2 * generic functions used by VFIO devices
4 * Copyright Red Hat, Inc. 2012
7 * Alex Williamson <alex.williamson@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Based on qemu-kvm device-assignment:
13 * Adapted for KVM by Qumranet.
14 * Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15 * Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16 * Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17 * Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18 * Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
24 #include <linux/kvm.h>
26 #include <linux/vfio.h>
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/vfio.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
33 #include "qemu/error-report.h"
34 #include "qemu/range.h"
35 #include "sysemu/kvm.h"
37 #include "qapi/error.h"
39 struct vfio_group_head vfio_group_list
=
40 QLIST_HEAD_INITIALIZER(vfio_group_list
);
41 struct vfio_as_head vfio_address_spaces
=
42 QLIST_HEAD_INITIALIZER(vfio_address_spaces
);
46 * We have a single VFIO pseudo device per KVM VM. Once created it lives
47 * for the life of the VM. Closing the file descriptor only drops our
48 * reference to it and the device's reference to kvm. Therefore once
49 * initialized, this file descriptor is only released on QEMU exit and
50 * we'll re-use it should another vfio device be attached before then.
52 static int vfio_kvm_device_fd
= -1;
56 * Common VFIO interrupt disable
58 void vfio_disable_irqindex(VFIODevice
*vbasedev
, int index
)
60 struct vfio_irq_set irq_set
= {
61 .argsz
= sizeof(irq_set
),
62 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_TRIGGER
,
68 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
71 void vfio_unmask_single_irqindex(VFIODevice
*vbasedev
, int index
)
73 struct vfio_irq_set irq_set
= {
74 .argsz
= sizeof(irq_set
),
75 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_UNMASK
,
81 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
84 void vfio_mask_single_irqindex(VFIODevice
*vbasedev
, int index
)
86 struct vfio_irq_set irq_set
= {
87 .argsz
= sizeof(irq_set
),
88 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_MASK
,
94 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
98 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
100 void vfio_region_write(void *opaque
, hwaddr addr
,
101 uint64_t data
, unsigned size
)
103 VFIORegion
*region
= opaque
;
104 VFIODevice
*vbasedev
= region
->vbasedev
;
117 buf
.word
= cpu_to_le16(data
);
120 buf
.dword
= cpu_to_le32(data
);
123 buf
.qword
= cpu_to_le64(data
);
126 hw_error("vfio: unsupported write size, %d bytes", size
);
130 if (pwrite(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
131 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", 0x%"PRIx64
133 __func__
, vbasedev
->name
, region
->nr
,
137 trace_vfio_region_write(vbasedev
->name
, region
->nr
, addr
, data
, size
);
140 * A read or write to a BAR always signals an INTx EOI. This will
141 * do nothing if not pending (including not in INTx mode). We assume
142 * that a BAR access is in response to an interrupt and that BAR
143 * accesses will service the interrupt. Unfortunately, we don't know
144 * which access will service the interrupt, so we're potentially
145 * getting quite a few host interrupts per guest interrupt.
147 vbasedev
->ops
->vfio_eoi(vbasedev
);
150 uint64_t vfio_region_read(void *opaque
,
151 hwaddr addr
, unsigned size
)
153 VFIORegion
*region
= opaque
;
154 VFIODevice
*vbasedev
= region
->vbasedev
;
163 if (pread(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
164 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", %d) failed: %m",
165 __func__
, vbasedev
->name
, region
->nr
,
174 data
= le16_to_cpu(buf
.word
);
177 data
= le32_to_cpu(buf
.dword
);
180 data
= le64_to_cpu(buf
.qword
);
183 hw_error("vfio: unsupported read size, %d bytes", size
);
187 trace_vfio_region_read(vbasedev
->name
, region
->nr
, addr
, size
, data
);
189 /* Same as write above */
190 vbasedev
->ops
->vfio_eoi(vbasedev
);
195 const MemoryRegionOps vfio_region_ops
= {
196 .read
= vfio_region_read
,
197 .write
= vfio_region_write
,
198 .endianness
= DEVICE_LITTLE_ENDIAN
,
200 .min_access_size
= 1,
201 .max_access_size
= 8,
204 .min_access_size
= 1,
205 .max_access_size
= 8,
210 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
