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_stopped_and_saving(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 ((migration
->device_state
& VFIO_DEVICE_STATE_SAVING
) &&
333 !(migration
->device_state
& VFIO_DEVICE_STATE_RUNNING
)) {
343 static bool vfio_devices_all_running_and_saving(VFIOContainer
*container
)
346 VFIODevice
*vbasedev
;
347 MigrationState
*ms
= migrate_get_current();
349 if (!migration_is_setup_or_active(ms
->state
)) {
353 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
354 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
355 VFIOMigration
*migration
= vbasedev
->migration
;
361 if ((migration
->device_state
& VFIO_DEVICE_STATE_SAVING
) &&
362 (migration
->device_state
& VFIO_DEVICE_STATE_RUNNING
)) {
372 static int vfio_dma_unmap_bitmap(VFIOContainer
*container
,
373 hwaddr iova
, ram_addr_t size
,
374 IOMMUTLBEntry
*iotlb
)
376 struct vfio_iommu_type1_dma_unmap
*unmap
;
377 struct vfio_bitmap
*bitmap
;
378 uint64_t pages
= TARGET_PAGE_ALIGN(size
) >> TARGET_PAGE_BITS
;
381 unmap
= g_malloc0(sizeof(*unmap
) + sizeof(*bitmap
));
383 unmap
->argsz
= sizeof(*unmap
) + sizeof(*bitmap
);
386 unmap
->flags
|= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP
;
387 bitmap
= (struct vfio_bitmap
*)&unmap
->data
;
390 * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
391 * TARGET_PAGE_SIZE to mark those dirty. Hence set bitmap_pgsize to
395 bitmap
->pgsize
= TARGET_PAGE_SIZE
;
396 bitmap
->size
= ROUND_UP(pages
, sizeof(__u64
) * BITS_PER_BYTE
) /
399 if (bitmap
->size
> container
->max_dirty_bitmap_size
) {
400 error_report("UNMAP: Size of bitmap too big 0x%"PRIx64
,
401 (uint64_t)bitmap
->size
);
406 bitmap
->data
= g_try_malloc0(bitmap
->size
);
412 ret
= ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, unmap
);
414 cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap
->data
,
415 iotlb
->translated_addr
, pages
);
417 error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
420 g_free(bitmap
->data
);
427 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
429 static int vfio_dma_unmap(VFIOContainer
*container
,
430 hwaddr iova
, ram_addr_t size
,
431 IOMMUTLBEntry
*iotlb
)
433 struct vfio_iommu_type1_dma_unmap unmap
= {
434 .argsz
= sizeof(unmap
),
440 if (iotlb
&& container
->dirty_pages_supported
&&
441 vfio_devices_all_running_and_saving(container
)) {
442 return vfio_dma_unmap_bitmap(container
, iova
, size
, iotlb
);
445 while (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
447 * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
448 * v4.15) where an overflow in its wrap-around check prevents us from
449 * unmapping the last page of the address space. Test for the error
450 * condition and re-try the unmap excluding the last page. The
451 * expectation is that we've never mapped the last page anyway and this
452 * unmap request comes via vIOMMU support which also makes it unlikely
453 * that this page is used. This bug was introduced well after type1 v2
454 * support was introduced, so we shouldn't need to test for v1. A fix
455 * is queued for kernel v5.0 so this workaround can be removed once
456 * affected kernels are sufficiently deprecated.
458 if (errno
== EINVAL
&& unmap
.size
&& !(unmap
.iova
+ unmap
.size
) &&
459 container
->iommu_type
== VFIO_TYPE1v2_IOMMU
) {
460 trace_vfio_dma_unmap_overflow_workaround();
461 unmap
.size
-= 1ULL << ctz64(container
->pgsizes
);
464 error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno
));
471 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
472 ram_addr_t size
, void *vaddr
, bool readonly
)
474 struct vfio_iommu_type1_dma_map map
= {
475 .argsz
= sizeof(map
),
476 .flags
= VFIO_DMA_MAP_FLAG_READ
,
477 .vaddr
= (__u64
)(uintptr_t)vaddr
,
483 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
487 * Try the mapping, if it fails with EBUSY, unmap the region and try
488 * again. This shouldn't be necessary, but we sometimes see it in
491 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
492 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
, NULL
) == 0 &&
493 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
497 error_report("VFIO_MAP_DMA failed: %s", strerror(errno
));
501 static void vfio_host_win_add(VFIOContainer
*container
,
502 hwaddr min_iova
, hwaddr max_iova
,
503 uint64_t iova_pgsizes
)
505 VFIOHostDMAWindow
*hostwin
;
507 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
508 if (ranges_overlap(hostwin
->min_iova
,
509 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
511 max_iova
- min_iova
+ 1)) {
512 hw_error("%s: Overlapped IOMMU are not enabled", __func__
);
516 hostwin
= g_malloc0(sizeof(*hostwin
));
518 hostwin
->min_iova
= min_iova
;
519 hostwin
->max_iova
= max_iova
;
520 hostwin
->iova_pgsizes
= iova_pgsizes
;
521 QLIST_INSERT_HEAD(&container
->hostwin_list
, hostwin
, hostwin_next
);
524 static int vfio_host_win_del(VFIOContainer
*container
, hwaddr min_iova
,
527 VFIOHostDMAWindow
*hostwin
;
529 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
530 if (hostwin
->min_iova
== min_iova
&& hostwin
->max_iova
== max_iova
) {
531 QLIST_REMOVE(hostwin
, hostwin_next
);
539 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
541 return (!memory_region_is_ram(section
->mr
) &&
542 !memory_region_is_iommu(section
->mr
)) ||
544 * Sizing an enabled 64-bit BAR can cause spurious mappings to
545 * addresses in the upper part of the 64-bit address space. These
546 * are never accessed by the CPU and beyond the address width of
547 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
549 section
->offset_within_address_space
& (1ULL << 63);
552 /* Called with rcu_read_lock held. */
553 static bool vfio_get_xlat_addr(IOMMUTLBEntry
*iotlb
, void **vaddr
,
554 ram_addr_t
*ram_addr
, bool *read_only
)
558 hwaddr len
= iotlb
->addr_mask
+ 1;
559 bool writable
= iotlb
->perm
& IOMMU_WO
;
562 * The IOMMU TLB entry we have just covers translation through
563 * this IOMMU to its immediate target. We need to translate
564 * it the rest of the way through to memory.
