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"
32 #include "exec/ram_addr.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/range.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
40 #include "qapi/error.h"
41 #include "migration/migration.h"
43 VFIOGroupList vfio_group_list
=
44 QLIST_HEAD_INITIALIZER(vfio_group_list
);
45 static QLIST_HEAD(, VFIOAddressSpace
) vfio_address_spaces
=
46 QLIST_HEAD_INITIALIZER(vfio_address_spaces
);
50 * We have a single VFIO pseudo device per KVM VM. Once created it lives
51 * for the life of the VM. Closing the file descriptor only drops our
52 * reference to it and the device's reference to kvm. Therefore once
53 * initialized, this file descriptor is only released on QEMU exit and
54 * we'll re-use it should another vfio device be attached before then.
56 static int vfio_kvm_device_fd
= -1;
60 * Common VFIO interrupt disable
62 void vfio_disable_irqindex(VFIODevice
*vbasedev
, int index
)
64 struct vfio_irq_set irq_set
= {
65 .argsz
= sizeof(irq_set
),
66 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_TRIGGER
,
72 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
75 void vfio_unmask_single_irqindex(VFIODevice
*vbasedev
, int index
)
77 struct vfio_irq_set irq_set
= {
78 .argsz
= sizeof(irq_set
),
79 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_UNMASK
,
85 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
88 void vfio_mask_single_irqindex(VFIODevice
*vbasedev
, int index
)
90 struct vfio_irq_set irq_set
= {
91 .argsz
= sizeof(irq_set
),
92 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_MASK
,
98 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
101 static inline const char *action_to_str(int action
)
104 case VFIO_IRQ_SET_ACTION_MASK
:
106 case VFIO_IRQ_SET_ACTION_UNMASK
:
108 case VFIO_IRQ_SET_ACTION_TRIGGER
:
111 return "UNKNOWN ACTION";
115 static const char *index_to_str(VFIODevice
*vbasedev
, int index
)
117 if (vbasedev
->type
!= VFIO_DEVICE_TYPE_PCI
) {
122 case VFIO_PCI_INTX_IRQ_INDEX
:
124 case VFIO_PCI_MSI_IRQ_INDEX
:
126 case VFIO_PCI_MSIX_IRQ_INDEX
:
128 case VFIO_PCI_ERR_IRQ_INDEX
:
130 case VFIO_PCI_REQ_IRQ_INDEX
:
137 int vfio_set_irq_signaling(VFIODevice
*vbasedev
, int index
, int subindex
,
138 int action
, int fd
, Error
**errp
)
140 struct vfio_irq_set
*irq_set
;
145 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
147 irq_set
= g_malloc0(argsz
);
148 irq_set
->argsz
= argsz
;
149 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| action
;
150 irq_set
->index
= index
;
151 irq_set
->start
= subindex
;
153 pfd
= (int32_t *)&irq_set
->data
;
156 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
)) {
165 error_setg_errno(errp
, -ret
, "VFIO_DEVICE_SET_IRQS failure");
167 name
= index_to_str(vbasedev
, index
);
169 error_prepend(errp
, "%s-%d: ", name
, subindex
);
171 error_prepend(errp
, "index %d-%d: ", index
, subindex
);
174 "Failed to %s %s eventfd signaling for interrupt ",
175 fd
< 0 ? "tear down" : "set up", action_to_str(action
));
180 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
182 void vfio_region_write(void *opaque
, hwaddr addr
,
183 uint64_t data
, unsigned size
)
185 VFIORegion
*region
= opaque
;
186 VFIODevice
*vbasedev
= region
->vbasedev
;
199 buf
.word
= cpu_to_le16(data
);
202 buf
.dword
= cpu_to_le32(data
);
205 buf
.qword
= cpu_to_le64(data
);
208 hw_error("vfio: unsupported write size, %u bytes", size
);
212 if (pwrite(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
213 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", 0x%"PRIx64
215 __func__
, vbasedev
->name
, region
->nr
,
219 trace_vfio_region_write(vbasedev
->name
, region
->nr
, addr
, data
, size
);
222 * A read or write to a BAR always signals an INTx EOI. This will
223 * do nothing if not pending (including not in INTx mode). We assume
224 * that a BAR access is in response to an interrupt and that BAR
225 * accesses will service the interrupt. Unfortunately, we don't know
226 * which access will service the interrupt, so we're potentially
227 * getting quite a few host interrupts per guest interrupt.
229 vbasedev
->ops
->vfio_eoi(vbasedev
);
232 uint64_t vfio_region_read(void *opaque
,
233 hwaddr addr
, unsigned size
)
235 VFIORegion
*region
= opaque
;
236 VFIODevice
*vbasedev
= region
->vbasedev
;
245 if (pread(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
246 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", %d) failed: %m",
247 __func__
, vbasedev
->name
, region
->nr
,
256 data
= le16_to_cpu(buf
.word
);
259 data
= le32_to_cpu(buf
.dword
);
262 data
= le64_to_cpu(buf
.qword
);
265 hw_error("vfio: unsupported read size, %u bytes", size
);
269 trace_vfio_region_read(vbasedev
->name
, region
->nr
, addr
, size
, data
);
271 /* Same as write above */
272 vbasedev
->ops
->vfio_eoi(vbasedev
);
277 const MemoryRegionOps vfio_region_ops
= {
278 .read
= vfio_region_read
,
279 .write
= vfio_region_write
,
280 .endianness
= DEVICE_LITTLE_ENDIAN
,
282 .min_access_size
= 1,
283 .max_access_size
= 8,
286 .min_access_size
= 1,
287 .max_access_size
= 8,
292 * Device state interfaces
295 bool vfio_mig_active(void)
298 VFIODevice
*vbasedev
;
300 if (QLIST_EMPTY(&vfio_group_list
)) {
304 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
305 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
306 if (vbasedev
->migration_blocker
) {
314 static bool vfio_devices_all_dirty_tracking(VFIOContainer
*container
)
317 VFIODevice
*vbasedev
;
318 MigrationState
*ms
= migrate_get_current();
320 if (!migration_is_setup_or_active(ms
->state
)) {
324 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
325 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
326 VFIOMigration
*migration
= vbasedev
->migration
;
332 if ((vbasedev
->pre_copy_dirty_page_tracking
== ON_OFF_AUTO_OFF
)
333 && (migration
->device_state
& VFIO_DEVICE_STATE_RUNNING
)) {
341 static bool vfio_devices_all_running_and_saving(VFIOContainer
*container
)
344 VFIODevice
*vbasedev
;
345 MigrationState
*ms
= migrate_get_current();
347 if (!migration_is_setup_or_active(ms
->state
)) {
351 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
352 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
353 VFIOMigration
*migration
= vbasedev
->migration
;
359 if ((migration
->device_state
& VFIO_DEVICE_STATE_SAVING
) &&
360 (migration
->device_state
& VFIO_DEVICE_STATE_RUNNING
)) {
370 static int vfio_dma_unmap_bitmap(VFIOContainer
*container
,
371 hwaddr iova
, ram_addr_t size
,
372 IOMMUTLBEntry
*iotlb
)
374 struct vfio_iommu_type1_dma_unmap
*unmap
;
375 struct vfio_bitmap
*bitmap
;
376 uint64_t pages
= REAL_HOST_PAGE_ALIGN(size
) / qemu_real_host_page_size
;
379 unmap
= g_malloc0(sizeof(*unmap
) + sizeof(*bitmap
));
381 unmap
->argsz
= sizeof(*unmap
) + sizeof(*bitmap
);
384 unmap
->flags
|= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP
;
385 bitmap
= (struct vfio_bitmap
*)&unmap
->data
;
388 * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
389 * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize
390 * to qemu_real_host_page_size.
