xref: /linux/drivers/vdpa/mlx5/net/mlx5_vnet.c (revision 280330fac48280e16454cfa46c368af4812ad79c)
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 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 	bool nb_registered;
168 	struct notifier_block nb;
169 	struct vdpa_callback config_cb;
170 	struct mlx5_vdpa_wq_ent cvq_ent;
171 	struct hlist_head macvlan_hash[MLX5V_MACVLAN_SIZE];
172 };
173 
174 struct macvlan_node {
175 	struct hlist_node hlist;
176 	struct mlx5_flow_handle *ucast_rule;
177 	struct mlx5_flow_handle *mcast_rule;
178 	u64 macvlan;
179 };
180 
181 static void free_resources(struct mlx5_vdpa_net *ndev);
182 static void init_mvqs(struct mlx5_vdpa_net *ndev);
183 static int setup_driver(struct mlx5_vdpa_dev *mvdev);
184 static void teardown_driver(struct mlx5_vdpa_net *ndev);
185 
186 static bool mlx5_vdpa_debug;
187 
188 #define MLX5_CVQ_MAX_ENT 16
189 
190 #define MLX5_LOG_VIO_FLAG(_feature)                                                                \
191 	do {                                                                                       \
192 		if (features & BIT_ULL(_feature))                                                  \
193 			mlx5_vdpa_info(mvdev, "%s\n", #_feature);                                  \
194 	} while (0)
195 
196 #define MLX5_LOG_VIO_STAT(_status)                                                                 \
197 	do {                                                                                       \
198 		if (status & (_status))                                                            \
199 			mlx5_vdpa_info(mvdev, "%s\n", #_status);                                   \
200 	} while (0)
201 
202 /* TODO: cross-endian support */
203 static inline bool mlx5_vdpa_is_little_endian(struct mlx5_vdpa_dev *mvdev)
204 {
205 	return virtio_legacy_is_little_endian() ||
206 		(mvdev->actual_features & BIT_ULL(VIRTIO_F_VERSION_1));
207 }
208 
209 static u16 mlx5vdpa16_to_cpu(struct mlx5_vdpa_dev *mvdev, __virtio16 val)
210 {
211 	return __virtio16_to_cpu(mlx5_vdpa_is_little_endian(mvdev), val);
212 }
213 
214 static __virtio16 cpu_to_mlx5vdpa16(struct mlx5_vdpa_dev *mvdev, u16 val)
215 {
216 	return __cpu_to_virtio16(mlx5_vdpa_is_little_endian(mvdev), val);
217 }
218 
219 static u16 ctrl_vq_idx(struct mlx5_vdpa_dev *mvdev)
220 {
221 	if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_MQ)))
222 		return 2;
223 
224 	return mvdev->max_vqs;
225 }
226 
227 static bool is_ctrl_vq_idx(struct mlx5_vdpa_dev *mvdev, u16 idx)
228 {
229 	return idx == ctrl_vq_idx(mvdev);
230 }
231 
232 static void print_status(struct mlx5_vdpa_dev *mvdev, u8 status, bool set)
233 {
234 	if (status & ~VALID_STATUS_MASK)
235 		mlx5_vdpa_warn(mvdev, "Warning: there are invalid status bits 0x%x\n",
236 			       status & ~VALID_STATUS_MASK);
237 
238 	if (!mlx5_vdpa_debug)
239 		return;
240 
241 	mlx5_vdpa_info(mvdev, "driver status %s", set ? "set" : "get");
242 	if (set && !status) {
243 		mlx5_vdpa_info(mvdev, "driver resets the device\n");
244 		return;
245 	}
246 
247 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_ACKNOWLEDGE);
248 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_DRIVER);
249 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_DRIVER_OK);
250 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_FEATURES_OK);
251 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_NEEDS_RESET);
252 	MLX5_LOG_VIO_STAT(VIRTIO_CONFIG_S_FAILED);
253 }
254 
255 static void print_features(struct mlx5_vdpa_dev *mvdev, u64 features, bool set)
256 {
257 	if (features & ~VALID_FEATURES_MASK)
258 		mlx5_vdpa_warn(mvdev, "There are invalid feature bits 0x%llx\n",
259 			       features & ~VALID_FEATURES_MASK);
260 
261 	if (!mlx5_vdpa_debug)
262 		return;
263 
264 	mlx5_vdpa_info(mvdev, "driver %s feature bits:\n", set ? "sets" : "reads");
265 	if (!features)
266 		mlx5_vdpa_info(mvdev, "all feature bits are cleared\n");
267 
268 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CSUM);
269 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_CSUM);
270 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS);
271 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MTU);
272 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MAC);
273 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_TSO4);
274 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_TSO6);
275 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_ECN);
276 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_UFO);
277 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_TSO4);
278 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_TSO6);
279 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_ECN);
280 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HOST_UFO);
281 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MRG_RXBUF);
282 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_STATUS);
283 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_VQ);
284 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_RX);
285 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_VLAN);
286 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_RX_EXTRA);
287 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_GUEST_ANNOUNCE);
288 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_MQ);
289 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_CTRL_MAC_ADDR);
290 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_HASH_REPORT);
291 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_RSS);
292 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_RSC_EXT);
293 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_STANDBY);
294 	MLX5_LOG_VIO_FLAG(VIRTIO_NET_F_SPEED_DUPLEX);
295 	MLX5_LOG_VIO_FLAG(VIRTIO_F_NOTIFY_ON_EMPTY);
296 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ANY_LAYOUT);
297 	MLX5_LOG_VIO_FLAG(VIRTIO_F_VERSION_1);
298 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ACCESS_PLATFORM);
299 	MLX5_LOG_VIO_FLAG(VIRTIO_F_RING_PACKED);
300 	MLX5_LOG_VIO_FLAG(VIRTIO_F_ORDER_PLATFORM);
301 	MLX5_LOG_VIO_FLAG(VIRTIO_F_SR_IOV);
302 }
303 
304 static int create_tis(struct mlx5_vdpa_net *ndev)
305 {
306 	struct mlx5_vdpa_dev *mvdev = &ndev->mvdev;
307 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
308 	void *tisc;
309 	int err;
310 
311 	tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
312 	MLX5_SET(tisc, tisc, transport_domain, ndev->res.tdn);
313 	err = mlx5_vdpa_create_tis(mvdev, in, &ndev->res.tisn);
314 	if (err)
315 		mlx5_vdpa_warn(mvdev, "create TIS (%d)\n", err);
316 
317 	return err;
318 }
319 
320 static void destroy_tis(struct mlx5_vdpa_net *ndev)
321 {
322 	mlx5_vdpa_destroy_tis(&ndev->mvdev, ndev->res.tisn);
323 }
324 
325 #define MLX5_VDPA_CQE_SIZE 64
326 #define MLX5_VDPA_LOG_CQE_SIZE ilog2(MLX5_VDPA_CQE_SIZE)
327 
328 static int cq_frag_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_cq_buf *buf, int nent)
329 {
330 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
331 	u8 log_wq_stride = MLX5_VDPA_LOG_CQE_SIZE;
332 	u8 log_wq_sz = MLX5_VDPA_LOG_CQE_SIZE;
333 	int err;
334 
335 	err = mlx5_frag_buf_alloc_node(ndev->mvdev.mdev, nent * MLX5_VDPA_CQE_SIZE, frag_buf,
336 				       ndev->mvdev.mdev->priv.numa_node);
337 	if (err)
338 		return err;
339 
340 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
341 
342 	buf->cqe_size = MLX5_VDPA_CQE_SIZE;
343 	buf->nent = nent;
344 
345 	return 0;
346 }
347 
348 static int umem_frag_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_umem *umem, int size)
349 {
350 	struct mlx5_frag_buf *frag_buf = &umem->frag_buf;
351 
352 	return mlx5_frag_buf_alloc_node(ndev->mvdev.mdev, size, frag_buf,
353 					ndev->mvdev.mdev->priv.numa_node);
354 }
355 
356 static void cq_frag_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_cq_buf *buf)
357 {
358 	mlx5_frag_buf_free(ndev->mvdev.mdev, &buf->frag_buf);
359 }
360 
361 static void *get_cqe(struct mlx5_vdpa_cq *vcq, int n)
362 {
363 	return mlx5_frag_buf_get_wqe(&vcq->buf.fbc, n);
364 }
365 
366 static void cq_frag_buf_init(struct mlx5_vdpa_cq *vcq, struct mlx5_vdpa_cq_buf *buf)
367 {
368 	struct mlx5_cqe64 *cqe64;
369 	void *cqe;
370 	int i;
371 
372 	for (i = 0; i < buf->nent; i++) {
373 		cqe = get_cqe(vcq, i);
374 		cqe64 = cqe;
375 		cqe64->op_own = MLX5_CQE_INVALID << 4;
376 	}
377 }
378 
379 static void *get_sw_cqe(struct mlx5_vdpa_cq *cq, int n)
380 {
381 	struct mlx5_cqe64 *cqe64 = get_cqe(cq, n & (cq->cqe - 1));
382 
383 	if (likely(get_cqe_opcode(cqe64) != MLX5_CQE_INVALID) &&
384 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & cq->cqe)))
385 		return cqe64;
386 
387 	return NULL;
388 }
389 
390 static void rx_post(struct mlx5_vdpa_qp *vqp, int n)
391 {
392 	vqp->head += n;
393 	vqp->db.db[0] = cpu_to_be32(vqp->head);
394 }
395 
396 static void qp_prepare(struct mlx5_vdpa_net *ndev, bool fw, void *in,
397 		       struct mlx5_vdpa_virtqueue *mvq, u32 num_ent)
398 {
399 	struct mlx5_vdpa_qp *vqp;
400 	__be64 *pas;
401 	void *qpc;
402 
403 	vqp = fw ? &mvq->fwqp : &mvq->vqqp;
404 	MLX5_SET(create_qp_in, in, uid, ndev->mvdev.res.uid);
405 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
406 	if (vqp->fw) {
407 		/* Firmware QP is allocated by the driver for the firmware's
408 		 * use so we can skip part of the params as they will be chosen by firmware
409 		 */
410 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
411 		MLX5_SET(qpc, qpc, rq_type, MLX5_ZERO_LEN_RQ);
412 		MLX5_SET(qpc, qpc, no_sq, 1);
413 		return;
414 	}
415 
416 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_RC);
417 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
418 	MLX5_SET(qpc, qpc, pd, ndev->mvdev.res.pdn);
419 	MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
420 	MLX5_SET(qpc, qpc, uar_page, ndev->mvdev.res.uar->index);
421 	MLX5_SET(qpc, qpc, log_page_size, vqp->frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
422 	MLX5_SET(qpc, qpc, no_sq, 1);
423 	MLX5_SET(qpc, qpc, cqn_rcv, mvq->cq.mcq.cqn);
424 	MLX5_SET(qpc, qpc, log_rq_size, ilog2(num_ent));
425 	MLX5_SET(qpc, qpc, rq_type, MLX5_NON_ZERO_RQ);
426 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, in, pas);
427 	mlx5_fill_page_frag_array(&vqp->frag_buf, pas);
428 }
429 
430 static int rq_buf_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp, u32 num_ent)
431 {
432 	return mlx5_frag_buf_alloc_node(ndev->mvdev.mdev,
433 					num_ent * sizeof(struct mlx5_wqe_data_seg), &vqp->frag_buf,
434 					ndev->mvdev.mdev->priv.numa_node);
435 }
436 
437 static void rq_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp)
438 {
439 	mlx5_frag_buf_free(ndev->mvdev.mdev, &vqp->frag_buf);
440 }
441 
442 static int qp_create(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
443 		     struct mlx5_vdpa_qp *vqp)
444 {
445 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
446 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
447 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
448 	void *qpc;
449 	void *in;
450 	int err;
451 
452 	if (!vqp->fw) {
453 		vqp = &mvq->vqqp;
454 		err = rq_buf_alloc(ndev, vqp, mvq->num_ent);
455 		if (err)
456 			return err;
457 
458 		err = mlx5_db_alloc(ndev->mvdev.mdev, &vqp->db);
459 		if (err)
460 			goto err_db;
461 		inlen += vqp->frag_buf.npages * sizeof(__be64);
462 	}
463 
464 	in = kzalloc(inlen, GFP_KERNEL);
465 	if (!in) {
466 		err = -ENOMEM;
467 		goto err_kzalloc;
468 	}
469 
470 	qp_prepare(ndev, vqp->fw, in, mvq, mvq->num_ent);
471 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
472 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_RC);
473 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
474 	MLX5_SET(qpc, qpc, pd, ndev->mvdev.res.pdn);
475 	MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
476 	if (!vqp->fw)
477 		MLX5_SET64(qpc, qpc, dbr_addr, vqp->db.dma);
478 	MLX5_SET(create_qp_in, in, opcode, MLX5_CMD_OP_CREATE_QP);
479 	err = mlx5_cmd_exec(mdev, in, inlen, out, sizeof(out));
480 	kfree(in);
481 	if (err)
482 		goto err_kzalloc;
483 
484 	vqp->mqp.uid = ndev->mvdev.res.uid;
485 	vqp->mqp.qpn = MLX5_GET(create_qp_out, out, qpn);
486 
487 	if (!vqp->fw)
488 		rx_post(vqp, mvq->num_ent);
489 
490 	return 0;
491 
492 err_kzalloc:
493 	if (!vqp->fw)
494 		mlx5_db_free(ndev->mvdev.mdev, &vqp->db);
495 err_db:
496 	if (!vqp->fw)
497 		rq_buf_free(ndev, vqp);
498 
499 	return err;
500 }
501 
502 static void qp_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_qp *vqp)
503 {
504 	u32 in[MLX5_ST_SZ_DW(destroy_qp_in)] = {};
505 
506 	MLX5_SET(destroy_qp_in, in, opcode, MLX5_CMD_OP_DESTROY_QP);
507 	MLX5_SET(destroy_qp_in, in, qpn, vqp->mqp.qpn);
508 	MLX5_SET(destroy_qp_in, in, uid, ndev->mvdev.res.uid);
509 	if (mlx5_cmd_exec_in(ndev->mvdev.mdev, destroy_qp, in))
510 		mlx5_vdpa_warn(&ndev->mvdev, "destroy qp 0x%x\n", vqp->mqp.qpn);
511 	if (!vqp->fw) {
512 		mlx5_db_free(ndev->mvdev.mdev, &vqp->db);
513 		rq_buf_free(ndev, vqp);
514 	}
515 }
516 
517 static void *next_cqe_sw(struct mlx5_vdpa_cq *cq)
518 {
519 	return get_sw_cqe(cq, cq->mcq.cons_index);
520 }
521 
522 static int mlx5_vdpa_poll_one(struct mlx5_vdpa_cq *vcq)
523 {
524 	struct mlx5_cqe64 *cqe64;
525 
526 	cqe64 = next_cqe_sw(vcq);
527 	if (!cqe64)
528 		return -EAGAIN;
529 
530 	vcq->mcq.cons_index++;
531 	return 0;
532 }
533 
534 static void mlx5_vdpa_handle_completions(struct mlx5_vdpa_virtqueue *mvq, int num)
535 {
536 	struct mlx5_vdpa_net *ndev = mvq->ndev;
537 	struct vdpa_callback *event_cb;
538 
539 	event_cb = &ndev->event_cbs[mvq->index];
540 	mlx5_cq_set_ci(&mvq->cq.mcq);
541 
542 	/* make sure CQ cosumer update is visible to the hardware before updating
543 	 * RX doorbell record.
