xref: /openbmc/qemu/net/vhost-vdpa.c (revision 76eb88b1)
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 "hw/virtio/vhost.h"
30 
31 /* Todo:need to add the multiqueue support here */
32 typedef struct VhostVDPAState {
33     NetClientState nc;
34     struct vhost_vdpa vhost_vdpa;
35     VHostNetState *vhost_net;
36 
37     /* Control commands shadow buffers */
38     void *cvq_cmd_out_buffer;
39     virtio_net_ctrl_ack *status;
40 
41     bool started;
42 } VhostVDPAState;
43 
44 const int vdpa_feature_bits[] = {
45     VIRTIO_F_NOTIFY_ON_EMPTY,
46     VIRTIO_RING_F_INDIRECT_DESC,
47     VIRTIO_RING_F_EVENT_IDX,
48     VIRTIO_F_ANY_LAYOUT,
49     VIRTIO_F_VERSION_1,
50     VIRTIO_NET_F_CSUM,
51     VIRTIO_NET_F_GUEST_CSUM,
52     VIRTIO_NET_F_GSO,
53     VIRTIO_NET_F_GUEST_TSO4,
54     VIRTIO_NET_F_GUEST_TSO6,
55     VIRTIO_NET_F_GUEST_ECN,
56     VIRTIO_NET_F_GUEST_UFO,
57     VIRTIO_NET_F_HOST_TSO4,
58     VIRTIO_NET_F_HOST_TSO6,
59     VIRTIO_NET_F_HOST_ECN,
60     VIRTIO_NET_F_HOST_UFO,
61     VIRTIO_NET_F_MRG_RXBUF,
62     VIRTIO_NET_F_MTU,
63     VIRTIO_NET_F_CTRL_RX,
64     VIRTIO_NET_F_CTRL_RX_EXTRA,
65     VIRTIO_NET_F_CTRL_VLAN,
66     VIRTIO_NET_F_GUEST_ANNOUNCE,
67     VIRTIO_NET_F_CTRL_MAC_ADDR,
68     VIRTIO_NET_F_RSS,
69     VIRTIO_NET_F_MQ,
70     VIRTIO_NET_F_CTRL_VQ,
71     VIRTIO_F_IOMMU_PLATFORM,
72     VIRTIO_F_RING_PACKED,
73     VIRTIO_NET_F_RSS,
74     VIRTIO_NET_F_HASH_REPORT,
75     VIRTIO_NET_F_GUEST_ANNOUNCE,
76     VIRTIO_NET_F_STATUS,
77     VHOST_INVALID_FEATURE_BIT
78 };
79 
80 /** Supported device specific feature bits with SVQ */
81 static const uint64_t vdpa_svq_device_features =
82     BIT_ULL(VIRTIO_NET_F_CSUM) |
83     BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) |
84     BIT_ULL(VIRTIO_NET_F_MTU) |
85     BIT_ULL(VIRTIO_NET_F_MAC) |
86     BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) |
87     BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) |
88     BIT_ULL(VIRTIO_NET_F_GUEST_ECN) |
89     BIT_ULL(VIRTIO_NET_F_GUEST_UFO) |
90     BIT_ULL(VIRTIO_NET_F_HOST_TSO4) |
91     BIT_ULL(VIRTIO_NET_F_HOST_TSO6) |
92     BIT_ULL(VIRTIO_NET_F_HOST_ECN) |
93     BIT_ULL(VIRTIO_NET_F_HOST_UFO) |
94     BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) |
95     BIT_ULL(VIRTIO_NET_F_STATUS) |
96     BIT_ULL(VIRTIO_NET_F_CTRL_VQ) |
97     BIT_ULL(VIRTIO_NET_F_MQ) |
98     BIT_ULL(VIRTIO_F_ANY_LAYOUT) |
99     BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) |
100     BIT_ULL(VIRTIO_NET_F_RSC_EXT) |
101     BIT_ULL(VIRTIO_NET_F_STANDBY);
102 
103 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc)
104 {
105     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
106     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
107     return s->vhost_net;
108 }
109 
110 static int vhost_vdpa_net_check_device_id(struct vhost_net *net)
111 {
112     uint32_t device_id;
113     int ret;
114     struct vhost_dev *hdev;
115 
116     hdev = (struct vhost_dev *)&net->dev;
117     ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id);
118     if (device_id != VIRTIO_ID_NET) {
119         return -ENOTSUP;
120     }
121     return ret;
122 }
123 
124 static int vhost_vdpa_add(NetClientState *ncs, void *be,
125                           int queue_pair_index, int nvqs)
126 {
127     VhostNetOptions options;
128     struct vhost_net *net = NULL;
