tce_iommu_clear(struct tce_container
*container
,
479 struct iommu_table
*tbl
,
480 unsigned long entry
, unsigned long pages
)
482 unsigned long oldhpa
;
484 enum dma_data_direction direction
;
485 unsigned long lastentry
= entry
+ pages
;
487 for ( ; entry
< lastentry
; ++entry
) {
488 if (tbl
->it_indirect_levels
&& tbl
->it_userspace
) {
490 * For multilevel tables, we can take a shortcut here
491 * and skip some TCEs as we know that the userspace
492 * addresses cache is a mirror of the real TCE table
493 * and if it is missing some indirect levels, then
494 * the hardware table does not have them allocated
495 * either and therefore does not require updating.
497 __be64
*pua
= IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl
,
500 /* align to level_size which is power of two */
501 entry
|= tbl
->it_level_size
- 1;
508 direction
= DMA_NONE
;
510 ret
= iommu_tce_xchg(container
->mm
, tbl
, entry
, &oldhpa
,
515 if (direction
== DMA_NONE
)
519 tce_iommu_unuse_page_v2(container
, tbl
, entry
);
523 tce_iommu_unuse_page(container
, oldhpa
);
529 static int tce_iommu_use_page(unsigned long tce
, unsigned long *hpa
)
531 struct page
*page
= NULL
;
532 enum dma_data_direction direction
= iommu_tce_direction(tce
);
534 if (get_user_pages_fast(tce
& PAGE_MASK
, 1,
535 direction
!= DMA_TO_DEVICE
, &page
) != 1)
538 *hpa
= __pa((unsigned long) page_address(page
));
543 static long tce_iommu_build(struct tce_container
*container
,
544 struct iommu_table
*tbl
,
545 unsigned long entry
, unsigned long tce
, unsigned long pages
,
546 enum dma_data_direction direction
)
550 enum dma_data_direction dirtmp
;
552 for (i
= 0; i
< pages
; ++i
) {
553 unsigned long offset
= tce
& IOMMU_PAGE_MASK(tbl
) & ~PAGE_MASK
;
555 ret
= tce_iommu_use_page(tce
, &hpa
);
559 if (!tce_page_is_contained(container
->mm
, hpa
,
560 tbl
->it_page_shift
)) {
567 ret
= iommu_tce_xchg(container
->mm
, tbl
, entry
+ i
, &hpa
,
570 tce_iommu_unuse_page(container
, hpa
);
571 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
572 __func__
, entry
<< tbl
->it_page_shift
,
577 if (dirtmp
!= DMA_NONE
)
578 tce_iommu_unuse_page(container
, hpa
);
580 tce
+= IOMMU_PAGE_SIZE(tbl
);
584 tce_iommu_clear(container
, tbl
, entry
, i
);
589 static long tce_iommu_build_v2(struct tce_container
*container
,
590 struct iommu_table
*tbl
,
591 unsigned long entry
, unsigned long tce
, unsigned long pages
,
592 enum dma_data_direction direction
)
596 enum dma_data_direction dirtmp
;
598 for (i
= 0; i
< pages
; ++i
) {
599 struct mm_iommu_table_group_mem_t
*mem
= NULL
;
600 __be64
*pua
= IOMMU_TABLE_USERSPACE_ENTRY(tbl
, entry
+ i
);
602 ret
= tce_iommu_prereg_ua_to_hpa(container
,
603 tce
, tbl
->it_page_shift
, &hpa
, &mem
);
607 if (!tce_page_is_contained(container
->mm
, hpa
,
608 tbl
->it_page_shift
)) {
613 /* Preserve offset within IOMMU page */
614 hpa
|= tce
& IOMMU_PAGE_MASK(tbl
) & ~PAGE_MASK
;
617 /* The registered region is being unregistered */
618 if (mm_iommu_mapped_inc(mem
))
621 ret
= iommu_tce_xchg(container
->mm
, tbl
, entry
+ i
, &hpa
,
624 /* dirtmp cannot be DMA_NONE here */
625 tce_iommu_unuse_page_v2(container
, tbl
, entry
+ i
);
626 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
627 __func__
, entry
<< tbl
->it_page_shift
,
632 if (dirtmp
