xref: /openbmc/linux/drivers/vdpa/mlx5/net/mlx5_vnet.c (revision 465191d6)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /* Copyright (c) 2020 Mellanox Technologies Ltd. */
3 
4 #include <linux/module.h>
5 #include <linux/vdpa.h>
6 #include <linux/vringh.h>
7 #include <uapi/linux/virtio_net.h>
8 #include <uapi/linux/virtio_ids.h>
9 #include <uapi/linux/vdpa.h>
10 #include <linux/virtio_config.h>
11 #include <linux/auxiliary_bus.h>
12 #include <linux/mlx5/cq.h>
13 #include <linux/mlx5/qp.h>
14 #include <linux/mlx5/device.h>
15 #include <linux/mlx5/driver.h>
16 #include <linux/mlx5/vport.h>
17 #include <linux/mlx5/fs.h>
18 #include <linux/mlx5/mlx5_ifc_vdpa.h>
19 #include <linux/mlx5/mpfs.h>
20 #include "mlx5_vdpa.h"
21 
22 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
23 MODULE_DESCRIPTION("Mellanox VDPA driver");
24 MODULE_LICENSE("Dual BSD/GPL");
25 
26 #define to_mlx5_vdpa_ndev(__mvdev)                                             \
27 	container_of(__mvdev, struct mlx5_vdpa_net, mvdev)
28 #define to_mvdev(__vdev) container_of((__vdev), struct mlx5_vdpa_dev, vdev)
29 
30 #define VALID_FEATURES_MASK                                                                        \
31 	(BIT_ULL(VIRTIO_NET_F_CSUM) | BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) |                                   \
32 	 BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | BIT_ULL(VIRTIO_NET_F_MTU) | BIT_ULL(VIRTIO_NET_F_MAC) |   \
33 	 BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) |                             \
34 	 BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | \
35 	 BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | BIT_ULL(VIRTIO_NET_F_HOST_ECN) | BIT_ULL(VIRTIO_NET_F_HOST_UFO) |   \
36 	 BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | BIT_ULL(VIRTIO_NET_F_STATUS) | BIT_ULL(VIRTIO_NET_F_CTRL_VQ) |      \
37 	 BIT_ULL(VIRTIO_NET_F_CTRL_RX) | BIT_ULL(VIRTIO_NET_F_CTRL_VLAN) |                                 \
38 	 BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) | BIT_ULL(VIRTIO_NET_F_GUEST_ANNOUNCE) |                      \
39 	 BIT_ULL(VIRTIO_NET_F_MQ) | BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | BIT_ULL(VIRTIO_NET_F_HASH_REPORT) |  \
40 	 BIT_ULL(VIRTIO_NET_F_RSS) | BIT_ULL(VIRTIO_NET_F_RSC_EXT) | BIT_ULL(VIRTIO_NET_F_STANDBY) |           \
41 	 BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX) | BIT_ULL(VIRTIO_F_NOTIFY_ON_EMPTY) |                          \
42 	 BIT_ULL(VIRTIO_F_ANY_LAYOUT) | BIT_ULL(VIRTIO_F_VERSION_1) | BIT_ULL(VIRTIO_F_ACCESS_PLATFORM) |      \
43 	 BIT_ULL(VIRTIO_F_RING_PACKED) | BIT_ULL(VIRTIO_F_ORDER_PLATFORM) | BIT_ULL(VIRTIO_F_SR_IOV))
44 
45 #define VALID_STATUS_MASK                                                                          \
46 	(VIRTIO_CONFIG_S_ACKNOWLEDGE | VIRTIO_CONFIG_S_DRIVER | VIRTIO_CONFIG_S_DRIVER_OK |        \
47 	 VIRTIO_CONFIG_S_FEATURES_OK | VIRTIO_CONFIG_S_NEEDS_RESET | VIRTIO_CONFIG_S_FAILED)
48 
49 #define MLX5_FEATURE(_mvdev, _feature) (!!((_mvdev)->actual_features & BIT_ULL(_feature)))
50 
51 #define MLX5V_UNTAGGED 0x1000
52 
53 struct mlx5_vdpa_net_resources {
54 	u32 tisn;
55 	u32 tdn;
56 	u32 tirn;
57 	u32 rqtn;
58 	bool valid;
59 };
60 
61 struct mlx5_vdpa_cq_buf {
62 	struct mlx5_frag_buf_ctrl fbc;
63 	struct mlx5_frag_buf frag_buf;
64 	int cqe_size;
65 	int nent;
66 };
67 
68 struct mlx5_vdpa_cq {
69 	struct mlx5_core_cq mcq;
70 	struct mlx5_vdpa_cq_buf buf;
71 	struct mlx5_db db;
72 	int cqe;
73 };
74 
75 struct mlx5_vdpa_umem {
76 	struct mlx5_frag_buf_ctrl fbc;
77 	struct mlx5_frag_buf frag_buf;
78 	int size;
79 	u32 id;
80 };
81 
82 struct mlx5_vdpa_qp {
83 	struct mlx5_core_qp mqp;
84 	struct mlx5_frag_buf frag_buf;
85 	struct mlx5_db db;
86 	u16 head;
87 	bool fw;
88 };
89 
90 struct mlx5_vq_restore_info {
91 	u32 num_ent;
92 	u64 desc_addr;
93 	u64 device_addr;
94 	u64 driver_addr;
95 	u16 avail_index;
96 	u16 used_index;
97 	bool ready;
98 	bool restore;
99 };
100 
101 struct mlx5_vdpa_virtqueue {
102 	bool ready;
103 	u64 desc_addr;
104 	u64 device_addr;
105 	u64 driver_addr;
106 	u32 num_ent;
107 
108 	/* Resources for implementing the notification channel from the device
109 	 * to the driver. fwqp is the firmware end of an RC connection; the
110 	 * other end is vqqp used by the driver. cq is is where completions are
111 	 * reported.
112 	 */
113 	struct mlx5_vdpa_cq cq;
114 	struct mlx5_vdpa_qp fwqp;
115 	struct mlx5_vdpa_qp vqqp;
116 
117 	/* umem resources are required for the virtqueue operation. They're use
118 	 * is internal and they must be provided by the driver.
119 	 */
120 	struct mlx5_vdpa_umem umem1;
121 	struct mlx5_vdpa_umem umem2;
122 	struct mlx5_vdpa_umem umem3;
123 
124 	u32 counter_set_id;
125 	bool initialized;
126 	int index;
127 	u32 virtq_id;
128 	struct mlx5_vdpa_net *ndev;
129 	u16 avail_idx;
130 	u16 used_idx;
131 	int fw_state;
132 
133 	/* keep last in the struct */
134 	struct mlx5_vq_restore_info ri;
135 };
136 
137 static bool is_index_valid(struct mlx5_vdpa_dev *mvdev, u16 idx)
138 {
139 	if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_MQ))) {
140 		if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
141 			return idx < 2;
142 		else
143 			return idx < 3;
144 	}
145 
146 	return idx <= mvdev->max_idx;
147 }
148 
149 #define MLX5V_MACVLAN_SIZE 256
150 
151 struct mlx5_vdpa_net {
152 	struct mlx5_vdpa_dev mvdev;
153 	struct mlx5_vdpa_net_resources res;
154 	struct virtio_net_config config;
155 	struct mlx5_vdpa_virtqueue *vqs;
156 	struct vdpa_callback *event_cbs;
157 
158 	/* Serialize vq resources creation and destruction. This is required
159 	 * since memory map might change and we need to destroy and create
160 	 * resources while driver in operational.
161 	 */
162 	struct rw_semaphore reslock;
163 	struct mlx5_flow_table *rxft;
164 	bool setup;
165 	u32 cur_num_vqs;
166 	u32 rqt_size;
167 	struct notifier_block nb;
168 	struct vdpa_callback config_cb;
169 	struct mlx5_vdpa_wq_ent cvq_ent;
170 	struct hlist_head macvlan_hash[MLX5V_MACVLAN_SIZE];
171 };
172 
173 struct macvlan_node {
174 	struct hlist_node hlist;
175 	struct mlx5_flow_handle *ucast_rule;
176 	struct mlx5_flow_handle *mcast_rule;
177 	u64 macvlan;
178 };
179 
180 static void free_resources(struct mlx5_vdpa_net *ndev);
181 static void init_mvqs(struct mlx5_vdpa_net *ndev);
182 static int setup_driver(struct mlx5_vdpa_dev *mvdev);
183 static void teardown_driver(struct mlx5_vdpa_net *ndev);
184 
185 static bool mlx5_vdpa_debug;
186 
187 #define MLX5_CVQ_MAX_ENT 16
188 
189 #define MLX5_LOG_VIO_FLAG(_feature)                                                                \
190 	do {                                                                                       \
191 		if (features & BIT_ULL(_feature))                                                  \
192 			mlx5_vdpa_info(mvdev, "%s\n", #_feature);                                  \
193 	} while (0)
194 
195 #define MLX5_LOG_VIO_STAT(_status)                                                                 \
196 	do {                                                                                       \
197 		if (status & (_status))                                                            \
198 			mlx5_vdpa_info(mvdev, "%s\n", #_status);                                   \
199 	} while (0)
200 
201 /* TODO: cross-endian support */
202 static inline bool mlx5_vdpa_is_little_endian(struct mlx5_vdpa_dev *mvdev)
203 {
204 	return virtio_legacy_is_little_endian() ||
205 		(mvdev->actual_features & BIT_ULL(VIRTIO_F_VERSION_1));
206 }
207 
208 static u16 mlx5vdpa16_to_cpu(struct mlx5_vdpa_dev *mvdev, __virtio16 val)
209 {
210 	return __virtio16_to_cpu(mlx5_vdpa_is_little_endian(mvdev), val);
211 }
212 
213 static __virtio16 cpu_to_mlx5vdpa16(struct mlx5_vdpa_dev *mvdev, u16 val)
214 {
215 	return __cpu_to_virtio16(mlx5_vdpa_is_little_endian(mvdev), val);
216 }
217 
218 static u16 ctrl_vq_idx(struct mlx5_vdpa_dev *mvdev)
219 {
220 	if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_MQ)))
221 		return 2;
222 
223 	return mvdev->max_vqs;
224 }
225 
226 static bool is_ctrl_vq_idx(struct mlx5_vdpa_dev *mvdev, u16 idx)
227 {
228 	return idx == ctrl_vq_idx(mvdev);
229 }
230 
231 static void print_status(struct mlx5_vdpa_dev *mvdev, u8 status, bool set)
232 {
233 	if (status & ~VALID_STATUS_MASK)
234 		mlx5_vdpa_warn(mvdev, "Warning: there are invalid status bits 0x%x\n",
235 			       status & ~VALID_STATUS_MASK);
236 
237 	if (!mlx5_vdpa_debug)
238 		return;
239 
240 	mlx5_vdpa_info(mvdev, "driver status %s", set ? "set" : "get");
241 	if (set && !status) {
242 		mlx5_vdpa_info(mvdev, "driver resets the device\n");
243 		return;
244 	}
245 
246 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_ACKNOWLEDGE);
247 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_DRIVER);
248 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_DRIVER_OK);
249 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_FEATURES_OK);
250 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_NEEDS_RESET);
251 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_FAILED);
252 }
253 
254 static void print_features(struct mlx5_vdpa_dev *mvdev, u64 features, bool set)
255 {
256 	if (features & ~VALID_FEATURES_MASK)
257 		mlx5_vdpa_warn(mvdev, "There are invalid feature bits 0x%llx\n",
258 			       features & ~VALID_FEATURES_MASK);
259 
260 	if (!mlx5_vdpa_debug)
261 		return;
262 
263 	mlx5_vdpa_info(mvdev, "driver %s feature bits:\n", set ? "sets" : "reads");
264 	if (!features)
265 		mlx5_vdpa_info(mvdev, "all feature bits are cleared\n");
266 
267 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CSUM);
268 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_CSUM);
269 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS);
270 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MTU);
271 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MAC);
272 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_TSO4);
273 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_TSO6);
274 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_ECN);
275 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_UFO);
276 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_TSO4);
277 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_TSO6);
278 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_ECN);
279 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_UFO);
280 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MRG_RXBUF);
281 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_STATUS);
282 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_VQ);
283 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_RX);
284 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_VLAN);
285 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_RX_EXTRA);
286 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_ANNOUNCE);
287 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MQ);
288 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_MAC_ADDR);
289 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HASH_REPORT);
290 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_RSS);
291 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_RSC_EXT);
292 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_STANDBY);
293 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_SPEED_DUPLEX);
294 	MLX5_LOG_VIO_FLAG(VIRTIO_F_NOTIFY_ON_EMPTY);
295 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ANY_LAYOUT);
296 	MLX5_LOG_VIO_FLAG(VIRTIO_F_VERSION_1);
297 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ACCESS_PLATFORM);
298 	MLX5_LOG_VIO_FLAG(VIRTIO_F_RING_PACKED);
299 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ORDER_PLATFORM);
300 	MLX5_LOG_VIO_FLAG(VIRTIO_F_SR_IOV);
301 }
302 
303 static int create_tis(struct mlx5_vdpa_net *ndev)
