2 * generic functions used by VFIO devices
4 * Copyright Red Hat, Inc. 2012
7 * Alex Williamson <alex.williamson@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Based on qemu-kvm device-assignment:
13 * Adapted for KVM by Qumranet.
14 * Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15 * Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16 * Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17 * Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18 * Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
24 #include <linux/kvm.h>
26 #include <linux/vfio.h>
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/vfio.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
33 #include "qemu/error-report.h"
34 #include "qemu/main-loop.h"
35 #include "qemu/range.h"
36 #include "sysemu/balloon.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
40 #include "qapi/error.h"
42 VFIOGroupList vfio_group_list
=
43 QLIST_HEAD_INITIALIZER(vfio_group_list
);
44 static QLIST_HEAD(, VFIOAddressSpace
) vfio_address_spaces
=
45 QLIST_HEAD_INITIALIZER(vfio_address_spaces
);
49 * We have a single VFIO pseudo device per KVM VM. Once created it lives
50 * for the life of the VM. Closing the file descriptor only drops our
51 * reference to it and the device's reference to kvm. Therefore once
52 * initialized, this file descriptor is only released on QEMU exit and
53 * we'll re-use it should another vfio device be attached before then.
55 static int vfio_kvm_device_fd
= -1;
59 * Common VFIO interrupt disable
61 void vfio_disable_irqindex(VFIODevice
*vbasedev
, int index
)
63 struct vfio_irq_set irq_set
= {
64 .argsz
= sizeof(irq_set
),
65 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_TRIGGER
,
71 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
74 void vfio_unmask_single_irqindex(VFIODevice
*vbasedev
, int index
)
76 struct vfio_irq_set irq_set
= {
77 .argsz
= sizeof(irq_set
),
78 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_UNMASK
,
84 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
87 void vfio_mask_single_irqindex(VFIODevice
*vbasedev
, int index
)
89 struct vfio_irq_set irq_set
= {
90 .argsz
= sizeof(irq_set
),
91 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_MASK
,
97 ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
100 static inline const char *action_to_str(int action
)
103 case VFIO_IRQ_SET_ACTION_MASK
:
105 case VFIO_IRQ_SET_ACTION_UNMASK
:
107 case VFIO_IRQ_SET_ACTION_TRIGGER
:
110 return "UNKNOWN ACTION";
114 static const char *index_to_str(VFIODevice
*vbasedev
, int index
)
116 if (vbasedev
->type
!= VFIO_DEVICE_TYPE_PCI
) {
121 case VFIO_PCI_INTX_IRQ_INDEX
:
123 case VFIO_PCI_MSI_IRQ_INDEX
:
125 case VFIO_PCI_MSIX_IRQ_INDEX
:
127 case VFIO_PCI_ERR_IRQ_INDEX
:
129 case VFIO_PCI_REQ_IRQ_INDEX
:
136 int vfio_set_irq_signaling(VFIODevice
*vbasedev
, int index
, int subindex
,
137 int action
, int fd
, Error
**errp
)
139 struct vfio_irq_set
*irq_set
;
144 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
146 irq_set
= g_malloc0(argsz
);
147 irq_set
->argsz
= argsz
;
148 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| action
;
149 irq_set
->index
= index
;
150 irq_set
->start
= subindex
;
152 pfd
= (int32_t *)&irq_set
->data
;
155 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
)) {
164 error_setg_errno(errp
, -ret
, "VFIO_DEVICE_SET_IRQS failure");
166 name
= index_to_str(vbasedev
, index
);
168 error_prepend(errp
, "%s-%d: ", name
, subindex
);
170 error_prepend(errp
, "index %d-%d: ", index
, subindex
);
173 "Failed to %s %s eventfd signaling for interrupt ",
174 fd
< 0 ? "tear down" : "set up", action_to_str(action
));
179 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
181 void vfio_region_write(void *opaque
, hwaddr addr
,
182 uint64_t data
, unsigned size
)
184 VFIORegion
*region
= opaque
;
185 VFIODevice
*vbasedev
= region
->vbasedev
;
198 buf
.word
= cpu_to_le16(data
);
201 buf
.dword
= cpu_to_le32(data
);
204 buf
.qword
= cpu_to_le64(data
);
207 hw_error("vfio: unsupported write size, %d bytes", size
);
211 if (pwrite(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
212 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", 0x%"PRIx64
214 __func__
, vbasedev
->name
, region
->nr
,
218 trace_vfio_region_write(vbasedev
->name
, region
->nr
, addr
, data
, size
);
221 * A read or write to a BAR always signals an INTx EOI. This will
222 * do nothing if not pending (including not in INTx mode). We assume
223 * that a BAR access is in response to an interrupt and that BAR
224 * accesses will service the interrupt. Unfortunately, we don't know
225 * which access will service the interrupt, so we're potentially
226 * getting quite a few host interrupts per guest interrupt.
