4 * Copyright(c) 2017-2018 Intel Corporation.
5 * Copyright(c) 2020 Red Hat, Inc.
7 * This work is licensed under the terms of the GNU GPL, version 2 or later.
8 * See the COPYING file in the top-level directory.
12 #include "qemu/osdep.h"
14 #include "hw/virtio/virtio-net.h"
15 #include "net/vhost_net.h"
16 #include "net/vhost-vdpa.h"
17 #include "hw/virtio/vhost-vdpa.h"
18 #include "qemu/config-file.h"
19 #include "qemu/error-report.h"
21 #include "qemu/memalign.h"
22 #include "qemu/option.h"
23 #include "qapi/error.h"
24 #include <linux/vhost.h>
25 #include <sys/ioctl.h>
27 #include "standard-headers/linux/virtio_net.h"
28 #include "monitor/monitor.h"
29 #include "migration/migration.h"
30 #include "migration/misc.h"
31 #include "hw/virtio/vhost.h"
33 /* Todo:need to add the multiqueue support here */
34 typedef struct VhostVDPAState
{
36 struct vhost_vdpa vhost_vdpa
;
37 Notifier migration_state
;
38 VHostNetState
*vhost_net
;
40 /* Control commands shadow buffers */
41 void *cvq_cmd_out_buffer
;
42 virtio_net_ctrl_ack
*status
;
44 /* The device always have SVQ enabled */
47 /* The device can isolate CVQ in its own ASID */
54 * The array is sorted alphabetically in ascending order,
55 * with the exception of VHOST_INVALID_FEATURE_BIT,
56 * which should always be the last entry.
58 const int vdpa_feature_bits
[] = {
60 VIRTIO_F_IOMMU_PLATFORM
,
61 VIRTIO_F_NOTIFY_ON_EMPTY
,
66 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
,
67 VIRTIO_NET_F_CTRL_MAC_ADDR
,
69 VIRTIO_NET_F_CTRL_RX_EXTRA
,
70 VIRTIO_NET_F_CTRL_VLAN
,
73 VIRTIO_NET_F_GUEST_CSUM
,
74 VIRTIO_NET_F_GUEST_ECN
,
75 VIRTIO_NET_F_GUEST_TSO4
,
76 VIRTIO_NET_F_GUEST_TSO6
,
77 VIRTIO_NET_F_GUEST_UFO
,
78 VIRTIO_NET_F_GUEST_USO4
,
79 VIRTIO_NET_F_GUEST_USO6
,
80 VIRTIO_NET_F_HASH_REPORT
,
81 VIRTIO_NET_F_HOST_ECN
,
82 VIRTIO_NET_F_HOST_TSO4
,
83 VIRTIO_NET_F_HOST_TSO6
,
84 VIRTIO_NET_F_HOST_UFO
,
85 VIRTIO_NET_F_HOST_USO
,
87 VIRTIO_NET_F_MRG_RXBUF
,
91 VIRTIO_RING_F_EVENT_IDX
,
92 VIRTIO_RING_F_INDIRECT_DESC
,
94 /* VHOST_INVALID_FEATURE_BIT should always be the last entry */
95 VHOST_INVALID_FEATURE_BIT
98 /** Supported device specific feature bits with SVQ */
99 static const uint64_t vdpa_svq_device_features
=
100 BIT_ULL(VIRTIO_NET_F_CSUM
) |
101 BIT_ULL(VIRTIO_NET_F_GUEST_CSUM
) |
102 BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
) |
103 BIT_ULL(VIRTIO_NET_F_MTU
) |
104 BIT_ULL(VIRTIO_NET_F_MAC
) |
105 BIT_ULL(VIRTIO_NET_F_GUEST_TSO4
) |
106 BIT_ULL(VIRTIO_NET_F_GUEST_TSO6
) |
107 BIT_ULL(VIRTIO_NET_F_GUEST_ECN
) |
108 BIT_ULL(VIRTIO_NET_F_GUEST_UFO
) |
109 BIT_ULL(VIRTIO_NET_F_HOST_TSO4
) |
110 BIT_ULL(VIRTIO_NET_F_HOST_TSO6
) |
111 BIT_ULL(VIRTIO_NET_F_HOST_ECN
) |
112 BIT_ULL(VIRTIO_NET_F_HOST_UFO
) |
113 BIT_ULL(VIRTIO_NET_F_MRG_RXBUF
) |
114 BIT_ULL(VIRTIO_NET_F_STATUS
) |
115 BIT_ULL(VIRTIO_NET_F_CTRL_VQ
) |
116 BIT_ULL(VIRTIO_NET_F_CTRL_RX
) |
117 BIT_ULL(VIRTIO_NET_F_CTRL_VLAN
) |
118 BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA
) |
119 BIT_ULL(VIRTIO_NET_F_MQ
) |
120 BIT_ULL(VIRTIO_F_ANY_LAYOUT
) |
121 BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR
) |
122 /* VHOST_F_LOG_ALL is exposed by SVQ */
123 BIT_ULL(VHOST_F_LOG_ALL
) |
124 BIT_ULL(VIRTIO_NET_F_RSC_EXT
) |
125 BIT_ULL(VIRTIO_NET_F_STANDBY
) |
126 BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX
);
128 #define VHOST_VDPA_NET_CVQ_ASID 1
130 VHostNetState
*vhost_vdpa_get_vhost_net(NetClientState
*nc
)
132 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
133 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
137 static size_t vhost_vdpa_net_cvq_cmd_len(void)
140 * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
141 * In buffer is always 1 byte, so it should fit here
143 return sizeof(struct virtio_net_ctrl_hdr
) +
144 2 * sizeof(struct virtio_net_ctrl_mac
) +
145 MAC_TABLE_ENTRIES
* ETH_ALEN
;
148 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
150 return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
153 static bool vhost_vdpa_net_valid_svq_features(uint64_t features
, Error
**errp
)
155 uint64_t invalid_dev_features
=
156 features
& ~vdpa_svq_device_features
&
157 /* Transport are all accepted at this point */
158 ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START
,
159 VIRTIO_TRANSPORT_F_END
- VIRTIO_TRANSPORT_F_START
);
161 if (invalid_dev_features
) {
162 error_setg(errp
, "vdpa svq does not work with features 0x%" PRIx64
,
163 invalid_dev_features
);
167 return vhost_svq_valid_features(features
, errp
);
170 static int vhost_vdpa_net_check_device_id(struct vhost_net
*net
)
174 struct vhost_dev
*hdev
;
176 hdev
= (struct vhost_dev
*)&net
->dev
;
177 ret
= hdev
->vhost_ops
->vhost_get_device_id(hdev
, &device_id
);
178 if (device_id
!= VIRTIO_ID_NET
) {
184 static int vhost_vdpa_add(NetClientState
*ncs
, void *be
,
185 int queue_pair_index
, int nvqs
)
187 VhostNetOptions options
;
188 struct vhost_net
*net
= NULL
;
192 options
.backend_type
= VHOST_BACKEND_TYPE_VDPA
;
193 assert(ncs
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
194 s
= DO_UPCAST(VhostVDPAState
, nc
, ncs
);
195 options
.net_backend
= ncs
;
197 options
.busyloop_timeout
= 0;
200 net
= vhost_net_init(&options
);
202 error_report("failed to init vhost_net for queue");
206 ret
= vhost_vdpa_net_check_device_id(net
);
212 vhost_net_cleanup(net
);
218 static void vhost_vdpa_cleanup(NetClientState
*nc
)
220 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
223 * If a peer NIC is attached, do not cleanup anything.
