xref: /openbmc/qemu/net/vhost-vdpa.c (revision c0ce7b4a)
1 /*
2  * vhost-vdpa.c
3  *
4  * Copyright(c) 2017-2018 Intel Corporation.
5  * Copyright(c) 2020 Red Hat, Inc.
6  *
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.
9  *
10  */
11 
12 #include "qemu/osdep.h"
13 #include "clients.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"
20 #include "qemu/log.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>
26 #include <err.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"
32 
33 /* Todo:need to add the multiqueue support here */
34 typedef struct VhostVDPAState {
35     NetClientState nc;
36     struct vhost_vdpa vhost_vdpa;
37     Notifier migration_state;
38     VHostNetState *vhost_net;
39 
40     /* Control commands shadow buffers */
41     void *cvq_cmd_out_buffer;
42     virtio_net_ctrl_ack *status;
43 
44     /* The device always have SVQ enabled */
45     bool always_svq;
46 
47     /* The device can isolate CVQ in its own ASID */
48     bool cvq_isolated;
49 
50     bool started;
51 } VhostVDPAState;
52 
53 /*
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.
57  */
58 const int vdpa_feature_bits[] = {
59     VIRTIO_F_ANY_LAYOUT,
60     VIRTIO_F_IOMMU_PLATFORM,
61     VIRTIO_F_NOTIFY_ON_EMPTY,
62     VIRTIO_F_RING_PACKED,
63     VIRTIO_F_RING_RESET,
64     VIRTIO_F_VERSION_1,
65     VIRTIO_NET_F_CSUM,
66     VIRTIO_NET_F_CTRL_GUEST_OFFLOADS,
67     VIRTIO_NET_F_CTRL_MAC_ADDR,
68     VIRTIO_NET_F_CTRL_RX,
69     VIRTIO_NET_F_CTRL_RX_EXTRA,
70     VIRTIO_NET_F_CTRL_VLAN,
71     VIRTIO_NET_F_CTRL_VQ,
72     VIRTIO_NET_F_GSO,
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_HASH_REPORT,
79     VIRTIO_NET_F_HOST_ECN,
80     VIRTIO_NET_F_HOST_TSO4,
81     VIRTIO_NET_F_HOST_TSO6,
82     VIRTIO_NET_F_HOST_UFO,
83     VIRTIO_NET_F_MQ,
84     VIRTIO_NET_F_MRG_RXBUF,
85     VIRTIO_NET_F_MTU,
86     VIRTIO_NET_F_RSS,
87     VIRTIO_NET_F_STATUS,
88     VIRTIO_RING_F_EVENT_IDX,
89     VIRTIO_RING_F_INDIRECT_DESC,
90 
91     /* VHOST_INVALID_FEATURE_BIT should always be the last entry */
92     VHOST_INVALID_FEATURE_BIT
93 };
94 
95 /** Supported device specific feature bits with SVQ */
96 static const uint64_t vdpa_svq_device_features =
97     BIT_ULL(VIRTIO_NET_F_CSUM) |
98     BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) |
99     BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) |
100     BIT_ULL(VIRTIO_NET_F_MTU) |
101     BIT_ULL(VIRTIO_NET_F_MAC) |
102     BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) |
103     BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) |
104     BIT_ULL(VIRTIO_NET_F_GUEST_ECN) |
105     BIT_ULL(VIRTIO_NET_F_GUEST_UFO) |
106     BIT_ULL(VIRTIO_NET_F_HOST_TSO4) |
107     BIT_ULL(VIRTIO_NET_F_HOST_TSO6) |
108     BIT_ULL(VIRTIO_NET_F_HOST_ECN) |
109     BIT_ULL(VIRTIO_NET_F_HOST_UFO) |
110     BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) |
111     BIT_ULL(VIRTIO_NET_F_STATUS) |
112     BIT_ULL(VIRTIO_NET_F_CTRL_VQ) |
113     BIT_ULL(VIRTIO_NET_F_CTRL_RX) |
114     BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) |
115     BIT_ULL(VIRTIO_NET_F_MQ) |
116     BIT_ULL(VIRTIO_F_ANY_LAYOUT) |
117     BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) |
118     /* VHOST_F_LOG_ALL is exposed by SVQ */
119     BIT_ULL(VHOST_F_LOG_ALL) |
120     BIT_ULL(VIRTIO_NET_F_RSC_EXT) |
121     BIT_ULL(VIRTIO_NET_F_STANDBY) |
122     BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX);
123 
124 #define VHOST_VDPA_NET_CVQ_ASID 1
125 
126 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc)
127 {
128     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
129     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
130     return s->vhost_net;
131 }
132 
133 static size_t vhost_vdpa_net_cvq_cmd_len(void)
134 {
135     /*
136      * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
137      * In buffer is always 1 byte, so it should fit here
138      */
139     return sizeof(struct virtio_net_ctrl_hdr) +
140            2 * sizeof(struct virtio_net_ctrl_mac) +
141            MAC_TABLE_ENTRIES * ETH_ALEN;
142 }
143 
144 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
145 {
146     return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
147 }
148 
149 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp)
150 {
151     uint64_t invalid_dev_features =
152         features & ~vdpa_svq_device_features &
153         /* Transport are all accepted at this point */
154         ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START,
155                          VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START);
156 
157     if (invalid_dev_features) {
158         error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64,
159                    invalid_dev_features);
160         return false;
161     }
162 
163     return vhost_svq_valid_features(features, errp);
164 }
165 
166 static int vhost_vdpa_net_check_device_id(struct vhost_net *net)
167 {
168     uint32_t device_id;
169     int ret;
170     struct vhost_dev *hdev;
171 
172     hdev = (struct vhost_dev *)&net->dev;
173     ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id);
174     if (device_id != VIRTIO_ID_NET) {
175         return -ENOTSUP;
176     }
177     return ret;
178 }
179 
180 static int vhost_vdpa_add(NetClientState *ncs, void *be,
181                           int queue_pair_index, int nvqs)
182 {
183     VhostNetOptions options;
184     struct vhost_net *net = NULL;
185     VhostVDPAState *s;
186     int ret;
187 
188     options.backend_type = VHOST_BACKEND_TYPE_VDPA;
189     assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
190     s = DO_UPCAST(VhostVDPAState, nc, ncs);
191     options.net_backend = ncs;
192     options.opaque      = be;
193     options.busyloop_timeout = 0;
194     options.nvqs = nvqs;
195 
196     net = vhost_net_init(&options);
197     if (!net) {
198         error_report("failed to init vhost_net for queue");
199         goto err_init;
200     }
201     s->vhost_net = net;
202     ret = vhost_vdpa_net_check_device_id(net);
203     if (ret) {
204         goto err_check;
205     }
206     return 0;
207 err_check:
208     vhost_net_cleanup(net);
209     g_free(net);
210 err_init:
211     return -1;
212 }
213 
214 static void vhost_vdpa_cleanup(NetClientState *nc)
215 {
216     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
217 
218     /*
219      * If a peer NIC is attached, do not cleanup anything.