544 	 */
545 	dma_wmb();
546 	rx_post(&mvq->vqqp, num);
547 	if (event_cb->callback)
548 		event_cb->callback(event_cb->private);
549 }
550 
551 static void mlx5_vdpa_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe)
552 {
553 	struct mlx5_vdpa_virtqueue *mvq = container_of(mcq, struct mlx5_vdpa_virtqueue, cq.mcq);
554 	struct mlx5_vdpa_net *ndev = mvq->ndev;
555 	void __iomem *uar_page = ndev->mvdev.res.uar->map;
556 	int num = 0;
557 
558 	while (!mlx5_vdpa_poll_one(&mvq->cq)) {
559 		num++;
560 		if (num > mvq->num_ent / 2) {
561 			/* If completions keep coming while we poll, we want to
562 			 * let the hardware know that we consumed them by
563 			 * updating the doorbell record.  We also let vdpa core
564 			 * know about this so it passes it on the virtio driver
565 			 * on the guest.
566 			 */
567 			mlx5_vdpa_handle_completions(mvq, num);
568 			num = 0;
569 		}
570 	}
571 
572 	if (num)
573 		mlx5_vdpa_handle_completions(mvq, num);
574 
575 	mlx5_cq_arm(&mvq->cq.mcq, MLX5_CQ_DB_REQ_NOT, uar_page, mvq->cq.mcq.cons_index);
576 }
577 
578 static int cq_create(struct mlx5_vdpa_net *ndev, u16 idx, u32 num_ent)
579 {
580 	struct mlx5_vdpa_virtqueue *mvq = &ndev->vqs[idx];
581 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
582 	void __iomem *uar_page = ndev->mvdev.res.uar->map;
583 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
584 	struct mlx5_vdpa_cq *vcq = &mvq->cq;
585 	__be64 *pas;
586 	int inlen;
587 	void *cqc;
588 	void *in;
589 	int err;
590 	int eqn;
591 
592 	err = mlx5_db_alloc(mdev, &vcq->db);
593 	if (err)
594 		return err;
595 
596 	vcq->mcq.set_ci_db = vcq->db.db;
597 	vcq->mcq.arm_db = vcq->db.db + 1;
598 	vcq->mcq.cqe_sz = 64;
599 
600 	err = cq_frag_buf_alloc(ndev, &vcq->buf, num_ent);
601 	if (err)
602 		goto err_db;
603 
604 	cq_frag_buf_init(vcq, &vcq->buf);
605 
606 	inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
607 		MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * vcq->buf.frag_buf.npages;
608 	in = kzalloc(inlen, GFP_KERNEL);
609 	if (!in) {
610 		err = -ENOMEM;
611 		goto err_vzalloc;
612 	}
613 
614 	MLX5_SET(create_cq_in, in, uid, ndev->mvdev.res.uid);
615 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas);
616 	mlx5_fill_page_frag_array(&vcq->buf.frag_buf, pas);
617 
618 	cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
619 	MLX5_SET(cqc, cqc, log_page_size, vcq->buf.frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
620 
621 	/* Use vector 0 by default. Consider adding code to choose least used
622 	 * vector.
623 	 */
624 	err = mlx5_vector2eqn(mdev, 0, &eqn);
625 	if (err)
626 		goto err_vec;
627 
628 	cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
629 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(num_ent));
630 	MLX5_SET(cqc, cqc, uar_page, ndev->mvdev.res.uar->index);
631 	MLX5_SET(cqc, cqc, c_eqn_or_apu_element, eqn);
632 	MLX5_SET64(cqc, cqc, dbr_addr, vcq->db.dma);
633 
634 	err = mlx5_core_create_cq(mdev, &vcq->mcq, in, inlen, out, sizeof(out));
635 	if (err)
636 		goto err_vec;
637 
638 	vcq->mcq.comp = mlx5_vdpa_cq_comp;
639 	vcq->cqe = num_ent;
640 	vcq->mcq.set_ci_db = vcq->db.db;
641 	vcq->mcq.arm_db = vcq->db.db + 1;
642 	mlx5_cq_arm(&mvq->cq.mcq, MLX5_CQ_DB_REQ_NOT, uar_page, mvq->cq.mcq.cons_index);
643 	kfree(in);
644 	return 0;
645 
646 err_vec:
647 	kfree(in);
648 err_vzalloc:
649 	cq_frag_buf_free(ndev, &vcq->buf);
650 err_db:
651 	mlx5_db_free(ndev->mvdev.mdev, &vcq->db);
652 	return err;
653 }
654 
655 static void cq_destroy(struct mlx5_vdpa_net *ndev, u16 idx)
656 {
657 	struct mlx5_vdpa_virtqueue *mvq = &ndev->vqs[idx];
658 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
659 	struct mlx5_vdpa_cq *vcq = &mvq->cq;
660 
661 	if (mlx5_core_destroy_cq(mdev, &vcq->mcq)) {
662 		mlx5_vdpa_warn(&ndev->mvdev, "destroy CQ 0x%x\n", vcq->mcq.cqn);
663 		return;
664 	}
665 	cq_frag_buf_free(ndev, &vcq->buf);
666 	mlx5_db_free(ndev->mvdev.mdev, &vcq->db);
667 }
668 
669 static void set_umem_size(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num,
670 			  struct mlx5_vdpa_umem **umemp)
671 {
672 	struct mlx5_core_dev *mdev = ndev->mvdev.mdev;
673 	int p_a;
674 	int p_b;
675 
676 	switch (num) {
677 	case 1:
678 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_1_buffer_param_a);
679 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_1_buffer_param_b);
680 		*umemp = &mvq->umem1;
681 		break;
682 	case 2:
683 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_2_buffer_param_a);
684 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_2_buffer_param_b);
685 		*umemp = &mvq->umem2;
686 		break;
687 	case 3:
688 		p_a = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_3_buffer_param_a);
689 		p_b = MLX5_CAP_DEV_VDPA_EMULATION(mdev, umem_3_buffer_param_b);
690 		*umemp = &mvq->umem3;
691 		break;
692 	}
693 	(*umemp)->size = p_a * mvq->num_ent + p_b;
694 }
695 
696 static void umem_frag_buf_free(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_umem *umem)
697 {
698 	mlx5_frag_buf_free(ndev->mvdev.mdev, &umem->frag_buf);
699 }
700 
701 static int create_umem(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num)
702 {
703 	int inlen;
704 	u32 out[MLX5_ST_SZ_DW(create_umem_out)] = {};
705 	void *um;
706 	void *in;
707 	int err;
708 	__be64 *pas;
709 	struct mlx5_vdpa_umem *umem;
710 
711 	set_umem_size(ndev, mvq, num, &umem);
712 	err = umem_frag_buf_alloc(ndev, umem, umem->size);
713 	if (err)
714 		return err;
715 
716 	inlen = MLX5_ST_SZ_BYTES(create_umem_in) + MLX5_ST_SZ_BYTES(mtt) * umem->frag_buf.npages;
717 
718 	in = kzalloc(inlen, GFP_KERNEL);
719 	if (!in) {
720 		err = -ENOMEM;
721 		goto err_in;
722 	}
723 
724 	MLX5_SET(create_umem_in, in, opcode, MLX5_CMD_OP_CREATE_UMEM);
725 	MLX5_SET(create_umem_in, in, uid, ndev->mvdev.res.uid);
726 	um = MLX5_ADDR_OF(create_umem_in, in, umem);
727 	MLX5_SET(umem, um, log_page_size, umem->frag_buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
728 	MLX5_SET64(umem, um, num_of_mtt, umem->frag_buf.npages);
729 
730 	pas = (__be64 *)MLX5_ADDR_OF(umem, um, mtt[0]);
731 	mlx5_fill_page_frag_array_perm(&umem->frag_buf, pas, MLX5_MTT_PERM_RW);
732 
733 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
734 	if (err) {
735 		mlx5_vdpa_warn(&ndev->mvdev, "create umem(%d)\n", err);
736 		goto err_cmd;
737 	}
738 
739 	kfree(in);
740 	umem->id = MLX5_GET(create_umem_out, out, umem_id);
741 
742 	return 0;
743 
744 err_cmd:
745 	kfree(in);
746 err_in:
747 	umem_frag_buf_free(ndev, umem);
748 	return err;
749 }
750 
751 static void umem_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int num)
752 {
753 	u32 in[MLX5_ST_SZ_DW(destroy_umem_in)] = {};
754 	u32 out[MLX5_ST_SZ_DW(destroy_umem_out)] = {};
755 	struct mlx5_vdpa_umem *umem;
756 
757 	switch (num) {
758 	case 1:
759 		umem = &mvq->umem1;
760 		break;
761 	case 2:
762 		umem = &mvq->umem2;
763 		break;
764 	case 3:
765 		umem = &mvq->umem3;
766 		break;
767 	}
768 
769 	MLX5_SET(destroy_umem_in, in, opcode, MLX5_CMD_OP_DESTROY_UMEM);
770 	MLX5_SET(destroy_umem_in, in, umem_id, umem->id);
771 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out)))
772 		return;
773 
774 	umem_frag_buf_free(ndev, umem);
775 }
776 
777 static int umems_create(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
778 {
779 	int num;
780 	int err;
781 
782 	for (num = 1; num <= 3; num++) {
783 		err = create_umem(ndev, mvq, num);
784 		if (err)
785 			goto err_umem;
786 	}
787 	return 0;
788 
789 err_umem:
790 	for (num--; num > 0; num--)
791 		umem_destroy(ndev, mvq, num);
792 
793 	return err;
794 }
795 
796 static void umems_destroy(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
797 {
798 	int num;
799 
800 	for (num = 3; num > 0; num--)
801 		umem_destroy(ndev, mvq, num);
802 }
803 
804 static int get_queue_type(struct mlx5_vdpa_net *ndev)
805 {
806 	u32 type_mask;
807 
808 	type_mask = MLX5_CAP_DEV_VDPA_EMULATION(ndev->mvdev.mdev, virtio_queue_type);
809 
810 	/* prefer split queue */
811 	if (type_mask & MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_SPLIT)
812 		return MLX5_VIRTIO_EMULATION_VIRTIO_QUEUE_TYPE_SPLIT;
813 
814 	WARN_ON(!(type_mask & MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_PACKED));
815 
816 	return MLX5_VIRTIO_EMULATION_VIRTIO_QUEUE_TYPE_PACKED;
817 }
818 
819 static bool vq_is_tx(u16 idx)
820 {
821 	return idx % 2;
822 }
823 
824 static u16 get_features_12_3(u64 features)
825 {
826 	return (!!(features & BIT_ULL(VIRTIO_NET_F_HOST_TSO4)) << 9) |
827 	       (!!(features & BIT_ULL(VIRTIO_NET_F_HOST_TSO6)) << 8) |
828 	       (!!(features & BIT_ULL(VIRTIO_NET_F_CSUM)) << 7) |
829 	       (!!(features & BIT_ULL(VIRTIO_NET_F_GUEST_CSUM)) << 6);
830 }
831 
832 static bool counters_supported(const struct mlx5_vdpa_dev *mvdev)
833 {
834 	return MLX5_CAP_GEN_64(mvdev->mdev, general_obj_types) &
835 	       BIT_ULL(MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
836 }
837 
838 static int create_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
839 {
840 	int inlen = MLX5_ST_SZ_BYTES(create_virtio_net_q_in);
841 	u32 out[MLX5_ST_SZ_DW(create_virtio_net_q_out)] = {};
842 	void *obj_context;
843 	void *cmd_hdr;
844 	void *vq_ctx;
845 	void *in;
846 	int err;
847 
848 	err = umems_create(ndev, mvq);
849 	if (err)
850 		return err;
851 
852 	in = kzalloc(inlen, GFP_KERNEL);
853 	if (!in) {
854 		err = -ENOMEM;
855 		goto err_alloc;
856 	}
857 
858 	cmd_hdr = MLX5_ADDR_OF(create_virtio_net_q_in, in, general_obj_in_cmd_hdr);
859 
860 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_CREATE_GENERAL_OBJECT);
861 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
862 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
863 
864 	obj_context = MLX5_ADDR_OF(create_virtio_net_q_in, in, obj_context);
865 	MLX5_SET(virtio_net_q_object, obj_context, hw_available_index, mvq->avail_idx);
866 	MLX5_SET(virtio_net_q_object, obj_context, hw_used_index, mvq->used_idx);
867 	MLX5_SET(virtio_net_q_object, obj_context, queue_feature_bit_mask_12_3,
868 		 get_features_12_3(ndev->mvdev.actual_features));
869 	vq_ctx = MLX5_ADDR_OF(virtio_net_q_object, obj_context, virtio_q_context);
870 	MLX5_SET(virtio_q, vq_ctx, virtio_q_type, get_queue_type(ndev));
871 
872 	if (vq_is_tx(mvq->index))
873 		MLX5_SET(virtio_net_q_object, obj_context, tisn_or_qpn, ndev->res.tisn);
874 
875 	MLX5_SET(virtio_q, vq_ctx, event_mode, MLX5_VIRTIO_Q_EVENT_MODE_QP_MODE);
876 	MLX5_SET(virtio_q, vq_ctx, queue_index, mvq->index);
877 	MLX5_SET(virtio_q, vq_ctx, event_qpn_or_msix, mvq->fwqp.mqp.qpn);
878 	MLX5_SET(virtio_q, vq_ctx, queue_size, mvq->num_ent);
879 	MLX5_SET(virtio_q, vq_ctx, virtio_version_1_0,
880 		 !!(ndev->mvdev.actual_features & BIT_ULL(VIRTIO_F_VERSION_1)));
881 	MLX5_SET64(virtio_q, vq_ctx, desc_addr, mvq->desc_addr);
882 	MLX5_SET64(virtio_q, vq_ctx, used_addr, mvq->device_addr);
883 	MLX5_SET64(virtio_q, vq_ctx, available_addr, mvq->driver_addr);
884 	MLX5_SET(virtio_q, vq_ctx, virtio_q_mkey, ndev->mvdev.mr.mkey);
885 	MLX5_SET(virtio_q, vq_ctx, umem_1_id, mvq->umem1.id);
886 	MLX5_SET(virtio_q, vq_ctx, umem_1_size, mvq->umem1.size);
887 	MLX5_SET(virtio_q, vq_ctx, umem_2_id, mvq->umem2.id);
888 	MLX5_SET(virtio_q, vq_ctx, umem_2_size, mvq->umem2.size);
889 	MLX5_SET(virtio_q, vq_ctx, umem_3_id, mvq->umem3.id);
890 	MLX5_SET(virtio_q, vq_ctx, umem_3_size, mvq->umem3.size);
891 	MLX5_SET(virtio_q, vq_ctx, pd, ndev->mvdev.res.pdn);
892 	if (counters_supported(&ndev->mvdev))
893 		MLX5_SET(virtio_q, vq_ctx, counter_set_id, mvq->counter_set_id);
894 
895 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
896 	if (err)
897 		goto err_cmd;
898 
899 	mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT;
900 	kfree(in);
901 	mvq->virtq_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
902 
903 	return 0;
904 
905 err_cmd:
906 	kfree(in);
907 err_alloc:
908 	umems_destroy(ndev, mvq);
909 	return err;
910 }
911 
912 static void destroy_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
913 {
914 	u32 in[MLX5_ST_SZ_DW(destroy_virtio_net_q_in)] = {};
915 	u32 out[MLX5_ST_SZ_DW(destroy_virtio_net_q_out)] = {};
916 
917 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.opcode,
918 		 MLX5_CMD_OP_DESTROY_GENERAL_OBJECT);
919 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.obj_id, mvq->virtq_id);
920 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.uid, ndev->mvdev.res.uid);
921 	MLX5_SET(destroy_virtio_net_q_in, in, general_obj_out_cmd_hdr.obj_type,
922 		 MLX5_OBJ_TYPE_VIRTIO_NET_Q);