212 static int vfio_dma_unmap(VFIOContainer
*container
,
213 hwaddr iova
, ram_addr_t size
)
215 struct vfio_iommu_type1_dma_unmap unmap
= {
216 .argsz
= sizeof(unmap
),
222 if (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
223 error_report("VFIO_UNMAP_DMA: %d", -errno
);
230 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
231 ram_addr_t size
, void *vaddr
, bool readonly
)
233 struct vfio_iommu_type1_dma_map map
= {
234 .argsz
= sizeof(map
),
235 .flags
= VFIO_DMA_MAP_FLAG_READ
,
236 .vaddr
= (__u64
)(uintptr_t)vaddr
,
242 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
246 * Try the mapping, if it fails with EBUSY, unmap the region and try
247 * again. This shouldn't be necessary, but we sometimes see it in
250 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
251 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
) == 0 &&
252 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
256 error_report("VFIO_MAP_DMA: %d", -errno
);
260 static void vfio_host_win_add(VFIOContainer
*container
,
261 hwaddr min_iova
, hwaddr max_iova
,
262 uint64_t iova_pgsizes
)
264 VFIOHostDMAWindow
*hostwin
;
266 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
267 if (ranges_overlap(hostwin
->min_iova
,
268 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
270 max_iova
- min_iova
+ 1)) {
271 hw_error("%s: Overlapped IOMMU are not enabled", __func__
);
275 hostwin
= g_malloc0(sizeof(*hostwin
));
277 hostwin
->min_iova
= min_iova
;
278 hostwin
->max_iova
= max_iova
;
279 hostwin
->iova_pgsizes
= iova_pgsizes
;
280 QLIST_INSERT_HEAD(&container
->hostwin_list
, hostwin
, hostwin_next
);
283 static int vfio_host_win_del(VFIOContainer
*container
, hwaddr min_iova
,
286 VFIOHostDMAWindow
*hostwin
;
288 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
289 if (hostwin
->min_iova
== min_iova
&& hostwin
->max_iova
== max_iova
) {
290 QLIST_REMOVE(hostwin
, hostwin_next
);
298 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
300 return (!memory_region_is_ram(section
->mr
) &&
301 !memory_region_is_iommu(section
->mr
)) ||
303 * Sizing an enabled 64-bit BAR can cause spurious mappings to
304 * addresses in the upper part of the 64-bit address space. These
305 * are never accessed by the CPU and beyond the address width of
306 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
308 section
->offset_within_address_space
& (1ULL << 63);
311 /* Called with rcu_read_lock held. */
312 static bool vfio_get_vaddr(IOMMUTLBEntry
*iotlb
, void **vaddr
,
317 hwaddr len
= iotlb
->addr_mask
+ 1;
318 bool writable
= iotlb
->perm
& IOMMU_WO
;
321 * The IOMMU TLB entry we have just covers translation through
322 * this IOMMU to its immediate target. We need to translate
323 * it the rest of the way through to memory.
325 mr
= address_space_translate(&address_space_memory
,
326 iotlb
->translated_addr
,
327 &xlat
, &len
, writable
);
328 if (!memory_region_is_ram(mr
)) {
329 error_report("iommu map to non memory area %"HWADDR_PRIx
"",
335 * Translation truncates length to the IOMMU page size,
336 * check that it did not truncate too much.
338 if (len
& iotlb
->addr_mask
) {
339 error_report("iommu has granularity incompatible with target AS");
343 *vaddr
= memory_region_get_ram_ptr(mr
) + xlat
;
344 *read_only
= !writable
|| mr
->readonly
;
349 static void vfio_iommu_map_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
351 VFIOGuestIOMMU
*giommu
= container_of(n
, VFIOGuestIOMMU
, n
);
352 VFIOContainer
*container
= giommu
->container
;
353 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
358 trace_vfio_iommu_map_notify(iotlb
->perm
== IOMMU_NONE
? "UNMAP" : "MAP",
359 iova
, iova
+ iotlb
->addr_mask
);
361 if (iotlb
->target_as
!= &address_space_memory
) {
362 error_report("Wrong target AS \"%s\", only system memory is allowed",
363 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
369 if ((iotlb
->perm
& IOMMU_RW
) != IOMMU_NONE
) {
370 if (!vfio_get_vaddr(iotlb
, &vaddr
, &read_only
)) {
374 * vaddr is only valid until rcu_read_unlock(). But after
375 * vfio_dma_map has set up the mapping the pages will be
376 * pinned by the kernel. This makes sure that the RAM backend
377 * of vaddr will always be there, even if the memory object is
378 * destroyed and its backing memory munmap-ed.