224 * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup()
225 * when the guest is shutting down.
227 if (nc
->peer
&& nc
->peer
->info
->type
== NET_CLIENT_DRIVER_NIC
) {
230 munmap(s
->cvq_cmd_out_buffer
, vhost_vdpa_net_cvq_cmd_page_len());
231 munmap(s
->status
, vhost_vdpa_net_cvq_cmd_page_len());
233 vhost_net_cleanup(s
->vhost_net
);
234 g_free(s
->vhost_net
);
237 if (s
->vhost_vdpa
.device_fd
>= 0) {
238 qemu_close(s
->vhost_vdpa
.device_fd
);
239 s
->vhost_vdpa
.device_fd
= -1;
243 static bool vhost_vdpa_has_vnet_hdr(NetClientState
*nc
)
245 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
250 static bool vhost_vdpa_has_ufo(NetClientState
*nc
)
252 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
253 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
254 uint64_t features
= 0;
255 features
|= (1ULL << VIRTIO_NET_F_HOST_UFO
);
256 features
= vhost_net_get_features(s
->vhost_net
, features
);
257 return !!(features
& (1ULL << VIRTIO_NET_F_HOST_UFO
));
261 static bool vhost_vdpa_check_peer_type(NetClientState
*nc
, ObjectClass
*oc
,
264 const char *driver
= object_class_get_name(oc
);
266 if (!g_str_has_prefix(driver
, "virtio-net-")) {
267 error_setg(errp
, "vhost-vdpa requires frontend driver virtio-net-*");
274 /** Dummy receive in case qemu falls back to userland tap networking */
275 static ssize_t
vhost_vdpa_receive(NetClientState
*nc
, const uint8_t *buf
,
281 /** From any vdpa net client, get the netclient of the first queue pair */
282 static VhostVDPAState
*vhost_vdpa_net_first_nc_vdpa(VhostVDPAState
*s
)
284 NICState
*nic
= qemu_get_nic(s
->nc
.peer
);
285 NetClientState
*nc0
= qemu_get_peer(nic
->ncs
, 0);
287 return DO_UPCAST(VhostVDPAState
, nc
, nc0
);
290 static void vhost_vdpa_net_log_global_enable(VhostVDPAState
*s
, bool enable
)
292 struct vhost_vdpa
*v
= &s
->vhost_vdpa
;
295 int data_queue_pairs
, cvq
, r
;
297 /* We are only called on the first data vqs and only if x-svq is not set */
298 if (s
->vhost_vdpa
.shadow_vqs_enabled
== enable
) {
303 n
= VIRTIO_NET(vdev
);
304 if (!n
->vhost_started
) {
308 data_queue_pairs
= n
->multiqueue
? n
->max_queue_pairs
: 1;
309 cvq
= virtio_vdev_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) ?
310 n
->max_ncs
- n
->max_queue_pairs
: 0;
312 * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter
313 * in the future and resume the device if read-only operations between
314 * suspend and reset goes wrong.
316 vhost_net_stop(vdev
, n
->nic
->ncs
, data_queue_pairs
, cvq
);
318 /* Start will check migration setup_or_active to configure or not SVQ */
319 r
= vhost_net_start(vdev
, n
->nic
->ncs
, data_queue_pairs
, cvq
);
320 if (unlikely(r
< 0)) {
321 error_report("unable to start vhost net: %s(%d)", g_strerror(-r
), -r
);
325 static void vdpa_net_migration_state_notifier(Notifier
*notifier
, void *data
)
327 MigrationState
*migration
= data
;
328 VhostVDPAState
*s
= container_of(notifier
, VhostVDPAState
,
331 if (migration_in_setup(migration
)) {
332 vhost_vdpa_net_log_global_enable(s
, true);
333 } else if (migration_has_failed(migration
)) {
334 vhost_vdpa_net_log_global_enable(s
, false);
338 static void vhost_vdpa_net_data_start_first(VhostVDPAState
*s
)
340 struct vhost_vdpa
*v
= &s
->vhost_vdpa
;
342 add_migration_state_change_notifier(&s
->migration_state
);
343 if (v
->shadow_vqs_enabled
) {
344 v
->iova_tree
= vhost_iova_tree_new(v
->iova_range
.first
,
349 static int vhost_vdpa_net_data_start(NetClientState
*nc
)
351 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
352 struct vhost_vdpa
*v
= &s
->vhost_vdpa
;
354 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
357 migration_is_setup_or_active(migrate_get_current()->state
)) {
358 v
->shadow_vqs_enabled
= true;
359 v
->shadow_data
= true;
361 v
->shadow_vqs_enabled
= false;
362 v
->shadow_data
= false;
366 vhost_vdpa_net_data_start_first(s
);
370 if (v
->shadow_vqs_enabled
) {
371 VhostVDPAState
*s0
= vhost_vdpa_net_first_nc_vdpa(s
);
372 v
->iova_tree
= s0
->vhost_vdpa
.iova_tree
;
378 static void vhost_vdpa_net_client_stop(NetClientState
*nc
)
380 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
381 struct vhost_dev
*dev
;
383 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
385 if (s
->vhost_vdpa
.index
== 0) {
386 remove_migration_state_change_notifier(&s
