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
) {
440 /* For now intersections are not allowed, we may relax this later */
441 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
442 if (ranges_overlap(hostwin
->min_iova
,
443 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
444 section
->offset_within_address_space
,
445 int128_get64(section
->size
))) {
451 ret
= vfio_spapr_create_window(container
, section
, &pgsize
);
456 vfio_host_win_add(container
, section
->offset_within_address_space
,
457 section
->offset_within_address_space
+
458 int128_get64(section
->size
) - 1, pgsize
);
462 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
463 struct kvm_vfio_spapr_tce param
;
464 struct kvm_device_attr attr
= {
465 .group
= KVM_DEV_VFIO_GROUP
,
466 .attr
= KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE
,
467 .addr
= (uint64_t)(unsigned long)¶m
,
470 if (!memory_region_iommu_get_attr(iommu_mr
, IOMMU_ATTR_SPAPR_TCE_FD
,
472 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
473 param
.groupfd
= group
->fd
;
474 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
475 error_report("vfio: failed to setup fd %d "
476 "for a group with fd %d: %s",
477 param
.tablefd
, param
.groupfd
,
481 trace_vfio_spapr_group_attach(param
.groupfd
, param
.tablefd
);
488 hostwin_found
= false;
489 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
490 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
491 hostwin_found
= true;
496 if (!hostwin_found
) {
497 error_report("vfio: IOMMU container %p can't map guest IOVA region"
498 " 0x%"HWADDR_PRIx
"..0x%"HWADDR_PRIx
,
499 container
, iova
, end
);
504 memory_region_ref(section
->mr
);
506 if (memory_region_is_iommu(section
->mr
)) {
507 VFIOGuestIOMMU
*giommu
;
508 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
510 trace_vfio_listener_region_add_iommu(iova
, end
);
512 * FIXME: For VFIO iommu types which have KVM acceleration to
513 * avoid bouncing all map/unmaps through qemu this way, this
514 * would be the right place to wire that up (tell the KVM
515 * device emulation the VFIO iommu handles to use).
517 giommu
= g_malloc0(sizeof(*giommu
));
518 giommu
->iommu
= iommu_mr
;
519 giommu
->iommu_offset
= section
->offset_within_address_space
-
520 section
->offset_within_region
;
521 giommu
->container
= container
;
522 llend
= int128_add(int128_make64(section
->offset_within_region
),
524 llend
= int128_sub(llend
, int128_one());
525 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
527 section
->offset_within_region
,
528 int128_get64(llend
));
529 QLIST_INSERT_HEAD(&container
->giommu_list
, giommu
, giommu_next
);
531 memory_region_register_iommu_notifier(section
->mr
, &giommu
->n
);
532 memory_region_iommu_replay(giommu
->iommu
, &giommu
->n
);
537 /* Here we assume that memory_region_is_ram(section->mr)==true */
539 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
540 section
->offset_within_region
+
541 (iova
- section
->offset_within_address_space
);
543 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
545 llsize
= int128_sub(llend
, int128_make64(iova
));
547 if (memory_region_is_ram_device(section
->mr
)) {
548 hwaddr pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
550 if ((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
)) {
551 trace_vfio_listener_region_add_no_dma_map(
552 memory_region_name(section
->mr
),
553 section
->offset_within_address_space
,
554 int128_getlo(section
->size
),
560 ret
= vfio_dma_map(container
, iova
, int128_get64(llsize
),
561 vaddr
, section
->readonly
);
563 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
564 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
565 container
, iova
, int128_get64(llsize
), vaddr
, ret
);
566 if (memory_region_is_ram_device(section
->mr
)) {
567 /* Allow unexpected mappings not to be fatal for RAM devices */
576 if (memory_region_is_ram_device(section
->mr
)) {
577 error_report("failed to vfio_dma_map. pci p2p may not work");
581 * On the initfn path, store the first error in the container so we
582 * can gracefully fail. Runtime, there's not much we can do other
583 * than throw a hardware error.