566 mr
= address_space_translate(&address_space_memory
,
567 iotlb
->translated_addr
,
568 &xlat
, &len
, writable
,
569 MEMTXATTRS_UNSPECIFIED
);
570 if (!memory_region_is_ram(mr
)) {
571 error_report("iommu map to non memory area %"HWADDR_PRIx
"",
577 * Translation truncates length to the IOMMU page size,
578 * check that it did not truncate too much.
580 if (len
& iotlb
->addr_mask
) {
581 error_report("iommu has granularity incompatible with target AS");
586 *vaddr
= memory_region_get_ram_ptr(mr
) + xlat
;
590 *ram_addr
= memory_region_get_ram_addr(mr
) + xlat
;
594 *read_only
= !writable
|| mr
->readonly
;
600 static void vfio_iommu_map_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
602 VFIOGuestIOMMU
*giommu
= container_of(n
, VFIOGuestIOMMU
, n
);
603 VFIOContainer
*container
= giommu
->container
;
604 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
608 trace_vfio_iommu_map_notify(iotlb
->perm
== IOMMU_NONE
? "UNMAP" : "MAP",
609 iova
, iova
+ iotlb
->addr_mask
);
611 if (iotlb
->target_as
!= &address_space_memory
) {
612 error_report("Wrong target AS \"%s\", only system memory is allowed",
613 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
619 if ((iotlb
->perm
& IOMMU_RW
) != IOMMU_NONE
) {
622 if (!vfio_get_xlat_addr(iotlb
, &vaddr
, NULL
, &read_only
)) {
626 * vaddr is only valid until rcu_read_unlock(). But after
627 * vfio_dma_map has set up the mapping the pages will be
628 * pinned by the kernel. This makes sure that the RAM backend
629 * of vaddr will always be there, even if the memory object is
630 * destroyed and its backing memory munmap-ed.
632 ret
= vfio_dma_map(container
, iova
,
633 iotlb
->addr_mask
+ 1, vaddr
,
636 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
637 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
639 iotlb
->addr_mask
+ 1, vaddr
, ret
);
642 ret
= vfio_dma_unmap(container
, iova
, iotlb
->addr_mask
+ 1, iotlb
);
644 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
645 "0x%"HWADDR_PRIx
") = %d (%m)",
647 iotlb
->addr_mask
+ 1, ret
);
654 static void vfio_listener_region_add(MemoryListener
*listener
,
655 MemoryRegionSection
*section
)
657 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
659 Int128 llend
, llsize
;
662 VFIOHostDMAWindow
*hostwin
;
666 if (vfio_listener_skipped_section(section
)) {
667 trace_vfio_listener_region_add_skip(
668 section
->offset_within_address_space
,
669 section
->offset_within_address_space
+
670 int128_get64(int128_sub(section
->size
, int128_one())));
674 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
675 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
676 error_report("%s received unaligned region", __func__
);
680 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
681 llend
= int128_make64(section
->offset_within_address_space
);
682 llend
= int128_add(llend
, section
->size
);
683 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
685 if (int128_ge(int128_make64(iova
), llend
)) {
688 end
= int128_get64(int128_sub(llend
, int128_one()));
690 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
693 /* For now intersections are not allowed, we may relax this later */
694 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
695 if (ranges_overlap(hostwin
->min_iova
,
696 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
697 section
->offset_within_address_space
,
698 int128_get64(section
->size
))) {
700 "region [0x%"PRIx64
",0x%"PRIx64
"] overlaps with existing"
701 "host DMA window [0x%"PRIx64
",0x%"PRIx64
"]",
702 section
->offset_within_address_space
,
703 section
->offset_within_address_space
+
704 int128_get64(section
->size
) - 1,
705 hostwin
->min_iova
, hostwin
->max_iova
);
710 ret
= vfio_spapr_create_window(container
, section
, &pgsize
);
712 error_setg_errno(&err
, -ret
, "Failed to create SPAPR window");
716 vfio_host_win_add(container
, section
->offset_within_address_space
,
717 section
->offset_within_address_space
+
718 int128_get64(section
->size
) - 1, pgsize
);
722 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
723 struct kvm_vfio_spapr_tce param
;
724 struct kvm_device_attr attr
= {
725 .group
= KVM_DEV_VFIO_GROUP
,
726 .attr
= KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE
,
727 .addr
= (uint64_t)(unsigned long)¶m
,
730 if (!memory_region_iommu_get_attr(iommu_mr
, IOMMU_ATTR_SPAPR_TCE_FD
,
732 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
733 param
.groupfd
= group
->fd
;
734 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
735 error_report("vfio: failed to setup fd %d "
736 "for a group with fd %d: %s",
737 param
.tablefd
, param
.groupfd
,
741 trace_vfio_spapr_group_attach(param
.groupfd
, param
.tablefd
);
748 hostwin_found
= false;
749 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
750 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
751 hostwin_found
= true;
756 if (!hostwin_found
) {
757 error_setg(&err
, "Container %p can't map guest IOVA region"
758 " 0x%"HWADDR_PRIx
"..0x%"HWADDR_PRIx
, container
, iova
, end
);
762 memory_region_ref(section
->mr
);
764 if (memory_region_is_iommu(section
->mr
)) {
765 VFIOGuestIOMMU
*giommu
;
766 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
769 trace_vfio_listener_region_add_iommu(iova
, end
);
771 * FIXME: For VFIO iommu types which have KVM acceleration to
772 * avoid bouncing all map/unmaps through qemu this way, this
773 * would be the right place to wire that up (tell the KVM
774 * device emulation the VFIO iommu handles to use).