393 bitmap
->pgsize
= qemu_real_host_page_size
;
394 bitmap
->size
= ROUND_UP(pages
, sizeof(__u64
) * BITS_PER_BYTE
) /
397 if (bitmap
->size
> container
->max_dirty_bitmap_size
) {
398 error_report("UNMAP: Size of bitmap too big 0x%"PRIx64
,
399 (uint64_t)bitmap
->size
);
404 bitmap
->data
= g_try_malloc0(bitmap
->size
);
410 ret
= ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, unmap
);
412 cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap
->data
,
413 iotlb
->translated_addr
, pages
);
415 error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
418 g_free(bitmap
->data
);
425 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
427 static int vfio_dma_unmap(VFIOContainer
*container
,
428 hwaddr iova
, ram_addr_t size
,
429 IOMMUTLBEntry
*iotlb
)
431 struct vfio_iommu_type1_dma_unmap unmap
= {
432 .argsz
= sizeof(unmap
),
438 if (iotlb
&& container
->dirty_pages_supported
&&
439 vfio_devices_all_running_and_saving(container
)) {
440 return vfio_dma_unmap_bitmap(container
, iova
, size
, iotlb
);
443 while (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
445 * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
446 * v4.15) where an overflow in its wrap-around check prevents us from
447 * unmapping the last page of the address space. Test for the error
448 * condition and re-try the unmap excluding the last page. The
449 * expectation is that we've never mapped the last page anyway and this
450 * unmap request comes via vIOMMU support which also makes it unlikely
451 * that this page is used. This bug was introduced well after type1 v2
452 * support was introduced, so we shouldn't need to test for v1. A fix
453 * is queued for kernel v5.0 so this workaround can be removed once
454 * affected kernels are sufficiently deprecated.
456 if (errno
== EINVAL
&& unmap
.size
&& !(unmap
.iova
+ unmap
.size
) &&
457 container
->iommu_type
== VFIO_TYPE1v2_IOMMU
) {
458 trace_vfio_dma_unmap_overflow_workaround();
459 unmap
.size
-= 1ULL << ctz64(container
->pgsizes
);
462 error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno
));
469 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
470 ram_addr_t size
, void *vaddr
, bool readonly
)
472 struct vfio_iommu_type1_dma_map map
= {
473 .argsz
= sizeof(map
),
474 .flags
= VFIO_DMA_MAP_FLAG_READ
,
475 .vaddr
= (__u64
)(uintptr_t)vaddr
,
481 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
485 * Try the mapping, if it fails with EBUSY, unmap the region and try
486 * again. This shouldn't be necessary, but we sometimes see it in
489 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
490 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
, NULL
) == 0 &&
491 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
495 error_report("VFIO_MAP_DMA failed: %s", strerror(errno
));
499 static void vfio_host_win_add(VFIOContainer
*container
,
500 hwaddr min_iova
, hwaddr max_iova
,
501 uint64_t iova_pgsizes
)
503 VFIOHostDMAWindow
*hostwin
;
505 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
506 if (ranges_overlap(hostwin
->min_iova
,
507 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
509 max_iova
- min_iova
+ 1)) {
510 hw_error("%s: Overlapped IOMMU are not enabled", __func__
);
514 hostwin
= g_malloc0(sizeof(*hostwin
));
516 hostwin
->min_iova
= min_iova
;
517 hostwin
->max_iova
= max_iova
;
518 hostwin
->iova_pgsizes
= iova_pgsizes
;
519 QLIST_INSERT_HEAD(&container
->hostwin_list
, hostwin
, hostwin_next
);
522 static int vfio_host_win_del(VFIOContainer
*container
, hwaddr min_iova
,
525 VFIOHostDMAWindow
*hostwin
;
527 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
528 if (hostwin
->min_iova
== min_iova
&& hostwin
->max_iova
== max_iova
) {
529 QLIST_REMOVE(hostwin
, hostwin_next
);
537 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
539 return (!memory_region_is_ram(section
->mr
) &&
540 !memory_region_is_iommu(section
->mr
)) ||
542 * Sizing an enabled 64-bit BAR can cause spurious mappings to
543 * addresses in the upper part of the 64-bit address space. These
544 * are never accessed by the CPU and beyond the address width of
545 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
547 section
->offset_within_address_space
& (1ULL << 63);
550 /* Called with rcu_read_lock held. */
551 static bool vfio_get_xlat_addr(IOMMUTLBEntry
*iotlb
, void **vaddr
,
552 ram_addr_t
*ram_addr
, bool *read_only
)
556 hwaddr len
= iotlb
->addr_mask
+ 1;
557 bool writable
= iotlb
->perm
& IOMMU_WO
;
560 * The IOMMU TLB entry we have just covers translation through
561 * this IOMMU to its immediate target. We need to translate
562 * it the rest of the way through to memory.
564 mr
= address_space_translate(&address_space_memory
,
565 iotlb
->translated_addr
,
566 &xlat
, &len
, writable
,
567 MEMTXATTRS_UNSPECIFIED
);
568 if (!memory_region_is_ram(mr
)) {
569 error_report("iommu map to non memory area %"HWADDR_PRIx
"",
575 * Translation truncates length to the IOMMU page size,
576 * check that it did not truncate too much.
578 if (len
& iotlb
->addr_mask
) {
579 error_report("iommu has granularity incompatible with target AS");
584 *vaddr
= memory_region_get_ram_ptr(mr
) + xlat
;
588 *ram_addr
= memory_region_get_ram_addr(mr
) + xlat
;
592 *read_only
= !writable
|| mr
->readonly
;
598 static void vfio_iommu_map_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
600 VFIOGuestIOMMU
*giommu
= container_of(n
, VFIOGuestIOMMU
, n
);
601 VFIOContainer
*container
= giommu
->container
;
602 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
606 trace_vfio_iommu_map_notify(iotlb
->perm
== IOMMU_NONE
? "UNMAP" : "MAP",
607 iova
, iova
+ iotlb
->addr_mask
);
609 if (iotlb
->target_as
!= &address_space_memory
) {
610 error_report("Wrong target AS \"%s\", only system memory is allowed",
611 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
617 if ((iotlb
->perm
& IOMMU_RW
) != IOMMU_NONE
) {
620 if (!vfio_get_xlat_addr(iotlb
, &vaddr
, NULL
, &read_only
)) {
624 * vaddr is only valid until rcu_read_unlock(). But after
625 * vfio_dma_map has set up the mapping the pages will be
626 * pinned by the kernel. This makes sure that the RAM backend
627 * of vaddr will always be there, even if the memory object is
628 * destroyed and its backing memory munmap-ed.