585 if (!container
->initialized
) {
586 if (!container
->error
) {
587 container
->error
= ret
;
590 hw_error("vfio: DMA mapping failed, unable to continue");
594 static void vfio_listener_region_del(MemoryListener
*listener
,
595 MemoryRegionSection
*section
)
597 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
599 Int128 llend
, llsize
;
601 bool try_unmap
= true;
603 if (vfio_listener_skipped_section(section
)) {
604 trace_vfio_listener_region_del_skip(
605 section
->offset_within_address_space
,
606 section
->offset_within_address_space
+
607 int128_get64(int128_sub(section
->size
, int128_one())));
611 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
612 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
613 error_report("%s received unaligned region", __func__
);
617 if (memory_region_is_iommu(section
->mr
)) {
618 VFIOGuestIOMMU
*giommu
;
620 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
621 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
622 giommu
->n
.start
== section
->offset_within_region
) {
623 memory_region_unregister_iommu_notifier(section
->mr
,
625 QLIST_REMOVE(giommu
, giommu_next
);
632 * FIXME: We assume the one big unmap below is adequate to
633 * remove any individual page mappings in the IOMMU which
634 * might have been copied into VFIO. This works for a page table
635 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
636 * That may not be true for all IOMMU types.
640 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
641 llend
= int128_make64(section
->offset_within_address_space
);
642 llend
= int128_add(llend
, section
->size
);
643 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
645 if (int128_ge(int128_make64(iova
), llend
)) {
648 end
= int128_get64(int128_sub(llend
, int128_one()));
650 llsize
= int128_sub(llend
, int128_make64(iova
));
652 trace_vfio_listener_region_del(iova
, end
);
654 if (memory_region_is_ram_device(section
->mr
)) {
656 VFIOHostDMAWindow
*hostwin
;
657 bool hostwin_found
= false;
659 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
660 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
661 hostwin_found
= true;
665 assert(hostwin_found
); /* or region_add() would have failed */
667 pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
668 try_unmap
= !((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
));
672 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
));
674 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
675 "0x%"HWADDR_PRIx
") = %d (%m)",
676 container
, iova
, int128_get64(llsize
), ret
);
680 memory_region_unref(section
->mr
);
682 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
683 vfio_spapr_remove_window(container
,
684 section
->offset_within_address_space
);
685 if (vfio_host_win_del(container
,
686 section
->offset_within_address_space
,
687 section
->offset_within_address_space
+
688 int128_get64(section
->size
) - 1) < 0) {
689 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx
,
690 __func__
, section
->offset_within_address_space
);
695 static const MemoryListener vfio_memory_listener
= {
696 .region_add
= vfio_listener_region_add
,
697 .region_del
= vfio_listener_region_del
,
700 static void vfio_listener_release(VFIOContainer
*container
)
702 memory_listener_unregister(&container
->listener
);
703 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
704 memory_listener_unregister(&container
->prereg_listener
);
708 static struct vfio_info_cap_header
*
709 vfio_get_region_info_cap(struct vfio_region_info
*info
, uint16_t id
)
711 struct vfio_info_cap_header
*hdr
;
714 if (!(info
->flags
& VFIO_REGION_INFO_FLAG_CAPS
)) {
718 for (hdr
= ptr
+ info
->cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
727 static int vfio_setup_region_sparse_mmaps(VFIORegion
*region
,
728 struct vfio_region_info
*info
)
730 struct vfio_info_cap_header
*hdr
;
731 struct vfio_region_info_cap_sparse_mmap
*sparse
;
734 hdr
= vfio_get_region_info_cap(info
, VFIO_REGION_INFO_CAP_SPARSE_MMAP
);
739 sparse
= container_of(hdr