220      * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup()
221      * when the guest is shutting down.
222      */
223     if (nc->peer && nc->peer->info->type == NET_CLIENT_DRIVER_NIC) {
224         return;
225     }
226     munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len());
227     munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len());
228     if (s->vhost_net) {
229         vhost_net_cleanup(s->vhost_net);
230         g_free(s->vhost_net);
231         s->vhost_net = NULL;
232     }
233      if (s->vhost_vdpa.device_fd >= 0) {
234         qemu_close(s->vhost_vdpa.device_fd);
235         s->vhost_vdpa.device_fd = -1;
236     }
237 }
238 
239 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc)
240 {
241     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
242 
243     return true;
244 }
245 
246 static bool vhost_vdpa_has_ufo(NetClientState *nc)
247 {
248     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
249     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
250     uint64_t features = 0;
251     features |= (1ULL << VIRTIO_NET_F_HOST_UFO);
252     features = vhost_net_get_features(s->vhost_net, features);
253     return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO));
254 
255 }
256 
257 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc,
258                                        Error **errp)
259 {
260     const char *driver = object_class_get_name(oc);
261 
262     if (!g_str_has_prefix(driver, "virtio-net-")) {
263         error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*");
264         return false;
265     }
266 
267     return true;
268 }
269 
270 /** Dummy receive in case qemu falls back to userland tap networking */
271 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf,
272                                   size_t size)
273 {
274     return size;
275 }
276 
277 /** From any vdpa net client, get the netclient of the first queue pair */
278 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s)
279 {
280     NICState *nic = qemu_get_nic(s->nc.peer);
281     NetClientState *nc0 = qemu_get_peer(nic->ncs, 0);
282 
283     return DO_UPCAST(VhostVDPAState, nc, nc0);
284 }
285 
286 static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable)
287 {
288     struct vhost_vdpa *v = &s->vhost_vdpa;
289     VirtIONet *n;
290     VirtIODevice *vdev;
291     int data_queue_pairs, cvq, r;
292 
293     /* We are only called on the first data vqs and only if x-svq is not set */
294     if (s->vhost_vdpa.shadow_vqs_enabled == enable) {
295         return;
296     }
297 
298     vdev = v->dev->vdev;
299     n = VIRTIO_NET(vdev);
300     if (!n->vhost_started) {
301         return;
302     }
303 
304     data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1;
305     cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ?
306                                   n->max_ncs - n->max_queue_pairs : 0;
307     /*
308      * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter
309      * in the future and resume the device if read-only operations between
310      * suspend and reset goes wrong.
311      */
312     vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq);
313 
314     /* Start will check migration setup_or_active to configure or not SVQ */
315     r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq);
316     if (unlikely(r < 0)) {
317         error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r);
318     }
319 }
320 
321 static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data)
322 {
323     MigrationState *migration = data;
324     VhostVDPAState *s = container_of(notifier, VhostVDPAState,
325                                      migration_state);
326 
327     if (migration_in_setup(migration)) {
328         vhost_vdpa_net_log_global_enable(s, true);
329     } else if (migration_has_failed(migration)) {
330         vhost_vdpa_net_log_global_enable(s, false);
331     }
332 }
333 
334 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s)
335 {
336     struct vhost_vdpa *v = &s->vhost_vdpa;
337 
338     add_migration_state_change_notifier(&s->migration_state);
339     if (v->shadow_vqs_enabled) {
340         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
341                                            v->iova_range.last);
342     }
343 }
344 
345 static int vhost_vdpa_net_data_start(NetClientState *nc)
346 {
347     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
348     struct vhost_vdpa *v = &s->vhost_vdpa;
349 
350     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
351 
352     if (s->always_svq ||
353         migration_is_setup_or_active(migrate_get_current()->state)) {
354         v->shadow_vqs_enabled = true;
355         v->shadow_data = true;
356     } else {
357         v->shadow_vqs_enabled = false;
358         v->shadow_data = false;
359     }
360 
361     if (v->index == 0) {
362         vhost_vdpa_net_data_start_first(s);
363         return 0;
364     }
365 
366     if (v->shadow_vqs_enabled) {
367         VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s);
368         v->iova_tree = s0->vhost_vdpa.iova_tree;
369     }
370 
371     return 0;
372 }
373 
374 static void vhost_vdpa_net_client_stop(NetClientState *nc)
375 {
376     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
377     struct vhost_dev *dev;
378 
379     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
380 
381     if (s->vhost_vdpa.index == 0) {
382         remove_migration_state_change_notifier(&s->migration_state);
383     }
384 
385     dev = s->vhost_vdpa.dev;
386     if (dev->vq_index + dev->nvqs == dev->vq_index_end) {
387         g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete);
388     }
389 }
390 
391 static NetClientInfo net_vhost_vdpa_info = {
392         .type = NET_CLIENT_DRIVER_VHOST_VDPA,
393         .size = sizeof(VhostVDPAState),
394         .receive = vhost_vdpa_receive,
395         .start = vhost_vdpa_net_data_start,
396         .stop = vhost_vdpa_net_client_stop,
397         .cleanup = vhost_vdpa_cleanup,
398         .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
399         .has_ufo = vhost_vdpa_has_ufo,
400         .check_peer_type = vhost_vdpa_check_peer_type,
401 };
402 
403 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index,
404                                           Error **errp)
405 {
406     struct vhost_vring_state state = {
407         .index = vq_index,
408     };
409     int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state);
410 
411     if (unlikely(r < 0)) {
412         r = -errno;
413         error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index);
414         return r;
415     }
416 
417     return state.num;
418 }
419 
420 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v,
421                                            unsigned vq_group,
422                                            unsigned asid_num)
423 {
424     struct vhost_vring_state asid = {
425         .index = vq_group,
426         .num = asid_num,
427     };
428     int r;
429 
430     r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid);
431     if (unlikely(r < 0)) {
432         error_report("Can't set vq group %u asid %u, errno=%d (%s)",
433                      asid.index, asid.num, errno, g_strerror(errno));
434     }
435     return r;
436 }
437 
438 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr)
439 {
440     VhostIOVATree *tree = v->iova_tree;
441     DMAMap needle = {
442         /*
443          * No need to specify size or to look for more translations since
444          * this contiguous chunk was allocated by us.