->migration_state
);
389 dev
= s
->vhost_vdpa
.dev
;
390 if (dev
->vq_index
+ dev
->nvqs
== dev
->vq_index_end
) {
391 g_clear_pointer(&s
->vhost_vdpa
.iova_tree
, vhost_iova_tree_delete
);
395 static NetClientInfo net_vhost_vdpa_info
= {
396 .type
= NET_CLIENT_DRIVER_VHOST_VDPA
,
397 .size
= sizeof(VhostVDPAState
),
398 .receive
= vhost_vdpa_receive
,
399 .start
= vhost_vdpa_net_data_start
,
400 .stop
= vhost_vdpa_net_client_stop
,
401 .cleanup
= vhost_vdpa_cleanup
,
402 .has_vnet_hdr
= vhost_vdpa_has_vnet_hdr
,
403 .has_ufo
= vhost_vdpa_has_ufo
,
404 .check_peer_type
= vhost_vdpa_check_peer_type
,
407 static int64_t vhost_vdpa_get_vring_group(int device_fd
, unsigned vq_index
,
410 struct vhost_vring_state state
= {
413 int r
= ioctl(device_fd
, VHOST_VDPA_GET_VRING_GROUP
, &state
);
415 if (unlikely(r
< 0)) {
417 error_setg_errno(errp
, errno
, "Cannot get VQ %u group", vq_index
);
424 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa
*v
,
428 struct vhost_vring_state asid
= {
434 r
= ioctl(v
->device_fd
, VHOST_VDPA_SET_GROUP_ASID
, &asid
);
435 if (unlikely(r
< 0)) {
436 error_report("Can't set vq group %u asid %u, errno=%d (%s)",
437 asid
.index
, asid
.num
, errno
, g_strerror(errno
));
442 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa
*v
, void *addr
)
444 VhostIOVATree
*tree
= v
->iova_tree
;
447 * No need to specify size or to look for more translations since
448 * this contiguous chunk was allocated by us.
450 .translated_addr
= (hwaddr
)(uintptr_t)addr
,
452 const DMAMap
*map
= vhost_iova_tree_find_iova(tree
, &needle
);
455 if (unlikely(!map
)) {
456 error_report("Cannot locate expected map");
460 r
= vhost_vdpa_dma_unmap(v
, v
->address_space_id
, map
->iova
, map
->size
+ 1);
461 if (unlikely(r
!= 0)) {
462 error_report("Device cannot unmap: %s(%d)", g_strerror(r
), r
);
465 vhost_iova_tree_remove(tree
, *map
);
468 /** Map CVQ buffer. */
469 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa
*v
, void *buf
, size_t size
,
475 map
.translated_addr
= (hwaddr
)(uintptr_t)buf
;
477 map
.perm
= write
? IOMMU_RW
: IOMMU_RO
,
478 r
= vhost_iova_tree_map_alloc(v
->iova_tree
, &map
);
479 if (unlikely(r
!= IOVA_OK
)) {
480 error_report("Cannot map injected element");
484 r
= vhost_vdpa_dma_map(v
, v
->address_space_id
, map
.iova
,
485 vhost_vdpa_net_cvq_cmd_page_len(), buf
, !write
);
486 if (unlikely(r
< 0)) {
493 vhost_iova_tree_remove(v
->iova_tree
, map
);
497 static int vhost_vdpa_net_cvq_start(NetClientState
*nc
)
499 VhostVDPAState
*s
, *s0
;
500 struct vhost_vdpa
*v
;
505 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
507 s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
510 s0
= vhost_vdpa_net_first_nc_vdpa(s
);
511 v
->shadow_data
= s0
->vhost_vdpa
.shadow_vqs_enabled
;
512 v
->shadow_vqs_enabled
= s
->always_svq
;
513 s
->vhost_vdpa
.address_space_id
= VHOST_VDPA_GUEST_PA_ASID
;
515 if (s
->vhost_vdpa
.shadow_data
) {
516 /* SVQ is already configured for all virtqueues */
521 * If we early return in these cases SVQ will not be enabled. The migration
522 * will be blocked as long as vhost-vdpa backends will not offer _F_LOG.
524 if (!vhost_vdpa_net_valid_svq_features(v
->dev
->features
, NULL
)) {
528 if (!s
->cvq_isolated
) {
532 cvq_group
= vhost_vdpa_get_vring_group(v
->device_fd
,
533 v
->dev
->vq_index_end
- 1,
535 if (unlikely(cvq_group
< 0)) {
536 error_report_err(err
);
540 r
= vhost_vdpa_set_address_space_id(v
, cvq_group
, VHOST_VDPA_NET_CVQ_ASID
);
541 if (unlikely(r
< 0)) {
545 v
->shadow_vqs_enabled
= true;
546 s
->vhost_vdpa
.address_space_id
= VHOST_VDPA_NET_CVQ_ASID
;
549 if (!s
->vhost_vdpa
.shadow_vqs_enabled
) {
553 if (s0
->vhost_vdpa
.iova_tree
) {
555 * SVQ is already configured for all virtqueues. Reuse IOVA tree for
556 * simplicity, whether CVQ shares ASID with guest or not, because:
557 * - Memory listener need access to guest's memory addresses allocated
559 * - There should be plenty of IOVA address space for both ASID not to
560 * worry about collisions between them. Guest's translations are
561 * still validated with virtio virtqueue_pop so there is no risk for
562 * the guest to access memory that it shouldn't.
564 * To allocate a iova tree per ASID is doable but it complicates the
565 * code and it is not worth it for the moment.