!= DMA_NONE
)
633 tce_iommu_unuse_page_v2(container
, tbl
, entry
+ i
);
635 *pua
= cpu_to_be64(tce
);
637 tce
+= IOMMU_PAGE_SIZE(tbl
);
641 tce_iommu_clear(container
, tbl
, entry
, i
);
646 static long tce_iommu_create_table(struct tce_container
*container
,
647 struct iommu_table_group
*table_group
,
652 struct iommu_table
**ptbl
)
654 long ret
, table_size
;
656 table_size
= table_group
->ops
->get_table_size(page_shift
, window_size
,
661 ret
= try_increment_locked_vm(container
->mm
, table_size
>> PAGE_SHIFT
);
665 ret
= table_group
->ops
->create_table(table_group
, num
,
666 page_shift
, window_size
, levels
, ptbl
);
668 WARN_ON(!ret
&& !(*ptbl
)->it_ops
->free
);
669 WARN_ON(!ret
&& ((*ptbl
)->it_allocated_size
> table_size
));
674 static void tce_iommu_free_table(struct tce_container
*container
,
675 struct iommu_table
*tbl
)
677 unsigned long pages
= tbl
->it_allocated_size
>> PAGE_SHIFT
;
679 iommu_tce_table_put(tbl
);
680 decrement_locked_vm(container
->mm
, pages
);
683 static long tce_iommu_create_window(struct tce_container
*container
,
684 __u32 page_shift
, __u64 window_size
, __u32 levels
,
687 struct tce_iommu_group
*tcegrp
;
688 struct iommu_table_group
*table_group
;
689 struct iommu_table
*tbl
= NULL
;
692 num
= tce_iommu_find_free_table(container
);
696 /* Get the first group for ops::create_table */
697 tcegrp
= list_first_entry(&container
->group_list
,
698 struct tce_iommu_group
, next
);
699 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
703 if (!(table_group
->pgsizes
& (1ULL << page_shift
)))
706 if (!table_group
->ops
->set_window
|| !table_group
->ops
->unset_window
||
707 !table_group
->ops
->get_table_size
||
708 !table_group
->ops
->create_table
)
711 /* Create TCE table */
712 ret
= tce_iommu_create_table(container
, table_group
, num
,
713 page_shift
, window_size
, levels
, &tbl
);
717 BUG_ON(!tbl
->it_ops
->free
);
720 * Program the table to every group.
721 * Groups have been tested for compatibility at the attach time.
723 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
724 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
726 ret
= table_group
->ops
->set_window(table_group
, num
, tbl
);
731 container
->tables
[num
] = tbl
;
733 /* Return start address assigned by platform in create_table() */
734 *start_addr
= tbl
->it_offset
<< tbl
->it_page_shift
;
739 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
740 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
741 table_group
->ops
->unset_window(table_group
, num
);
743 tce_iommu_free_table(container
, tbl
);
748 static long tce_iommu_remove_window(struct tce_container
*container
,
751 struct iommu_table_group
*table_group
= NULL
;
752 struct iommu_table
*tbl
;
753 struct tce_iommu_group
*tcegrp
;
756 num
= tce_iommu_find_table(container
, start_addr
, &tbl
);
760 BUG_ON(!tbl
->it_size
);
762 /* Detach groups from IOMMUs */
763 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
764 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
767 * SPAPR TCE IOMMU exposes the default DMA window to
768 * the guest via dma32_window_start/size of
769 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
770 * the userspace to remove this window, some do not so
771 * here we check for the platform capability.
