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/pci.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"
39 #include "sysemu/runstate.h"
41 #include "qapi/error.h"
42 #include "migration/misc.h"
43 #include "migration/blocker.h"
44 #include "migration/qemu-file.h"
45 #include "sysemu/tpm.h"
47 VFIODeviceList vfio_device_list
=
48 QLIST_HEAD_INITIALIZER(vfio_device_list
);
49 static QLIST_HEAD(, VFIOAddressSpace
) vfio_address_spaces
=
50 QLIST_HEAD_INITIALIZER(vfio_address_spaces
);
54 * We have a single VFIO pseudo device per KVM VM. Once created it lives
55 * for the life of the VM. Closing the file descriptor only drops our
56 * reference to it and the device's reference to kvm. Therefore once
57 * initialized, this file descriptor is only released on QEMU exit and
58 * we'll re-use it should another vfio device be attached before then.
60 int vfio_kvm_device_fd
= -1;
64 * Device state interfaces
67 bool vfio_mig_active(void)
71 if (QLIST_EMPTY(&vfio_device_list
)) {
75 QLIST_FOREACH(vbasedev
, &vfio_device_list
, global_next
) {
76 if (vbasedev
->migration_blocker
) {
83 static Error
*multiple_devices_migration_blocker
;
86 * Multiple devices migration is allowed only if all devices support P2P
87 * migration. Single device migration is allowed regardless of P2P migration
90 static bool vfio_multiple_devices_migration_is_supported(void)
93 unsigned int device_num
= 0;
94 bool all_support_p2p
= true;
96 QLIST_FOREACH(vbasedev
, &vfio_device_list
, global_next
) {
97 if (vbasedev
->migration
) {
100 if (!(vbasedev
->migration
->mig_flags
& VFIO_MIGRATION_P2P
)) {
101 all_support_p2p
= false;
106 return all_support_p2p
|| device_num
<= 1;
109 int vfio_block_multiple_devices_migration(VFIODevice
*vbasedev
, Error
**errp
)
113 if (vfio_multiple_devices_migration_is_supported()) {
117 if (vbasedev
->enable_migration
== ON_OFF_AUTO_ON
) {
118 error_setg(errp
, "Multiple VFIO devices migration is supported only if "
119 "all of them support P2P migration");
123 if (multiple_devices_migration_blocker
) {
127 error_setg(&multiple_devices_migration_blocker
,
128 "Multiple VFIO devices migration is supported only if all of "
129 "them support P2P migration");
130 ret
= migrate_add_blocker_normal(&multiple_devices_migration_blocker
, errp
);
135 void vfio_unblock_multiple_devices_migration(void)
137 if (!multiple_devices_migration_blocker
||
138 !vfio_multiple_devices_migration_is_supported()) {
142 migrate_del_blocker(&multiple_devices_migration_blocker
);
145 bool vfio_viommu_preset(VFIODevice
*vbasedev
)
147 return vbasedev
->bcontainer
->space
->as
!= &address_space_memory
;
150 static void vfio_set_migration_error(int err
)
152 if (migration_is_setup_or_active()) {
153 migration_file_set_error(err
);
157 bool vfio_device_state_is_running(VFIODevice
*vbasedev
)
159 VFIOMigration
*migration
= vbasedev
->migration
;
161 return migration
->device_state
== VFIO_DEVICE_STATE_RUNNING
||
162 migration
->device_state
== VFIO_DEVICE_STATE_RUNNING_P2P
;
165 bool vfio_device_state_is_precopy(VFIODevice
*vbasedev
)
167 VFIOMigration
*migration
= vbasedev
->migration
;
169 return migration
->device_state
== VFIO_DEVICE_STATE_PRE_COPY
||
170 migration
->device_state
== VFIO_DEVICE_STATE_PRE_COPY_P2P
;
173 static bool vfio_devices_all_dirty_tracking(VFIOContainerBase
*bcontainer
)
175 VFIODevice
*vbasedev
;
177 if (!migration_is_active() && !migration_is_device()) {
181 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
182 VFIOMigration
*migration
= vbasedev
->migration
;
188 if (vbasedev
->pre_copy_dirty_page_tracking
== ON_OFF_AUTO_OFF
&&
189 (vfio_device_state_is_running(vbasedev
) ||
190 vfio_device_state_is_precopy(vbasedev
))) {
197 bool vfio_devices_all_device_dirty_tracking(const VFIOContainerBase
*bcontainer
)
199 VFIODevice
*vbasedev
;
201 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
202 if (!vbasedev
->dirty_pages_supported
) {
211 * Check if all VFIO devices are running and migration is active, which is
212 * essentially equivalent to the migration being in pre-copy phase.
215 vfio_devices_all_running_and_mig_active(const VFIOContainerBase
*bcontainer
)
217 VFIODevice
*vbasedev
;
219 if (!migration_is_active()) {
223 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
224 VFIOMigration
*migration
= vbasedev
->migration
;
230 if (vfio_device_state_is_running(vbasedev
) ||
231 vfio_device_state_is_precopy(vbasedev
)) {
240 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
242 return (!memory_region_is_ram(section
->mr
) &&
243 !memory_region_is_iommu(section
->mr
)) ||
244 memory_region_is_protected(section
->mr
) ||
246 * Sizing an enabled 64-bit BAR can cause spurious mappings to
247 * addresses in the upper part of the 64-bit address space. These
248 * are never accessed by the CPU and beyond the address width of
249 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
251 section
->offset_within_address_space
& (1ULL << 63);
254 /* Called with rcu_read_lock held. */
255 static bool vfio_get_xlat_addr(IOMMUTLBEntry
*iotlb
, void **vaddr
,
256 ram_addr_t
*ram_addr
, bool *read_only
)
258 bool ret
, mr_has_discard_manager
;
260 ret
= memory_get_xlat_addr(iotlb
, vaddr
, ram_addr
, read_only
,
261 &mr_has_discard_manager
);
262 if (ret
&& mr_has_discard_manager
) {
264 * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
265 * pages will remain pinned inside vfio until unmapped, resulting in a
266 * higher memory consumption than expected. If memory would get
267 * populated again later, there would be an inconsistency between pages
268 * pinned by vfio and pages seen by QEMU. This is the case until
269 * unmapped from the IOMMU (e.g., during device reset).