, struct vfio_region_info_cap_sparse_mmap
, header
);
741 trace_vfio_region_sparse_mmap_header(region
->vbasedev
->name
,
742 region
->nr
, sparse
->nr_areas
);
744 region
->mmaps
= g_new0(VFIOMmap
, sparse
->nr_areas
);
746 for (i
= 0, j
= 0; i
< sparse
->nr_areas
; i
++) {
747 trace_vfio_region_sparse_mmap_entry(i
, sparse
->areas
[i
].offset
,
748 sparse
->areas
[i
].offset
+
749 sparse
->areas
[i
].size
);
751 if (sparse
->areas
[i
].size
) {
752 region
->mmaps
[j
].offset
= sparse
->areas
[i
].offset
;
753 region
->mmaps
[j
].size
= sparse
->areas
[i
].size
;
758 region
->nr_mmaps
= j
;
759 region
->mmaps
= g_realloc(region
->mmaps
, j
* sizeof(VFIOMmap
));
764 int vfio_region_setup(Object
*obj
, VFIODevice
*vbasedev
, VFIORegion
*region
,
765 int index
, const char *name
)
767 struct vfio_region_info
*info
;
770 ret
= vfio_get_region_info(vbasedev
, index
, &info
);
775 region
->vbasedev
= vbasedev
;
776 region
->flags
= info
->flags
;
777 region
->size
= info
->size
;
778 region
->fd_offset
= info
->offset
;
782 region
->mem
= g_new0(MemoryRegion
, 1);
783 memory_region_init_io(region
->mem
, obj
, &vfio_region_ops
,
784 region
, name
, region
->size
);
786 if (!vbasedev
->no_mmap
&&
787 region
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
789 ret
= vfio_setup_region_sparse_mmaps(region
, info
);
792 region
->nr_mmaps
= 1;
793 region
->mmaps
= g_new0(VFIOMmap
, region
->nr_mmaps
);
794 region
->mmaps
[0].offset
= 0;
795 region
->mmaps
[0].size
= region
->size
;
802 trace_vfio_region_setup(vbasedev
->name
, index
, name
,
803 region
->flags
, region
->fd_offset
, region
->size
);
807 int vfio_region_mmap(VFIORegion
*region
)
816 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_READ
? PROT_READ
: 0;
817 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_WRITE
? PROT_WRITE
: 0;
819 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
820 region
->mmaps
[i
].mmap
= mmap(NULL
, region
->mmaps
[i
].size
, prot
,
821 MAP_SHARED
, region
->vbasedev
->fd
,
823 region
->mmaps
[i
].offset
);
824 if (region
->mmaps
[i
].mmap
== MAP_FAILED
) {
827 trace_vfio_region_mmap_fault(memory_region_name(region
->mem
), i
,
829 region
->mmaps
[i
].offset
,
831 region
->mmaps
[i
].offset
+
832 region
->mmaps
[i
].size
- 1, ret
);
834 region
->mmaps
[i
].mmap
= NULL
;
836 for (i
--; i
>= 0; i
--) {
837 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
838 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
839 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
840 region
->mmaps
[i
].mmap
= NULL
;
846 name
= g_strdup_printf("%s mmaps[%d]",
847 memory_region_name(region
->mem
), i
);
848 memory_region_init_ram_device_ptr(®ion
->mmaps
[i
].mem
,
849 memory_region_owner(region
->mem
),
850 name
, region
->mmaps
[i
].size
,
851 region
->mmaps
[i
].mmap
);
853 memory_region_add_subregion(region
->mem
, region
->mmaps
[i
].offset
,
854 ®ion
->mmaps
[i
].mem
);
856 trace_vfio_region_mmap(memory_region_name(®ion
->mmaps
[i
].mem
),
857 region
->mmaps
[i
].offset
,
858 region
->mmaps
[i
].offset
+
859 region
->mmaps
[i
].size
- 1);
865 void vfio_region_exit(VFIORegion
*region
)
873 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
874 if (region
->mmaps
[i
].mmap
) {
875 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
879 trace_vfio_region_exit(region
->vbasedev
->name
, region
->nr
);
882 void vfio_region_finalize(VFIORegion
*region
)
890 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
891 if (region
->mmaps
[i
].mmap
) {
892 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
893 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
897 object_unparent(OBJECT(region
->mem
));
900 g_free(region
->mmaps
);
902 trace_vfio_region_finalize(region
->vbasedev
->name
, region
->nr
);
905 region
->mmaps
= NULL
;
906 region
->nr_mmaps
= 0;
912 void vfio_region_mmaps_set_enabled(VFIORegion
*region
, bool enabled
)
920 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
921 if (region
->mmaps
[i
].mmap
) {
922 memory_region_set_enabled(®ion
->mmaps
[i
].mem
, enabled
);