567 v
->iova_tree
= s0
->vhost_vdpa
.iova_tree
;
569 v
->iova_tree
= vhost_iova_tree_new(v
->iova_range
.first
,
573 r
= vhost_vdpa_cvq_map_buf(&s
->vhost_vdpa
, s
->cvq_cmd_out_buffer
,
574 vhost_vdpa_net_cvq_cmd_page_len(), false);
575 if (unlikely(r
< 0)) {
579 r
= vhost_vdpa_cvq_map_buf(&s
->vhost_vdpa
, s
->status
,
580 vhost_vdpa_net_cvq_cmd_page_len(), true);
581 if (unlikely(r
< 0)) {
582 vhost_vdpa_cvq_unmap_buf(&s
->vhost_vdpa
, s
->cvq_cmd_out_buffer
);
588 static void vhost_vdpa_net_cvq_stop(NetClientState
*nc
)
590 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
592 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
594 if (s
->vhost_vdpa
.shadow_vqs_enabled
) {
595 vhost_vdpa_cvq_unmap_buf(&s
->vhost_vdpa
, s
->cvq_cmd_out_buffer
);
596 vhost_vdpa_cvq_unmap_buf(&s
->vhost_vdpa
, s
->status
);
599 vhost_vdpa_net_client_stop(nc
);
602 static ssize_t
vhost_vdpa_net_cvq_add(VhostVDPAState
*s
, size_t out_len
,
605 /* Buffers for the device */
606 const struct iovec out
= {
607 .iov_base
= s
->cvq_cmd_out_buffer
,
610 const struct iovec in
= {
611 .iov_base
= s
->status
,
612 .iov_len
= sizeof(virtio_net_ctrl_ack
),
614 VhostShadowVirtqueue
*svq
= g_ptr_array_index(s
->vhost_vdpa
.shadow_vqs
, 0);
617 r
= vhost_svq_add(svq
, &out
, 1, &in
, 1, NULL
);
618 if (unlikely(r
!= 0)) {
619 if (unlikely(r
== -ENOSPC
)) {
620 qemu_log_mask(LOG_GUEST_ERROR
, "%s: No space on device queue\n",
627 * We can poll here since we've had BQL from the time we sent the
628 * descriptor. Also, we need to take the answer before SVQ pulls by itself,
629 * when BQL is released
631 return vhost_svq_poll(svq
);
634 static ssize_t
vhost_vdpa_net_load_cmd(VhostVDPAState
*s
, uint8_t class,
635 uint8_t cmd
, const struct iovec
*data_sg
,
638 const struct virtio_net_ctrl_hdr ctrl
= {
642 size_t data_size
= iov_size(data_sg
, data_num
);
644 assert(data_size
< vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl
));
646 /* pack the CVQ command header */
647 memcpy(s
->cvq_cmd_out_buffer
, &ctrl
, sizeof(ctrl
));
649 /* pack the CVQ command command-specific-data */
650 iov_to_buf(data_sg
, data_num
, 0,
651 s
->cvq_cmd_out_buffer
+ sizeof(ctrl
), data_size
);
653 return vhost_vdpa_net_cvq_add(s
, data_size
+ sizeof(ctrl
),
654 sizeof(virtio_net_ctrl_ack
));
657 static int vhost_vdpa_net_load_mac(VhostVDPAState
*s
, const VirtIONet
*n
)
659 if (virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
660 const struct iovec data
= {
661 .iov_base
= (void *)n
->mac
,
662 .iov_len
= sizeof(n
->mac
),
664 ssize_t dev_written
= vhost_vdpa_net_load_cmd(s
, VIRTIO_NET_CTRL_MAC
,
665 VIRTIO_NET_CTRL_MAC_ADDR_SET
,
667 if (unlikely(dev_written
< 0)) {
670 if (*s
->status
!= VIRTIO_NET_OK
) {
676 * According to VirtIO standard, "The device MUST have an
677 * empty MAC filtering table on reset.".
679 * Therefore, there is no need to send this CVQ command if the
680 * driver also sets an empty MAC filter table, which aligns with
681 * the device's defaults.
683 * Note that the device's defaults can mismatch the driver's
684 * configuration only at live migration.
686 if (!virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_CTRL_RX
) ||
687 n
->mac_table
.in_use
== 0) {
691 uint32_t uni_entries
= n
->mac_table
.first_multi
,
692 uni_macs_size
= uni_entries
* ETH_ALEN
,
693 mul_entries
= n
->mac_table
.in_use
- uni_entries
,
694 mul_macs_size
= mul_entries
* ETH_ALEN
;
695 struct virtio_net_ctrl_mac uni
= {
696 .entries
= cpu_to_le32(uni_entries
),
698 struct virtio_net_ctrl_mac mul
= {
699 .entries
= cpu_to_le32(mul_entries
),
701 const struct iovec data
[] = {
704 .iov_len
= sizeof(uni
),
706 .iov_base
= n
->mac_table
.macs
,
707 .iov_len
= uni_macs_size
,
710 .iov_len
= sizeof(mul
),
712 .iov_base
= &n
->mac_table
.macs
[uni_macs_size
],
713 .iov_len
= mul_macs_size
,
716 ssize_t dev_written
= vhost_vdpa_net_load_cmd(s
,
718 VIRTIO_NET_CTRL_MAC_TABLE_SET
,
719 data
, ARRAY_SIZE(data
));
720 if (unlikely(dev_written
< 0)) {
723 if (*s
->status
!= VIRTIO_NET_OK
) {
730 static int vhost_vdpa_net_load_mq(VhostVDPAState
*s
,
733 struct virtio_net_ctrl_mq mq
;
736 if (!virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_MQ
)) {
740 mq
.virtqueue_pairs
= cpu_to_le16(n
->curr_queue_pairs
);
741 const struct iovec data
= {
743 .iov_len
= sizeof(mq
),
745 dev_written
= vhost_vdpa_net_load_cmd(s
, VIRTIO_NET_CTRL_MQ
,
746 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
,
748 if (unlikely(dev_written
< 0)) {
751 if (*s
->status
!= VIRTIO_NET_OK
) {
758 static int vhost_vdpa_net_load_offloads(VhostVDPAState
*s
,
764 if (!virtio_vdev_has_feature(&n
->parent_obj
,
765 VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)) {
769 if (n
->curr_guest_offloads
== virtio_net_supported_guest_offloads(n
)) {
771 * According to VirtIO standard, "Upon feature negotiation
772 * corresponding offload gets enabled to preserve
773 * backward compatibility.".
775 * Therefore, there is no need to send this CVQ command if the
776 * driver also enables all supported offloads, which aligns with
777 * the device's defaults.
779 * Note that the device's defaults can mismatch the driver's
780 * configuration only at live migration.
