xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_cq.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
1 /*-
2  * Copyright (c) 2013-2020, Mellanox Technologies.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  */
25 
26 #include "opt_rss.h"
27 #include "opt_ratelimit.h"
28 
29 #include <linux/kref.h>
30 #include <rdma/ib_umem.h>
31 #include <rdma/ib_user_verbs.h>
32 #include <rdma/ib_cache.h>
33 #include <rdma/uverbs_ioctl.h>
34 #include <dev/mlx5/mlx5_ib/mlx5_ib.h>
35 
36 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe __unused)
37 {
38 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
39 
40 	ibcq->comp_handler(ibcq, ibcq->cq_context);
41 }
42 
43 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, int type)
44 {
45 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
46 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
47 	struct ib_cq *ibcq = &cq->ibcq;
48 	struct ib_event event;
49 
50 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
51 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
52 			     type, mcq->cqn);
53 		return;
54 	}
55 
56 	if (ibcq->event_handler) {
57 		event.device     = &dev->ib_dev;
58 		event.event      = IB_EVENT_CQ_ERR;
59 		event.element.cq = ibcq;
60 		ibcq->event_handler(&event, ibcq->cq_context);
61 	}
62 }
63 
64 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
65 {
66 	return mlx5_buf_offset(&buf->buf, n * size);
67 }
68 
69 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
70 {
71 	return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
72 }
73 
74 static u8 sw_ownership_bit(int n, int nent)
75 {
76 	return (n & nent) ? 1 : 0;
77 }
78 
79 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
80 {
81 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
82 	struct mlx5_cqe64 *cqe64;
83 
84 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
85 
86 	if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
87 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
88 		return cqe;
89 	} else {
90 		return NULL;
91 	}
92 }
93 
94 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
95 {
96 	return get_sw_cqe(cq, cq->mcq.cons_index);
97 }
98 
99 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
100 {
101 	switch (wq->wr_data[idx]) {
102 	case MLX5_IB_WR_UMR:
103 		return 0;
104 
105 	case IB_WR_LOCAL_INV:
106 		return IB_WC_LOCAL_INV;
107 
108 	case IB_WR_REG_MR:
109 		return IB_WC_REG_MR;
110 
111 	default:
112 		pr_warn("unknown completion status\n");
113 		return 0;
114 	}
115 }
116 
117 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
118 			    struct mlx5_ib_wq *wq, int idx)
119 {
120 	wc->wc_flags = 0;
121 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
122 	case MLX5_OPCODE_RDMA_WRITE_IMM:
123 		wc->wc_flags |= IB_WC_WITH_IMM;
124 	case MLX5_OPCODE_RDMA_WRITE:
125 		wc->opcode    = IB_WC_RDMA_WRITE;
126 		break;
127 	case MLX5_OPCODE_SEND_IMM:
128 		wc->wc_flags |= IB_WC_WITH_IMM;
129 	case MLX5_OPCODE_SEND:
130 	case MLX5_OPCODE_SEND_INVAL:
131 		wc->opcode    = IB_WC_SEND;
132 		break;
133 	case MLX5_OPCODE_RDMA_READ:
134 		wc->opcode    = IB_WC_RDMA_READ;
135 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
136 		break;
137 	case MLX5_OPCODE_ATOMIC_CS:
138 		wc->opcode    = IB_WC_COMP_SWAP;
139 		wc->byte_len  = 8;
140 		break;
141 	case MLX5_OPCODE_ATOMIC_FA:
142 		wc->opcode    = IB_WC_FETCH_ADD;
143 		wc->byte_len  = 8;
144 		break;
145 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
146 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
147 		wc->byte_len  = 8;
148 		break;
149 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
150 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
151 		wc->byte_len  = 8;
152 		break;
153 	case MLX5_OPCODE_UMR:
154 		wc->opcode = get_umr_comp(wq, idx);
155 		break;
156 	}
157 }
158 
159 enum {
160 	MLX5_GRH_IN_BUFFER = 1,
161 	MLX5_GRH_IN_CQE	   = 2,
162 };
163 
164 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
165 			     struct mlx5_ib_qp *qp)
166 {
167 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
168 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
169 	struct mlx5_ib_srq *srq;
170 	struct mlx5_ib_wq *wq;
171 	u16 wqe_ctr;
172 	u8  roce_packet_type;
173 	bool vlan_present;
174 	u8 g;
175 
176 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
177 		struct mlx5_core_srq *msrq = NULL;
178 
179 		if (qp->ibqp.xrcd) {
180 			msrq = mlx5_core_get_srq(dev->mdev,
181 						 be32_to_cpu(cqe->srqn));
182 			srq = to_mibsrq(msrq);
183 		} else {
184 			srq = to_msrq(qp->ibqp.srq);
185 		}
186 		if (srq) {
187 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
188 			wc->wr_id = srq->wrid[wqe_ctr];
189 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
190 			if (msrq && atomic_dec_and_test(&msrq->refcount))
191 				complete(&msrq->free);
192 		}
193 	} else {
194 		wq	  = &qp->rq;
195 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
196 		++wq->tail;