129     VhostVDPAState *s;
130     int ret;
131 
132     options.backend_type = VHOST_BACKEND_TYPE_VDPA;
133     assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
134     s = DO_UPCAST(VhostVDPAState, nc, ncs);
135     options.net_backend = ncs;
136     options.opaque      = be;
137     options.busyloop_timeout = 0;
138     options.nvqs = nvqs;
139 
140     net = vhost_net_init(&options);
141     if (!net) {
142         error_report("failed to init vhost_net for queue");
143         goto err_init;
144     }
145     s->vhost_net = net;
146     ret = vhost_vdpa_net_check_device_id(net);
147     if (ret) {
148         goto err_check;
149     }
150     return 0;
151 err_check:
152     vhost_net_cleanup(net);
153     g_free(net);
154 err_init:
155     return -1;
156 }
157 
158 static void vhost_vdpa_cleanup(NetClientState *nc)
159 {
160     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
161     struct vhost_dev *dev = &s->vhost_net->dev;
162 
163     qemu_vfree(s->cvq_cmd_out_buffer);
164     qemu_vfree(s->status);
165     if (dev->vq_index + dev->nvqs == dev->vq_index_end) {
166         g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete);
167     }
168     if (s->vhost_net) {
169         vhost_net_cleanup(s->vhost_net);
170         g_free(s->vhost_net);
171         s->vhost_net = NULL;
172     }
173      if (s->vhost_vdpa.device_fd >= 0) {
174         qemu_close(s->vhost_vdpa.device_fd);
175         s->vhost_vdpa.device_fd = -1;
176     }
177 }
178 
179 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc)
180 {
181     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
182 
183     return true;
184 }
185 
186 static bool vhost_vdpa_has_ufo(NetClientState *nc)
187 {
188     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
189     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
190     uint64_t features = 0;
191     features |= (1ULL << VIRTIO_NET_F_HOST_UFO);
192     features = vhost_net_get_features(s->vhost_net, features);
193     return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO));
194 
195 }
196 
197 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc,
198                                        Error **errp)
199 {
200     const char *driver = object_class_get_name(oc);
201 
202     if (!g_str_has_prefix(driver, "virtio-net-")) {
203         error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*");
204         return false;
205     }
206 
207     return true;
208 }
209 
210 /** Dummy receive in case qemu falls back to userland tap networking */
211 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf,
212                                   size_t size)
213 {
214     return 0;
215 }
216 
217 static NetClientInfo net_vhost_vdpa_info = {
218         .type = NET_CLIENT_DRIVER_VHOST_VDPA,
219         .size = sizeof(VhostVDPAState),
220         .receive = vhost_vdpa_receive,
221         .cleanup = vhost_vdpa_cleanup,
222         .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
223         .has_ufo = vhost_vdpa_has_ufo,
224         .check_peer_type = vhost_vdpa_check_peer_type,
225 };
226 
227 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr)
228 {
229     VhostIOVATree *tree = v->iova_tree;
230     DMAMap needle = {
231         /*
232          * No need to specify size or to look for more translations since
233          * this contiguous chunk was allocated by us.