228 vbasedev
->ops
->vfio_eoi(vbasedev
);
231 uint64_t vfio_region_read(void *opaque
,
232 hwaddr addr
, unsigned size
)
234 VFIORegion
*region
= opaque
;
235 VFIODevice
*vbasedev
= region
->vbasedev
;
244 if (pread(vbasedev
->fd
, &buf
, size
, region
->fd_offset
+ addr
) != size
) {
245 error_report("%s(%s:region%d+0x%"HWADDR_PRIx
", %d) failed: %m",
246 __func__
, vbasedev
->name
, region
->nr
,
255 data
= le16_to_cpu(buf
.word
);
258 data
= le32_to_cpu(buf
.dword
);
261 data
= le64_to_cpu(buf
.qword
);
264 hw_error("vfio: unsupported read size, %d bytes", size
);
268 trace_vfio_region_read(vbasedev
->name
, region
->nr
, addr
, size
, data
);
270 /* Same as write above */
271 vbasedev
->ops
->vfio_eoi(vbasedev
);
276 const MemoryRegionOps vfio_region_ops
= {
277 .read
= vfio_region_read
,
278 .write
= vfio_region_write
,
279 .endianness
= DEVICE_LITTLE_ENDIAN
,
281 .min_access_size
= 1,
282 .max_access_size
= 8,
285 .min_access_size
= 1,
286 .max_access_size
= 8,
291 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
293 static int vfio_dma_unmap(VFIOContainer
*container
,
294 hwaddr iova
, ram_addr_t size
)
296 struct vfio_iommu_type1_dma_unmap unmap
= {
297 .argsz
= sizeof(unmap
),
303 while (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
305 * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
306 * v4.15) where an overflow in its wrap-around check prevents us from
307 * unmapping the last page of the address space. Test for the error
308 * condition and re-try the unmap excluding the last page. The
309 * expectation is that we've never mapped the last page anyway and this
310 * unmap request comes via vIOMMU support which also makes it unlikely
311 * that this page is used. This bug was introduced well after type1 v2
312 * support was introduced, so we shouldn't need to test for v1. A fix
313 * is queued for kernel v5.0 so this workaround can be removed once
314 * affected kernels are sufficiently deprecated.
316 if (errno
== EINVAL
&& unmap
.size
&& !(unmap
.iova
+ unmap
.size
) &&
317 container
->iommu_type
== VFIO_TYPE1v2_IOMMU
) {
318 trace_vfio_dma_unmap_overflow_workaround();
319 unmap
.size
-= 1ULL << ctz64(container
->pgsizes
);
322 error_report("VFIO_UNMAP_DMA: %d", -errno
);
329 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
330 ram_addr_t size
, void *vaddr
, bool readonly
)
332 struct vfio_iommu_type1_dma_map map
= {
333 .argsz
= sizeof(map
),
334 .flags
= VFIO_DMA_MAP_FLAG_READ
,
335 .vaddr
= (__u64
)(uintptr_t)vaddr
,
341 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
345 * Try the mapping, if it fails with EBUSY, unmap the region and try
346 * again. This shouldn't be necessary, but we sometimes see it in
349 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
350 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
) == 0 &&
351 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
355 error_report("VFIO_MAP_DMA: %d", -errno
);
359 static void vfio_host_win_add(VFIOContainer
*container
,
360 hwaddr min_iova
, hwaddr max_iova
,
361 uint64_t iova_pgsizes
)
363 VFIOHostDMAWindow
*hostwin
;
365 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
366 if (ranges_overlap(hostwin
->min_iova
,
367 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
369 max_iova
- min_iova
+ 1)) {
370 hw_error("%s: Overlapped IOMMU are not enabled", __func__
);
374 hostwin
= g_malloc0(sizeof(*hostwin
));
376 hostwin
->min_iova
= min_iova
;
377 hostwin
->max_iova
= max_iova
;
378 hostwin
->iova_pgsizes
= iova_pgsizes
;
379 QLIST_INSERT_HEAD(&container
->hostwin_list
, hostwin
, hostwin_next
);
382 static int vfio_host_win_del(VFIOContainer
*container
, hwaddr min_iova
,
385 VFIOHostDMAWindow
*hostwin
;
387 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
388 if (hostwin
->min_iova
== min_iova
&& hostwin
->max_iova
== max_iova
) {
389 QLIST_REMOVE(hostwin
, hostwin_next
);
397 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
399 return (!memory_region_is_ram(section
->mr
) &&
400 !memory_region_is_iommu(section
->mr
)) ||
402 * Sizing an enabled 64-bit BAR can cause spurious mappings to
403 * addresses in the upper part of the 64-bit address space. These
404 * are never accessed by the CPU and beyond the address width of
405 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
407 section
->offset_within_address_space
& (1ULL << 63);
410 /* Called with rcu_read_lock held. */
411 static bool vfio_get_vaddr(IOMMUTLBEntry
*iotlb
, void **vaddr
,
416 hwaddr len
= iotlb
->addr_mask
+ 1;
417 bool writable
= iotlb
->perm
& IOMMU_WO
;
420 * The IOMMU TLB entry we have just covers translation through
421 * this IOMMU to its immediate target. We need to translate
422 * it the rest of the way through to memory.
424 mr
= address_space_translate(&address_space_memory
,
425 iotlb
->translated_addr
,
426 &xlat
, &len
, writable
,
427 MEMTXATTRS_UNSPECIFIED
);
428 if (!memory_region_is_ram(mr
)) {
429 error_report("iommu map to non memory area %"HWADDR_PRIx
"",
435 * Translation truncates length to the IOMMU page size,
436 * check that it did not truncate too much.
438 if (len
& iotlb
->addr_mask
) {
439 error_report("iommu has granularity incompatible with target AS");
443 *vaddr
= memory_region_get_ram_ptr(mr
) + xlat
;
444 *read_only
= !writable
|| mr
->readonly
;
449 static void vfio_iommu_map_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
451 VFIOGuestIOMMU
*giommu
= container_of(n
, VFIOGuestIOMMU
, n
);
452 VFIOContainer
*container
= giommu
->container
;
453 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
458 trace_vfio_iommu_map_notify(iotlb
->perm
== IOMMU_NONE
? "UNMAP" : "MAP",
459 iova
, iova
+ iotlb
->addr_mask
);
461 if (iotlb
->target_as
!= &address_space_memory
) {
462 error_report("Wrong target AS \"%s\", only system memory is allowed",
463 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
469 if ((iotlb
->perm
& IOMMU_RW
) != IOMMU_NONE
) {
470 if (!vfio_get_vaddr(iotlb
, &vaddr
, &read_only
)) {
474 * vaddr is only valid until rcu_read_unlock(). But after
475 * vfio_dma_map has set up the mapping the pages will be
476 * pinned by the kernel. This makes sure that the RAM backend
477 * of vaddr will always be there, even if the memory object is
478 * destroyed and its backing memory munmap-ed.