304 {
305 	struct mlx5_vdpa_dev *mvdev = &ndev->mvdev;
306 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
307 	void *tisc;
308 	int err;
309 
310 	tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
311 	MLX5_SET(tisc, tisc, transport_domain, ndev->res.tdn);
312 	err = mlx5_vdpa_create_tis(mvdev, in, &ndev->res.tisn);
313 	if (err)
314 		mlx5_vdpa_warn(mvdev, "create TIS (%d)\n", err);
315 
316 	return err;
317 }
318 
319 static void destroy_tis(struct mlx5_vdpa_net *ndev)
320 {
321 	mlx5_vdpa_destroy_tis(&ndev->mvdev, ndev->res.tisn);
322 }
323 
324 #define MLX5_VDPA_CQE_SIZE 64
325 #define MLX5_VDPA_LOG_CQE_SIZE ilog2(MLX5_VDPA_CQE_SIZE)
326 
327 static int cq_frag_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_cq_buf *buf, int nent)
328 {
329 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
330 	u8 log_wq_stride = MLX5_VDPA_LOG_CQE_SIZE;
331 	u8 log_wq_sz = MLX5_VDPA_LOG_CQE_SIZE;
332 	int err;
333 
334 	err = mlx5_frag_buf_alloc_node(ndev->mvdev.mdev, nent * MLX5_VDPA_CQE_SIZE, frag_buf,
335 				       ndev->mvdev.mdev->priv.numa_node);
336 	if (err)
337 		return err;
338 
339 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
340 
341 	buf->cqe_size = MLX5_VDPA_CQE_SIZE;
342 	buf->nent = nent;
343 
344 	return 0;
345 }
346 
347 static int umem_frag_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_umem *umem, int size)
348 {
349 	struct mlx5_frag_buf *frag_buf = &umem->frag_buf;
350 
351 	return mlx5_frag_buf_alloc_node(ndev->mvdev.mdev, size, frag_buf,
352 					ndev->mvdev.mdev->priv.numa_node);
353 }
354 
355 static void cq_frag_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_cq_buf *buf)
356 {
357 	mlx5_frag_buf_free(ndev->mvdev.mdev, &buf->frag_buf);
358 }
359 
360 static void *get_cqe(struct mlx5_vdpa_cq *vcq, int n)
361 {
362 	return mlx5_frag_buf_get_wqe(&vcq->buf.fbc, n);
363 }
364 
365 static void cq_frag_buf_init(struct mlx5_vdpa_cq *vcq, struct mlx5_vdpa_cq_buf *buf)
366 {
367 	struct mlx5_cqe64 *cqe64;
368 	void *cqe;
369 	int i;
370 
371 	for (i = 0; i < buf->nent; i++) {
372 		cqe = get_cqe(vcq, i);
373 		cqe64 = cqe;
374 		cqe64->op_own = MLX5_CQE_INVALID << 4;
375 	}
376 }
377 
378 static void *get_sw_cqe(struct mlx5_vdpa_cq *cq, int n)
379 {
380 	struct mlx5_cqe64 *cqe64 = get_cqe(cq, n & (cq->cqe - 1));
381 
382 	if (likely(get_cqe_opcode(cqe64) != MLX5_CQE_INVALID) &&
383 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & cq->cqe)))
384 		return cqe64;
385 
386 	return NULL;
387 }
388 
389 static void rx_post(struct mlx5_vdpa_qp *vqp, int n)
390 {
391 	vqp->head += n;
392 	vqp->db.db[0] = cpu_to_be32(vqp->head);
393 }
394 
395 static void qp_prepare(struct mlx5_vdpa_net *ndev, bool fw, void *in,
396 		       struct mlx5_vdpa_virtqueue *mvq, u32 num_ent)
397 {
398 	struct mlx5_vdpa_qp *vqp;
399 	__be64 *pas;
400 	void *qpc;
401 
402 	vqp = fw ? &mvq->fwqp : &mvq->vqqp;
403 	MLX5_SET(create_qp_in, in, uid, ndev->mvdev.res.uid);
404 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
405 	if (vqp->fw) {
406 		/* Firmware QP is allocated by the driver for the firmware's
407 		 * use so we can skip part of the params as they will be chosen by firmware
408 		 */
409 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
410 		MLX5_SET(qpc, qpc, rq_type, MLX5_ZERO_LEN_RQ);
411 		MLX5_SET(qpc, qpc, no_sq, 1);
412 		return;
413 	}
414 
415 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_RC);
416 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
417 	MLX5_SET(qpc, qpc, pd, ndev->mvdev.res.pdn);
418 	MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
419 	MLX5_SET(qpc, qpc, uar_page, ndev->mvdev.res.uar->index);
420 	MLX5_SET(qpc, qpc, log_page_size, vqp->frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
421 	MLX5_SET(qpc, qpc, no_sq, 1);
422 	MLX5_SET(qpc, qpc, cqn_rcv, mvq->cq.mcq.cqn);
423 	MLX5_SET(qpc, qpc, log_rq_size, ilog2(num_ent));
424 	MLX5_SET(qpc, qpc, rq_type, MLX5_NON_ZERO_RQ);
425 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, in, pas);
426 	mlx5_fill_page_frag_array(&vqp->frag_buf, pas);
427 }
428 
429 static int rq_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp, u32 num_ent)
430 {
431 	return mlx5_frag_buf_alloc_node(ndev->mvdev.mdev,
432 					num_ent * sizeof(struct mlx5_wqe_data_seg), &vqp->frag_buf,
433 					ndev->mvdev.mdev->priv.numa_node);
434 }
435 
436 static void rq_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp)
437 {
438 	mlx5_frag_buf_free(ndev->mvdev.mdev, &vqp->frag_buf);
439 }
440 
441 static int qp_create(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
442 		     struct mlx5_vdpa_qp *vqp)
443 {
444 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
445 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
446 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
447 	void *qpc;
448 	void *in;
449 	int err;
450 
451 	if (!vqp->fw) {
452 		vqp = &mvq->vqqp;
453 		err = rq_buf_alloc(ndev, vqp, mvq->num_ent);
454 		if (err)
455 			return err;
456 
457 		err = mlx5_db_alloc(ndev->mvdev.mdev, &vqp->db);
458 		if (err)
459 			goto err_db;
460 		inlen += vqp->frag_buf.npages * sizeof(__be64);
461 	}
462 
463 	in = kzalloc(inlen, GFP_KERNEL);
464 	if (!in) {
465 		err = -ENOMEM;
466 		goto err_kzalloc;
467 	}
468 
469 	qp_prepare(ndev, vqp->fw, in, mvq, mvq->num_ent);
470 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
471 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_RC);
472 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
473 	MLX5_SET(qpc, qpc, pd, ndev->mvdev.res.pdn);
474 	MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
475 	if (!vqp->fw)
476 		MLX5_SET64(qpc, qpc, dbr_addr, vqp->db.dma);
477 	MLX5_SET(create_qp_in, in, opcode, MLX5_CMD_OP_CREATE_QP);
478 	err = mlx5_cmd_exec(mdev, in, inlen, out, sizeof(out));
479 	kfree(in);
480 	if (err)
481 		goto err_kzalloc;
482 
483 	vqp->mqp.uid = ndev->mvdev.res.uid;
484 	vqp->mqp.qpn = MLX5_GET(create_qp_out, out, qpn);
485 
486 	if (!vqp->fw)
487 		rx_post(vqp, mvq->num_ent);
488 
489 	return 0;
490 
491 err_kzalloc:
492 	if (!vqp->fw)
493 		mlx5_db_free(ndev->mvdev.mdev, &vqp->db);
494 err_db:
495 	if (!vqp->fw)
496 		rq_buf_free(ndev, vqp);
497 
498 	return err;
499 }
500 
501 static void qp_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp)
502 {
503 	u32 in[MLX5_ST_SZ_DW(destroy_qp_in)] = {};
504 
505 	MLX5_SET(destroy_qp_in, in, opcode, MLX5_CMD_OP_DESTROY_QP);
506 	MLX5_SET(destroy_qp_in, in, qpn, vqp->mqp.qpn);
507 	MLX5_SET(destroy_qp_in, in, uid, ndev->mvdev.res.uid);
508 	if (mlx5_cmd_exec_in(ndev->mvdev.mdev, destroy_qp, in))
509 		mlx5_vdpa_warn(&ndev->mvdev, "destroy qp 0x%x\n", vqp->mqp.qpn);
510 	if (!vqp->fw) {
511 		mlx5_db_free(ndev->mvdev.mdev, &vqp->db);
512 		rq_buf_free(ndev, vqp);
513 	}
514 }
515 
516 static void *next_cqe_sw(struct mlx5_vdpa_cq *cq)
517 {
518 	return get_sw_cqe(cq, cq->mcq.cons_index);
519 }
520 
521 static int mlx5_vdpa_poll_one(struct mlx5_vdpa_cq *vcq)
522 {
523 	struct mlx5_cqe64 *cqe64;
524 
525 	cqe64 = next_cqe_sw(vcq);
526 	if (!cqe64)
527 		return -EAGAIN;
528 
529 	vcq->mcq.cons_index++;
530 	return 0;
531 }
532 
533 static void mlx5_vdpa_handle_completions(struct mlx5_vdpa_virtqueue *mvq, int num)
534 {
535 	struct mlx5_vdpa_net *ndev = mvq->ndev;
536 	struct vdpa_callback *event_cb;
537 
538 	event_cb = &ndev->event_cbs[mvq->index];
539 	mlx5_cq_set_ci(&mvq->cq.mcq);
540 
541 	/* make sure CQ cosumer update is visible to the hardware before updating
542 	 * RX doorbell record.
543 	 */
544 	dma_wmb();
545 	rx_post(&mvq->vqqp, num);
546 	if (event_cb->callback)
547 		event_cb->callback(event_cb->private);
548 }
549 
550 static void mlx5_vdpa_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe)
551 {
552 	struct mlx5_vdpa_virtqueue *mvq = container_of(mcq, struct mlx5_vdpa_virtqueue, cq.mcq);
553 	struct mlx5_vdpa_net *ndev = mvq->ndev;
554 	void __iomem *uar_page = ndev->mvdev.res.uar->map;
555 	int num = 0;
556 
557 	while (!mlx5_vdpa_poll_one(&mvq->cq)) {
558 		num++;
559 		if (num > mvq->num_ent / 2) {
560 			/* If completions keep coming while we poll, we want to
561 			 * let the hardware know that we consumed them by
562 			 * updating the doorbell record.  We also let vdpa core
563 			 * know about this so it passes it on the virtio driver
564 			 * on the guest.
565 			 */
566 			mlx5_vdpa_handle_completions(mvq, num);
567 			num = 0;
568 		}
569 	}
570 
571 	if (num)
572 		mlx5_vdpa_handle_completions(mvq, num);
573 
574 	mlx5_cq_arm(&mvq->cq.mcq, MLX5_CQ_DB_REQ_NOT, uar_page, mvq->cq.mcq.cons_index);
575 }
576 
577 static int cq_create(struct mlx5_vdpa_net *ndev, u16 idx, u32 num_ent)
578 {
579 	struct mlx5_vdpa_virtqueue *mvq = &ndev->vqs[idx];
580 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
581 	void __iomem *uar_page = ndev->mvdev.res.uar->map;
582 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
583 	struct mlx5_vdpa_cq *vcq = &mvq->cq;
584 	__be64 *pas;
585 	int inlen;
586 	void *cqc;
587 	void *in;
588 	int err;
589 	int eqn;
590 
591 	err = mlx5_db_alloc(mdev, &vcq->db);
592 	if (err)
593 		return err;
594 
595 	vcq->mcq.set_ci_db = vcq->db.db;
596 	vcq->mcq.arm_db = vcq->db.db + 1;
597 	vcq->mcq.cqe_sz = 64;
598 
599 	err = cq_frag_buf_alloc(ndev, &vcq->buf, num_ent);
600 	if (err)
601 		goto err_db;
602 
603 	cq_frag_buf_init(vcq, &vcq->buf);
604 
605 	inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
606 		MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * vcq->buf.frag_buf.npages;
607 	in = kzalloc(inlen, GFP_KERNEL);
608 	if (!in) {
609 		err = -ENOMEM;
610 		goto err_vzalloc;
611 	}
612 
613 	MLX5_SET(create_cq_in, in, uid, ndev->mvdev.res.uid);
614 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas);
615 	mlx5_fill_page_frag_array(&vcq->buf.frag_buf, pas);
616 
617 	cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
618 	MLX5_SET(cqc, cqc, log_page_size, vcq->buf.frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
619 
620 	/* Use vector 0 by default. Consider adding code to choose least used
621 	 * vector.
622 	 */
623 	err = mlx5_vector2eqn(mdev, 0, &eqn);
624 	if (err)
625 		goto err_vec;
626 
627 	cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
628 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(num_ent));
629 	MLX5_SET(cqc, cqc, uar_page, ndev->mvdev.res.uar->index);
630 	MLX5_SET(cqc, cqc, c_eqn_or_apu_element, eqn);
631 	MLX5_SET64(cqc, cqc, dbr_addr, vcq->db.dma);
632 
633 	err = mlx5_core_create_cq(mdev, &vcq->mcq, in, inlen, out, sizeof(out));
634 	if (err)
635 		goto err_vec;
636 
637 	vcq->mcq.comp = mlx5_vdpa_cq_comp;
638 	vcq->cqe = num_ent;
639 	vcq->mcq.set_ci_db = vcq->db.db;
640 	vcq->mcq.arm_db = vcq->db.db + 1;
641 	mlx5_cq_arm(&mvq->cq.mcq, MLX5_CQ_DB_REQ_NOT, uar_page, mvq->cq.mcq.cons_index);
642 	kfree(in);
643 	return 0;
644 
645 err_vec:
646 	kfree(in);
647 err_vzalloc:
648 	cq_frag_buf_free(ndev, &vcq->buf);
649 err_db:
650 	mlx5_db_free(ndev->mvdev.mdev, &vcq->db);
651 	return err;
652 }
653 
654 static void cq_destroy(struct mlx5_vdpa_net *ndev, u16 idx)
655 {
656 	struct mlx5_vdpa_virtqueue *mvq = &ndev->vqs[idx];
657 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
658 	struct mlx5_vdpa_cq *vcq = &mvq->cq;
659 
660 	if (mlx5_core_destroy_cq(mdev, &vcq->mcq)) {
661 		mlx5_vdpa_warn(&ndev->mvdev, "destroy CQ 0x%x\n", vcq->mcq.cqn);
662 		return;
663 	}
664 	cq_frag_buf_free(ndev, &vcq->buf);
665 	mlx5_db_free(ndev->mvdev.mdev, &vcq->db);
666 }
667 
668 static void set_umem_size(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num,
669 			  struct mlx5_vdpa_umem **umemp)
670 {
671 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
672 	int p_a;
673 	int p_b;
674 
675 	switch (num) {
676 	case 1:
677 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_1_buffer_param_a);
678 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_1_buffer_param_b);