923 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out))) {
924 		mlx5_vdpa_warn(&ndev->mvdev, "destroy virtqueue 0x%x\n", mvq->virtq_id);
925 		return;
926 	}
927 	mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_NONE;
928 	umems_destroy(ndev, mvq);
929 }
930 
931 static u32 get_rqpn(struct mlx5_vdpa_virtqueue *mvq, bool fw)
932 {
933 	return fw ? mvq->vqqp.mqp.qpn : mvq->fwqp.mqp.qpn;
934 }
935 
936 static u32 get_qpn(struct mlx5_vdpa_virtqueue *mvq, bool fw)
937 {
938 	return fw ? mvq->fwqp.mqp.qpn : mvq->vqqp.mqp.qpn;
939 }
940 
941 static void alloc_inout(struct mlx5_vdpa_net *ndev, int cmd, void **in, int *inlen, void **out,
942 			int *outlen, u32 qpn, u32 rqpn)
943 {
944 	void *qpc;
945 	void *pp;
946 
947 	switch (cmd) {
948 	case MLX5_CMD_OP_2RST_QP:
949 		*inlen = MLX5_ST_SZ_BYTES(qp_2rst_in);
950 		*outlen = MLX5_ST_SZ_BYTES(qp_2rst_out);
951 		*in = kzalloc(*inlen, GFP_KERNEL);
952 		*out = kzalloc(*outlen, GFP_KERNEL);
953 		if (!*in || !*out)
954 			goto outerr;
955 
956 		MLX5_SET(qp_2rst_in, *in, opcode, cmd);
957 		MLX5_SET(qp_2rst_in, *in, uid, ndev->mvdev.res.uid);
958 		MLX5_SET(qp_2rst_in, *in, qpn, qpn);
959 		break;
960 	case MLX5_CMD_OP_RST2INIT_QP:
961 		*inlen = MLX5_ST_SZ_BYTES(rst2init_qp_in);
962 		*outlen = MLX5_ST_SZ_BYTES(rst2init_qp_out);
963 		*in = kzalloc(*inlen, GFP_KERNEL);
964 		*out = kzalloc(MLX5_ST_SZ_BYTES(rst2init_qp_out), GFP_KERNEL);
965 		if (!*in || !*out)
966 			goto outerr;
967 
968 		MLX5_SET(rst2init_qp_in, *in, opcode, cmd);
969 		MLX5_SET(rst2init_qp_in, *in, uid, ndev->mvdev.res.uid);
970 		MLX5_SET(rst2init_qp_in, *in, qpn, qpn);
971 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
972 		MLX5_SET(qpc, qpc, remote_qpn, rqpn);
973 		MLX5_SET(qpc, qpc, rwe, 1);
974 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
975 		MLX5_SET(ads, pp, vhca_port_num, 1);
976 		break;
977 	case MLX5_CMD_OP_INIT2RTR_QP:
978 		*inlen = MLX5_ST_SZ_BYTES(init2rtr_qp_in);
979 		*outlen = MLX5_ST_SZ_BYTES(init2rtr_qp_out);
980 		*in = kzalloc(*inlen, GFP_KERNEL);
981 		*out = kzalloc(MLX5_ST_SZ_BYTES(init2rtr_qp_out), GFP_KERNEL);
982 		if (!*in || !*out)
983 			goto outerr;
984 
985 		MLX5_SET(init2rtr_qp_in, *in, opcode, cmd);
986 		MLX5_SET(init2rtr_qp_in, *in, uid, ndev->mvdev.res.uid);
987 		MLX5_SET(init2rtr_qp_in, *in, qpn, qpn);
988 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
989 		MLX5_SET(qpc, qpc, mtu, MLX5_QPC_MTU_256_BYTES);
990 		MLX5_SET(qpc, qpc, log_msg_max, 30);
991 		MLX5_SET(qpc, qpc, remote_qpn, rqpn);
992 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
993 		MLX5_SET(ads, pp, fl, 1);
994 		break;
995 	case MLX5_CMD_OP_RTR2RTS_QP:
996 		*inlen = MLX5_ST_SZ_BYTES(rtr2rts_qp_in);
997 		*outlen = MLX5_ST_SZ_BYTES(rtr2rts_qp_out);
998 		*in = kzalloc(*inlen, GFP_KERNEL);
999 		*out = kzalloc(MLX5_ST_SZ_BYTES(rtr2rts_qp_out), GFP_KERNEL);
1000 		if (!*in || !*out)
1001 			goto outerr;
1002 
1003 		MLX5_SET(rtr2rts_qp_in, *in, opcode, cmd);
1004 		MLX5_SET(rtr2rts_qp_in, *in, uid, ndev->mvdev.res.uid);
1005 		MLX5_SET(rtr2rts_qp_in, *in, qpn, qpn);
1006 		qpc = MLX5_ADDR_OF(rst2init_qp_in, *in, qpc);
1007 		pp = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
1008 		MLX5_SET(ads, pp, ack_timeout, 14);
1009 		MLX5_SET(qpc, qpc, retry_count, 7);
1010 		MLX5_SET(qpc, qpc, rnr_retry, 7);
1011 		break;
1012 	default:
1013 		goto outerr_nullify;
1014 	}
1015 
1016 	return;
1017 
1018 outerr:
1019 	kfree(*in);
1020 	kfree(*out);
1021 outerr_nullify:
1022 	*in = NULL;
1023 	*out = NULL;
1024 }
1025 
1026 static void free_inout(void *in, void *out)
1027 {
1028 	kfree(in);
1029 	kfree(out);
1030 }
1031 
1032 /* Two QPs are used by each virtqueue. One is used by the driver and one by
1033  * firmware. The fw argument indicates whether the subjected QP is the one used
1034  * by firmware.
1035  */
1036 static int modify_qp(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, bool fw, int cmd)
1037 {
1038 	int outlen;
1039 	int inlen;
1040 	void *out;
1041 	void *in;
1042 	int err;
1043 
1044 	alloc_inout(ndev, cmd, &in, &inlen, &out, &outlen, get_qpn(mvq, fw), get_rqpn(mvq, fw));
1045 	if (!in || !out)
1046 		return -ENOMEM;
1047 
1048 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, outlen);
1049 	free_inout(in, out);
1050 	return err;
1051 }
1052 
1053 static int connect_qps(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1054 {
1055 	int err;
1056 
1057 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_2RST_QP);
1058 	if (err)
1059 		return err;
1060 
1061 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_2RST_QP);
1062 	if (err)
1063 		return err;
1064 
1065 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_RST2INIT_QP);
1066 	if (err)
1067 		return err;
1068 
1069 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_RST2INIT_QP);
1070 	if (err)
1071 		return err;
1072 
1073 	err = modify_qp(ndev, mvq, true, MLX5_CMD_OP_INIT2RTR_QP);
1074 	if (err)
1075 		return err;
1076 
1077 	err = modify_qp(ndev, mvq, false, MLX5_CMD_OP_INIT2RTR_QP);
1078 	if (err)
1079 		return err;
1080 
1081 	return modify_qp(ndev, mvq, true, MLX5_CMD_OP_RTR2RTS_QP);
1082 }
1083 
1084 struct mlx5_virtq_attr {
1085 	u8 state;
1086 	u16 available_index;
1087 	u16 used_index;
1088 };
1089 
1090 static int query_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
1091 			   struct mlx5_virtq_attr *attr)
1092 {
1093 	int outlen = MLX5_ST_SZ_BYTES(query_virtio_net_q_out);
1094 	u32 in[MLX5_ST_SZ_DW(query_virtio_net_q_in)] = {};
1095 	void *out;
1096 	void *obj_context;
1097 	void *cmd_hdr;
1098 	int err;
1099 
1100 	out = kzalloc(outlen, GFP_KERNEL);
1101 	if (!out)
1102 		return -ENOMEM;
1103 
1104 	cmd_hdr = MLX5_ADDR_OF(query_virtio_net_q_in, in, general_obj_in_cmd_hdr);
1105 
1106 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_QUERY_GENERAL_OBJECT);
1107 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
1108 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->virtq_id);
1109 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1110 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, outlen);
1111 	if (err)
1112 		goto err_cmd;
1113 
1114 	obj_context = MLX5_ADDR_OF(query_virtio_net_q_out, out, obj_context);
1115 	memset(attr, 0, sizeof(*attr));
1116 	attr->state = MLX5_GET(virtio_net_q_object, obj_context, state);
1117 	attr->available_index = MLX5_GET(virtio_net_q_object, obj_context, hw_available_index);
1118 	attr->used_index = MLX5_GET(virtio_net_q_object, obj_context, hw_used_index);
1119 	kfree(out);
1120 	return 0;
1121 
1122 err_cmd:
1123 	kfree(out);
1124 	return err;
1125 }
1126 
1127 static bool is_valid_state_change(int oldstate, int newstate)
1128 {
1129 	switch (oldstate) {
1130 	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT:
1131 		return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY;
1132 	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY:
1133 		return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND;
1134 	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND:
1135 	case MLX5_VIRTIO_NET_Q_OBJECT_STATE_ERR:
1136 	default:
1137 		return false;
1138 	}
1139 }
1140 
1141 static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int state)
1142 {
1143 	int inlen = MLX5_ST_SZ_BYTES(modify_virtio_net_q_in);
1144 	u32 out[MLX5_ST_SZ_DW(modify_virtio_net_q_out)] = {};
1145 	void *obj_context;
1146 	void *cmd_hdr;
1147 	void *in;
1148 	int err;
1149 
1150 	if (mvq->fw_state == MLX5_VIRTIO_NET_Q_OBJECT_NONE)
1151 		return 0;
1152 
1153 	if (!is_valid_state_change(mvq->fw_state, state))
1154 		return -EINVAL;
1155 
1156 	in = kzalloc(inlen, GFP_KERNEL);
1157 	if (!in)
1158 		return -ENOMEM;
1159 
1160 	cmd_hdr = MLX5_ADDR_OF(modify_virtio_net_q_in, in, general_obj_in_cmd_hdr);
1161 
1162 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_MODIFY_GENERAL_OBJECT);
1163 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_NET_Q);
1164 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->virtq_id);
1165 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1166 
1167 	obj_context = MLX5_ADDR_OF(modify_virtio_net_q_in, in, obj_context);
1168 	MLX5_SET64(virtio_net_q_object, obj_context, modify_field_select,
1169 		   MLX5_VIRTQ_MODIFY_MASK_STATE);
1170 	MLX5_SET(virtio_net_q_object, obj_context, state, state);
1171 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, inlen, out, sizeof(out));
1172 	kfree(in);
1173 	if (!err)
1174 		mvq->fw_state = state;
1175 
1176 	return err;
1177 }
1178 
1179 static int counter_set_alloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1180 {
1181 	u32 in[MLX5_ST_SZ_DW(create_virtio_q_counters_in)] = {};
1182 	u32 out[MLX5_ST_SZ_DW(create_virtio_q_counters_out)] = {};
1183 	void *cmd_hdr;
1184 	int err;
1185 
1186 	if (!counters_supported(&ndev->mvdev))
1187 		return 0;
1188 
1189 	cmd_hdr = MLX5_ADDR_OF(create_virtio_q_counters_in, in, hdr);
1190 
1191 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_CREATE_GENERAL_OBJECT);
1192 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
1193 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
1194 
1195 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out));
1196 	if (err)
1197 		return err;
1198 
1199 	mvq->counter_set_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
1200 
1201 	return 0;
1202 }
1203 
1204 static void counter_set_dealloc(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1205 {
1206 	u32 in[MLX5_ST_SZ_DW(destroy_virtio_q_counters_in)] = {};
1207 	u32 out[MLX5_ST_SZ_DW(destroy_virtio_q_counters_out)] = {};
1208 
1209 	if (!counters_supported(&ndev->mvdev))
1210 		return;
1211 
1212 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.opcode, MLX5_CMD_OP_DESTROY_GENERAL_OBJECT);
1213 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.obj_id, mvq->counter_set_id);
1214 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.uid, ndev->mvdev.res.uid);
1215 	MLX5_SET(destroy_virtio_q_counters_in, in, hdr.obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
1216 	if (mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out)))
1217 		mlx5_vdpa_warn(&ndev->mvdev, "dealloc counter set 0x%x\n", mvq->counter_set_id);
1218 }
1219 
1220 static int setup_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1221 {
1222 	u16 idx = mvq->index;
1223 	int err;
1224 
1225 	if (!mvq->num_ent)
1226 		return 0;
1227 
1228 	if (mvq->initialized)
1229 		return 0;
1230 
1231 	err = cq_create(ndev, idx, mvq->num_ent);
1232 	if (err)
1233 		return err;
1234 
1235 	err = qp_create(ndev, mvq, &mvq->fwqp);
1236 	if (err)
1237 		goto err_fwqp;
1238 
1239 	err = qp_create(ndev, mvq, &mvq->vqqp);
1240 	if (err)
1241 		goto err_vqqp;
1242 
1243 	err = connect_qps(ndev, mvq);
1244 	if (err)
1245 		goto err_connect;
1246 
1247 	err = counter_set_alloc(ndev, mvq);
1248 	if (err)
1249 		goto err_counter;
1250 
1251 	err = create_virtqueue(ndev, mvq);
1252 	if (err)
1253 		goto err_connect;
1254 
1255 	if (mvq->ready) {
1256 		err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
1257 		if (err) {
1258 			mlx5_vdpa_warn(&ndev->mvdev, "failed to modify to ready vq idx %d(%d)\n",
1259 				       idx, err);
1260 			goto err_connect;
1261 		}
1262 	}
1263 
1264 	mvq->initialized = true;
1265 	return 0;
1266 
1267 err_connect:
1268 	counter_set_dealloc(ndev, mvq);
1269 err_counter:
1270 	qp_destroy(ndev, &mvq->vqqp);
1271 err_vqqp:
1272 	qp_destroy(ndev, &mvq->fwqp);
1273 err_fwqp:
1274 	cq_destroy(ndev, idx);
1275 	return err;
1276 }
1277 
1278 static void suspend_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1279 {
1280 	struct mlx5_virtq_attr attr;
1281 
1282 	if (!mvq->initialized)
1283 		return;
1284 
1285 	if (mvq->fw_state != MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY)
1286 		return;
1287 
1288 	if (modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND))
1289 		mlx5_vdpa_warn(&ndev->mvdev, "modify to suspend failed\n");
1290 
1291 	if (query_virtqueue(ndev, mvq, &attr)) {
1292 		mlx5_vdpa_warn(&ndev->mvdev, "failed to query virtqueue\n");
1293 		return;
1294 	}
1295 	mvq->avail_idx = attr.available_index;
1296 	mvq->used_idx = attr.used_index;
1297 }
1298 
1299 static void suspend_vqs(struct mlx5_vdpa_net *ndev)
1300 {
1301 	int i;
1302 
1303 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
1304 		suspend_vq(ndev, &ndev->vqs[i]);
1305 }
1306 
1307 static void teardown_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
1308 {
1309 	if (!mvq->initialized)
1310 		return;
1311 
1312 	suspend_vq(ndev, mvq);
1313 	destroy_virtqueue(ndev, mvq);