630 ret
= vfio_dma_map(container
, iova
,
631 iotlb
->addr_mask
+ 1, vaddr
,
634 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
635 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
637 iotlb
->addr_mask
+ 1, vaddr
, ret
);
640 ret
= vfio_dma_unmap(container
, iova
, iotlb
->addr_mask
+ 1, iotlb
);
642 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
643 "0x%"HWADDR_PRIx
") = %d (%m)",
645 iotlb
->addr_mask
+ 1, ret
);
652 static void vfio_listener_region_add(MemoryListener
*listener
,
653 MemoryRegionSection
*section
)
655 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
657 Int128 llend
, llsize
;
660 VFIOHostDMAWindow
*hostwin
;
664 if (vfio_listener_skipped_section(section
)) {
665 trace_vfio_listener_region_add_skip(
666 section
->offset_within_address_space
,
667 section
->offset_within_address_space
+
668 int128_get64(int128_sub(section
->size
, int128_one())));
672 if (unlikely((section
->offset_within_address_space
&
673 ~qemu_real_host_page_mask
) !=
674 (section
->offset_within_region
& ~qemu_real_host_page_mask
))) {
675 error_report("%s received unaligned region", __func__
);
679 iova
= REAL_HOST_PAGE_ALIGN(section
->offset_within_address_space
);
680 llend
= int128_make64(section
->offset_within_address_space
);
681 llend
= int128_add(llend
, section
->size
);
682 llend
= int128_and(llend
, int128_exts64(qemu_real_host_page_mask
));
684 if (int128_ge(int128_make64(iova
), llend
)) {
687 end
= int128_get64(int128_sub(llend
, int128_one()));
689 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
692 /* For now intersections are not allowed, we may relax this later */
693 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
694 if (ranges_overlap(hostwin
->min_iova
,
695 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
696 section
->offset_within_address_space
,
697 int128_get64(section
->size
))) {
699 "region [0x%"PRIx64
",0x%"PRIx64
"] overlaps with existing"
700 "host DMA window [0x%"PRIx64
",0x%"PRIx64
"]",
701 section
->offset_within_address_space
,
702 section
->offset_within_address_space
+
703 int128_get64(section
->size
) - 1,
704 hostwin
->min_iova
, hostwin
->max_iova
);
709 ret
= vfio_spapr_create_window(container
, section
, &pgsize
);
711 error_setg_errno(&err
, -ret
, "Failed to create SPAPR window");
715 vfio_host_win_add(container
, section
->offset_within_address_space
,
716 section
->offset_within_address_space
+
717 int128_get64(section
->size
) - 1, pgsize
);
721 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
722 struct kvm_vfio_spapr_tce param
;
723 struct kvm_device_attr attr
= {
724 .group
= KVM_DEV_VFIO_GROUP
,
725 .attr
= KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE
,
726 .addr
= (uint64_t)(unsigned long)¶m
,
729 if (!memory_region_iommu_get_attr(iommu_mr
, IOMMU_ATTR_SPAPR_TCE_FD
,
731 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
732 param
.groupfd
= group
->fd
;
733 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
734 error_report("vfio: failed to setup fd %d "
735 "for a group with fd %d: %s",
736 param
.tablefd
, param
.groupfd
,
740 trace_vfio_spapr_group_attach(param
.groupfd
, param
.tablefd
);
747 hostwin_found
= false;
748 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
749 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
750 hostwin_found
= true;
755 if (!hostwin_found
) {
756 error_setg(&err
, "Container %p can't map guest IOVA region"
757 " 0x%"HWADDR_PRIx
"..0x%"HWADDR_PRIx
, container
, iova
, end
);
761 memory_region_ref(section
->mr
);
763 if (memory_region_is_iommu(section
->mr
)) {
764 VFIOGuestIOMMU
*giommu
;
765 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
768 trace_vfio_listener_region_add_iommu(iova
, end
);
770 * FIXME: For VFIO iommu types which have KVM acceleration to
771 * avoid bouncing all map/unmaps through qemu this way, this
772 * would be the right place to wire that up (tell the KVM
773 * device emulation the VFIO iommu handles to use).
775 giommu
= g_malloc0(sizeof(*giommu
));
776 giommu
->iommu
= iommu_mr
;
777 giommu
->iommu_offset
= section
->offset_within_address_space
-
778 section
->offset_within_region
;
779 giommu
->container
= container
;
780 llend
= int128_add(int128_make64(section
->offset_within_region
),
782 llend
= int128_sub(llend
, int128_one());
783 iommu_idx
= memory_region_iommu_attrs_to_index(iommu_mr
,
784 MEMTXATTRS_UNSPECIFIED
);
785 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
786 IOMMU_NOTIFIER_IOTLB_EVENTS
,
787 section
->offset_within_region
,
791 ret
= memory_region_iommu_set_page_size_mask(giommu
->iommu
,
799 ret
= memory_region_register_iommu_notifier(section
->mr
, &giommu
->n
,
805 QLIST_INSERT_HEAD(&container
->giommu_list
, giommu
, giommu_next
);
806 memory_region_iommu_replay(giommu
->iommu
, &giommu
->n
);
811 /* Here we assume that memory_region_is_ram(section->mr)==true */
813 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
814 section
->offset_within_region
+
815 (iova
- section
->offset_within_address_space
);
817 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
819 llsize
= int128_sub(llend
, int128_make64(iova
));
821 if (memory_region_is_ram_device(section
->mr
)) {
822 hwaddr pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
824 if ((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
)) {
825 trace_vfio_listener_region_add_no_dma_map(
826 memory_region_name(section
->mr
),
827 section
->offset_within_address_space
,
828 int128_getlo(section
->size
),
834 ret
= vfio_dma_map(container
, iova
, int128_get64(llsize
),
835 vaddr
, section
->readonly
);
837 error_setg(&err
, "vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
838 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
839 container
, iova
, int128_get64(llsize
), vaddr
, ret
);
840 if (memory_region_is_ram_device(section
->mr
)) {
841 /* Allow unexpected mappings not to be fatal for RAM devices */
842 error_report_err(err
);
851 if (memory_region_is_ram_device(section
->mr
)) {
852 error_report("failed to vfio_dma_map. pci p2p may not work");
856 * On the initfn path, store the first error in the container so we
857 * can gracefully fail. Runtime, there's not much we can do other
858 * than throw a hardware error.
860 if (!container
->initialized
) {
861 if (!container
->error
) {
862 error_propagate_prepend(&container
->error
, err
,
864 memory_region_name(section
->mr
));
869 error_report_err(err
);
870 hw_error("vfio: DMA mapping failed, unable to continue");
874 static void vfio_listener_region_del(MemoryListener
*listener
,
875 MemoryRegionSection
*section
)
877 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
879 Int128 llend
, llsize
;
881 bool try_unmap
= true;
883 if (vfio_listener_skipped_section(section
)) {
884 trace_vfio_listener_region_del_skip(
885 section
->offset_within_address_space
,
886 section
->offset_within_address_space
+
887 int128_get64(int128_sub(section
->size
, int128_one())));
891 if (unlikely((section
->offset_within_address_space
&
892 ~qemu_real_host_page_mask
) !=
893 (section
->offset_within_region
& ~qemu_real_host_page_mask
))) {
894 error_report("%s received unaligned region", __func__
);
898 if (memory_region_is_iommu(section
->mr
)) {
899 VFIOGuestIOMMU
*giommu
;
901 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
902 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
903 giommu
->n
.start
== section
->offset_within_region
) {
904 memory_region_unregister_iommu_notifier(section
->mr
,
906 QLIST_REMOVE(giommu
, giommu_next
);
913 * FIXME: We assume the one big unmap below is adequate to
914 * remove any individual page mappings in the IOMMU which
915 * might have been copied into VFIO. This works for a page table
916 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
917 * That may not be true for all IOMMU types.