380 ret
= vfio_dma_map(container
, iova
,
381 iotlb
->addr_mask
+ 1, vaddr
,
384 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
385 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
387 iotlb
->addr_mask
+ 1, vaddr
, ret
);
390 ret
= vfio_dma_unmap(container
, iova
, iotlb
->addr_mask
+ 1);
392 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
393 "0x%"HWADDR_PRIx
") = %d (%m)",
395 iotlb
->addr_mask
+ 1, ret
);
402 static void vfio_listener_region_add(MemoryListener
*listener
,
403 MemoryRegionSection
*section
)
405 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
407 Int128 llend
, llsize
;
410 VFIOHostDMAWindow
*hostwin
;
413 if (vfio_listener_skipped_section(section
)) {
414 trace_vfio_listener_region_add_skip(
415 section
->offset_within_address_space
,
416 section
->offset_within_address_space
+
417 int128_get64(int128_sub(section
->size
, int128_one())));
421 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
422 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
423 error_report("%s received unaligned region", __func__
);
427 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
428 llend
= int128_make64(section
->offset_within_address_space
);
429 llend
= int128_add(llend
, section
->size
);
430 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
432 if (int128_ge(int128_make64(iova
), llend
)) {
435 end
= int128_get64(int128_sub(llend
, int128_one()));
437 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
438 VFIOHostDMAWindow
*hostwin
;
441 /* For now intersections are not allowed, we may relax this later */
442 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
443 if (ranges_overlap(hostwin
->min_iova
,
444 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
445 section
->offset_within_address_space
,
446 int128_get64(section
->size
))) {
452 ret
= vfio_spapr_create_window(container
, section
, &pgsize
);
457 vfio_host_win_add(container
, section
->offset_within_address_space
,
458 section
->offset_within_address_space
+
459 int128_get64(section
->size
) - 1, pgsize
);
462 hostwin_found
= false;
463 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
464 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
465 hostwin_found
= true;
470 if (!hostwin_found
) {
471 error_report("vfio: IOMMU container %p can't map guest IOVA region"
472 " 0x%"HWADDR_PRIx
"..0x%"HWADDR_PRIx
,
473 container
, iova
, end
);
478 memory_region_ref(section
->mr
);
480 if (memory_region_is_iommu(section
->mr
)) {
481 VFIOGuestIOMMU
*giommu
;
483 trace_vfio_listener_region_add_iommu(iova
, end
);
485 * FIXME: For VFIO iommu types which have KVM acceleration to
486 * avoid bouncing all map/unmaps through qemu this way, this
487 * would be the right place to wire that up (tell the KVM
488 * device emulation the VFIO iommu handles to use).
490 giommu
= g_malloc0(sizeof(*giommu
));
491 giommu
->iommu
= section
->mr
;
492 giommu
->iommu_offset
= section
->offset_within_address_space
-
493 section
->offset_within_region
;
494 giommu
->container
= container
;
495 llend
= int128_add(int128_make64(section
->offset_within_region
),
497 llend
= int128_sub(llend
, int128_one());
498 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
500 section
->offset_within_region
,
501 int128_get64(llend
));
502 QLIST_INSERT_HEAD(&container
->giommu_list
, giommu
, giommu_next
);
504 memory_region_register_iommu_notifier(giommu
->iommu
, &giommu
->n
);
505 memory_region_iommu_replay(giommu
->iommu
, &giommu
->n
, false);
510 /* Here we assume that memory_region_is_ram(section->mr)==true */
512 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
513 section
->offset_within_region
+
514 (iova
- section
->offset_within_address_space
);
516 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
518 llsize
= int128_sub(llend
, int128_make64(iova
));
520 ret
= vfio_dma_map(container
, iova
, int128_get64(llsize
),
521 vaddr
, section
->readonly
);
523 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
524 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
525 container
, iova
, int128_get64(llsize
), vaddr
, ret
);
533 * On the initfn path, store the first error in the container so we
534 * can gracefully fail. Runtime, there's not much we can do other
535 * than throw a hardware error.
537 if (!container
->initialized
) {
538 if (!container
->error
) {
539 container
->error
= ret
;
542 hw_error("vfio: DMA mapping failed, unable to continue");
546 static void vfio_listener_region_del(MemoryListener
*listener
,
547 MemoryRegionSection
*section
)
549 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
551 Int128 llend
, llsize
;
554 if (vfio_listener_skipped_section(section
)) {
555 trace_vfio_listener_region_del_skip(
556 section
->offset_within_address_space
,
557 section
->offset_within_address_space
+
558 int128_get64(int128_sub(section
->size
, int128_one())));
562 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
563 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
564 error_report("%s received unaligned region", __func__
);
568 if (memory_region_is_iommu(section
->mr
)) {
569 VFIOGuestIOMMU
*giommu
;
571 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
572 if (giommu
->iommu
== section
->mr
&&
573 giommu
->n
.start
== section
->offset_within_region
) {
574 memory_region_unregister_iommu_notifier(giommu
->iommu
,
576 QLIST_REMOVE(giommu
, giommu_next
);
583 * FIXME: We assume the one big unmap below is adequate to
584 * remove any individual page mappings in the IOMMU which
585 * might have been copied into VFIO. This works for a page table
586 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
587 * That may not be true for all IOMMU types.