773 if (!table_group
->ops
|| !table_group
->ops
->unset_window
)
776 table_group
->ops
->unset_window(table_group
, num
);
780 tce_iommu_clear(container
, tbl
, tbl
->it_offset
, tbl
->it_size
);
781 tce_iommu_free_table(container
, tbl
);
782 container
->tables
[num
] = NULL
;
787 static long tce_iommu_create_default_window(struct tce_container
*container
)
790 __u64 start_addr
= 0;
791 struct tce_iommu_group
*tcegrp
;
792 struct iommu_table_group
*table_group
;
794 if (!container
->def_window_pending
)
797 if (!tce_groups_attached(container
))
800 tcegrp
= list_first_entry(&container
->group_list
,
801 struct tce_iommu_group
, next
);
802 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
806 ret
= tce_iommu_create_window(container
, IOMMU_PAGE_SHIFT_4K
,
807 table_group
->tce32_size
, 1, &start_addr
);
808 WARN_ON_ONCE(!ret
&& start_addr
);
811 container
->def_window_pending
= false;
816 static long tce_iommu_ioctl(void *iommu_data
,
817 unsigned int cmd
, unsigned long arg
)
819 struct tce_container
*container
= iommu_data
;
820 unsigned long minsz
, ddwsz
;
824 case VFIO_CHECK_EXTENSION
:
826 case VFIO_SPAPR_TCE_IOMMU
:
827 case VFIO_SPAPR_TCE_v2_IOMMU
:
831 ret
= vfio_spapr_iommu_eeh_ioctl(NULL
, cmd
, arg
);
835 return (ret
< 0) ? 0 : ret
;
839 * Sanity check to prevent one userspace from manipulating
840 * another userspace mm.
843 if (container
->mm
&& container
->mm
!= current
->mm
)
847 case VFIO_IOMMU_SPAPR_TCE_GET_INFO
: {
848 struct vfio_iommu_spapr_tce_info info
;
849 struct tce_iommu_group
*tcegrp
;
850 struct iommu_table_group
*table_group
;
852 if (!tce_groups_attached(container
))
855 tcegrp
= list_first_entry(&container
->group_list
,
856 struct tce_iommu_group
, next
);
857 table_group
= iommu_group_get_iommudata(tcegrp
->grp
);
862 minsz
= offsetofend(struct vfio_iommu_spapr_tce_info
,
865 if (copy_from_user(&info
, (void __user
*)arg
, minsz
))
868 if (info
.argsz
< minsz
)
871 info
.dma32_window_start
= table_group
->tce32_start
;
872 info
.dma32_window_size
= table_group
->tce32_size
;
874 memset(&info
.ddw
, 0, sizeof(info
.ddw
));
876 if (table_group
->max_dynamic_windows_supported
&&
878 info
.flags
|= VFIO_IOMMU_SPAPR_INFO_DDW
;
879 info
.ddw
.pgsizes
= table_group
->pgsizes
;
880 info
.ddw
.max_dynamic_windows_supported
=
881 table_group
->max_dynamic_windows_supported
;
882 info
.ddw
.levels
= table_group
->max_levels
;
885 ddwsz
= offsetofend(struct vfio_iommu_spapr_tce_info
, ddw
);
887 if (info
.argsz
>= ddwsz
)
890 if (copy_to_user((void __user
*)arg
, &info
, minsz
))
895 case VFIO_IOMMU_MAP_DMA
: {
896 struct vfio_iommu_type1_dma_map param
;
897 struct iommu_table
*tbl
= NULL
;
899 enum dma_data_direction direction
;
901 if (!container
->enabled
)
904 minsz
= offsetofend(struct vfio_iommu_type1_dma_map
, size
);
906 if (copy_from_user(¶m
, (void __user
*)arg
, minsz
))
909 if (param
.argsz
< minsz
)
912 if (param
.flags
& ~(VFIO_DMA_MAP_FLAG_READ
|
913 VFIO_DMA_MAP_FLAG_WRITE
))
916 ret
= tce_iommu_create_default_window(container
);
920 num
= tce_iommu_find_table(container
, param
.iova
, &tbl
);
924 if ((param
.size
& ~IOMMU_PAGE_MASK(tbl
)) ||
925 (param
.vaddr