480 ret
= vfio_dma_map(container
, iova
,
481 iotlb
->addr_mask
+ 1, vaddr
,
484 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
485 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
487 iotlb
->addr_mask
+ 1, vaddr
, ret
);
490 ret
= vfio_dma_unmap(container
, iova
, iotlb
->addr_mask
+ 1);
492 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
493 "0x%"HWADDR_PRIx
") = %d (%m)",
495 iotlb
->addr_mask
+ 1, ret
);
502 static void vfio_listener_region_add(MemoryListener
*listener
,
503 MemoryRegionSection
*section
)
505 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
507 Int128 llend
, llsize
;
510 VFIOHostDMAWindow
*hostwin
;
513 if (vfio_listener_skipped_section(section
)) {
514 trace_vfio_listener_region_add_skip(
515 section
->offset_within_address_space
,
516 section
->offset_within_address_space
+
517 int128_get64(int128_sub(section
->size
, int128_one())));
521 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
522 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
523 error_report("%s received unaligned region", __func__
);
527 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
528 llend
= int128_make64(section
->offset_within_address_space
);
529 llend
= int128_add(llend
, section
->size
);
530 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
532 if (int128_ge(int128_make64(iova
), llend
)) {
535 end
= int128_get64(int128_sub(llend
, int128_one()));
537 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
540 /* For now intersections are not allowed, we may relax this later */
541 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
542 if (ranges_overlap(hostwin
->min_iova
,
543 hostwin
->max_iova
- hostwin
->min_iova
+ 1,
544 section
->offset_within_address_space
,
545 int128_get64(section
->size
))) {
551 ret
= vfio_spapr_create_window(container
, section
, &pgsize
);
556 vfio_host_win_add(container
, section
->offset_within_address_space
,
557 section
->offset_within_address_space
+
558 int128_get64(section
->size
) - 1, pgsize
);
562 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
563 struct kvm_vfio_spapr_tce param
;
564 struct kvm_device_attr attr
= {
565 .group
= KVM_DEV_VFIO_GROUP
,
566 .attr
= KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE
,
567 .addr
= (uint64_t)(unsigned long)¶m
,
570 if (!memory_region_iommu_get_attr(iommu_mr
, IOMMU_ATTR_SPAPR_TCE_FD
,
572 QLIST_FOREACH(group
, &container
->group_list
, container_next
) {
573 param
.groupfd
= group
->fd
;
574 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
575 error_report("vfio: failed to setup fd %d "
576 "for a group with fd %d: %s",
577 param
.tablefd
, param
.groupfd
,
581 trace_vfio_spapr_group_attach(param
.groupfd
, param
.tablefd
);
588 hostwin_found
= false;
589 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
590 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
591 hostwin_found
= true;
596 if (!hostwin_found
) {
597 error_report("vfio: IOMMU container %p can't map guest IOVA region"
598 " 0x%"HWADDR_PRIx
"..0x%"HWADDR_PRIx
,
599 container
, iova
, end
);
604 memory_region_ref(section
->mr
);
606 if (memory_region_is_iommu(section
->mr
)) {
607 VFIOGuestIOMMU
*giommu
;
608 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
611 trace_vfio_listener_region_add_iommu(iova
, end
);
613 * FIXME: For VFIO iommu types which have KVM acceleration to
614 * avoid bouncing all map/unmaps through qemu this way, this
615 * would be the right place to wire that up (tell the KVM
616 * device emulation the VFIO iommu handles to use).
618 giommu
= g_malloc0(sizeof(*giommu
));
619 giommu
->iommu
= iommu_mr
;
620 giommu
->iommu_offset
= section
->offset_within_address_space
-
621 section
->offset_within_region
;
622 giommu
->container
= container
;
623 llend
= int128_add(int128_make64(section
->offset_within_region
),
625 llend
= int128_sub(llend
, int128_one());
626 iommu_idx
= memory_region_iommu_attrs_to_index(iommu_mr
,
627 MEMTXATTRS_UNSPECIFIED
);
628 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
630 section
->offset_within_region
,
633 QLIST_INSERT_HEAD(&container
->giommu_list
, giommu
, giommu_next
);
635 memory_region_register_iommu_notifier(section
->mr
, &giommu
->n
);
636 memory_region_iommu_replay(giommu
->iommu
, &giommu
->n
);
641 /* Here we assume that memory_region_is_ram(section->mr)==true */
643 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
644 section
->offset_within_region
+
645 (iova
- section
->offset_within_address_space
);
647 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
649 llsize
= int128_sub(llend
, int128_make64(iova
));
651 if (memory_region_is_ram_device(section
->mr
)) {
652 hwaddr pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
654 if ((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
)) {
655 trace_vfio_listener_region_add_no_dma_map(
656 memory_region_name(section
->mr
),
657 section
->offset_within_address_space
,
658 int128_getlo(section
->size
),
664 ret
= vfio_dma_map(container
, iova
, int128_get64(llsize
),
665 vaddr
, section
->readonly
);
667 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
668 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
669 container
, iova
, int128_get64(llsize
), vaddr
, ret
);
670 if (memory_region_is_ram_device(section
->mr
)) {
671 /* Allow unexpected mappings not to be fatal for RAM devices */
680 if (memory_region_is_ram_device(section
->mr
)) {
681 error_report("failed to vfio_dma_map. pci p2p may not work");
685 * On the initfn path, store the first error in the container so we
686 * can gracefully fail. Runtime, there's not much we can do other
687 * than throw a hardware error.