776 giommu
= g_malloc0(sizeof(*giommu
));
777 giommu
->iommu
= iommu_mr
;
778 giommu
->iommu_offset
= section
->offset_within_address_space
-
779 section
->offset_within_region
;
780 giommu
->container
= container
;
781 llend
= int128_add(int128_make64(section
->offset_within_region
),
783 llend
= int128_sub(llend
, int128_one());
784 iommu_idx
= memory_region_iommu_attrs_to_index(iommu_mr
,
785 MEMTXATTRS_UNSPECIFIED
);
786 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
788 section
->offset_within_region
,
792 ret
= memory_region_register_iommu_notifier(section
->mr
, &giommu
->n
,
798 QLIST_INSERT_HEAD(&container
->giommu_list
, giommu
, giommu_next
);
799 memory_region_iommu_replay(giommu
->iommu
, &giommu
->n
);
804 /* Here we assume that memory_region_is_ram(section->mr)==true */
806 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
807 section
->offset_within_region
+
808 (iova
- section
->offset_within_address_space
);
810 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
812 llsize
= int128_sub(llend
, int128_make64(iova
));
814 if (memory_region_is_ram_device(section
->mr
)) {
815 hwaddr pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
817 if ((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
)) {
818 trace_vfio_listener_region_add_no_dma_map(
819 memory_region_name(section
->mr
),
820 section
->offset_within_address_space
,
821 int128_getlo(section
->size
),
827 ret
= vfio_dma_map(container
, iova
, int128_get64(llsize
),
828 vaddr
, section
->readonly
);
830 error_setg(&err
, "vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
831 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
832 container
, iova
, int128_get64(llsize
), vaddr
, ret
);
833 if (memory_region_is_ram_device(section
->mr
)) {
834 /* Allow unexpected mappings not to be fatal for RAM devices */
835 error_report_err(err
);
844 if (memory_region_is_ram_device(section
->mr
)) {
845 error_report("failed to vfio_dma_map. pci p2p may not work");
849 * On the initfn path, store the first error in the container so we
850 * can gracefully fail. Runtime, there's not much we can do other
851 * than throw a hardware error.
853 if (!container
->initialized
) {
854 if (!container
->error
) {
855 error_propagate_prepend(&container
->error
, err
,
857 memory_region_name(section
->mr
));
862 error_report_err(err
);
863 hw_error("vfio: DMA mapping failed, unable to continue");
867 static void vfio_listener_region_del(MemoryListener
*listener
,
868 MemoryRegionSection
*section
)
870 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
872 Int128 llend
, llsize
;
874 bool try_unmap
= true;
876 if (vfio_listener_skipped_section(section
)) {
877 trace_vfio_listener_region_del_skip(
878 section
->offset_within_address_space
,
879 section
->offset_within_address_space
+
880 int128_get64(int128_sub(section
->size
, int128_one())));
884 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
885 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
886 error_report("%s received unaligned region", __func__
);
890 if (memory_region_is_iommu(section
->mr
)) {
891 VFIOGuestIOMMU
*giommu
;
893 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
894 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
895 giommu
->n
.start
== section
->offset_within_region
) {
896 memory_region_unregister_iommu_notifier(section
->mr
,
898 QLIST_REMOVE(giommu
, giommu_next
);
905 * FIXME: We assume the one big unmap below is adequate to
906 * remove any individual page mappings in the IOMMU which
907 * might have been copied into VFIO. This works for a page table
908 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
909 * That may not be true for all IOMMU types.
913 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
914 llend
= int128_make64(section
->offset_within_address_space
);
915 llend
= int128_add(llend
, section
->size
);
916 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
918 if (int128_ge(int128_make64(iova
), llend
)) {
921 end
= int128_get64(int128_sub(llend
, int128_one()));
923 llsize
= int128_sub(llend
, int128_make64(iova
));
925 trace_vfio_listener_region_del(iova
, end
);
927 if (memory_region_is_ram_device(section
->mr
)) {
929 VFIOHostDMAWindow
*hostwin
;
930 bool hostwin_found
= false;
932 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
933 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
934 hostwin_found
= true;
938 assert(hostwin_found
); /* or region_add() would have failed */
940 pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
941 try_unmap
= !((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
));
945 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
), NULL
);
947 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
948 "0x%"HWADDR_PRIx
") = %d (%m)",
949 container
, iova
, int128_get64(llsize
), ret
);
953 memory_region_unref(section
->mr
);
955 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
956 vfio_spapr_remove_window(container
,
957 section
->offset_within_address_space
);
958 if (vfio_host_win_del(container
,
959 section
->offset_within_address_space
,
960 section
->offset_within_address_space
+
961 int128_get64(section
->size
) - 1) < 0) {
962 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx
,
963 __func__
, section
->offset_within_address_space
);
968 static int vfio_get_dirty_bitmap(VFIOContainer
*container
, uint64_t iova
,
969 uint64_t size
, ram_addr_t ram_addr
)
971 struct vfio_iommu_type1_dirty_bitmap
*dbitmap
;
972 struct vfio_iommu_type1_dirty_bitmap_get
*range
;
976 dbitmap
= g_malloc0(sizeof(*dbitmap
) + sizeof(*range
));
978 dbitmap
->argsz
= sizeof(*dbitmap
) + sizeof(*range
);
979 dbitmap
->flags
= VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP
;
980 range
= (struct vfio_iommu_type1_dirty_bitmap_get
*)&dbitmap
->data
;
985 * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
986 * TARGET_PAGE_SIZE to mark those dirty. Hence set bitmap's pgsize to
989 range
->bitmap
.pgsize
= TARGET_PAGE_SIZE
;
991 pages
= TARGET_PAGE_ALIGN(range
->size
) >> TARGET_PAGE_BITS
;
992 range
->bitmap
.size
= ROUND_UP(pages
, sizeof(__u64
) * BITS_PER_BYTE
) /
994 range
->bitmap
.data
= g_try_malloc0(range
->bitmap
.size
);
995 if (!range
->bitmap
.data
) {
1000 ret
= ioctl(container
->fd
, VFIO_IOMMU_DIRTY_PAGES
, dbitmap
);
1002 error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1003 " size: 0x%"PRIx64
" err: %d", (uint64_t)range
->iova
,