445          */
446         .translated_addr = (hwaddr)(uintptr_t)addr,
447     };
448     const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle);
449     int r;
450 
451     if (unlikely(!map)) {
452         error_report("Cannot locate expected map");
453         return;
454     }
455 
456     r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1);
457     if (unlikely(r != 0)) {
458         error_report("Device cannot unmap: %s(%d)", g_strerror(r), r);
459     }
460 
461     vhost_iova_tree_remove(tree, *map);
462 }
463 
464 /** Map CVQ buffer. */
465 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size,
466                                   bool write)
467 {
468     DMAMap map = {};
469     int r;
470 
471     map.translated_addr = (hwaddr)(uintptr_t)buf;
472     map.size = size - 1;
473     map.perm = write ? IOMMU_RW : IOMMU_RO,
474     r = vhost_iova_tree_map_alloc(v->iova_tree, &map);
475     if (unlikely(r != IOVA_OK)) {
476         error_report("Cannot map injected element");
477         return r;
478     }
479 
480     r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova,
481                            vhost_vdpa_net_cvq_cmd_page_len(), buf, !write);
482     if (unlikely(r < 0)) {
483         goto dma_map_err;
484     }
485 
486     return 0;
487 
488 dma_map_err:
489     vhost_iova_tree_remove(v->iova_tree, map);
490     return r;
491 }
492 
493 static int vhost_vdpa_net_cvq_start(NetClientState *nc)
494 {
495     VhostVDPAState *s, *s0;
496     struct vhost_vdpa *v;
497     int64_t cvq_group;
498     int r;
499     Error *err = NULL;
500 
501     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
502 
503     s = DO_UPCAST(VhostVDPAState, nc, nc);
504     v = &s->vhost_vdpa;
505 
506     s0 = vhost_vdpa_net_first_nc_vdpa(s);
507     v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled;
508     v->shadow_vqs_enabled = s->always_svq;
509     s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID;
510 
511     if (s->vhost_vdpa.shadow_data) {
512         /* SVQ is already configured for all virtqueues */
513         goto out;
514     }
515 
516     /*
517      * If we early return in these cases SVQ will not be enabled. The migration
518      * will be blocked as long as vhost-vdpa backends will not offer _F_LOG.
519      */
520     if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) {
521         return 0;
522     }
523 
524     if (!s->cvq_isolated) {
525         return 0;
526     }
527 
528     cvq_group = vhost_vdpa_get_vring_group(v->device_fd,
529                                            v->dev->vq_index_end - 1,
530                                            &err);
531     if (unlikely(cvq_group < 0)) {
532         error_report_err(err);
533         return cvq_group;
534     }
535 
536     r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID);
537     if (unlikely(r < 0)) {
538         return r;
539     }
540 
541     v->shadow_vqs_enabled = true;
542     s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID;
543 
544 out:
545     if (!s->vhost_vdpa.shadow_vqs_enabled) {
546         return 0;
547     }
548 
549     if (s0->vhost_vdpa.iova_tree) {
550         /*
551          * SVQ is already configured for all virtqueues.  Reuse IOVA tree for
552          * simplicity, whether CVQ shares ASID with guest or not, because:
553          * - Memory listener need access to guest's memory addresses allocated
554          *   in the IOVA tree.
555          * - There should be plenty of IOVA address space for both ASID not to
556          *   worry about collisions between them.  Guest's translations are
557          *   still validated with virtio virtqueue_pop so there is no risk for
558          *   the guest to access memory that it shouldn't.
559          *
560          * To allocate a iova tree per ASID is doable but it complicates the
561          * code and it is not worth it for the moment.
562          */
563         v->iova_tree = s0->vhost_vdpa.iova_tree;
564     } else {
565         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
566                                            v->iova_range.last);
567     }
568 
569     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer,
570                                vhost_vdpa_net_cvq_cmd_page_len(), false);
571     if (unlikely(r < 0)) {
572         return r;
573     }
574 
575     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status,
576                                vhost_vdpa_net_cvq_cmd_page_len(), true);
577     if (unlikely(r < 0)) {
578         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
579     }
580 
581     return r;
582 }
583 
584 static void vhost_vdpa_net_cvq_stop(NetClientState *nc)
585 {
586     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
587 
588     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
589 
590     if (s->vhost_vdpa.shadow_vqs_enabled) {
591         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
592         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status);
593     }
594 
595     vhost_vdpa_net_client_stop(nc);
596 }
597 
598 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len,
599                                       size_t in_len)
600 {
601     /* Buffers for the device */
602     const struct iovec out = {
603         .iov_base = s->cvq_cmd_out_buffer,
604         .iov_len = out_len,
605     };
606     const struct iovec in = {
607         .iov_base = s->status,
608         .iov_len = sizeof(virtio_net_ctrl_ack),
609     };
610     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
611     int r;
612 
613     r = vhost_svq_add(svq, &out, 1, &in, 1, NULL);
614     if (unlikely(r != 0)) {
615         if (unlikely(r == -ENOSPC)) {
616             qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n",
617                           __func__);
618         }
619         return r;
620     }
621 
622     /*
623      * We can poll here since we've had BQL from the time we sent the
624      * descriptor. Also, we need to take the answer before SVQ pulls by itself,
625      * when BQL is released
626      */
627     return vhost_svq_poll(svq);
628 }
629 
630 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class,
631                                        uint8_t cmd, const struct iovec *data_sg,
632                                        size_t data_num)
633 {
634     const struct virtio_net_ctrl_hdr ctrl = {
635         .class = class,
636         .cmd = cmd,
637     };
638     size_t data_size = iov_size(data_sg, data_num);
639 
640     assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl));
641 
642     /* pack the CVQ command header */
643     memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl));
644 
645     /* pack the CVQ command command-specific-data */
646     iov_to_buf(data_sg, data_num, 0,
647                s->cvq_cmd_out_buffer + sizeof(ctrl), data_size);
648 
649     return vhost_vdpa_net_cvq_add(s, data_size + sizeof(ctrl),
650                                   sizeof(virtio_net_ctrl_ack));
651 }
652 
653 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n)
654 {
655     if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
656         const struct iovec data = {
657             .iov_base = (void *)n->mac,
658             .iov_len = sizeof(n->mac),
659         };
660         ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC,
661                                                   VIRTIO_NET_CTRL_MAC_ADDR_SET,
662                                                   &data, 1);
663         if (unlikely(dev_written < 0)) {
664             return dev_written;
665         }
666         if (*s->status != VIRTIO_NET_OK) {
667             return -EIO;
668         }
669     }
670 
671     /*
672      * According to VirtIO standard, "The device MUST have an
673      * empty MAC filtering table on reset.".