591 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
592 llend
= int128_make64(section
->offset_within_address_space
);
593 llend
= int128_add(llend
, section
->size
);
594 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
596 if (int128_ge(int128_make64(iova
), llend
)) {
599 end
= int128_get64(int128_sub(llend
, int128_one()));
601 llsize
= int128_sub(llend
, int128_make64(iova
));
603 trace_vfio_listener_region_del(iova
, end
);
605 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
));
606 memory_region_unref(section
->mr
);
608 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
609 "0x%"HWADDR_PRIx
") = %d (%m)",
610 container
, iova
, int128_get64(llsize
), ret
);
613 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
614 vfio_spapr_remove_window(container
,
615 section
->offset_within_address_space
);
616 if (vfio_host_win_del(container
,
617 section
->offset_within_address_space
,
618 section
->offset_within_address_space
+
619 int128_get64(section
->size
) - 1) < 0) {
620 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx
,
621 __func__
, section
->offset_within_address_space
);
626 static const MemoryListener vfio_memory_listener
= {
627 .region_add
= vfio_listener_region_add
,
628 .region_del
= vfio_listener_region_del
,
631 static void vfio_listener_release(VFIOContainer
*container
)
633 memory_listener_unregister(&container
->listener
);
634 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
635 memory_listener_unregister(&container
->prereg_listener
);
639 static struct vfio_info_cap_header
*
640 vfio_get_region_info_cap(struct vfio_region_info
*info
, uint16_t id
)
642 struct vfio_info_cap_header
*hdr
;
645 if (!(info
->flags
& VFIO_REGION_INFO_FLAG_CAPS
)) {
649 for (hdr
= ptr
+ info
->cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
658 static int vfio_setup_region_sparse_mmaps(VFIORegion
*region
,
659 struct vfio_region_info
*info
)
661 struct vfio_info_cap_header
*hdr
;
662 struct vfio_region_info_cap_sparse_mmap
*sparse
;
665 hdr
= vfio_get_region_info_cap(info
, VFIO_REGION_INFO_CAP_SPARSE_MMAP
);
670 sparse
= container_of(hdr
, struct vfio_region_info_cap_sparse_mmap
, header
);
672 trace_vfio_region_sparse_mmap_header(region
->vbasedev
->name
,
673 region
->nr
, sparse
->nr_areas
);
675 region
->mmaps
= g_new0(VFIOMmap
, sparse
->nr_areas
);
677 for (i
= 0, j
= 0; i
< sparse
->nr_areas
; i
++) {
678 trace_vfio_region_sparse_mmap_entry(i
, sparse
->areas
[i
].offset
,
679 sparse
->areas
[i
].offset
+
680 sparse
->areas
[i
].size
);
682 if (sparse
->areas
[i
].size
) {
683 region
->mmaps
[j
].offset
= sparse
->areas
[i
].offset
;
684 region
->mmaps
[j
].size
= sparse
->areas
[i
].size
;
689 region
->nr_mmaps
= j
;
690 region
->mmaps
= g_realloc(region
->mmaps
, j
* sizeof(VFIOMmap
));
695 int vfio_region_setup(Object
*obj
, VFIODevice
*vbasedev
, VFIORegion
*region
,
696 int index
, const char *name
)
698 struct vfio_region_info
*info
;
701 ret
= vfio_get_region_info(vbasedev
, index
, &info
);
706 region
->vbasedev
= vbasedev
;
707 region
->flags
= info
->flags
;
708 region
->size
= info
->size
;
709 region
->fd_offset
= info
->offset
;
713 region
->mem
= g_new0(MemoryRegion
, 1);
714 memory_region_init_io(region
->mem
, obj
, &vfio_region_ops
,
715 region
, name
, region
->size
);
717 if (!vbasedev
->no_mmap
&&
718 region
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
720 ret
= vfio_setup_region_sparse_mmaps(region
, info
);
723 region
->nr_mmaps
= 1;
724 region
->mmaps
= g_new0(VFIOMmap
, region
->nr_mmaps
);
725 region
->mmaps
[0].offset
= 0;
726 region
->mmaps
[0].size
= region
->size
;
733 trace_vfio_region_setup(vbasedev
->name
, index
, name
,
734 region
->flags
, region
->fd_offset
, region
->size
);
738 int vfio_region_mmap(VFIORegion
*region
)
747 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_READ
? PROT_READ
: 0;
748 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_WRITE
? PROT_WRITE
: 0;
750 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
751 region
->mmaps