1004 (uint64_t)range
->size
, errno
);
1008 cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range
->bitmap
.data
,
1011 trace_vfio_get_dirty_bitmap(container
->fd
, range
->iova
, range
->size
,
1012 range
->bitmap
.size
, ram_addr
);
1014 g_free(range
->bitmap
.data
);
1022 VFIOGuestIOMMU
*giommu
;
1023 } vfio_giommu_dirty_notifier
;
1025 static void vfio_iommu_map_dirty_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
1027 vfio_giommu_dirty_notifier
*gdn
= container_of(n
,
1028 vfio_giommu_dirty_notifier
, n
);
1029 VFIOGuestIOMMU
*giommu
= gdn
->giommu
;
1030 VFIOContainer
*container
= giommu
->container
;
1031 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
1032 ram_addr_t translated_addr
;
1034 trace_vfio_iommu_map_dirty_notify(iova
, iova
+ iotlb
->addr_mask
);
1036 if (iotlb
->target_as
!= &address_space_memory
) {
1037 error_report("Wrong target AS \"%s\", only system memory is allowed",
1038 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
1043 if (vfio_get_xlat_addr(iotlb
, NULL
, &translated_addr
, NULL
)) {
1046 ret
= vfio_get_dirty_bitmap(container
, iova
, iotlb
->addr_mask
+ 1,
1049 error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx
", "
1050 "0x%"HWADDR_PRIx
") = %d (%m)",
1052 iotlb
->addr_mask
+ 1, ret
);
1058 static int vfio_sync_dirty_bitmap(VFIOContainer
*container
,
1059 MemoryRegionSection
*section
)
1061 ram_addr_t ram_addr
;
1063 if (memory_region_is_iommu(section
->mr
)) {
1064 VFIOGuestIOMMU
*giommu
;
1066 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
1067 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
1068 giommu
->n
.start
== section
->offset_within_region
) {
1070 vfio_giommu_dirty_notifier gdn
= { .giommu
= giommu
};
1071 int idx
= memory_region_iommu_attrs_to_index(giommu
->iommu
,
1072 MEMTXATTRS_UNSPECIFIED
);
1074 llend
= int128_add(int128_make64(section
->offset_within_region
),
1076 llend
= int128_sub(llend
, int128_one());
1078 iommu_notifier_init(&gdn
.n
,
1079 vfio_iommu_map_dirty_notify
,
1081 section
->offset_within_region
,
1082 int128_get64(llend
),
1084 memory_region_iommu_replay(giommu
->iommu
, &gdn
.n
);
1091 ram_addr
= memory_region_get_ram_addr(section
->mr
) +
1092 section
->offset_within_region
;
1094 return vfio_get_dirty_bitmap(container
,
1095 TARGET_PAGE_ALIGN(section
->offset_within_address_space
),
1096 int128_get64(section
->size
), ram_addr
);
1099 static void vfio_listerner_log_sync(MemoryListener
*listener
,
1100 MemoryRegionSection
*section
)
1102 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
1104 if (vfio_listener_skipped_section(section
) ||
1105 !container
->dirty_pages_supported
) {
1109 if (vfio_devices_all_stopped_and_saving(container
)) {
1110 vfio_sync_dirty_bitmap(container
, section
);
1114 static const MemoryListener vfio_memory_listener
= {
1115 .region_add
= vfio_listener_region_add
,
1116 .region_del
= vfio_listener_region_del
,
1117 .log_sync
= vfio_listerner_log_sync
,
1120 static void vfio_listener_release(VFIOContainer
*container
)
1122 memory_listener_unregister(&container
->listener
);
1123 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
1124 memory_listener_unregister(&container
->prereg_listener
);
1128 static struct vfio_info_cap_header
*
1129 vfio_get_cap(void *ptr
, uint32_t cap_offset
, uint16_t id
)
1131 struct vfio_info_cap_header
*hdr
;
1133 for (hdr
= ptr
+ cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
1134 if (hdr
->id
== id
) {
1142 struct vfio_info_cap_header
*
1143 vfio_get_region_info_cap(struct vfio_region_info
*info
, uint16_t id
)
1145 if (!(info
->flags
& VFIO_REGION_INFO_FLAG_CAPS
)) {
1149 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1152 static struct vfio_info_cap_header
*
1153 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info
*info
, uint16_t id
)
1155 if (!(info
->flags
& VFIO_IOMMU_INFO_CAPS
)) {
1159 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1162 struct vfio_info_cap_header
*
1163 vfio_get_device_info_cap(struct vfio_device_info
*info
, uint16_t id
)
1165 if (!(info
->flags
& VFIO_DEVICE_FLAGS_CAPS
)) {
1169 return vfio_get_cap((void *)info
, info
->cap_offset
, id
);
1172 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info
*info
,
1173 unsigned int *avail
)
1175 struct vfio_info_cap_header
*hdr
;
1176 struct vfio_iommu_type1_info_dma_avail
*cap
;
1178 /* If the capability cannot be found, assume no DMA limiting */
1179 hdr
= vfio_get_iommu_type1_info_cap(info
,
1180 VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL
);
1185 if (avail
!= NULL
) {
1187 *avail
= cap
->avail
;
1193 static int vfio_setup_region_sparse_mmaps(VFIORegion
*region
,
1194 struct vfio_region_info
*info
)
1196 struct vfio_info_cap_header
*hdr
;
1197 struct vfio_region_info_cap_sparse_mmap
*sparse
;
1200 hdr
= vfio_get_region_info_cap(info
, VFIO_REGION_INFO_CAP_SPARSE_MMAP
);
1205 sparse
= container_of(hdr
, struct vfio_region_info_cap_sparse_mmap
, header
);
1207 trace_vfio_region_sparse_mmap_header(region
->vbasedev
->name
,
1208 region
->nr
, sparse
->nr_areas
);
1210 region
->mmaps
= g_new0(VFIOMmap
, sparse
->nr_areas
);
1212 for (i
= 0, j
= 0; i
< sparse
->nr_areas
; i
++) {
1213 trace_vfio_region_sparse_mmap_entry(i
, sparse
->areas
[i
].offset
,
1214 sparse
->areas
[i
].offset
+
1215 sparse
->areas
[i
].size
);
1217 if (sparse
->areas
[i
].size
) {
1218 region
->mmaps
[j
].offset
= sparse
->areas
[i
].offset
;
1219 region
->mmaps
[j
].size
= sparse
->areas
[i
].size
;
1224 region
->nr_mmaps
= j
;
1225 region
->mmaps
= g_realloc(region
->mmaps
, j
* sizeof(VFIOMmap
));
1230 int vfio_region_setup(Object
*obj
, VFIODevice
*vbasedev
, VFIORegion
*region
,
1231 int index
, const char *name
)
1233 struct vfio_region_info
*info
;
1236 ret
= vfio_get_region_info(vbasedev
, index
, &info
);
1241 region
->vbasedev
= vbasedev
;
1242 region
->flags
= info
->flags
;
1243 region
->size
= info
->size
;
1244 region
->fd_offset
= info
->offset
;
1248 region
->mem
= g_new0(MemoryRegion
, 1);
1249 memory_region_init_io(region
->mem
, obj
, &vfio_region_ops
,
1250 region
, name
, region
->size
);
1252 if (!vbasedev
->no_mmap
&&
1253 region
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
1255 ret
= vfio_setup_region_sparse_mmaps(region
, info
);
1258 region
->nr_mmaps
= 1;
1259 region
->mmaps
= g_new0(VFIOMmap
, region
->nr_mmaps
);
1260 region
->mmaps
[0].offset
= 0;
1261 region
->mmaps
[0].size
= region
->size
;
1268 trace_vfio_region_setup(vbasedev
