xref: /openbmc/qemu/net/vhost-vdpa.c (revision bac4711b)
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_MQ) |
114     BIT_ULL(VIRTIO_F_ANY_LAYOUT) |
115     BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) |
116     /* VHOST_F_LOG_ALL is exposed by SVQ */
117     BIT_ULL(VHOST_F_LOG_ALL) |
118     BIT_ULL(VIRTIO_NET_F_RSC_EXT) |
119     BIT_ULL(VIRTIO_NET_F_STANDBY) |
120     BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX);
121 
122 #define VHOST_VDPA_NET_CVQ_ASID 1
123 
124 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc)
125 {
126     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
127     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
128     return s->vhost_net;
129 }
130 
131 static size_t vhost_vdpa_net_cvq_cmd_len(void)
132 {
133     /*
134      * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
135      * In buffer is always 1 byte, so it should fit here
136      */
137     return sizeof(struct virtio_net_ctrl_hdr) +
138            2 * sizeof(struct virtio_net_ctrl_mac) +
139            MAC_TABLE_ENTRIES * ETH_ALEN;
140 }
141 
142 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
143 {
144     return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
145 }
146 
147 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp)
148 {
149     uint64_t invalid_dev_features =
150         features & ~vdpa_svq_device_features &
151         /* Transport are all accepted at this point */
152         ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START,
153                          VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START);
154 
155     if (invalid_dev_features) {
156         error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64,
157                    invalid_dev_features);
158         return false;
159     }
160 
161     return vhost_svq_valid_features(features, errp);
162 }
163 
164 static int vhost_vdpa_net_check_device_id(struct vhost_net *net)
165 {
166     uint32_t device_id;
167     int ret;
168     struct vhost_dev *hdev;
169 
170     hdev = (struct vhost_dev *)&net->dev;
171     ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id);
172     if (device_id != VIRTIO_ID_NET) {
173         return -ENOTSUP;
174     }
175     return ret;
176 }
177 
178 static int vhost_vdpa_add(NetClientState *ncs, void *be,
179                           int queue_pair_index, int nvqs)
180 {
181     VhostNetOptions options;
182     struct vhost_net *net = NULL;
183     VhostVDPAState *s;
184     int ret;
185 
186     options.backend_type = VHOST_BACKEND_TYPE_VDPA;
187     assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
188     s = DO_UPCAST(VhostVDPAState, nc, ncs);
189     options.net_backend = ncs;
190     options.opaque      = be;
191     options.busyloop_timeout = 0;
192     options.nvqs = nvqs;
193 
194     net = vhost_net_init(&options);
195     if (!net) {
196         error_report("failed to init vhost_net for queue");
197         goto err_init;
198     }
199     s->vhost_net = net;
200     ret = vhost_vdpa_net_check_device_id(net);
201     if (ret) {
202         goto err_check;
203     }
204     return 0;
205 err_check:
206     vhost_net_cleanup(net);
207     g_free(net);
208 err_init:
209     return -1;
210 }
211 
212 static void vhost_vdpa_cleanup(NetClientState *nc)
213 {
214     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
215 
216     /*
217      * If a peer NIC is attached, do not cleanup anything.
218      * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup()
219      * when the guest is shutting down.