689 if (!container
->initialized
) {
690 if (!container
->error
) {
691 container
->error
= ret
;
694 hw_error("vfio: DMA mapping failed, unable to continue");
698 static void vfio_listener_region_del(MemoryListener
*listener
,
699 MemoryRegionSection
*section
)
701 VFIOContainer
*container
= container_of(listener
, VFIOContainer
, listener
);
703 Int128 llend
, llsize
;
705 bool try_unmap
= true;
707 if (vfio_listener_skipped_section(section
)) {
708 trace_vfio_listener_region_del_skip(
709 section
->offset_within_address_space
,
710 section
->offset_within_address_space
+
711 int128_get64(int128_sub(section
->size
, int128_one())));
715 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
716 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
717 error_report("%s received unaligned region", __func__
);
721 if (memory_region_is_iommu(section
->mr
)) {
722 VFIOGuestIOMMU
*giommu
;
724 QLIST_FOREACH(giommu
, &container
->giommu_list
, giommu_next
) {
725 if (MEMORY_REGION(giommu
->iommu
) == section
->mr
&&
726 giommu
->n
.start
== section
->offset_within_region
) {
727 memory_region_unregister_iommu_notifier(section
->mr
,
729 QLIST_REMOVE(giommu
, giommu_next
);
736 * FIXME: We assume the one big unmap below is adequate to
737 * remove any individual page mappings in the IOMMU which
738 * might have been copied into VFIO. This works for a page table
739 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
740 * That may not be true for all IOMMU types.
744 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
745 llend
= int128_make64(section
->offset_within_address_space
);
746 llend
= int128_add(llend
, section
->size
);
747 llend
= int128_and(llend
, int128_exts64(TARGET_PAGE_MASK
));
749 if (int128_ge(int128_make64(iova
), llend
)) {
752 end
= int128_get64(int128_sub(llend
, int128_one()));
754 llsize
= int128_sub(llend
, int128_make64(iova
));
756 trace_vfio_listener_region_del(iova
, end
);
758 if (memory_region_is_ram_device(section
->mr
)) {
760 VFIOHostDMAWindow
*hostwin
;
761 bool hostwin_found
= false;
763 QLIST_FOREACH(hostwin
, &container
->hostwin_list
, hostwin_next
) {
764 if (hostwin
->min_iova
<= iova
&& end
<= hostwin
->max_iova
) {
765 hostwin_found
= true;
769 assert(hostwin_found
); /* or region_add() would have failed */
771 pgmask
= (1ULL << ctz64(hostwin
->iova_pgsizes
)) - 1;
772 try_unmap
= !((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
));
776 ret
= vfio_dma_unmap(container
, iova
, int128_get64(llsize
));
778 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
779 "0x%"HWADDR_PRIx
") = %d (%m)",
780 container
, iova
, int128_get64(llsize
), ret
);
784 memory_region_unref(section
->mr
);
786 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
787 vfio_spapr_remove_window(container
,
788 section
->offset_within_address_space
);
789 if (vfio_host_win_del(container
,
790 section
->offset_within_address_space
,
791 section
->offset_within_address_space
+
792 int128_get64(section
->size
) - 1) < 0) {
793 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx
,
794 __func__
, section
->offset_within_address_space
);
799 static const MemoryListener vfio_memory_listener
= {
800 .region_add
= vfio_listener_region_add
,
801 .region_del
= vfio_listener_region_del
,
804 static void vfio_listener_release(VFIOContainer
*container
)
806 memory_listener_unregister(&container
->listener
);
807 if (container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
808 memory_listener_unregister(&container
->prereg_listener
);
812 struct vfio_info_cap_header
*
813 vfio_get_region_info_cap(struct vfio_region_info
*info
, uint16_t id
)
815 struct vfio_info_cap_header
*hdr
;
818 if (!(info
->flags
& VFIO_REGION_INFO_FLAG_CAPS
)) {
822 for (hdr
= ptr
+ info
->cap_offset
; hdr
!= ptr
; hdr
= ptr
+ hdr
->next
) {
831 static int vfio_setup_region_sparse_mmaps(VFIORegion
*region
,
832 struct vfio_region_info
*info
)
834 struct vfio_info_cap_header
*hdr
;
835 struct vfio_region_info_cap_sparse_mmap
*sparse
;
838 hdr
= vfio_get_region_info_cap(info
, VFIO_REGION_INFO_CAP_SPARSE_MMAP
);
843 sparse
= container_of(hdr
, struct vfio_region_info_cap_sparse_mmap
, header
);
845 trace_vfio_region_sparse_mmap_header(region
->vbasedev
->name
,
846 region
->nr
, sparse
->nr_areas
);
848 region
->mmaps
= g_new0(VFIOMmap
, sparse
->nr_areas
);
850 for (i
= 0, j
= 0; i
< sparse
->nr_areas
; i
++) {
851 trace_vfio_region_sparse_mmap_entry(i
, sparse
->areas
[i
].offset
,
852 sparse
->areas
[i
].offset
+
853 sparse
->areas
[i
].size
);
855 if (sparse
->areas
[i
].size
) {
856 region
->mmaps
[j
].offset
= sparse
->areas
[i
].offset
;
857 region
->mmaps
[j
].size
= sparse
->areas
[i
].size
;
862 region
->nr_mmaps
= j
;
863 region
->mmaps
= g_realloc(region
->mmaps
, j
* sizeof(VFIOMmap
));
868 int vfio_region_setup(Object
*obj
, VFIODevice
*vbasedev
, VFIORegion
*region
,
869 int index
, const char *name
)
871 struct vfio_region_info
*info
;
874 ret
= vfio_get_region_info(vbasedev
, index
, &info
);
879 region
->vbasedev
= vbasedev
;
880 region
->flags
= info
->flags
;
881 region
->size
= info
->size
;
882 region
->fd_offset
= info
->offset
;
886 region
->mem
= g_new0(MemoryRegion
, 1);
887 memory_region_init_io(region
->mem
, obj
, &vfio_region_ops
,
888 region
, name
, region
->size
);
890 if (!vbasedev
->no_mmap
&&
891 region
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
893 ret
= vfio_setup_region_sparse_mmaps(region
, info
);
896 region
->nr_mmaps
= 1;
897 region
->mmaps
= g_new0(VFIOMmap
, region
->nr_mmaps
);
898 region
->mmaps
[0].offset
= 0;
899 region
->mmaps
[0].size
= region
->size
;
906 trace_vfio_region_setup(vbasedev
->name
, index
, name
,
907 region
->flags
, region
->fd_offset
, region
->size
);
911 int vfio_region_mmap(VFIORegion