->name
, index
, name
,
1269 region
->flags
, region
->fd_offset
, region
->size
);
1273 static void vfio_subregion_unmap(VFIORegion
*region
, int index
)
1275 trace_vfio_region_unmap(memory_region_name(®ion
->mmaps
[index
].mem
),
1276 region
->mmaps
[index
].offset
,
1277 region
->mmaps
[index
].offset
+
1278 region
->mmaps
[index
].size
- 1);
1279 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[index
].mem
);
1280 munmap(region
->mmaps
[index
].mmap
, region
->mmaps
[index
].size
);
1281 object_unparent(OBJECT(®ion
->mmaps
[index
].mem
));
1282 region
->mmaps
[index
].mmap
= NULL
;
1285 int vfio_region_mmap(VFIORegion
*region
)
1294 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_READ
? PROT_READ
: 0;
1295 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_WRITE
? PROT_WRITE
: 0;
1297 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1298 region
->mmaps
[i
].mmap
= mmap(NULL
, region
->mmaps
[i
].size
, prot
,
1299 MAP_SHARED
, region
->vbasedev
->fd
,
1301 region
->mmaps
[i
].offset
);
1302 if (region
->mmaps
[i
].mmap
== MAP_FAILED
) {
1305 trace_vfio_region_mmap_fault(memory_region_name(region
->mem
), i
,
1307 region
->mmaps
[i
].offset
,
1309 region
->mmaps
[i
].offset
+
1310 region
->mmaps
[i
].size
- 1, ret
);
1312 region
->mmaps
[i
].mmap
= NULL
;
1314 for (i
--; i
>= 0; i
--) {
1315 vfio_subregion_unmap(region
, i
);
1321 name
= g_strdup_printf("%s mmaps[%d]",
1322 memory_region_name(region
->mem
), i
);
1323 memory_region_init_ram_device_ptr(®ion
->mmaps
[i
].mem
,
1324 memory_region_owner(region
->mem
),
1325 name
, region
->mmaps
[i
].size
,
1326 region
->mmaps
[i
].mmap
);
1328 memory_region_add_subregion(region
->mem
, region
->mmaps
[i
].offset
,
1329 ®ion
->mmaps
[i
].mem
);
1331 trace_vfio_region_mmap(memory_region_name(®ion
->mmaps
[i
].mem
),
1332 region
->mmaps
[i
].offset
,
1333 region
->mmaps
[i
].offset
+
1334 region
->mmaps
[i
].size
- 1);
1340 void vfio_region_unmap(VFIORegion
*region
)
1348 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1349 if (region
->mmaps
[i
].mmap
) {
1350 vfio_subregion_unmap(region
, i
);
1355 void vfio_region_exit(VFIORegion
*region
)
1363 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1364 if (region
->mmaps
[i
].mmap
) {
1365 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
1369 trace_vfio_region_exit(region
->vbasedev
->name
, region
->nr
);
1372 void vfio_region_finalize(VFIORegion
*region
)
1380 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1381 if (region
->mmaps
[i
].mmap
) {
1382 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
1383 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
1387 object_unparent(OBJECT(region
->mem
));
1389 g_free(region
->mem
);
1390 g_free(region
->mmaps
);
1392 trace_vfio_region_finalize(region
->vbasedev
->name
, region
->nr
);
1395 region
->mmaps
= NULL
;
1396 region
->nr_mmaps
= 0;
1402 void vfio_region_mmaps_set_enabled(VFIORegion
*region
, bool enabled
)
1410 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1411 if (region
->mmaps
[i
].mmap
) {
1412 memory_region_set_enabled(®ion
->mmaps
[i
].mem
, enabled
);
1416 trace_vfio_region_mmaps_set_enabled(memory_region_name(region
->mem
),
1420 void vfio_reset_handler(void *opaque
)
1423 VFIODevice
*vbasedev
;
1425 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1426 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1427 if (vbasedev
->dev
->realized
) {
1428 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
1433 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1434 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1435 if (vbasedev
->dev
->realized
&& vbasedev
->needs_reset
) {
1436 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
1442 static void vfio_kvm_device_add_group(VFIOGroup
*group
)
1445 struct kvm_device_attr attr
= {
1446 .group
= KVM_DEV_VFIO_GROUP
,
1447 .attr
= KVM_DEV_VFIO_GROUP_ADD
,
1448 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1451 if (!kvm_enabled()) {
1455 if (vfio_kvm_device_fd
< 0) {
1456 struct kvm_create_device cd
= {
1457 .type
= KVM_DEV_TYPE_VFIO
,
1460 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
1461 error_report("Failed to create KVM VFIO device: %m");
1465 vfio_kvm_device_fd
= cd
.fd
;
1468 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1469 error_report("Failed to add group %d to KVM VFIO device: %m",
1475 static void vfio_kvm_device_del_group(VFIOGroup
*group
)
1478 struct kvm_device_attr attr
= {
1479 .group
= KVM_DEV_VFIO_GROUP
,
1480 .attr
= KVM_DEV_VFIO_GROUP_DEL
,
1481 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1484 if (vfio_kvm_device_fd
< 0) {
1488 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1489 error_report("Failed to remove group %d from KVM VFIO device: %m",
1495 static VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
1497 VFIOAddressSpace
*space
;
1499 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
1500 if (space
->as
== as
) {
1505 /* No suitable VFIOAddressSpace, create a new one */
1506 space
= g_malloc0(sizeof(*space
));
1508 QLIST_INIT(&space
->containers
);
1510 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
1515 static void vfio_put_address_space(VFIOAddressSpace
*space
)
1517 if (QLIST_EMPTY(&space
->containers
)) {
1518 QLIST_REMOVE(space
, list
);
1524 * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1526 static int vfio_get_iommu_type(VFIOContainer
*container
,
1529 int iommu_types
[] = { VFIO_TYPE1v2_IOMMU
, VFIO_TYPE1_IOMMU
,
1530 VFIO_SPAPR_TCE_v2_IOMMU
, VFIO_SPAPR_TCE_IOMMU
};
1533 for (i
= 0; i
< ARRAY_SIZE(iommu_types
); i
++) {
1534 if (ioctl(container
->fd
, VFIO_CHECK_EXTENSION
, iommu_types
[i
])) {
1535 return iommu_types
[i
];
1538 error_setg(errp
, "No available IOMMU models");
1542 static int vfio_init_container(VFIOContainer
*container
, int group_fd
,
1545 int iommu_type
, ret
;
1547 iommu_type
= vfio_get_iommu_type(container
, errp
);
1548 if (iommu_type
< 0) {
1552 ret
= ioctl(group_fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
);
1554 error_setg_errno(errp
, errno
, "Failed to set group container");
1558 while (ioctl(container
->fd
, VFIO_SET_IOMMU
, iommu_type
)) {
1559 if (iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
1561 * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1562 * v2, the running platform may not support v2 and there is no
1563 * way to guess it until an IOMMU group gets added to the container.