220      */
221     if (nc->peer && nc->peer->info->type == NET_CLIENT_DRIVER_NIC) {
222         return;
223     }
224     munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len());
225     munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len());
226     if (s->vhost_net) {
227         vhost_net_cleanup(s->vhost_net);
228         g_free(s->vhost_net);
229         s->vhost_net = NULL;
230     }
231      if (s->vhost_vdpa.device_fd >= 0) {
232         qemu_close(s->vhost_vdpa.device_fd);
233         s->vhost_vdpa.device_fd = -1;
234     }
235 }
236 
237 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc)
238 {
239     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
240 
241     return true;
242 }
243 
244 static bool vhost_vdpa_has_ufo(NetClientState *nc)
245 {
246     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
247     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
248     uint64_t features = 0;
249     features |= (1ULL << VIRTIO_NET_F_HOST_UFO);
250     features = vhost_net_get_features(s->vhost_net, features);
251     return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO));
252 
253 }
254 
255 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc,
256                                        Error **errp)
257 {
258     const char *driver = object_class_get_name(oc);
259 
260     if (!g_str_has_prefix(driver, "virtio-net-")) {
261         error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*");
262         return false;
263     }
264 
265     return true;
266 }
267 
268 /** Dummy receive in case qemu falls back to userland tap networking */
269 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf,
270                                   size_t size)
271 {
272     return size;
273 }
274 
275 /** From any vdpa net client, get the netclient of the first queue pair */
276 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s)
277 {
278     NICState *nic = qemu_get_nic(s->nc.peer);
279     NetClientState *nc0 = qemu_get_peer(nic->ncs, 0);
280 
281     return DO_UPCAST(VhostVDPAState, nc, nc0);
282 }
283 
284 static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable)
285 {
286     struct vhost_vdpa *v = &s->vhost_vdpa;
287     VirtIONet *n;
288     VirtIODevice *vdev;
289     int data_queue_pairs, cvq, r;
290 
291     /* We are only called on the first data vqs and only if x-svq is not set */
292     if (s->vhost_vdpa.shadow_vqs_enabled == enable) {
293         return;
294     }
295 
296     vdev = v->dev->vdev;
297     n = VIRTIO_NET(vdev);
298     if (!n->vhost_started) {
299         return;
300     }
301 
302     data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1;
303     cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ?
304                                   n->max_ncs - n->max_queue_pairs : 0;
305     /*
306      * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter
307      * in the future and resume the device if read-only operations between
308      * suspend and reset goes wrong.
309      */
310     vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq);
311 
312     /* Start will check migration setup_or_active to configure or not SVQ */
313     r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq);
314     if (unlikely(r < 0)) {
315         error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r);
316     }
317 }
318 
319 static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data)
320 {
321     MigrationState *migration = data;
322     VhostVDPAState *s = container_of(notifier, VhostVDPAState,
323                                      migration_state);
324 
325     if (migration_in_setup(migration)) {
326         vhost_vdpa_net_log_global_enable(s, true);
327     } else if (migration_has_failed(migration)) {
328         vhost_vdpa_net_log_global_enable(s, false);
329     }
330 }
331 
332 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s)
333 {
334     struct vhost_vdpa *v = &s->vhost_vdpa;
335 
336     add_migration_state_change_notifier(&s->migration_state);
337     if (v->shadow_vqs_enabled) {
338         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
339                                            v->iova_range.last);
340     }
341 }
342 
343 static int vhost_vdpa_net_data_start(NetClientState *nc)
344 {
345     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
346     struct vhost_vdpa *v = &s->vhost_vdpa;
347 
348     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
349 
350     if (s->always_svq ||
351         migration_is_setup_or_active(migrate_get_current()->state)) {
352         v->shadow_vqs_enabled = true;
353         v->shadow_data = true;
354     } else {
355         v->shadow_vqs_enabled = false;
356         v->shadow_data = false;
357     }
358 
359     if (v->index == 0) {
360         vhost_vdpa_net_data_start_first(s);
361         return 0;
362     }
363 
364     if (v->shadow_vqs_enabled) {
365         VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s);
366         v->iova_tree = s0->vhost_vdpa.iova_tree;
367     }
368 
369     return 0;
370 }
371 
372 static void vhost_vdpa_net_client_stop(NetClientState *nc)
373 {
374     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
375     struct vhost_dev *dev;
376 
377     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
378 
379     if (s->vhost_vdpa.index == 0) {
380         remove_migration_state_change_notifier(&s->migration_state);
381     }
382 
383     dev = s->vhost_vdpa.dev;
384     if (dev->vq_index + dev->nvqs == dev->vq_index_end) {
385         g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete);
386     }
387 }
388 
389 static NetClientInfo net_vhost_vdpa_info = {
390         .type = NET_CLIENT_DRIVER_VHOST_VDPA,
391         .size = sizeof(VhostVDPAState),
392         .receive = vhost_vdpa_receive,
393         .start = vhost_vdpa_net_data_start,
394         .stop = vhost_vdpa_net_client_stop,
395         .cleanup = vhost_vdpa_cleanup,
396         .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
397         .has_ufo = vhost_vdpa_has_ufo,
398         .check_peer_type = vhost_vdpa_check_peer_type,
399 };
400 
401 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index,
402                                           Error **errp)
403 {
404     struct vhost_vring_state state = {
405         .index = vq_index,
406     };
407     int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state);
408 
409     if (unlikely(r < 0)) {
410         r = -errno;
411         error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index);
412         return r;
413     }
414 
415     return state.num;
416 }
417 
418 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v,
419                                            unsigned vq_group,
420                                            unsigned asid_num)
421 {
422     struct vhost_vring_state asid = {
423         .index = vq_group,
424         .num = asid_num,
425     };
426     int r;
427 
428     r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid);
429     if (unlikely(r < 0)) {
430         error_report("Can't set vq group %u asid %u, errno=%d (%s)",
431                      asid.index, asid.num, errno, g_strerror(errno));
432     }
433     return r;
434 }
435 
436 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr)
437 {
438     VhostIOVATree *tree = v->iova_tree;
439     DMAMap needle = {
440         /*
441          * No need to specify size or to look for more translations since
442          * this contiguous chunk was allocated by us.