271 * With malicious guests, we really only care about pinning more memory
272 * than expected. RLIMIT_MEMLOCK set for the user/process can never be
273 * exceeded and can be used to mitigate this problem.
275 warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
276 " RAM (e.g., virtio-mem) works, however, malicious"
277 " guests can trigger pinning of more memory than"
278 " intended via an IOMMU. It's possible to mitigate "
279 " by setting/adjusting RLIMIT_MEMLOCK.");
284 static void vfio_iommu_map_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
286 VFIOGuestIOMMU
*giommu
= container_of(n
, VFIOGuestIOMMU
, n
);
287 VFIOContainerBase
*bcontainer
= giommu
->bcontainer
;
288 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
292 trace_vfio_iommu_map_notify(iotlb
->perm
== IOMMU_NONE
? "UNMAP" : "MAP",
293 iova
, iova
+ iotlb
->addr_mask
);
295 if (iotlb
->target_as
!= &address_space_memory
) {
296 error_report("Wrong target AS \"%s\", only system memory is allowed",
297 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
298 vfio_set_migration_error(-EINVAL
);
304 if ((iotlb
->perm
& IOMMU_RW
) != IOMMU_NONE
) {
307 if (!vfio_get_xlat_addr(iotlb
, &vaddr
, NULL
, &read_only
)) {
311 * vaddr is only valid until rcu_read_unlock(). But after
312 * vfio_dma_map has set up the mapping the pages will be
313 * pinned by the kernel. This makes sure that the RAM backend
314 * of vaddr will always be there, even if the memory object is
315 * destroyed and its backing memory munmap-ed.
317 ret
= vfio_container_dma_map(bcontainer
, iova
,
318 iotlb
->addr_mask
+ 1, vaddr
,
321 error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx
", "
322 "0x%"HWADDR_PRIx
", %p) = %d (%s)",
324 iotlb
->addr_mask
+ 1, vaddr
, ret
, strerror(-ret
));
327 ret
= vfio_container_dma_unmap(bcontainer
, iova
,
328 iotlb
->addr_mask
+ 1, iotlb
);
330 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
331 "0x%"HWADDR_PRIx
") = %d (%s)",
333 iotlb
->addr_mask
+ 1, ret
, strerror(-ret
));
334 vfio_set_migration_error(ret
);
341 static void vfio_ram_discard_notify_discard(RamDiscardListener
*rdl
,
342 MemoryRegionSection
*section
)
344 VFIORamDiscardListener
*vrdl
= container_of(rdl
, VFIORamDiscardListener
,
346 VFIOContainerBase
*bcontainer
= vrdl
->bcontainer
;
347 const hwaddr size
= int128_get64(section
->size
);
348 const hwaddr iova
= section
->offset_within_address_space
;
351 /* Unmap with a single call. */
352 ret
= vfio_container_dma_unmap(bcontainer
, iova
, size
, NULL
);
354 error_report("%s: vfio_container_dma_unmap() failed: %s", __func__
,
359 static int vfio_ram_discard_notify_populate(RamDiscardListener
*rdl
,
360 MemoryRegionSection
*section
)
362 VFIORamDiscardListener
*vrdl
= container_of(rdl
, VFIORamDiscardListener
,
364 VFIOContainerBase
*bcontainer
= vrdl
->bcontainer
;
365 const hwaddr end
= section
->offset_within_region
+
366 int128_get64(section
->size
);
367 hwaddr start
, next
, iova
;
372 * Map in (aligned within memory region) minimum granularity, so we can
373 * unmap in minimum granularity later.
375 for (start
= section
->offset_within_region
; start
< end
; start
= next
) {
376 next
= ROUND_UP(start
+ 1, vrdl
->granularity
);
377 next
= MIN(next
, end
);
379 iova
= start
- section
->offset_within_region
+
380 section
->offset_within_address_space
;
381 vaddr
= memory_region_get_ram_ptr(section
->mr
) + start
;
383 ret
= vfio_container_dma_map(bcontainer
, iova
, next
- start
,
384 vaddr
, section
->readonly
);
387 vfio_ram_discard_notify_discard(rdl
, section
);
394 static void vfio_register_ram_discard_listener(VFIOContainerBase
*bcontainer
,
395 MemoryRegionSection
*section
)
397 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(section
->mr
);
398 VFIORamDiscardListener
*vrdl
;
400 /* Ignore some corner cases not relevant in practice. */
401 g_assert(QEMU_IS_ALIGNED(section
->offset_within_region
, TARGET_PAGE_SIZE
));
402 g_assert(QEMU_IS_ALIGNED(section
->offset_within_address_space
,
404 g_assert(QEMU_IS_ALIGNED(int128_get64(section
->size
), TARGET_PAGE_SIZE
));
406 vrdl
= g_new0(VFIORamDiscardListener
, 1);
407 vrdl
->bcontainer
= bcontainer
;
408 vrdl
->mr
= section
->mr
;
409 vrdl
->offset_within_address_space
= section
->offset_within_address_space
;
410 vrdl
->size
= int128_get64(section
->size
);
411 vrdl
->granularity
= ram_discard_manager_get_min_granularity(rdm
,
414 g_assert(vrdl
->granularity
&& is_power_of_2(vrdl
->granularity
));
415 g_assert(bcontainer
->pgsizes
&&
416 vrdl
->granularity
>= 1ULL << ctz64(bcontainer
->pgsizes
));
418 ram_discard_listener_init(&vrdl
->listener
,
419 vfio_ram_discard_notify_populate
,
420 vfio_ram_discard_notify_discard
, true);
421 ram_discard_manager_register_listener(rdm
, &vrdl
->listener
, section
);
422 QLIST_INSERT_HEAD(&bcontainer
->vrdl_list
, vrdl
, next
);
425 * Sanity-check if we have a theoretically problematic setup where we could
426 * exceed the maximum number of possible DMA mappings over time. We assume
427 * that each mapped section in the same address space as a RamDiscardManager
428 * section consumes exactly one DMA mapping, with the exception of
429 * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
430 * in the same address space as RamDiscardManager sections.