785 offloads
= cpu_to_le64(n
->curr_guest_offloads
);
786 const struct iovec data
= {
787 .iov_base
= &offloads
,
788 .iov_len
= sizeof(offloads
),
790 dev_written
= vhost_vdpa_net_load_cmd(s
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
791 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
,
793 if (unlikely(dev_written
< 0)) {
796 if (*s
->status
!= VIRTIO_NET_OK
) {
803 static int vhost_vdpa_net_load_rx_mode(VhostVDPAState
*s
,
807 const struct iovec data
= {
809 .iov_len
= sizeof(on
),
811 return vhost_vdpa_net_load_cmd(s
, VIRTIO_NET_CTRL_RX
,
815 static int vhost_vdpa_net_load_rx(VhostVDPAState
*s
,
820 if (!virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_CTRL_RX
)) {
825 * According to virtio_net_reset(), device turns promiscuous mode
828 * Additionally, according to VirtIO standard, "Since there are
829 * no guarantees, it can use a hash filter or silently switch to
830 * allmulti or promiscuous mode if it is given too many addresses.".
831 * QEMU marks `n->mac_table.uni_overflow` if guest sets too many
832 * non-multicast MAC addresses, indicating that promiscuous mode
835 * Therefore, QEMU should only send this CVQ command if the
836 * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off,
837 * which sets promiscuous mode on, different from the device's defaults.
839 * Note that the device's defaults can mismatch the driver's
840 * configuration only at live migration.
842 if (!n
->mac_table
.uni_overflow
&& !n
->promisc
) {
843 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
844 VIRTIO_NET_CTRL_RX_PROMISC
, 0);
845 if (unlikely(dev_written
< 0)) {
848 if (*s
->status
!= VIRTIO_NET_OK
) {
854 * According to virtio_net_reset(), device turns all-multicast mode
857 * According to VirtIO standard, "Since there are no guarantees,
858 * it can use a hash filter or silently switch to allmulti or
859 * promiscuous mode if it is given too many addresses.". QEMU marks
860 * `n->mac_table.multi_overflow` if guest sets too many
861 * non-multicast MAC addresses.
863 * Therefore, QEMU should only send this CVQ command if the
864 * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on,
865 * which sets all-multicast mode on, different from the device's defaults.
867 * Note that the device's defaults can mismatch the driver's
868 * configuration only at live migration.
870 if (n
->mac_table
.multi_overflow
|| n
->allmulti
) {
871 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
872 VIRTIO_NET_CTRL_RX_ALLMULTI
, 1);
873 if (unlikely(dev_written
< 0)) {
876 if (*s
->status
!= VIRTIO_NET_OK
) {
881 if (!virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_CTRL_RX_EXTRA
)) {
886 * According to virtio_net_reset(), device turns all-unicast mode
889 * Therefore, QEMU should only send this CVQ command if the driver
890 * sets all-unicast mode on, different from the device's defaults.
892 * Note that the device's defaults can mismatch the driver's
893 * configuration only at live migration.
896 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
897 VIRTIO_NET_CTRL_RX_ALLUNI
, 1);
898 if (dev_written
< 0) {
901 if (*s
->status
!= VIRTIO_NET_OK
) {
907 * According to virtio_net_reset(), device turns non-multicast mode
910 * Therefore, QEMU should only send this CVQ command if the driver
911 * sets non-multicast mode on, different from the device's defaults.
913 * Note that the device's defaults can mismatch the driver's
914 * configuration only at live migration.
917 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
918 VIRTIO_NET_CTRL_RX_NOMULTI
, 1);
919 if (dev_written
< 0) {
922 if (*s
->status
!= VIRTIO_NET_OK
) {
928 * According to virtio_net_reset(), device turns non-unicast mode
931 * Therefore, QEMU should only send this CVQ command if the driver
932 * sets non-unicast mode on, different from the device's defaults.
934 * Note that the device's defaults can mismatch the driver's
935 * configuration only at live migration.
938 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
939 VIRTIO_NET_CTRL_RX_NOUNI
, 1);
940 if (dev_written
< 0) {
943 if (*s
->status
!= VIRTIO_NET_OK
) {
949 * According to virtio_net_reset(), device turns non-broadcast mode
952 * Therefore, QEMU should only send this CVQ command if the driver
953 * sets non-broadcast mode on, different from the device's defaults.
955 * Note that the device's defaults can mismatch the driver's
956 * configuration only at live migration.
959 dev_written
= vhost_vdpa_net_load_rx_mode(s
,
960 VIRTIO_NET_CTRL_RX_NOBCAST
, 1);
961 if (dev_written
< 0) {
964 if (*s
->status
!= VIRTIO_NET_OK
) {
972 static int vhost_vdpa_net_load_single_vlan(VhostVDPAState
*s
,
976 const struct iovec data
= {
978 .iov_len
= sizeof(vid
),
980 ssize_t dev_written
= vhost_vdpa_net_load_cmd(s
, VIRTIO_NET_CTRL_VLAN
,
981 VIRTIO_NET_CTRL_VLAN_ADD
,
983 if (unlikely(dev_written
< 0)) {
986 if (unlikely(*s
->status
!= VIRTIO_NET_OK
)) {
993 static int vhost_vdpa_net_load_vlan(VhostVDPAState
*s
,
998 if (!virtio_vdev_has_feature(&n
->parent_obj
, VIRTIO_NET_F_CTRL_VLAN
)) {
1002 for (int i
= 0; i
< MAX_VLAN
>> 5; i
++) {
1003 for (int j
= 0; n
->vlans
[i
] && j
<= 0x1f; j
++) {
1004 if (n
->vlans
[i
] & (1U << j
)) {
1005 r
= vhost_vdpa_net_load_single_vlan(s
, n
, (i
<< 5) + j
);
1006 if (unlikely(r
!= 0)) {
1016 static int vhost_vdpa_net_load(NetClientState
*nc
)
1018 VhostVDPAState
*s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
1019 struct vhost_vdpa
*v
= &s
->vhost_vdpa
;
1023 assert(nc
->info
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
1025 if (!v
->shadow_vqs_enabled
) {
1029 n
= VIRTIO_NET(v
->dev
->vdev
);
1030 r
= vhost_vdpa_net_load_mac(s
, n
);
1031 if (unlikely(r
< 0)) {
1034 r
= vhost_vdpa_net_load_mq(s
, n
);
1038 r
= vhost_vdpa_net_load_offloads(s
, n
);
1042 r
= vhost_vdpa_net_load_rx(s
, n
);
1046 r
= vhost_vdpa_net_load_vlan(s
, n
);
1054 static NetClientInfo net_vhost_vdpa_cvq_info
= {
1055 .type
= NET_CLIENT_DRIVER_VHOST_VDPA
,
1056 .size
= sizeof(VhostVDPAState
),
1057 .receive
= vhost_vdpa_receive
,
1058 .start
= vhost_vdpa_net_cvq_start
,
1059 .load
= vhost_vdpa_net_load
,
1060 .stop
= vhost_vdpa_net_cvq_stop
,
1061 .cleanup
= vhost_vdpa_cleanup
,
1062 .has_vnet_hdr
= vhost_vdpa_has_vnet_hdr
,
1063 .has_ufo
= vhost_vdpa_has_ufo
,
1064 .check_peer_type
= vhost_vdpa_check_peer_type
,
1068 * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to
1071 * Considering that QEMU cannot send the entire filter table to the
1072 * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ
1073 * command to enable promiscuous mode to receive all packets,
1074 * according to VirtIO standard, "Since there are no guarantees,
1075 * it can use a hash filter or silently switch to allmulti or
1076 * promiscuous mode if it is given too many addresses.".