679 		*umemp = &mvq->umem1;
680 		break;
681 	case 2:
682 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_2_buffer_param_a);
683 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_2_buffer_param_b);
684 		*umemp = &mvq->umem2;
685 		break;
686 	case 3:
687 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_3_buffer_param_a);
688 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_3_buffer_param_b);
689 		*umemp = &mvq->umem3;
690 		break;
691 	}
692 	(*umemp)->size = p_a * mvq->num_ent + p_b;
693 }
694 
695 static void umem_frag_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_umem *umem)
696 {
697 	mlx5_frag_buf_free(ndev->mvdev.mdev, &umem->frag_buf);
698 }
699 
700 static int create_umem(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num)
701 {
702 	int inlen;
703 	u32 out[MLX5_ST_SZ_DW(create_umem_out)] = {};
704 	void *um;
705 	void *in;
706 	int err;
707 	__be64 *pas;
708 	struct mlx5_vdpa_umem *umem;
709 
710 	set_umem_size(ndev, mvq, num, &umem);
711 	err = umem_frag_buf_alloc(ndev, umem, umem->size);
712 	if (err)
713 		return err;
714 
715 	inlen = MLX5_ST_SZ_BYTES(create_umem_in) + MLX5_ST_SZ_BYTES(mtt) * umem->frag_buf.npages;
716 
717 	in = kzalloc(inlen, GFP_KERNEL);
718 	if (!in) {
719 		err = -ENOMEM;
720 		goto err_in;
721 	}
722 
723 	MLX5_SET(create_umem_in, in, opcode, MLX5_CMD_OP_CREATE_UMEM);
724 	MLX5_SET(create_umem_in, in, uid, ndev->mvdev.res.uid);
725 	um = MLX5_ADDR_OF(create_umem_in, in, umem);
726 	MLX5_SET(umem, um, log_page_size, umem->frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
727 	MLX5_SET64(umem, um, num_of_mtt, umem->frag_buf.npages);
728 
729 	pas = (__be64 *)MLX5_ADDR_OF(umem, um, mtt[0]);
730 	mlx5_fill_page_frag_array_perm(&umem->frag_buf, pas, MLX5_MTT_PERM_RW);
731 
732 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
733 	if (err) {
734 		mlx5_vdpa_warn(&ndev->mvdev, "create umem(%d)\n", err);
735 		goto err_cmd;
736 	}
737 
738 	kfree(in);
739 	umem->id = MLX5_GET(create_umem_out, out, umem_id);
740 
741 	return 0;
742 
743 err_cmd:
744 	kfree(in);
745 err_in:
746 	umem_frag_buf_free(ndev, umem);
747 	return err;
748 }
749 
750 static void umem_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num)
751 {
752 	u32 in[MLX5_ST_SZ_DW(destroy_umem_in)] = {};
753 	u32 out[MLX5_ST_SZ_DW(destroy_umem_out)] = {};
754 	struct mlx5_vdpa_umem *umem;
755 
756 	switch (num) {
757 	case 1:
758 		umem = &mvq->umem1;
759 		break;
760 	case 2:
761 		umem = &mvq->umem2;
762 		break;
763 	case 3:
764 		umem = &mvq->umem3;
765 		break;
766 	}
767 
768 	MLX5_SET(destroy_umem_in, in, opcode, MLX5_CMD_OP_DESTROY_UMEM);
769 	MLX5_SET(destroy_umem_in, in, umem_id, umem->id);
770 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out)))
771 		return;
772 
773 	umem_frag_buf_free(ndev, umem);
774 }
775 
776 static int umems_create(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
777 {
778 	int num;
779 	int err;
780 
781 	for (num = 1; num <= 3; num++) {
782 		err = create_umem(ndev, mvq, num);
783 		if (err)
784 			goto err_umem;
785 	}
786 	return 0;
787 
788 err_umem:
789 	for (num--; num > 0; num--)
790 		umem_destroy(ndev, mvq, num);
791 
792 	return err;
793 }
794 
795 static void umems_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
796 {
797 	int num;
798 
799 	for (num = 3; num > 0; num--)
800 		umem_destroy(ndev, mvq, num);
801 }
802 
803 static int get_queue_type(struct mlx5_vdpa_net *ndev)
804 {
805 	u32 type_mask;
806 
807 	type_mask = MLX5_CAP_DEV_VDPA_EMULATION(ndev->mvdev.mdev, virtio_queue_type);
808 
809 	/* prefer split queue */
810 	if (type_mask & MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_SPLIT)
811 		return MLX5_VIRTIO_EMULATION_VIRTIO_QUEUE_TYPE_SPLIT;
812 
813 	WARN_ON(!(type_mask & MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_PACKED));
814 
815 	return MLX5_VIRTIO_EMULATION_VIRTIO_QUEUE_TYPE_PACKED;
816 }
817 
818 static bool vq_is_tx(u16 idx)
819 {
820 	return idx % 2;
821 }
822 
823 static u16 get_features_12_3(u64 features)
824 {
825 	return (!!(features & BIT_ULL(VIRTIO_NET_F_HOST_TSO4)) << 9) |
826 	       (!!(features & BIT_ULL(VIRTIO_NET_F_HOST_TSO6)) << 8) |
827 	       (!!(features & BIT_ULL(VIRTIO_NET_F_CSUM)) << 7) |
828 	       (!!(features & BIT_ULL(VIRTIO_NET_F_GUEST_CSUM)) << 6);
829 }
830 
831 static bool counters_supported(const struct mlx5_vdpa_dev *mvdev)
832 {
833 	return MLX5_CAP_GEN_64(mvdev->mdev, general_obj_types) &
834 	       BIT_ULL(MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
835 }
836 
837 static int create_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
838 {
839 	int inlen = MLX5_ST_SZ_BYTES(create_virtio_net_q_in);
840 	u32 out[MLX5_ST_SZ_DW(create_virtio_net_q_out)] = {};
841 	void *obj_context;
842 	void *cmd_hdr;
843 	void *vq_ctx;
844 	void *in;
845 	int err;
846 
847 	err = umems_create(ndev, mvq);
848 	if (err)
849 		return err;
850 
851 	in = kzalloc(inlen, GFP_KERNEL);
852 	if (!in) {
853 		err = -ENOMEM;
854 		goto err_alloc;
855 	}
856 
857 	cmd_hdr = MLX5_ADDR_OF(create_virtio_net_q_in, in, general_obj_in_cmd_hdr);
858 
859 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_CREATE_GENERAL_OBJECT);
860 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
861 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
862 
863 	obj_context = MLX5_ADDR_OF(create_virtio_net_q_in, in, obj_context);
864 	MLX5_SET(virtio_net_q_object, obj_context, hw_available_index, mvq->avail_idx);
865 	MLX5_SET(virtio_net_q_object, obj_context, hw_used_index, mvq->used_idx);
866 	MLX5_SET(virtio_net_q_object, obj_context, queue_feature_bit_mask_12_3,
867 		 get_features_12_3(ndev->mvdev.actual_features));
868 	vq_ctx = MLX5_ADDR_OF(virtio_net_q_object, obj_context, virtio_q_context);
869 	MLX5_SET(virtio_q, vq_ctx, virtio_q_type, get_queue_type(ndev));
870 
871 	if (vq_is_tx(mvq->index))
872 		MLX5_SET(virtio_net_q_object, obj_context, tisn_or_qpn, ndev->res.tisn);
873 
874 	MLX5_SET(virtio_q, vq_ctx, event_mode, MLX5_VIRTIO_Q_EVENT_MODE_QP_MODE);
875 	MLX5_SET(virtio_q, vq_ctx, queue_index, mvq->index);
876 	MLX5_SET(virtio_q, vq_ctx, event_qpn_or_msix, mvq->fwqp.mqp.qpn);
877 	MLX5_SET(virtio_q, vq_ctx, queue_size, mvq->num_ent);
878 	MLX5_SET(virtio_q, vq_ctx, virtio_version_1_0,
879 		 !!(ndev->mvdev.actual_features & BIT_ULL(VIRTIO_F_VERSION_1)));
880 	MLX5_SET64(virtio_q, vq_ctx, desc_addr, mvq->desc_addr);
881 	MLX5_SET64(virtio_q, vq_ctx, used_addr, mvq->device_addr);
882 	MLX5_SET64(virtio_q, vq_ctx, available_addr, mvq->driver_addr);
883 	MLX5_SET(virtio_q, vq_ctx, virtio_q_mkey, ndev->mvdev.mr.mkey);
884 	MLX5_SET(virtio_q, vq_ctx, umem_1_id, mvq->umem1.id);
885 	MLX5_SET(virtio_q, vq_ctx, umem_1_size, mvq->umem1.size);
886 	MLX5_SET(virtio_q, vq_ctx, umem_2_id, mvq->umem2.id);
887 	MLX5_SET(virtio_q, vq_ctx, umem_2_size, mvq->umem2.size);
888 	MLX5_SET(virtio_q, vq_ctx, umem_3_id, mvq->umem3.id);
889 	MLX5_SET(virtio_q, vq_ctx, umem_3_size, mvq->umem3.size);
890 	MLX5_SET(virtio_q, vq_ctx, pd, ndev->mvdev.res.pdn);
891 	if (counters_supported(&ndev->mvdev))
892 		MLX5_SET(virtio_q, vq_ctx, counter_set_id, mvq->counter_set_id);
893 
894 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
895 	if (err)
896 		goto err_cmd;
897 
898 	kfree(in);
899 	mvq->virtq_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
900 
901 	return 0;
902 
903 err_cmd:
904 	kfree(in);
905 err_alloc:
906 	umems_destroy(ndev, mvq);
907 	return err;
908 }
909 
910 static void destroy_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
911 {
912 	u32 in[MLX5_ST_SZ_DW(destroy_virtio_net_q_in)] = {};
913 	u32 out[MLX5_ST_SZ_DW(destroy_virtio_net_q_out)] = {};
914 
915 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.opcode,
916 		 MLX5_CMD_OP_DESTROY_GENERAL_OBJECT);
917 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.obj_id, mvq->virtq_id);
918 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.uid, ndev->mvdev.res.uid);
919 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.obj_type,
920 		 MLX5_OBJ_TYPE_VIRTIO_NET_Q);
921 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out))) {
922 		mlx5_vdpa_warn(&ndev->mvdev, "destroy virtqueue 0x%x\n", mvq->virtq_id);
923 		return;
924 	}
925 	umems_destroy(ndev, mvq);
926 }
927 
928 static u32 get_rqpn(struct mlx5_vdpa_virtqueue *mvq, bool fw)
929 {
930 	return fw ? mvq->vqqp.mqp.qpn : mvq->fwqp.mqp.qpn;
931 }
932 
933 static u32 get_qpn(struct mlx5_vdpa_virtqueue *mvq, bool fw)
934 {
935 	return fw ? mvq->fwqp.mqp.qpn : mvq->vqqp.mqp.qpn;
936 }
937 
938 static void alloc_inout(struct mlx5_vdpa_net *ndev, int cmd, void **in, int *inlen, void **out,
939 			int *outlen, u32 qpn, u32 rqpn)
940 {
941 	void *qpc;
942 	void *pp;
943 
944 	switch (cmd) {
945 	case MLX5_CMD_OP_2RST_QP:
946 		*inlen = MLX5_ST_SZ_BYTES(qp_2rst_in);
947 		*outlen = MLX5_ST_SZ_BYTES(qp_2rst_out);
948 		*in = kzalloc(*inlen, GFP_KERNEL);
949 		*out = kzalloc(*outlen, GFP_KERNEL);
950 		if (!*in || !*out)
951 			goto outerr;
952 
953 		MLX5_SET(qp_2rst_in, *in, opcode, cmd);
954 		MLX5_SET(qp_2rst_in, *in, uid, ndev->mvdev.res.uid);
955 		MLX5_SET(qp_2rst_in, *in, qpn, qpn);
956 		break;
957 	case MLX5_CMD_OP_RST2INIT_QP:
958 		*inlen = MLX5_ST_SZ_BYTES(rst2init_qp_in);
959 		*outlen = MLX5_ST_SZ_BYTES(rst2init_qp_out);
960 		*in = kzalloc(*inlen, GFP_KERNEL);
961 		*out = kzalloc(MLX5_ST_SZ_BYTES(rst2init_qp_out), GFP_KERNEL);
962 		if (!*in || !*out)
963 			goto outerr;
964 
965 		MLX5_SET(rst2init_qp_in, *in, opcode, cmd);
966 		MLX5_SET(rst2init_qp_in, *in, uid, ndev->mvdev.res.uid);
967 		MLX5_SET(rst2init_qp_in, *in, qpn, qpn);
968 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
969 		MLX5_SET(qpc, qpc, remote_qpn, rqpn);
970 		MLX5_SET(qpc, qpc, rwe, 1);
971 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
972 		MLX5_SET(ads, pp, vhca_port_num, 1);
973 		break;
974 	case MLX5_CMD_OP_INIT2RTR_QP:
975 		*inlen = MLX5_ST_SZ_BYTES(init2rtr_qp_in);
976 		*outlen = MLX5_ST_SZ_BYTES(init2rtr_qp_out);
977 		*in = kzalloc(*inlen, GFP_KERNEL);
978 		*out = kzalloc(MLX5_ST_SZ_BYTES(init2rtr_qp_out), GFP_KERNEL);
979 		if (!*in || !*out)
980 			goto outerr;
981 
982 		MLX5_SET(init2rtr_qp_in, *in, opcode, cmd);
983 		MLX5_SET(init2rtr_qp_in, *in, uid, ndev->mvdev.res.uid);
984 		MLX5_SET(init2rtr_qp_in, *in, qpn, qpn);
985 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
986 		MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
987 		MLX5_SET(qpc, qpc, log_msg_max, 30);
988 		MLX5_SET(qpc, qpc, remote_qpn, rqpn);
989 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
990 		MLX5_SET(ads, pp, fl, 1);
991 		break;
992 	case MLX5_CMD_OP_RTR2RTS_QP:
993 		*inlen = MLX5_ST_SZ_BYTES(rtr2rts_qp_in);
994 		*outlen = MLX5_ST_SZ_BYTES(rtr2rts_qp_out);
995 		*in = kzalloc(*inlen, GFP_KERNEL);
996 		*out = kzalloc(MLX5_ST_SZ_BYTES(rtr2rts_qp_out), GFP_KERNEL);
997 		if (!*in || !*out)
998 			goto outerr;
999 
1000 		MLX5_SET(rtr2rts_qp_in, *in, opcode, cmd);
1001 		MLX5_SET(rtr2rts_qp_in, *in, uid, ndev->mvdev.res.uid);
1002 		MLX5_SET(rtr2rts_qp_in, *in, qpn, qpn);
1003 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
1004 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
1005 		MLX5_SET(ads, pp, ack_timeout, 14);
1006 		MLX5_SET(qpc, qpc, retry_count, 7);
1007 		MLX5_SET(qpc, qpc, rnr_retry, 7);
1008 		break;
1009 	default:
1010 		goto outerr_nullify;
1011 	}
1012 
1013 	return;
1014 
1015 outerr:
1016 	kfree(*in);
1017 	kfree(*out);
1018 outerr_nullify:
1019 	*in = NULL;
1020 	*out = NULL;
1021 }
1022 
1023 static void free_inout(void *in, void *out)
1024 {
1025 	kfree(in);
1026 	kfree(out);
1027 }
1028 
1029 /* Two QPs are used by each virtqueue. One is used by the driver and one by
1030  * firmware. The fw argument indicates whether the subjected QP is the one used
1031  * by firmware.