926 trace_vfio_region_mmaps_set_enabled(memory_region_name(region
->mem
),
930 void vfio_reset_handler(void *opaque
)
933 VFIODevice
*vbasedev
;
935 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
936 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
937 if (vbasedev
->dev
->realized
) {
938 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
943 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
944 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
945 if (vbasedev
->dev
->realized
&& vbasedev
->needs_reset
) {
946 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
952 static void vfio_kvm_device_add_group(VFIOGroup
*group
)
955 struct kvm_device_attr attr
= {
956 .group
= KVM_DEV_VFIO_GROUP
,
957 .attr
= KVM_DEV_VFIO_GROUP_ADD
,
958 .addr
= (uint64_t)(unsigned long)&group
->fd
,
961 if (!kvm_enabled()) {
965 if (vfio_kvm_device_fd
< 0) {
966 struct kvm_create_device cd
= {
967 .type
= KVM_DEV_TYPE_VFIO
,
970 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
971 error_report("Failed to create KVM VFIO device: %m");
975 vfio_kvm_device_fd
= cd
.fd
;
978 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
979 error_report("Failed to add group %d to KVM VFIO device: %m",
985 static void vfio_kvm_device_del_group(VFIOGroup
*group
)
988 struct kvm_device_attr attr
= {
989 .group
= KVM_DEV_VFIO_GROUP
,
990 .attr
= KVM_DEV_VFIO_GROUP_DEL
,
991 .addr
= (uint64_t)(unsigned long)&group
->fd
,
994 if (vfio_kvm_device_fd
< 0) {
998 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
999 error_report("Failed to remove group %d from KVM VFIO device: %m",
1005 static VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
1007 VFIOAddressSpace
*space
;
1009 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
1010 if (space
->as
== as
) {
1015 /* No suitable VFIOAddressSpace, create a new one */
1016 space
= g_malloc0(sizeof(*space
));
1018 QLIST_INIT(&space
->containers
);
1020 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
1025 static void vfio_put_address_space(VFIOAddressSpace
*space
)
1027 if (QLIST_EMPTY(&space
->containers
)) {
1028 QLIST_REMOVE(space
, list
);
1033 static int vfio_connect_container(VFIOGroup
*group
, AddressSpace
*as
,
1036 VFIOContainer
*container
;
1038 VFIOAddressSpace
*space
;
1040 space
= vfio_get_address_space(as
);
1042 QLIST_FOREACH(container
, &space
->containers
, next
) {
1043 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
1044 group
->container
= container
;
1045 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1046 vfio_kvm_device_add_group(group
);
1051 fd
= qemu_open("/dev/vfio/vfio", O_RDWR
);
1053 error_setg_errno(errp
, errno
, "failed to open /dev/vfio/vfio");
1055 goto put_space_exit
;
1058 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
1059 if (ret
!= VFIO_API_VERSION
) {
1060 error_setg(errp
, "supported vfio version: %d, "
1061 "reported version: %d", VFIO_API_VERSION
, ret
);
1066 container
= g_malloc0(sizeof(*container
));
1067 container
->space
= space
;
1069 QLIST_INIT(&container
->giommu_list
);
1070 QLIST_INIT(&container
->hostwin_list
);
1071 if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1_IOMMU
) ||
1072 ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1v2_IOMMU
)) {
1073 bool v2
= !!ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1v2_IOMMU
);
1074 struct vfio_iommu_type1_info info
;
1076 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
1078 error_setg_errno(errp
, errno
, "failed to set group container");
1080 goto free_container_exit
;
1083 container
->iommu_type
= v2
? VFIO_TYPE1v2_IOMMU
: VFIO_TYPE1_IOMMU
;
1084 ret
= ioctl(fd
, VFIO_SET_IOMMU
, container
->iommu_type
);
1086 error_setg_errno(errp
, errno
, "failed to set iommu for container");
1088 goto free_container_exit
;
1092 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1093 * IOVA whatsoever. That's not actually true, but the current
1094 * kernel interface doesn't tell us what it can map, and the
1095 * existing Type1 IOMMUs generally support any IOVA we're
1096 * going to actually try in practice.