234          */
235         .translated_addr = (hwaddr)(uintptr_t)addr,
236     };
237     const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle);
238     int r;
239 
240     if (unlikely(!map)) {
241         error_report("Cannot locate expected map");
242         return;
243     }
244 
245     r = vhost_vdpa_dma_unmap(v, map->iova, map->size + 1);
246     if (unlikely(r != 0)) {
247         error_report("Device cannot unmap: %s(%d)", g_strerror(r), r);
248     }
249 
250     vhost_iova_tree_remove(tree, *map);
251 }
252 
253 static size_t vhost_vdpa_net_cvq_cmd_len(void)
254 {
255     /*
256      * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
257      * In buffer is always 1 byte, so it should fit here
258      */
259     return sizeof(struct virtio_net_ctrl_hdr) +
260            2 * sizeof(struct virtio_net_ctrl_mac) +
261            MAC_TABLE_ENTRIES * ETH_ALEN;
262 }
263 
264 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
265 {
266     return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
267 }
268 
269 /** Map CVQ buffer. */
270 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size,
271                                   bool write)
272 {
273     DMAMap map = {};
274     int r;
275 
276     map.translated_addr = (hwaddr)(uintptr_t)buf;
277     map.size = size - 1;
278     map.perm = write ? IOMMU_RW : IOMMU_RO,
279     r = vhost_iova_tree_map_alloc(v->iova_tree, &map);
280     if (unlikely(r != IOVA_OK)) {
281         error_report("Cannot map injected element");
282         return r;
283     }
284 
285     r = vhost_vdpa_dma_map(v, map.iova, vhost_vdpa_net_cvq_cmd_page_len(), buf,
286                            !write);
287     if (unlikely(r < 0)) {
288         goto dma_map_err;
289     }
290 
291     return 0;
292 
293 dma_map_err:
294     vhost_iova_tree_remove(v->iova_tree, map);
295     return r;
296 }
297 
298 static int vhost_vdpa_net_cvq_start(NetClientState *nc)
299 {
300     VhostVDPAState *s;
301     int r;
302 
303     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
304 
305     s = DO_UPCAST(VhostVDPAState, nc, nc);
306     if (!s->vhost_vdpa.shadow_vqs_enabled) {
307         return 0;
308     }
309 
310     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer,
311                                vhost_vdpa_net_cvq_cmd_page_len(), false);
312     if (unlikely(r < 0)) {
313         return r;
314     }
315 
316     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status,
317                                vhost_vdpa_net_cvq_cmd_page_len(), true);
318     if (unlikely(r < 0)) {
319         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
320     }
321 
322     return r;
323 }
324 
325 static void vhost_vdpa_net_cvq_stop(NetClientState *nc)
326 {
327     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
328 
329     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
330 
331     if (s->vhost_vdpa.shadow_vqs_enabled) {
332         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
333         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status);
334     }
335 }
336 
337 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len,
338                                       size_t in_len)
339 {
340     /* Buffers for the device */
341     const struct iovec out = {
342         .iov_base = s->cvq_cmd_out_buffer,
343         .iov_len = out_len,
344     };
345     const struct iovec in = {
346         .iov_base = s->status,
347         .iov_len = sizeof(virtio_net_ctrl_ack),
348     };
349     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
350     int r;
351 
352     r = vhost_svq_add(svq, &out, 1, &in, 1, NULL);
353     if (unlikely(r != 0)) {
354         if (unlikely(r == -ENOSPC)) {
355             qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n",
356                           __func__);
357         }
358         return r;
359     }
360 
361     /*
362      * We can poll here since we've had BQL from the time we sent the
363      * descriptor. Also, we need to take the answer before SVQ pulls by itself,
364      * when BQL is released
365      */
366     return vhost_svq_poll(svq);
367 }
368 
369 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class,
370                                        uint8_t cmd, const void *data,
371                                        size_t data_size)
372 {
373     const struct virtio_net_ctrl_hdr ctrl = {
374         .class = class,
375         .cmd = cmd,
376     };