1314 	counter_set_dealloc(ndev, mvq);
1315 	qp_destroy(ndev, &mvq->vqqp);
1316 	qp_destroy(ndev, &mvq->fwqp);
1317 	cq_destroy(ndev, mvq->index);
1318 	mvq->initialized = false;
1319 }
1320 
1321 static int create_rqt(struct mlx5_vdpa_net *ndev)
1322 {
1323 	int rqt_table_size = roundup_pow_of_two(ndev->rqt_size);
1324 	int act_sz = roundup_pow_of_two(ndev->cur_num_vqs / 2);
1325 	__be32 *list;
1326 	void *rqtc;
1327 	int inlen;
1328 	void *in;
1329 	int i, j;
1330 	int err;
1331 
1332 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + rqt_table_size * MLX5_ST_SZ_BYTES(rq_num);
1333 	in = kzalloc(inlen, GFP_KERNEL);
1334 	if (!in)
1335 		return -ENOMEM;
1336 
1337 	MLX5_SET(create_rqt_in, in, uid, ndev->mvdev.res.uid);
1338 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
1339 
1340 	MLX5_SET(rqtc, rqtc, list_q_type, MLX5_RQTC_LIST_Q_TYPE_VIRTIO_NET_Q);
1341 	MLX5_SET(rqtc, rqtc, rqt_max_size, rqt_table_size);
1342 	list = MLX5_ADDR_OF(rqtc, rqtc, rq_num[0]);
1343 	for (i = 0, j = 0; i < act_sz; i++, j += 2)
1344 		list[i] = cpu_to_be32(ndev->vqs[j % ndev->cur_num_vqs].virtq_id);
1345 
1346 	MLX5_SET(rqtc, rqtc, rqt_actual_size, act_sz);
1347 	err = mlx5_vdpa_create_rqt(&ndev->mvdev, in, inlen, &ndev->res.rqtn);
1348 	kfree(in);
1349 	if (err)
1350 		return err;
1351 
1352 	return 0;
1353 }
1354 
1355 #define MLX5_MODIFY_RQT_NUM_RQS ((u64)1)
1356 
1357 static int modify_rqt(struct mlx5_vdpa_net *ndev, int num)
1358 {
1359 	int act_sz = roundup_pow_of_two(num / 2);
1360 	__be32 *list;
1361 	void *rqtc;
1362 	int inlen;
1363 	void *in;
1364 	int i, j;
1365 	int err;
1366 
1367 	inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + act_sz * MLX5_ST_SZ_BYTES(rq_num);
1368 	in = kzalloc(inlen, GFP_KERNEL);
1369 	if (!in)
1370 		return -ENOMEM;
1371 
1372 	MLX5_SET(modify_rqt_in, in, uid, ndev->mvdev.res.uid);
1373 	MLX5_SET64(modify_rqt_in, in, bitmask, MLX5_MODIFY_RQT_NUM_RQS);
1374 	rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
1375 	MLX5_SET(rqtc, rqtc, list_q_type, MLX5_RQTC_LIST_Q_TYPE_VIRTIO_NET_Q);
1376 
1377 	list = MLX5_ADDR_OF(rqtc, rqtc, rq_num[0]);
1378 	for (i = 0, j = 0; i < act_sz; i++, j = j + 2)
1379 		list[i] = cpu_to_be32(ndev->vqs[j % num].virtq_id);
1380 
1381 	MLX5_SET(rqtc, rqtc, rqt_actual_size, act_sz);
1382 	err = mlx5_vdpa_modify_rqt(&ndev->mvdev, in, inlen, ndev->res.rqtn);
1383 	kfree(in);
1384 	if (err)
1385 		return err;
1386 
1387 	return 0;
1388 }
1389 
1390 static void destroy_rqt(struct mlx5_vdpa_net *ndev)
1391 {
1392 	mlx5_vdpa_destroy_rqt(&ndev->mvdev, ndev->res.rqtn);
1393 }
1394 
1395 static int create_tir(struct mlx5_vdpa_net *ndev)
1396 {
1397 #define HASH_IP_L4PORTS                                                                            \
1398 	(MLX5_HASH_FIELD_SEL_SRC_IP | MLX5_HASH_FIELD_SEL_DST_IP | MLX5_HASH_FIELD_SEL_L4_SPORT |  \
1399 	 MLX5_HASH_FIELD_SEL_L4_DPORT)
1400 	static const u8 rx_hash_toeplitz_key[] = { 0x2c, 0xc6, 0x81, 0xd1, 0x5b, 0xdb, 0xf4, 0xf7,
1401 						   0xfc, 0xa2, 0x83, 0x19, 0xdb, 0x1a, 0x3e, 0x94,
1402 						   0x6b, 0x9e, 0x38, 0xd9, 0x2c, 0x9c, 0x03, 0xd1,
1403 						   0xad, 0x99, 0x44, 0xa7, 0xd9, 0x56, 0x3d, 0x59,
1404 						   0x06, 0x3c, 0x25, 0xf3, 0xfc, 0x1f, 0xdc, 0x2a };
1405 	void *rss_key;
1406 	void *outer;
1407 	void *tirc;
1408 	void *in;
1409 	int err;
1410 
1411 	in = kzalloc(MLX5_ST_SZ_BYTES(create_tir_in), GFP_KERNEL);
1412 	if (!in)
1413 		return -ENOMEM;
1414 
1415 	MLX5_SET(create_tir_in, in, uid, ndev->mvdev.res.uid);
1416 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1417 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
1418 
1419 	MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1420 	MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1421 	rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1422 	memcpy(rss_key, rx_hash_toeplitz_key, sizeof(rx_hash_toeplitz_key));
1423 
1424 	outer = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1425 	MLX5_SET(rx_hash_field_select, outer, l3_prot_type, MLX5_L3_PROT_TYPE_IPV4);
1426 	MLX5_SET(rx_hash_field_select, outer, l4_prot_type, MLX5_L4_PROT_TYPE_TCP);
1427 	MLX5_SET(rx_hash_field_select, outer, selected_fields, HASH_IP_L4PORTS);
1428 
1429 	MLX5_SET(tirc, tirc, indirect_table, ndev->res.rqtn);
1430 	MLX5_SET(tirc, tirc, transport_domain, ndev->res.tdn);
1431 
1432 	err = mlx5_vdpa_create_tir(&ndev->mvdev, in, &ndev->res.tirn);
1433 	kfree(in);
1434 	return err;
1435 }
1436 
1437 static void destroy_tir(struct mlx5_vdpa_net *ndev)
1438 {
1439 	mlx5_vdpa_destroy_tir(&ndev->mvdev, ndev->res.tirn);
1440 }
1441 
1442 #define MAX_STEERING_ENT 0x8000
1443 #define MAX_STEERING_GROUPS 2
1444 
1445 static int mlx5_vdpa_add_mac_vlan_rules(struct mlx5_vdpa_net *ndev, u8 *mac,
1446 					u16 vid, bool tagged,
1447 					struct mlx5_flow_handle **ucast,
1448 					struct mlx5_flow_handle **mcast)
1449 {
1450 	struct mlx5_flow_destination dest = {};
1451 	struct mlx5_flow_act flow_act = {};
1452 	struct mlx5_flow_handle *rule;
1453 	struct mlx5_flow_spec *spec;
1454 	void *headers_c;
1455 	void *headers_v;
1456 	u8 *dmac_c;
1457 	u8 *dmac_v;
1458 	int err;
1459 
1460 	spec = kvzalloc(sizeof(*spec), GFP_KERNEL);
1461 	if (!spec)
1462 		return -ENOMEM;
1463 
1464 	spec->match_criteria_enable = MLX5_MATCH_OUTER_HEADERS;
1465 	headers_c = MLX5_ADDR_OF(fte_match_param, spec->match_criteria, outer_headers);
1466 	headers_v = MLX5_ADDR_OF(fte_match_param, spec->match_value, outer_headers);
1467 	dmac_c = MLX5_ADDR_OF(fte_match_param, headers_c, outer_headers.dmac_47_16);
1468 	dmac_v = MLX5_ADDR_OF(fte_match_param, headers_v, outer_headers.dmac_47_16);
1469 	eth_broadcast_addr(dmac_c);
1470 	ether_addr_copy(dmac_v, mac);
1471 	MLX5_SET(fte_match_set_lyr_2_4, headers_c, cvlan_tag, 1);
1472 	if (tagged) {
1473 		MLX5_SET(fte_match_set_lyr_2_4, headers_v, cvlan_tag, 1);
1474 		MLX5_SET_TO_ONES(fte_match_set_lyr_2_4, headers_c, first_vid);
1475 		MLX5_SET(fte_match_set_lyr_2_4, headers_c, first_vid, vid);
1476 	}
1477 	flow_act.action = MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
1478 	dest.type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1479 	dest.tir_num = ndev->res.tirn;
1480 	rule = mlx5_add_flow_rules(ndev->rxft, spec, &flow_act, &dest, 1);
1481 	if (IS_ERR(rule))
1482 		return PTR_ERR(rule);
1483 
1484 	*ucast = rule;
1485 
1486 	memset(dmac_c, 0, ETH_ALEN);
1487 	memset(dmac_v, 0, ETH_ALEN);
1488 	dmac_c[0] = 1;
1489 	dmac_v[0] = 1;
1490 	rule = mlx5_add_flow_rules(ndev->rxft, spec, &flow_act, &dest, 1);
1491 	kvfree(spec);
1492 	if (IS_ERR(rule)) {
1493 		err = PTR_ERR(rule);
1494 		goto err_mcast;
1495 	}
1496 
1497 	*mcast = rule;
1498 	return 0;
1499 
1500 err_mcast:
1501 	mlx5_del_flow_rules(*ucast);
1502 	return err;
1503 }
1504 
1505 static void mlx5_vdpa_del_mac_vlan_rules(struct mlx5_vdpa_net *ndev,
1506 					 struct mlx5_flow_handle *ucast,
1507 					 struct mlx5_flow_handle *mcast)
1508 {
1509 	mlx5_del_flow_rules(ucast);
1510 	mlx5_del_flow_rules(mcast);
1511 }
1512 
1513 static u64 search_val(u8 *mac, u16 vlan, bool tagged)
1514 {
1515 	u64 val;
1516 
1517 	if (!tagged)
1518 		vlan = MLX5V_UNTAGGED;
1519 
1520 	val = (u64)vlan << 48 |
1521 	      (u64)mac[0] << 40 |
1522 	      (u64)mac[1] << 32 |
1523 	      (u64)mac[2] << 24 |
1524 	      (u64)mac[3] << 16 |
1525 	      (u64)mac[4] << 8 |
1526 	      (u64)mac[5];
1527 
1528 	return val;
1529 }
1530 
1531 static struct macvlan_node *mac_vlan_lookup(struct mlx5_vdpa_net *ndev, u64 value)
1532 {
1533 	struct macvlan_node *pos;
1534 	u32 idx;
1535 
1536 	idx = hash_64(value, 8); // tbd 8
1537 	hlist_for_each_entry(pos, &ndev->macvlan_hash[idx], hlist) {
1538 		if (pos->macvlan == value)
1539 			return pos;
1540 	}
1541 	return NULL;
1542 }
1543 
1544 static int mac_vlan_add(struct mlx5_vdpa_net *ndev, u8 *mac, u16 vlan, bool tagged) // vlan -> vid
1545 {
1546 	struct macvlan_node *ptr;
1547 	u64 val;
1548 	u32 idx;
1549 	int err;
1550 
1551 	val = search_val(mac, vlan, tagged);
1552 	if (mac_vlan_lookup(ndev, val))
1553 		return -EEXIST;
1554 
1555 	ptr = kzalloc(sizeof(*ptr), GFP_KERNEL);
1556 	if (!ptr)
1557 		return -ENOMEM;
1558 
1559 	err = mlx5_vdpa_add_mac_vlan_rules(ndev, ndev->config.mac, vlan, tagged,
1560 					   &ptr->ucast_rule, &ptr->mcast_rule);
1561 	if (err)
1562 		goto err_add;
1563 
1564 	ptr->macvlan = val;
1565 	idx = hash_64(val, 8);
1566 	hlist_add_head(&ptr->hlist, &ndev->macvlan_hash[idx]);
1567 	return 0;
1568 
1569 err_add:
1570 	kfree(ptr);
1571 	return err;
1572 }
1573 
1574 static void mac_vlan_del(struct mlx5_vdpa_net *ndev, u8 *mac, u16 vlan, bool tagged)
1575 {
1576 	struct macvlan_node *ptr;
1577 
1578 	ptr = mac_vlan_lookup(ndev, search_val(mac, vlan, tagged));
1579 	if (!ptr)
1580 		return;
1581 
1582 	hlist_del(&ptr->hlist);
1583 	mlx5_vdpa_del_mac_vlan_rules(ndev, ptr->ucast_rule, ptr->mcast_rule);
1584 	kfree(ptr);
1585 }
1586 
1587 static void clear_mac_vlan_table(struct mlx5_vdpa_net *ndev)
1588 {
1589 	struct macvlan_node *pos;
1590 	struct hlist_node *n;
1591 	int i;
1592 
1593 	for (i = 0; i < MLX5V_MACVLAN_SIZE; i++) {
1594 		hlist_for_each_entry_safe(pos, n, &ndev->macvlan_hash[i], hlist) {
1595 			hlist_del(&pos->hlist);
1596 			mlx5_vdpa_del_mac_vlan_rules(ndev, pos->ucast_rule, pos->mcast_rule);
1597 			kfree(pos);
1598 		}
1599 	}
1600 }
1601 
1602 static int setup_steering(struct mlx5_vdpa_net *ndev)
1603 {
1604 	struct mlx5_flow_table_attr ft_attr = {};
1605 	struct mlx5_flow_namespace *ns;
1606 	int err;
1607 
1608 	ft_attr.max_fte = MAX_STEERING_ENT;
1609 	ft_attr.autogroup.max_num_groups = MAX_STEERING_GROUPS;
1610 
1611 	ns = mlx5_get_flow_namespace(ndev->mvdev.mdev, MLX5_FLOW_NAMESPACE_BYPASS);
1612 	if (!ns) {
1613 		mlx5_vdpa_warn(&ndev->mvdev, "failed to get flow namespace\n");
1614 		return -EOPNOTSUPP;
1615 	}
1616 
1617 	ndev->rxft = mlx5_create_auto_grouped_flow_table(ns, &ft_attr);
1618 	if (IS_ERR(ndev->rxft)) {
1619 		mlx5_vdpa_warn(&ndev->mvdev, "failed to create flow table\n");
1620 		return PTR_ERR(ndev->rxft);
1621 	}
1622 
1623 	err = mac_vlan_add(ndev, ndev->config.mac, 0, false);
1624 	if (err)
1625 		goto err_add;
1626 
1627 	return 0;
1628 
1629 err_add:
1630 	mlx5_destroy_flow_table(ndev->rxft);
1631 	return err;
1632 }
1633 
1634 static void teardown_steering(struct mlx5_vdpa_net *ndev)
1635 {
1636 	clear_mac_vlan_table(ndev);
1637 	mlx5_destroy_flow_table(ndev->rxft);
1638 }
1639 
1640 static virtio_net_ctrl_ack handle_ctrl_mac(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1641 {
1642 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1643 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1644 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1645 	struct mlx5_core_dev *pfmdev;
1646 	size_t read;
1647 	u8 mac[ETH_ALEN], mac_back[ETH_ALEN];
1648 
1649 	pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
1650 	switch (cmd) {
1651 	case VIRTIO_NET_CTRL_MAC_ADDR_SET:
1652 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, (void *)mac, ETH_ALEN);
1653 		if (read != ETH_ALEN)
1654 			break;
1655 
1656 		if (!memcmp(ndev->config.mac, mac, 6)) {
1657 			status = VIRTIO_NET_OK;
1658 			break;
1659 		}
1660 
1661 		if (is_zero_ether_addr(mac))
1662 			break;
1663 
1664 		if (!is_zero_ether_addr(ndev->config.mac)) {
1665 			if (mlx5_mpfs_del_mac(pfmdev, ndev->config.mac)) {
1666 				mlx5_vdpa_warn(mvdev, "failed to delete old MAC %pM from MPFS table\n",
1667 					       ndev->config.mac);
1668 				break;
1669 			}
1670 		}
1671 
1672 		if (mlx5_mpfs_add_mac(pfmdev, mac)) {
1673 			mlx5_vdpa_warn(mvdev, "failed to insert new MAC %pM into MPFS table\n",
1674 				       mac);
1675 			break;
1676 		}
1677 
1678 		/* backup the original mac address so that if failed to add the forward rules
1679 		 * we could restore it
1680 		 */
1681 		memcpy(mac_back, ndev->config.mac, ETH_ALEN);
1682 
1683 		memcpy(ndev->config.mac, mac, ETH_ALEN);
1684 
1685 		/* Need recreate the flow table entry, so that the packet could forward back
1686 		 */
1687 		mac_vlan_del(ndev, ndev->config.mac, 0, false);
1688 
1689 		if (mac_vlan_add(ndev, ndev->config.mac, 0, false)) {
1690 			mlx5_vdpa_warn(mvdev, "failed to insert forward rules, try to restore\n");
1691 
1692 			/* Although it hardly run here, we still need double check */
1693 			if (is_zero_ether_addr(mac_back)) {
1694 				mlx5_vdpa_warn(mvdev, "restore mac failed: Original MAC is zero\n");
1695 				break;
1696 			}
1697 
1698 			/* Try to restore original mac address to MFPS table, and try to restore
1699 			 * the forward rule entry.