1098 info
.argsz
= sizeof(info
);
1099 ret
= ioctl(fd
, VFIO_IOMMU_GET_INFO
, &info
);
1101 if (ret
|| !(info
.flags
& VFIO_IOMMU_INFO_PGSIZES
)) {
1102 /* Assume 4k IOVA page size */
1103 info
.iova_pgsizes
= 4096;
1105 vfio_host_win_add(container
, 0, (hwaddr
)-1, info
.iova_pgsizes
);
1106 } else if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_IOMMU
) ||
1107 ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_v2_IOMMU
)) {
1108 struct vfio_iommu_spapr_tce_info info
;
1109 bool v2
= !!ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_SPAPR_TCE_v2_IOMMU
);
1111 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
1113 error_setg_errno(errp
, errno
, "failed to set group container");
1115 goto free_container_exit
;
1117 container
->iommu_type
=
1118 v2
? VFIO_SPAPR_TCE_v2_IOMMU
: VFIO_SPAPR_TCE_IOMMU
;
1119 ret
= ioctl(fd
, VFIO_SET_IOMMU
, container
->iommu_type
);
1121 container
->iommu_type
= VFIO_SPAPR_TCE_IOMMU
;
1123 ret
= ioctl(fd
, VFIO_SET_IOMMU
, container
->iommu_type
);
1126 error_setg_errno(errp
, errno
, "failed to set iommu for container");
1128 goto free_container_exit
;
1132 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1133 * when container fd is closed so we do not call it explicitly
1137 ret
= ioctl(fd
, VFIO_IOMMU_ENABLE
);
1139 error_setg_errno(errp
, errno
, "failed to enable container");
1141 goto free_container_exit
;
1144 container
->prereg_listener
= vfio_prereg_listener
;
1146 memory_listener_register(&container
->prereg_listener
,
1147 &address_space_memory
);
1148 if (container
->error
) {
1149 memory_listener_unregister(&container
->prereg_listener
);
1150 ret
= container
->error
;
1152 "RAM memory listener initialization failed for container");
1153 goto free_container_exit
;
1157 info
.argsz
= sizeof(info
);
1158 ret
= ioctl(fd
, VFIO_IOMMU_SPAPR_TCE_GET_INFO
, &info
);
1160 error_setg_errno(errp
, errno
,
1161 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1164 memory_listener_unregister(&container
->prereg_listener
);
1166 goto free_container_exit
;
1171 * There is a default window in just created container.
1172 * To make region_add/del simpler, we better remove this
1173 * window now and let those iommu_listener callbacks
1174 * create/remove them when needed.
1176 ret
= vfio_spapr_remove_window(container
, info
.dma32_window_start
);
1178 error_setg_errno(errp
, -ret
,
1179 "failed to remove existing window");
1180 goto free_container_exit
;
1183 /* The default table uses 4K pages */
1184 vfio_host_win_add(container
, info
.dma32_window_start
,
1185 info
.dma32_window_start
+
1186 info
.dma32_window_size
- 1,
1190 error_setg(errp
, "No available IOMMU models");
1192 goto free_container_exit
;
1195 vfio_kvm_device_add_group(group
);
1197 QLIST_INIT(&container
->group_list
);
1198 QLIST_INSERT_HEAD(&space
->containers
, container
, next
);
1200 group
->container
= container
;
1201 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1203 container
->listener
= vfio_memory_listener
;
1205 memory_listener_register(&container
->listener
, container
->space
->as
);
1207 if (container
->error
) {
1208 ret
= container
->error
;
1209 error_setg_errno(errp
, -ret
,
1210 "memory listener initialization failed for container");
1211 goto listener_release_exit
;
1214 container
->initialized
= true;
1217 listener_release_exit
:
1218 QLIST_REMOVE(group
, container_next
);
1219 QLIST_REMOVE(container
, next
);
1220 vfio_kvm_device_del_group(group
);
1221 vfio_listener_release(container
);
1223 free_container_exit
:
1230 vfio_put_address_space(space
);
1235 static void vfio_disconnect_container(VFIOGroup
*group
)
1237 VFIOContainer
*container
= group
->container