674      *
675      * Therefore, there is no need to send this CVQ command if the
676      * driver also sets an empty MAC filter table, which aligns with
677      * the device's defaults.
678      *
679      * Note that the device's defaults can mismatch the driver's
680      * configuration only at live migration.
681      */
682     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) ||
683         n->mac_table.in_use == 0) {
684         return 0;
685     }
686 
687     uint32_t uni_entries = n->mac_table.first_multi,
688              uni_macs_size = uni_entries * ETH_ALEN,
689              mul_entries = n->mac_table.in_use - uni_entries,
690              mul_macs_size = mul_entries * ETH_ALEN;
691     struct virtio_net_ctrl_mac uni = {
692         .entries = cpu_to_le32(uni_entries),
693     };
694     struct virtio_net_ctrl_mac mul = {
695         .entries = cpu_to_le32(mul_entries),
696     };
697     const struct iovec data[] = {
698         {
699             .iov_base = &uni,
700             .iov_len = sizeof(uni),
701         }, {
702             .iov_base = n->mac_table.macs,
703             .iov_len = uni_macs_size,
704         }, {
705             .iov_base = &mul,
706             .iov_len = sizeof(mul),
707         }, {
708             .iov_base = &n->mac_table.macs[uni_macs_size],
709             .iov_len = mul_macs_size,
710         },
711     };
712     ssize_t dev_written = vhost_vdpa_net_load_cmd(s,
713                                 VIRTIO_NET_CTRL_MAC,
714                                 VIRTIO_NET_CTRL_MAC_TABLE_SET,
715                                 data, ARRAY_SIZE(data));
716     if (unlikely(dev_written < 0)) {
717         return dev_written;
718     }
719     if (*s->status != VIRTIO_NET_OK) {
720         return -EIO;
721     }
722 
723     return 0;
724 }
725 
726 static int vhost_vdpa_net_load_mq(VhostVDPAState *s,
727                                   const VirtIONet *n)
728 {
729     struct virtio_net_ctrl_mq mq;
730     ssize_t dev_written;
731 
732     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) {
733         return 0;
734     }
735 
736     mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs);
737     const struct iovec data = {
738         .iov_base = &mq,
739         .iov_len = sizeof(mq),
740     };
741     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ,
742                                           VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET,
743                                           &data, 1);
744     if (unlikely(dev_written < 0)) {
745         return dev_written;
746     }
747     if (*s->status != VIRTIO_NET_OK) {
748         return -EIO;
749     }
750 
751     return 0;
752 }
753 
754 static int vhost_vdpa_net_load_offloads(VhostVDPAState *s,
755                                         const VirtIONet *n)
756 {
757     uint64_t offloads;
758     ssize_t dev_written;
759 
760     if (!virtio_vdev_has_feature(&n->parent_obj,
761                                  VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) {
762         return 0;
763     }
764 
765     if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) {
766         /*
767          * According to VirtIO standard, "Upon feature negotiation
768          * corresponding offload gets enabled to preserve
769          * backward compatibility.".
770          *
771          * Therefore, there is no need to send this CVQ command if the
772          * driver also enables all supported offloads, which aligns with
773          * the device's defaults.
774          *
775          * Note that the device's defaults can mismatch the driver's
776          * configuration only at live migration.
777          */
778         return 0;
779     }
780 
781     offloads = cpu_to_le64(n->curr_guest_offloads);
782     const struct iovec data = {
783         .iov_base = &offloads,
784         .iov_len = sizeof(offloads),
785     };
786     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
787                                           VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET,
788                                           &data, 1);
789     if (unlikely(dev_written < 0)) {
790         return dev_written;
791     }
792     if (*s->status != VIRTIO_NET_OK) {
793         return -EIO;
794     }
795 
796     return 0;
797 }
798 
799 static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s,
800                                        uint8_t cmd,
801                                        uint8_t on)
802 {
803     const struct iovec data = {
804         .iov_base = &on,
805         .iov_len = sizeof(on),
806     };
807     return vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_RX,
808                                    cmd, &data, 1);
809 }
810 
811 static int vhost_vdpa_net_load_rx(VhostVDPAState *s,
812                                   const VirtIONet *n)
813 {
814     ssize_t dev_written;
815 
816     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) {
817         return 0;
818     }
819 
820     /*
821      * According to virtio_net_reset(), device turns promiscuous mode
822      * on by default.
823      *
824      * Addtionally, according to VirtIO standard, "Since there are
825      * no guarantees, it can use a hash filter or silently switch to
826      * allmulti or promiscuous mode if it is given too many addresses.".
827      * QEMU marks `n->mac_table.uni_overflow` if guest sets too many
828      * non-multicast MAC addresses, indicating that promiscuous mode
829      * should be enabled.