1078 * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and
1079 * marks `n->mac_table.x_overflow` accordingly, it should have
1080 * the same effect on the device model to receive
1081 * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses.
1082 * The same applies to multicast MAC addresses.
1084 * Therefore, QEMU can provide the device model with a fake
1085 * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1)
1086 * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast
1087 * MAC addresses. This ensures that the device model marks
1088 * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`,
1089 * allowing all packets to be received, which aligns with the
1090 * state of the vdpa device.
1092 static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState
*s
,
1093 VirtQueueElement
*elem
,
1096 struct virtio_net_ctrl_mac mac_data
, *mac_ptr
;
1097 struct virtio_net_ctrl_hdr
*hdr_ptr
;
1101 /* parse the non-multicast MAC address entries from CVQ command */
1102 cursor
= sizeof(*hdr_ptr
);
1103 r
= iov_to_buf(elem
->out_sg
, elem
->out_num
, cursor
,
1104 &mac_data
, sizeof(mac_data
));
1105 if (unlikely(r
!= sizeof(mac_data
))) {
1107 * If the CVQ command is invalid, we should simulate the vdpa device
1108 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1110 *s
->status
= VIRTIO_NET_ERR
;
1111 return sizeof(*s
->status
);
1113 cursor
+= sizeof(mac_data
) + le32_to_cpu(mac_data
.entries
) * ETH_ALEN
;
1115 /* parse the multicast MAC address entries from CVQ command */
1116 r
= iov_to_buf(elem
->out_sg
, elem
->out_num
, cursor
,
1117 &mac_data
, sizeof(mac_data
));
1118 if (r
!= sizeof(mac_data
)) {
1120 * If the CVQ command is invalid, we should simulate the vdpa device
1121 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1123 *s
->status
= VIRTIO_NET_ERR
;
1124 return sizeof(*s
->status
);
1126 cursor
+= sizeof(mac_data
) + le32_to_cpu(mac_data
.entries
) * ETH_ALEN
;
1128 /* validate the CVQ command */
1129 if (iov_size(elem
->out_sg
, elem
->out_num
) != cursor
) {
1131 * If the CVQ command is invalid, we should simulate the vdpa device
1132 * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1134 *s
->status
= VIRTIO_NET_ERR
;
1135 return sizeof(*s
->status
);
1139 * According to VirtIO standard, "Since there are no guarantees,
1140 * it can use a hash filter or silently switch to allmulti or
1141 * promiscuous mode if it is given too many addresses.".
1143 * Therefore, considering that QEMU is unable to send the entire
1144 * filter table to the vdpa device, it should send the
1145 * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode
1147 r
= vhost_vdpa_net_load_rx_mode(s
, VIRTIO_NET_CTRL_RX_PROMISC
, 1);
1148 if (unlikely(r
< 0)) {
1151 if (*s
->status
!= VIRTIO_NET_OK
) {
1152 return sizeof(*s
->status
);
1156 * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ
1157 * command to the device model, including (`MAC_TABLE_ENTRIES` + 1)
1158 * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1)
1159 * multicast MAC addresses.
1161 * By doing so, the device model can mark `n->mac_table.uni_overflow`
1162 * and `n->mac_table.multi_overflow`, enabling all packets to be
1163 * received, which aligns with the state of the vdpa device.
1166 uint32_t fake_uni_entries
= MAC_TABLE_ENTRIES
+ 1,
1167 fake_mul_entries
= MAC_TABLE_ENTRIES
+ 1,
1168 fake_cvq_size
= sizeof(struct virtio_net_ctrl_hdr
) +
1169 sizeof(mac_data
) + fake_uni_entries
* ETH_ALEN
+
1170 sizeof(mac_data
) + fake_mul_entries
* ETH_ALEN
;
1172 assert(fake_cvq_size
< vhost_vdpa_net_cvq_cmd_page_len());
1173 out
->iov_len
= fake_cvq_size
;
1175 /* pack the header for fake CVQ command */
1176 hdr_ptr
= out
->iov_base
+ cursor
;
1177 hdr_ptr
->class = VIRTIO_NET_CTRL_MAC
;
1178 hdr_ptr
->cmd
= VIRTIO_NET_CTRL_MAC_TABLE_SET
;
1179 cursor
+= sizeof(*hdr_ptr
);
1182 * Pack the non-multicast MAC addresses part for fake CVQ command.
1184 * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1185 * addresses provided in CVQ command. Therefore, only the entries
1186 * field need to be prepared in the CVQ command.
1188 mac_ptr
= out
->iov_base
+ cursor
;
1189 mac_ptr
->entries
= cpu_to_le32(fake_uni_entries
);
1190 cursor
+= sizeof(*mac_ptr
) + fake_uni_entries
* ETH_ALEN
;
1193 * Pack the multicast MAC addresses part for fake CVQ command.
1195 * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1196 * addresses provided in CVQ command. Therefore, only the entries
1197 * field need to be prepared in the CVQ command.
1199 mac_ptr
= out
->iov_base
+ cursor
;
1200 mac_ptr
->entries
= cpu_to_le32(fake_mul_entries
);
1203 * Simulating QEMU poll a vdpa device used buffer
1204 * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1206 return sizeof(*s
->status
);
1210 * Validate and copy control virtqueue commands.
1212 * Following QEMU guidelines, we offer a copy of the buffers to the device to
1213 * prevent TOCTOU bugs.