1564 * So in case it fails with v2, try v1 as a fallback.
1566 iommu_type
= VFIO_SPAPR_TCE_IOMMU
;
1569 error_setg_errno(errp
, errno
, "Failed to set iommu for container");
1573 container
->iommu_type
= iommu_type
;
1577 static int vfio_get_iommu_info(VFIOContainer
*container
,
1578 struct vfio_iommu_type1_info
**info
)
1581 size_t argsz
= sizeof(struct vfio_iommu_type1_info
);
1583 *info
= g_new0(struct vfio_iommu_type1_info
, 1);
1585 (*info
)->argsz
= argsz
;
1587 if (ioctl(container
->fd
, VFIO_IOMMU_GET_INFO
, *info
)) {
1593 if (((*info
)->argsz
> argsz
)) {
1594 argsz
= (*info
)->argsz
;
1595 *info
= g_realloc(*info
, argsz
);
1602 static struct vfio_info_cap_header
*
1603 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info
*info
, uint16_t id
)
1605 struct vfio_info_cap_header
*hdr
;
1608 if (!(info
->flags
& VFIO_IOMMU_INFO_CAPS
)) {
1612 for (hdr
= ptr
+ info
->cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
1613 if (hdr
->id
== id
) {
1621 static void vfio_get_iommu_info_migration(VFIOContainer
*container
,
1622 struct vfio_iommu_type1_info
*info
)
1624 struct vfio_info_cap_header
*hdr
;
1625 struct vfio_iommu_type1_info_cap_migration
*cap_mig
;
1627 hdr
= vfio_get_iommu_info_cap(info
, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION
);
1632 cap_mig
= container_of(hdr
, struct vfio_iommu_type1_info_cap_migration
,
1636 * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
1637 * TARGET_PAGE_SIZE to mark those dirty.
1639 if (cap_mig
->pgsize_bitmap
& TARGET_PAGE_SIZE
) {
1640 container
->dirty_pages_supported
= true;
1641 container
->max_dirty_bitmap_size
= cap_mig
->max_dirty_bitmap_size
;
1642 container
->dirty_pgsizes
= cap_mig
->pgsize_bitmap
;
1646 static int vfio_connect_container(VFIOGroup
*group
, AddressSpace
*as
,
1649 VFIOContainer
*container
;
1651 VFIOAddressSpace
*space
;
1653 space
= vfio_get_address_space(as
);
1656 * VFIO is currently incompatible with discarding of RAM insofar as the
1657 * madvise to purge (zap) the page from QEMU's address space does not
1658 * interact with the memory API and therefore leaves stale virtual to
1659 * physical mappings in the IOMMU if the page was previously pinned. We
1660 * therefore set discarding broken for each group added to a container,
1661 * whether the container is used individually or shared. This provides
1662 * us with options to allow devices within a group to opt-in and allow
1663 * discarding, so long as it is done consistently for a group (for instance
1664 * if the device is an mdev device where it is known that the host vendor
1665 * driver will never pin pages outside of the working set of the guest
1666 * driver, which would thus not be discarding candidates).
1668 * The first opportunity to induce pinning occurs here where we attempt to
1669 * attach the group to existing containers within the AddressSpace. If any
1670 * pages are already zapped from the virtual address space, such as from
1671 * previous discards, new pinning will cause valid mappings to be
1672 * re-established. Likewise, when the overall MemoryListener for a new
1673 * container is registered, a replay of mappings within the AddressSpace
1674 * will occur, re-establishing any previously zapped pages as well.
1676 * Especially virtio-balloon is currently only prevented from discarding
1677 * new memory, it will not yet set ram_block_discard_set_required() and
1678 * therefore, neither stops us here or deals with the sudden memory
1679 * consumption of inflated memory.
1681 ret
= ram_block_discard_disable(true);
1683 error_setg_errno(errp
, -ret
, "Cannot set discarding of RAM broken");
1687 QLIST_FOREACH(container
, &space
->containers
, next
) {
1688 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
1689 group
->container
= container
;
1690 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1691 vfio_kvm_device_add_group(group
);
1696 fd
= qemu_open_old("/dev/vfio/vfio", O_RDWR
);
1698 error_setg_errno(errp
, errno
, "failed to open /dev/vfio/vfio");
1700 goto put_space_exit
;
1703 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
1704 if (ret
!= VFIO_API_VERSION
) {
1705 error_setg(errp
, "supported vfio version: %d, "
1706 "reported version: %d", VFIO_API_VERSION
, ret
);
1711 container
= g_malloc0(sizeof(*container
));
1712 container
->space
= space
;
1714 container
->error
= NULL
;
1715 container
->dirty_pages_supported
= false;
1716 QLIST_INIT(&container
->giommu_list
);
1717 QLIST_INIT(&container
->hostwin_list
);
1719 ret
= vfio_init_container(container
, group
->fd
, errp
);
1721 goto free_container_exit
;
1724 switch (container
->iommu_type
) {
1725 case VFIO_TYPE1v2_IOMMU
:
1726 case VFIO_TYPE1_IOMMU
:
1728 struct vfio_iommu_type1_info
*info
;
1731 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1732 * IOVA whatsoever. That's not actually true, but the current
1733 * kernel interface doesn't tell us what it can map, and the
1734 * existing Type1 IOMMUs generally support any IOVA we're
1735 * going to actually try in practice.