443          */
444         .translated_addr = (hwaddr)(uintptr_t)addr,
445     };
446     const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle);
447     int r;
448 
449     if (unlikely(!map)) {
450         error_report("Cannot locate expected map");
451         return;
452     }
453 
454     r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1);
455     if (unlikely(r != 0)) {
456         error_report("Device cannot unmap: %s(%d)", g_strerror(r), r);
457     }
458 
459     vhost_iova_tree_remove(tree, *map);
460 }
461 
462 /** Map CVQ buffer. */
463 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size,
464                                   bool write)
465 {
466     DMAMap map = {};
467     int r;
468 
469     map.translated_addr = (hwaddr)(uintptr_t)buf;
470     map.size = size - 1;
471     map.perm = write ? IOMMU_RW : IOMMU_RO,
472     r = vhost_iova_tree_map_alloc(v->iova_tree, &map);
473     if (unlikely(r != IOVA_OK)) {
474         error_report("Cannot map injected element");
475         return r;
476     }
477 
478     r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova,
479                            vhost_vdpa_net_cvq_cmd_page_len(), buf, !write);
480     if (unlikely(r < 0)) {
481         goto dma_map_err;
482     }
483 
484     return 0;
485 
486 dma_map_err:
487     vhost_iova_tree_remove(v->iova_tree, map);
488     return r;
489 }
490 
491 static int vhost_vdpa_net_cvq_start(NetClientState *nc)
492 {
493     VhostVDPAState *s, *s0;
494     struct vhost_vdpa *v;
495     int64_t cvq_group;
496     int r;
497     Error *err = NULL;
498 
499     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
500 
501     s = DO_UPCAST(VhostVDPAState, nc, nc);
502     v = &s->vhost_vdpa;
503 
504     s0 = vhost_vdpa_net_first_nc_vdpa(s);
505     v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled;
506     v->shadow_vqs_enabled = s->always_svq;
507     s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID;
508 
509     if (s->vhost_vdpa.shadow_data) {
510         /* SVQ is already configured for all virtqueues */
511         goto out;
512     }
513 
514     /*
515      * If we early return in these cases SVQ will not be enabled. The migration
516      * will be blocked as long as vhost-vdpa backends will not offer _F_LOG.
517      */
518     if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) {
519         return 0;
520     }
521 
522     if (!s->cvq_isolated) {
523         return 0;
524     }
525 
526     cvq_group = vhost_vdpa_get_vring_group(v->device_fd,
527                                            v->dev->vq_index_end - 1,
528                                            &err);
529     if (unlikely(cvq_group < 0)) {
530         error_report_err(err);
531         return cvq_group;
532     }
533 
534     r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID);
535     if (unlikely(r < 0)) {
536         return r;
537     }
538 
539     v->shadow_vqs_enabled = true;
540     s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID;
541 
542 out:
543     if (!s->vhost_vdpa.shadow_vqs_enabled) {
544         return 0;
545     }
546 
547     if (s0->vhost_vdpa.iova_tree) {
548         /*
549          * SVQ is already configured for all virtqueues.  Reuse IOVA tree for
550          * simplicity, whether CVQ shares ASID with guest or not, because:
551          * - Memory listener need access to guest's memory addresses allocated
552          *   in the IOVA tree.