921 iova
= REAL_HOST_PAGE_ALIGN(section
->offset_within_address_space
);
922 llend
= int128_make64(section
->offset_within_address_space
);
923 llend
= int128_add(llend
, section
->size
);
924 llend
= int128_and(llend
, int128_exts64(qemu_real_host_page_mask
));
926 if (int128_ge(int128_make64(iova
), llend
)) {
929 end
= int128_get64(int128_sub(llend
, int128_one()));
931 llsize
= int128_sub(llend
, int128_make64(iova
));
933 trace_vfio_listener_region_del(iova
, end
);
935 if (memory_region_is_ram_device(section
->mr
)) {
937 VFIOHostDMAWindow
*hostwin
;
938 bool hostwin_found
= false;
940 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
941 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
942 hostwin_found
= true;
946 assert(hostwin_found
); /* or region_add() would have failed */
948 pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
949 try_unmap
= !((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
));
953 if (int128_eq(llsize
, int128_2_64())) {
954 /* The unmap ioctl doesn't accept a full 64-bit span. */
955 llsize
= int128_rshift(llsize
, 1);
956 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
), NULL
);
958 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
959 "0x%"HWADDR_PRIx
") = %d (%m)",
960 container
, iova
, int128_get64(llsize
), ret
);
962 iova
+= int128_get64(llsize
);
964 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
), NULL
);
966 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
967 "0x%"HWADDR_PRIx
") = %d (%m)",
968 container
, iova
, int128_get64(llsize
), ret
);
972 memory_region_unref(section
->mr
);
974 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
975 vfio_spapr_remove_window(container
,
976 section
->offset_within_address_space
);
977 if (vfio_host_win_del(container
,
978 section
->offset_within_address_space
,
979 section
->offset_within_address_space
+
980 int128_get64(section
->size
) - 1) < 0) {
981 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx
,
982 __func__
, section
->offset_within_address_space
);
987 static void vfio_set_dirty_page_tracking(VFIOContainer
*container
, bool start
)
990 struct vfio_iommu_type1_dirty_bitmap dirty
= {
991 .argsz
= sizeof(dirty
),
995 dirty
.flags
= VFIO_IOMMU_DIRTY_PAGES_FLAG_START
;
997 dirty
.flags
= VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP
;
1000 ret
= ioctl(container
->fd
, VFIO_IOMMU_DIRTY_PAGES
, &dirty
);
1002 error_report("Failed to set dirty tracking flag 0x%x errno: %d",
1003 dirty
.flags
, errno
);
1007 static void vfio_listener_log_global_start(MemoryListener
*listener
)
1009 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
1011 vfio_set_dirty_page_tracking(container
, true);
1014 static void vfio_listener_log_global_stop(MemoryListener
*listener
)
1016 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
1018 vfio_set_dirty_page_tracking(container
, false);
1021 static int vfio_get_dirty_bitmap(VFIOContainer
*container
, uint64_t iova
,
1022 uint64_t size
, ram_addr_t ram_addr
)
1024 struct vfio_iommu_type1_dirty_bitmap
*dbitmap
;
1025 struct vfio_iommu_type1_dirty_bitmap_get
*range
;
1029 dbitmap
= g_malloc0(sizeof(*dbitmap
) + sizeof(*range
));
1031 dbitmap
->argsz
= sizeof(*dbitmap
) + sizeof(*range
);
1032 dbitmap
->flags
= VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP
;
1033 range
= (struct vfio_iommu_type1_dirty_bitmap_get
*)&dbitmap
->data
;
1038 * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1039 * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize
1040 * to qemu_real_host_page_size.
1042 range
->bitmap
.pgsize
= qemu_real_host_page_size
;
1044 pages
= REAL_HOST_PAGE_ALIGN(range
->size
) / qemu_real_host_page_size
;
1045 range
->bitmap
.size
= ROUND_UP(pages
, sizeof(__u64
) * BITS_PER_BYTE
) /
1047 range
->bitmap
.data
= g_try_malloc0(range
->bitmap
.size
);
1048 if (!range
->bitmap
.data
) {
1053 ret
= ioctl(container
->fd
, VFIO_IOMMU_DIRTY_PAGES
, dbitmap
);
1055 error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1056 " size: 0x%"PRIx64
" err: %d", (uint64_t)range
->iova
,
1057 (uint64_t)range
->size
, errno
);
1061 cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range
->bitmap
.data
,
1064 trace_vfio_get_dirty_bitmap(container
->fd
, range
->iova
, range
->size
,
1065 range
->bitmap
.size
, ram_addr
);
1067 g_free(range
->bitmap
.data
);
1075 VFIOGuestIOMMU
*giommu
;
1076 } vfio_giommu_dirty_notifier
;
1078 static void vfio_iommu_map_dirty_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
1080 vfio_giommu_dirty_notifier
*gdn
= container_of(n
,
1081 vfio_giommu_dirty_notifier
, n
);
1082 VFIOGuestIOMMU
*giommu
= gdn
->giommu
;
1083 VFIOContainer
*container
= giommu
->container
;
1084 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
1085 ram_addr_t translated_addr
;
1087 trace_vfio_iommu_map_dirty_notify(iova
, iova
+ iotlb
->addr_mask
);
1089 if (iotlb
->target_as
!= &address_space_memory
) {
1090 error_report("Wrong target AS \"%s\", only system memory is allowed",
1091 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
1096 if (vfio_get_xlat_addr(iotlb
, NULL
, &translated_addr
, NULL
)) {
1099 ret
= vfio_get_dirty_bitmap(container
, iova
, iotlb
->addr_mask
+ 1,
1102 error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx
", "
1103 "0x%"HWADDR_PRIx
") = %d (%m)",
1105 iotlb
->addr_mask
+ 1, ret
);
1111 static int vfio_sync_dirty_bitmap(VFIOContainer
*container
,
1112 MemoryRegionSection
*section
)
1114 ram_addr_t ram_addr
;
1116 if (memory_region_is_iommu(section
->mr
)) {
1117 VFIOGuestIOMMU
*giommu
;
1119 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
1120 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
1121 giommu
->n
.start
== section
->offset_within_region
) {
1123 vfio_giommu_dirty_notifier gdn
= { .giommu
= giommu
};
1124 int idx
= memory_region_iommu_attrs_to_index(giommu
->iommu
,
1125 MEMTXATTRS_UNSPECIFIED
);
1127 llend
= int128_add(int128_make64(section
->offset_within_region
),
1129 llend
= int128_sub(llend
, int128_one());
1131 iommu_notifier_init(&gdn
.n
,
1132 vfio_iommu_map_dirty_notify
,
1134 section
->offset_within_region
,
1135 int128_get64(llend
),
1137 memory_region_iommu_replay(giommu
->iommu
, &gdn
.n
);
1144 ram_addr
= memory_region_get_ram_addr(section
->mr
) +
1145 section
->offset_within_region
;
1147 return vfio_get_dirty_bitmap(container
,
1148 REAL_HOST_PAGE_ALIGN(section
->offset_within_address_space
),
1149 int128_get64(section
->size
), ram_addr
);
1152 static void vfio_listener_log_sync(MemoryListener
*listener
,
1153 MemoryRegionSection
*section
)
1155 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