197 	}
198 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
199 
200 	switch (cqe->op_own >> 4) {
201 	case MLX5_CQE_RESP_WR_IMM:
202 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
203 		wc->wc_flags	= IB_WC_WITH_IMM;
204 		wc->ex.imm_data = cqe->imm_inval_pkey;
205 		break;
206 	case MLX5_CQE_RESP_SEND:
207 		wc->opcode   = IB_WC_RECV;
208 		wc->wc_flags = IB_WC_IP_CSUM_OK;
209 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
210 			       (cqe->hds_ip_ext & CQE_L4_OK))))
211 			wc->wc_flags = 0;
212 		break;
213 	case MLX5_CQE_RESP_SEND_IMM:
214 		wc->opcode	= IB_WC_RECV;
215 		wc->wc_flags	= IB_WC_WITH_IMM;
216 		wc->ex.imm_data = cqe->imm_inval_pkey;
217 		break;
218 	case MLX5_CQE_RESP_SEND_INV:
219 		wc->opcode	= IB_WC_RECV;
220 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
221 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
222 		break;
223 	}
224 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
225 	wc->dlid_path_bits = cqe->ml_path;
226 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
227 	wc->wc_flags |= g ? IB_WC_GRH : 0;
228 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
229 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
230 
231 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
232 				    &wc->pkey_index);
233 	} else {
234 		wc->pkey_index = 0;
235 	}
236 
237 	if (ll != IB_LINK_LAYER_ETHERNET) {
238 		wc->slid = be16_to_cpu(cqe->slid);
239 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
240 		return;
241 	}
242 
243 	wc->slid = 0;
244 	vlan_present = cqe_has_vlan(cqe);
245 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
246 	if (vlan_present) {
247 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
248 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
249 		wc->wc_flags |= IB_WC_WITH_VLAN;
250 	} else {
251 		wc->sl = 0;
252 	}
253 
254 	switch (roce_packet_type) {
255 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
256 		wc->network_hdr_type = RDMA_NETWORK_IB;
257 		break;
258 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
259 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
260 		break;
261 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
262 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
263 		break;
264 	}
265 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
266 }
267 
268 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
269 {
270 	__be32 *p = (__be32 *)cqe;
271 	int i;
272 
273 	mlx5_ib_warn(dev, "dump error cqe\n");
274 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
275 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
276 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
277 			be32_to_cpu(p[3]));
278 }
279 
280 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
281 				  struct mlx5_err_cqe *cqe,
282 				  struct ib_wc *wc)
283 {
284 	int dump = 1;
285 
286 	switch (cqe->syndrome) {
287 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
288 		wc->status = IB_WC_LOC_LEN_ERR;
289 		break;
290 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
291 		wc->status = IB_WC_LOC_QP_OP_ERR;
292 		break;
293 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
294 		wc->status = IB_WC_LOC_PROT_ERR;
295 		break;
296 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
297 		dump = 0;
298 		wc->status = IB_WC_WR_FLUSH_ERR;
299 		break;
300 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
301 		wc->status = IB_WC_MW_BIND_ERR;
302 		break;
303 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
304 		wc->status = IB_WC_BAD_RESP_ERR;
305 		break;
306 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
307 		wc->status = IB_WC_LOC_ACCESS_ERR;
308 		break;
309 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
310 		wc->status = IB_WC_REM_INV_REQ_ERR;
311 		break;
312 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
313 		wc->status = IB_WC_REM_ACCESS_ERR;
314 		break;
315 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
316 		wc->status = IB_WC_REM_OP_ERR;
317 		break;
318 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
319 		wc->status = IB_WC_RETRY_EXC_ERR;
320 		dump = 0;
321 		break;
322 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
323 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
324 		dump = 0;
325 		break;
326 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
327 		wc->status = IB_WC_REM_ABORT_ERR;
328 		break;
329 	default:
330 		wc->status = IB_WC_GENERAL_ERR;
331 		break;
332 	}
333 
334 	wc->vendor_err = cqe->vendor_err_synd;
335 	if (dump)
336 		dump_cqe(dev, cqe);
337 }
338 
339 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
340 {
341 	/* TBD: waiting decision
342 	*/
343 	return 0;
344 }
345 
346 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
347 {
348 	struct mlx5_wqe_data_seg *dpseg;
349 	void *addr;
350 
351 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
352 		sizeof(struct mlx5_wqe_raddr_seg) +
353 		sizeof(struct mlx5_wqe_atomic_seg);