830      *
831      * Therefore, QEMU should only send this CVQ command if the
832      * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off,
833      * which sets promiscuous mode on, different from the device's defaults.
834      *
835      * Note that the device's defaults can mismatch the driver's
836      * configuration only at live migration.
837      */
838     if (!n->mac_table.uni_overflow && !n->promisc) {
839         dev_written = vhost_vdpa_net_load_rx_mode(s,
840                                             VIRTIO_NET_CTRL_RX_PROMISC, 0);
841         if (unlikely(dev_written < 0)) {
842             return dev_written;
843         }
844         if (*s->status != VIRTIO_NET_OK) {
845             return -EIO;
846         }
847     }
848 
849     /*
850      * According to virtio_net_reset(), device turns all-multicast mode
851      * off by default.
852      *
853      * According to VirtIO standard, "Since there are no guarantees,
854      * it can use a hash filter or silently switch to allmulti or
855      * promiscuous mode if it is given too many addresses.". QEMU marks
856      * `n->mac_table.multi_overflow` if guest sets too many
857      * non-multicast MAC addresses.
858      *
859      * Therefore, QEMU should only send this CVQ command if the
860      * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on,
861      * which sets all-multicast mode on, different from the device's defaults.
862      *
863      * Note that the device's defaults can mismatch the driver's
864      * configuration only at live migration.
865      */
866     if (n->mac_table.multi_overflow || n->allmulti) {
867         dev_written = vhost_vdpa_net_load_rx_mode(s,
868                                             VIRTIO_NET_CTRL_RX_ALLMULTI, 1);
869         if (unlikely(dev_written < 0)) {
870             return dev_written;
871         }
872         if (*s->status != VIRTIO_NET_OK) {
873             return -EIO;
874         }
875     }
876 
877     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX_EXTRA)) {
878         return 0;
879     }
880 
881     /*
882      * According to virtio_net_reset(), device turns all-unicast mode
883      * off by default.
884      *
885      * Therefore, QEMU should only send this CVQ command if the driver
886      * sets all-unicast mode on, different from the device's defaults.
887      *
888      * Note that the device's defaults can mismatch the driver's
889      * configuration only at live migration.
890      */
891     if (n->alluni) {
892         dev_written = vhost_vdpa_net_load_rx_mode(s,
893                                             VIRTIO_NET_CTRL_RX_ALLUNI, 1);
894         if (dev_written < 0) {
895             return dev_written;
896         }
897         if (*s->status != VIRTIO_NET_OK) {
898             return -EIO;
899         }
900     }
901 
902     /*
903      * According to virtio_net_reset(), device turns non-multicast mode
904      * off by default.
905      *
906      * Therefore, QEMU should only send this CVQ command if the driver
907      * sets non-multicast mode on, different from the device's defaults.
908      *
909      * Note that the device's defaults can mismatch the driver's
910      * configuration only at live migration.
911      */
912     if (n->nomulti) {
913         dev_written = vhost_vdpa_net_load_rx_mode(s,
914                                             VIRTIO_NET_CTRL_RX_NOMULTI, 1);
915         if (dev_written < 0) {
916             return dev_written;
917         }
918         if (*s->status != VIRTIO_NET_OK) {
919             return -EIO;
920         }
921     }
922 
923     /*
924      * According to virtio_net_reset(), device turns non-unicast mode
925      * off by default.
926      *
927      * Therefore, QEMU should only send this CVQ command if the driver
928      * sets non-unicast mode on, different from the device's defaults.
929      *
930      * Note that the device's defaults can mismatch the driver's
931      * configuration only at live migration.
932      */
933     if (n->nouni) {
934         dev_written = vhost_vdpa_net_load_rx_mode(s,
935                                             VIRTIO_NET_CTRL_RX_NOUNI, 1);
936         if (dev_written < 0) {
937             return dev_written;
938         }
939         if (*s->status != VIRTIO_NET_OK) {
940             return -EIO;
941         }
942     }
943 
944     /*
945      * According to virtio_net_reset(), device turns non-broadcast mode
946      * off by default.
947      *
948      * Therefore, QEMU should only send this CVQ command if the driver
949      * sets non-broadcast mode on, different from the device's defaults.
950      *
951      * Note that the device's defaults can mismatch the driver's
952      * configuration only at live migration.
953      */
954     if (n->nobcast) {
955         dev_written = vhost_vdpa_net_load_rx_mode(s,
956                                             VIRTIO_NET_CTRL_RX_NOBCAST, 1);
957         if (dev_written < 0) {
958             return dev_written;
959         }
960         if (*s->status != VIRTIO_NET_OK) {
961             return -EIO;
962         }
963     }
964 
965     return 0;
966 }
967 
968 static int vhost_vdpa_net_load(NetClientState *nc)
969 {
970     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
971     struct vhost_vdpa *v = &s->vhost_vdpa;
972     const VirtIONet *n;
973     int r;
974 
975     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
976 
977     if (!v->shadow_vqs_enabled) {
978         return 0;
979     }
980 
981     n = VIRTIO_NET(v->dev->vdev);
982     r = vhost_vdpa_net_load_mac(s, n);
983     if (unlikely(r < 0)) {
984         return r;
985     }
986     r = vhost_vdpa_net_load_mq(s, n);
987     if (unlikely(r)) {
988         return r;
989     }
990     r = vhost_vdpa_net_load_offloads(s, n);
991     if (unlikely(r)) {
992         return r;
993     }
994     r = vhost_vdpa_net_load_rx(s, n);
995     if (unlikely(r)) {
996         return r;
997     }
998 
999     return 0;
1000 }
1001 
1002 static NetClientInfo net_vhost_vdpa_cvq_info = {
1003     .type = NET_CLIENT_DRIVER_VHOST_VDPA,
1004     .size = sizeof(VhostVDPAState),
1005     .receive = vhost_vdpa_receive,
1006     .start = vhost_vdpa_net_cvq_start,
1007     .load = vhost_vdpa_net_load,
1008     .stop = vhost_vdpa_net_cvq_stop,
1009     .cleanup = vhost_vdpa_cleanup,
1010     .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
1011     .has_ufo = vhost_vdpa_has_ufo,
1012     .check_peer_type = vhost_vdpa_check_peer_type,
1013 };
1014 
1015 /*
1016  * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to
1017  * vdpa device.