& ~IOMMU_PAGE_MASK(tbl
)))
928 /* iova is checked by the IOMMU API */
929 if (param
.flags
& VFIO_DMA_MAP_FLAG_READ
) {
930 if (param
.flags
& VFIO_DMA_MAP_FLAG_WRITE
)
931 direction
= DMA_BIDIRECTIONAL
;
933 direction
= DMA_TO_DEVICE
;
935 if (param
.flags
& VFIO_DMA_MAP_FLAG_WRITE
)
936 direction
= DMA_FROM_DEVICE
;
941 ret
= iommu_tce_put_param_check(tbl
, param
.iova
, param
.vaddr
);
946 ret
= tce_iommu_build_v2(container
, tbl
,
947 param
.iova
>> tbl
->it_page_shift
,
949 param
.size
>> tbl
->it_page_shift
,
952 ret
= tce_iommu_build(container
, tbl
,
953 param
.iova
>> tbl
->it_page_shift
,
955 param
.size
>> tbl
->it_page_shift
,
958 iommu_flush_tce(tbl
);
962 case VFIO_IOMMU_UNMAP_DMA
: {
963 struct vfio_iommu_type1_dma_unmap param
;
964 struct iommu_table
*tbl
= NULL
;
967 if (!container
->enabled
)
970 minsz
= offsetofend(struct vfio_iommu_type1_dma_unmap
,
973 if (copy_from_user(¶m
, (void __user
*)arg
, minsz
))
976 if (param
.argsz
< minsz
)
979 /* No flag is supported now */
983 ret
= tce_iommu_create_default_window(container
);
987 num
= tce_iommu_find_table(container
, param
.iova
, &tbl
);
991 if (param
.size
& ~IOMMU_PAGE_MASK(tbl
))
994 ret
= iommu_tce_clear_param_check(tbl
, param
.iova
, 0,
995 param
.size
>> tbl
->it_page_shift
);
999 ret
= tce_iommu_clear(container
, tbl
,
1000 param
.iova
>> tbl
->it_page_shift
,
1001 param
.size
>> tbl
->it_page_shift
);
1002 iommu_flush_tce(tbl
);
1006 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY
: {
1007 struct vfio_iommu_spapr_register_memory param
;
1012 minsz
= offsetofend(struct vfio_iommu_spapr_register_memory
,
1015 ret
= tce_iommu_mm_set(container
);
1019 if (copy_from_user(¶m
, (void __user
*)arg
, minsz
))
1022 if (param
.argsz
< minsz
)
1025 /* No flag is supported now */
1029 mutex_lock(&container
->lock
);
1030 ret
= tce_iommu_register_pages(container
, param
.vaddr
,
1032 mutex_unlock(&container
->lock
);
1036 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY
: {
1037 struct vfio_iommu_spapr_register_memory param
;
1045 minsz
= offsetofend(struct vfio_iommu_spapr_register_memory
,
1048 if (copy_from_user(¶m
, (void __user
*)arg
, minsz
))
1051 if (param
.argsz
< minsz
)
1054 /* No flag is supported now */
1058 mutex_lock(&container
->lock
);
1059 ret
= tce_iommu_unregister_pages(container
, param
.vaddr
,
1061 mutex_unlock(&container
->lock
);
1065 case VFIO_IOMMU_ENABLE
:
1069 mutex_lock(&container
->lock
);
1070 ret
= tce_iommu_enable(container
);
1071 mutex_unlock(&container
->lock
);
1075 case VFIO_IOMMU_DISABLE
:
1079 mutex_lock(&container
->lock
);
1080 tce_iommu_disable(container
);
1081 mutex_unlock(&container
->lock
);
1084 case VFIO_EEH_PE_OP
: {
1085 struct tce_iommu_group
*tcegrp
;
1088 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
1089 ret
= vfio_spapr_iommu_eeh_ioctl(tcegrp
->grp
,
1097 case VFIO_IOMMU_SPAPR_TCE_CREATE
: {
1098 struct vfio_iommu_spapr_tce_create create
;
1103 ret
= tce_iommu_mm_set(container
);
1107 if (!tce_groups_attached(container
))
1110 minsz
= offsetofend(struct vfio_iommu_spapr_tce_create
,
1113 if (copy_from_user(&create
, (void __user
*)arg
, minsz
))
1116 if (create