377 
378     assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl));
379 
380     memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl));
381     memcpy(s->cvq_cmd_out_buffer + sizeof(ctrl), data, data_size);
382 
383     return vhost_vdpa_net_cvq_add(s, sizeof(ctrl) + data_size,
384                                   sizeof(virtio_net_ctrl_ack));
385 }
386 
387 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n)
388 {
389     uint64_t features = n->parent_obj.guest_features;
390     if (features & BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR)) {
391         ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC,
392                                                   VIRTIO_NET_CTRL_MAC_ADDR_SET,
393                                                   n->mac, sizeof(n->mac));
394         if (unlikely(dev_written < 0)) {
395             return dev_written;
396         }
397 
398         return *s->status != VIRTIO_NET_OK;
399     }
400 
401     return 0;
402 }
403 
404 static int vhost_vdpa_net_load_mq(VhostVDPAState *s,
405                                   const VirtIONet *n)
406 {
407     struct virtio_net_ctrl_mq mq;
408     uint64_t features = n->parent_obj.guest_features;
409     ssize_t dev_written;
410 
411     if (!(features & BIT_ULL(VIRTIO_NET_F_MQ))) {
412         return 0;
413     }
414 
415     mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs);
416     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ,
417                                           VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &mq,
418                                           sizeof(mq));
419     if (unlikely(dev_written < 0)) {
420         return dev_written;
421     }
422 
423     return *s->status != VIRTIO_NET_OK;
424 }
425 
426 static int vhost_vdpa_net_load(NetClientState *nc)
427 {
428     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
429     struct vhost_vdpa *v = &s->vhost_vdpa;
430     const VirtIONet *n;
431     int r;
432 
433     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
434 
435     if (!v->shadow_vqs_enabled) {
436         return 0;
437     }
438 
439     n = VIRTIO_NET(v->dev->vdev);
440     r = vhost_vdpa_net_load_mac(s, n);
441     if (unlikely(r < 0)) {
442         return r;
443     }
444     r = vhost_vdpa_net_load_mq(s, n);
445     if (unlikely(r)) {
446         return r;
447     }
448 
449     return 0;
450 }
451 
452 static NetClientInfo net_vhost_vdpa_cvq_info = {
453     .type = NET_CLIENT_DRIVER_VHOST_VDPA,
454     .size = sizeof(VhostVDPAState),
455     .receive = vhost_vdpa_receive,
456     .start = vhost_vdpa_net_cvq_start,
457     .load = vhost_vdpa_net_load,
458     .stop = vhost_vdpa_net_cvq_stop,
459     .cleanup = vhost_vdpa_cleanup,
460     .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
461     .has_ufo = vhost_vdpa_has_ufo,
462     .check_peer_type = vhost_vdpa_check_peer_type,
463 };
464 
465 /**
466  * Do not forward commands not supported by SVQ. Otherwise, the device could
467  * accept it and qemu would not know how to update the device model.
468  */
469 static bool vhost_vdpa_net_cvq_validate_cmd(const void *out_buf, size_t len)
470 {
471     struct virtio_net_ctrl_hdr ctrl;
472 
473     if (unlikely(len < sizeof(ctrl))) {
474         qemu_log_mask(LOG_GUEST_ERROR,
475                       "%s: invalid legnth of out buffer %zu\n", __func__, len);
476         return false;
477     }
478 
479     memcpy(&ctrl, out_buf, sizeof(ctrl));
480     switch (ctrl.class) {
481     case VIRTIO_NET_CTRL_MAC:
482         switch (ctrl.cmd) {
483         case VIRTIO_NET_CTRL_MAC_ADDR_SET:
484             return true;
485         default:
486             qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid mac cmd %u\n",
487                           __func__, ctrl.cmd);
488         };
489         break;
490     case VIRTIO_NET_CTRL_MQ:
491         switch (ctrl.cmd) {
492         case VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET:
493             return true;
494         default:
495             qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid mq cmd %u\n",
496                           __func__, ctrl.cmd);
497         };
498         break;
499     default:
500         qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid control class %u\n",
501                       __func__, ctrl.class);
502     };
503 
504     return false;
505 }
506 
507 /**
508  * Validate and copy control virtqueue commands.