1018  *
1019  * Considering that QEMU cannot send the entire filter table to the
1020  * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ
1021  * command to enable promiscuous mode to receive all packets,
1022  * according to VirtIO standard, "Since there are no guarantees,
1023  * it can use a hash filter or silently switch to allmulti or
1024  * promiscuous mode if it is given too many addresses.".
1025  *
1026  * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and
1027  * marks `n->mac_table.x_overflow` accordingly, it should have
1028  * the same effect on the device model to receive
1029  * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses.
1030  * The same applies to multicast MAC addresses.
1031  *
1032  * Therefore, QEMU can provide the device model with a fake
1033  * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1)
1034  * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast
1035  * MAC addresses. This ensures that the device model marks
1036  * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`,
1037  * allowing all packets to be received, which aligns with the
1038  * state of the vdpa device.
1039  */
1040 static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState *s,
1041                                                        VirtQueueElement *elem,
1042                                                        struct iovec *out)
1043 {
1044     struct virtio_net_ctrl_mac mac_data, *mac_ptr;
1045     struct virtio_net_ctrl_hdr *hdr_ptr;
1046     uint32_t cursor;
1047     ssize_t r;
1048 
1049     /* parse the non-multicast MAC address entries from CVQ command */
1050     cursor = sizeof(*hdr_ptr);
1051     r = iov_to_buf(elem->out_sg, elem->out_num, cursor,
1052                    &mac_data, sizeof(mac_data));
1053     if (unlikely(r != sizeof(mac_data))) {
1054         /*
1055          * If the CVQ command is invalid, we should simulate the vdpa device
1056          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1057          */
1058         *s->status = VIRTIO_NET_ERR;
1059         return sizeof(*s->status);
1060     }
1061     cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN;
1062 
1063     /* parse the multicast MAC address entries from CVQ command */
1064     r = iov_to_buf(elem->out_sg, elem->out_num, cursor,
1065                    &mac_data, sizeof(mac_data));
1066     if (r != sizeof(mac_data)) {
1067         /*
1068          * If the CVQ command is invalid, we should simulate the vdpa device
1069          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1070          */
1071         *s->status = VIRTIO_NET_ERR;
1072         return sizeof(*s->status);
1073     }
1074     cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN;
1075 
1076     /* validate the CVQ command */
1077     if (iov_size(elem->out_sg, elem->out_num) != cursor) {
1078         /*
1079          * If the CVQ command is invalid, we should simulate the vdpa device
1080          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1081          */
1082         *s->status = VIRTIO_NET_ERR;
1083         return sizeof(*s->status);
1084     }
1085 
1086     /*
1087      * According to VirtIO standard, "Since there are no guarantees,
1088      * it can use a hash filter or silently switch to allmulti or
1089      * promiscuous mode if it is given too many addresses.".
1090      *
1091      * Therefore, considering that QEMU is unable to send the entire
1092      * filter table to the vdpa device, it should send the
1093      * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode
1094      */
1095     r = vhost_vdpa_net_load_rx_mode(s, VIRTIO_NET_CTRL_RX_PROMISC, 1);
1096     if (unlikely(r < 0)) {
1097         return r;
1098     }
1099     if (*s->status != VIRTIO_NET_OK) {
1100         return sizeof(*s->status);
1101     }
1102 
1103     /*
1104      * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ
1105      * command to the device model, including (`MAC_TABLE_ENTRIES` + 1)
1106      * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1)
1107      * multicast MAC addresses.
1108      *
1109      * By doing so, the device model can mark `n->mac_table.uni_overflow`
1110      * and `n->mac_table.multi_overflow`, enabling all packets to be
1111      * received, which aligns with the state of the vdpa device.
1112      */
1113     cursor = 0;
1114     uint32_t fake_uni_entries = MAC_TABLE_ENTRIES + 1,
1115              fake_mul_entries = MAC_TABLE_ENTRIES + 1,
1116              fake_cvq_size = sizeof(struct virtio_net_ctrl_hdr) +
1117                              sizeof(mac_data) + fake_uni_entries * ETH_ALEN +
1118                              sizeof(mac_data) + fake_mul_entries * ETH_ALEN;
1119 
1120     assert(fake_cvq_size < vhost_vdpa_net_cvq_cmd_page_len());
1121     out->iov_len = fake_cvq_size;
1122 
1123     /* pack the header for fake CVQ command */
1124     hdr_ptr = out->iov_base + cursor;
1125     hdr_ptr->class = VIRTIO_NET_CTRL_MAC;
1126     hdr_ptr->cmd = VIRTIO_NET_CTRL_MAC_TABLE_SET;
1127     cursor += sizeof(*hdr_ptr);
1128 
1129     /*
1130      * Pack the non-multicast MAC addresses part for fake CVQ command.
1131      *
1132      * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1133      * addresses provieded in CVQ command. Therefore, only the entries
1134      * field need to be prepared in the CVQ command.
1135      */
1136     mac_ptr = out->iov_base + cursor;
1137     mac_ptr->entries = cpu_to_le32(fake_uni_entries);
1138     cursor += sizeof(*mac_ptr) + fake_uni_entries * ETH_ALEN;
1139 
1140     /*
1141      * Pack the multicast MAC addresses part for fake CVQ command.
1142      *
1143      * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1144      * addresses provieded in CVQ command. Therefore, only the entries
1145      * field need to be prepared in the CVQ command.
1146      */
1147     mac_ptr = out->iov_base + cursor;
1148     mac_ptr->entries = cpu_to_le32(fake_mul_entries);
1149 
1150     /*
1151      * Simulating QEMU poll a vdpa device used buffer
1152      * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1153      */
1154     return sizeof(*s->status);
1155 }
1156 
1157 /**
1158  * Validate and copy control virtqueue commands.
1159  *
1160  * Following QEMU guidelines, we offer a copy of the buffers to the device to
1161  * prevent TOCTOU bugs.