[i
].mmap
= mmap(NULL
, region
->mmaps
[i
].size
, prot
,
752 MAP_SHARED
, region
->vbasedev
->fd
,
754 region
->mmaps
[i
].offset
);
755 if (region
->mmaps
[i
].mmap
== MAP_FAILED
) {
758 trace_vfio_region_mmap_fault(memory_region_name(region
->mem
), i
,
760 region
->mmaps
[i
].offset
,
762 region
->mmaps
[i
].offset
+
763 region
->mmaps
[i
].size
- 1, ret
);
765 region
->mmaps
[i
].mmap
= NULL
;
767 for (i
--; i
>= 0; i
--) {
768 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
769 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
770 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
771 region
->mmaps
[i
].mmap
= NULL
;
777 name
= g_strdup_printf("%s mmaps[%d]",
778 memory_region_name(region
->mem
), i
);
779 memory_region_init_ram_device_ptr(®ion
->mmaps
[i
].mem
,
780 memory_region_owner(region
->mem
),
781 name
, region
->mmaps
[i
].size
,
782 region
->mmaps
[i
].mmap
);
784 memory_region_add_subregion(region
->mem
, region
->mmaps
[i
].offset
,
785 ®ion
->mmaps
[i
].mem
);
787 trace_vfio_region_mmap(memory_region_name(®ion
->mmaps
[i
].mem
),
788 region
->mmaps
[i
].offset
,
789 region
->mmaps
[i
].offset
+
790 region
->mmaps
[i
].size
- 1);
796 void vfio_region_exit(VFIORegion
*region
)
804 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
805 if (region
->mmaps
[i
].mmap
) {
806 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
810 trace_vfio_region_exit(region
->vbasedev
->name
, region
->nr
);
813 void vfio_region_finalize(VFIORegion
*region
)
821 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
822 if (region
->mmaps
[i
].mmap
) {
823 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
824 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
828 object_unparent(OBJECT(region
->mem
));
831 g_free(region
->mmaps
);
833 trace_vfio_region_finalize(region
->vbasedev
->name
, region
->nr
);
836 void vfio_region_mmaps_set_enabled(VFIORegion
*region
, bool enabled
)
844 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
845 if (region
->mmaps
[i
].mmap
) {
846 memory_region_set_enabled(®ion
->mmaps
[i
].mem
, enabled
);
850 trace_vfio_region_mmaps_set_enabled(memory_region_name(region
->mem
),
854 void vfio_reset_handler(void *opaque
)
857 VFIODevice
*vbasedev
;
859 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
860 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
861 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
865 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
866 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
867 if (vbasedev
->needs_reset
) {
868 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
874 static void vfio_kvm_device_add_group(VFIOGroup
*group
)
877 struct kvm_device_attr attr
= {
878 .group
= KVM_DEV_VFIO_GROUP
,
879 .attr
= KVM_DEV_VFIO_GROUP_ADD
,
880 .addr
= (uint64_t)(unsigned long)&group
->fd
,
883 if (!kvm_enabled()) {
887 if (vfio_kvm_device_fd
< 0) {
888 struct kvm_create_device cd
= {
889 .type
= KVM_DEV_TYPE_VFIO
,
892 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
893 error_report("Failed to create KVM VFIO device: %m");
897 vfio_kvm_device_fd
= cd
.fd
;
900 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
901 error_report("Failed to add group %d to KVM VFIO device: %m",
907 static void vfio_kvm_device_del_group(VFIOGroup
*group
)
910 struct kvm_device_attr attr
= {
911 .group
= KVM_DEV_VFIO_GROUP
,
912 .attr
= KVM_DEV_VFIO_GROUP_DEL
,
913 .addr
= (uint64_t)(unsigned long)&group
->fd
,
916 if (vfio_kvm_device_fd
< 0) {
920 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
921 error_report("Failed to remove group %d from KVM VFIO device: %m",
927 static VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
929 VFIOAddressSpace
*space
;
931 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
932 if (space
->as
== as
) {
937 /* No suitable VFIOAddressSpace, create a new one */
938 space
= g_malloc0(sizeof(*space
));
940 QLIST_INIT(&space
->containers
);
942 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