354 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
355 	return addr;
356 }
357 
358 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
359 			  uint16_t idx)
360 {
361 	void *addr;
362 	int byte_count;
363 	int i;
364 
365 	if (!is_atomic_response(qp, idx))
366 		return;
367 
368 	byte_count = be32_to_cpu(cqe64->byte_cnt);
369 	addr = mlx5_get_atomic_laddr(qp, idx);
370 
371 	if (byte_count == 4) {
372 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
373 	} else {
374 		for (i = 0; i < byte_count; i += 8) {
375 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
376 			addr += 8;
377 		}
378 	}
379 
380 	return;
381 }
382 
383 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
384 			   u16 tail, u16 head)
385 {
386 	u16 idx;
387 
388 	do {
389 		idx = tail & (qp->sq.wqe_cnt - 1);
390 		handle_atomic(qp, cqe64, idx);
391 		if (idx == head)
392 			break;
393 
394 		tail = qp->sq.w_list[idx].next;
395 	} while (1);
396 	tail = qp->sq.w_list[idx].next;
397 	qp->sq.last_poll = tail;
398 }
399 
400 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
401 {
402 	mlx5_buf_free(dev->mdev, &buf->buf);
403 }
404 
405 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
406 			     struct ib_sig_err *item)
407 {
408 	u16 syndrome = be16_to_cpu(cqe->syndrome);
409 
410 #define GUARD_ERR   (1 << 13)
411 #define APPTAG_ERR  (1 << 12)
412 #define REFTAG_ERR  (1 << 11)
413 
414 	if (syndrome & GUARD_ERR) {
415 		item->err_type = IB_SIG_BAD_GUARD;
416 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
417 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
418 	} else
419 	if (syndrome & REFTAG_ERR) {
420 		item->err_type = IB_SIG_BAD_REFTAG;
421 		item->expected = be32_to_cpu(cqe->expected_reftag);
422 		item->actual = be32_to_cpu(cqe->actual_reftag);
423 	} else
424 	if (syndrome & APPTAG_ERR) {
425 		item->err_type = IB_SIG_BAD_APPTAG;
426 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
427 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
428 	} else {
429 		pr_err("Got signature completion error with bad syndrome %04x\n",
430 		       syndrome);
431 	}
432 
433 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
434 	item->key = be32_to_cpu(cqe->mkey);
435 }
436 
437 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
438 			 struct ib_wc *wc, int *npolled)
439 {
440 	struct mlx5_ib_wq *wq;
441 	unsigned int cur;
442 	unsigned int idx;
443 	int np;
444 	int i;
445 
446 	wq = &qp->sq;
447 	cur = wq->head - wq->tail;
448 	np = *npolled;
449 
450 	if (cur == 0)
451 		return;
452 
453 	for (i = 0;  i < cur && np < num_entries; i++) {
454 		idx = wq->last_poll & (wq->wqe_cnt - 1);
455 		wc->wr_id = wq->wrid[idx];
456 		wc->status = IB_WC_WR_FLUSH_ERR;
457 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
458 		wq->tail++;
459 		np++;
460 		wc->qp = &qp->ibqp;
461 		wc++;
462 		wq->last_poll = wq->w_list[idx].next;
463 	}
464 	*npolled = np;
465 }
466 
467 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
468 			 struct ib_wc *wc, int *npolled)
469 {
470 	struct mlx5_ib_wq *wq;
471 	unsigned int cur;
472 	int np;
473 	int i;
474 
475 	wq = &qp->rq;
476 	cur = wq->head - wq->tail;
477 	np = *npolled;
478 
479 	if (cur == 0)
480 		return;
481 
482 	for (i = 0;  i < cur && np < num_entries; i++) {
483 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
484 		wc->status = IB_WC_WR_FLUSH_ERR;
485 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
486 		wq->tail++;
487 		np++;
488 		wc->qp = &qp->ibqp;
489 		wc++;
490 	}
491 	*npolled = np;
492 }
493 
494 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
495 				 struct ib_wc *wc, int *npolled)
496 {
497 	struct mlx5_ib_qp *qp;
498 
499 	*npolled = 0;
500 	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
501 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
502 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
503 		if (*npolled >= num_entries)
504 			return;
505 	}
506 
507 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
508 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
509 		if (*npolled >= num_entries)
510 			return;
511 	}
512 }
513 
514 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
515 			 struct mlx5_ib_qp **cur_qp,
516 			 struct ib_wc *wc)
517 {
518 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
519 	struct mlx5_err_cqe *err_cqe;
520 	struct mlx5_cqe64 *cqe64;
521 	struct mlx5_core_qp *mqp;
522 	struct mlx5_ib_wq *wq;
523 	struct mlx5_sig_err_cqe *sig_err_cqe;
524 	struct mlx5_core_mkey *mmkey;
525 	struct mlx5_ib_mr *mr;
526 	unsigned long flags;
527 	uint8_t opcode;
528 	uint32_t qpn;
529 	u16 wqe_ctr;
530 	void *cqe;
531 	int idx;
532 
533 repoll:
534 	cqe = next_cqe_sw(cq);
535 	if (!cqe)
536 		return -EAGAIN;
537 
538 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
539 
540 	++cq->mcq.cons_index;
541 
542 	/* Make sure we read CQ entry contents after we've checked the
543 	 * ownership bit.