1700 			 */
1701 			if (mlx5_mpfs_del_mac(pfmdev, ndev->config.mac)) {
1702 				mlx5_vdpa_warn(mvdev, "restore mac failed: delete MAC %pM from MPFS table failed\n",
1703 					       ndev->config.mac);
1704 			}
1705 
1706 			if (mlx5_mpfs_add_mac(pfmdev, mac_back)) {
1707 				mlx5_vdpa_warn(mvdev, "restore mac failed: insert old MAC %pM into MPFS table failed\n",
1708 					       mac_back);
1709 			}
1710 
1711 			memcpy(ndev->config.mac, mac_back, ETH_ALEN);
1712 
1713 			if (mac_vlan_add(ndev, ndev->config.mac, 0, false))
1714 				mlx5_vdpa_warn(mvdev, "restore forward rules failed: insert forward rules failed\n");
1715 
1716 			break;
1717 		}
1718 
1719 		status = VIRTIO_NET_OK;
1720 		break;
1721 
1722 	default:
1723 		break;
1724 	}
1725 
1726 	return status;
1727 }
1728 
1729 static int change_num_qps(struct mlx5_vdpa_dev *mvdev, int newqps)
1730 {
1731 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1732 	int cur_qps = ndev->cur_num_vqs / 2;
1733 	int err;
1734 	int i;
1735 
1736 	if (cur_qps > newqps) {
1737 		err = modify_rqt(ndev, 2 * newqps);
1738 		if (err)
1739 			return err;
1740 
1741 		for (i = ndev->cur_num_vqs - 1; i >= 2 * newqps; i--)
1742 			teardown_vq(ndev, &ndev->vqs[i]);
1743 
1744 		ndev->cur_num_vqs = 2 * newqps;
1745 	} else {
1746 		ndev->cur_num_vqs = 2 * newqps;
1747 		for (i = cur_qps * 2; i < 2 * newqps; i++) {
1748 			err = setup_vq(ndev, &ndev->vqs[i]);
1749 			if (err)
1750 				goto clean_added;
1751 		}
1752 		err = modify_rqt(ndev, 2 * newqps);
1753 		if (err)
1754 			goto clean_added;
1755 	}
1756 	return 0;
1757 
1758 clean_added:
1759 	for (--i; i >= 2 * cur_qps; --i)
1760 		teardown_vq(ndev, &ndev->vqs[i]);
1761 
1762 	ndev->cur_num_vqs = 2 * cur_qps;
1763 
1764 	return err;
1765 }
1766 
1767 static virtio_net_ctrl_ack handle_ctrl_mq(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1768 {
1769 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1770 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1771 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1772 	struct virtio_net_ctrl_mq mq;
1773 	size_t read;
1774 	u16 newqps;
1775 
1776 	switch (cmd) {
1777 	case VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET:
1778 		/* This mq feature check aligns with pre-existing userspace
1779 		 * implementation.
1780 		 *
1781 		 * Without it, an untrusted driver could fake a multiqueue config
1782 		 * request down to a non-mq device that may cause kernel to
1783 		 * panic due to uninitialized resources for extra vqs. Even with
1784 		 * a well behaving guest driver, it is not expected to allow
1785 		 * changing the number of vqs on a non-mq device.
1786 		 */
1787 		if (!MLX5_FEATURE(mvdev, VIRTIO_NET_F_MQ))
1788 			break;
1789 
1790 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, (void *)&mq, sizeof(mq));
1791 		if (read != sizeof(mq))
1792 			break;
1793 
1794 		newqps = mlx5vdpa16_to_cpu(mvdev, mq.virtqueue_pairs);
1795 		if (newqps < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1796 		    newqps > ndev->rqt_size)
1797 			break;
1798 
1799 		if (ndev->cur_num_vqs == 2 * newqps) {
1800 			status = VIRTIO_NET_OK;
1801 			break;
1802 		}
1803 
1804 		if (!change_num_qps(mvdev, newqps))
1805 			status = VIRTIO_NET_OK;
1806 
1807 		break;
1808 	default:
1809 		break;
1810 	}
1811 
1812 	return status;
1813 }
1814 
1815 static virtio_net_ctrl_ack handle_ctrl_vlan(struct mlx5_vdpa_dev *mvdev, u8 cmd)
1816 {
1817 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1818 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1819 	struct mlx5_control_vq *cvq = &mvdev->cvq;
1820 	__virtio16 vlan;
1821 	size_t read;
1822 	u16 id;
1823 
1824 	switch (cmd) {
1825 	case VIRTIO_NET_CTRL_VLAN_ADD:
1826 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &vlan, sizeof(vlan));
1827 		if (read != sizeof(vlan))
1828 			break;
1829 
1830 		id = mlx5vdpa16_to_cpu(mvdev, vlan);
1831 		if (mac_vlan_add(ndev, ndev->config.mac, id, true))
1832 			break;
1833 
1834 		status = VIRTIO_NET_OK;
1835 		break;
1836 	case VIRTIO_NET_CTRL_VLAN_DEL:
1837 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &vlan, sizeof(vlan));
1838 		if (read != sizeof(vlan))
1839 			break;
1840 
1841 		id = mlx5vdpa16_to_cpu(mvdev, vlan);
1842 		mac_vlan_del(ndev, ndev->config.mac, id, true);
1843 		status = VIRTIO_NET_OK;
1844 		break;
1845 	default:
1846 		break;
1847 	}
1848 
1849 	return status;
1850 }
1851 
1852 static void mlx5_cvq_kick_handler(struct work_struct *work)
1853 {
1854 	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1855 	struct virtio_net_ctrl_hdr ctrl;
1856 	struct mlx5_vdpa_wq_ent *wqent;
1857 	struct mlx5_vdpa_dev *mvdev;
1858 	struct mlx5_control_vq *cvq;
1859 	struct mlx5_vdpa_net *ndev;
1860 	size_t read, write;
1861 	int err;
1862 
1863 	wqent = container_of(work, struct mlx5_vdpa_wq_ent, work);
1864 	mvdev = wqent->mvdev;
1865 	ndev = to_mlx5_vdpa_ndev(mvdev);
1866 	cvq = &mvdev->cvq;
1867 
1868 	down_write(&ndev->reslock);
1869 
1870 	if (!(mvdev->status & VIRTIO_CONFIG_S_DRIVER_OK))
1871 		goto out;
1872 
1873 	if (!(ndev->mvdev.actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
1874 		goto out;
1875 
1876 	if (!cvq->ready)
1877 		goto out;
1878 
1879 	while (true) {
1880 		err = vringh_getdesc_iotlb(&cvq->vring, &cvq->riov, &cvq->wiov, &cvq->head,
1881 					   GFP_ATOMIC);
1882 		if (err <= 0)
1883 			break;
1884 
1885 		read = vringh_iov_pull_iotlb(&cvq->vring, &cvq->riov, &ctrl, sizeof(ctrl));
1886 		if (read != sizeof(ctrl))
1887 			break;
1888 
1889 		cvq->received_desc++;
1890 		switch (ctrl.class) {
1891 		case VIRTIO_NET_CTRL_MAC:
1892 			status = handle_ctrl_mac(mvdev, ctrl.cmd);
1893 			break;
1894 		case VIRTIO_NET_CTRL_MQ:
1895 			status = handle_ctrl_mq(mvdev, ctrl.cmd);
1896 			break;
1897 		case VIRTIO_NET_CTRL_VLAN:
1898 			status = handle_ctrl_vlan(mvdev, ctrl.cmd);
1899 			break;
1900 		default:
1901 			break;
1902 		}
1903 
1904 		/* Make sure data is written before advancing index */
1905 		smp_wmb();
1906 
1907 		write = vringh_iov_push_iotlb(&cvq->vring, &cvq->wiov, &status, sizeof(status));
1908 		vringh_complete_iotlb(&cvq->vring, cvq->head, write);
1909 		vringh_kiov_cleanup(&cvq->riov);
1910 		vringh_kiov_cleanup(&cvq->wiov);
1911 
1912 		if (vringh_need_notify_iotlb(&cvq->vring))
1913 			vringh_notify(&cvq->vring);
1914 
1915 		cvq->completed_desc++;
1916 		queue_work(mvdev->wq, &wqent->work);
1917 		break;
1918 	}
1919 
1920 out:
1921 	up_write(&ndev->reslock);
1922 }
1923 
1924 static void mlx5_vdpa_kick_vq(struct vdpa_device *vdev, u16 idx)
1925 {
1926 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1927 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1928 	struct mlx5_vdpa_virtqueue *mvq;
1929 
1930 	if (!is_index_valid(mvdev, idx))
1931 		return;
1932 
1933 	if (unlikely(is_ctrl_vq_idx(mvdev, idx))) {
1934 		if (!mvdev->wq || !mvdev->cvq.ready)
1935 			return;
1936 
1937 		queue_work(mvdev->wq, &ndev->cvq_ent.work);
1938 		return;
1939 	}
1940 
1941 	mvq = &ndev->vqs[idx];
1942 	if (unlikely(!mvq->ready))
1943 		return;
1944 
1945 	iowrite16(idx, ndev->mvdev.res.kick_addr);
1946 }
1947 
1948 static int mlx5_vdpa_set_vq_address(struct vdpa_device *vdev, u16 idx, u64 desc_area,
1949 				    u64 driver_area, u64 device_area)
1950 {
1951 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1952 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1953 	struct mlx5_vdpa_virtqueue *mvq;
1954 
1955 	if (!is_index_valid(mvdev, idx))
1956 		return -EINVAL;
1957 
1958 	if (is_ctrl_vq_idx(mvdev, idx)) {
1959 		mvdev->cvq.desc_addr = desc_area;
1960 		mvdev->cvq.device_addr = device_area;
1961 		mvdev->cvq.driver_addr = driver_area;
1962 		return 0;
1963 	}
1964 
1965 	mvq = &ndev->vqs[idx];
1966 	mvq->desc_addr = desc_area;
1967 	mvq->device_addr = device_area;
1968 	mvq->driver_addr = driver_area;
1969 	return 0;
1970 }
1971 
1972 static void mlx5_vdpa_set_vq_num(struct vdpa_device *vdev, u16 idx, u32 num)
1973 {
1974 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1975 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1976 	struct mlx5_vdpa_virtqueue *mvq;
1977 
1978 	if (!is_index_valid(mvdev, idx) || is_ctrl_vq_idx(mvdev, idx))
1979 		return;
1980 
1981 	mvq = &ndev->vqs[idx];
1982 	mvq->num_ent = num;
1983 }
1984 
1985 static void mlx5_vdpa_set_vq_cb(struct vdpa_device *vdev, u16 idx, struct vdpa_callback *cb)
1986 {
1987 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
1988 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
1989 
1990 	ndev->event_cbs[idx] = *cb;
1991 	if (is_ctrl_vq_idx(mvdev, idx))
1992 		mvdev->cvq.event_cb = *cb;
1993 }
1994 
1995 static void mlx5_cvq_notify(struct vringh *vring)
1996 {
1997 	struct mlx5_control_vq *cvq = container_of(vring, struct mlx5_control_vq, vring);
1998 
1999 	if (!cvq->event_cb.callback)
2000 		return;
2001 
2002 	cvq->event_cb.callback(cvq->event_cb.private);
2003 }
2004 
2005 static void set_cvq_ready(struct mlx5_vdpa_dev *mvdev, bool ready)
2006 {
2007 	struct mlx5_control_vq *cvq = &mvdev->cvq;
2008 
2009 	cvq->ready = ready;
2010 	if (!ready)
2011 		return;
2012 
2013 	cvq->vring.notify = mlx5_cvq_notify;
2014 }
2015 
2016 static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready)
2017 {
2018 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2019 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2020 	struct mlx5_vdpa_virtqueue *mvq;
2021 	int err;
2022 
2023 	if (!mvdev->actual_features)
2024 		return;
2025 
2026 	if (!is_index_valid(mvdev, idx))
2027 		return;
2028 
2029 	if (is_ctrl_vq_idx(mvdev, idx)) {
2030 		set_cvq_ready(mvdev, ready);
2031 		return;
2032 	}
2033 
2034 	mvq = &ndev->vqs[idx];
2035 	if (!ready) {
2036 		suspend_vq(ndev, mvq);
2037 	} else {
2038 		err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
2039 		if (err) {
2040 			mlx5_vdpa_warn(mvdev, "modify VQ %d to ready failed (%d)\n", idx, err);
2041 			ready = false;
2042 		}
2043 	}
2044 
2045 
2046 	mvq->ready = ready;
2047 }
2048 
2049 static bool mlx5_vdpa_get_vq_ready(struct vdpa_device *vdev, u16 idx)
2050 {
2051 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2052 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2053 
2054 	if (!is_index_valid(mvdev, idx))
2055 		return false;
2056 
2057 	if (is_ctrl_vq_idx(mvdev, idx))
2058 		return mvdev->cvq.ready;
2059 
2060 	return ndev->vqs[idx].ready;
2061 }
2062 
2063 static int mlx5_vdpa_set_vq_state(struct vdpa_device *vdev, u16 idx,
2064 				  const struct vdpa_vq_state *state)
2065 {
2066 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2067 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2068 	struct mlx5_vdpa_virtqueue *mvq;
2069 
2070 	if (!is_index_valid(mvdev, idx))
2071 		return -EINVAL;
2072 
2073 	if (is_ctrl_vq_idx(mvdev, idx)) {
2074 		mvdev->cvq.vring.last_avail_idx = state->split.avail_index;
2075 		return 0;
2076 	}
2077 
2078 	mvq = &ndev->vqs[idx];
2079 	if (mvq->fw_state == MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY) {
2080 		mlx5_vdpa_warn(mvdev, "can't modify available index\n");
2081 		return -EINVAL;
2082 	}
2083 
2084 	mvq->used_idx = state->split.avail_index;
2085 	mvq->avail_idx = state->split.avail_index;
2086 	return 0;
2087 }
2088 
2089 static int mlx5_vdpa_get_vq_state(struct vdpa_device *vdev, u16 idx, struct vdpa_vq_state *state)
2090 {
2091 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2092 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2093 	struct mlx5_vdpa_virtqueue *mvq;
2094 	struct mlx5_virtq_attr attr;
2095 	int err;
2096 
2097 	if (!is_index_valid(mvdev, idx))
2098 		return -EINVAL;
2099 
2100 	if (is_ctrl_vq_idx(mvdev, idx)) {
2101 		state->split.avail_index = mvdev->cvq.vring.last_avail_idx;
2102 		return 0;
2103 	}
2104 
2105 	mvq = &ndev->vqs[idx];
2106 	/* If the virtq object was destroyed, use the value saved at
2107 	 * the last minute of suspend_vq. This caters for userspace
2108 	 * that cares about emulating the index after vq is stopped.
2109 	 */
2110 	if (!mvq->initialized) {
2111 		/* Firmware returns a wrong value for the available index.
2112 		 * Since both values should be identical, we take the value of
2113 		 * used_idx which is reported correctly.