553          * - There should be plenty of IOVA address space for both ASID not to
554          *   worry about collisions between them.  Guest's translations are
555          *   still validated with virtio virtqueue_pop so there is no risk for
556          *   the guest to access memory that it shouldn't.
557          *
558          * To allocate a iova tree per ASID is doable but it complicates the
559          * code and it is not worth it for the moment.
560          */
561         v->iova_tree = s0->vhost_vdpa.iova_tree;
562     } else {
563         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
564                                            v->iova_range.last);
565     }
566 
567     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer,
568                                vhost_vdpa_net_cvq_cmd_page_len(), false);
569     if (unlikely(r < 0)) {
570         return r;
571     }
572 
573     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status,
574                                vhost_vdpa_net_cvq_cmd_page_len(), true);
575     if (unlikely(r < 0)) {
576         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
577     }
578 
579     return r;
580 }
581 
582 static void vhost_vdpa_net_cvq_stop(NetClientState *nc)
583 {
584     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
585 
586     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
587 
588     if (s->vhost_vdpa.shadow_vqs_enabled) {
589         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
590         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status);
591     }
592 
593     vhost_vdpa_net_client_stop(nc);
594 }
595 
596 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len,
597                                       size_t in_len)
598 {
599     /* Buffers for the device */
600     const struct iovec out = {
601         .iov_base = s->cvq_cmd_out_buffer,
602         .iov_len = out_len,
603     };
604     const struct iovec in = {
605         .iov_base = s->status,
606         .iov_len = sizeof(virtio_net_ctrl_ack),
607     };
608     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
609     int r;
610 
611     r = vhost_svq_add(svq, &out, 1, &in, 1, NULL);
612     if (unlikely(r != 0)) {
613         if (unlikely(r == -ENOSPC)) {
614             qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n",
615                           __func__);
616         }
617         return r;
618     }
619 
620     /*
621      * We can poll here since we've had BQL from the time we sent the
622      * descriptor. Also, we need to take the answer before SVQ pulls by itself,
623      * when BQL is released
624      */
625     return vhost_svq_poll(svq);
626 }
627 
628 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class,
629                                        uint8_t cmd, const void *data,
630                                        size_t data_size)
631 {
632     const struct virtio_net_ctrl_hdr ctrl = {
633         .class = class,
634         .cmd = cmd,
635     };
636 
637     assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl));
638 
639     memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl));
640     memcpy(s->cvq_cmd_out_buffer + sizeof(ctrl), data, data_size);
641 
642     return vhost_vdpa_net_cvq_add(s, sizeof(ctrl) + data_size,
643                                   sizeof(virtio_net_ctrl_ack));
644 }
645 
646 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n)
647 {
648     if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
649         ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC,
650                                                   VIRTIO_NET_CTRL_MAC_ADDR_SET,
651                                                   n->mac, sizeof(n->mac));
652         if (unlikely(dev_written < 0)) {
653             return dev_written;
654         }
655         if (*s->status != VIRTIO_NET_OK) {
656             return -EIO;
657         }
658     }
659 
660     return 0;
661 }
662 
663 static int vhost_vdpa_net_load_mq(VhostVDPAState *s,
664                                   const VirtIONet *n)
665 {
666     struct virtio_net_ctrl_mq mq;
667     ssize_t dev_written;
668 
669     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) {
670         return 0;
671     }
672 
673     mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs);
674     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ,
675                                           VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &mq,
676                                           sizeof(mq));
677     if (unlikely(dev_written < 0)) {
678         return dev_written;
679     }
680     if (*s->status != VIRTIO_NET_OK) {
681         return -EIO;
682     }
683 
684     return 0;
685 }
686 
687 static int vhost_vdpa_net_load_offloads(VhostVDPAState *s,
688                                         const VirtIONet *n)
689 {
690     uint64_t offloads;
691     ssize_t dev_written;
692 
693     if (!virtio_vdev_has_feature(&n->parent_obj,
694                                  VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) {
695         return 0;
696     }
697 
698     if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) {
699         /*
700          * According to VirtIO standard, "Upon feature negotiation
701          * corresponding offload gets enabled to preserve
702          * backward compatibility.".
703          *
704          * Therefore, there is no need to send this CVQ command if the
705          * driver also enables all supported offloads, which aligns with
706          * the device's defaults.
707          *
708          * Note that the device's defaults can mismatch the driver's
709          * configuration only at live migration.