432 * We assume that each section in the address space consumes one memslot.
433 * We take the number of KVM memory slots as a best guess for the maximum
434 * number of sections in the address space we could have over time,
435 * also consuming DMA mappings.
437 if (bcontainer
->dma_max_mappings
) {
438 unsigned int vrdl_count
= 0, vrdl_mappings
= 0, max_memslots
= 512;
442 max_memslots
= kvm_get_max_memslots();
446 QLIST_FOREACH(vrdl
, &bcontainer
->vrdl_list
, next
) {
449 start
= QEMU_ALIGN_DOWN(vrdl
->offset_within_address_space
,
451 end
= ROUND_UP(vrdl
->offset_within_address_space
+ vrdl
->size
,
453 vrdl_mappings
+= (end
- start
) / vrdl
->granularity
;
457 if (vrdl_mappings
+ max_memslots
- vrdl_count
>
458 bcontainer
->dma_max_mappings
) {
459 warn_report("%s: possibly running out of DMA mappings. E.g., try"
460 " increasing the 'block-size' of virtio-mem devies."
461 " Maximum possible DMA mappings: %d, Maximum possible"
462 " memslots: %d", __func__
, bcontainer
->dma_max_mappings
,
468 static void vfio_unregister_ram_discard_listener(VFIOContainerBase
*bcontainer
,
469 MemoryRegionSection
*section
)
471 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(section
->mr
);
472 VFIORamDiscardListener
*vrdl
= NULL
;
474 QLIST_FOREACH(vrdl
, &bcontainer
->vrdl_list
, next
) {
475 if (vrdl
->mr
== section
->mr
&&
476 vrdl
->offset_within_address_space
==
477 section
->offset_within_address_space
) {
483 hw_error("vfio: Trying to unregister missing RAM discard listener");
486 ram_discard_manager_unregister_listener(rdm
, &vrdl
->listener
);
487 QLIST_REMOVE(vrdl
, next
);
491 static bool vfio_known_safe_misalignment(MemoryRegionSection
*section
)
493 MemoryRegion
*mr
= section
->mr
;
495 if (!TPM_IS_CRB(mr
->owner
)) {
499 /* this is a known safe misaligned region, just trace for debug purpose */
500 trace_vfio_known_safe_misalignment(memory_region_name(mr
),
501 section
->offset_within_address_space
,
502 section
->offset_within_region
,
503 qemu_real_host_page_size());
507 static bool vfio_listener_valid_section(MemoryRegionSection
*section
,
510 if (vfio_listener_skipped_section(section
)) {
511 trace_vfio_listener_region_skip(name
,
512 section
->offset_within_address_space
,
513 section
->offset_within_address_space
+
514 int128_get64(int128_sub(section
->size
, int128_one())));
518 if (unlikely((section
->offset_within_address_space
&
519 ~qemu_real_host_page_mask()) !=
520 (section
->offset_within_region
& ~qemu_real_host_page_mask()))) {
521 if (!vfio_known_safe_misalignment(section
)) {
522 error_report("%s received unaligned region %s iova=0x%"PRIx64
523 " offset_within_region=0x%"PRIx64
524 " qemu_real_host_page_size=0x%"PRIxPTR
,
525 __func__
, memory_region_name(section
->mr
),
526 section
->offset_within_address_space
,
527 section
->offset_within_region
,
528 qemu_real_host_page_size());
536 static bool vfio_get_section_iova_range(VFIOContainerBase
*bcontainer
,
537 MemoryRegionSection
*section
,
538 hwaddr
*out_iova
, hwaddr
*out_end
,
544 iova
= REAL_HOST_PAGE_ALIGN(section
->offset_within_address_space
);
545 llend
= int128_make64(section
->offset_within_address_space
);
546 llend
= int128_add(llend
, section
->size
);
547 llend
= int128_and(llend
, int128_exts64(qemu_real_host_page_mask()));
549 if (int128_ge(int128_make64(iova
), llend
)) {
554 *out_end
= int128_get64(int128_sub(llend
, int128_one()));
561 static void vfio_listener_region_add(MemoryListener
*listener
,
562 MemoryRegionSection
*section
)
564 VFIOContainerBase
*bcontainer
= container_of(listener
, VFIOContainerBase
,
567 Int128 llend
, llsize
;
572 if (!vfio_listener_valid_section(section
, "region_add")) {
576 if (!vfio_get_section_iova_range(bcontainer
, section
, &iova
, &end
,
578 if (memory_region_is_ram_device(section
->mr
)) {
579 trace_vfio_listener_region_add_no_dma_map(
580 memory_region_name(section
->mr
),
581 section
->offset_within_address_space
,
582 int128_getlo(section
->size
),
583 qemu_real_host_page_size());
588 if (vfio_container_add_section_window(bcontainer
, section
, &err
)) {
592 memory_region_ref(section
->mr
);
594 if (memory_region_is_iommu(section
->mr
)) {
595 VFIOGuestIOMMU
*giommu
;
596 IOMMUMemoryRegion
*iommu_mr
= IOMMU_MEMORY_REGION(section
->mr
);
599 trace_vfio_listener_region_add_iommu(iova
, end
);
601 * FIXME: For VFIO iommu types which have KVM acceleration to
602 * avoid bouncing all map/unmaps through qemu this way, this
603 * would be the right place to wire that up (tell the KVM
604 * device emulation the VFIO iommu handles to use).