1032  */
1033 static int modify_qp(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, bool fw, int cmd)
1034 {
1035 	int outlen;
1036 	int inlen;
1037 	void *out;
1038 	void *in;
1039 	int err;
1040 
1041 	alloc_inout(ndev, cmd, &in, &inlen, &out, &outlen, get_qpn(mvq, fw), get_rqpn(mvq, fw));
1042 	if (!in || !out)
1043 		return -ENOMEM;
1044 
1045 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, outlen);
1046 	free_inout(in, out);
1047 	return err;
1048 }
1049 
1050 static int connect_qps(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1051 {
1052 	int err;
1053 
1054 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_2RST_QP);
1055 	if (err)
1056 		return err;
1057 
1058 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_2RST_QP);
1059 	if (err)
1060 		return err;
1061 
1062 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_RST2INIT_QP);
1063 	if (err)
1064 		return err;
1065 
1066 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_RST2INIT_QP);
1067 	if (err)
1068 		return err;
1069 
1070 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_INIT2RTR_QP);
1071 	if (err)
1072 		return err;
1073 
1074 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_INIT2RTR_QP);
1075 	if (err)
1076 		return err;
1077 
1078 	return modify_qp(ndev, mvq, true, MLX5_CMD_OP_RTR2RTS_QP);
1079 }
1080 
1081 struct mlx5_virtq_attr {
1082 	u8 state;
1083 	u16 available_index;
1084 	u16 used_index;
1085 };
1086 
1087 static int query_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
1088 			   struct mlx5_virtq_attr *attr)
1089 {
1090 	int outlen = MLX5_ST_SZ_BYTES(query_virtio_net_q_out);
1091 	u32 in[MLX5_ST_SZ_DW(query_virtio_net_q_in)] = {};
1092 	void *out;
1093 	void *obj_context;
1094 	void *cmd_hdr;
1095 	int err;
1096 
1097 	out = kzalloc(outlen, GFP_KERNEL);
1098 	if (!out)
1099 		return -ENOMEM;
1100 
1101 	cmd_hdr = MLX5_ADDR_OF(query_virtio_net_q_in, in, general_obj_in_cmd_hdr);
1102 
1103 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_QUERY_GENERAL_OBJECT);
1104 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
1105 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->virtq_id);
1106 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1107 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, outlen);
1108 	if (err)
1109 		goto err_cmd;
1110 
1111 	obj_context = MLX5_ADDR_OF(query_virtio_net_q_out, out, obj_context);
1112 	memset(attr, 0, sizeof(*attr));
1113 	attr->state = MLX5_GET(virtio_net_q_object, obj_context, state);
1114 	attr->available_index = MLX5_GET(virtio_net_q_object, obj_context, hw_available_index);
1115 	attr->used_index = MLX5_GET(virtio_net_q_object, obj_context, hw_used_index);
1116 	kfree(out);
1117 	return 0;
1118 
1119 err_cmd:
1120 	kfree(out);
1121 	return err;
1122 }
1123 
1124 static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int state)
1125 {
1126 	int inlen = MLX5_ST_SZ_BYTES(modify_virtio_net_q_in);
1127 	u32 out[MLX5_ST_SZ_DW(modify_virtio_net_q_out)] = {};
1128 	void *obj_context;
1129 	void *cmd_hdr;
1130 	void *in;
1131 	int err;
1132 
1133 	in = kzalloc(inlen, GFP_KERNEL);
1134 	if (!in)
1135 		return -ENOMEM;
1136 
1137 	cmd_hdr = MLX5_ADDR_OF(modify_virtio_net_q_in, in, general_obj_in_cmd_hdr);
1138 
1139 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_MODIFY_GENERAL_OBJECT);
1140 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
1141 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->virtq_id);
1142 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1143 
1144 	obj_context = MLX5_ADDR_OF(modify_virtio_net_q_in, in, obj_context);
1145 	MLX5_SET64(virtio_net_q_object, obj_context, modify_field_select,
1146 		   MLX5_VIRTQ_MODIFY_MASK_STATE);
1147 	MLX5_SET(virtio_net_q_object, obj_context, state, state);
1148 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
1149 	kfree(in);
1150 	if (!err)
1151 		mvq->fw_state = state;
1152 
1153 	return err;
1154 }
1155 
1156 static int counter_set_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1157 {
1158 	u32 in[MLX5_ST_SZ_DW(create_virtio_q_counters_in)] = {};
1159 	u32 out[MLX5_ST_SZ_DW(create_virtio_q_counters_out)] = {};
1160 	void *cmd_hdr;
1161 	int err;
1162 
1163 	if (!counters_supported(&ndev->mvdev))
1164 		return 0;
1165 
1166 	cmd_hdr = MLX5_ADDR_OF(create_virtio_q_counters_in, in, hdr);
1167 
1168 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_CREATE_GENERAL_OBJECT);
1169 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
1170 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1171 
1172 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out));
1173 	if (err)
1174 		return err;
1175 
1176 	mvq->counter_set_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
1177 
1178 	return 0;
1179 }
1180 
1181 static void counter_set_dealloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1182 {
1183 	u32 in[MLX5_ST_SZ_DW(destroy_virtio_q_counters_in)] = {};
1184 	u32 out[MLX5_ST_SZ_DW(destroy_virtio_q_counters_out)] = {};
1185 
1186 	if (!counters_supported(&ndev->mvdev))
1187 		return;
1188 
1189 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.opcode, MLX5_CMD_OP_DESTROY_GENERAL_OBJECT);
1190 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.obj_id, mvq->counter_set_id);
1191 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.uid, ndev->mvdev.res.uid);
1192 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
1193 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out)))
1194 		mlx5_vdpa_warn(&ndev->mvdev, "dealloc counter set 0x%x\n", mvq->counter_set_id);
1195 }
1196 
1197 static int setup_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1198 {
1199 	u16 idx = mvq->index;
1200 	int err;
1201 
1202 	if (!mvq->num_ent)
1203 		return 0;
1204 
1205 	if (mvq->initialized)
1206 		return 0;
1207 
1208 	err = cq_create(ndev, idx, mvq->num_ent);
1209 	if (err)
1210 		return err;
1211 
1212 	err = qp_create(ndev, mvq, &mvq->fwqp);
1213 	if (err)
1214 		goto err_fwqp;
1215 
1216 	err = qp_create(ndev, mvq, &mvq->vqqp);
1217 	if (err)
1218 		goto err_vqqp;
1219 
1220 	err = connect_qps(ndev, mvq);
1221 	if (err)
1222 		goto err_connect;
1223 
1224 	err = counter_set_alloc(ndev, mvq);
1225 	if (err)
1226 		goto err_counter;
1227 
1228 	err = create_virtqueue(ndev, mvq);
1229 	if (err)
1230 		goto err_connect;
1231 
1232 	if (mvq->ready) {
1233 		err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
1234 		if (err) {
1235 			mlx5_vdpa_warn(&ndev->mvdev, "failed to modify to ready vq idx %d(%d)\n",
1236 				       idx, err);
1237 			goto err_connect;
1238 		}
1239 	}
1240 
1241 	mvq->initialized = true;
1242 	return 0;
1243 
1244 err_connect:
1245 	counter_set_dealloc(ndev, mvq);
1246 err_counter:
1247 	qp_destroy(ndev, &mvq->vqqp);
1248 err_vqqp:
1249 	qp_destroy(ndev, &mvq->fwqp);
1250 err_fwqp:
1251 	cq_destroy(ndev, idx);
1252 	return err;
1253 }
1254 
1255 static void suspend_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1256 {
1257 	struct mlx5_virtq_attr attr;
1258 
1259 	if (!mvq->initialized)
1260 		return;
1261 
1262 	if (mvq->fw_state != MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY)
1263 		return;
1264 
1265 	if (modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND))
1266 		mlx5_vdpa_warn(&ndev->mvdev, "modify to suspend failed\n");
1267 
1268 	if (query_virtqueue(ndev, mvq, &attr)) {
1269 		mlx5_vdpa_warn(&ndev->mvdev, "failed to query virtqueue\n");
1270 		return;
1271 	}
1272 	mvq->avail_idx = attr.available_index;
1273 	mvq->used_idx = attr.used_index;
1274 }
1275 
1276 static void suspend_vqs(struct mlx5_vdpa_net *ndev)
1277 {
1278 	int i;
1279 
1280 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
1281 		suspend_vq(ndev, &ndev->vqs[i]);
1282 }
1283 
1284 static void teardown_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1285 {
1286 	if (!mvq->initialized)
1287 		return;
1288 
1289 	suspend_vq(ndev, mvq);
1290 	destroy_virtqueue(ndev, mvq);
1291 	counter_set_dealloc(ndev, mvq);
1292 	qp_destroy(ndev, &mvq->vqqp);
1293 	qp_destroy(ndev, &mvq->fwqp);
1294 	cq_destroy(ndev, mvq->index);
1295 	mvq->initialized = false;
1296 }
1297 
1298 static int create_rqt(struct mlx5_vdpa_net *ndev)
1299 {
1300 	__be32 *list;
1301 	void *rqtc;
1302 	int inlen;
1303 	void *in;
1304 	int i, j;
1305 	int err;
1306 
1307 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + ndev->rqt_size * MLX5_ST_SZ_BYTES(rq_num);
1308 	in = kzalloc(inlen, GFP_KERNEL);
1309 	if (!in)
1310 		return -ENOMEM;
1311 
1312 	MLX5_SET(create_rqt_in, in, uid, ndev->mvdev.res.uid);
1313 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
1314 
1315 	MLX5_SET(rqtc, rqtc, list_q_type, MLX5_RQTC_LIST_Q_TYPE_VIRTIO_NET_Q);
1316 	MLX5_SET(rqtc, rqtc, rqt_max_size, ndev->rqt_size);
1317 	list = MLX5_ADDR_OF(rqtc, rqtc, rq_num[0]);
1318 	for (i = 0, j = 0; i < ndev->rqt_size; i++, j += 2)
1319 		list[i] = cpu_to_be32(ndev->vqs[j % ndev->cur_num_vqs].virtq_id);
1320 
1321 	MLX5_SET(rqtc, rqtc, rqt_actual_size, ndev->rqt_size);
1322 	err = mlx5_vdpa_create_rqt(&ndev->mvdev, in, inlen, &ndev->res.rqtn);
1323 	kfree(in);
1324 	if (err)
1325 		return err;
1326 
1327 	return 0;
1328 }
1329 
1330 #define MLX5_MODIFY_RQT_NUM_RQS ((u64)1)
1331 
1332 static int modify_rqt(struct mlx5_vdpa_net *ndev, int num)
1333 {
1334 	__be32 *list;
1335 	void *rqtc;
1336 	int inlen;
1337 	void *in;
1338 	int i, j;
1339 	int err;
1340 
1341 	inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + ndev->rqt_size * MLX5_ST_SZ_BYTES(rq_num);
1342 	in = kzalloc(inlen, GFP_KERNEL);
1343 	if (!in)
1344 		return -ENOMEM;
1345 
1346 	MLX5_SET(modify_rqt_in, in, uid, ndev->mvdev.res.uid);
1347 	MLX5_SET64(modify_rqt_in, in, bitmask, MLX5_MODIFY_RQT_NUM_RQS);
1348 	rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
1349 	MLX5_SET(rqtc, rqtc, list_q_type, MLX5_RQTC_LIST_Q_TYPE_VIRTIO_NET_Q);
1350 
1351 	list = MLX5_ADDR_OF(rqtc, rqtc, rq_num[0]);
1352 	for (i = 0, j = 0; i < ndev->rqt_size; i++, j += 2)
1353 		list[i] = cpu_to_be32(ndev->vqs[j % num].virtq_id);
1354 
1355 	MLX5_SET(rqtc, rqtc, rqt_actual_size, ndev->rqt_size);
1356 	err = mlx5_vdpa_modify_rqt(&ndev->mvdev, in, inlen, ndev->res.rqtn);
1357 	kfree(in);
1358 	if (err)
1359 		return err;
1360 
1361 	return 0;
1362 }
1363 
1364 static void destroy_rqt(struct mlx5_vdpa_net *ndev)
1365 {
1366 	mlx5_vdpa_destroy_rqt(&ndev->mvdev, ndev->res.rqtn);
1367 }
1368 
1369 static int create_tir(struct mlx5_vdpa_net *ndev)
1370 {
1371 #define HASH_IP_L4PORTS                                                                            \
1372 	(MLX5_HASH_FIELD_SEL_SRC_IP | MLX5_HASH_FIELD_SEL_DST_IP | MLX5_HASH_FIELD_SEL_L4_SPORT |  \
1373 	 MLX5_HASH_FIELD_SEL_L4_DPORT)
1374 	static const u8 rx_hash_toeplitz_key[] = { 0x2c, 0xc6, 0x81, 0xd1, 0x5b, 0xdb, 0xf4, 0xf7,
1375 						   0xfc, 0xa2, 0x83, 0x19, 0xdb, 0x1a, 0x3e, 0x94,
1376 						   0x6b, 0x9e, 0x38, 0xd9, 0x2c, 0x9c, 0x03, 0xd1,
1377 						   0xad, 0x99, 0x44, 0xa7, 0xd9, 0x56, 0x3d, 0x59,
1378 						   0x06, 0x3c, 0x25, 0xf3, 0xfc, 0x1f, 0xdc, 0x2a };
1379 	void *rss_key;
1380 	void *outer;
1381 	void *tirc;
1382 	void *in;
1383 	int err;
1384 
1385 	in = kzalloc(MLX5_ST_SZ_BYTES(create_tir_in), GFP_KERNEL);
1386 	if (!in)
1387 		return -ENOMEM;
1388 
1389 	MLX5_SET(create_tir_in, in, uid, ndev->mvdev.res.uid);
1390 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1391 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
1392 
1393 	MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1394 	MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1395 	rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1396 	memcpy(rss_key, rx_hash_toeplitz_key, sizeof(rx_hash_toeplitz_key));
1397 
1398 	outer = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1399 	MLX5_SET(rx_hash_field_select, outer, l3_prot_type, MLX5_L3_PROT_TYPE_IPV4);
1400 	MLX5_SET(rx_hash_field_select, outer, l4_prot_type, MLX5_L4_PROT_TYPE_TCP);
1401 	MLX5_SET(rx_hash_field_select, outer, selected_fields, HASH_IP_L4PORTS);
1402 
1403 	MLX5_SET(tirc, tirc, indirect_table, ndev->res.rqtn);
1404 	MLX5_SET(tirc, tirc, transport_domain, ndev->res.tdn);
1405 
1406 	err = mlx5_vdpa_create_tir(&ndev->mvdev, in, &ndev->res.tirn);
1407 	kfree(in);
1408 	return err;
1409 }
1410 
1411 static void destroy_tir(struct mlx5_vdpa_net *ndev)
1412 {
1413 	mlx5_vdpa_destroy_tir(&ndev->mvdev, ndev->res.tirn);
1414 }
1415 
1416 #define MAX_STEERING_ENT 0x8000
1417 #define MAX_STEERING_GROUPS 2
1418 
1419 static int mlx5_vdpa_add_mac_vlan_rules(struct mlx5_vdpa_net *ndev, u8 *mac,
1420 					u16 vid, bool tagged,
1421 					struct mlx5_flow_handle **ucast,
1422 					struct mlx5_flow_handle **mcast)
1423 {
1424 	struct mlx5_flow_destination dest = {};
1425 	struct mlx5_flow_act flow_act = {};
1426 	struct mlx5_flow_handle *rule;
1427 	struct mlx5_flow_spec *spec;
1428 	void *headers_c;
1429 	void *headers_v;
1430 	u8 *dmac_c;
1431 	u8 *dmac_v;
1432 	int err;
1433 
1434 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
1435 	if (!spec)
1436 		return -ENOMEM;
1437 
1438 	spec->match_criteria_enable = MLX5_MATCH_OUTER_HEADERS;
1439 	headers_c = MLX5_ADDR_OF(fte_match_param, spec->match_criteria, outer_headers);
1440 	headers_v = MLX5_ADDR_OF(fte_match_param, spec->match_value, outer_headers);
1441 	dmac_c = MLX5_ADDR_OF(fte_match_param, headers_c, outer_headers.dmac_47_16);
1442 	dmac_v = MLX5_ADDR_OF(fte_match_param, headers_v, outer_headers.dmac_47_16);
1443 	memset(dmac_c, 0xff, ETH_ALEN);
1444 	ether_addr_copy(dmac_v, mac);
1445 	MLX5_SET(fte_match_set_lyr_2_4, headers_c, cvlan_tag, 1);
1446 	if (tagged) {
1447 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, cvlan_tag, 1);
1448 		MLX5_SET_TO_ONES(fte_match_set_lyr_2_4, headers_c, first_vid);
1449 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, first_vid, vid);
1450 	}
1451 	flow_act.action = MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
1452 	dest.type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1453 	dest.tir_num = ndev->res.tirn;
1454 	rule = mlx5_add_flow_rules(ndev->rxft, spec, &flow_act, &dest, 1);
1455 	if (IS_ERR(rule))
1456 		return PTR_ERR(rule);
1457 
1458 	*ucast = rule;
1459 
1460 	memset(dmac_c, 0, ETH_ALEN);
1461 	memset(dmac_v, 0, ETH_ALEN);
1462 	dmac_c[0] = 1;
1463 	dmac_v[0] = 1;
1464 	rule = mlx5_add_flow_rules(ndev->rxft, spec, &flow_act, &dest, 1);
1465 	kvfree(spec);
1466 	if (IS_ERR(rule)) {
1467 		err = PTR_ERR(rule);
1468 		goto err_mcast;
1469 	}
1470 
1471 	*mcast = rule;
1472 	return 0;
1473 
1474 err_mcast:
1475 	mlx5_del_flow_rules(*ucast);
1476 	return err;
1477 }
1478 
1479 static void mlx5_vdpa_del_mac_vlan_rules(struct mlx5_vdpa_net *ndev,
1480 					 struct mlx5_flow_handle *ucast,
1481 					 struct mlx5_flow_handle *mcast)
1482 {
1483 	mlx5_del_flow_rules(ucast);
1484 	mlx5_del_flow_rules(mcast);
1485 }
1486 
1487 static u64 search_val(u8 *mac, u16 vlan, bool tagged)
1488 {
1489 	u64 val;
1490 
1491 	if (!tagged)
1492 		vlan = MLX5V_UNTAGGED;
1493 
1494 	val = (u64)vlan << 48 |
1495 	      (u64)mac[0] << 40 |
1496 	      (u64)mac[1] << 32 |
1497 	      (u64)mac[2] << 24 |
1498 	      (u64)mac[3] << 16 |
1499 	      (u64)mac[4] << 8 |
1500 	      (u64)mac[5];
1501 
1502 	return val;
1503 }
1504 
1505 static struct macvlan_node *mac_vlan_lookup(struct mlx5_vdpa_net *ndev, u64 value)
1506 {
1507 	struct macvlan_node *pos;
1508 	u32 idx;
1509 
1510 	idx = hash_64(value, 8); // tbd 8
1511 	hlist_for_each_entry(pos, &ndev->macvlan_hash[idx], hlist) {
1512 		if (pos->macvlan == value)
1513 			return pos;
1514 	}
1515 	return NULL;
1516 }
1517 
1518 static int mac_vlan_add(struct mlx5_vdpa_net *ndev, u8 *mac, u16 vlan, bool tagged) // vlan -> vid
1519 {
1520 	struct macvlan_node *ptr;
1521 	u64 val;
1522 	u32 idx;
1523 	int err;
1524 
1525 	val = search_val(mac, vlan, tagged);
1526 	if (mac_vlan_lookup(ndev, val))
1527 		return -EEXIST;
1528 
1529 	ptr = kzalloc(sizeof(*ptr), GFP_KERNEL);
1530 	if (!ptr)
1531 		return -ENOMEM;
1532 
1533 	err = mlx5_vdpa_add_mac_vlan_rules(ndev, ndev->config.mac, vlan, tagged,