1215 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue
*svq
,
1216 VirtQueueElement
*elem
,
1219 VhostVDPAState
*s
= opaque
;
1221 const struct virtio_net_ctrl_hdr
*ctrl
;
1222 virtio_net_ctrl_ack status
= VIRTIO_NET_ERR
;
1223 /* Out buffer sent to both the vdpa device and the device model */
1224 struct iovec out
= {
1225 .iov_base
= s
->cvq_cmd_out_buffer
,
1227 /* in buffer used for device model */
1228 const struct iovec in
= {
1229 .iov_base
= &status
,
1230 .iov_len
= sizeof(status
),
1232 ssize_t dev_written
= -EINVAL
;
1234 out
.iov_len
= iov_to_buf(elem
->out_sg
, elem
->out_num
, 0,
1235 s
->cvq_cmd_out_buffer
,
1236 vhost_vdpa_net_cvq_cmd_page_len());
1238 ctrl
= s
->cvq_cmd_out_buffer
;
1239 if (ctrl
->class == VIRTIO_NET_CTRL_ANNOUNCE
) {
1241 * Guest announce capability is emulated by qemu, so don't forward to
1244 dev_written
= sizeof(status
);
1245 *s
->status
= VIRTIO_NET_OK
;
1246 } else if (unlikely(ctrl
->class == VIRTIO_NET_CTRL_MAC
&&
1247 ctrl
->cmd
== VIRTIO_NET_CTRL_MAC_TABLE_SET
&&
1248 iov_size(elem
->out_sg
, elem
->out_num
) > out
.iov_len
)) {
1250 * Due to the size limitation of the out buffer sent to the vdpa device,
1251 * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive
1252 * MAC addresses set by the driver for the filter table can cause
1253 * truncation of the CVQ command in QEMU. As a result, the vdpa device
1254 * rejects the flawed CVQ command.
1256 * Therefore, QEMU must handle this situation instead of sending
1257 * the CVQ command directly.
1259 dev_written
= vhost_vdpa_net_excessive_mac_filter_cvq_add(s
, elem
,
1261 if (unlikely(dev_written
< 0)) {
1265 dev_written
= vhost_vdpa_net_cvq_add(s
, out
.iov_len
, sizeof(status
));
1266 if (unlikely(dev_written
< 0)) {
1271 if (unlikely(dev_written
< sizeof(status
))) {
1272 error_report("Insufficient written data (%zu)", dev_written
);
1276 if (*s
->status
!= VIRTIO_NET_OK
) {
1280 status
= VIRTIO_NET_ERR
;
1281 virtio_net_handle_ctrl_iov(svq
->vdev
, &in
, 1, &out
, 1);
1282 if (status
!= VIRTIO_NET_OK
) {
1283 error_report("Bad CVQ processing in model");
1287 in_len
= iov_from_buf(elem
->in_sg
, elem
->in_num
, 0, &status
,
1289 if (unlikely(in_len
< sizeof(status
))) {
1290 error_report("Bad device CVQ written length");
1292 vhost_svq_push_elem(svq
, elem
, MIN(in_len
, sizeof(status
)));
1294 * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when
1295 * the function successfully forwards the CVQ command, indicated
1296 * by a non-negative value of `dev_written`. Otherwise, it still
1298 * This function should only free the `elem` when it owns.
1300 if (dev_written
>= 0) {
1303 return dev_written
< 0 ? dev_written
: 0;
1306 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops
= {
1307 .avail_handler
= vhost_vdpa_net_handle_ctrl_avail
,
1311 * Probe if CVQ is isolated
1313 * @device_fd The vdpa device fd
1314 * @features Features offered by the device.
1315 * @cvq_index The control vq pair index
1317 * Returns <0 in case of failure, 0 if false and 1 if true.
1319 static int vhost_vdpa_probe_cvq_isolation(int device_fd
, uint64_t features
,
1320 int cvq_index
, Error
**errp
)
1322 uint64_t backend_features
;
1324 uint8_t status
= VIRTIO_CONFIG_S_ACKNOWLEDGE
|
1325 VIRTIO_CONFIG_S_DRIVER
|
1326 VIRTIO_CONFIG_S_FEATURES_OK
;
1331 r
= ioctl(device_fd
, VHOST_GET_BACKEND_FEATURES
, &backend_features
);
1332 if (unlikely(r
< 0)) {
1333 error_setg_errno(errp
, errno
, "Cannot get vdpa backend_features");
1337 if (!(backend_features
& BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID
))) {
1341 r
= ioctl(device_fd
, VHOST_SET_FEATURES
, &features
);
1343 error_setg_errno(errp
, errno
, "Cannot set features");
1346 r
= ioctl(device_fd
, VHOST_VDPA_SET_STATUS
, &status
);
1348 error_setg_errno(errp
, -r
, "Cannot set device features");
1352 cvq_group
= vhost_vdpa_get_vring_group(device_fd
, cvq_index
, errp
);
1353 if (unlikely(cvq_group
< 0)) {
1354 if (cvq_group
!= -ENOTSUP
) {
1360 * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend
1361 * support ASID even if the parent driver does not. The CVQ cannot be
1362 * isolated in this case.