947 static void vfio_put_address_space(VFIOAddressSpace
*space
)
949 if (QLIST_EMPTY(&space
->containers
)) {
950 QLIST_REMOVE(space
, list
);
955 static int vfio_connect_container(VFIOGroup
*group
, AddressSpace
*as
,
958 VFIOContainer
*container
;
960 VFIOAddressSpace
*space
;
962 space
= vfio_get_address_space(as
);
964 QLIST_FOREACH(container
, &space
->containers
, next
) {
965 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
966 group
->container
= container
;
967 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
972 fd
= qemu_open("/dev/vfio/vfio", O_RDWR
);
974 error_setg_errno(errp
, errno
, "failed to open /dev/vfio/vfio");
979 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
980 if (ret
!= VFIO_API_VERSION
) {
981 error_setg(errp
, "supported vfio version: %d, "
982 "reported version: %d", VFIO_API_VERSION
, ret
);
987 container
= g_malloc0(sizeof(*container
));
988 container
->space
= space
;
990 if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1_IOMMU
) ||
991 ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1v2_IOMMU
)) {
992 bool v2
= !!ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1v2_IOMMU
);
993 struct vfio_iommu_type1_info info
;
995 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
997 error_setg_errno(errp
, errno
, "failed to set group container");
999 goto free_container_exit
;
1002 container
->iommu_type
= v2
? VFIO_TYPE1v2_IOMMU
: VFIO_TYPE1_IOMMU
;
1003 ret
= ioctl(fd
, VFIO_SET_IOMMU
, container
->iommu_type
);
1005 error_setg_errno(errp
, errno
, "failed to set iommu for container");
1007 goto free_container_exit
;
1011 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1012 * IOVA whatsoever. That's not actually true, but the current
1013 * kernel interface doesn't tell us what it can map, and the
1014 * existing Type1 IOMMUs generally support any IOVA we're
1015 * going to actually try in practice.
1017 info
.argsz
= sizeof(info
);
1018 ret
= ioctl(fd
, VFIO_IOMMU_GET_INFO
, &info
);
1020 if (ret
|| !(info
.flags
& VFIO_IOMMU_INFO_PGSIZES
)) {
1021 /* Assume 4k IOVA page size */
1022 info
.iova_pgsizes
= 4096;
1024 vfio_host_win_add(container
, 0, (hwaddr
)-1, info
.iova_pgsizes
);
1025 } else if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_IOMMU
) ||
1026 ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_v2_IOMMU
)) {
1027 struct vfio_iommu_spapr_tce_info info
;
1028 bool v2
= !!ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_v2_IOMMU
);
1030 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
1032 error_setg_errno(errp
, errno
, "failed to set group container");
1034 goto free_container_exit
;
1036 container
->iommu_type
=
1037 v2
? VFIO_SPAPR_TCE_v2_IOMMU
: VFIO_SPAPR_TCE_IOMMU
;
1038 ret
= ioctl(fd
, VFIO_SET_IOMMU
, container
->iommu_type
);
1040 error_setg_errno(errp
, errno
, "failed to set iommu for container");
1042 goto free_container_exit
;
1046 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1047 * when container fd is closed so we do not call it explicitly
1051 ret
= ioctl(fd
, VFIO_IOMMU_ENABLE
);
1053 error_setg_errno(errp
, errno
, "failed to enable container");
1055 goto free_container_exit
;
1058 container
->prereg_listener
= vfio_prereg_listener
;
1060 memory_listener_register(&container
->prereg_listener
,
1061 &address_space_memory
);
1062 if (container
->error
) {
1063 memory_listener_unregister(&container
->prereg_listener
);
1064 ret
= container
->error
;
1066 "RAM memory listener initialization failed for container");
1067 goto free_container_exit
;
1071 info
.argsz
= sizeof(info
);
1072 ret
= ioctl(fd
, VFIO_IOMMU_SPAPR_TCE_GET_INFO
, &info
);
1074 error_setg_errno(errp
, errno
,
1075 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1078 memory_listener_unregister(&container
->prereg_listener
);
1080 goto free_container_exit
;
1085 * There is a default window in just created container.
1086 * To make region_add/del simpler, we better remove this
1087 * window now and let those iommu_listener callbacks
1088 * create/remove them when needed.