;
1239 QLIST_REMOVE(group
, container_next
);
1240 group
->container
= NULL
;
1243 * Explicitly release the listener first before unset container,
1244 * since unset may destroy the backend container if it's the last
1247 if (QLIST_EMPTY(&container
->group_list
)) {
1248 vfio_listener_release(container
);
1251 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1252 error_report("vfio: error disconnecting group %d from container",
1256 if (QLIST_EMPTY(&container
->group_list
)) {
1257 VFIOAddressSpace
*space
= container
->space
;
1258 VFIOGuestIOMMU
*giommu
, *tmp
;
1260 QLIST_REMOVE(container
, next
);
1262 QLIST_FOREACH_SAFE(giommu
, &container
->giommu_list
, giommu_next
, tmp
) {
1263 memory_region_unregister_iommu_notifier(
1264 MEMORY_REGION(giommu
->iommu
), &giommu
->n
);
1265 QLIST_REMOVE(giommu
, giommu_next
);
1269 trace_vfio_disconnect_container(container
->fd
);
1270 close(container
->fd
);
1273 vfio_put_address_space(space
);
1277 VFIOGroup
*vfio_get_group(int groupid
, AddressSpace
*as
, Error
**errp
)
1281 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1283 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1284 if (group
->groupid
== groupid
) {
1285 /* Found it. Now is it already in the right context? */
1286 if (group
->container
->space
->as
== as
) {
1289 error_setg(errp
, "group %d used in multiple address spaces",
1296 group
= g_malloc0(sizeof(*group
));
1298 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1299 group
->fd
= qemu_open(path
, O_RDWR
);
1300 if (group
->fd
< 0) {
1301 error_setg_errno(errp
, errno
, "failed to open %s", path
);
1302 goto free_group_exit
;
1305 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1306 error_setg_errno(errp
, errno
, "failed to get group %d status", groupid
);
1310 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1311 error_setg(errp
, "group %d is not viable", groupid
);
1312 error_append_hint(errp
,
1313 "Please ensure all devices within the iommu_group "
1314 "are bound to their vfio bus driver.\n");
1318 group
->groupid
= groupid
;
1319 QLIST_INIT(&group
->device_list
);
1321 if (vfio_connect_container(group
, as
, errp
)) {
1322 error_prepend(errp
, "failed to setup container for group %d: ",
1327 if (QLIST_EMPTY(&vfio_group_list
)) {
1328 qemu_register_reset(vfio_reset_handler
, NULL
);
1331 QLIST_INSERT_HEAD(&vfio_group_list
, group
, next
);
1344 void vfio_put_group(VFIOGroup
*group
)
1346 if (!group
|| !QLIST_EMPTY(&group
->device_list
)) {
1350 vfio_kvm_device_del_group(group
);
1351 vfio_disconnect_container(group
);
1352 QLIST_REMOVE(group
, next
);
1353 trace_vfio_put_group(group
->fd
);
1357 if (QLIST_EMPTY(&vfio_group_list
)) {
1358 qemu_unregister_reset(vfio_reset_handler
, NULL
);
1362 int vfio_get_device(VFIOGroup
*group
, const char *name
,
1363 VFIODevice
*vbasedev
, Error
**errp
)
1365 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
1368 fd
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
1370 error_setg_errno(errp
, errno
, "error getting device from group %d",
1372 error_append_hint(errp
,
1373 "Verify all devices in group %d are bound to vfio-<bus> "
1374 "or pci-stub and not already in use\n", group
->groupid
);
1378 ret
= ioctl(fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
1380 error_setg_errno(errp
, errno
, "error getting device info");
1386 vbasedev
->group
= group
;
1387 QLIST_INSERT_HEAD(&group
->device_list
, vbasedev
, next
);
1389 vbasedev
->num_irqs
= dev_info
.num_irqs
;
1390 vbasedev
->num_regions
= dev_info
.num_regions
;
1391 vbasedev
->flags
= dev_info
.flags
;
1393 trace_vfio_get_device(name
, dev_info
.flags
, dev_info
.num_regions
,