606 giommu
= g_malloc0(sizeof(*giommu
));
607 giommu
->iommu_mr
= iommu_mr
;
608 giommu
->iommu_offset
= section
->offset_within_address_space
-
609 section
->offset_within_region
;
610 giommu
->bcontainer
= bcontainer
;
611 llend
= int128_add(int128_make64(section
->offset_within_region
),
613 llend
= int128_sub(llend
, int128_one());
614 iommu_idx
= memory_region_iommu_attrs_to_index(iommu_mr
,
615 MEMTXATTRS_UNSPECIFIED
);
616 iommu_notifier_init(&giommu
->n
, vfio_iommu_map_notify
,
617 IOMMU_NOTIFIER_IOTLB_EVENTS
,
618 section
->offset_within_region
,
622 ret
= memory_region_iommu_set_page_size_mask(giommu
->iommu_mr
,
630 if (bcontainer
->iova_ranges
) {
631 ret
= memory_region_iommu_set_iova_ranges(giommu
->iommu_mr
,
632 bcontainer
->iova_ranges
,
640 ret
= memory_region_register_iommu_notifier(section
->mr
, &giommu
->n
,
646 QLIST_INSERT_HEAD(&bcontainer
->giommu_list
, giommu
, giommu_next
);
647 memory_region_iommu_replay(giommu
->iommu_mr
, &giommu
->n
);
652 /* Here we assume that memory_region_is_ram(section->mr)==true */
655 * For RAM memory regions with a RamDiscardManager, we only want to map the
656 * actually populated parts - and update the mapping whenever we're notified
659 if (memory_region_has_ram_discard_manager(section
->mr
)) {
660 vfio_register_ram_discard_listener(bcontainer
, section
);
664 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
665 section
->offset_within_region
+
666 (iova
- section
->offset_within_address_space
);
668 trace_vfio_listener_region_add_ram(iova
, end
, vaddr
);
670 llsize
= int128_sub(llend
, int128_make64(iova
));
672 if (memory_region_is_ram_device(section
->mr
)) {
673 hwaddr pgmask
= (1ULL << ctz64(bcontainer
->pgsizes
)) - 1;
675 if ((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
)) {
676 trace_vfio_listener_region_add_no_dma_map(
677 memory_region_name(section
->mr
),
678 section
->offset_within_address_space
,
679 int128_getlo(section
->size
),
685 ret
= vfio_container_dma_map(bcontainer
, iova
, int128_get64(llsize
),
686 vaddr
, section
->readonly
);
688 error_setg(&err
, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx
", "
689 "0x%"HWADDR_PRIx
", %p) = %d (%s)",
690 bcontainer
, iova
, int128_get64(llsize
), vaddr
, ret
,
692 if (memory_region_is_ram_device(section
->mr
)) {
693 /* Allow unexpected mappings not to be fatal for RAM devices */
694 error_report_err(err
);
703 if (memory_region_is_ram_device(section
->mr
)) {
704 error_reportf_err(err
, "PCI p2p may not work: ");
708 * On the initfn path, store the first error in the container so we
709 * can gracefully fail. Runtime, there's not much we can do other
710 * than throw a hardware error.
712 if (!bcontainer
->initialized
) {
713 if (!bcontainer
->error
) {
714 error_propagate_prepend(&bcontainer
->error
, err
,
716 memory_region_name(section
->mr
));
721 error_report_err(err
);
722 hw_error("vfio: DMA mapping failed, unable to continue");
726 static void vfio_listener_region_del(MemoryListener
*listener
,
727 MemoryRegionSection
*section
)
729 VFIOContainerBase
*bcontainer
= container_of(listener
, VFIOContainerBase
,
732 Int128 llend
, llsize
;
734 bool try_unmap
= true;
736 if (!vfio_listener_valid_section(section
, "region_del")) {
740 if (memory_region_is_iommu(section
->mr
)) {
741 VFIOGuestIOMMU
*giommu
;
743 QLIST_FOREACH(giommu
, &bcontainer
->giommu_list
, giommu_next
) {
744 if (MEMORY_REGION(giommu
->iommu_mr
) == section
->mr
&&
745 giommu
->n
.start
== section
->offset_within_region
) {
746 memory_region_unregister_iommu_notifier(section
->mr
,
748 QLIST_REMOVE(giommu
, giommu_next
);
755 * FIXME: We assume the one big unmap below is adequate to
756 * remove any individual page mappings in the IOMMU which
757 * might have been copied into VFIO. This works for a page table
758 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
759 * That may not be true for all IOMMU types.