2114 		 */
2115 		state->split.avail_index = mvq->used_idx;
2116 		return 0;
2117 	}
2118 
2119 	err = query_virtqueue(ndev, mvq, &attr);
2120 	if (err) {
2121 		mlx5_vdpa_warn(mvdev, "failed to query virtqueue\n");
2122 		return err;
2123 	}
2124 	state->split.avail_index = attr.used_index;
2125 	return 0;
2126 }
2127 
2128 static u32 mlx5_vdpa_get_vq_align(struct vdpa_device *vdev)
2129 {
2130 	return PAGE_SIZE;
2131 }
2132 
2133 static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdev, u16 idx)
2134 {
2135 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2136 
2137 	if (is_ctrl_vq_idx(mvdev, idx))
2138 		return MLX5_VDPA_CVQ_GROUP;
2139 
2140 	return MLX5_VDPA_DATAVQ_GROUP;
2141 }
2142 
2143 enum { MLX5_VIRTIO_NET_F_GUEST_CSUM = 1 << 9,
2144 	MLX5_VIRTIO_NET_F_CSUM = 1 << 10,
2145 	MLX5_VIRTIO_NET_F_HOST_TSO6 = 1 << 11,
2146 	MLX5_VIRTIO_NET_F_HOST_TSO4 = 1 << 12,
2147 };
2148 
2149 static u64 mlx_to_vritio_features(u16 dev_features)
2150 {
2151 	u64 result = 0;
2152 
2153 	if (dev_features & MLX5_VIRTIO_NET_F_GUEST_CSUM)
2154 		result |= BIT_ULL(VIRTIO_NET_F_GUEST_CSUM);
2155 	if (dev_features & MLX5_VIRTIO_NET_F_CSUM)
2156 		result |= BIT_ULL(VIRTIO_NET_F_CSUM);
2157 	if (dev_features & MLX5_VIRTIO_NET_F_HOST_TSO6)
2158 		result |= BIT_ULL(VIRTIO_NET_F_HOST_TSO6);
2159 	if (dev_features & MLX5_VIRTIO_NET_F_HOST_TSO4)
2160 		result |= BIT_ULL(VIRTIO_NET_F_HOST_TSO4);
2161 
2162 	return result;
2163 }
2164 
2165 static u64 get_supported_features(struct mlx5_core_dev *mdev)
2166 {
2167 	u64 mlx_vdpa_features = 0;
2168 	u16 dev_features;
2169 
2170 	dev_features = MLX5_CAP_DEV_VDPA_EMULATION(mdev, device_features_bits_mask);
2171 	mlx_vdpa_features |= mlx_to_vritio_features(dev_features);
2172 	if (MLX5_CAP_DEV_VDPA_EMULATION(mdev, virtio_version_1_0))
2173 		mlx_vdpa_features |= BIT_ULL(VIRTIO_F_VERSION_1);
2174 	mlx_vdpa_features |= BIT_ULL(VIRTIO_F_ACCESS_PLATFORM);
2175 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_VQ);
2176 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR);
2177 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MQ);
2178 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_STATUS);
2179 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_MTU);
2180 	mlx_vdpa_features |= BIT_ULL(VIRTIO_NET_F_CTRL_VLAN);
2181 
2182 	return mlx_vdpa_features;
2183 }
2184 
2185 static u64 mlx5_vdpa_get_device_features(struct vdpa_device *vdev)
2186 {
2187 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2188 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2189 
2190 	print_features(mvdev, ndev->mvdev.mlx_features, false);
2191 	return ndev->mvdev.mlx_features;
2192 }
2193 
2194 static int verify_driver_features(struct mlx5_vdpa_dev *mvdev, u64 features)
2195 {
2196 	/* Minimum features to expect */
2197 	if (!(features & BIT_ULL(VIRTIO_F_ACCESS_PLATFORM)))
2198 		return -EOPNOTSUPP;
2199 
2200 	/* Double check features combination sent down by the driver.
2201 	 * Fail invalid features due to absence of the depended feature.
2202 	 *
2203 	 * Per VIRTIO v1.1 specification, section 5.1.3.1 Feature bit
2204 	 * requirements: "VIRTIO_NET_F_MQ Requires VIRTIO_NET_F_CTRL_VQ".
2205 	 * By failing the invalid features sent down by untrusted drivers,
2206 	 * we're assured the assumption made upon is_index_valid() and
2207 	 * is_ctrl_vq_idx() will not be compromised.
2208 	 */
2209 	if ((features & (BIT_ULL(VIRTIO_NET_F_MQ) | BIT_ULL(VIRTIO_NET_F_CTRL_VQ))) ==
2210             BIT_ULL(VIRTIO_NET_F_MQ))
2211 		return -EINVAL;
2212 
2213 	return 0;
2214 }
2215 
2216 static int setup_virtqueues(struct mlx5_vdpa_dev *mvdev)
2217 {
2218 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2219 	int err;
2220 	int i;
2221 
2222 	for (i = 0; i < mvdev->max_vqs; i++) {
2223 		err = setup_vq(ndev, &ndev->vqs[i]);
2224 		if (err)
2225 			goto err_vq;
2226 	}
2227 
2228 	return 0;
2229 
2230 err_vq:
2231 	for (--i; i >= 0; i--)
2232 		teardown_vq(ndev, &ndev->vqs[i]);
2233 
2234 	return err;
2235 }
2236 
2237 static void teardown_virtqueues(struct mlx5_vdpa_net *ndev)
2238 {
2239 	struct mlx5_vdpa_virtqueue *mvq;
2240 	int i;
2241 
2242 	for (i = ndev->mvdev.max_vqs - 1; i >= 0; i--) {
2243 		mvq = &ndev->vqs[i];
2244 		if (!mvq->initialized)
2245 			continue;
2246 
2247 		teardown_vq(ndev, mvq);
2248 	}
2249 }
2250 
2251 static void update_cvq_info(struct mlx5_vdpa_dev *mvdev)
2252 {
2253 	if (MLX5_FEATURE(mvdev, VIRTIO_NET_F_CTRL_VQ)) {
2254 		if (MLX5_FEATURE(mvdev, VIRTIO_NET_F_MQ)) {
2255 			/* MQ supported. CVQ index is right above the last data virtqueue's */
2256 			mvdev->max_idx = mvdev->max_vqs;
2257 		} else {
2258 			/* Only CVQ supportted. data virtqueues occupy indices 0 and 1.
2259 			 * CVQ gets index 2
2260 			 */
2261 			mvdev->max_idx = 2;
2262 		}
2263 	} else {
2264 		/* Two data virtqueues only: one for rx and one for tx */
2265 		mvdev->max_idx = 1;
2266 	}
2267 }
2268 
2269 static int mlx5_vdpa_set_driver_features(struct vdpa_device *vdev, u64 features)
2270 {
2271 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2272 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2273 	int err;
2274 
2275 	print_features(mvdev, features, true);
2276 
2277 	err = verify_driver_features(mvdev, features);
2278 	if (err)
2279 		return err;
2280 
2281 	ndev->mvdev.actual_features = features & ndev->mvdev.mlx_features;
2282 	if (ndev->mvdev.actual_features & BIT_ULL(VIRTIO_NET_F_MQ))
2283 		ndev->rqt_size = mlx5vdpa16_to_cpu(mvdev, ndev->config.max_virtqueue_pairs);
2284 	else
2285 		ndev->rqt_size = 1;
2286 
2287 	ndev->cur_num_vqs = 2 * ndev->rqt_size;
2288 
2289 	update_cvq_info(mvdev);
2290 	return err;
2291 }
2292 
2293 static void mlx5_vdpa_set_config_cb(struct vdpa_device *vdev, struct vdpa_callback *cb)
2294 {
2295 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2296 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2297 
2298 	ndev->config_cb = *cb;
2299 }
2300 
2301 #define MLX5_VDPA_MAX_VQ_ENTRIES 256
2302 static u16 mlx5_vdpa_get_vq_num_max(struct vdpa_device *vdev)
2303 {
2304 	return MLX5_VDPA_MAX_VQ_ENTRIES;
2305 }
2306 
2307 static u32 mlx5_vdpa_get_device_id(struct vdpa_device *vdev)
2308 {
2309 	return VIRTIO_ID_NET;
2310 }
2311 
2312 static u32 mlx5_vdpa_get_vendor_id(struct vdpa_device *vdev)
2313 {
2314 	return PCI_VENDOR_ID_MELLANOX;
2315 }
2316 
2317 static u8 mlx5_vdpa_get_status(struct vdpa_device *vdev)
2318 {
2319 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2320 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2321 
2322 	print_status(mvdev, ndev->mvdev.status, false);
2323 	return ndev->mvdev.status;
2324 }
2325 
2326 static int save_channel_info(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
2327 {
2328 	struct mlx5_vq_restore_info *ri = &mvq->ri;
2329 	struct mlx5_virtq_attr attr = {};
2330 	int err;
2331 
2332 	if (mvq->initialized) {
2333 		err = query_virtqueue(ndev, mvq, &attr);
2334 		if (err)
2335 			return err;
2336 	}
2337 
2338 	ri->avail_index = attr.available_index;
2339 	ri->used_index = attr.used_index;
2340 	ri->ready = mvq->ready;
2341 	ri->num_ent = mvq->num_ent;
2342 	ri->desc_addr = mvq->desc_addr;
2343 	ri->device_addr = mvq->device_addr;
2344 	ri->driver_addr = mvq->driver_addr;
2345 	ri->restore = true;
2346 	return 0;
2347 }
2348 
2349 static int save_channels_info(struct mlx5_vdpa_net *ndev)
2350 {
2351 	int i;
2352 
2353 	for (i = 0; i < ndev->mvdev.max_vqs; i++) {
2354 		memset(&ndev->vqs[i].ri, 0, sizeof(ndev->vqs[i].ri));
2355 		save_channel_info(ndev, &ndev->vqs[i]);
2356 	}
2357 	return 0;
2358 }
2359 
2360 static void mlx5_clear_vqs(struct mlx5_vdpa_net *ndev)
2361 {
2362 	int i;
2363 
2364 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
2365 		memset(&ndev->vqs[i], 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2366 }
2367 
2368 static void restore_channels_info(struct mlx5_vdpa_net *ndev)
2369 {
2370 	struct mlx5_vdpa_virtqueue *mvq;
2371 	struct mlx5_vq_restore_info *ri;
2372 	int i;
2373 
2374 	mlx5_clear_vqs(ndev);
2375 	init_mvqs(ndev);
2376 	for (i = 0; i < ndev->mvdev.max_vqs; i++) {
2377 		mvq = &ndev->vqs[i];
2378 		ri = &mvq->ri;
2379 		if (!ri->restore)
2380 			continue;
2381 
2382 		mvq->avail_idx = ri->avail_index;
2383 		mvq->used_idx = ri->used_index;
2384 		mvq->ready = ri->ready;
2385 		mvq->num_ent = ri->num_ent;
2386 		mvq->desc_addr = ri->desc_addr;
2387 		mvq->device_addr = ri->device_addr;
2388 		mvq->driver_addr = ri->driver_addr;
2389 	}
2390 }
2391 
2392 static int mlx5_vdpa_change_map(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
2393 {
2394 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2395 	int err;
2396 
2397 	suspend_vqs(ndev);
2398 	err = save_channels_info(ndev);
2399 	if (err)
2400 		goto err_mr;
2401 
2402 	teardown_driver(ndev);
2403 	mlx5_vdpa_destroy_mr(mvdev);
2404 	err = mlx5_vdpa_create_mr(mvdev, iotlb);
2405 	if (err)
2406 		goto err_mr;
2407 
2408 	if (!(mvdev->status & VIRTIO_CONFIG_S_DRIVER_OK))
2409 		goto err_mr;
2410 
2411 	restore_channels_info(ndev);
2412 	err = setup_driver(mvdev);
2413 	if (err)
2414 		goto err_setup;
2415 
2416 	return 0;
2417 
2418 err_setup:
2419 	mlx5_vdpa_destroy_mr(mvdev);
2420 err_mr:
2421 	return err;
2422 }
2423 
2424 /* reslock must be held for this function */
2425 static int setup_driver(struct mlx5_vdpa_dev *mvdev)
2426 {
2427 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2428 	int err;
2429 
2430 	WARN_ON(!rwsem_is_locked(&ndev->reslock));
2431 
2432 	if (ndev->setup) {
2433 		mlx5_vdpa_warn(mvdev, "setup driver called for already setup driver\n");
2434 		err = 0;
2435 		goto out;
2436 	}
2437 	err = setup_virtqueues(mvdev);
2438 	if (err) {
2439 		mlx5_vdpa_warn(mvdev, "setup_virtqueues\n");
2440 		goto out;
2441 	}
2442 
2443 	err = create_rqt(ndev);
2444 	if (err) {
2445 		mlx5_vdpa_warn(mvdev, "create_rqt\n");
2446 		goto err_rqt;
2447 	}
2448 
2449 	err = create_tir(ndev);
2450 	if (err) {
2451 		mlx5_vdpa_warn(mvdev, "create_tir\n");
2452 		goto err_tir;
2453 	}
2454 
2455 	err = setup_steering(ndev);
2456 	if (err) {
2457 		mlx5_vdpa_warn(mvdev, "setup_steering\n");
2458 		goto err_fwd;
2459 	}
2460 	ndev->setup = true;
2461 
2462 	return 0;
2463 
2464 err_fwd:
2465 	destroy_tir(ndev);
2466 err_tir:
2467 	destroy_rqt(ndev);
2468 err_rqt:
2469 	teardown_virtqueues(ndev);
2470 out:
2471 	return err;
2472 }
2473 
2474 /* reslock must be held for this function */
2475 static void teardown_driver(struct mlx5_vdpa_net *ndev)
2476 {
2477 
2478 	WARN_ON(!rwsem_is_locked(&ndev->reslock));
2479 
2480 	if (!ndev->setup)
2481 		return;
2482 
2483 	teardown_steering(ndev);
2484 	destroy_tir(ndev);
2485 	destroy_rqt(ndev);
2486 	teardown_virtqueues(ndev);
2487 	ndev->setup = false;
2488 }
2489 
2490 static void clear_vqs_ready(struct mlx5_vdpa_net *ndev)
2491 {
2492 	int i;
2493 
2494 	for (i = 0; i < ndev->mvdev.max_vqs; i++)
2495 		ndev->vqs[i].ready = false;
2496 
2497 	ndev->mvdev.cvq.ready = false;
2498 }
2499 
2500 static int setup_cvq_vring(struct mlx5_vdpa_dev *mvdev)
2501 {
2502 	struct mlx5_control_vq *cvq = &mvdev->cvq;
2503 	int err = 0;
2504 
2505 	if (mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ))
2506 		err = vringh_init_iotlb(&cvq->vring, mvdev->actual_features,
2507 					MLX5_CVQ_MAX_ENT, false,
2508 					(struct vring_desc *)(uintptr_t)cvq->desc_addr,
2509 					(struct vring_avail *)(uintptr_t)cvq->driver_addr,
2510 					(struct vring_used *)(uintptr_t)cvq->device_addr);
2511 
2512 	return err;
2513 }
2514 
2515 static void mlx5_vdpa_set_status(struct vdpa_device *vdev, u8 status)
2516 {
2517 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2518 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2519 	int err;
2520 
2521 	print_status(mvdev, status, true);
2522 
2523 	down_write(&ndev->reslock);
2524 
2525 	if ((status ^ ndev->mvdev.status) & VIRTIO_CONFIG_S_DRIVER_OK) {
2526 		if (status & VIRTIO_CONFIG_S_DRIVER_OK) {
2527 			err = setup_cvq_vring(mvdev);
2528 			if (err) {
2529 				mlx5_vdpa_warn(mvdev, "failed to setup control VQ vring\n");
2530 				goto err_setup;
2531 			}
2532 			err = setup_driver(mvdev);
2533 			if (err) {
2534 				mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
2535 				goto err_setup;
2536 			}