710          */
711         return 0;
712     }
713 
714     offloads = cpu_to_le64(n->curr_guest_offloads);
715     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
716                                           VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET,
717                                           &offloads, sizeof(offloads));
718     if (unlikely(dev_written < 0)) {
719         return dev_written;
720     }
721     if (*s->status != VIRTIO_NET_OK) {
722         return -EIO;
723     }
724 
725     return 0;
726 }
727 
728 static int vhost_vdpa_net_load(NetClientState *nc)
729 {
730     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
731     struct vhost_vdpa *v = &s->vhost_vdpa;
732     const VirtIONet *n;
733     int r;
734 
735     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
736 
737     if (!v->shadow_vqs_enabled) {
738         return 0;
739     }
740 
741     n = VIRTIO_NET(v->dev->vdev);
742     r = vhost_vdpa_net_load_mac(s, n);
743     if (unlikely(r < 0)) {
744         return r;
745     }
746     r = vhost_vdpa_net_load_mq(s, n);
747     if (unlikely(r)) {
748         return r;
749     }
750     r = vhost_vdpa_net_load_offloads(s, n);
751     if (unlikely(r)) {
752         return r;
753     }
754 
755     return 0;
756 }
757 
758 static NetClientInfo net_vhost_vdpa_cvq_info = {
759     .type = NET_CLIENT_DRIVER_VHOST_VDPA,
760     .size = sizeof(VhostVDPAState),
761     .receive = vhost_vdpa_receive,
762     .start = vhost_vdpa_net_cvq_start,
763     .load = vhost_vdpa_net_load,
764     .stop = vhost_vdpa_net_cvq_stop,
765     .cleanup = vhost_vdpa_cleanup,
766     .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
767     .has_ufo = vhost_vdpa_has_ufo,
768     .check_peer_type = vhost_vdpa_check_peer_type,
769 };
770 
771 /**
772  * Validate and copy control virtqueue commands.
773  *
774  * Following QEMU guidelines, we offer a copy of the buffers to the device to
775  * prevent TOCTOU bugs.
776  */
777 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq,
778                                             VirtQueueElement *elem,
779                                             void *opaque)
780 {
781     VhostVDPAState *s = opaque;
782     size_t in_len;
783     virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
784     /* Out buffer sent to both the vdpa device and the device model */
785     struct iovec out = {
786         .iov_base = s->cvq_cmd_out_buffer,
787     };
788     /* in buffer used for device model */
789     const struct iovec in = {
790         .iov_base = &status,
791         .iov_len = sizeof(status),
792     };
793     ssize_t dev_written = -EINVAL;
794 
795     out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0,
796                              s->cvq_cmd_out_buffer,
797                              vhost_vdpa_net_cvq_cmd_len());
798     if (*(uint8_t *)s->cvq_cmd_out_buffer == VIRTIO_NET_CTRL_ANNOUNCE) {
799         /*
800          * Guest announce capability is emulated by qemu, so don't forward to
801          * the device.
802          */
803         dev_written = sizeof(status);
804         *s->status = VIRTIO_NET_OK;
805     } else {
806         dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status));
807         if (unlikely(dev_written < 0)) {
808             goto out;
809         }
810     }
811 
812     if (unlikely(dev_written < sizeof(status))) {
813         error_report("Insufficient written data (%zu)", dev_written);
814         goto out;
815     }
816 
817     if (*s->status != VIRTIO_NET_OK) {
818         goto out;
819     }
820 
821     status = VIRTIO_NET_ERR;
822     virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1);
823     if (status != VIRTIO_NET_OK) {
824         error_report("Bad CVQ processing in model");
825     }
826 
827 out:
828     in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status,
829                           sizeof(status));
830     if (unlikely(in_len < sizeof(status))) {
831         error_report("Bad device CVQ written length");
832     }
833     vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status)));
834     g_free(elem);
835     return dev_written < 0 ? dev_written : 0;
836 }
837 
838 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = {
839     .avail_handler = vhost_vdpa_net_handle_ctrl_avail,
840 };
841 
842 /**
843  * Probe if CVQ is isolated
844  *
845  * @device_fd         The vdpa device fd
846  * @features          Features offered by the device.
847  * @cvq_index         The control vq pair index
848  *
849  * Returns <0 in case of failure, 0 if false and 1 if true.