763 if (!vfio_get_section_iova_range(bcontainer
, section
, &iova
, &end
,
768 llsize
= int128_sub(llend
, int128_make64(iova
));
770 trace_vfio_listener_region_del(iova
, end
);
772 if (memory_region_is_ram_device(section
->mr
)) {
775 pgmask
= (1ULL << ctz64(bcontainer
->pgsizes
)) - 1;
776 try_unmap
= !((iova
& pgmask
) || (int128_get64(llsize
) & pgmask
));
777 } else if (memory_region_has_ram_discard_manager(section
->mr
)) {
778 vfio_unregister_ram_discard_listener(bcontainer
, section
);
779 /* Unregistering will trigger an unmap. */
784 if (int128_eq(llsize
, int128_2_64())) {
785 /* The unmap ioctl doesn't accept a full 64-bit span. */
786 llsize
= int128_rshift(llsize
, 1);
787 ret
= vfio_container_dma_unmap(bcontainer
, iova
,
788 int128_get64(llsize
), NULL
);
790 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
791 "0x%"HWADDR_PRIx
") = %d (%s)",
792 bcontainer
, iova
, int128_get64(llsize
), ret
,
795 iova
+= int128_get64(llsize
);
797 ret
= vfio_container_dma_unmap(bcontainer
, iova
,
798 int128_get64(llsize
), NULL
);
800 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
801 "0x%"HWADDR_PRIx
") = %d (%s)",
802 bcontainer
, iova
, int128_get64(llsize
), ret
,
807 memory_region_unref(section
->mr
);
809 vfio_container_del_section_window(bcontainer
, section
);
812 typedef struct VFIODirtyRanges
{
821 typedef struct VFIODirtyRangesListener
{
822 VFIOContainerBase
*bcontainer
;
823 VFIODirtyRanges ranges
;
824 MemoryListener listener
;
825 } VFIODirtyRangesListener
;
827 static bool vfio_section_is_vfio_pci(MemoryRegionSection
*section
,
828 VFIOContainerBase
*bcontainer
)
830 VFIOPCIDevice
*pcidev
;
831 VFIODevice
*vbasedev
;
834 owner
= memory_region_owner(section
->mr
);
836 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
837 if (vbasedev
->type
!= VFIO_DEVICE_TYPE_PCI
) {
840 pcidev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
841 if (OBJECT(pcidev
) == owner
) {
849 static void vfio_dirty_tracking_update(MemoryListener
*listener
,
850 MemoryRegionSection
*section
)
852 VFIODirtyRangesListener
*dirty
= container_of(listener
,
853 VFIODirtyRangesListener
,
855 VFIODirtyRanges
*range
= &dirty
->ranges
;
856 hwaddr iova
, end
, *min
, *max
;
858 if (!vfio_listener_valid_section(section
, "tracking_update") ||
859 !vfio_get_section_iova_range(dirty
->bcontainer
, section
,
860 &iova
, &end
, NULL
)) {
865 * The address space passed to the dirty tracker is reduced to three ranges:
866 * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
869 * The underlying reports of dirty will query a sub-interval of each of
872 * The purpose of the three range handling is to handle known cases of big
873 * holes in the address space, like the x86 AMD 1T hole, and firmware (like
874 * OVMF) which may relocate the pci-hole64 to the end of the address space.
875 * The latter would otherwise generate large ranges for tracking, stressing
876 * the limits of supported hardware. The pci-hole32 will always be below 4G
877 * (overlapping or not) so it doesn't need special handling and is part of
880 * The alternative would be an IOVATree but that has a much bigger runtime
881 * overhead and unnecessary complexity.
883 if (vfio_section_is_vfio_pci(section
, dirty
->bcontainer
) &&
884 iova
>= UINT32_MAX
) {
885 min
= &range
->minpci64
;
886 max
= &range
->maxpci64
;
888 min
= (end
<= UINT32_MAX
) ? &range
->min32
: &range
->min64
;
889 max
= (end
<= UINT32_MAX
) ? &range
->max32
: &range
->max64
;
898 trace_vfio_device_dirty_tracking_update(iova
, end
, *min
, *max
);
902 static const MemoryListener vfio_dirty_tracking_listener
= {
903 .name
= "vfio-tracking",
904 .region_add
= vfio_dirty_tracking_update
,
907 static void vfio_dirty_tracking_init(VFIOContainerBase
*bcontainer
,
908 VFIODirtyRanges
*ranges
)
910 VFIODirtyRangesListener dirty
;
912 memset(&dirty
, 0, sizeof(dirty
));
913 dirty
.ranges
.min32
= UINT32_MAX
;
914 dirty
.ranges
.min64
= UINT64_MAX
;
915 dirty
.ranges
.minpci64
= UINT64_MAX
;
916 dirty
.listener
= vfio_dirty_tracking_listener
;
917 dirty
.bcontainer
= bcontainer
;
919 memory_listener_register(&dirty
.listener
,
920 bcontainer
->space
->as
);
922 *ranges
= dirty
.ranges
;
925 * The memory listener is synchronous, and used to calculate the range
926 * to dirty tracking. Unregister it after we are done as we are not
927 * interested in any follow-up updates.
929 memory_listener_unregister(&dirty
.listener
);
932 static void vfio_devices_dma_logging_stop(VFIOContainerBase
*bcontainer
)
934 uint64_t buf
[DIV_ROUND_UP(sizeof(struct vfio_device_feature
),
935 sizeof(uint64_t))] = {};
936 struct vfio_device_feature
*feature
= (struct vfio_device_feature
*)buf
;
937 VFIODevice
*vbasedev
;
939 feature
->argsz
= sizeof(buf
);
940 feature
->flags
= VFIO_DEVICE_FEATURE_SET
|
941 VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP
;
943 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
944 if (!vbasedev
->dirty_tracking
) {
948 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_FEATURE
, feature
)) {
949 warn_report("%s: Failed to stop DMA logging, err %d (%s)",
950 vbasedev
->name
, -errno
, strerror(errno
));
952 vbasedev
->dirty_tracking
= false;
956 static struct vfio_device_feature
*
957 vfio_device_feature_dma_logging_start_create(VFIOContainerBase
*bcontainer
,
958 VFIODirtyRanges
*tracking
)
960 struct vfio_device_feature
*feature
;
962 struct vfio_device_feature_dma_logging_control
*control
;
963 struct vfio_device_feature_dma_logging_range
*ranges
;
965 feature_size
= sizeof(struct vfio_device_feature
) +
966 sizeof(struct vfio_device_feature_dma_logging_control
);
967 feature
= g_try_malloc0(feature_size
);
972 feature
->argsz
= feature_size
;
973 feature
->flags
= VFIO_DEVICE_FEATURE_SET
|
974 VFIO_DEVICE_FEATURE_DMA_LOGGING_START
;
976 control
= (struct vfio_device_feature_dma_logging_control
*)feature
->data
;
977 control
->page_size
= qemu_real_host_page_size();
980 * DMA logging uAPI guarantees to support at least a number of ranges that
981 * fits into a single host kernel base page.