2537 		} else {
2538 			mlx5_vdpa_warn(mvdev, "did not expect DRIVER_OK to be cleared\n");
2539 			goto err_clear;
2540 		}
2541 	}
2542 
2543 	ndev->mvdev.status = status;
2544 	up_write(&ndev->reslock);
2545 	return;
2546 
2547 err_setup:
2548 	mlx5_vdpa_destroy_mr(&ndev->mvdev);
2549 	ndev->mvdev.status |= VIRTIO_CONFIG_S_FAILED;
2550 err_clear:
2551 	up_write(&ndev->reslock);
2552 }
2553 
2554 static void init_group_to_asid_map(struct mlx5_vdpa_dev *mvdev)
2555 {
2556 	int i;
2557 
2558 	/* default mapping all groups are mapped to asid 0 */
2559 	for (i = 0; i < MLX5_VDPA_NUMVQ_GROUPS; i++)
2560 		mvdev->group2asid[i] = 0;
2561 }
2562 
2563 static int mlx5_vdpa_reset(struct vdpa_device *vdev)
2564 {
2565 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2566 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2567 
2568 	print_status(mvdev, 0, true);
2569 	mlx5_vdpa_info(mvdev, "performing device reset\n");
2570 
2571 	down_write(&ndev->reslock);
2572 	teardown_driver(ndev);
2573 	clear_vqs_ready(ndev);
2574 	mlx5_vdpa_destroy_mr(&ndev->mvdev);
2575 	ndev->mvdev.status = 0;
2576 	ndev->cur_num_vqs = 0;
2577 	ndev->mvdev.cvq.received_desc = 0;
2578 	ndev->mvdev.cvq.completed_desc = 0;
2579 	memset(ndev->event_cbs, 0, sizeof(*ndev->event_cbs) * (mvdev->max_vqs + 1));
2580 	ndev->mvdev.actual_features = 0;
2581 	init_group_to_asid_map(mvdev);
2582 	++mvdev->generation;
2583 
2584 	if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
2585 		if (mlx5_vdpa_create_mr(mvdev, NULL))
2586 			mlx5_vdpa_warn(mvdev, "create MR failed\n");
2587 	}
2588 	up_write(&ndev->reslock);
2589 
2590 	return 0;
2591 }
2592 
2593 static size_t mlx5_vdpa_get_config_size(struct vdpa_device *vdev)
2594 {
2595 	return sizeof(struct virtio_net_config);
2596 }
2597 
2598 static void mlx5_vdpa_get_config(struct vdpa_device *vdev, unsigned int offset, void *buf,
2599 				 unsigned int len)
2600 {
2601 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2602 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2603 
2604 	if (offset + len <= sizeof(struct virtio_net_config))
2605 		memcpy(buf, (u8 *)&ndev->config + offset, len);
2606 }
2607 
2608 static void mlx5_vdpa_set_config(struct vdpa_device *vdev, unsigned int offset, const void *buf,
2609 				 unsigned int len)
2610 {
2611 	/* not supported */
2612 }
2613 
2614 static u32 mlx5_vdpa_get_generation(struct vdpa_device *vdev)
2615 {
2616 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2617 
2618 	return mvdev->generation;
2619 }
2620 
2621 static int set_map_control(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
2622 {
2623 	u64 start = 0ULL, last = 0ULL - 1;
2624 	struct vhost_iotlb_map *map;
2625 	int err = 0;
2626 
2627 	spin_lock(&mvdev->cvq.iommu_lock);
2628 	vhost_iotlb_reset(mvdev->cvq.iotlb);
2629 
2630 	for (map = vhost_iotlb_itree_first(iotlb, start, last); map;
2631 	     map = vhost_iotlb_itree_next(map, start, last)) {
2632 		err = vhost_iotlb_add_range(mvdev->cvq.iotlb, map->start,
2633 					    map->last, map->addr, map->perm);
2634 		if (err)
2635 			goto out;
2636 	}
2637 
2638 out:
2639 	spin_unlock(&mvdev->cvq.iommu_lock);
2640 	return err;
2641 }
2642 
2643 static int set_map_data(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
2644 {
2645 	bool change_map;
2646 	int err;
2647 
2648 	err = mlx5_vdpa_handle_set_map(mvdev, iotlb, &change_map);
2649 	if (err) {
2650 		mlx5_vdpa_warn(mvdev, "set map failed(%d)\n", err);
2651 		return err;
2652 	}
2653 
2654 	if (change_map)
2655 		err = mlx5_vdpa_change_map(mvdev, iotlb);
2656 
2657 	return err;
2658 }
2659 
2660 static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
2661 			     struct vhost_iotlb *iotlb)
2662 {
2663 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2664 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2665 	int err = -EINVAL;
2666 
2667 	down_write(&ndev->reslock);
2668 	if (mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] == asid) {
2669 		err = set_map_data(mvdev, iotlb);
2670 		if (err)
2671 			goto out;
2672 	}
2673 
2674 	if (mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] == asid)
2675 		err = set_map_control(mvdev, iotlb);
2676 
2677 out:
2678 	up_write(&ndev->reslock);
2679 	return err;
2680 }
2681 
2682 static void mlx5_vdpa_free(struct vdpa_device *vdev)
2683 {
2684 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2685 	struct mlx5_core_dev *pfmdev;
2686 	struct mlx5_vdpa_net *ndev;
2687 
2688 	ndev = to_mlx5_vdpa_ndev(mvdev);
2689 
2690 	free_resources(ndev);
2691 	mlx5_vdpa_destroy_mr(mvdev);
2692 	if (!is_zero_ether_addr(ndev->config.mac)) {
2693 		pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
2694 		mlx5_mpfs_del_mac(pfmdev, ndev->config.mac);
2695 	}
2696 	mlx5_vdpa_free_resources(&ndev->mvdev);
2697 	kfree(ndev->event_cbs);
2698 	kfree(ndev->vqs);
2699 }
2700 
2701 static struct vdpa_notification_area mlx5_get_vq_notification(struct vdpa_device *vdev, u16 idx)
2702 {
2703 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2704 	struct vdpa_notification_area ret = {};
2705 	struct mlx5_vdpa_net *ndev;
2706 	phys_addr_t addr;
2707 
2708 	if (!is_index_valid(mvdev, idx) || is_ctrl_vq_idx(mvdev, idx))
2709 		return ret;
2710 
2711 	/* If SF BAR size is smaller than PAGE_SIZE, do not use direct
2712 	 * notification to avoid the risk of mapping pages that contain BAR of more
2713 	 * than one SF
2714 	 */
2715 	if (MLX5_CAP_GEN(mvdev->mdev, log_min_sf_size) + 12 < PAGE_SHIFT)
2716 		return ret;
2717 
2718 	ndev = to_mlx5_vdpa_ndev(mvdev);
2719 	addr = (phys_addr_t)ndev->mvdev.res.phys_kick_addr;
2720 	ret.addr = addr;
2721 	ret.size = PAGE_SIZE;
2722 	return ret;
2723 }
2724 
2725 static int mlx5_get_vq_irq(struct vdpa_device *vdv, u16 idx)
2726 {
2727 	return -EOPNOTSUPP;
2728 }
2729 
2730 static u64 mlx5_vdpa_get_driver_features(struct vdpa_device *vdev)
2731 {
2732 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2733 
2734 	return mvdev->actual_features;
2735 }
2736 
2737 static int counter_set_query(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq,
2738 			     u64 *received_desc, u64 *completed_desc)
2739 {
2740 	u32 in[MLX5_ST_SZ_DW(query_virtio_q_counters_in)] = {};
2741 	u32 out[MLX5_ST_SZ_DW(query_virtio_q_counters_out)] = {};
2742 	void *cmd_hdr;
2743 	void *ctx;
2744 	int err;
2745 
2746 	if (!counters_supported(&ndev->mvdev))
2747 		return -EOPNOTSUPP;
2748 
2749 	if (mvq->fw_state != MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY)
2750 		return -EAGAIN;
2751 
2752 	cmd_hdr = MLX5_ADDR_OF(query_virtio_q_counters_in, in, hdr);
2753 
2754 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, opcode, MLX5_CMD_OP_QUERY_GENERAL_OBJECT);
2755 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_type, MLX5_OBJ_TYPE_VIRTIO_Q_COUNTERS);
2756 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, uid, ndev->mvdev.res.uid);
2757 	MLX5_SET(general_obj_in_cmd_hdr, cmd_hdr, obj_id, mvq->counter_set_id);
2758 
2759 	err = mlx5_cmd_exec(ndev->mvdev.mdev, in, sizeof(in), out, sizeof(out));
2760 	if (err)
2761 		return err;
2762 
2763 	ctx = MLX5_ADDR_OF(query_virtio_q_counters_out, out, counters);
2764 	*received_desc = MLX5_GET64(virtio_q_counters, ctx, received_desc);
2765 	*completed_desc = MLX5_GET64(virtio_q_counters, ctx, completed_desc);
2766 	return 0;
2767 }
2768 
2769 static int mlx5_vdpa_get_vendor_vq_stats(struct vdpa_device *vdev, u16 idx,
2770 					 struct sk_buff *msg,
2771 					 struct netlink_ext_ack *extack)
2772 {
2773 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2774 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2775 	struct mlx5_vdpa_virtqueue *mvq;
2776 	struct mlx5_control_vq *cvq;
2777 	u64 received_desc;
2778 	u64 completed_desc;
2779 	int err = 0;
2780 
2781 	down_read(&ndev->reslock);
2782 	if (!is_index_valid(mvdev, idx)) {
2783 		NL_SET_ERR_MSG_MOD(extack, "virtqueue index is not valid");
2784 		err = -EINVAL;
2785 		goto out_err;
2786 	}
2787 
2788 	if (idx == ctrl_vq_idx(mvdev)) {
2789 		cvq = &mvdev->cvq;
2790 		received_desc = cvq->received_desc;
2791 		completed_desc = cvq->completed_desc;
2792 		goto out;
2793 	}
2794 
2795 	mvq = &ndev->vqs[idx];
2796 	err = counter_set_query(ndev, mvq, &received_desc, &completed_desc);
2797 	if (err) {
2798 		NL_SET_ERR_MSG_MOD(extack, "failed to query hardware");
2799 		goto out_err;
2800 	}
2801 
2802 out:
2803 	err = -EMSGSIZE;
2804 	if (nla_put_string(msg, VDPA_ATTR_DEV_VENDOR_ATTR_NAME, "received_desc"))
2805 		goto out_err;
2806 
2807 	if (nla_put_u64_64bit(msg, VDPA_ATTR_DEV_VENDOR_ATTR_VALUE, received_desc,
2808 			      VDPA_ATTR_PAD))
2809 		goto out_err;
2810 
2811 	if (nla_put_string(msg, VDPA_ATTR_DEV_VENDOR_ATTR_NAME, "completed_desc"))
2812 		goto out_err;
2813 
2814 	if (nla_put_u64_64bit(msg, VDPA_ATTR_DEV_VENDOR_ATTR_VALUE, completed_desc,
2815 			      VDPA_ATTR_PAD))
2816 		goto out_err;
2817 
2818 	err = 0;
2819 out_err:
2820 	up_read(&ndev->reslock);
2821 	return err;
2822 }
2823 
2824 static void mlx5_vdpa_cvq_suspend(struct mlx5_vdpa_dev *mvdev)
2825 {
2826 	struct mlx5_control_vq *cvq;
2827 
2828 	if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
2829 		return;
2830 
2831 	cvq = &mvdev->cvq;
2832 	cvq->ready = false;
2833 }
2834 
2835 static int mlx5_vdpa_suspend(struct vdpa_device *vdev)
2836 {
2837 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2838 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
2839 	struct mlx5_vdpa_virtqueue *mvq;
2840 	int i;
2841 
2842 	down_write(&ndev->reslock);
2843 	mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
2844 	ndev->nb_registered = false;
2845 	flush_workqueue(ndev->mvdev.wq);
2846 	for (i = 0; i < ndev->cur_num_vqs; i++) {
2847 		mvq = &ndev->vqs[i];
2848 		suspend_vq(ndev, mvq);
2849 	}
2850 	mlx5_vdpa_cvq_suspend(mvdev);
2851 	up_write(&ndev->reslock);
2852 	return 0;
2853 }
2854 
2855 static int mlx5_set_group_asid(struct vdpa_device *vdev, u32 group,
2856 			       unsigned int asid)
2857 {
2858 	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
2859 
2860 	if (group >= MLX5_VDPA_NUMVQ_GROUPS)
2861 		return -EINVAL;
2862 
2863 	mvdev->group2asid[group] = asid;
2864 	return 0;
2865 }
2866 
2867 static const struct vdpa_config_ops mlx5_vdpa_ops = {
2868 	.set_vq_address = mlx5_vdpa_set_vq_address,
2869 	.set_vq_num = mlx5_vdpa_set_vq_num,
2870 	.kick_vq = mlx5_vdpa_kick_vq,
2871 	.set_vq_cb = mlx5_vdpa_set_vq_cb,
2872 	.set_vq_ready = mlx5_vdpa_set_vq_ready,
2873 	.get_vq_ready = mlx5_vdpa_get_vq_ready,
2874 	.set_vq_state = mlx5_vdpa_set_vq_state,
2875 	.get_vq_state = mlx5_vdpa_get_vq_state,
2876 	.get_vendor_vq_stats = mlx5_vdpa_get_vendor_vq_stats,
2877 	.get_vq_notification = mlx5_get_vq_notification,
2878 	.get_vq_irq = mlx5_get_vq_irq,
2879 	.get_vq_align = mlx5_vdpa_get_vq_align,
2880 	.get_vq_group = mlx5_vdpa_get_vq_group,
2881 	.get_device_features = mlx5_vdpa_get_device_features,
2882 	.set_driver_features = mlx5_vdpa_set_driver_features,
2883 	.get_driver_features = mlx5_vdpa_get_driver_features,
2884 	.set_config_cb = mlx5_vdpa_set_config_cb,
2885 	.get_vq_num_max = mlx5_vdpa_get_vq_num_max,
2886 	.get_device_id = mlx5_vdpa_get_device_id,
2887 	.get_vendor_id = mlx5_vdpa_get_vendor_id,
2888 	.get_status = mlx5_vdpa_get_status,
2889 	.set_status = mlx5_vdpa_set_status,
2890 	.reset = mlx5_vdpa_reset,
2891 	.get_config_size = mlx5_vdpa_get_config_size,
2892 	.get_config = mlx5_vdpa_get_config,
2893 	.set_config = mlx5_vdpa_set_config,
2894 	.get_generation = mlx5_vdpa_get_generation,
2895 	.set_map = mlx5_vdpa_set_map,
2896 	.set_group_asid = mlx5_set_group_asid,
2897 	.free = mlx5_vdpa_free,
2898 	.suspend = mlx5_vdpa_suspend,
2899 };
2900 
2901 static int query_mtu(struct mlx5_core_dev *mdev, u16 *mtu)
2902 {
2903 	u16 hw_mtu;
2904 	int err;
2905 
2906 	err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2907 	if (err)
2908 		return err;
2909 
2910 	*mtu = hw_mtu - MLX5V_ETH_HARD_MTU;
2911 	return 0;
2912 }
2913 
2914 static int alloc_resources(struct mlx5_vdpa_net *ndev)
2915 {
2916 	struct mlx5_vdpa_net_resources *res = &ndev->res;
2917 	int err;
2918 
2919 	if (res->valid) {
2920 		mlx5_vdpa_warn(&ndev->mvdev, "resources already allocated\n");
2921 		return -EEXIST;
2922 	}
2923 