1534 					   &ptr->ucast_rule, &ptr->mcast_rule);
1535 	if (err)
1536 		goto err_add;
1537 
1538 	ptr->macvlan = val;
1539 	idx = hash_64(val, 8);
1540 	hlist_add_head(&ptr->hlist, &ndev->macvlan_hash[idx]);
1541 	return 0;
1542 
1543 err_add:
1544 	kfree(ptr);
1545 	return err;
1546 }
1547 
1548 static void mac_vlan_del(struct mlx5_vdpa_net *ndev, u8 *mac, u16 vlan, bool tagged)
1549 {
1550 	struct macvlan_node *ptr;
1551 
1552 	ptr = mac_vlan_lookup(ndev, search_val(mac, vlan, tagged));
1553 	if (!ptr)
1554 		return;
1555 
1556 	hlist_del(&ptr->hlist);
1557 	mlx5_vdpa_del_mac_vlan_rules(ndev, ptr->ucast_rule, ptr->mcast_rule);
1558 	kfree(ptr);
1559 }
1560 
1561 static void clear_mac_vlan_table(struct mlx5_vdpa_net *ndev)
1562 {
1563 	struct macvlan_node *pos;
1564 	struct hlist_node *n;
1565 	int i;
1566 
1567 	for (i = 0; i < MLX5V_MACVLAN_SIZE; i++) {
1568 		hlist_for_each_entry_safe(pos, n, &ndev->macvlan_hash[i], hlist) {
1569 			hlist_del(&pos->hlist);
1570 			mlx5_vdpa_del_mac_vlan_rules(ndev, pos->ucast_rule, pos->mcast_rule);
1571 			kfree(pos);
1572 		}
1573 	}
1574 }
1575 
1576 static int setup_steering(struct mlx5_vdpa_net *ndev)
1577 {
1578 	struct mlx5_flow_table_attr ft_attr = {};
1579 	struct mlx5_flow_namespace *ns;
1580 	int err;
1581 
1582 	ft_attr.max_fte = MAX_STEERING_ENT;
1583 	ft_attr.autogroup.max_num_groups = MAX_STEERING_GROUPS;
1584 
1585 	ns = mlx5_get_flow_namespace(ndev->mvdev.mdev, MLX5_FLOW_NAMESPACE_BYPASS);
1586 	if (!ns) {
1587 		mlx5_vdpa_warn(&ndev->mvdev, "failed to get flow namespace\n");
1588 		return -EOPNOTSUPP;
1589 	}
1590 
1591 	ndev->rxft = mlx5_create_auto_grouped_flow_table(ns, &ft_attr);
1592 	if (IS_ERR(ndev->rxft)) {
1593 		mlx5_vdpa_warn(&ndev->mvdev, "failed to create flow table\n");
1594 		return PTR_ERR(ndev->rxft);
1595 	}
1596 
1597 	err = mac_vlan_add(ndev, ndev->config.mac, 0, false);
1598 	if (err)
1599 		goto err_add;
1600 
1601 	return 0;
1602 
1603 err_add:
1604 	mlx5_destroy_flow_table(ndev->rxft);
1605 	return err;
1606 }
1607 
1608 static void teardown_steering(struct mlx5_vdpa_net *ndev)
1609 {
1610 	clear_mac_vlan_table(ndev);
1611 	mlx5_destroy_flow_table(ndev->rxft);
1612 }
1613 
1614 static virtio_net_ctrl_ack handle_ctrl_mac(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1615 {
1616 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1617 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1618 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1619 	struct mlx5_core_dev *pfmdev;
1620 	size_t read;
1621 	u8 mac[ETH_ALEN], mac_back[ETH_ALEN];
1622 
1623 	pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
1624 	switch (cmd) {
1625 	case VIRTIO_NET_CTRL_MAC_ADDR_SET:
1626 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, (void *)mac, ETH_ALEN);
1627 		if (read != ETH_ALEN)
1628 			break;
1629 
1630 		if (!memcmp(ndev->config.mac, mac, 6)) {
1631 			status = VIRTIO_NET_OK;
1632 			break;
1633 		}
1634 
1635 		if (is_zero_ether_addr(mac))
1636 			break;
1637 
1638 		if (!is_zero_ether_addr(ndev->config.mac)) {
1639 			if (mlx5_mpfs_del_mac(pfmdev, ndev->config.mac)) {
1640 				mlx5_vdpa_warn(mvdev, "failed to delete old MAC %pM from MPFS table\n",
1641 					       ndev->config.mac);
1642 				break;
1643 			}
1644 		}
1645 
1646 		if (mlx5_mpfs_add_mac(pfmdev, mac)) {
1647 			mlx5_vdpa_warn(mvdev, "failed to insert new MAC %pM into MPFS table\n",
1648 				       mac);
1649 			break;
1650 		}
1651 
1652 		/* backup the original mac address so that if failed to add the forward rules
1653 		 * we could restore it
1654 		 */
1655 		memcpy(mac_back, ndev->config.mac, ETH_ALEN);
1656 
1657 		memcpy(ndev->config.mac, mac, ETH_ALEN);
1658 
1659 		/* Need recreate the flow table entry, so that the packet could forward back
1660 		 */
1661 		mac_vlan_del(ndev, ndev->config.mac, 0, false);
1662 
1663 		if (mac_vlan_add(ndev, ndev->config.mac, 0, false)) {
1664 			mlx5_vdpa_warn(mvdev, "failed to insert forward rules, try to restore\n");
1665 
1666 			/* Although it hardly run here, we still need double check */
1667 			if (is_zero_ether_addr(mac_back)) {
1668 				mlx5_vdpa_warn(mvdev, "restore mac failed: Original MAC is zero\n");
1669 				break;
1670 			}
1671 
1672 			/* Try to restore original mac address to MFPS table, and try to restore
1673 			 * the forward rule entry.
1674 			 */
1675 			if (mlx5_mpfs_del_mac(pfmdev, ndev->config.mac)) {
1676 				mlx5_vdpa_warn(mvdev, "restore mac failed: delete MAC %pM from MPFS table failed\n",
1677 					       ndev->config.mac);
1678 			}
1679 
1680 			if (mlx5_mpfs_add_mac(pfmdev, mac_back)) {
1681 				mlx5_vdpa_warn(mvdev, "restore mac failed: insert old MAC %pM into MPFS table failed\n",
1682 					       mac_back);
1683 			}
1684 
1685 			memcpy(ndev->config.mac, mac_back, ETH_ALEN);
1686 
1687 			if (mac_vlan_add(ndev, ndev->config.mac, 0, false))
1688 				mlx5_vdpa_warn(mvdev, "restore forward rules failed: insert forward rules failed\n");
1689 
1690 			break;
1691 		}
1692 
1693 		status = VIRTIO_NET_OK;
1694 		break;
1695 
1696 	default:
1697 		break;
1698 	}
1699 
1700 	return status;
1701 }
1702 
1703 static int change_num_qps(struct mlx5_vdpa_dev *mvdev, int newqps)
1704 {
1705 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1706 	int cur_qps = ndev->cur_num_vqs / 2;
1707 	int err;
1708 	int i;
1709 
1710 	if (cur_qps > newqps) {
1711 		err = modify_rqt(ndev, 2 * newqps);
1712 		if (err)
1713 			return err;
1714 
1715 		for (i = ndev->cur_num_vqs - 1; i >= 2 * newqps; i--)
1716 			teardown_vq(ndev, &ndev->vqs[i]);
1717 
1718 		ndev->cur_num_vqs = 2 * newqps;
1719 	} else {
1720 		ndev->cur_num_vqs = 2 * newqps;
1721 		for (i = cur_qps * 2; i < 2 * newqps; i++) {
1722 			err = setup_vq(ndev, &ndev->vqs[i]);
1723 			if (err)
1724 				goto clean_added;
1725 		}
1726 		err = modify_rqt(ndev, 2 * newqps);
1727 		if (err)
1728 			goto clean_added;
1729 	}
1730 	return 0;
1731 
1732 clean_added:
1733 	for (--i; i >= 2 * cur_qps; --i)
1734 		teardown_vq(ndev, &ndev->vqs[i]);
1735 
1736 	ndev->cur_num_vqs = 2 * cur_qps;
1737 
1738 	return err;
1739 }
1740 
1741 static virtio_net_ctrl_ack handle_ctrl_mq(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1742 {
1743 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1744 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1745 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1746 	struct virtio_net_ctrl_mq mq;
1747 	size_t read;
1748 	u16 newqps;
1749 
1750 	switch (cmd) {
1751 	case VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET:
1752 		/* This mq feature check aligns with pre-existing userspace
1753 		 * implementation.
1754 		 *
1755 		 * Without it, an untrusted driver could fake a multiqueue config
1756 		 * request down to a non-mq device that may cause kernel to
1757 		 * panic due to uninitialized resources for extra vqs. Even with
1758 		 * a well behaving guest driver, it is not expected to allow
1759 		 * changing the number of vqs on a non-mq device.
1760 		 */
1761 		if (!MLX5_FEATURE(mvdev, VIRTIO_NET_F_MQ))
1762 			break;
1763 
1764 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, (void *)&mq, sizeof(mq));
1765 		if (read != sizeof(mq))
1766 			break;
1767 
1768 		newqps = mlx5vdpa16_to_cpu(mvdev, mq.virtqueue_pairs);
1769 		if (newqps < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1770 		    newqps > ndev->rqt_size)
1771 			break;
1772 
1773 		if (ndev->cur_num_vqs == 2 * newqps) {
1774 			status = VIRTIO_NET_OK;
1775 			break;
1776 		}
1777 
1778 		if (!change_num_qps(mvdev, newqps))
1779 			status = VIRTIO_NET_OK;
1780 
1781 		break;
1782 	default:
1783 		break;
1784 	}
1785 
1786 	return status;
1787 }
1788 
1789 static virtio_net_ctrl_ack handle_ctrl_vlan(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1790 {
1791 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1792 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1793 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1794 	__virtio16 vlan;
1795 	size_t read;
1796 	u16 id;
1797 
1798 	switch (cmd) {
1799 	case VIRTIO_NET_CTRL_VLAN_ADD:
1800 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &vlan, sizeof(vlan));
1801 		if (read != sizeof(vlan))
1802 			break;
1803 
1804 		id = mlx5vdpa16_to_cpu(mvdev, vlan);
1805 		if (mac_vlan_add(ndev, ndev->config.mac, id, true))
1806 			break;
1807 
1808 		status = VIRTIO_NET_OK;
1809 		break;
1810 	case VIRTIO_NET_CTRL_VLAN_DEL:
1811 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &vlan, sizeof(vlan));
1812 		if (read != sizeof(vlan))
1813 			break;
1814 
1815 		id = mlx5vdpa16_to_cpu(mvdev, vlan);
1816 		mac_vlan_del(ndev, ndev->config.mac, id, true);
1817 		break;
1818 	default:
1819 	break;
1820 }
1821 
1822 return status;
1823 }
1824 
1825 static void mlx5_cvq_kick_handler(struct work_struct *work)
1826 {
1827 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1828 	struct virtio_net_ctrl_hdr ctrl;
1829 	struct mlx5_vdpa_wq_ent *wqent;
1830 	struct mlx5_vdpa_dev *mvdev;
1831 	struct mlx5_control_vq *cvq;
1832 	struct mlx5_vdpa_net *ndev;
1833 	size_t read, write;
1834 	int err;
1835 
1836 	wqent = container_of(work, struct mlx5_vdpa_wq_ent, work);
1837 	mvdev = wqent->mvdev;
1838 	ndev = to_mlx5_vdpa_ndev(mvdev);
1839 	cvq = &mvdev->cvq;
1840 
1841 	down_write(&ndev->reslock);
1842 
1843 	if (!(mvdev->status & VIRTIO_CONFIG_S_DRIVER_OK))
1844 		goto out;
1845 
1846 	if (!(ndev->mvdev.actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
1847 		goto out;
1848 
1849 	if (!cvq->ready)
1850 		goto out;
1851 
1852 	while (true) {
1853 		err = vringh_getdesc_iotlb(&cvq->vring, &cvq->riov, &cvq->wiov, &cvq->head,
1854 					   GFP_ATOMIC);
1855 		if (err <= 0)
1856 			break;
1857 
1858 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &ctrl, sizeof(ctrl));
1859 		if (read != sizeof(ctrl))
1860 			break;
1861 
1862 		cvq->received_desc++;
1863 		switch (ctrl.class) {
1864 		case VIRTIO_NET_CTRL_MAC:
1865 			status = handle_ctrl_mac(mvdev, ctrl.cmd);
1866 			break;
1867 		case VIRTIO_NET_CTRL_MQ:
1868 			status = handle_ctrl_mq(mvdev, ctrl.cmd);
1869 			break;
1870 		case VIRTIO_NET_CTRL_VLAN:
1871 			status = handle_ctrl_vlan(mvdev, ctrl.cmd);
1872 			break;
1873 		default:
1874 			break;
1875 		}
1876 
1877 		/* Make sure data is written before advancing index */
1878 		smp_wmb();
1879 
1880 		write = vringh_iov_push_iotlb(&cvq->vring, &cvq->wiov, &status, sizeof(status));
1881 		vringh_complete_iotlb(&cvq->vring, cvq->head, write);
1882 		vringh_kiov_cleanup(&cvq->riov);
1883 		vringh_kiov_cleanup(&cvq->wiov);
1884 
1885 		if (vringh_need_notify_iotlb(&cvq->vring))
1886 			vringh_notify(&cvq->vring);
1887 
1888 		cvq->completed_desc++;
1889 		queue_work(mvdev->wq, &wqent->work);
1890 		break;
1891 	}
1892 
1893 out:
1894 	up_write(&ndev->reslock);
1895 }
1896 
1897 static void mlx5_vdpa_kick_vq(struct vdpa_device *vdev, u16 idx)
1898 {
1899 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1900 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1901 	struct mlx5_vdpa_virtqueue *mvq;
1902 
1903 	if (!is_index_valid(mvdev, idx))
1904 		return;
1905 
1906 	if (unlikely(is_ctrl_vq_idx(mvdev, idx))) {
1907 		if (!mvdev->wq || !mvdev->cvq.ready)
1908 			return;
1909 
1910 		queue_work(mvdev->wq, &ndev->cvq_ent.work);
1911 		return;
1912 	}
1913 
1914 	mvq = &ndev->vqs[idx];
1915 	if (unlikely(!mvq->ready))
1916 		return;
1917 
1918 	iowrite16(idx, ndev->mvdev.res.kick_addr);
1919 }
1920 
1921 static int mlx5_vdpa_set_vq_address(struct vdpa_device *vdev, u16 idx, u64 desc_area,
1922 				    u64 driver_area, u64 device_area)
1923 {
1924 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1925 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1926 	struct mlx5_vdpa_virtqueue *mvq;
1927 
1928 	if (!is_index_valid(mvdev, idx))
1929 		return -EINVAL;
1930 
1931 	if (is_ctrl_vq_idx(mvdev, idx)) {
1932 		mvdev->cvq.desc_addr = desc_area;
1933 		mvdev->cvq.device_addr = device_area;
1934 		mvdev->cvq.driver_addr = driver_area;
1935 		return 0;
1936 	}
1937 
1938 	mvq = &ndev->vqs[idx];
1939 	mvq->desc_addr = desc_area;
1940 	mvq->device_addr = device_area;
1941 	mvq->driver_addr = driver_area;
1942 	return 0;
1943 }
1944 
1945 static void mlx5_vdpa_set_vq_num(struct vdpa_device *vdev, u16 idx, u32 num)
1946 {
1947 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1948 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1949 	struct mlx5_vdpa_virtqueue *mvq;
1950 
1951 	if (!is_index_valid(mvdev, idx) || is_ctrl_vq_idx(mvdev, idx))
1952 		return;
1953 
1954 	mvq = &ndev->vqs[idx];
1955 	mvq->num_ent = num;
1956 }
1957 
1958 static void mlx5_vdpa_set_vq_cb(struct vdpa_device *vdev, u16 idx, struct vdpa_callback *cb)
1959 {
1960 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1961 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1962 
1963 	ndev->event_cbs[idx] = *cb;
1964 }
1965 
1966 static void mlx5_cvq_notify(struct vringh *vring)
1967 {
1968 	struct mlx5_control_vq *cvq = container_of(vring, struct mlx5_control_vq, vring);
1969 
1970 	if (!cvq->event_cb.callback)
1971 		return;
1972 
1973 	cvq->event_cb.callback(cvq->event_cb.private);
1974 }
1975 
1976 static void set_cvq_ready(struct mlx5_vdpa_dev *mvdev, bool ready)
1977 {
1978 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1979 
1980 	cvq->ready = ready;
1981 	if (!ready)
1982 		return;
1983 
1984 	cvq->vring.notify = mlx5_cvq_notify;
1985 }
1986 
1987 static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready)
1988 {
1989 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1990 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1991 	struct mlx5_vdpa_virtqueue *mvq;
1992 
1993 	if (!mvdev->actual_features)
1994 		return;
1995 
1996 	if (!is_index_valid(mvdev, idx))
1997 		return;
1998 
1999 	if (is_ctrl_vq_idx(mvdev, idx)) {
2000 		set_cvq_ready(mvdev, ready);
2001 		return;
2002 	}
2003 
2004 	mvq = &ndev->vqs[idx];
2005 	if (!ready)
2006 		suspend_vq(ndev, mvq);
2007 
2008 	mvq->ready = ready;
2009 }
2010 
2011 static bool mlx5_vdpa_get_vq_ready(struct vdpa_device *vdev, u16 idx)
2012 {
2013 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2014 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2015 
2016 	if (!is_index_valid(mvdev, idx))
2017 		return false;
2018 
2019 	if (is_ctrl_vq_idx(mvdev, idx))
2020 		return mvdev->cvq.ready;
2021 
2022 	return ndev->vqs[idx].ready;
2023 }
2024 
2025 static int mlx5_vdpa_set_vq_state(struct vdpa_device *vdev, u16 idx,
2026 				  const struct vdpa_vq_state *state)
2027 {
2028 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2029 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2030 	struct mlx5_vdpa_virtqueue *mvq;
2031 
2032 	if (!is_index_valid(mvdev, idx))
2033 		return -EINVAL;
2034 
2035 	if (is_ctrl_vq_idx(mvdev, idx)) {
2036 		mvdev->cvq.vring.last_avail_idx = state->split.avail_index;
2037 		return 0;
2038 	}
2039 
2040 	mvq = &ndev->vqs[idx];
2041 	if (mvq->fw_state == MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY) {
2042 		mlx5_vdpa_warn(mvdev, "can't modify available index\n");
2043 		return -EINVAL;
2044 	}
2045 
2046 	mvq->used_idx = state->split.avail_index;
2047 	mvq->avail_idx = state->split.avail_index;
2048 	return 0;
2049 }
2050 
2051 static int mlx5_vdpa_get_vq_state(struct vdpa_device *vdev, u16 idx, struct vdpa_vq_state *state)
2052 {
2053 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2054 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2055 	struct mlx5_vdpa_virtqueue *mvq;
2056 	struct mlx5_virtq_attr attr;
2057 	int err;
2058 
2059 	if (!is_index_valid(mvdev, idx))
2060 		return -EINVAL;
2061 
2062 	if (is_ctrl_vq_idx(mvdev, idx)) {
2063 		state->split.avail_index = mvdev->cvq.vring.last_avail_idx;
2064 		return 0;
2065 	}
2066 
2067 	mvq = &ndev->vqs[idx];
2068 	/* If the virtq object was destroyed, use the value saved at
2069 	 * the last minute of suspend_vq. This caters for userspace
2070 	 * that cares about emulating the index after vq is stopped.