1396 vbasedev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
1400 void vfio_put_base_device(VFIODevice
*vbasedev
)
1402 if (!vbasedev
->group
) {
1405 QLIST_REMOVE(vbasedev
, next
);
1406 vbasedev
->group
= NULL
;
1407 trace_vfio_put_base_device(vbasedev
->fd
);
1408 close(vbasedev
->fd
);
1411 int vfio_get_region_info(VFIODevice
*vbasedev
, int index
,
1412 struct vfio_region_info
**info
)
1414 size_t argsz
= sizeof(struct vfio_region_info
);
1416 *info
= g_malloc0(argsz
);
1418 (*info
)->index
= index
;
1420 (*info
)->argsz
= argsz
;
1422 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, *info
)) {
1428 if ((*info
)->argsz
> argsz
) {
1429 argsz
= (*info
)->argsz
;
1430 *info
= g_realloc(*info
, argsz
);
1438 int vfio_get_dev_region_info(VFIODevice
*vbasedev
, uint32_t type
,
1439 uint32_t subtype
, struct vfio_region_info
**info
)
1443 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
1444 struct vfio_info_cap_header
*hdr
;
1445 struct vfio_region_info_cap_type
*cap_type
;
1447 if (vfio_get_region_info(vbasedev
, i
, info
)) {
1451 hdr
= vfio_get_region_info_cap(*info
, VFIO_REGION_INFO_CAP_TYPE
);
1457 cap_type
= container_of(hdr
, struct vfio_region_info_cap_type
, header
);
1459 trace_vfio_get_dev_region(vbasedev
->name
, i
,
1460 cap_type
->type
, cap_type
->subtype
);
1462 if (cap_type
->type
== type
&& cap_type
->subtype
== subtype
) {
1473 bool vfio_has_region_cap(VFIODevice
*vbasedev
, int region
, uint16_t cap_type
)
1475 struct vfio_region_info
*info
= NULL
;
1478 if (!vfio_get_region_info(vbasedev
, region
, &info
)) {
1479 if (vfio_get_region_info_cap(info
, cap_type
)) {
1489 * Interfaces for IBM EEH (Enhanced Error Handling)
1491 static bool vfio_eeh_container_ok(VFIOContainer
*container
)
1494 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1495 * implementation is broken if there are multiple groups in a
1496 * container. The hardware works in units of Partitionable
1497 * Endpoints (== IOMMU groups) and the EEH operations naively
1498 * iterate across all groups in the container, without any logic
1499 * to make sure the groups have their state synchronized. For
1500 * certain operations (ENABLE) that might be ok, until an error
1501 * occurs, but for others (GET_STATE) it's clearly broken.
1505 * XXX Once fixed kernels exist, test for them here
1508 if (QLIST_EMPTY(&container
->group_list
)) {
1512 if (QLIST_NEXT(QLIST_FIRST(&container
->group_list
), container_next
)) {
1519 static int vfio_eeh_container_op(VFIOContainer
*container
, uint32_t op
)
1521 struct vfio_eeh_pe_op pe_op
= {
1522 .argsz
= sizeof(pe_op
),
1527 if (!vfio_eeh_container_ok(container
)) {
1528 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1529 "kernel requires a container with exactly one group", op
);
1533 ret
= ioctl(container
->fd
, VFIO_EEH_PE_OP
, &pe_op
);
1535 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op
);
1542 static VFIOContainer
*vfio_eeh_as_container(AddressSpace
*as
)
1544 VFIOAddressSpace
*space
= vfio_get_address_space(as
);
1545 VFIOContainer
*container
= NULL
;
1547 if (QLIST_EMPTY(&space
->containers
)) {
1548 /* No containers to act on */
1552 container
= QLIST_FIRST(&space
->containers
);
1554 if (QLIST_NEXT(container
, next
)) {
1555 /* We don't yet have logic to synchronize EEH state across
1556 * multiple containers */
1562 vfio_put_address_space(space
);
1566 bool vfio_eeh_as_ok(AddressSpace
*as
)
1568 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1570 return (container
!= NULL
) && vfio_eeh_container_ok(container
);
1573 int vfio_eeh_as_op(AddressSpace
*as
, uint32_t op
)
1575 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1580 return vfio_eeh_container_op(container
, op
);