850  */
851 static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features,
852                                           int cvq_index, Error **errp)
853 {
854     uint64_t backend_features;
855     int64_t cvq_group;
856     uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE |
857                      VIRTIO_CONFIG_S_DRIVER |
858                      VIRTIO_CONFIG_S_FEATURES_OK;
859     int r;
860 
861     ERRP_GUARD();
862 
863     r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features);
864     if (unlikely(r < 0)) {
865         error_setg_errno(errp, errno, "Cannot get vdpa backend_features");
866         return r;
867     }
868 
869     if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) {
870         return 0;
871     }
872 
873     r = ioctl(device_fd, VHOST_SET_FEATURES, &features);
874     if (unlikely(r)) {
875         error_setg_errno(errp, errno, "Cannot set features");
876     }
877 
878     r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
879     if (unlikely(r)) {
880         error_setg_errno(errp, -r, "Cannot set device features");
881         goto out;
882     }
883 
884     cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp);
885     if (unlikely(cvq_group < 0)) {
886         if (cvq_group != -ENOTSUP) {
887             r = cvq_group;
888             goto out;
889         }
890 
891         /*
892          * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend
893          * support ASID even if the parent driver does not.  The CVQ cannot be
894          * isolated in this case.
895          */
896         error_free(*errp);
897         *errp = NULL;
898         r = 0;
899         goto out;
900     }
901 
902     for (int i = 0; i < cvq_index; ++i) {
903         int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp);
904         if (unlikely(group < 0)) {
905             r = group;
906             goto out;
907         }
908 
909         if (group == (int64_t)cvq_group) {
910             r = 0;
911             goto out;
912         }
913     }
914 
915     r = 1;
916 
917 out:
918     status = 0;
919     ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
920     return r;
921 }
922 
923 static NetClientState *net_vhost_vdpa_init(NetClientState *peer,
924                                        const char *device,
925                                        const char *name,
926                                        int vdpa_device_fd,
927                                        int queue_pair_index,
928                                        int nvqs,
929                                        bool is_datapath,
930                                        bool svq,
931                                        struct vhost_vdpa_iova_range iova_range,
932                                        uint64_t features,
933                                        Error **errp)
934 {
935     NetClientState *nc = NULL;
936     VhostVDPAState *s;
937     int ret = 0;
938     assert(name);
939     int cvq_isolated;
940 
941     if (is_datapath) {
942         nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device,
943                                  name);
944     } else {
945         cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features,
946                                                       queue_pair_index * 2,
947                                                       errp);
948         if (unlikely(cvq_isolated < 0)) {
949             return NULL;
950         }
951 
952         nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer,
953                                          device, name);
954     }
955     qemu_set_info_str(nc, TYPE_VHOST_VDPA);
956     s = DO_UPCAST(VhostVDPAState, nc, nc);
957 
958     s->vhost_vdpa.device_fd = vdpa_device_fd;
959     s->vhost_vdpa.index = queue_pair_index;
960     s->always_svq = svq;
961     s->migration_state.notify = vdpa_net_migration_state_notifier;
962     s->vhost_vdpa.shadow_vqs_enabled = svq;
963     s->vhost_vdpa.iova_range = iova_range;
964     s->vhost_vdpa.shadow_data = svq;
965     if (queue_pair_index == 0) {
966         vhost_vdpa_net_valid_svq_features(features,
967                                           &s->vhost_vdpa.migration_blocker);
968     } else if (!is_datapath) {
969         s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
970                                      PROT_READ | PROT_WRITE,
971                                      MAP_SHARED | MAP_ANONYMOUS, -1, 0);
972         s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
973                          PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS,
974                          -1, 0);
975 
976         s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops;
977         s->vhost_vdpa.shadow_vq_ops_opaque = s;
978         s->cvq_isolated = cvq_isolated;
979 
980         /*
981          * TODO: We cannot migrate devices with CVQ and no x-svq enabled as
982          * there is no way to set the device state (MAC, MQ, etc) before
983          * starting the datapath.
984          *
985          * Migration blocker ownership now belongs to s->vhost_vdpa.