509  *
510  * Following QEMU guidelines, we offer a copy of the buffers to the device to
511  * prevent TOCTOU bugs.
512  */
513 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq,
514                                             VirtQueueElement *elem,
515                                             void *opaque)
516 {
517     VhostVDPAState *s = opaque;
518     size_t in_len;
519     virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
520     /* Out buffer sent to both the vdpa device and the device model */
521     struct iovec out = {
522         .iov_base = s->cvq_cmd_out_buffer,
523     };
524     /* in buffer used for device model */
525     const struct iovec in = {
526         .iov_base = &status,
527         .iov_len = sizeof(status),
528     };
529     ssize_t dev_written = -EINVAL;
530     bool ok;
531 
532     out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0,
533                              s->cvq_cmd_out_buffer,
534                              vhost_vdpa_net_cvq_cmd_len());
535     ok = vhost_vdpa_net_cvq_validate_cmd(s->cvq_cmd_out_buffer, out.iov_len);
536     if (unlikely(!ok)) {
537         goto out;
538     }
539 
540     dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status));
541     if (unlikely(dev_written < 0)) {
542         goto out;
543     }
544 
545     if (unlikely(dev_written < sizeof(status))) {
546         error_report("Insufficient written data (%zu)", dev_written);
547         goto out;
548     }
549 
550     if (*s->status != VIRTIO_NET_OK) {
551         return VIRTIO_NET_ERR;
552     }
553 
554     status = VIRTIO_NET_ERR;
555     virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1);
556     if (status != VIRTIO_NET_OK) {
557         error_report("Bad CVQ processing in model");
558     }
559 
560 out:
561     in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status,
562                           sizeof(status));
563     if (unlikely(in_len < sizeof(status))) {
564         error_report("Bad device CVQ written length");
565     }
566     vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status)));
567     g_free(elem);
568     return dev_written < 0 ? dev_written : 0;
569 }
570 
571 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = {
572     .avail_handler = vhost_vdpa_net_handle_ctrl_avail,
573 };
574 
575 static NetClientState *net_vhost_vdpa_init(NetClientState *peer,
576                                            const char *device,
577                                            const char *name,
578                                            int vdpa_device_fd,
579                                            int queue_pair_index,
580                                            int nvqs,
581                                            bool is_datapath,
582                                            bool svq,
583                                            VhostIOVATree *iova_tree)
584 {
585     NetClientState *nc = NULL;
586     VhostVDPAState *s;
587     int ret = 0;
588     assert(name);
589     if (is_datapath) {
590         nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device,
591                                  name);
592     } else {
593         nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer,
594                                          device, name);
595     }
596     snprintf(nc->info_str, sizeof(nc->info_str), TYPE_VHOST_VDPA);
597     s = DO_UPCAST(VhostVDPAState, nc, nc);
598 
599     s->vhost_vdpa.device_fd = vdpa_device_fd;
600     s->vhost_vdpa.index = queue_pair_index;
601     s->vhost_vdpa.shadow_vqs_enabled = svq;
602     s->vhost_vdpa.iova_tree = iova_tree;
603     if (!is_datapath) {
604         s->cvq_cmd_out_buffer = qemu_memalign(qemu_real_host_page_size(),
605                                             vhost_vdpa_net_cvq_cmd_page_len());
606         memset(s->cvq_cmd_out_buffer, 0, vhost_vdpa_net_cvq_cmd_page_len());
607         s->status = qemu_memalign(qemu_real_host_page_size(),
608                                   vhost_vdpa_net_cvq_cmd_page_len());
609         memset(s->status, 0, vhost_vdpa_net_cvq_cmd_page_len());
610 
611         s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops;
612         s->vhost_vdpa.shadow_vq_ops_opaque = s;
613     }
614     ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);
615     if (ret) {
616         qemu_del_net_client(nc);
617         return NULL;
618     }
619     return nc;
620 }
621 
622 static int vhost_vdpa_get_iova_range(int fd,
623                                      struct vhost_vdpa_iova_range *iova_range)