*region
)
920 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_READ
? PROT_READ
: 0;
921 prot
|= region
->flags
& VFIO_REGION_INFO_FLAG_WRITE
? PROT_WRITE
: 0;
923 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
924 region
->mmaps
[i
].mmap
= mmap(NULL
, region
->mmaps
[i
].size
, prot
,
925 MAP_SHARED
, region
->vbasedev
->fd
,
927 region
->mmaps
[i
].offset
);
928 if (region
->mmaps
[i
].mmap
== MAP_FAILED
) {
931 trace_vfio_region_mmap_fault(memory_region_name(region
->mem
), i
,
933 region
->mmaps
[i
].offset
,
935 region
->mmaps
[i
].offset
+
936 region
->mmaps
[i
].size
- 1, ret
);
938 region
->mmaps
[i
].mmap
= NULL
;
940 for (i
--; i
>= 0; i
--) {
941 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
942 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
943 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
944 region
->mmaps
[i
].mmap
= NULL
;
950 name
= g_strdup_printf("%s mmaps[%d]",
951 memory_region_name(region
->mem
), i
);
952 memory_region_init_ram_device_ptr(®ion
->mmaps
[i
].mem
,
953 memory_region_owner(region
->mem
),
954 name
, region
->mmaps
[i
].size
,
955 region
->mmaps
[i
].mmap
);
957 memory_region_add_subregion(region
->mem
, region
->mmaps
[i
].offset
,
958 ®ion
->mmaps
[i
].mem
);
960 trace_vfio_region_mmap(memory_region_name(®ion
->mmaps
[i
].mem
),
961 region
->mmaps
[i
].offset
,
962 region
->mmaps
[i
].offset
+
963 region
->mmaps
[i
].size
- 1);
969 void vfio_region_exit(VFIORegion
*region
)
977 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
978 if (region
->mmaps
[i
].mmap
) {
979 memory_region_del_subregion(region
->mem
, ®ion
->mmaps
[i
].mem
);
983 trace_vfio_region_exit(region
->vbasedev
->name
, region
->nr
);
986 void vfio_region_finalize(VFIORegion
*region
)
994 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
995 if (region
->mmaps
[i
].mmap
) {
996 munmap(region
->mmaps
[i
].mmap
, region
->mmaps
[i
].size
);
997 object_unparent(OBJECT(®ion
->mmaps
[i
].mem
));
1001 object_unparent(OBJECT(region
->mem
));
1003 g_free(region
->mem
);
1004 g_free(region
->mmaps
);
1006 trace_vfio_region_finalize(region
->vbasedev
->name
, region
->nr
);
1009 region
->mmaps
= NULL
;
1010 region
->nr_mmaps
= 0;
1016 void vfio_region_mmaps_set_enabled(VFIORegion
*region
, bool enabled
)
1024 for (i
= 0; i
< region
->nr_mmaps
; i
++) {
1025 if (region
->mmaps
[i
].mmap
) {
1026 memory_region_set_enabled(®ion
->mmaps
[i
].mem
, enabled
);
1030 trace_vfio_region_mmaps_set_enabled(memory_region_name(region
->mem
),
1034 void vfio_reset_handler(void *opaque
)
1037 VFIODevice
*vbasedev
;
1039 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1040 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1041 if (vbasedev
->dev
->realized
) {
1042 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
1047 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1048 QLIST_FOREACH(vbasedev
, &group
->device_list
, next
) {
1049 if (vbasedev
->dev
->realized
&& vbasedev
->needs_reset
) {
1050 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
1056 static void vfio_kvm_device_add_group(VFIOGroup
*group
)
1059 struct kvm_device_attr attr
= {
1060 .group
= KVM_DEV_VFIO_GROUP
,
1061 .attr
= KVM_DEV_VFIO_GROUP_ADD
,
1062 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1065 if (!kvm_enabled()) {
1069 if (vfio_kvm_device_fd
< 0) {
1070 struct kvm_create_device cd
= {
1071 .type
= KVM_DEV_TYPE_VFIO
,
1074 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
1075 error_report("Failed to create KVM VFIO device: %m");
1079 vfio_kvm_device_fd
= cd
.fd
;
1082 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1083 error_report("Failed to add group %d to KVM VFIO device: %m",
1089 static void vfio_kvm_device_del_group(VFIOGroup
*group
)
1092 struct kvm_device_attr attr
= {
1093 .group
= KVM_DEV_VFIO_GROUP
,
1094 .attr
= KVM_DEV_VFIO_GROUP_DEL
,
1095 .addr
= (uint64_t)(unsigned long)&group
->fd
,
1098 if (vfio_kvm_device_fd
< 0) {
1102 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1103 error_report("Failed to remove group %d from KVM VFIO device: %m",
1109 static VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
1111 VFIOAddressSpace
*space
;
1113 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
1114 if (space
->as
== as
) {
1119 /* No suitable VFIOAddressSpace, create a new one */
1120 space
= g_malloc0(sizeof(*space
));
1122 QLIST_INIT(&space
->containers
);
1124 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
1129 static void vfio_put_address_space(VFIOAddressSpace
*space
)
1131 if (QLIST_EMPTY(&space
->containers
)) {
1132 QLIST_REMOVE(space
, list
);
1138 * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1140 static int vfio_get_iommu_type(VFIOContainer
*container
,
1143 int iommu_types
[] = { VFIO_TYPE1v2_IOMMU
, VFIO_TYPE1_IOMMU
,
1144 VFIO_SPAPR_TCE_v2_IOMMU
, VFIO_SPAPR_TCE_IOMMU
};
1147 for (i
= 0; i
< ARRAY_SIZE(iommu_types
); i
++) {
1148 if (ioctl(container
->fd
, VFIO_CHECK_EXTENSION
, iommu_types
[i
])) {
1149 return iommu_types
[i
];
1152 error_setg(errp
, "No available IOMMU models");
1156 static int vfio_init_container(VFIOContainer
*container
, int group_fd
,
1159 int iommu_type
, ret
;
1161 iommu_type
= vfio_get_iommu_type(container
, errp
);
1162 if (iommu_type
< 0) {
1166 ret
= ioctl(group_fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
);
1168 error_setg_errno(errp
, errno
, "Failed to set group container");
1172 while (ioctl(container
->fd
, VFIO_SET_IOMMU
, iommu_type
)) {
1173 if (iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
) {
1175 * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1176 * v2, the running platform may not support v2 and there is no
1177 * way to guess it until an IOMMU group gets added to the container.