1090 ret
= vfio_spapr_remove_window(container
, info
.dma32_window_start
);
1092 error_setg_errno(errp
, -ret
,
1093 "failed to remove existing window");
1094 goto free_container_exit
;
1097 /* The default table uses 4K pages */
1098 vfio_host_win_add(container
, info
.dma32_window_start
,
1099 info
.dma32_window_start
+
1100 info
.dma32_window_size
- 1,
1104 error_setg(errp
, "No available IOMMU models");
1106 goto free_container_exit
;
1109 container
->listener
= vfio_memory_listener
;
1111 memory_listener_register(&container
->listener
, container
->space
->as
);
1113 if (container
->error
) {
1114 ret
= container
->error
;
1115 error_setg_errno(errp
, -ret
,
1116 "memory listener initialization failed for container");
1117 goto listener_release_exit
;
1120 container
->initialized
= true;
1122 QLIST_INIT(&container
->group_list
);
1123 QLIST_INSERT_HEAD(&space
->containers
, container
, next
);
1125 group
->container
= container
;
1126 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1129 listener_release_exit
:
1130 vfio_listener_release(container
);
1132 free_container_exit
:
1139 vfio_put_address_space(space
);
1144 static void vfio_disconnect_container(VFIOGroup
*group
)
1146 VFIOContainer
*container
= group
->container
;
1148 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1149 error_report("vfio: error disconnecting group %d from container",
1153 QLIST_REMOVE(group
, container_next
);
1154 group
->container
= NULL
;
1156 if (QLIST_EMPTY(&container
->group_list
)) {
1157 VFIOAddressSpace
*space
= container
->space
;
1158 VFIOGuestIOMMU
*giommu
, *tmp
;
1160 vfio_listener_release(container
);
1161 QLIST_REMOVE(container
, next
);
1163 QLIST_FOREACH_SAFE(giommu
, &container
->giommu_list
, giommu_next
, tmp
) {
1164 memory_region_unregister_iommu_notifier(giommu
->iommu
, &giommu
->n
);
1165 QLIST_REMOVE(giommu
, giommu_next
);
1169 trace_vfio_disconnect_container(container
->fd
);
1170 close(container
->fd
);
1173 vfio_put_address_space(space
);
1177 VFIOGroup
*vfio_get_group(int groupid
, AddressSpace
*as
, Error
**errp
)
1181 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1183 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1184 if (group
->groupid
== groupid
) {
1185 /* Found it. Now is it already in the right context? */
1186 if (group
->container
->space
->as
== as
) {
1189 error_setg(errp
, "group %d used in multiple address spaces",
1196 group
= g_malloc0(sizeof(*group
));
1198 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1199 group
->fd
= qemu_open(path
, O_RDWR
);
1200 if (group
->fd
< 0) {
1201 error_setg_errno(errp
, errno
, "failed to open %s", path
);
1202 goto free_group_exit
;
1205 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1206 error_setg_errno(errp
, errno
, "failed to get group %d status", groupid
);
1210 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1211 error_setg(errp
, "group %d is not viable", groupid
);
1212 error_append_hint(errp
,
1213 "Please ensure all devices within the iommu_group "
1214 "are bound to their vfio bus driver.\n");
1218 group
->groupid
= groupid
;
1219 QLIST_INIT(&group
->device_list
);
1221 if (vfio_connect_container(group
, as
, errp
)) {
1222 error_prepend(errp
, "failed to setup container for group %d: ",
1227 if (QLIST_EMPTY(&vfio_group_list
)) {
1228 qemu_register_reset(vfio_reset_handler
, NULL
);
1231 QLIST_INSERT_HEAD(&vfio_group_list
, group
, next
);
1233 vfio_kvm_device_add_group(group
);
1246 void vfio_put_group(VFIOGroup
*group
)
1248 if (!group
|| !QLIST_EMPTY(&group
->device_list
)) {
1252 vfio_kvm_device_del_group(group
);
1253 vfio_disconnect_container(group
);
1254 QLIST_REMOVE(group
, next
);
1255 trace_vfio_put_group(group
->fd
);
1259 if (QLIST_EMPTY(&vfio_group_list
)) {
1260 qemu_unregister_reset(vfio_reset_handler
, NULL
);
1264 int vfio_get_device(VFIOGroup
*group
, const char *name
,
1265 VFIODevice
*vbasedev
, Error
**errp
)
1267 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
1270 fd
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
1272 error_setg_errno(errp
, errno