1157 if (vfio_listener_skipped_section(section
) ||
1158 !container
->dirty_pages_supported
) {
1162 if (vfio_devices_all_dirty_tracking(container
)) {
1163 vfio_sync_dirty_bitmap(container
, section
);
1167 static const MemoryListener vfio_memory_listener
= {
1168 .region_add
= vfio_listener_region_add
,
1169 .region_del
= vfio_listener_region_del
,
1170 .log_global_start
= vfio_listener_log_global_start
,
1171 .log_global_stop
= vfio_listener_log_global_stop
,
1172 .log_sync
= vfio_listener_log_sync
,
1175 static void vfio_listener_release(VFIOContainer
*container
)
1177 memory_listener_unregister(&container
->listener
);
1178 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
1179 memory_listener_unregister(&container
->prereg_listener
);
1183 static struct vfio_info_cap_header
*
1184 vfio_get_cap(void *ptr
, uint32_t cap_offset
, uint16_t id
)
1186 struct vfio_info_cap_header
*hdr
;
1188 for (hdr
= ptr
+ cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
1189 if (hdr
->id
== id
) {
1197 struct vfio_info_cap_header
*
1198 vfio_get_region_info_cap(struct vfio_region_info
*info
, uint16_t id
)
1200 if (!(info
->flags
& VFIO_REGION_INFO_FLAG_CAPS
)) {
1204 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1207 static struct vfio_info_cap_header
*
1208 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info
*info
, uint16_t id
)
1210 if (!(info
->flags
& VFIO_IOMMU_INFO_CAPS
)) {
1214 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1217 struct vfio_info_cap_header
*
1218 vfio_get_device_info_cap(struct vfio_device_info
*info
, uint16_t id
)
1220 if (!(info
->flags
& VFIO_DEVICE_FLAGS_CAPS
)) {
1224 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1227 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info
*info
,
1228 unsigned int *avail
)
1230 struct vfio_info_cap_header
*hdr
;
1231 struct vfio_iommu_type1_info_dma_avail
*cap
;
1233 /* If the capability cannot be found, assume no DMA limiting */
1234 hdr
= vfio_get_iommu_type1_info_cap(info
,
1235 VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL
);
1240 if (avail
!= NULL
) {
1242 *avail
= cap
->avail
;
1248 static int vfio_setup_region_sparse_mmaps(VFIORegion
*region
,
1249 struct vfio_region_info
*info
)
1251 struct vfio_info_cap_header
*hdr
;
1252 struct vfio_region_info_cap_sparse_mmap
*sparse
;
1255 hdr
= vfio_get_region_info_cap(info
, VFIO_REGION_INFO_CAP_SPARSE_MMAP
);
1260 sparse
= container_of(hdr
, struct vfio_region_info_cap_sparse_mmap
, header
);
1262 trace_vfio_region_sparse_mmap_header(region
->vbasedev
->name
,
1263 region
->nr
, sparse
->nr_areas
);
1265 region
->mmaps
= g_new0(VFIOMmap
, sparse
->nr_areas
);
1267 for (i
= 0, j
= 0; i
< sparse
->nr_areas
; i
++) {
1268 trace_vfio_region_sparse_mmap_entry(i
, sparse
->areas
[i
].offset
,
1269 sparse
->areas
[i
].offset
+
1270 sparse
->areas
[i
].size
);
1272 if (sparse
->areas
[i
].size
) {
1273 region
->mmaps
[j
].offset
= sparse
->areas
[i
].offset
;
1274 region
->mmaps
[j
].size
= sparse
->areas
[i
].size
;
1279 region
->nr_mmaps
= j
;
1280 region
->mmaps
= g_realloc(region
->mmaps
, j
* sizeof(VFIOMmap
));
1285 int vfio_region_setup(Object
*obj
, VFIODevice
*vbasedev
, VFIORegion
*region
,
1286 int index
, const char *name
)
1288 struct vfio_region_info
*info
;
1291 ret
= vfio_get_region_info(vbasedev
, index
, &info
);
1296 region
->vbasedev
= vbasedev
;
1297 region
->flags
= info
->flags
;
1298 region
->size
= info
->size
;
1299 region
->fd_offset
= info
->offset
;
1303 region
->mem
= g_new0(MemoryRegion
, 1);
1304 memory_region_init_io(region
->mem
, obj
, &vfio_region_ops
,
1305 region
, name
, region
->size
);
1307 if (!vbasedev
->no_mmap
&&
1308 region
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
1310 ret
= vfio_setup_region_sparse_mmaps(region
, info
);
1313 region
->nr_mmaps
= 1;
1314 region
->mmaps
= g_new0(VFIOMmap
, region
->nr_mmaps
);
1315 region
->mmaps
[0].offset
= 0;
1316 region
->mmaps
[0].size
= region
->size
;
1323 trace_vfio_region_setup(vbasedev
->name
, index
, name
,
1324 region
->flags
, region
->fd_offset
, region
->size
);
1328 static void vfio_subregion_unmap(VFIORegion
*region
, int index
)
1330 trace_vfio_region_unmap(memory_region_name(®ion
->mmaps
[index
].mem
),
1331 region
->mmaps
[index
].offset
,
1332 region
->mmaps
[index
].offset
+
1333 region
->mmaps
[index
].size
- 1);
1334 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[index
].mem
);
1335 munmap(region
->mmaps
[index
].mmap
, region
->mmaps
[index
].size
);
1336 object_unparent(OBJECT(®ion
->mmaps
[index
].mem
));
1337 region
->mmaps
[index
].mmap
= NULL
;
1340 int vfio_region_mmap(VFIORegion
*region
)
1349 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_READ
? PROT_READ
: 0;
1350 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_WRITE
? PROT_WRITE
: 0;
1352 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1353 region
->mmaps
[i
].mmap
= mmap(NULL
, region
->mmaps
[i
].size
, prot
,
1354 MAP_SHARED
, region
->vbasedev
->fd
,
1356 region
->mmaps
[i
].offset
);
1357 if (region
->mmaps
[i
].mmap
== MAP_FAILED
) {
1360 trace_vfio_region_mmap_fault(memory_region_name(region
->mem
), i
,
1362 region
->mmaps
[i
].offset
,
1364 region
->mmaps
[i
].offset
+
1365 region
->mmaps
[i
].size
- 1, ret
);
1367 region
->mmaps
[i
].mmap
= NULL
;
1369 for (i
--; i
>= 0; i
--) {
1370 vfio_subregion_unmap(region
, i
);
1376 name
= g_strdup_printf("%s mmaps[%d]",
1377 memory_region_name(region
->mem
), i
);
1378 memory_region_init_ram_device_ptr(®ion
->mmaps
[i
].mem
,
1379 memory_region_owner(region
->mem
),
1380 name
, region
->mmaps
[i
].size
,
1381 region
->mmaps
[i
].mmap
);
1383 memory_region_add_subregion(region
->mem
, region
->mmaps
[i
].offset
,
1384 ®ion
->mmaps
[i
].mem
);
1386 trace_vfio_region_mmap(memory_region_name(®ion
->mmaps
[i
].mem
),
1387 region
->mmaps
[i
].offset
,
1388 region
->mmaps
[i
].offset
+
1389 region
->mmaps
[i
].size
- 1);
1395 void vfio_region_unmap(VFIORegion
*region
)
1403 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1404 if (region
->mmaps
[i
].mmap
) {
1405 vfio_subregion_unmap(region
, i
);
1410 void vfio_region_exit(VFIORegion
*region
)
1418 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1419 if (region
->mmaps
[i
].mmap
) {
1420 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
1424 trace_vfio_region_exit(region
->vbasedev
->name
, region
->nr
);
1427 void vfio_region_finalize(VFIORegion
*region
)
1435 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1436 if (region
->mmaps
[i
].mmap
) {
1437 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