1162  */
1163 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq,
1164                                             VirtQueueElement *elem,
1165                                             void *opaque)
1166 {
1167     VhostVDPAState *s = opaque;
1168     size_t in_len;
1169     const struct virtio_net_ctrl_hdr *ctrl;
1170     virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1171     /* Out buffer sent to both the vdpa device and the device model */
1172     struct iovec out = {
1173         .iov_base = s->cvq_cmd_out_buffer,
1174     };
1175     /* in buffer used for device model */
1176     const struct iovec in = {
1177         .iov_base = &status,
1178         .iov_len = sizeof(status),
1179     };
1180     ssize_t dev_written = -EINVAL;
1181 
1182     out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0,
1183                              s->cvq_cmd_out_buffer,
1184                              vhost_vdpa_net_cvq_cmd_page_len());
1185 
1186     ctrl = s->cvq_cmd_out_buffer;
1187     if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) {
1188         /*
1189          * Guest announce capability is emulated by qemu, so don't forward to
1190          * the device.
1191          */
1192         dev_written = sizeof(status);
1193         *s->status = VIRTIO_NET_OK;
1194     } else if (unlikely(ctrl->class == VIRTIO_NET_CTRL_MAC &&
1195                         ctrl->cmd == VIRTIO_NET_CTRL_MAC_TABLE_SET &&
1196                         iov_size(elem->out_sg, elem->out_num) > out.iov_len)) {
1197         /*
1198          * Due to the size limitation of the out buffer sent to the vdpa device,
1199          * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive
1200          * MAC addresses set by the driver for the filter table can cause
1201          * truncation of the CVQ command in QEMU. As a result, the vdpa device
1202          * rejects the flawed CVQ command.
1203          *
1204          * Therefore, QEMU must handle this situation instead of sending
1205          * the CVQ command direclty.
1206          */
1207         dev_written = vhost_vdpa_net_excessive_mac_filter_cvq_add(s, elem,
1208                                                                   &out);
1209         if (unlikely(dev_written < 0)) {
1210             goto out;
1211         }
1212     } else {
1213         dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status));
1214         if (unlikely(dev_written < 0)) {
1215             goto out;
1216         }
1217     }
1218 
1219     if (unlikely(dev_written < sizeof(status))) {
1220         error_report("Insufficient written data (%zu)", dev_written);
1221         goto out;
1222     }
1223 
1224     if (*s->status != VIRTIO_NET_OK) {
1225         goto out;
1226     }
1227 
1228     status = VIRTIO_NET_ERR;
1229     virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1);
1230     if (status != VIRTIO_NET_OK) {
1231         error_report("Bad CVQ processing in model");
1232     }
1233 
1234 out:
1235     in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status,
1236                           sizeof(status));
1237     if (unlikely(in_len < sizeof(status))) {
1238         error_report("Bad device CVQ written length");
1239     }
1240     vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status)));
1241     /*
1242      * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when
1243      * the function successfully forwards the CVQ command, indicated
1244      * by a non-negative value of `dev_written`. Otherwise, it still
1245      * belongs to SVQ.
1246      * This function should only free the `elem` when it owns.
1247      */
1248     if (dev_written >= 0) {
1249         g_free(elem);
1250     }
1251     return dev_written < 0 ? dev_written : 0;
1252 }
1253 
1254 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = {
1255     .avail_handler = vhost_vdpa_net_handle_ctrl_avail,
1256 };
1257 
1258 /**
1259  * Probe if CVQ is isolated
1260  *
1261  * @device_fd         The vdpa device fd
1262  * @features          Features offered by the device.
1263  * @cvq_index         The control vq pair index
1264  *
1265  * Returns <0 in case of failure, 0 if false and 1 if true.
1266  */
1267 static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features,
1268                                           int cvq_index, Error **errp)
1269 {
1270     uint64_t backend_features;
1271     int64_t cvq_group;
1272     uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE |
1273                      VIRTIO_CONFIG_S_DRIVER |
1274                      VIRTIO_CONFIG_S_FEATURES_OK;
1275     int r;
1276 
1277     ERRP_GUARD();
1278 
1279     r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features);
1280     if (unlikely(r < 0)) {
1281         error_setg_errno(errp, errno, "Cannot get vdpa backend_features");
1282         return r;
1283     }
1284 
1285     if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) {
1286         return 0;
1287     }
1288 
1289     r = ioctl(device_fd, VHOST_SET_FEATURES, &features);
1290     if (unlikely(r)) {
1291         error_setg_errno(errp, errno, "Cannot set features");
1292     }
1293 
1294     r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
1295     if (unlikely(r)) {
1296         error_setg_errno(errp, -r, "Cannot set device features");
1297         goto out;
1298     }
1299 
1300     cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp);
1301     if (unlikely(cvq_group < 0)) {
1302         if (cvq_group != -ENOTSUP) {
1303             r = cvq_group;
1304             goto out;
1305         }
1306 
1307         /*
1308          * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend
1309          * support ASID even if the parent driver does not.  The CVQ cannot be
1310          * isolated in this case.