983 control
->num_ranges
= !!tracking
->max32
+ !!tracking
->max64
+
984 !!tracking
->maxpci64
;
985 ranges
= g_try_new0(struct vfio_device_feature_dma_logging_range
,
986 control
->num_ranges
);
994 control
->ranges
= (uintptr_t)ranges
;
995 if (tracking
->max32
) {
996 ranges
->iova
= tracking
->min32
;
997 ranges
->length
= (tracking
->max32
- tracking
->min32
) + 1;
1000 if (tracking
->max64
) {
1001 ranges
->iova
= tracking
->min64
;
1002 ranges
->length
= (tracking
->max64
- tracking
->min64
) + 1;
1005 if (tracking
->maxpci64
) {
1006 ranges
->iova
= tracking
->minpci64
;
1007 ranges
->length
= (tracking
->maxpci64
- tracking
->minpci64
) + 1;
1010 trace_vfio_device_dirty_tracking_start(control
->num_ranges
,
1011 tracking
->min32
, tracking
->max32
,
1012 tracking
->min64
, tracking
->max64
,
1013 tracking
->minpci64
, tracking
->maxpci64
);
1018 static void vfio_device_feature_dma_logging_start_destroy(
1019 struct vfio_device_feature
*feature
)
1021 struct vfio_device_feature_dma_logging_control
*control
=
1022 (struct vfio_device_feature_dma_logging_control
*)feature
->data
;
1023 struct vfio_device_feature_dma_logging_range
*ranges
=
1024 (struct vfio_device_feature_dma_logging_range
*)(uintptr_t)control
->ranges
;
1030 static int vfio_devices_dma_logging_start(VFIOContainerBase
*bcontainer
)
1032 struct vfio_device_feature
*feature
;
1033 VFIODirtyRanges ranges
;
1034 VFIODevice
*vbasedev
;
1037 vfio_dirty_tracking_init(bcontainer
, &ranges
);
1038 feature
= vfio_device_feature_dma_logging_start_create(bcontainer
,
1044 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
1045 if (vbasedev
->dirty_tracking
) {
1049 ret
= ioctl(vbasedev
->fd
, VFIO_DEVICE_FEATURE
, feature
);
1052 error_report("%s: Failed to start DMA logging, err %d (%s)",
1053 vbasedev
->name
, ret
, strerror(errno
));
1056 vbasedev
->dirty_tracking
= true;
1061 vfio_devices_dma_logging_stop(bcontainer
);
1064 vfio_device_feature_dma_logging_start_destroy(feature
);
1069 static void vfio_listener_log_global_start(MemoryListener
*listener
)
1071 VFIOContainerBase
*bcontainer
= container_of(listener
, VFIOContainerBase
,
1075 if (vfio_devices_all_device_dirty_tracking(bcontainer
)) {
1076 ret
= vfio_devices_dma_logging_start(bcontainer
);
1078 ret
= vfio_container_set_dirty_page_tracking(bcontainer
, true);
1082 error_report("vfio: Could not start dirty page tracking, err: %d (%s)",
1083 ret
, strerror(-ret
));
1084 vfio_set_migration_error(ret
);
1088 static void vfio_listener_log_global_stop(MemoryListener
*listener
)
1090 VFIOContainerBase
*bcontainer
= container_of(listener
, VFIOContainerBase
,
1094 if (vfio_devices_all_device_dirty_tracking(bcontainer
)) {
1095 vfio_devices_dma_logging_stop(bcontainer
);
1097 ret
= vfio_container_set_dirty_page_tracking(bcontainer
, false);
1101 error_report("vfio: Could not stop dirty page tracking, err: %d (%s)",
1102 ret
, strerror(-ret
));
1103 vfio_set_migration_error(ret
);
1107 static int vfio_device_dma_logging_report(VFIODevice
*vbasedev
, hwaddr iova
,
1108 hwaddr size
, void *bitmap
)
1110 uint64_t buf
[DIV_ROUND_UP(sizeof(struct vfio_device_feature
) +
1111 sizeof(struct vfio_device_feature_dma_logging_report
),
1112 sizeof(uint64_t))] = {};
1113 struct vfio_device_feature
*feature
= (struct vfio_device_feature
*)buf
;
1114 struct vfio_device_feature_dma_logging_report
*report
=
1115 (struct vfio_device_feature_dma_logging_report
*)feature
->data
;
1117 report
->iova
= iova
;
1118 report
->length
= size
;
1119 report
->page_size
= qemu_real_host_page_size();
1120 report
->bitmap
= (uintptr_t)bitmap
;
1122 feature
->argsz
= sizeof(buf
);
1123 feature
->flags
= VFIO_DEVICE_FEATURE_GET
|