1737 ret
= vfio_get_iommu_info(container
, &info
);
1739 if (ret
|| !(info
->flags
& VFIO_IOMMU_INFO_PGSIZES
)) {
1740 /* Assume 4k IOVA page size */
1741 info
->iova_pgsizes
= 4096;
1743 vfio_host_win_add(container
, 0, (hwaddr
)-1, info
->iova_pgsizes
);
1744 container
->pgsizes
= info
->iova_pgsizes
;
1747 vfio_get_iommu_info_migration(container
, info
);
1752 case VFIO_SPAPR_TCE_v2_IOMMU
:
1753 case VFIO_SPAPR_TCE_IOMMU
:
1755 struct vfio_iommu_spapr_tce_info info
;
1756 bool v2
= container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
;
1759 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1760 * when container fd is closed so we do not call it explicitly
1764 ret
= ioctl(fd
, VFIO_IOMMU_ENABLE
);
1766 error_setg_errno(errp
, errno
, "failed to enable container");
1768 goto free_container_exit
;
1771 container
->prereg_listener
= vfio_prereg_listener
;
1773 memory_listener_register(&container
->prereg_listener
,
1774 &address_space_memory
);
1775 if (container
->error
) {
1776 memory_listener_unregister(&container
->prereg_listener
);
1778 error_propagate_prepend(errp
, container
->error
,
1779 "RAM memory listener initialization failed: ");
1780 goto free_container_exit
;
1784 info
.argsz
= sizeof(info
);
1785 ret
= ioctl(fd
, VFIO_IOMMU_SPAPR_TCE_GET_INFO
, &info
);
1787 error_setg_errno(errp
, errno
,
1788 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1791 memory_listener_unregister(&container
->prereg_listener
);
1793 goto free_container_exit
;
1797 container
->pgsizes
= info
.ddw
.pgsizes
;
1799 * There is a default window in just created container.
1800 * To make region_add/del simpler, we better remove this
1801 * window now and let those iommu_listener callbacks
1802 * create/remove them when needed.
1804 ret
= vfio_spapr_remove_window(container
, info
.dma32_window_start
);
1806 error_setg_errno(errp
, -ret
,
1807 "failed to remove existing window");
1808 goto free_container_exit
;
1811 /* The default table uses 4K pages */
1812 container
->pgsizes
= 0x1000;
1813 vfio_host_win_add(container
, info
.dma32_window_start
,
1814 info
.dma32_window_start
+
1815 info
.dma32_window_size
- 1,
1821 vfio_kvm_device_add_group(group
);
1823 QLIST_INIT(&container
->group_list
);
1824 QLIST_INSERT_HEAD(&space
->containers
, container
, next
);
1826 group
->container
= container
;
1827 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1829 container
->listener
= vfio_memory_listener
;
1831 memory_listener_register(&container
->listener
, container
->space
->as
);
1833 if (container
->error
) {
1835 error_propagate_prepend(errp
, container
->error
,
1836 "memory listener initialization failed: ");
1837 goto listener_release_exit
;
1840 container
->initialized
= true;
1843 listener_release_exit
:
1844 QLIST_REMOVE(group
, container_next
);
1845 QLIST_REMOVE(container
, next
);
1846 vfio_kvm_device_del_group(group
);
1847 vfio_listener_release(container
);
1849 free_container_exit
:
1856 ram_block_discard_disable(false);
1857 vfio_put_address_space(space
);
1862 static void vfio_disconnect_container(VFIOGroup
*group
)
1864 VFIOContainer
*container
= group
->container
;
1866 QLIST_REMOVE(group
, container_next
);
1867 group
->container
= NULL
;
1870 * Explicitly release the listener first before unset container,
1871 * since unset may destroy the backend container if it's the last
1874 if (QLIST_EMPTY(&container
->group_list
)) {
1875 vfio_listener_release(container
);
1878 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1879 error_report("vfio: error disconnecting group %d from container",
1883 if (QLIST_EMPTY(&container
->group_list
)) {
1884 VFIOAddressSpace
*space
= container
->space
;
1885 VFIOGuestIOMMU
*giommu
, *tmp
;
1887 QLIST_REMOVE(container
, next
);
1889 QLIST_FOREACH_SAFE(giommu
, &container
->giommu_list
, giommu_next
, tmp
) {
1890 memory_region_unregister_iommu_notifier(
1891 MEMORY_REGION(giommu
->iommu
), &giommu
->n
);
1892 QLIST_REMOVE(giommu
, giommu_next
);
1896 trace_vfio_disconnect_container(container
->fd
);
1897 close(container
->fd
);
1900 vfio_put_address_space(space
);
1904 VFIOGroup
*vfio_get_group(int groupid
, AddressSpace
*as
, Error
**errp
)
1908 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1910 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1911 if (group
->groupid
== groupid
) {
1912 /* Found it. Now is it already in the right context? */
1913 if (group
->container
->space
->as
== as
) {
1916 error_setg(errp
, "group %d used in multiple address spaces",
1923 group
= g_malloc0(sizeof(*group
));
1925 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1926 group
->fd
= qemu_open_old(path
, O_RDWR
);
1927 if (group
->fd
< 0) {
1928 error_setg_errno(errp
, errno
, "failed to open %s", path
);
1929 goto free_group_exit
;
1932 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1933 error_setg_errno(errp
, errno
, "failed to get group %d status", groupid
);
1937 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1938 error_setg(errp
, "group %d is not viable", groupid
);
1939 error_append_hint(errp
,
1940 "Please ensure all devices within the iommu_group "
1941 "are bound to their vfio bus driver.\n");
1945 group
->groupid
= groupid
;
1946 QLIST_INIT(&group
->device_list
);
1948 if (vfio_connect_container(group
, as
, errp
)) {
1949 error_prepend(errp
, "failed to setup container for group %d: ",
1954 if (QLIST_EMPTY(&vfio_group_list
)) {
1955 qemu_register_reset(vfio_reset_handler
, NULL
);
1958 QLIST_INSERT_HEAD(&vfio_group_list
, group
, next
);
1971 void vfio_put_group(VFIOGroup
*group
)
1973 if (!group
|| !QLIST_EMPTY(&group
->device_list
)) {
1977 if (!group
->ram_block_discard_allowed
) {
1978 ram_block_discard_disable(false);
1980 vfio_kvm_device_del_group(group
);
1981 vfio_disconnect_container(group
);
1982 QLIST_REMOVE(group
, next
);
1983 trace_vfio_put_group(group
->fd
);
1987 if (QLIST_EMPTY(&vfio_group_list
)) {
1988 qemu_unregister_reset(vfio_reset_handler
, NULL
);
1992 int vfio_get_device(VFIOGroup
*group
, const char *name
,
1993 VFIODevice
*vbasedev
, Error
**errp
)
1995 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
1998 fd
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
2000 error_setg_errno(errp
, errno
, "error getting device from group %d",
2002 error_append_hint(errp
,
2003 "Verify all devices in group %d are bound to vfio-<bus> "
2004 "or pci-stub and not already in use\n", group
->groupid
);
2008 ret
= ioctl(fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
2010 error_setg_errno(errp
, errno
, "error getting device info");
2016 * Set discarding of RAM as not broken for this group if the driver knows
2017 * the device operates compatibly with discarding. Setting must be
2018 * consistent per group, but since compatibility is really only possible
2019 * with mdev currently, we expect singleton groups.