2924 	err = mlx5_vdpa_alloc_transport_domain(&ndev->mvdev, &res->tdn);
2925 	if (err)
2926 		return err;
2927 
2928 	err = create_tis(ndev);
2929 	if (err)
2930 		goto err_tis;
2931 
2932 	res->valid = true;
2933 
2934 	return 0;
2935 
2936 err_tis:
2937 	mlx5_vdpa_dealloc_transport_domain(&ndev->mvdev, res->tdn);
2938 	return err;
2939 }
2940 
2941 static void free_resources(struct mlx5_vdpa_net *ndev)
2942 {
2943 	struct mlx5_vdpa_net_resources *res = &ndev->res;
2944 
2945 	if (!res->valid)
2946 		return;
2947 
2948 	destroy_tis(ndev);
2949 	mlx5_vdpa_dealloc_transport_domain(&ndev->mvdev, res->tdn);
2950 	res->valid = false;
2951 }
2952 
2953 static void init_mvqs(struct mlx5_vdpa_net *ndev)
2954 {
2955 	struct mlx5_vdpa_virtqueue *mvq;
2956 	int i;
2957 
2958 	for (i = 0; i < ndev->mvdev.max_vqs; ++i) {
2959 		mvq = &ndev->vqs[i];
2960 		memset(mvq, 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2961 		mvq->index = i;
2962 		mvq->ndev = ndev;
2963 		mvq->fwqp.fw = true;
2964 		mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_NONE;
2965 	}
2966 	for (; i < ndev->mvdev.max_vqs; i++) {
2967 		mvq = &ndev->vqs[i];
2968 		memset(mvq, 0, offsetof(struct mlx5_vdpa_virtqueue, ri));
2969 		mvq->index = i;
2970 		mvq->ndev = ndev;
2971 	}
2972 }
2973 
2974 struct mlx5_vdpa_mgmtdev {
2975 	struct vdpa_mgmt_dev mgtdev;
2976 	struct mlx5_adev *madev;
2977 	struct mlx5_vdpa_net *ndev;
2978 };
2979 
2980 static u8 query_vport_state(struct mlx5_core_dev *mdev, u8 opmod, u16 vport)
2981 {
2982 	u32 out[MLX5_ST_SZ_DW(query_vport_state_out)] = {};
2983 	u32 in[MLX5_ST_SZ_DW(query_vport_state_in)] = {};
2984 	int err;
2985 
2986 	MLX5_SET(query_vport_state_in, in, opcode, MLX5_CMD_OP_QUERY_VPORT_STATE);
2987 	MLX5_SET(query_vport_state_in, in, op_mod, opmod);
2988 	MLX5_SET(query_vport_state_in, in, vport_number, vport);
2989 	if (vport)
2990 		MLX5_SET(query_vport_state_in, in, other_vport, 1);
2991 
2992 	err = mlx5_cmd_exec_inout(mdev, query_vport_state, in, out);
2993 	if (err)
2994 		return 0;
2995 
2996 	return MLX5_GET(query_vport_state_out, out, state);
2997 }
2998 
2999 static bool get_link_state(struct mlx5_vdpa_dev *mvdev)
3000 {
3001 	if (query_vport_state(mvdev->mdev, MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT, 0) ==
3002 	    VPORT_STATE_UP)
3003 		return true;
3004 
3005 	return false;
3006 }
3007 
3008 static void update_carrier(struct work_struct *work)
3009 {
3010 	struct mlx5_vdpa_wq_ent *wqent;
3011 	struct mlx5_vdpa_dev *mvdev;
3012 	struct mlx5_vdpa_net *ndev;
3013 
3014 	wqent = container_of(work, struct mlx5_vdpa_wq_ent, work);
3015 	mvdev = wqent->mvdev;
3016 	ndev = to_mlx5_vdpa_ndev(mvdev);
3017 	if (get_link_state(mvdev))
3018 		ndev->config.status |= cpu_to_mlx5vdpa16(mvdev, VIRTIO_NET_S_LINK_UP);
3019 	else
3020 		ndev->config.status &= cpu_to_mlx5vdpa16(mvdev, ~VIRTIO_NET_S_LINK_UP);
3021 
3022 	if (ndev->config_cb.callback)
3023 		ndev->config_cb.callback(ndev->config_cb.private);
3024 
3025 	kfree(wqent);
3026 }
3027 
3028 static int event_handler(struct notifier_block *nb, unsigned long event, void *param)
3029 {
3030 	struct mlx5_vdpa_net *ndev = container_of(nb, struct mlx5_vdpa_net, nb);
3031 	struct mlx5_eqe *eqe = param;
3032 	int ret = NOTIFY_DONE;
3033 	struct mlx5_vdpa_wq_ent *wqent;
3034 
3035 	if (event == MLX5_EVENT_TYPE_PORT_CHANGE) {
3036 		switch (eqe->sub_type) {
3037 		case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
3038 		case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
3039 			down_read(&ndev->reslock);
3040 			if (!ndev->nb_registered) {
3041 				up_read(&ndev->reslock);
3042 				return NOTIFY_DONE;
3043 			}
3044 			wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
3045 			if (!wqent) {
3046 				up_read(&ndev->reslock);
3047 				return NOTIFY_DONE;
3048 			}
3049 
3050 			wqent->mvdev = &ndev->mvdev;
3051 			INIT_WORK(&wqent->work, update_carrier);
3052 			queue_work(ndev->mvdev.wq, &wqent->work);
3053 			up_read(&ndev->reslock);
3054 			ret = NOTIFY_OK;
3055 			break;
3056 		default:
3057 			return NOTIFY_DONE;
3058 		}
3059 		return ret;
3060 	}
3061 	return ret;
3062 }
3063 
3064 static int config_func_mtu(struct mlx5_core_dev *mdev, u16 mtu)
3065 {
3066 	int inlen = MLX5_ST_SZ_BYTES(modify_nic_vport_context_in);
3067 	void *in;
3068 	int err;
3069 
3070 	in = kvzalloc(inlen, GFP_KERNEL);
3071 	if (!in)
3072 		return -ENOMEM;
3073 
3074 	MLX5_SET(modify_nic_vport_context_in, in, field_select.mtu, 1);
3075 	MLX5_SET(modify_nic_vport_context_in, in, nic_vport_context.mtu,
3076 		 mtu + MLX5V_ETH_HARD_MTU);
3077 	MLX5_SET(modify_nic_vport_context_in, in, opcode,
3078 		 MLX5_CMD_OP_MODIFY_NIC_VPORT_CONTEXT);
3079 
3080 	err = mlx5_cmd_exec_in(mdev, modify_nic_vport_context, in);
3081 
3082 	kvfree(in);
3083 	return err;
3084 }
3085 
3086 static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
3087 			     const struct vdpa_dev_set_config *add_config)
3088 {
3089 	struct mlx5_vdpa_mgmtdev *mgtdev = container_of(v_mdev, struct mlx5_vdpa_mgmtdev, mgtdev);
3090 	struct virtio_net_config *config;
3091 	struct mlx5_core_dev *pfmdev;
3092 	struct mlx5_vdpa_dev *mvdev;
3093 	struct mlx5_vdpa_net *ndev;
3094 	struct mlx5_core_dev *mdev;
3095 	u32 max_vqs;
3096 	u16 mtu;
3097 	int err;
3098 
3099 	if (mgtdev->ndev)
3100 		return -ENOSPC;
3101 
3102 	mdev = mgtdev->madev->mdev;
3103 	if (!(MLX5_CAP_DEV_VDPA_EMULATION(mdev, virtio_queue_type) &
3104 	    MLX5_VIRTIO_EMULATION_CAP_VIRTIO_QUEUE_TYPE_SPLIT)) {
3105 		dev_warn(mdev->device, "missing support for split virtqueues\n");
3106 		return -EOPNOTSUPP;
3107 	}
3108 
3109 	max_vqs = min_t(int, MLX5_CAP_DEV_VDPA_EMULATION(mdev, max_num_virtio_queues),
3110 			1 << MLX5_CAP_GEN(mdev, log_max_rqt_size));
3111 	if (max_vqs < 2) {
3112 		dev_warn(mdev->device,
3113 			 "%d virtqueues are supported. At least 2 are required\n",
3114 			 max_vqs);
3115 		return -EAGAIN;
3116 	}
3117 
3118 	if (add_config->mask & BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MAX_VQP)) {
3119 		if (add_config->net.max_vq_pairs > max_vqs / 2)
3120 			return -EINVAL;
3121 		max_vqs = min_t(u32, max_vqs, 2 * add_config->net.max_vq_pairs);
3122 	} else {
3123 		max_vqs = 2;
3124 	}
3125 
3126 	ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev->device, &mlx5_vdpa_ops,
3127 				 MLX5_VDPA_NUMVQ_GROUPS, MLX5_VDPA_NUM_AS, name, false);
3128 	if (IS_ERR(ndev))
3129 		return PTR_ERR(ndev);
3130 
3131 	ndev->mvdev.mlx_features = mgtdev->mgtdev.supported_features;
3132 	ndev->mvdev.max_vqs = max_vqs;
3133 	mvdev = &ndev->mvdev;
3134 	mvdev->mdev = mdev;
3135 
3136 	ndev->vqs = kcalloc(max_vqs, sizeof(*ndev->vqs), GFP_KERNEL);
3137 	ndev->event_cbs = kcalloc(max_vqs + 1, sizeof(*ndev->event_cbs), GFP_KERNEL);
3138 	if (!ndev->vqs || !ndev->event_cbs) {
3139 		err = -ENOMEM;
3140 		goto err_alloc;
3141 	}
3142 
3143 	init_mvqs(ndev);
3144 	init_rwsem(&ndev->reslock);
3145 	config = &ndev->config;
3146 
3147 	if (add_config->mask & BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MTU)) {
3148 		err = config_func_mtu(mdev, add_config->net.mtu);
3149 		if (err)
3150 			goto err_alloc;
3151 	}
3152 
3153 	err = query_mtu(mdev, &mtu);
3154 	if (err)
3155 		goto err_alloc;
3156 
3157 	ndev->config.mtu = cpu_to_mlx5vdpa16(mvdev, mtu);
3158 
3159 	if (get_link_state(mvdev))
3160 		ndev->config.status |= cpu_to_mlx5vdpa16(mvdev, VIRTIO_NET_S_LINK_UP);
3161 	else
3162 		ndev->config.status &= cpu_to_mlx5vdpa16(mvdev, ~VIRTIO_NET_S_LINK_UP);
3163 
3164 	if (add_config->mask & (1 << VDPA_ATTR_DEV_NET_CFG_MACADDR)) {
3165 		memcpy(ndev->config.mac, add_config->net.mac, ETH_ALEN);
3166 	} else {
3167 		err = mlx5_query_nic_vport_mac_address(mdev, 0, 0, config->mac);
3168 		if (err)
3169 			goto err_alloc;
3170 	}
3171 
3172 	if (!is_zero_ether_addr(config->mac)) {
3173 		pfmdev = pci_get_drvdata(pci_physfn(mdev->pdev));
3174 		err = mlx5_mpfs_add_mac(pfmdev, config->mac);
3175 		if (err)
3176 			goto err_alloc;
3177 
3178 		ndev->mvdev.mlx_features |= BIT_ULL(VIRTIO_NET_F_MAC);
3179 	}
3180 
3181 	config->max_virtqueue_pairs = cpu_to_mlx5vdpa16(mvdev, max_vqs / 2);
3182 	mvdev->vdev.dma_dev = &mdev->pdev->dev;
3183 	err = mlx5_vdpa_alloc_resources(&ndev->mvdev);
3184 	if (err)
3185 		goto err_mpfs;
3186 
3187 	if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
3188 		err = mlx5_vdpa_create_mr(mvdev, NULL);
3189 		if (err)
3190 			goto err_res;
3191 	}
3192 
3193 	err = alloc_resources(ndev);
3194 	if (err)
3195 		goto err_mr;
3196 
3197 	ndev->cvq_ent.mvdev = mvdev;
3198 	INIT_WORK(&ndev->cvq_ent.work, mlx5_cvq_kick_handler);
3199 	mvdev->wq = create_singlethread_workqueue("mlx5_vdpa_wq");
3200 	if (!mvdev->wq) {
3201 		err = -ENOMEM;
3202 		goto err_res2;
3203 	}
3204 
3205 	ndev->nb.notifier_call = event_handler;
3206 	mlx5_notifier_register(mdev, &ndev->nb);
3207 	ndev->nb_registered = true;
3208 	mvdev->vdev.mdev = &mgtdev->mgtdev;
3209 	err = _vdpa_register_device(&mvdev->vdev, max_vqs + 1);
3210 	if (err)
3211 		goto err_reg;
3212 
3213 	mgtdev->ndev = ndev;
3214 	return 0;
3215 
3216 err_reg:
3217 	destroy_workqueue(mvdev->wq);
3218 err_res2:
3219 	free_resources(ndev);
3220 err_mr:
3221 	mlx5_vdpa_destroy_mr(mvdev);
3222 err_res:
3223 	mlx5_vdpa_free_resources(&ndev->mvdev);
3224 err_mpfs:
3225 	if (!is_zero_ether_addr(config->mac))
3226 		mlx5_mpfs_del_mac(pfmdev, config->mac);
3227 err_alloc:
3228 	put_device(&mvdev->vdev.dev);
3229 	return err;
3230 }
3231 
3232 static void mlx5_vdpa_dev_del(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *dev)
3233 {
3234 	struct mlx5_vdpa_mgmtdev *mgtdev = container_of(v_mdev, struct mlx5_vdpa_mgmtdev, mgtdev);
3235 	struct mlx5_vdpa_dev *mvdev = to_mvdev(dev);
3236 	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
3237 	struct workqueue_struct *wq;
3238 
3239 	if (ndev->nb_registered) {
3240 		mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
3241 		ndev->nb_registered = false;
3242 	}
3243 	wq = mvdev->wq;
3244 	mvdev->wq = NULL;
3245 	destroy_workqueue(wq);
3246 	_vdpa_unregister_device(dev);
3247 	mgtdev->ndev = NULL;
3248 }
3249 
3250 static const struct vdpa_mgmtdev_ops mdev_ops = {
3251 	.dev_add = mlx5_vdpa_dev_add,
3252 	.dev_del = mlx5_vdpa_dev_del,
3253 };
3254 
3255 static struct virtio_device_id id_table[] = {
3256 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3257 	{ 0 },
3258 };
3259 
3260 static int mlx5v_probe(struct auxiliary_device *adev,
3261 		       const struct auxiliary_device_id *id)
3262 
3263 {
3264 	struct mlx5_adev *madev = container_of(adev, struct mlx5_adev, adev);
3265 	struct mlx5_core_dev *mdev = madev->mdev;
3266 	struct mlx5_vdpa_mgmtdev *mgtdev;
3267 	int err;
3268 
3269 	mgtdev = kzalloc(sizeof(*mgtdev), GFP_KERNEL);
3270 	if (!mgtdev)
3271 		return -ENOMEM;
3272 
3273 	mgtdev->mgtdev.ops = &mdev_ops;
3274 	mgtdev->mgtdev.device = mdev->device;
3275 	mgtdev->mgtdev.id_table = id_table;
3276 	mgtdev->mgtdev.config_attr_mask = BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MACADDR) |
3277 					  BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MAX_VQP) |
3278 					  BIT_ULL(VDPA_ATTR_DEV_NET_CFG_MTU);
3279 	mgtdev->mgtdev.max_supported_vqs =
3280 		MLX5_CAP_DEV_VDPA_EMULATION(mdev, max_num_virtio_queues) + 1;
3281 	mgtdev->mgtdev.supported_features = get_supported_features(mdev);
3282 	mgtdev->madev = madev;
3283 
3284 	err = vdpa_mgmtdev_register(&mgtdev->mgtdev);
3285 	if (err)
3286 		goto reg_err;
3287 
3288 	auxiliary_set_drvdata(adev, mgtdev);
3289 
3290 	return 0;
3291 
3292 reg_err:
3293 	kfree(mgtdev);
3294 	return err;
3295 }
3296 
3297 static void mlx5v_remove(struct auxiliary_device *adev)
3298 {
3299 	struct mlx5_vdpa_mgmtdev *mgtdev;
3300 
3301 	mgtdev = auxiliary_get_drvdata(adev);
3302 	vdpa_mgmtdev_unregister(&mgtdev->mgtdev);
3303 	kfree(mgtdev);
3304 }
3305 
3306 static const struct auxiliary_device_id mlx5v_id_table[] = {
3307 	{ .name = MLX5_ADEV_NAME ".vnet", },
3308 	{},
3309 };
3310 
3311 MODULE_DEVICE_TABLE(auxiliary, mlx5v_id_table);
3312 
3313 static struct auxiliary_driver mlx5v_driver = {
3314 	.name = "vnet",
3315 	.probe = mlx5v_probe,
3316 	.remove = mlx5v_remove,
3317 	.id_table = mlx5v_id_table,
3318 };
3319 
3320 module_auxiliary_driver(mlx5v_driver);
3321