1124 VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT
;
1126 if (ioctl(vbasedev
->fd
, VFIO_DEVICE_FEATURE
, feature
)) {
1133 int vfio_devices_query_dirty_bitmap(const VFIOContainerBase
*bcontainer
,
1134 VFIOBitmap
*vbmap
, hwaddr iova
,
1137 VFIODevice
*vbasedev
;
1140 QLIST_FOREACH(vbasedev
, &bcontainer
->device_list
, container_next
) {
1141 ret
= vfio_device_dma_logging_report(vbasedev
, iova
, size
,
1144 error_report("%s: Failed to get DMA logging report, iova: "
1145 "0x%" HWADDR_PRIx
", size: 0x%" HWADDR_PRIx
1147 vbasedev
->name
, iova
, size
, ret
, strerror(-ret
));
1156 int vfio_get_dirty_bitmap(const VFIOContainerBase
*bcontainer
, uint64_t iova
,
1157 uint64_t size
, ram_addr_t ram_addr
)
1159 bool all_device_dirty_tracking
=
1160 vfio_devices_all_device_dirty_tracking(bcontainer
);
1161 uint64_t dirty_pages
;
1165 if (!bcontainer
->dirty_pages_supported
&& !all_device_dirty_tracking
) {
1166 cpu_physical_memory_set_dirty_range(ram_addr
, size
,
1167 tcg_enabled() ? DIRTY_CLIENTS_ALL
:
1168 DIRTY_CLIENTS_NOCODE
);
1172 ret
= vfio_bitmap_alloc(&vbmap
, size
);
1177 if (all_device_dirty_tracking
) {
1178 ret
= vfio_devices_query_dirty_bitmap(bcontainer
, &vbmap
, iova
, size
);
1180 ret
= vfio_container_query_dirty_bitmap(bcontainer
, &vbmap
, iova
, size
);
1187 dirty_pages
= cpu_physical_memory_set_dirty_lebitmap(vbmap
.bitmap
, ram_addr
,
1190 trace_vfio_get_dirty_bitmap(iova
, size
, vbmap
.size
, ram_addr
, dirty_pages
);
1192 g_free(vbmap
.bitmap
);
1199 VFIOGuestIOMMU
*giommu
;
1200 } vfio_giommu_dirty_notifier
;
1202 static void vfio_iommu_map_dirty_notify(IOMMUNotifier
*n
, IOMMUTLBEntry
*iotlb
)
1204 vfio_giommu_dirty_notifier
*gdn
= container_of(n
,
1205 vfio_giommu_dirty_notifier
, n
);
1206 VFIOGuestIOMMU
*giommu
= gdn
->giommu
;
1207 VFIOContainerBase
*bcontainer
= giommu
->bcontainer
;
1208 hwaddr iova
= iotlb
->iova
+ giommu
->iommu_offset
;
1209 ram_addr_t translated_addr
;
1212 trace_vfio_iommu_map_dirty_notify(iova
, iova
+ iotlb
->addr_mask
);
1214 if (iotlb
->target_as
!= &address_space_memory
) {
1215 error_report("Wrong target AS \"%s\", only system memory is allowed",
1216 iotlb
->target_as
->name
? iotlb
->target_as
->name
: "none");
1221 if (vfio_get_xlat_addr(iotlb
, NULL
, &translated_addr
, NULL
)) {
1222 ret
= vfio_get_dirty_bitmap(bcontainer
, iova
, iotlb
->addr_mask
+ 1,
1225 error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx
", "
1226 "0x%"HWADDR_PRIx
") = %d (%s)",
1227 bcontainer
, iova
, iotlb
->addr_mask
+ 1, ret
,
1235 vfio_set_migration_error(ret
);
1239 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection
*section
,
1242 const hwaddr size
= int128_get64(section
->size
);
1243 const hwaddr iova
= section
->offset_within_address_space
;
1244 const ram_addr_t ram_addr
= memory_region_get_ram_addr(section
->mr
) +
1245 section
->offset_within_region
;
1246 VFIORamDiscardListener
*vrdl
= opaque
;
1249 * Sync the whole mapped region (spanning multiple individual mappings)
1252 return vfio_get_dirty_bitmap(vrdl
->bcontainer
, iova
, size
, ram_addr
);
1256 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase
*bcontainer
,
1257 MemoryRegionSection
*section
)
1259 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(section
->mr
);
1260 VFIORamDiscardListener
*vrdl
= NULL
;
1262 QLIST_FOREACH(vrdl
, &bcontainer
->vrdl_list
, next
) {
1263 if (vrdl
->mr
== section
->mr
&&
1264 vrdl
->offset_within_address_space
==
1265 section
->offset_within_address_space
) {
1271 hw_error("vfio: Trying to sync missing RAM discard listener");
1275 * We only want/can synchronize the bitmap for actually mapped parts -
1276 * which correspond to populated parts. Replay all populated parts.