1438 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
1442 object_unparent(OBJECT(region
->mem
));
1444 g_free(region
->mem
);
1445 g_free(region
->mmaps
);
1447 trace_vfio_region_finalize(region
->vbasedev
->name
, region
->nr
);
1450 region
->mmaps
= NULL
;
1451 region
->nr_mmaps
= 0;
1457 void vfio_region_mmaps_set_enabled(VFIORegion
*region
, bool enabled
)
1465 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1466 if (region
->mmaps
[i
].mmap
) {
1467 memory_region_set_enabled(®ion
->mmaps
[i
].mem
, enabled
);
1471 trace_vfio_region_mmaps_set_enabled(memory_region_name(region
->mem
),
1475 void vfio_reset_handler(void *opaque
)
1478 VFIODevice
*vbasedev
;
1480 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1481 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1482 if (vbasedev
->dev
->realized
) {
1483 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
1488 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1489 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1490 if (vbasedev
->dev
->realized
&& vbasedev
->needs_reset
) {
1491 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
1497 static void vfio_kvm_device_add_group(VFIOGroup
*group
)
1500 struct kvm_device_attr attr
= {
1501 .group
= KVM_DEV_VFIO_GROUP
,
1502 .attr
= KVM_DEV_VFIO_GROUP_ADD
,
1503 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1506 if (!kvm_enabled()) {
1510 if (vfio_kvm_device_fd
< 0) {
1511 struct kvm_create_device cd
= {
1512 .type
= KVM_DEV_TYPE_VFIO
,
1515 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
1516 error_report("Failed to create KVM VFIO device: %m");
1520 vfio_kvm_device_fd
= cd
.fd
;
1523 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1524 error_report("Failed to add group %d to KVM VFIO device: %m",
1530 static void vfio_kvm_device_del_group(VFIOGroup
*group
)
1533 struct kvm_device_attr attr
= {
1534 .group
= KVM_DEV_VFIO_GROUP
,
1535 .attr
= KVM_DEV_VFIO_GROUP_DEL
,
1536 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1539 if (vfio_kvm_device_fd
< 0) {
1543 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1544 error_report("Failed to remove group %d from KVM VFIO device: %m",
1550 static VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
1552 VFIOAddressSpace
*space
;
1554 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
1555 if (space
->as
== as
) {
1560 /* No suitable VFIOAddressSpace, create a new one */
1561 space
= g_malloc0(sizeof(*space
));
1563 QLIST_INIT(&space
->containers
);
1565 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
1570 static void vfio_put_address_space(VFIOAddressSpace
*space
)
1572 if (QLIST_EMPTY(&space
->containers
)) {
1573 QLIST_REMOVE(space
, list
);
1579 * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1581 static int vfio_get_iommu_type(VFIOContainer
*container
,
1584 int iommu_types
[] = { VFIO_TYPE1v2_IOMMU
, VFIO_TYPE1_IOMMU
,
1585 VFIO_SPAPR_TCE_v2_IOMMU
, VFIO_SPAPR_TCE_IOMMU
};
1588 for (i
= 0; i
< ARRAY_SIZE(iommu_types
); i
++) {
1589 if (ioctl(container
->fd
, VFIO_CHECK_EXTENSION
, iommu_types
[i
])) {
1590 return iommu_types
[i
];
1593 error_setg(errp
, "No available IOMMU models");
1597 static int vfio_init_container(VFIOContainer
*container
, int group_fd
,
1600 int iommu_type
, ret
;
1602 iommu_type
= vfio_get_iommu_type(container
, errp
);
1603 if (iommu_type
< 0) {
1607 ret
= ioctl(group_fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
);
1609 error_setg_errno(errp
, errno
, "Failed to set group container");
1613 while (ioctl(container
->fd
, VFIO_SET_IOMMU
, iommu_type
)) {
1614 if (iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
1616 * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1617 * v2, the running platform may not support v2 and there is no
1618 * way to guess it until an IOMMU group gets added to the container.
1619 * So in case it fails with v2, try v1 as a fallback.
1621 iommu_type
= VFIO_SPAPR_TCE_IOMMU
;
1624 error_setg_errno(errp
, errno
, "Failed to set iommu for container");
1628 container
->iommu_type
= iommu_type
;
1632 static int vfio_get_iommu_info(VFIOContainer
*container
,
1633 struct vfio_iommu_type1_info
**info
)
1636 size_t argsz
= sizeof(struct vfio_iommu_type1_info
);
1638 *info
= g_new0(struct vfio_iommu_type1_info
, 1);
1640 (*info
)->argsz
= argsz
;
1642 if (ioctl(container
->fd
, VFIO_IOMMU_GET_INFO
, *info
)) {
1648 if (((*info
)->argsz
> argsz
)) {
1649 argsz
= (*info
)->argsz
;
1650 *info
= g_realloc(*info
, argsz
);
1657 static struct vfio_info_cap_header
*
1658 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info
*info
, uint16_t id
)
1660 struct vfio_info_cap_header
*hdr
;
1663 if (!(info
->flags
& VFIO_IOMMU_INFO_CAPS
)) {
1667 for (hdr
= ptr
+ info
->cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
1668 if (hdr
->id
== id
) {
1676 static void vfio_get_iommu_info_migration(VFIOContainer
*container
,
1677 struct vfio_iommu_type1_info
*info
)
1679 struct vfio_info_cap_header
*hdr
;
1680 struct vfio_iommu_type1_info_cap_migration
*cap_mig
;
1682 hdr
= vfio_get_iommu_info_cap(info
, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION
);
1687 cap_mig
= container_of(hdr
, struct vfio_iommu_type1_info_cap_migration
,
1691 * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1692 * qemu_real_host_page_size to mark those dirty.
1694 if (cap_mig
->pgsize_bitmap
& qemu_real_host_page_size
) {
1695 container
->dirty_pages_supported
= true;
1696 container
->max_dirty_bitmap_size
= cap_mig
->max_dirty_bitmap_size
;
1697 container
->dirty_pgsizes
= cap_mig
->pgsize_bitmap
;
1701 static int vfio_connect_container(VFIOGroup
*group
, AddressSpace
*as
,
1704 VFIOContainer
*container
;
1706 VFIOAddressSpace
*space
;
1708 space
= vfio_get_address_space(as
);
1711 * VFIO is currently incompatible with discarding of RAM insofar as the
1712 * madvise to purge (zap) the page from QEMU's address space does not
1713 * interact with the memory API and therefore leaves stale virtual to
1714 * physical mappings in the IOMMU if the page was previously pinned. We
1715 * therefore set discarding broken for each group added to a container,
1716 * whether the container is used individually or shared. This provides
1717 * us with options to allow devices within a group to opt-in and allow
1718 * discarding, so long as it is done consistently for a group (for instance
1719 * if the device is an mdev device where it is known that the host vendor
1720 * driver will never pin pages outside of the working set of the guest
1721 * driver, which would thus not be discarding candidates).