1178 * So in case it fails with v2, try v1 as a fallback.
1180 iommu_type
= VFIO_SPAPR_TCE_IOMMU
;
1183 error_setg_errno(errp
, errno
, "Failed to set iommu for container");
1187 container
->iommu_type
= iommu_type
;
1191 static int vfio_connect_container(VFIOGroup
*group
, AddressSpace
*as
,
1194 VFIOContainer
*container
;
1196 VFIOAddressSpace
*space
;
1198 space
= vfio_get_address_space(as
);
1201 * VFIO is currently incompatible with memory ballooning insofar as the
1202 * madvise to purge (zap) the page from QEMU's address space does not
1203 * interact with the memory API and therefore leaves stale virtual to
1204 * physical mappings in the IOMMU if the page was previously pinned. We
1205 * therefore add a balloon inhibit for each group added to a container,
1206 * whether the container is used individually or shared. This provides
1207 * us with options to allow devices within a group to opt-in and allow
1208 * ballooning, so long as it is done consistently for a group (for instance
1209 * if the device is an mdev device where it is known that the host vendor
1210 * driver will never pin pages outside of the working set of the guest
1211 * driver, which would thus not be ballooning candidates).
1213 * The first opportunity to induce pinning occurs here where we attempt to
1214 * attach the group to existing containers within the AddressSpace. If any
1215 * pages are already zapped from the virtual address space, such as from a
1216 * previous ballooning opt-in, new pinning will cause valid mappings to be
1217 * re-established. Likewise, when the overall MemoryListener for a new
1218 * container is registered, a replay of mappings within the AddressSpace
1219 * will occur, re-establishing any previously zapped pages as well.
1221 * NB. Balloon inhibiting does not currently block operation of the
1222 * balloon driver or revoke previously pinned pages, it only prevents
1223 * calling madvise to modify the virtual mapping of ballooned pages.
1225 qemu_balloon_inhibit(true);
1227 QLIST_FOREACH(container
, &space
->containers
, next
) {
1228 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
1229 group
->container
= container
;
1230 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1231 vfio_kvm_device_add_group(group
);
1236 fd
= qemu_open("/dev/vfio/vfio", O_RDWR
);
1238 error_setg_errno(errp
, errno
, "failed to open /dev/vfio/vfio");
1240 goto put_space_exit
;
1243 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
1244 if (ret
!= VFIO_API_VERSION
) {
1245 error_setg(errp
, "supported vfio version: %d, "
1246 "reported version: %d", VFIO_API_VERSION
, ret
);
1251 container
= g_malloc0(sizeof(*container
));
1252 container
->space
= space
;
1254 QLIST_INIT(&container
->giommu_list
);
1255 QLIST_INIT(&container
->hostwin_list
);
1257 ret
= vfio_init_container(container
, group
->fd
, errp
);
1259 goto free_container_exit
;
1262 switch (container
->iommu_type
) {
1263 case VFIO_TYPE1v2_IOMMU
:
1264 case VFIO_TYPE1_IOMMU
:
1266 struct vfio_iommu_type1_info info
;
1269 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1270 * IOVA whatsoever. That's not actually true, but the current
1271 * kernel interface doesn't tell us what it can map, and the
1272 * existing Type1 IOMMUs generally support any IOVA we're
1273 * going to actually try in practice.
1275 info
.argsz
= sizeof(info
);
1276 ret
= ioctl(fd
, VFIO_IOMMU_GET_INFO
, &info
);
1278 if (ret
|| !(info
.flags
& VFIO_IOMMU_INFO_PGSIZES
)) {
1279 /* Assume 4k IOVA page size */
1280 info
.iova_pgsizes
= 4096;
1282 vfio_host_win_add(container
, 0, (hwaddr
)-1, info
.iova_pgsizes
);
1283 container
->pgsizes
= info
.iova_pgsizes
;
1286 case VFIO_SPAPR_TCE_v2_IOMMU
:
1287 case VFIO_SPAPR_TCE_IOMMU
:
1289 struct vfio_iommu_spapr_tce_info info
;
1290 bool v2
= container
->iommu_type
== VFIO_SPAPR_TCE_v2_IOMMU
;
1293 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1294 * when container fd is closed so we do not call it explicitly
1298 ret
= ioctl(fd
, VFIO_IOMMU_ENABLE
);
1300 error_setg_errno(errp
, errno
, "failed to enable container");
1302 goto free_container_exit
;
1305 container
->prereg_listener
= vfio_prereg_listener
;
1307 memory_listener_register(&container
->prereg_listener
,
1308 &address_space_memory
);
1309 if (container
->error
) {
1310 memory_listener_unregister(&container
->prereg_listener
);
1311 ret
= container
->error
;
1313 "RAM memory listener initialization failed for container");
1314 goto free_container_exit
;
1318 info
.argsz
= sizeof(info
);
1319 ret
= ioctl(fd
, VFIO_IOMMU_SPAPR_TCE_GET_INFO
, &info
);
1321 error_setg_errno(errp
, errno
,
1322 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1325 memory_listener_unregister(&container
->prereg_listener
);
1327 goto free_container_exit
;
1331 container
->pgsizes
= info
.ddw
.pgsizes
;
1333 * There is a default window in just created container.