986          */
987         if (!svq) {
988             error_setg(&s->vhost_vdpa.migration_blocker,
989                        "net vdpa cannot migrate with CVQ feature");
990         }
991     }
992     ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);
993     if (ret) {
994         qemu_del_net_client(nc);
995         return NULL;
996     }
997     return nc;
998 }
999 
1000 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp)
1001 {
1002     int ret = ioctl(fd, VHOST_GET_FEATURES, features);
1003     if (unlikely(ret < 0)) {
1004         error_setg_errno(errp, errno,
1005                          "Fail to query features from vhost-vDPA device");
1006     }
1007     return ret;
1008 }
1009 
1010 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features,
1011                                           int *has_cvq, Error **errp)
1012 {
1013     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
1014     g_autofree struct vhost_vdpa_config *config = NULL;
1015     __virtio16 *max_queue_pairs;
1016     int ret;
1017 
1018     if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) {
1019         *has_cvq = 1;
1020     } else {
1021         *has_cvq = 0;
1022     }
1023 
1024     if (features & (1 << VIRTIO_NET_F_MQ)) {
1025         config = g_malloc0(config_size + sizeof(*max_queue_pairs));
1026         config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs);
1027         config->len = sizeof(*max_queue_pairs);
1028 
1029         ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config);
1030         if (ret) {
1031             error_setg(errp, "Fail to get config from vhost-vDPA device");
1032             return -ret;
1033         }
1034 
1035         max_queue_pairs = (__virtio16 *)&config->buf;
1036 
1037         return lduw_le_p(max_queue_pairs);
1038     }
1039 
1040     return 1;
1041 }
1042 
1043 int net_init_vhost_vdpa(const Netdev *netdev, const char *name,
1044                         NetClientState *peer, Error **errp)
1045 {
1046     const NetdevVhostVDPAOptions *opts;
1047     uint64_t features;
1048     int vdpa_device_fd;
1049     g_autofree NetClientState **ncs = NULL;
1050     struct vhost_vdpa_iova_range iova_range;
1051     NetClientState *nc;
1052     int queue_pairs, r, i = 0, has_cvq = 0;
1053 
1054     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA);
1055     opts = &netdev->u.vhost_vdpa;
1056     if (!opts->vhostdev && !opts->vhostfd) {
1057         error_setg(errp,
1058                    "vhost-vdpa: neither vhostdev= nor vhostfd= was specified");
1059         return -1;
1060     }
1061 
1062     if (opts->vhostdev && opts->vhostfd) {
1063         error_setg(errp,
1064                    "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive");
1065         return -1;
1066     }
1067 
1068     if (opts->vhostdev) {
1069         vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp);
1070         if (vdpa_device_fd == -1) {
1071             return -errno;
1072         }
1073     } else {
1074         /* has_vhostfd */
1075         vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp);
1076         if (vdpa_device_fd == -1) {
1077             error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: ");
1078             return -1;
1079         }
1080     }
1081 
1082     r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp);
1083     if (unlikely(r < 0)) {
1084         goto err;
1085     }
1086 
1087     queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features,
1088                                                  &has_cvq, errp);
1089     if (queue_pairs < 0) {
1090         qemu_close(vdpa_device_fd);
1091         return queue_pairs;
1092     }
1093 
1094     r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range);
1095     if (unlikely(r < 0)) {
1096         error_setg(errp, "vhost-vdpa: get iova range failed: %s",
1097                    strerror(-r));
1098         goto err;
1099     }
1100 
1101     if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) {
1102         goto err;
1103     }
1104 
1105     ncs = g_malloc0(sizeof(*ncs) * queue_pairs);
1106 
1107     for (i = 0; i < queue_pairs; i++) {
1108         ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1109                                      vdpa_device_fd, i, 2, true, opts->x_svq,
1110                                      iova_range, features, errp);
1111         if (!ncs[i])
1112             goto err;
1113     }
1114 
1115     if (has_cvq) {
1116         nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1117                                  vdpa_device_fd, i, 1, false,
1118                                  opts->x_svq, iova_range, features, errp);
1119         if (!nc)
1120             goto err;
1121     }
1122 
1123     return 0;
1124 
1125 err:
1126     if (i) {
1127         for (i--; i >= 0; i--) {
1128             qemu_del_net_client(ncs[i]);
1129         }
1130     }
1131 
1132     qemu_close(vdpa_device_fd);
1133 
1134     return -1;
1135 }
1136