1723 * The first opportunity to induce pinning occurs here where we attempt to
1724 * attach the group to existing containers within the AddressSpace. If any
1725 * pages are already zapped from the virtual address space, such as from
1726 * previous discards, new pinning will cause valid mappings to be
1727 * re-established. Likewise, when the overall MemoryListener for a new
1728 * container is registered, a replay of mappings within the AddressSpace
1729 * will occur, re-establishing any previously zapped pages as well.
1731 * Especially virtio-balloon is currently only prevented from discarding
1732 * new memory, it will not yet set ram_block_discard_set_required() and
1733 * therefore, neither stops us here or deals with the sudden memory
1734 * consumption of inflated memory.
1736 ret
= ram_block_discard_disable(true);
1738 error_setg_errno(errp
, -ret
, "Cannot set discarding of RAM broken");
1742 QLIST_FOREACH(container
, &space
->containers
, next
) {
1743 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
1744 group
->container
= container
;
1745 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1746 vfio_kvm_device_add_group(group
);
1751 fd
= qemu_open_old("/dev/vfio/vfio", O_RDWR
);
1753 error_setg_errno(errp
, errno
, "failed to open /dev/vfio/vfio");
1755 goto put_space_exit
;
1758 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
1759 if (ret
!= VFIO_API_VERSION
) {
1760 error_setg(errp
, "supported vfio version: %d, "
1761 "reported version: %d", VFIO_API_VERSION
, ret
);
1766 container
= g_malloc0(sizeof(*container
));
1767 container
->space
= space
;
1769 container
->error
= NULL
;
1770 container
->dirty_pages_supported
= false;
1771 QLIST_INIT(&container
->giommu_list
);
1772 QLIST_INIT(&container
->hostwin_list
);
1774 ret
= vfio_init_container(container
, group
->fd
, errp
);
1776 goto free_container_exit
;
1779 switch (container
->iommu_type
) {
1780 case VFIO_TYPE1v2_IOMMU
:
1781 case VFIO_TYPE1_IOMMU
:
1783 struct vfio_iommu_type1_info
*info
;
1786 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1787 * IOVA whatsoever. That's not actually true, but the current
1788 * kernel interface doesn't tell us what it can map, and the
1789 * existing Type1 IOMMUs generally support any IOVA we're
1790 * going to actually try in practice.
1792 ret
= vfio_get_iommu_info(container
, &info
);
1794 if (ret
|| !(info
->flags
& VFIO_IOMMU_INFO_PGSIZES
)) {
1795 /* Assume 4k IOVA page size */
1796 info
->iova_pgsizes
= 4096;
1798 vfio_host_win_add(container
, 0, (hwaddr
)-1, info
->iova_pgsizes
);
1799 container
->pgsizes
= info
->iova_pgsizes
;
1802 vfio_get_iommu_info_migration(container
, info
);
1807 case VFIO_SPAPR_TCE_v2_IOMMU
:
1808 case VFIO_SPAPR_TCE_IOMMU
:
1810 struct vfio_iommu_spapr_tce_info info
;
1811 bool v2
= container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
;
1814 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1815 * when container fd is closed so we do not call it explicitly
1819 ret
= ioctl(fd
, VFIO_IOMMU_ENABLE
);
1821 error_setg_errno(errp
, errno
, "failed to enable container");
1823 goto free_container_exit
;
1826 container
->prereg_listener
= vfio_prereg_listener
;
1828 memory_listener_register(&container
->prereg_listener
,
1829 &address_space_memory
);
1830 if (container
->error
) {
1831 memory_listener_unregister(&container
->prereg_listener
);
1833 error_propagate_prepend(errp
, container
->error
,
1834 "RAM memory listener initialization failed: ");
1835 goto free_container_exit
;
1839 info
.argsz
= sizeof(info
);
1840 ret
= ioctl(fd
, VFIO_IOMMU_SPAPR_TCE_GET_INFO
, &info
);
1842 error_setg_errno(errp
, errno
,
1843 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1846 memory_listener_unregister(&container
->prereg_listener
);
1848 goto free_container_exit
;
1852 container
->pgsizes
= info
.ddw
.pgsizes
;
1854 * There is a default window in just created container.
1855 * To make region_add/del simpler, we better remove this
1856 * window now and let those iommu_listener callbacks
1857 * create/remove them when needed.
1859 ret
= vfio_spapr_remove_window(container
, info
.dma32_window_start
);
1861 error_setg_errno(errp
, -ret
,
1862 "failed to remove existing window");
1863 goto free_container_exit
;
1866 /* The default table uses 4K pages */
1867 container
->pgsizes
= 0x1000;
1868 vfio_host_win_add(container
, info
.dma32_window_start
,
1869 info
.dma32_window_start
+
1870 info
.dma32_window_size
- 1,
1876 vfio_kvm_device_add_group(group
);
1878 QLIST_INIT(&container
->group_list
);
1879 QLIST_INSERT_HEAD(&space
->containers
, container
, next
);
1881 group
->container
= container
;
1882 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1884 container
->listener
= vfio_memory_listener
;
1886 memory_listener_register(&container
->listener
, container
->space
->as
);
1888 if (container
->error
) {
1890 error_propagate_prepend(errp
, container
->error
,
1891 "memory listener initialization failed: ");
1892 goto listener_release_exit
;
1895 container
->initialized
= true;
1898 listener_release_exit
:
1899 QLIST_REMOVE(group
, container_next
);
1900 QLIST_REMOVE(container
, next
);
1901 vfio_kvm_device_del_group(group
);
1902 vfio_listener_release(container
);
1904 free_container_exit
:
1911 ram_block_discard_disable(false);
1912 vfio_put_address_space(space
);
1917 static void vfio_disconnect_container(VFIOGroup
*group
)
1919 VFIOContainer
*container
= group
->container
;
1921 QLIST_REMOVE(group
, container_next
);
1922 group
->container
= NULL
;
1925 * Explicitly release the listener first before unset container,
1926 * since unset may destroy the backend container if it's the last
1929 if (QLIST_EMPTY(&container
->group_list
)) {
1930 vfio_listener_release(container
);
1933 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1934 error_report("vfio: error disconnecting group %d from container",
1938 if (QLIST_EMPTY(&container
->group_list
)) {
1939 VFIOAddressSpace
*space
= container
->space
;
1940 VFIOGuestIOMMU
*giommu
, *tmp
;
1942 QLIST_REMOVE(container
, next
);
1944 QLIST_FOREACH_SAFE(giommu
, &container
->giommu_list
, giommu_next
, tmp
) {
1945 memory_region_unregister_iommu_notifier(
1946 MEMORY_REGION(giommu
->iommu
), &giommu
->n
);
1947 QLIST_REMOVE(giommu
, giommu_next
);
1951 trace_vfio_disconnect_container(container
->fd
);
1952 close(container
->fd
);
1955 vfio_put_address_space(space
);
1959 VFIOGroup
*vfio_get_group(int groupid
, AddressSpace
*as
, Error
**errp
)
1963 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1965 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1966 if (group
->groupid
== groupid
) {
1967 /* Found it. Now is it already in the right context? */
1968 if (group
->container
->space
->as
== as
) {
1971 error_setg(errp
, "group %d used in multiple address spaces",
1978 group
= g_malloc0(sizeof(*group
));
1980 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1981 group
->fd
= qemu_open_old(path
, O_RDWR
);
1982 if (group
->fd
< 0) {
1983 error_setg_errno(errp
, errno
, "failed to open %s", path
);
1984 goto free_group_exit
;
1987 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1988 error_setg_errno(errp
, errno
, "failed to get group %d status", groupid
);
1992 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1993 error_setg(errp
, "group %d is not viable", groupid
);
1994 error_append_hint(errp
,
1995 "Please ensure all devices within the iommu_group "
1996 "are bound to their vfio bus driver.\n");
2000 group
->groupid
= groupid
;
2001 QLIST_INIT(&group
->device_list
);
2003 if (vfio_connect_container(group
, as
, errp
)) {
2004 error_prepend(errp
, "failed to setup container for group %d: ",
2009 if (QLIST_EMPTY(&vfio_group_list
)) {
2010 qemu_register_reset(vfio_reset_handler
, NULL
);
2013 QLIST_INSERT_HEAD(&vfio_group_list
, group
, next
);
2026 void vfio_put_group(VFIOGroup
*group
)
2028 if (!group
|| !QLIST_EMPTY(&group
->device_list
)) {
2032 if (!group
->ram_block_discard_allowed
) {
2033 ram_block_discard_disable(false);
2035 vfio_kvm_device_del_group(group
);
2036 vfio_disconnect_container(group
);
2037 QLIST_REMOVE(group
, next
);
2038 trace_vfio_put_group(group
->fd
);
2042 if (QLIST_EMPTY(&vfio_group_list
)) {
2043 qemu_unregister_reset(vfio_reset_handler
, NULL
);
2047 int vfio_get_device(VFIOGroup
*group
, const char *name
,
2048 VFIODevice
*vbasedev
, Error
**errp
)
2050 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
2053 fd
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
2055 error_setg_errno(errp
, errno
, "error getting device from group %d",
2057 error_append_hint(errp
,
2058 "Verify all devices in group %d are bound to vfio-<bus> "
2059 "or pci-stub and not already in use\n", group
->groupid
);
2063 ret
= ioctl(fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
2065 error_setg_errno(errp
, errno
, "error getting device info");
2071 * Set discarding of RAM as not broken for this group if the driver knows
2072 * the device operates compatibly with discarding. Setting must be
2073 * consistent per group, but since compatibility is really only possible
2074 * with mdev currently, we expect singleton groups.