624 {
625     int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range);
626 
627     return ret < 0 ? -errno : 0;
628 }
629 
630 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp)
631 {
632     int ret = ioctl(fd, VHOST_GET_FEATURES, features);
633     if (unlikely(ret < 0)) {
634         error_setg_errno(errp, errno,
635                          "Fail to query features from vhost-vDPA device");
636     }
637     return ret;
638 }
639 
640 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features,
641                                           int *has_cvq, Error **errp)
642 {
643     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
644     g_autofree struct vhost_vdpa_config *config = NULL;
645     __virtio16 *max_queue_pairs;
646     int ret;
647 
648     if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) {
649         *has_cvq = 1;
650     } else {
651         *has_cvq = 0;
652     }
653 
654     if (features & (1 << VIRTIO_NET_F_MQ)) {
655         config = g_malloc0(config_size + sizeof(*max_queue_pairs));
656         config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs);
657         config->len = sizeof(*max_queue_pairs);
658 
659         ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config);
660         if (ret) {
661             error_setg(errp, "Fail to get config from vhost-vDPA device");
662             return -ret;
663         }
664 
665         max_queue_pairs = (__virtio16 *)&config->buf;
666 
667         return lduw_le_p(max_queue_pairs);
668     }
669 
670     return 1;
671 }
672 
673 int net_init_vhost_vdpa(const Netdev *netdev, const char *name,
674                         NetClientState *peer, Error **errp)
675 {
676     const NetdevVhostVDPAOptions *opts;
677     uint64_t features;
678     int vdpa_device_fd;
679     g_autofree NetClientState **ncs = NULL;
680     g_autoptr(VhostIOVATree) iova_tree = NULL;
681     NetClientState *nc;
682     int queue_pairs, r, i = 0, has_cvq = 0;
683 
684     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA);
685     opts = &netdev->u.vhost_vdpa;
686     if (!opts->vhostdev) {
687         error_setg(errp, "vdpa character device not specified with vhostdev");
688         return -1;
689     }
690 
691     vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp);
692     if (vdpa_device_fd == -1) {
693         return -errno;
694     }
695 
696     r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp);
697     if (unlikely(r < 0)) {
698         goto err;
699     }
700 
701     queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features,
702                                                  &has_cvq, errp);
703     if (queue_pairs < 0) {
704         qemu_close(vdpa_device_fd);
705         return queue_pairs;
706     }
707 
708     if (opts->x_svq) {
709         struct vhost_vdpa_iova_range iova_range;
710 
711         uint64_t invalid_dev_features =
712             features & ~vdpa_svq_device_features &
713             /* Transport are all accepted at this point */
714             ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START,
715                              VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START);
716 
717         if (invalid_dev_features) {
718             error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64,
719                        invalid_dev_features);
720             goto err_svq;
721         }
722 
723         vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range);
724         iova_tree = vhost_iova_tree_new(iova_range.first, iova_range.last);
725     }
726 
727     ncs = g_malloc0(sizeof(*ncs) * queue_pairs);
728 
729     for (i = 0; i < queue_pairs; i++) {
730         ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
731                                      vdpa_device_fd, i, 2, true, opts->x_svq,
732                                      iova_tree);
733         if (!ncs[i])
734             goto err;
735     }
736 
737     if (has_cvq) {
738         nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
739                                  vdpa_device_fd, i, 1, false,
740                                  opts->x_svq, iova_tree);
741         if (!nc)
742             goto err;
743     }
744 
745     /* iova_tree ownership belongs to last NetClientState */
746     g_steal_pointer(&iova_tree);
747     return 0;
748 
749 err:
750     if (i) {
751         for (i--; i >= 0; i--) {
752             qemu_del_net_client(ncs[i]);
753         }
754     }
755 
756 err_svq:
757     qemu_close(vdpa_device_fd);
758 
759     return -1;
760 }
761