.argsz
< minsz
)
1122 mutex_lock(&container
->lock
);
1124 ret
= tce_iommu_create_default_window(container
);
1126 ret
= tce_iommu_create_window(container
,
1128 create
.window_size
, create
.levels
,
1129 &create
.start_addr
);
1131 mutex_unlock(&container
->lock
);
1133 if (!ret
&& copy_to_user((void __user
*)arg
, &create
, minsz
))
1138 case VFIO_IOMMU_SPAPR_TCE_REMOVE
: {
1139 struct vfio_iommu_spapr_tce_remove remove
;
1144 ret
= tce_iommu_mm_set(container
);
1148 if (!tce_groups_attached(container
))
1151 minsz
= offsetofend(struct vfio_iommu_spapr_tce_remove
,
1154 if (copy_from_user(&remove
, (void __user
*)arg
, minsz
))
1157 if (remove
.argsz
< minsz
)
1163 if (container
->def_window_pending
&& !remove
.start_addr
) {
1164 container
->def_window_pending
= false;
1168 mutex_lock(&container
->lock
);
1170 ret
= tce_iommu_remove_window(container
, remove
.start_addr
);
1172 mutex_unlock(&container
->lock
);
1181 static void tce_iommu_release_ownership(struct tce_container
*container
,
1182 struct iommu_table_group
*table_group
)
1186 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
1187 struct iommu_table
*tbl
= container
->tables
[i
];
1192 tce_iommu_clear(container
, tbl
, tbl
->it_offset
, tbl
->it_size
);
1194 iommu_release_ownership(tbl
);
1196 container
->tables
[i
] = NULL
;
1200 static int tce_iommu_take_ownership(struct tce_container
*container
,
1201 struct iommu_table_group
*table_group
)
1205 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
1206 struct iommu_table
*tbl
= table_group
->tables
[i
];
1208 if (!tbl
|| !tbl
->it_map
)
1211 rc
= iommu_take_ownership(tbl
);
1213 for (j
= 0; j
< i
; ++j
)
1214 iommu_release_ownership(
1215 table_group
->tables
[j
]);
1221 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
)
1222 container
->tables
[i
] = table_group
->tables
[i
];
1227 static void tce_iommu_release_ownership_ddw(struct tce_container
*container
,
1228 struct iommu_table_group
*table_group
)
1232 if (!table_group
->ops
->unset_window
) {
1237 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
)
1238 if (container
->tables
[i
])
1239 table_group
->ops
->unset_window(table_group
, i
);
1241 table_group
->ops
->release_ownership(table_group
);
1244 static long tce_iommu_take_ownership_ddw(struct tce_container
*container
,
1245 struct iommu_table_group
*table_group
)
1249 if (!table_group
->ops
->create_table
|| !table_group
->ops
->set_window
||
1250 !table_group
->ops
->release_ownership
) {
1255 table_group
->ops
->take_ownership(table_group
);
1257 /* Set all windows to the new group */
1258 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
) {
1259 struct iommu_table
*tbl
= container
->tables
[i
];
1264 ret
= table_group
->ops
->set_window(table_group
, i
, tbl
);
1272 for (i
= 0; i
< IOMMU_TABLE_GROUP_MAX_TABLES
; ++i
)
1273 table_group
->ops
->unset_window(table_group
, i
);
1275 table_group
->ops
->release_ownership(table_group
);
1280 static int tce_iommu_attach_group(void *iommu_data
,
1281 struct iommu_group
*iommu_group
)
1284 struct tce_container
*container
= iommu_data
;
1285 struct iommu_table_group
*table_group
;
1286 struct tce_iommu_group
*tcegrp