2076 if (vbasedev
->ram_block_discard_allowed
!=
2077 group
->ram_block_discard_allowed
) {
2078 if (!QLIST_EMPTY(&group
->device_list
)) {
2079 error_setg(errp
, "Inconsistent setting of support for discarding "
2080 "RAM (e.g., balloon) within group");
2085 if (!group
->ram_block_discard_allowed
) {
2086 group
->ram_block_discard_allowed
= true;
2087 ram_block_discard_disable(false);
2092 vbasedev
->group
= group
;
2093 QLIST_INSERT_HEAD(&group
->device_list
, vbasedev
, next
);
2095 vbasedev
->num_irqs
= dev_info
.num_irqs
;
2096 vbasedev
->num_regions
= dev_info
.num_regions
;
2097 vbasedev
->flags
= dev_info
.flags
;
2099 trace_vfio_get_device(name
, dev_info
.flags
, dev_info
.num_regions
,
2102 vbasedev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
2106 void vfio_put_base_device(VFIODevice
*vbasedev
)
2108 if (!vbasedev
->group
) {
2111 QLIST_REMOVE(vbasedev
, next
);
2112 vbasedev
->group
= NULL
;
2113 trace_vfio_put_base_device(vbasedev
->fd
);
2114 close(vbasedev
->fd
);
2117 int vfio_get_region_info(VFIODevice
*vbasedev
, int index
,
2118 struct vfio_region_info
**info
)
2120 size_t argsz
= sizeof(struct vfio_region_info
);
2122 *info
= g_malloc0(argsz
);
2124 (*info
)->index
= index
;
2126 (*info
)->argsz
= argsz
;
2128 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, *info
)) {
2134 if ((*info
)->argsz
> argsz
) {
2135 argsz
= (*info
)->argsz
;
2136 *info
= g_realloc(*info
, argsz
);
2144 int vfio_get_dev_region_info(VFIODevice
*vbasedev
, uint32_t type
,
2145 uint32_t subtype
, struct vfio_region_info
**info
)
2149 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
2150 struct vfio_info_cap_header
*hdr
;
2151 struct vfio_region_info_cap_type
*cap_type
;
2153 if (vfio_get_region_info(vbasedev
, i
, info
)) {
2157 hdr
= vfio_get_region_info_cap(*info
, VFIO_REGION_INFO_CAP_TYPE
);
2163 cap_type
= container_of(hdr
, struct vfio_region_info_cap_type
, header
);
2165 trace_vfio_get_dev_region(vbasedev
->name
, i
,
2166 cap_type
->type
, cap_type
->subtype
);
2168 if (cap_type
->type
== type
&& cap_type
->subtype
== subtype
) {
2179 bool vfio_has_region_cap(VFIODevice
*vbasedev
, int region
, uint16_t cap_type
)
2181 struct vfio_region_info
*info
= NULL
;
2184 if (!vfio_get_region_info(vbasedev
, region
, &info
)) {
2185 if (vfio_get_region_info_cap(info
, cap_type
)) {
2195 * Interfaces for IBM EEH (Enhanced Error Handling)
2197 static bool vfio_eeh_container_ok(VFIOContainer
*container
)
2200 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2201 * implementation is broken if there are multiple groups in a
2202 * container. The hardware works in units of Partitionable
2203 * Endpoints (== IOMMU groups) and the EEH operations naively
2204 * iterate across all groups in the container, without any logic
2205 * to make sure the groups have their state synchronized. For
2206 * certain operations (ENABLE) that might be ok, until an error
2207 * occurs, but for others (GET_STATE) it's clearly broken.
2211 * XXX Once fixed kernels exist, test for them here
2214 if (QLIST_EMPTY(&container
->group_list
)) {
2218 if (QLIST_NEXT(QLIST_FIRST(&container
->group_list
), container_next
)) {
2225 static int vfio_eeh_container_op(VFIOContainer
*container
, uint32_t op
)
2227 struct vfio_eeh_pe_op pe_op
= {
2228 .argsz
= sizeof(pe_op
),
2233 if (!vfio_eeh_container_ok(container
)) {
2234 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
2235 "kernel requires a container with exactly one group", op
);
2239 ret
= ioctl(container
->fd
, VFIO_EEH_PE_OP
, &pe_op
);
2241 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op
);
2248 static VFIOContainer
*vfio_eeh_as_container(AddressSpace
*as
)
2250 VFIOAddressSpace
*space
= vfio_get_address_space(as
);
2251 VFIOContainer
*container
= NULL
;
2253 if (QLIST_EMPTY(&space
->containers
)) {
2254 /* No containers to act on */
2258 container
= QLIST_FIRST(&space
->containers
);
2260 if (QLIST_NEXT(container
, next
)) {
2261 /* We don't yet have logic to synchronize EEH state across
2262 * multiple containers */
2268 vfio_put_address_space(space
);
2272 bool vfio_eeh_as_ok(AddressSpace
*as
)
2274 VFIOContainer
*container
= vfio_eeh_as_container(as
);
2276 return (container
!= NULL
) && vfio_eeh_container_ok(container
);
2279 int vfio_eeh_as_op(AddressSpace
*as
, uint32_t op
)
2281 VFIOContainer
*container
= vfio_eeh_as_container(as
);
2286 return vfio_eeh_container_op(container
, op
);