2071 	 */
2072 	if (!mvq->initialized) {
2073 		/* Firmware returns a wrong value for the available index.
2074 		 * Since both values should be identical, we take the value of
2075 		 * used_idx which is reported correctly.
2076 		 */
2077 		state->split.avail_index = mvq->used_idx;
2078 		return 0;
2079 	}
2080 
2081 	err = query_virtqueue(ndev, mvq, &attr);
2082 	if (err) {
2083 		mlx5_vdpa_warn(mvdev, "failed to query virtqueue\n");
2084 		return err;
2085 	}
2086 	state->split.avail_index = attr.used_index;
2087 	return 0;
2088 }
2089 
2090 static u32 mlx5_vdpa_get_vq_align(struct vdpa_device *vdev)
2091 {
2092 	return PAGE_SIZE;
2093 }
2094 
2095 static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdpa, u16 idx)
2096 {
2097 	return 0;
2098 }
2099 
2100 enum { MLX5_VIRTIO_NET_F_GUEST_CSUM = 1 << 9,
2101 	MLX5_VIRTIO_NET_F_CSUM = 1 << 10,
2102 	MLX5_VIRTIO_NET_F_HOST_TSO6 = 1 << 11,
2103 	MLX5_VIRTIO_NET_F_HOST_TSO4 = 1 << 12,
2104 };
2105 
2106 static u64 mlx_to_vritio_features(u16 dev_features)
2107 {
2108 	u64 result = 0;
2109 
2110 	if (dev_features & MLX5_VIRTIO_NET_F_GUEST_CSUM)
2111 		result |= BIT_ULL(VIRTIO_NET_F_GUEST_CSUM);
2112 	if (dev_features & MLX5_VIRTIO_NET_F_CSUM)
2113 		result |= BIT_ULL(VIRTIO_NET_F_CSUM);
2114 	if (dev_features & MLX5_VIRTIO_NET_F_HOST_TSO6)
2115 		result |= BIT_ULL(VIRTIO_NET_F_HOST_TSO6);
2116 	if (dev_features & MLX5_VIRTIO_NET_F_HOST_TSO4)
2117 		result |= BIT_ULL(VIRTIO_NET_F_HOST_TSO4);
2118 
2119 	return result;
2120 }
2121 
2122 static u64 get_supported_features(struct mlx5_core_dev *mdev)
2123 {
2124 	u64 mlx_vdpa_features = 0;
2125 	u16 dev_features;
2126 
2127 	dev_features = MLX5_CAP_DEV_VDPA_EMULATION(mdev, device_features_bits_mask);
2128 	mlx_vdpa_features |= mlx_to_vritio_features(dev_features);
2129 	if (MLX5_CAP_DEV_VDPA_EMULATION(mdev, virtio_version_1_0))
2130 		mlx_vdpa_features |= BIT_ULL(VIRTIO_F_VERSION_1);
2131 	mlx_vdpa_features |= BIT_ULL(VIRTIO_F_ACCESS_PLATFORM);
2132 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_VQ);
2133 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR);
2134 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MQ);
2135 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_STATUS);
2136 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MTU);
2137 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_VLAN);
2138 
2139 	return mlx_vdpa_features;
2140 }
2141 
2142 static u64 mlx5_vdpa_get_device_features(struct vdpa_device *vdev)
2143 {
2144 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2145 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2146 
2147 	print_features(mvdev, ndev->mvdev.mlx_features, false);
2148 	return ndev->mvdev.mlx_features;
2149 }
2150 
2151 static int verify_driver_features(struct mlx5_vdpa_dev *mvdev, u64 features)
2152 {
2153 	/* Minimum features to expect */
2154 	if (!(features & BIT_ULL(VIRTIO_F_ACCESS_PLATFORM)))
2155 		return -EOPNOTSUPP;
2156 
2157 	/* Double check features combination sent down by the driver.
2158 	 * Fail invalid features due to absence of the depended feature.
2159 	 *
2160 	 * Per VIRTIO v1.1 specification, section 5.1.3.1 Feature bit
2161 	 * requirements: "VIRTIO_NET_F_MQ Requires VIRTIO_NET_F_CTRL_VQ".
2162 	 * By failing the invalid features sent down by untrusted drivers,
2163 	 * we're assured the assumption made upon is_index_valid() and
2164 	 * is_ctrl_vq_idx() will not be compromised.
2165 	 */
2166 	if ((features & (BIT_ULL(VIRTIO_NET_F_MQ) | BIT_ULL(VIRTIO_NET_F_CTRL_VQ))) ==
2167             BIT_ULL(VIRTIO_NET_F_MQ))
2168 		return -EINVAL;
2169 
2170 	return 0;
2171 }
2172 
2173 static int setup_virtqueues(struct mlx5_vdpa_dev *mvdev)
2174 {
2175 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2176 	struct mlx5_control_vq *cvq = &mvdev->cvq;
2177 	int err;
2178 	int i;
2179 
2180 	for (i = 0; i < mvdev->max_vqs; i++) {
2181 		err = setup_vq(ndev, &ndev->vqs[i]);
2182 		if (err)
2183 			goto err_vq;
2184 	}
2185 
2186 	if (mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)) {
2187 		err = vringh_init_iotlb(&cvq->vring, mvdev->actual_features,
2188 					MLX5_CVQ_MAX_ENT, false,
2189 					(struct vring_desc *)(uintptr_t)cvq->desc_addr,
2190 					(struct vring_avail *)(uintptr_t)cvq->driver_addr,
2191 					(struct vring_used *)(uintptr_t)cvq->device_addr);
2192 		if (err)
2193 			goto err_vq;
2194 	}
2195 
2196 	return 0;
2197 
2198 err_vq:
2199 	for (--i; i >= 0; i--)
2200 		teardown_vq(ndev, &ndev->vqs[i]);
2201 
2202 	return err;
2203 }
2204 
2205 static void teardown_virtqueues(struct mlx5_vdpa_net *ndev)
2206 {
2207 	struct mlx5_vdpa_virtqueue *mvq;
2208 	int i;
2209 
2210 	for (i = ndev->mvdev.max_vqs - 1; i >= 0; i--) {
2211 		mvq = &ndev->vqs[i];
2212 		if (!mvq->initialized)
2213 			continue;
2214 
2215 		teardown_vq(ndev, mvq);
2216 	}
2217 }
2218 
2219 static void update_cvq_info(struct mlx5_vdpa_dev *mvdev)
2220 {
2221 	if (MLX5_FEATURE(mvdev, VIRTIO_NET_F_CTRL_VQ)) {
2222 		if (MLX5_FEATURE(mvdev, VIRTIO_NET_F_MQ)) {
2223 			/* MQ supported. CVQ index is right above the last data virtqueue's */
2224 			mvdev->max_idx = mvdev->max_vqs;
2225 		} else {
2226 			/* Only CVQ supportted. data virtqueues occupy indices 0 and 1.
2227 			 * CVQ gets index 2
2228 			 */
2229 			mvdev->max_idx = 2;
2230 		}
2231 	} else {
2232 		/* Two data virtqueues only: one for rx and one for tx */
2233 		mvdev->max_idx = 1;
2234 	}
2235 }
2236 
2237 static int mlx5_vdpa_set_driver_features(struct vdpa_device *vdev, u64 features)
2238 {
2239 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2240 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2241 	int err;
2242 
2243 	print_features(mvdev, features, true);
2244 
2245 	err = verify_driver_features(mvdev, features);
2246 	if (err)
2247 		return err;
2248 
2249 	ndev->mvdev.actual_features = features & ndev->mvdev.mlx_features;
2250 	if (ndev->mvdev.actual_features & BIT_ULL(VIRTIO_NET_F_MQ))
2251 		ndev->rqt_size = mlx5vdpa16_to_cpu(mvdev, ndev->config.max_virtqueue_pairs);
2252 	else
2253 		ndev->rqt_size = 1;
2254 
2255 	ndev->cur_num_vqs = 2 * ndev->rqt_size;
2256 
2257 	update_cvq_info(mvdev);
2258 	return err;
2259 }
2260 
2261 static void mlx5_vdpa_set_config_cb(struct vdpa_device *vdev, struct vdpa_callback *cb)
2262 {
2263 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2264 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2265 
2266 	ndev->config_cb = *cb;
2267 }
2268 
2269 #define MLX5_VDPA_MAX_VQ_ENTRIES 256
2270 static u16 mlx5_vdpa_get_vq_num_max(struct vdpa_device *vdev)
2271 {
2272 	return MLX5_VDPA_MAX_VQ_ENTRIES;
2273 }
2274 
2275 static u32 mlx5_vdpa_get_device_id(struct vdpa_device *vdev)
2276 {
2277 	return VIRTIO_ID_NET;
2278 }
2279 
2280 static u32 mlx5_vdpa_get_vendor_id(struct vdpa_device *vdev)
2281 {
2282 	return PCI_VENDOR_ID_MELLANOX;
2283 }
2284 
2285 static u8 mlx5_vdpa_get_status(struct vdpa_device *vdev)
2286 {
2287 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2288 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2289 
2290 	print_status(mvdev, ndev->mvdev.status, false);
2291 	return ndev->mvdev.status;
2292 }
2293 
2294 static int save_channel_info(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
2295 {
2296 	struct mlx5_vq_restore_info *ri = &mvq->ri;
2297 	struct mlx5_virtq_attr attr = {};
2298 	int err;
2299 
2300 	if (mvq->initialized) {
2301 		err = query_virtqueue(ndev, mvq, &attr);
2302 		if (err)
2303 			return err;
2304 	}
2305 
2306 	ri->avail_index = attr.available_index;
2307 	ri->used_index = attr.used_index;
2308 	ri->ready = mvq->ready;
2309 	ri->num_ent = mvq->num_ent;
2310 	ri->desc_addr = mvq->desc_addr;
2311 	ri->device_addr = mvq->device_addr;
2312 	ri->driver_addr = mvq->driver_addr;
2313 	ri->restore = true;
2314 	return 0;
2315 }
2316 
2317 static int save_channels_info(struct mlx5_vdpa_net *ndev)
2318 {
2319 	int i;
2320 
2321 	for (i = 0; i < ndev->mvdev.max_vqs; i++) {
2322 		memset(&ndev->vqs[i].ri, 0, sizeof(ndev->vqs[i].ri));
2323 		save_channel_info(ndev, &ndev->vqs[i]);
2324 	}
2325 	return 0;
2326 }
2327 
2328 static void mlx5_clear_vqs(struct mlx5_vdpa_net *ndev)
2329 {
2330 	int i;
2331 
2332 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
2333 		memset(&ndev->vqs[i], 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2334 }
2335 
2336 static void restore_channels_info(struct mlx5_vdpa_net *ndev)
2337 {
2338 	struct mlx5_vdpa_virtqueue *mvq;
2339 	struct mlx5_vq_restore_info *ri;
2340 	int i;
2341 
2342 	mlx5_clear_vqs(ndev);
2343 	init_mvqs(ndev);
2344 	for (i = 0; i < ndev->mvdev.max_vqs; i++) {
2345 		mvq = &ndev->vqs[i];
2346 		ri = &mvq->ri;
2347 		if (!ri->restore)
2348 			continue;
2349 
2350 		mvq->avail_idx = ri->avail_index;
2351 		mvq->used_idx = ri->used_index;
2352 		mvq->ready = ri->ready;
2353 		mvq->num_ent = ri->num_ent;
2354 		mvq->desc_addr = ri->desc_addr;
2355 		mvq->device_addr = ri->device_addr;
2356 		mvq->driver_addr = ri->driver_addr;
2357 	}
2358 }
2359 
2360 static int mlx5_vdpa_change_map(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
2361 {
2362 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2363 	int err;
2364 
2365 	suspend_vqs(ndev);
2366 	err = save_channels_info(ndev);
2367 	if (err)
2368 		goto err_mr;
2369 
2370 	teardown_driver(ndev);
2371 	mlx5_vdpa_destroy_mr(mvdev);
2372 	err = mlx5_vdpa_create_mr(mvdev, iotlb);
2373 	if (err)
2374 		goto err_mr;
2375 
2376 	if (!(mvdev->status & VIRTIO_CONFIG_S_DRIVER_OK))
2377 		goto err_mr;
2378 
2379 	restore_channels_info(ndev);
2380 	err = setup_driver(mvdev);
2381 	if (err)
2382 		goto err_setup;
2383 
2384 	return 0;
2385 
2386 err_setup:
2387 	mlx5_vdpa_destroy_mr(mvdev);
2388 err_mr:
2389 	return err;
2390 }
2391 
2392 /* reslock must be held for this function */
2393 static int setup_driver(struct mlx5_vdpa_dev *mvdev)
2394 {
2395 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2396 	int err;
2397 
2398 	WARN_ON(!rwsem_is_locked(&ndev->reslock));
2399 
2400 	if (ndev->setup) {
2401 		mlx5_vdpa_warn(mvdev, "setup driver called for already setup driver\n");
2402 		err = 0;
2403 		goto out;
2404 	}
2405 	err = setup_virtqueues(mvdev);
2406 	if (err) {
2407 		mlx5_vdpa_warn(mvdev, "setup_virtqueues\n");
2408 		goto out;
2409 	}
2410 
2411 	err = create_rqt(ndev);
2412 	if (err) {
2413 		mlx5_vdpa_warn(mvdev, "create_rqt\n");
2414 		goto err_rqt;
2415 	}
2416 
2417 	err = create_tir(ndev);
2418 	if (err) {
2419 		mlx5_vdpa_warn(mvdev, "create_tir\n");
2420 		goto err_tir;
2421 	}
2422 
2423 	err = setup_steering(ndev);
2424 	if (err) {
2425 		mlx5_vdpa_warn(mvdev, "setup_steering\n");
2426 		goto err_fwd;
2427 	}
2428 	ndev->setup = true;
2429 
2430 	return 0;
2431 
2432 err_fwd:
2433 	destroy_tir(ndev);
2434 err_tir:
2435 	destroy_rqt(ndev);
2436 err_rqt:
2437 	teardown_virtqueues(ndev);
2438 out:
2439 	return err;
2440 }
2441 
2442 /* reslock must be held for this function */
2443 static void teardown_driver(struct mlx5_vdpa_net *ndev)
2444 {
2445 
2446 	WARN_ON(!rwsem_is_locked(&ndev->reslock));
2447 
2448 	if (!ndev->setup)
2449 		return;
2450 
2451 	teardown_steering(ndev);