2021 if (vbasedev
->ram_block_discard_allowed
!=
2022 group
->ram_block_discard_allowed
) {
2023 if (!QLIST_EMPTY(&group
->device_list
)) {
2024 error_setg(errp
, "Inconsistent setting of support for discarding "
2025 "RAM (e.g., balloon) within group");
2030 if (!group
->ram_block_discard_allowed
) {
2031 group
->ram_block_discard_allowed
= true;
2032 ram_block_discard_disable(false);
2037 vbasedev
->group
= group
;
2038 QLIST_INSERT_HEAD(&group
->device_list
, vbasedev
, next
);
2040 vbasedev
->num_irqs
= dev_info
.num_irqs
;
2041 vbasedev
->num_regions
= dev_info
.num_regions
;
2042 vbasedev
->flags
= dev_info
.flags
;
2044 trace_vfio_get_device(name
, dev_info
.flags
, dev_info
.num_regions
,
2047 vbasedev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
2051 void vfio_put_base_device(VFIODevice
*vbasedev
)
2053 if (!vbasedev
->group
) {
2056 QLIST_REMOVE(vbasedev
, next
);
2057 vbasedev
->group
= NULL
;
2058 trace_vfio_put_base_device(vbasedev
->fd
);
2059 close(vbasedev
->fd
);
2062 int vfio_get_region_info(VFIODevice
*vbasedev
, int index
,
2063 struct vfio_region_info
**info
)
2065 size_t argsz
= sizeof(struct vfio_region_info
);
2067 *info
= g_malloc0(argsz
);
2069 (*info
)->index
= index
;
2071 (*info
)->argsz
= argsz
;
2073 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, *info
)) {
2079 if ((*info
)->argsz
> argsz
) {
2080 argsz
= (*info
)->argsz
;
2081 *info
= g_realloc(*info
, argsz
);
2089 int vfio_get_dev_region_info(VFIODevice
*vbasedev
, uint32_t type
,
2090 uint32_t subtype
, struct vfio_region_info
**info
)
2094 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
2095 struct vfio_info_cap_header
*hdr
;
2096 struct vfio_region_info_cap_type
*cap_type
;
2098 if (vfio_get_region_info(vbasedev
, i
, info
)) {
2102 hdr
= vfio_get_region_info_cap(*info
, VFIO_REGION_INFO_CAP_TYPE
);
2108 cap_type
= container_of(hdr
, struct vfio_region_info_cap_type
, header
);
2110 trace_vfio_get_dev_region(vbasedev
->name
, i
,
2111 cap_type
->type
, cap_type
->subtype
);
2113 if (cap_type
->type
== type
&& cap_type
->subtype
== subtype
) {
2124 bool vfio_has_region_cap(VFIODevice
*vbasedev
, int region
, uint16_t cap_type
)
2126 struct vfio_region_info
*info
= NULL
;
2129 if (!vfio_get_region_info(vbasedev
, region
, &info
)) {
2130 if (vfio_get_region_info_cap(info
, cap_type
)) {
2140 * Interfaces for IBM EEH (Enhanced Error Handling)
2142 static bool vfio_eeh_container_ok(VFIOContainer
*container
)
2145 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2146 * implementation is broken if there are multiple groups in a
2147 * container. The hardware works in units of Partitionable
2148 * Endpoints (== IOMMU groups) and the EEH operations naively
2149 * iterate across all groups in the container, without any logic
2150 * to make sure the groups have their state synchronized. For
2151 * certain operations (ENABLE) that might be ok, until an error
2152 * occurs, but for others (GET_STATE) it's clearly broken.
2156 * XXX Once fixed kernels exist, test for them here
2159 if (QLIST_EMPTY(&container
->group_list
)) {
2163 if (QLIST_NEXT(QLIST_FIRST(&container
->group_list
), container_next
)) {
2170 static int vfio_eeh_container_op(VFIOContainer
*container
, uint32_t op
)
2172 struct vfio_eeh_pe_op pe_op
= {
2173 .argsz
= sizeof(pe_op
),
2178 if (!vfio_eeh_container_ok(container
)) {
2179 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
2180 "kernel requires a container with exactly one group", op
);
2184 ret
= ioctl(container
->fd
, VFIO_EEH_PE_OP
, &pe_op
);
2186 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op
);
2193 static VFIOContainer
*vfio_eeh_as_container(AddressSpace
*as
)
2195 VFIOAddressSpace
*space
= vfio_get_address_space(as
);
2196 VFIOContainer
*container
= NULL
;
2198 if (QLIST_EMPTY(&space
->containers
)) {
2199 /* No containers to act on */
2203 container
= QLIST_FIRST(&space
->containers
);
2205 if (QLIST_NEXT(container
, next
)) {
2206 /* We don't yet have logic to synchronize EEH state across
2207 * multiple containers */
2213 vfio_put_address_space(space
);
2217 bool vfio_eeh_as_ok(AddressSpace
*as
)
2219 VFIOContainer
*container
= vfio_eeh_as_container(as
);
2221 return (container
!= NULL
) && vfio_eeh_container_ok(container
);
2224 int vfio_eeh_as_op(AddressSpace
*as
, uint32_t op
)
2226 VFIOContainer
*container
= vfio_eeh_as_container(as
);
2231 return vfio_eeh_container_op(container
, op
);