1278 return ram_discard_manager_replay_populated(rdm
, section
,
1279 vfio_ram_discard_get_dirty_bitmap
,
1283 static int vfio_sync_dirty_bitmap(VFIOContainerBase
*bcontainer
,
1284 MemoryRegionSection
*section
)
1286 ram_addr_t ram_addr
;
1288 if (memory_region_is_iommu(section
->mr
)) {
1289 VFIOGuestIOMMU
*giommu
;
1291 QLIST_FOREACH(giommu
, &bcontainer
->giommu_list
, giommu_next
) {
1292 if (MEMORY_REGION(giommu
->iommu_mr
) == section
->mr
&&
1293 giommu
->n
.start
== section
->offset_within_region
) {
1295 vfio_giommu_dirty_notifier gdn
= { .giommu
= giommu
};
1296 int idx
= memory_region_iommu_attrs_to_index(giommu
->iommu_mr
,
1297 MEMTXATTRS_UNSPECIFIED
);
1299 llend
= int128_add(int128_make64(section
->offset_within_region
),
1301 llend
= int128_sub(llend
, int128_one());
1303 iommu_notifier_init(&gdn
.n
,
1304 vfio_iommu_map_dirty_notify
,
1306 section
->offset_within_region
,
1307 int128_get64(llend
),
1309 memory_region_iommu_replay(giommu
->iommu_mr
, &gdn
.n
);
1314 } else if (memory_region_has_ram_discard_manager(section
->mr
)) {
1315 return vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer
, section
);
1318 ram_addr
= memory_region_get_ram_addr(section
->mr
) +
1319 section
->offset_within_region
;
1321 return vfio_get_dirty_bitmap(bcontainer
,
1322 REAL_HOST_PAGE_ALIGN(section
->offset_within_address_space
),
1323 int128_get64(section
->size
), ram_addr
);
1326 static void vfio_listener_log_sync(MemoryListener
*listener
,
1327 MemoryRegionSection
*section
)
1329 VFIOContainerBase
*bcontainer
= container_of(listener
, VFIOContainerBase
,
1333 if (vfio_listener_skipped_section(section
)) {
1337 if (vfio_devices_all_dirty_tracking(bcontainer
)) {
1338 ret
= vfio_sync_dirty_bitmap(bcontainer
, section
);
1340 error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret
,
1342 vfio_set_migration_error(ret
);
1347 const MemoryListener vfio_memory_listener
= {
1349 .region_add
= vfio_listener_region_add
,
1350 .region_del
= vfio_listener_region_del
,
1351 .log_global_start
= vfio_listener_log_global_start
,
1352 .log_global_stop
= vfio_listener_log_global_stop
,
1353 .log_sync
= vfio_listener_log_sync
,
1356 void vfio_reset_handler(void *opaque
)
1358 VFIODevice
*vbasedev
;
1360 QLIST_FOREACH(vbasedev
, &vfio_device_list
, global_next
) {
1361 if (vbasedev
->dev
->realized
) {
1362 vbasedev
->ops
->vfio_compute_needs_reset(vbasedev
);
1366 QLIST_FOREACH(vbasedev
, &vfio_device_list
, global_next
) {
1367 if (vbasedev
->dev
->realized
&& vbasedev
->needs_reset
) {
1368 vbasedev
->ops
->vfio_hot_reset_multi(vbasedev
);
1373 int vfio_kvm_device_add_fd(int fd
, Error
**errp
)
1376 struct kvm_device_attr attr
= {
1377 .group
= KVM_DEV_VFIO_FILE
,
1378 .attr
= KVM_DEV_VFIO_FILE_ADD
,
1379 .addr
= (uint64_t)(unsigned long)&fd
,
1382 if (!kvm_enabled()) {
1386 if (vfio_kvm_device_fd
< 0) {
1387 struct kvm_create_device cd
= {
1388 .type
= KVM_DEV_TYPE_VFIO
,
1391 if (kvm_vm_ioctl(kvm_state
, KVM_CREATE_DEVICE
, &cd
)) {
1392 error_setg_errno(errp
, errno
, "Failed to create KVM VFIO device");
1396 vfio_kvm_device_fd
= cd
.fd
;
1399 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1400 error_setg_errno(errp
, errno
, "Failed to add fd %d to KVM VFIO device",
1408 int vfio_kvm_device_del_fd(int fd
, Error
**errp
)
1411 struct kvm_device_attr attr
= {
1412 .group
= KVM_DEV_VFIO_FILE
,
1413 .attr
= KVM_DEV_VFIO_FILE_DEL
,
1414 .addr
= (uint64_t)(unsigned long)&fd
,
1417 if (vfio_kvm_device_fd
< 0) {
1418 error_setg(errp
, "KVM VFIO device isn't created yet");
1422 if (ioctl(vfio_kvm_device_fd
, KVM_SET_DEVICE_ATTR
, &attr
)) {
1423 error_setg_errno(errp
, errno
,
1424 "Failed to remove fd %d from KVM VFIO device", fd
);
1431 VFIOAddressSpace
*vfio_get_address_space(AddressSpace
*as
)
1433 VFIOAddressSpace
*space
;
1435 QLIST_FOREACH(space
, &vfio_address_spaces
, list
) {
1436 if (space
->as
== as
) {
1441 /* No suitable VFIOAddressSpace, create a new one */
1442 space
= g_malloc0(sizeof(*space
));
1444 QLIST_INIT(&space
->containers
);
1446 if (QLIST_EMPTY(&vfio_address_spaces
)) {
1447 qemu_register_reset(vfio_reset_handler
, NULL
);
1450 QLIST_INSERT_HEAD(&vfio_address_spaces
, space
, list
);
1455 void vfio_put_address_space(VFIOAddressSpace
*space
)
1457 if (!QLIST_EMPTY(&space
->containers
)) {
1461 QLIST_REMOVE(space
, list
);
1464 if (QLIST_EMPTY(&vfio_address_spaces
)) {
1465 qemu_unregister_reset(vfio_reset_handler
, NULL
);
1469 struct vfio_device_info
*vfio_get_device_info(int fd
)
1471 struct vfio_device_info
*info
;
1472 uint32_t argsz
= sizeof(*info
);
1474 info
= g_malloc0(argsz
);
1477 info
->argsz
= argsz
;
1479 if (ioctl(fd
, VFIO_DEVICE_GET_INFO
, info
)) {
1484 if (info
->argsz
> argsz
) {
1485 argsz
= info
->argsz
;
1486 info
= g_realloc(info
, argsz
);
1493 int vfio_attach_device(char *name
, VFIODevice
*vbasedev
,
1494 AddressSpace
*as
, Error
**errp
)
1496 const VFIOIOMMUClass
*ops
=
1497 VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_LEGACY
));
1499 if (vbasedev
->iommufd
) {
1500 ops
= VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD
));
1505 return ops
->attach_device(name
, vbasedev
, as
, errp
);
1508 void vfio_detach_device(VFIODevice
*vbasedev
)
1510 if (!vbasedev
->bcontainer
) {
1513 vbasedev
->bcontainer
->ops
->detach_device(vbasedev
);