1370 for (int i
= 0; i
< cvq_index
; ++i
) {
1371 int64_t group
= vhost_vdpa_get_vring_group(device_fd
, i
, errp
);
1372 if (unlikely(group
< 0)) {
1377 if (group
== (int64_t)cvq_group
) {
1387 ioctl(device_fd
, VHOST_VDPA_SET_STATUS
, &status
);
1391 static NetClientState
*net_vhost_vdpa_init(NetClientState
*peer
,
1395 int queue_pair_index
,
1399 struct vhost_vdpa_iova_range iova_range
,
1403 NetClientState
*nc
= NULL
;
1410 nc
= qemu_new_net_client(&net_vhost_vdpa_info
, peer
, device
,
1413 cvq_isolated
= vhost_vdpa_probe_cvq_isolation(vdpa_device_fd
, features
,
1414 queue_pair_index
* 2,
1416 if (unlikely(cvq_isolated
< 0)) {
1420 nc
= qemu_new_net_control_client(&net_vhost_vdpa_cvq_info
, peer
,
1423 qemu_set_info_str(nc
, TYPE_VHOST_VDPA
);
1424 s
= DO_UPCAST(VhostVDPAState
, nc
, nc
);
1426 s
->vhost_vdpa
.device_fd
= vdpa_device_fd
;
1427 s
->vhost_vdpa
.index
= queue_pair_index
;
1428 s
->always_svq
= svq
;
1429 s
->migration_state
.notify
= vdpa_net_migration_state_notifier
;
1430 s
->vhost_vdpa
.shadow_vqs_enabled
= svq
;
1431 s
->vhost_vdpa
.iova_range
= iova_range
;
1432 s
->vhost_vdpa
.shadow_data
= svq
;
1433 if (queue_pair_index
== 0) {
1434 vhost_vdpa_net_valid_svq_features(features
,
1435 &s
->vhost_vdpa
.migration_blocker
);
1436 } else if (!is_datapath
) {
1437 s
->cvq_cmd_out_buffer
= mmap(NULL
, vhost_vdpa_net_cvq_cmd_page_len(),
1438 PROT_READ
| PROT_WRITE
,
1439 MAP_SHARED
| MAP_ANONYMOUS
, -1, 0);
1440 s
->status
= mmap(NULL
, vhost_vdpa_net_cvq_cmd_page_len(),
1441 PROT_READ
| PROT_WRITE
, MAP_SHARED
| MAP_ANONYMOUS
,
1444 s
->vhost_vdpa
.shadow_vq_ops
= &vhost_vdpa_net_svq_ops
;
1445 s
->vhost_vdpa
.shadow_vq_ops_opaque
= s
;
1446 s
->cvq_isolated
= cvq_isolated
;
1449 * TODO: We cannot migrate devices with CVQ and no x-svq enabled as
1450 * there is no way to set the device state (MAC, MQ, etc) before
1451 * starting the datapath.
1453 * Migration blocker ownership now belongs to s->vhost_vdpa.
1456 error_setg(&s
->vhost_vdpa
.migration_blocker
,
1457 "net vdpa cannot migrate with CVQ feature");
1460 ret
= vhost_vdpa_add(nc
, (void *)&s
->vhost_vdpa
, queue_pair_index
, nvqs
);
1462 qemu_del_net_client(nc
);
1468 static int vhost_vdpa_get_features(int fd
, uint64_t *features
, Error
**errp
)
1470 int ret
= ioctl(fd
, VHOST_GET_FEATURES
, features
);
1471 if (unlikely(ret
< 0)) {
1472 error_setg_errno(errp
, errno
,
1473 "Fail to query features from vhost-vDPA device");
1478 static int vhost_vdpa_get_max_queue_pairs(int fd
, uint64_t features
,
1479 int *has_cvq
, Error
**errp
)
1481 unsigned long config_size
= offsetof(struct vhost_vdpa_config
, buf
);
1482 g_autofree
struct vhost_vdpa_config
*config
= NULL
;
1483 __virtio16
*max_queue_pairs
;
1486 if (features
& (1 << VIRTIO_NET_F_CTRL_VQ
)) {
1492 if (features
& (1 << VIRTIO_NET_F_MQ
)) {
1493 config
= g_malloc0(config_size
+ sizeof(*max_queue_pairs
));
1494 config
->off
= offsetof(struct virtio_net_config
, max_virtqueue_pairs
);
1495 config
->len
= sizeof(*max_queue_pairs
);
1497 ret
= ioctl(fd
, VHOST_VDPA_GET_CONFIG
, config
);
1499 error_setg(errp
, "Fail to get config from vhost-vDPA device");
1503 max_queue_pairs
= (__virtio16
*)&config
->buf
;
1505 return lduw_le_p(max_queue_pairs
);
1511 int net_init_vhost_vdpa(const Netdev
*netdev
, const char *name
,
1512 NetClientState
*peer
, Error
**errp
)
1514 const NetdevVhostVDPAOptions
*opts
;
1517 g_autofree NetClientState
**ncs
= NULL
;
1518 struct vhost_vdpa_iova_range iova_range
;
1520 int queue_pairs
, r
, i
= 0, has_cvq
= 0;
1522 assert(netdev
->type
== NET_CLIENT_DRIVER_VHOST_VDPA
);
1523 opts
= &netdev
->u
.vhost_vdpa
;
1524 if (!opts
->vhostdev
&& !opts
->vhostfd
) {
1526 "vhost-vdpa: neither vhostdev= nor vhostfd= was specified");
1530 if (opts
->vhostdev
&& opts
->vhostfd
) {
1532 "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive");
1536 if (opts
->vhostdev
) {
1537 vdpa_device_fd
= qemu_open(opts
->vhostdev
, O_RDWR
, errp
);
1538 if (vdpa_device_fd
== -1) {
1543 vdpa_device_fd
= monitor_fd_param(monitor_cur(), opts
->vhostfd
, errp
);
1544 if (vdpa_device_fd
== -1) {
1545 error_prepend(errp
, "vhost-vdpa: unable to parse vhostfd: ");
1550 r
= vhost_vdpa_get_features(vdpa_device_fd
, &features
, errp
);
1551 if (unlikely(r
< 0)) {
1555 queue_pairs
= vhost_vdpa_get_max_queue_pairs(vdpa_device_fd
, features
,
1557 if (queue_pairs
< 0) {
1558 qemu_close(vdpa_device_fd
);
1562 r
= vhost_vdpa_get_iova_range(vdpa_device_fd
, &iova_range
);
1563 if (unlikely(r
< 0)) {
1564 error_setg(errp
, "vhost-vdpa: get iova range failed: %s",
1569 if (opts
->x_svq
&& !vhost_vdpa_net_valid_svq_features(features
, errp
)) {
1573 ncs
= g_malloc0(sizeof(*ncs
) * queue_pairs
);
1575 for (i
= 0; i
< queue_pairs
; i
++) {
1576 ncs
[i
] = net_vhost_vdpa_init(peer
, TYPE_VHOST_VDPA
, name
,
1577 vdpa_device_fd
, i
, 2, true, opts
->x_svq
,
1578 iova_range
, features
, errp
);
1584 nc
= net_vhost_vdpa_init(peer
, TYPE_VHOST_VDPA
, name
,
1585 vdpa_device_fd
, i
, 1, false,
1586 opts
->x_svq
, iova_range
, features
, errp
);
1595 for (i
--; i
>= 0; i
--) {
1596 qemu_del_net_client(ncs
[i
]);
1600 qemu_close(vdpa_device_fd
);