1334 * To make region_add/del simpler, we better remove this
1335 * window now and let those iommu_listener callbacks
1336 * create/remove them when needed.
1338 ret
= vfio_spapr_remove_window(container
, info
.dma32_window_start
);
1340 error_setg_errno(errp
, -ret
,
1341 "failed to remove existing window");
1342 goto free_container_exit
;
1345 /* The default table uses 4K pages */
1346 container
->pgsizes
= 0x1000;
1347 vfio_host_win_add(container
, info
.dma32_window_start
,
1348 info
.dma32_window_start
+
1349 info
.dma32_window_size
- 1,
1355 vfio_kvm_device_add_group(group
);
1357 QLIST_INIT(&container
->group_list
);
1358 QLIST_INSERT_HEAD(&space
->containers
, container
, next
);
1360 group
->container
= container
;
1361 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1363 container
->listener
= vfio_memory_listener
;
1365 memory_listener_register(&container
->listener
, container
->space
->as
);
1367 if (container
->error
) {
1368 ret
= container
->error
;
1369 error_setg_errno(errp
, -ret
,
1370 "memory listener initialization failed for container");
1371 goto listener_release_exit
;
1374 container
->initialized
= true;
1377 listener_release_exit
:
1378 QLIST_REMOVE(group
, container_next
);
1379 QLIST_REMOVE(container
, next
);
1380 vfio_kvm_device_del_group(group
);
1381 vfio_listener_release(container
);
1383 free_container_exit
:
1390 qemu_balloon_inhibit(false);
1391 vfio_put_address_space(space
);
1396 static void vfio_disconnect_container(VFIOGroup
*group
)
1398 VFIOContainer
*container
= group
->container
;
1400 QLIST_REMOVE(group
, container_next
);
1401 group
->container
= NULL
;
1404 * Explicitly release the listener first before unset container,
1405 * since unset may destroy the backend container if it's the last
1408 if (QLIST_EMPTY(&container
->group_list
)) {
1409 vfio_listener_release(container
);
1412 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1413 error_report("vfio: error disconnecting group %d from container",
1417 if (QLIST_EMPTY(&container
->group_list
)) {
1418 VFIOAddressSpace
*space
= container
->space
;
1419 VFIOGuestIOMMU
*giommu
, *tmp
;
1421 QLIST_REMOVE(container
, next
);
1423 QLIST_FOREACH_SAFE(giommu
, &container
->giommu_list
, giommu_next
, tmp
) {
1424 memory_region_unregister_iommu_notifier(
1425 MEMORY_REGION(giommu
->iommu
), &giommu
->n
);
1426 QLIST_REMOVE(giommu
, giommu_next
);
1430 trace_vfio_disconnect_container(container
->fd
);
1431 close(container
->fd
);
1434 vfio_put_address_space(space
);
1438 VFIOGroup
*vfio_get_group(int groupid
, AddressSpace
*as
, Error
**errp
)
1442 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1444 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
1445 if (group
->groupid
== groupid
) {
1446 /* Found it. Now is it already in the right context? */
1447 if (group
->container
->space
->as
== as
) {
1450 error_setg(errp
, "group %d used in multiple address spaces",
1457 group
= g_malloc0(sizeof(*group
));
1459 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1460 group
->fd
= qemu_open(path
, O_RDWR
);
1461 if (group
->fd
< 0) {
1462 error_setg_errno(errp
, errno
, "failed to open %s", path
);
1463 goto free_group_exit
;
1466 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1467 error_setg_errno(errp
, errno
, "failed to get group %d status", groupid
);
1471 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1472 error_setg(errp
, "group %d is not viable", groupid
);
1473 error_append_hint(errp
,
1474 "Please ensure all devices within the iommu_group "
1475 "are bound to their vfio bus driver.\n");
1479 group
->groupid
= groupid
;
1480 QLIST_INIT(&group
->device_list
);
1482 if (vfio_connect_container(group
, as
, errp
)) {
1483 error_prepend(errp
, "failed to setup container for group %d: ",
1488 if (QLIST_EMPTY(&vfio_group_list
)) {
1489 qemu_register_reset(vfio_reset_handler
, NULL
);
1492 QLIST_INSERT_HEAD(&vfio_group_list
, group
, next
);
1505 void vfio_put_group(VFIOGroup
*group
)
1507 if (!group
|| !QLIST_EMPTY(&group
->device_list
)) {
1511 if (!group
->balloon_allowed
) {
1512 qemu_balloon_inhibit(false);
1514 vfio_kvm_device_del_group(group
);
1515 vfio_disconnect_container(group
);
1516 QLIST_REMOVE(group
, next
);
1517 trace_vfio_put_group(group
->fd
);
1521 if (QLIST_EMPTY(&vfio_group_list
)) {
1522 qemu_unregister_reset(vfio_reset_handler
, NULL
);
1526 int vfio_get_device(VFIOGroup
*group
, const char *name
,
1527 VFIODevice
*vbasedev
, Error
**errp
)
1529 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
1532 fd
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
1534 error_setg_errno(errp
, errno
, "error getting device from group %d",
1536 error_append_hint(errp
,
1537 "Verify all devices in group %d are bound to vfio-<bus> "
1538 "or pci-stub and not already in use\n", group
->groupid
);
1542 ret
= ioctl(fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
1544 error_setg_errno(errp
, errno
, "error getting device info");
1550 * Clear the balloon inhibitor for this group if the driver knows the
1551 * device operates compatibly with ballooning. Setting must be consistent
1552 * per group, but since compatibility is really only possible with mdev
1553 * currently, we expect singleton groups.