2452 	destroy_tir(ndev);
2453 	destroy_rqt(ndev);
2454 	teardown_virtqueues(ndev);
2455 	ndev->setup = false;
2456 }
2457 
2458 static void clear_vqs_ready(struct mlx5_vdpa_net *ndev)
2459 {
2460 	int i;
2461 
2462 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
2463 		ndev->vqs[i].ready = false;
2464 
2465 	ndev->mvdev.cvq.ready = false;
2466 }
2467 
2468 static void mlx5_vdpa_set_status(struct vdpa_device *vdev, u8 status)
2469 {
2470 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2471 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2472 	int err;
2473 
2474 	print_status(mvdev, status, true);
2475 
2476 	down_write(&ndev->reslock);
2477 
2478 	if ((status ^ ndev->mvdev.status) & VIRTIO_CONFIG_S_DRIVER_OK) {
2479 		if (status & VIRTIO_CONFIG_S_DRIVER_OK) {
2480 			err = setup_driver(mvdev);
2481 			if (err) {
2482 				mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
2483 				goto err_setup;
2484 			}
2485 		} else {
2486 			mlx5_vdpa_warn(mvdev, "did not expect DRIVER_OK to be cleared\n");
2487 			goto err_clear;
2488 		}
2489 	}
2490 
2491 	ndev->mvdev.status = status;
2492 	up_write(&ndev->reslock);
2493 	return;
2494 
2495 err_setup:
2496 	mlx5_vdpa_destroy_mr(&ndev->mvdev);
2497 	ndev->mvdev.status |= VIRTIO_CONFIG_S_FAILED;
2498 err_clear:
2499 	up_write(&ndev->reslock);
2500 }
2501 
2502 static int mlx5_vdpa_reset(struct vdpa_device *vdev)
2503 {
2504 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2505 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2506 
2507 	print_status(mvdev, 0, true);
2508 	mlx5_vdpa_info(mvdev, "performing device reset\n");
2509 
2510 	down_write(&ndev->reslock);
2511 	teardown_driver(ndev);
2512 	clear_vqs_ready(ndev);
2513 	mlx5_vdpa_destroy_mr(&ndev->mvdev);
2514 	ndev->mvdev.status = 0;
2515 	ndev->cur_num_vqs = 0;
2516 	ndev->mvdev.cvq.received_desc = 0;
2517 	ndev->mvdev.cvq.completed_desc = 0;
2518 	memset(ndev->event_cbs, 0, sizeof(*ndev->event_cbs) * (mvdev->max_vqs + 1));
2519 	ndev->mvdev.actual_features = 0;
2520 	++mvdev->generation;
2521 	if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
2522 		if (mlx5_vdpa_create_mr(mvdev, NULL))
2523 			mlx5_vdpa_warn(mvdev, "create MR failed\n");
2524 	}
2525 	up_write(&ndev->reslock);
2526 
2527 	return 0;
2528 }
2529 
2530 static size_t mlx5_vdpa_get_config_size(struct vdpa_device *vdev)
2531 {
2532 	return sizeof(struct virtio_net_config);
2533 }
2534 
2535 static void mlx5_vdpa_get_config(struct vdpa_device *vdev, unsigned int offset, void *buf,
2536 				 unsigned int len)
2537 {
2538 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2539 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2540 
2541 	if (offset + len <= sizeof(struct virtio_net_config))
2542 		memcpy(buf, (u8 *)&ndev->config + offset, len);
2543 }
2544 
2545 static void mlx5_vdpa_set_config(struct vdpa_device *vdev, unsigned int offset, const void *buf,
2546 				 unsigned int len)
2547 {
2548 	/* not supported */
2549 }
2550 
2551 static u32 mlx5_vdpa_get_generation(struct vdpa_device *vdev)
2552 {
2553 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2554 
2555 	return mvdev->generation;
2556 }
2557 
2558 static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
2559 			     struct vhost_iotlb *iotlb)
2560 {
2561 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2562 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2563 	bool change_map;
2564 	int err;
2565 
2566 	down_write(&ndev->reslock);
2567 
2568 	err = mlx5_vdpa_handle_set_map(mvdev, iotlb, &change_map);
2569 	if (err) {
2570 		mlx5_vdpa_warn(mvdev, "set map failed(%d)\n", err);
2571 		goto err;
2572 	}
2573 
2574 	if (change_map)
2575 		err = mlx5_vdpa_change_map(mvdev, iotlb);
2576 
2577 err:
2578 	up_write(&ndev->reslock);
2579 	return err;
2580 }
2581 
2582 static void mlx5_vdpa_free(struct vdpa_device *vdev)
2583 {
2584 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2585 	struct mlx5_core_dev *pfmdev;
2586 	struct mlx5_vdpa_net *ndev;
2587 
2588 	ndev = to_mlx5_vdpa_ndev(mvdev);
2589 
2590 	free_resources(ndev);
2591 	mlx5_vdpa_destroy_mr(mvdev);
2592 	if (!is_zero_ether_addr(ndev->config.mac)) {
2593 		pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
2594 		mlx5_mpfs_del_mac(pfmdev, ndev->config.mac);
2595 	}
2596 	mlx5_vdpa_free_resources(&ndev->mvdev);
2597 	kfree(ndev->event_cbs);
2598 	kfree(ndev->vqs);
2599 }
2600 
2601 static struct vdpa_notification_area mlx5_get_vq_notification(struct vdpa_device *vdev, u16 idx)
2602 {
2603 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2604 	struct vdpa_notification_area ret = {};
2605 	struct mlx5_vdpa_net *ndev;
2606 	phys_addr_t addr;
2607 
2608 	if (!is_index_valid(mvdev, idx) || is_ctrl_vq_idx(mvdev, idx))
2609 		return ret;
2610 
2611 	/* If SF BAR size is smaller than PAGE_SIZE, do not use direct
2612 	 * notification to avoid the risk of mapping pages that contain BAR of more
2613 	 * than one SF
2614 	 */
2615 	if (MLX5_CAP_GEN(mvdev->mdev, log_min_sf_size) + 12 < PAGE_SHIFT)
2616 		return ret;
2617 
2618 	ndev = to_mlx5_vdpa_ndev(mvdev);
2619 	addr = (phys_addr_t)ndev->mvdev.res.phys_kick_addr;
2620 	ret.addr = addr;
2621 	ret.size = PAGE_SIZE;
2622 	return ret;
2623 }
2624 
2625 static int mlx5_get_vq_irq(struct vdpa_device *vdv, u16 idx)
2626 {
2627 	return -EOPNOTSUPP;
2628 }
2629 
2630 static u64 mlx5_vdpa_get_driver_features(struct vdpa_device *vdev)
2631 {
2632 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2633 
2634 	return mvdev->actual_features;
2635 }
2636 
2637 static int counter_set_query(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
2638 			     u64 *received_desc, u64 *completed_desc)
2639 {
2640 	u32 in[MLX5_ST_SZ_DW(query_virtio_q_counters_in)] = {};
2641 	u32 out[MLX5_ST_SZ_DW(query_virtio_q_counters_out)] = {};
2642 	void *cmd_hdr;
2643 	void *ctx;
2644 	int err;
2645 
2646 	if (!counters_supported(&ndev->mvdev))
2647 		return -EOPNOTSUPP;
2648 
2649 	if (mvq->fw_state != MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY)
2650 		return -EAGAIN;
2651 
2652 	cmd_hdr = MLX5_ADDR_OF(query_virtio_q_counters_in, in, hdr);
2653 
2654 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_QUERY_GENERAL_OBJECT);
2655 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
2656 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
2657 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->counter_set_id);
2658 
2659 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out));
2660 	if (err)
2661 		return err;
2662 
2663 	ctx = MLX5_ADDR_OF(query_virtio_q_counters_out, out, counters);
2664 	*received_desc = MLX5_GET64(virtio_q_counters, ctx, received_desc);
2665 	*completed_desc = MLX5_GET64(virtio_q_counters, ctx, completed_desc);
2666 	return 0;
2667 }
2668 
2669 static int mlx5_vdpa_get_vendor_vq_stats(struct vdpa_device *vdev, u16 idx,
2670 					 struct sk_buff *msg,
2671 					 struct netlink_ext_ack *extack)
2672 {
2673 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2674 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2675 	struct mlx5_vdpa_virtqueue *mvq;
2676 	struct mlx5_control_vq *cvq;
2677 	u64 received_desc;
2678 	u64 completed_desc;
2679 	int err = 0;
2680 
2681 	down_read(&ndev->reslock);
2682 	if (!is_index_valid(mvdev, idx)) {
2683 		NL_SET_ERR_MSG_MOD(extack, "virtqueue index is not valid");
2684 		err = -EINVAL;
2685 		goto out_err;
2686 	}
2687 
2688 	if (idx == ctrl_vq_idx(mvdev)) {
2689 		cvq = &mvdev->cvq;
2690 		received_desc = cvq->received_desc;
2691 		completed_desc = cvq->completed_desc;
2692 		goto out;
2693 	}
2694 
2695 	mvq = &ndev->vqs[idx];
2696 	err = counter_set_query(ndev, mvq, &received_desc, &completed_desc);
2697 	if (err) {
2698 		NL_SET_ERR_MSG_MOD(extack, "failed to query hardware");
2699 		goto out_err;
2700 	}
2701 
2702 out:
2703 	err = -EMSGSIZE;
2704 	if (nla_put_string(msg, VDPA_ATTR_DEV_VENDOR_ATTR_NAME, "received_desc"))
2705 		goto out_err;
2706 
2707 	if (nla_put_u64_64bit(msg, VDPA_ATTR_DEV_VENDOR_ATTR_VALUE, received_desc,
2708 			      VDPA_ATTR_PAD))
2709 		goto out_err;
2710 
2711 	if (nla_put_string(msg, VDPA_ATTR_DEV_VENDOR_ATTR_NAME, "completed_desc"))
2712 		goto out_err;
2713 
2714 	if (nla_put_u64_64bit(msg, VDPA_ATTR_DEV_VENDOR_ATTR_VALUE, completed_desc,
2715 			      VDPA_ATTR_PAD))
2716 		goto out_err;
2717 
2718 	err = 0;
2719 out_err:
2720 	up_read(&ndev->reslock);
2721 	return err;
2722 }
2723 
2724 static const struct vdpa_config_ops mlx5_vdpa_ops = {
2725 	.set_vq_address = mlx5_vdpa_set_vq_address,
2726 	.set_vq_num = mlx5_vdpa_set_vq_num,
2727 	.kick_vq = mlx5_vdpa_kick_vq,
2728 	.set_vq_cb = mlx5_vdpa_set_vq_cb,
2729 	.set_vq_ready = mlx5_vdpa_set_vq_ready,
2730 	.get_vq_ready = mlx5_vdpa_get_vq_ready,
2731 	.set_vq_state = mlx5_vdpa_set_vq_state,
2732 	.get_vq_state = mlx5_vdpa_get_vq_state,
2733 	.get_vendor_vq_stats = mlx5_vdpa_get_vendor_vq_stats,
2734 	.get_vq_notification = mlx5_get_vq_notification,
2735 	.get_vq_irq = mlx5_get_vq_irq,
2736 	.get_vq_align = mlx5_vdpa_get_vq_align,
2737 	.get_vq_group = mlx5_vdpa_get_vq_group,
2738 	.get_device_features = mlx5_vdpa_get_device_features,
2739 	.set_driver_features = mlx5_vdpa_set_driver_features,
2740 	.get_driver_features = mlx5_vdpa_get_driver_features,
2741 	.set_config_cb = mlx5_vdpa_set_config_cb,
2742 	.get_vq_num_max = mlx5_vdpa_get_vq_num_max,
2743 	.get_device_id = mlx5_vdpa_get_device_id,
2744 	.get_vendor_id = mlx5_vdpa_get_vendor_id,
2745 	.get_status = mlx5_vdpa_get_status,
2746 	.set_status = mlx5_vdpa_set_status,
2747 	.reset = mlx5_vdpa_reset,
2748 	.get_config_size = mlx5_vdpa_get_config_size,
2749 	.get_config = mlx5_vdpa_get_config,
2750 	.set_config = mlx5_vdpa_set_config,
2751 	.get_generation = mlx5_vdpa_get_generation,
2752 	.set_map = mlx5_vdpa_set_map,
2753 	.free = mlx5_vdpa_free,
2754 };
2755 
2756 static int query_mtu(struct mlx5_core_dev *mdev, u16 *mtu)
2757 {
2758 	u16 hw_mtu;
2759 	int err;
2760 
2761 	err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2762 	if (err)
2763 		return err;
2764 
2765 	*mtu = hw_mtu - MLX5V_ETH_HARD_MTU;
2766 	return 0;
2767 }
2768 
2769 static int alloc_resources(struct mlx5_vdpa_net *ndev)
2770 {
2771 	struct mlx5_vdpa_net_resources *res = &ndev->res;
2772 	int err;
2773 
2774 	if (res->valid) {
2775 		mlx5_vdpa_warn(&ndev->mvdev, "resources already allocated\n");
2776 		return -EEXIST;
2777 	}
2778 
2779 	err = mlx5_vdpa_alloc_transport_domain(&ndev->mvdev, &res->tdn);
2780 	if (err)
2781 		return err;
2782 
2783 	err = create_tis(ndev);
2784 	if (err)
2785 		goto err_tis;
2786 
2787 	res->valid = true;
2788 
2789 	return 0;
2790 
2791 err_tis:
2792 	mlx5_vdpa_dealloc_transport_domain(&ndev->mvdev, res->tdn);
2793 	return err;
2794 }
2795 
2796 static void free_resources(struct mlx5_vdpa_net *ndev)
2797 {
2798 	struct mlx5_vdpa_net_resources *res = &ndev->res;
2799 
2800 	if (!res->valid)
2801 		return;
2802 
2803 	destroy_tis(ndev);
2804 	mlx5_vdpa_dealloc_transport_domain(&ndev->mvdev, res->tdn);
2805 	res->valid = false;
2806 }
2807 
2808 static void init_mvqs(struct mlx5_vdpa_net *ndev)
2809 {