1311          */
1312         error_free(*errp);
1313         *errp = NULL;
1314         r = 0;
1315         goto out;
1316     }
1317 
1318     for (int i = 0; i < cvq_index; ++i) {
1319         int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp);
1320         if (unlikely(group < 0)) {
1321             r = group;
1322             goto out;
1323         }
1324 
1325         if (group == (int64_t)cvq_group) {
1326             r = 0;
1327             goto out;
1328         }
1329     }
1330 
1331     r = 1;
1332 
1333 out:
1334     status = 0;
1335     ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
1336     return r;
1337 }
1338 
1339 static NetClientState *net_vhost_vdpa_init(NetClientState *peer,
1340                                        const char *device,
1341                                        const char *name,
1342                                        int vdpa_device_fd,
1343                                        int queue_pair_index,
1344                                        int nvqs,
1345                                        bool is_datapath,
1346                                        bool svq,
1347                                        struct vhost_vdpa_iova_range iova_range,
1348                                        uint64_t features,
1349                                        Error **errp)
1350 {
1351     NetClientState *nc = NULL;
1352     VhostVDPAState *s;
1353     int ret = 0;
1354     assert(name);
1355     int cvq_isolated;
1356 
1357     if (is_datapath) {
1358         nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device,
1359                                  name);
1360     } else {
1361         cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features,
1362                                                       queue_pair_index * 2,
1363                                                       errp);
1364         if (unlikely(cvq_isolated < 0)) {
1365             return NULL;
1366         }
1367 
1368         nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer,
1369                                          device, name);
1370     }
1371     qemu_set_info_str(nc, TYPE_VHOST_VDPA);
1372     s = DO_UPCAST(VhostVDPAState, nc, nc);
1373 
1374     s->vhost_vdpa.device_fd = vdpa_device_fd;
1375     s->vhost_vdpa.index = queue_pair_index;
1376     s->always_svq = svq;
1377     s->migration_state.notify = vdpa_net_migration_state_notifier;
1378     s->vhost_vdpa.shadow_vqs_enabled = svq;
1379     s->vhost_vdpa.iova_range = iova_range;
1380     s->vhost_vdpa.shadow_data = svq;
1381     if (queue_pair_index == 0) {
1382         vhost_vdpa_net_valid_svq_features(features,
1383                                           &s->vhost_vdpa.migration_blocker);
1384     } else if (!is_datapath) {
1385         s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
1386                                      PROT_READ | PROT_WRITE,
1387                                      MAP_SHARED | MAP_ANONYMOUS, -1, 0);
1388         s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
1389                          PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS,
1390                          -1, 0);
1391 
1392         s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops;
1393         s->vhost_vdpa.shadow_vq_ops_opaque = s;
1394         s->cvq_isolated = cvq_isolated;
1395 
1396         /*
1397          * TODO: We cannot migrate devices with CVQ and no x-svq enabled as
1398          * there is no way to set the device state (MAC, MQ, etc) before
1399          * starting the datapath.
1400          *
1401          * Migration blocker ownership now belongs to s->vhost_vdpa.
1402          */
1403         if (!svq) {
1404             error_setg(&s->vhost_vdpa.migration_blocker,
1405                        "net vdpa cannot migrate with CVQ feature");
1406         }
1407     }
1408     ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);
1409     if (ret) {
1410         qemu_del_net_client(nc);
1411         return NULL;
1412     }
1413     return nc;
1414 }
1415 
1416 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp)
1417 {
1418     int ret = ioctl(fd, VHOST_GET_FEATURES, features);
1419     if (unlikely(ret < 0)) {
1420         error_setg_errno(errp, errno,
1421                          "Fail to query features from vhost-vDPA device");
1422     }
1423     return ret;
1424 }
1425 
1426 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features,
1427                                           int *has_cvq, Error **errp)
1428 {
1429     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
1430     g_autofree struct vhost_vdpa_config *config = NULL;
1431     __virtio16 *max_queue_pairs;
1432     int ret;
1433 
1434     if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) {
1435         *has_cvq = 1;
1436     } else {
1437         *has_cvq = 0;
1438     }
1439 
1440     if (features & (1 << VIRTIO_NET_F_MQ)) {
1441         config = g_malloc0(config_size + sizeof(*max_queue_pairs));
1442         config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs);
1443         config->len = sizeof(*max_queue_pairs);
1444 
1445         ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config);
1446         if (ret) {
1447             error_setg(errp, "Fail to get config from vhost-vDPA device");
1448             return -ret;
1449         }
1450 
1451         max_queue_pairs = (__virtio16 *)&config->buf;
1452 
1453         return lduw_le_p(max_queue_pairs);
1454     }
1455 
1456     return 1;
1457 }
1458 
1459 int net_init_vhost_vdpa(const Netdev *netdev, const char *name,
1460                         NetClientState *peer, Error **errp)
1461 {
1462     const NetdevVhostVDPAOptions *opts;
1463     uint64_t features;
1464     int vdpa_device_fd;
1465     g_autofree NetClientState **ncs = NULL;
1466     struct vhost_vdpa_iova_range iova_range;
1467     NetClientState *nc;
1468     int queue_pairs, r, i = 0, has_cvq = 0;
1469 
1470     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA);
1471     opts = &netdev->u.vhost_vdpa;
1472     if (!opts->vhostdev && !opts->vhostfd) {
1473         error_setg(errp,
1474                    "vhost-vdpa: neither vhostdev= nor vhostfd= was specified");
1475         return -1;
1476     }
1477 
1478     if (opts->vhostdev && opts->vhostfd) {
1479         error_setg(errp,
1480                    "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive");
1481         return -1;
1482     }
1483 
1484     if (opts->vhostdev) {
1485         vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp);
1486         if (vdpa_device_fd == -1) {
1487             return -errno;
1488         }
1489     } else {
1490         /* has_vhostfd */
1491         vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp);
1492         if (vdpa_device_fd == -1) {
1493             error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: ");
1494             return -1;
1495         }
1496     }
1497 
1498     r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp);
1499     if (unlikely(r < 0)) {
1500         goto err;
1501     }
1502 
1503     queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features,
1504                                                  &has_cvq, errp);
1505     if (queue_pairs < 0) {
1506         qemu_close(vdpa_device_fd);
1507         return queue_pairs;
1508     }
1509 
1510     r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range);
1511     if (unlikely(r < 0)) {
1512         error_setg(errp, "vhost-vdpa: get iova range failed: %s",
1513                    strerror(-r));
1514         goto err;
1515     }
1516 
1517     if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) {
1518         goto err;
1519     }
1520 
1521     ncs = g_malloc0(sizeof(*ncs) * queue_pairs);
1522 
1523     for (i = 0; i < queue_pairs; i++) {
1524         ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1525                                      vdpa_device_fd, i, 2, true, opts->x_svq,
1526                                      iova_range, features, errp);
1527         if (!ncs[i])
1528             goto err;
1529     }
1530 
1531     if (has_cvq) {
1532         nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1533                                  vdpa_device_fd, i, 1, false,
1534                                  opts->x_svq, iova_range, features, errp);
1535         if (!nc)
1536             goto err;
1537     }
1538 
1539     return 0;
1540 
1541 err:
1542     if (i) {
1543         for (i--; i >= 0; i--) {
1544             qemu_del_net_client(ncs[i]);
1545         }
1546     }
1547 
1548     qemu_close(vdpa_device_fd);
1549 
1550     return -1;
1551 }
1552