, "error getting device from group %d",
1274 error_append_hint(errp
,
1275 "Verify all devices in group %d are bound to vfio-<bus> "
1276 "or pci-stub and not already in use\n", group
->groupid
);
1280 ret
= ioctl(fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
1282 error_setg_errno(errp
, errno
, "error getting device info");
1288 vbasedev
->group
= group
;
1289 QLIST_INSERT_HEAD(&group
->device_list
, vbasedev
, next
);
1291 vbasedev
->num_irqs
= dev_info
.num_irqs
;
1292 vbasedev
->num_regions
= dev_info
.num_regions
;
1293 vbasedev
->flags
= dev_info
.flags
;
1295 trace_vfio_get_device(name
, dev_info
.flags
, dev_info
.num_regions
,
1298 vbasedev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
1302 void vfio_put_base_device(VFIODevice
*vbasedev
)
1304 if (!vbasedev
->group
) {
1307 QLIST_REMOVE(vbasedev
, next
);
1308 vbasedev
->group
= NULL
;
1309 trace_vfio_put_base_device(vbasedev
->fd
);
1310 close(vbasedev
->fd
);
1313 int vfio_get_region_info(VFIODevice
*vbasedev
, int index
,
1314 struct vfio_region_info
**info
)
1316 size_t argsz
= sizeof(struct vfio_region_info
);
1318 *info
= g_malloc0(argsz
);
1320 (*info
)->index
= index
;
1322 (*info
)->argsz
= argsz
;
1324 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, *info
)) {
1330 if ((*info
)->argsz
> argsz
) {
1331 argsz
= (*info
)->argsz
;
1332 *info
= g_realloc(*info
, argsz
);
1340 int vfio_get_dev_region_info(VFIODevice
*vbasedev
, uint32_t type
,
1341 uint32_t subtype
, struct vfio_region_info
**info
)
1345 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
1346 struct vfio_info_cap_header
*hdr
;
1347 struct vfio_region_info_cap_type
*cap_type
;
1349 if (vfio_get_region_info(vbasedev
, i
, info
)) {
1353 hdr
= vfio_get_region_info_cap(*info
, VFIO_REGION_INFO_CAP_TYPE
);
1359 cap_type
= container_of(hdr
, struct vfio_region_info_cap_type
, header
);
1361 trace_vfio_get_dev_region(vbasedev
->name
, i
,
1362 cap_type
->type
, cap_type
->subtype
);
1364 if (cap_type
->type
== type
&& cap_type
->subtype
== subtype
) {
1376 * Interfaces for IBM EEH (Enhanced Error Handling)
1378 static bool vfio_eeh_container_ok(VFIOContainer
*container
)
1381 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1382 * implementation is broken if there are multiple groups in a
1383 * container. The hardware works in units of Partitionable
1384 * Endpoints (== IOMMU groups) and the EEH operations naively
1385 * iterate across all groups in the container, without any logic
1386 * to make sure the groups have their state synchronized. For
1387 * certain operations (ENABLE) that might be ok, until an error
1388 * occurs, but for others (GET_STATE) it's clearly broken.
1392 * XXX Once fixed kernels exist, test for them here
1395 if (QLIST_EMPTY(&container
->group_list
)) {
1399 if (QLIST_NEXT(QLIST_FIRST(&container
->group_list
), container_next
)) {
1406 static int vfio_eeh_container_op(VFIOContainer
*container
, uint32_t op
)
1408 struct vfio_eeh_pe_op pe_op
= {
1409 .argsz
= sizeof(pe_op
),
1414 if (!vfio_eeh_container_ok(container
)) {
1415 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1416 "kernel requires a container with exactly one group", op
);
1420 ret
= ioctl(container
->fd
, VFIO_EEH_PE_OP
, &pe_op
);
1422 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op
);
1429 static VFIOContainer
*vfio_eeh_as_container(AddressSpace
*as
)
1431 VFIOAddressSpace
*space
= vfio_get_address_space(as
);
1432 VFIOContainer
*container
= NULL
;
1434 if (QLIST_EMPTY(&space
->containers
)) {
1435 /* No containers to act on */
1439 container
= QLIST_FIRST(&space
->containers
);
1441 if (QLIST_NEXT(container
, next
)) {
1442 /* We don't yet have logic to synchronize EEH state across
1443 * multiple containers */
1449 vfio_put_address_space(space
);
1453 bool vfio_eeh_as_ok(AddressSpace
*as
)
1455 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1457 return (container
!= NULL
) && vfio_eeh_container_ok(container
);
1460 int vfio_eeh_as_op(AddressSpace
*as
, uint32_t op
)
1462 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1467 return vfio_eeh_container_op(container
, op
);