2810 	struct mlx5_vdpa_virtqueue *mvq;
2811 	int i;
2812 
2813 	for (i = 0; i < ndev->mvdev.max_vqs; ++i) {
2814 		mvq = &ndev->vqs[i];
2815 		memset(mvq, 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2816 		mvq->index = i;
2817 		mvq->ndev = ndev;
2818 		mvq->fwqp.fw = true;
2819 	}
2820 	for (; i < ndev->mvdev.max_vqs; i++) {
2821 		mvq = &ndev->vqs[i];
2822 		memset(mvq, 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2823 		mvq->index = i;
2824 		mvq->ndev = ndev;
2825 	}
2826 }
2827 
2828 struct mlx5_vdpa_mgmtdev {
2829 	struct vdpa_mgmt_dev mgtdev;
2830 	struct mlx5_adev *madev;
2831 	struct mlx5_vdpa_net *ndev;
2832 };
2833 
2834 static u8 query_vport_state(struct mlx5_core_dev *mdev, u8 opmod, u16 vport)
2835 {
2836 	u32 out[MLX5_ST_SZ_DW(query_vport_state_out)] = {};
2837 	u32 in[MLX5_ST_SZ_DW(query_vport_state_in)] = {};
2838 	int err;
2839 
2840 	MLX5_SET(query_vport_state_in, in, opcode, MLX5_CMD_OP_QUERY_VPORT_STATE);
2841 	MLX5_SET(query_vport_state_in, in, op_mod, opmod);
2842 	MLX5_SET(query_vport_state_in, in, vport_number, vport);
2843 	if (vport)
2844 		MLX5_SET(query_vport_state_in, in, other_vport, 1);
2845 
2846 	err = mlx5_cmd_exec_inout(mdev, query_vport_state, in, out);
2847 	if (err)
2848 		return 0;
2849 
2850 	return MLX5_GET(query_vport_state_out, out, state);
2851 }
2852 
2853 static bool get_link_state(struct mlx5_vdpa_dev *mvdev)
2854 {
2855 	if (query_vport_state(mvdev->mdev, MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT, 0) ==
2856 	    VPORT_STATE_UP)
2857 		return true;
2858 
2859 	return false;
2860 }
2861 
2862 static void update_carrier(struct work_struct *work)
2863 {
2864 	struct mlx5_vdpa_wq_ent *wqent;
2865 	struct mlx5_vdpa_dev *mvdev;
2866 	struct mlx5_vdpa_net *ndev;
2867 
2868 	wqent = container_of(work, struct mlx5_vdpa_wq_ent, work);
2869 	mvdev = wqent->mvdev;
2870 	ndev = to_mlx5_vdpa_ndev(mvdev);
2871 	if (get_link_state(mvdev))
2872 		ndev->config.status |= cpu_to_mlx5vdpa16(mvdev, VIRTIO_NET_S_LINK_UP);
2873 	else
2874 		ndev->config.status &= cpu_to_mlx5vdpa16(mvdev, ~VIRTIO_NET_S_LINK_UP);
2875 
2876 	if (ndev->config_cb.callback)
2877 		ndev->config_cb.callback(ndev->config_cb.private);
2878 
2879 	kfree(wqent);
2880 }
2881 
2882 static int event_handler(struct notifier_block *nb, unsigned long event, void *param)
2883 {
2884 	struct mlx5_vdpa_net *ndev = container_of(nb, struct mlx5_vdpa_net, nb);
2885 	struct mlx5_eqe *eqe = param;
2886 	int ret = NOTIFY_DONE;
2887 	struct mlx5_vdpa_wq_ent *wqent;
2888 
2889 	if (event == MLX5_EVENT_TYPE_PORT_CHANGE) {
2890 		switch (eqe->sub_type) {
2891 		case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2892 		case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2893 			wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
2894 			if (!wqent)
2895 				return NOTIFY_DONE;
2896 
2897 			wqent->mvdev = &ndev->mvdev;
2898 			INIT_WORK(&wqent->work, update_carrier);
2899 			queue_work(ndev->mvdev.wq, &wqent->work);
2900 			ret = NOTIFY_OK;
2901 			break;
2902 		default:
2903 			return NOTIFY_DONE;
2904 		}
2905 		return ret;
2906 	}
2907 	return ret;
2908 }
2909 
2910 static int config_func_mtu(struct mlx5_core_dev *mdev, u16 mtu)
2911 {
2912 	int inlen = MLX5_ST_SZ_BYTES(modify_nic_vport_context_in);
2913 	void *in;
2914 	int err;
2915 
2916 	in = kvzalloc(inlen, GFP_KERNEL);
2917 	if (!in)
2918 		return -ENOMEM;
2919 
2920 	MLX5_SET(modify_nic_vport_context_in, in, field_select.mtu, 1);
2921 	MLX5_SET(modify_nic_vport_context_in, in, nic_vport_context.mtu,
2922 		 mtu + MLX5V_ETH_HARD_MTU);
2923 	MLX5_SET(modify_nic_vport_context_in, in, opcode,
2924 		 MLX5_CMD_OP_MODIFY_NIC_VPORT_CONTEXT);
2925 
2926 	err = mlx5_cmd_exec_in(mdev, modify_nic_vport_context, in);
2927 
2928 	kvfree(in);
2929 	return err;
2930 }
2931 
2932 static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
2933 			     const struct vdpa_dev_set_config *add_config)
2934 {
2935 	struct mlx5_vdpa_mgmtdev *mgtdev = container_of(v_mdev, struct mlx5_vdpa_mgmtdev, mgtdev);
2936 	struct virtio_net_config *config;
2937 	struct mlx5_core_dev *pfmdev;
2938 	struct mlx5_vdpa_dev *mvdev;
2939 	struct mlx5_vdpa_net *ndev;
2940 	struct mlx5_core_dev *mdev;
2941 	u32 max_vqs;
2942 	u16 mtu;
2943 	int err;
2944 
2945 	if (mgtdev->ndev)
2946 		return -ENOSPC;
2947 
2948 	mdev = mgtdev->madev->mdev;
2949 	if (!(MLX5_CAP_DEV_VDPA_EMULATION(mdev, virtio_queue_type) &
2950 	    MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_SPLIT)) {
2951 		dev_warn(mdev->device, "missing support for split virtqueues\n");
2952 		return -EOPNOTSUPP;
2953 	}
2954 
2955 	max_vqs = min_t(int, MLX5_CAP_DEV_VDPA_EMULATION(mdev, max_num_virtio_queues),
2956 			1 << MLX5_CAP_GEN(mdev, log_max_rqt_size));
2957 	if (max_vqs < 2) {
2958 		dev_warn(mdev->device,
2959 			 "%d virtqueues are supported. At least 2 are required\n",
2960 			 max_vqs);
2961 		return -EAGAIN;
2962 	}
2963 
2964 	if (add_config->mask & BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MAX_VQP)) {
2965 		if (add_config->net.max_vq_pairs > max_vqs / 2)
2966 			return -EINVAL;
2967 		max_vqs = min_t(u32, max_vqs, 2 * add_config->net.max_vq_pairs);
2968 	} else {
2969 		max_vqs = 2;
2970 	}
2971 
2972 	ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev->device, &mlx5_vdpa_ops,
2973 				 1, 1, name, false);
2974 	if (IS_ERR(ndev))
2975 		return PTR_ERR(ndev);
2976 
2977 	ndev->mvdev.mlx_features = mgtdev->mgtdev.supported_features;
2978 	ndev->mvdev.max_vqs = max_vqs;
2979 	mvdev = &ndev->mvdev;
2980 	mvdev->mdev = mdev;
2981 
2982 	ndev->vqs = kcalloc(max_vqs, sizeof(*ndev->vqs), GFP_KERNEL);
2983 	ndev->event_cbs = kcalloc(max_vqs + 1, sizeof(*ndev->event_cbs), GFP_KERNEL);
2984 	if (!ndev->vqs || !ndev->event_cbs) {
2985 		err = -ENOMEM;
2986 		goto err_alloc;
2987 	}
2988 
2989 	init_mvqs(ndev);
2990 	init_rwsem(&ndev->reslock);
2991 	config = &ndev->config;
2992 
2993 	if (add_config->mask & BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MTU)) {
2994 		err = config_func_mtu(mdev, add_config->net.mtu);
2995 		if (err)
2996 			goto err_alloc;
2997 	}
2998 
2999 	err = query_mtu(mdev, &mtu);
3000 	if (err)
3001 		goto err_alloc;
3002 
3003 	ndev->config.mtu = cpu_to_mlx5vdpa16(mvdev, mtu);
3004 
3005 	if (get_link_state(mvdev))
3006 		ndev->config.status |= cpu_to_mlx5vdpa16(mvdev, VIRTIO_NET_S_LINK_UP);
3007 	else
3008 		ndev->config.status &= cpu_to_mlx5vdpa16(mvdev, ~VIRTIO_NET_S_LINK_UP);
3009 
3010 	if (add_config->mask & (1 << VDPA_ATTR_DEV_NET_CFG_MACADDR)) {
3011 		memcpy(ndev->config.mac, add_config->net.mac, ETH_ALEN);
3012 	} else {
3013 		err = mlx5_query_nic_vport_mac_address(mdev, 0, 0, config->mac);
3014 		if (err)
3015 			goto err_alloc;
3016 	}
3017 
3018 	if (!is_zero_ether_addr(config->mac)) {
3019 		pfmdev = pci_get_drvdata(pci_physfn(mdev->pdev));
3020 		err = mlx5_mpfs_add_mac(pfmdev, config->mac);
3021 		if (err)
3022 			goto err_alloc;
3023 
3024 		ndev->mvdev.mlx_features |= BIT_ULL(VIRTIO_NET_F_MAC);
3025 	}
3026 
3027 	config->max_virtqueue_pairs = cpu_to_mlx5vdpa16(mvdev, max_vqs / 2);
3028 	mvdev->vdev.dma_dev = &mdev->pdev->dev;
3029 	err = mlx5_vdpa_alloc_resources(&ndev->mvdev);
3030 	if (err)
3031 		goto err_mpfs;
3032 
3033 	if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
3034 		err = mlx5_vdpa_create_mr(mvdev, NULL);
3035 		if (err)
3036 			goto err_res;
3037 	}
3038 
3039 	err = alloc_resources(ndev);
3040 	if (err)
3041 		goto err_mr;
3042 
3043 	ndev->cvq_ent.mvdev = mvdev;
3044 	INIT_WORK(&ndev->cvq_ent.work, mlx5_cvq_kick_handler);
3045 	mvdev->wq = create_singlethread_workqueue("mlx5_vdpa_wq");
3046 	if (!mvdev->wq) {
3047 		err = -ENOMEM;
3048 		goto err_res2;
3049 	}
3050 
3051 	ndev->nb.notifier_call = event_handler;
3052 	mlx5_notifier_register(mdev, &ndev->nb);
3053 	mvdev->vdev.mdev = &mgtdev->mgtdev;
3054 	err = _vdpa_register_device(&mvdev->vdev, max_vqs + 1);
3055 	if (err)
3056 		goto err_reg;
3057 
3058 	mgtdev->ndev = ndev;
3059 	return 0;
3060 
3061 err_reg:
3062 	destroy_workqueue(mvdev->wq);
3063 err_res2:
3064 	free_resources(ndev);
3065 err_mr:
3066 	mlx5_vdpa_destroy_mr(mvdev);
3067 err_res:
3068 	mlx5_vdpa_free_resources(&ndev->mvdev);
3069 err_mpfs:
3070 	if (!is_zero_ether_addr(config->mac))
3071 		mlx5_mpfs_del_mac(pfmdev, config->mac);
3072 err_alloc:
3073 	put_device(&mvdev->vdev.dev);
3074 	return err;
3075 }
3076 
3077 static void mlx5_vdpa_dev_del(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *dev)
3078 {
3079 	struct mlx5_vdpa_mgmtdev *mgtdev = container_of(v_mdev, struct mlx5_vdpa_mgmtdev, mgtdev);
3080 	struct mlx5_vdpa_dev *mvdev = to_mvdev(dev);
3081 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
3082 	struct workqueue_struct *wq;
3083 
3084 	mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
3085 	wq = mvdev->wq;
3086 	mvdev->wq = NULL;
3087 	destroy_workqueue(wq);
3088 	_vdpa_unregister_device(dev);
3089 	mgtdev->ndev = NULL;
3090 }
3091 
3092 static const struct vdpa_mgmtdev_ops mdev_ops = {
3093 	.dev_add = mlx5_vdpa_dev_add,
3094 	.dev_del = mlx5_vdpa_dev_del,
3095 };
3096 
3097 static struct virtio_device_id id_table[] = {
3098 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3099 	{ 0 },
3100 };
3101 
3102 static int mlx5v_probe(struct auxiliary_device *adev,
3103 		       const struct auxiliary_device_id *id)
3104 
3105 {
3106 	struct mlx5_adev *madev = container_of(adev, struct mlx5_adev, adev);
3107 	struct mlx5_core_dev *mdev = madev->mdev;
3108 	struct mlx5_vdpa_mgmtdev *mgtdev;
3109 	int err;
3110 
3111 	mgtdev = kzalloc(sizeof(*mgtdev), GFP_KERNEL);
3112 	if (!mgtdev)
3113 		return -ENOMEM;
3114 
3115 	mgtdev->mgtdev.ops = &mdev_ops;
3116 	mgtdev->mgtdev.device = mdev->device;
3117 	mgtdev->mgtdev.id_table = id_table;
3118 	mgtdev->mgtdev.config_attr_mask = BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MACADDR) |
3119 					  BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MAX_VQP) |
3120 					  BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MTU);
3121 	mgtdev->mgtdev.max_supported_vqs =
3122 		MLX5_CAP_DEV_VDPA_EMULATION(mdev, max_num_virtio_queues) + 1;
3123 	mgtdev->mgtdev.supported_features = get_supported_features(mdev);
3124 	mgtdev->madev = madev;
3125 
3126 	err = vdpa_mgmtdev_register(&mgtdev->mgtdev);
3127 	if (err)
3128 		goto reg_err;
3129 
3130 	auxiliary_set_drvdata(adev, mgtdev);
3131 
3132 	return 0;
3133 
3134 reg_err:
3135 	kfree(mgtdev);
3136 	return err;
3137 }
3138 
3139 static void mlx5v_remove(struct auxiliary_device *adev)
3140 {
3141 	struct mlx5_vdpa_mgmtdev *mgtdev;
3142 
3143 	mgtdev = auxiliary_get_drvdata(adev);
3144 	vdpa_mgmtdev_unregister(&mgtdev->mgtdev);
3145 	kfree(mgtdev);
3146 }
3147 
3148 static const struct auxiliary_device_id mlx5v_id_table[] = {
3149 	{ .name = MLX5_ADEV_NAME ".vnet", },
3150 	{},
3151 };
3152 
3153 MODULE_DEVICE_TABLE(auxiliary, mlx5v_id_table);
3154 
3155 static struct auxiliary_driver mlx5v_driver = {
3156 	.name = "vnet",
3157 	.probe = mlx5v_probe,
3158 	.remove = mlx5v_remove,
3159 	.id_table = mlx5v_id_table,
3160 };
3161 
3162 module_auxiliary_driver(mlx5v_driver);
3163