1555 if (vbasedev
->balloon_allowed
!= group
->balloon_allowed
) {
1556 if (!QLIST_EMPTY(&group
->device_list
)) {
1558 "Inconsistent device balloon setting within group");
1563 if (!group
->balloon_allowed
) {
1564 group
->balloon_allowed
= true;
1565 qemu_balloon_inhibit(false);
1570 vbasedev
->group
= group
;
1571 QLIST_INSERT_HEAD(&group
->device_list
, vbasedev
, next
);
1573 vbasedev
->num_irqs
= dev_info
.num_irqs
;
1574 vbasedev
->num_regions
= dev_info
.num_regions
;
1575 vbasedev
->flags
= dev_info
.flags
;
1577 trace_vfio_get_device(name
, dev_info
.flags
, dev_info
.num_regions
,
1580 vbasedev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
1584 void vfio_put_base_device(VFIODevice
*vbasedev
)
1586 if (!vbasedev
->group
) {
1589 QLIST_REMOVE(vbasedev
, next
);
1590 vbasedev
->group
= NULL
;
1591 trace_vfio_put_base_device(vbasedev
->fd
);
1592 close(vbasedev
->fd
);
1595 int vfio_get_region_info(VFIODevice
*vbasedev
, int index
,
1596 struct vfio_region_info
**info
)
1598 size_t argsz
= sizeof(struct vfio_region_info
);
1600 *info
= g_malloc0(argsz
);
1602 (*info
)->index
= index
;
1604 (*info
)->argsz
= argsz
;
1606 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, *info
)) {
1612 if ((*info
)->argsz
> argsz
) {
1613 argsz
= (*info
)->argsz
;
1614 *info
= g_realloc(*info
, argsz
);
1622 int vfio_get_dev_region_info(VFIODevice
*vbasedev
, uint32_t type
,
1623 uint32_t subtype
, struct vfio_region_info
**info
)
1627 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
1628 struct vfio_info_cap_header
*hdr
;
1629 struct vfio_region_info_cap_type
*cap_type
;
1631 if (vfio_get_region_info(vbasedev
, i
, info
)) {
1635 hdr
= vfio_get_region_info_cap(*info
, VFIO_REGION_INFO_CAP_TYPE
);
1641 cap_type
= container_of(hdr
, struct vfio_region_info_cap_type
, header
);
1643 trace_vfio_get_dev_region(vbasedev
->name
, i
,
1644 cap_type
->type
, cap_type
->subtype
);
1646 if (cap_type
->type
== type
&& cap_type
->subtype
== subtype
) {
1657 bool vfio_has_region_cap(VFIODevice
*vbasedev
, int region
, uint16_t cap_type
)
1659 struct vfio_region_info
*info
= NULL
;
1662 if (!vfio_get_region_info(vbasedev
, region
, &info
)) {
1663 if (vfio_get_region_info_cap(info
, cap_type
)) {
1673 * Interfaces for IBM EEH (Enhanced Error Handling)
1675 static bool vfio_eeh_container_ok(VFIOContainer
*container
)
1678 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1679 * implementation is broken if there are multiple groups in a
1680 * container. The hardware works in units of Partitionable
1681 * Endpoints (== IOMMU groups) and the EEH operations naively
1682 * iterate across all groups in the container, without any logic
1683 * to make sure the groups have their state synchronized. For
1684 * certain operations (ENABLE) that might be ok, until an error
1685 * occurs, but for others (GET_STATE) it's clearly broken.
1689 * XXX Once fixed kernels exist, test for them here
1692 if (QLIST_EMPTY(&container
->group_list
)) {
1696 if (QLIST_NEXT(QLIST_FIRST(&container
->group_list
), container_next
)) {
1703 static int vfio_eeh_container_op(VFIOContainer
*container
, uint32_t op
)
1705 struct vfio_eeh_pe_op pe_op
= {
1706 .argsz
= sizeof(pe_op
),
1711 if (!vfio_eeh_container_ok(container
)) {
1712 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1713 "kernel requires a container with exactly one group", op
);
1717 ret
= ioctl(container
->fd
, VFIO_EEH_PE_OP
, &pe_op
);
1719 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op
);
1726 static VFIOContainer
*vfio_eeh_as_container(AddressSpace
*as
)
1728 VFIOAddressSpace
*space
= vfio_get_address_space(as
);
1729 VFIOContainer
*container
= NULL
;
1731 if (QLIST_EMPTY(&space
->containers
)) {
1732 /* No containers to act on */
1736 container
= QLIST_FIRST(&space
->containers
);
1738 if (QLIST_NEXT(container
, next
)) {
1739 /* We don't yet have logic to synchronize EEH state across
1740 * multiple containers */
1746 vfio_put_address_space(space
);
1750 bool vfio_eeh_as_ok(AddressSpace
*as
)
1752 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1754 return (container
!= NULL
) && vfio_eeh_container_ok(container
);
1757 int vfio_eeh_as_op(AddressSpace
*as
, uint32_t op
)
1759 VFIOContainer
*container
= vfio_eeh_as_container(as
);
1764 return vfio_eeh_container_op(container
, op
);