= NULL
;
1288 mutex_lock(&container
->lock
);
1290 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1291 iommu_group_id(iommu_group), iommu_group); */
1292 table_group
= iommu_group_get_iommudata(iommu_group
);
1298 if (tce_groups_attached(container
) && (!table_group
->ops
||
1299 !table_group
->ops
->take_ownership
||
1300 !table_group
->ops
->release_ownership
)) {
1305 /* Check if new group has the same iommu_ops (i.e. compatible) */
1306 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
1307 struct iommu_table_group
*table_group_tmp
;
1309 if (tcegrp
->grp
== iommu_group
) {
1310 pr_warn("tce_vfio: Group %d is already attached\n",
1311 iommu_group_id(iommu_group
));
1315 table_group_tmp
= iommu_group_get_iommudata(tcegrp
->grp
);
1316 if (table_group_tmp
->ops
->create_table
!=
1317 table_group
->ops
->create_table
) {
1318 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1319 iommu_group_id(iommu_group
),
1320 iommu_group_id(tcegrp
->grp
));
1326 tcegrp
= kzalloc(sizeof(*tcegrp
), GFP_KERNEL
);
1332 if (!table_group
->ops
|| !table_group
->ops
->take_ownership
||
1333 !table_group
->ops
->release_ownership
) {
1334 if (container
->v2
) {
1338 ret
= tce_iommu_take_ownership(container
, table_group
);
1340 if (!container
->v2
) {
1344 ret
= tce_iommu_take_ownership_ddw(container
, table_group
);
1345 if (!tce_groups_attached(container
) && !container
->tables
[0])
1346 container
->def_window_pending
= true;
1350 tcegrp
->grp
= iommu_group
;
1351 list_add(&tcegrp
->next
, &container
->group_list
);
1358 mutex_unlock(&container
->lock
);
1363 static void tce_iommu_detach_group(void *iommu_data
,
1364 struct iommu_group
*iommu_group
)
1366 struct tce_container
*container
= iommu_data
;
1367 struct iommu_table_group
*table_group
;
1369 struct tce_iommu_group
*tcegrp
;
1371 mutex_lock(&container
->lock
);
1373 list_for_each_entry(tcegrp
, &container
->group_list
, next
) {
1374 if (tcegrp
->grp
== iommu_group
) {
1381 pr_warn("tce_vfio: detaching unattached group #%u\n",
1382 iommu_group_id(iommu_group
));
1386 list_del(&tcegrp
->next
);
1389 table_group
= iommu_group_get_iommudata(iommu_group
);
1390 BUG_ON(!table_group
);
1392 if (!table_group
->ops
|| !table_group
->ops
->release_ownership
)
1393 tce_iommu_release_ownership(container
, table_group
);
1395 tce_iommu_release_ownership_ddw(container
, table_group
);
1398 mutex_unlock(&container
->lock
);
1401 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops
= {
1402 .name
= "iommu-vfio-powerpc",
1403 .owner
= THIS_MODULE
,
1404 .open
= tce_iommu_open
,
1405 .release
= tce_iommu_release
,
1406 .ioctl
= tce_iommu_ioctl
,
1407 .attach_group
= tce_iommu_attach_group
,
1408 .detach_group
= tce_iommu_detach_group
,
1411 static int __init
tce_iommu_init(void)
1413 return vfio_register_iommu_driver(&tce_iommu_driver_ops
);
1416 static void __exit
tce_iommu_cleanup(void)
1418 vfio_unregister_iommu_driver(&tce_iommu_driver_ops
);
1421 module_init(tce_iommu_init
);
1422 module_exit(tce_iommu_cleanup
);
1424 MODULE_VERSION(DRIVER_VERSION
);
1425 MODULE_LICENSE("GPL v2");
1426 MODULE_AUTHOR(DRIVER_AUTHOR
);
1427 MODULE_DESCRIPTION(DRIVER_DESC
);