544 	 */
545 	rmb();
546 
547 	opcode = cqe64->op_own >> 4;
548 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
549 		if (likely(cq->resize_buf)) {
550 			free_cq_buf(dev, &cq->buf);
551 			cq->buf = *cq->resize_buf;
552 			kfree(cq->resize_buf);
553 			cq->resize_buf = NULL;
554 			goto repoll;
555 		} else {
556 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
557 		}
558 	}
559 
560 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
561 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
562 		/* We do not have to take the QP table lock here,
563 		 * because CQs will be locked while QPs are removed
564 		 * from the table.
565 		 */
566 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
567 		*cur_qp = to_mibqp(mqp);
568 	}
569 
570 	wc->qp  = &(*cur_qp)->ibqp;
571 	switch (opcode) {
572 	case MLX5_CQE_REQ:
573 		wq = &(*cur_qp)->sq;
574 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
575 		idx = wqe_ctr & (wq->wqe_cnt - 1);
576 		handle_good_req(wc, cqe64, wq, idx);
577 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
578 		wc->wr_id = wq->wrid[idx];
579 		wq->tail = wq->wqe_head[idx] + 1;
580 		wc->status = IB_WC_SUCCESS;
581 		break;
582 	case MLX5_CQE_RESP_WR_IMM:
583 	case MLX5_CQE_RESP_SEND:
584 	case MLX5_CQE_RESP_SEND_IMM:
585 	case MLX5_CQE_RESP_SEND_INV:
586 		handle_responder(wc, cqe64, *cur_qp);
587 		wc->status = IB_WC_SUCCESS;
588 		break;
589 	case MLX5_CQE_RESIZE_CQ:
590 		break;
591 	case MLX5_CQE_REQ_ERR:
592 	case MLX5_CQE_RESP_ERR:
593 		err_cqe = (struct mlx5_err_cqe *)cqe64;
594 		mlx5_handle_error_cqe(dev, err_cqe, wc);
595 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
596 			    opcode == MLX5_CQE_REQ_ERR ?
597 			    "Requestor" : "Responder", cq->mcq.cqn);
598 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
599 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
600 		if (opcode == MLX5_CQE_REQ_ERR) {
601 			wq = &(*cur_qp)->sq;
602 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
603 			idx = wqe_ctr & (wq->wqe_cnt - 1);
604 			wc->wr_id = wq->wrid[idx];
605 			wq->tail = wq->wqe_head[idx] + 1;
606 		} else {
607 			struct mlx5_ib_srq *srq;
608 
609 			if ((*cur_qp)->ibqp.srq) {
610 				srq = to_msrq((*cur_qp)->ibqp.srq);
611 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
612 				wc->wr_id = srq->wrid[wqe_ctr];
613 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
614 			} else {
615 				wq = &(*cur_qp)->rq;
616 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
617 				++wq->tail;
618 			}
619 		}
620 		break;
621 	case MLX5_CQE_SIG_ERR:
622 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
623 
624 		spin_lock_irqsave(&dev->mdev->priv.mr_table.lock, flags);
625 		mmkey = __mlx5_mr_lookup(dev->mdev,
626 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
627 		mr = to_mibmr(mmkey);
628 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
629 		mr->sig->sig_err_exists = true;
630 		mr->sig->sigerr_count++;
631 
632 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
633 			     cq->mcq.cqn, mr->sig->err_item.key,
634 			     mr->sig->err_item.err_type,
635 			     (long long)mr->sig->err_item.sig_err_offset,
636 			     mr->sig->err_item.expected,
637 			     mr->sig->err_item.actual);
638 
639 		spin_unlock_irqrestore(&dev->mdev->priv.mr_table.lock, flags);
640 		goto repoll;
641 	}
642 
643 	return 0;
644 }
645 
646 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
647 			struct ib_wc *wc)
648 {
649 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
650 	struct mlx5_ib_wc *soft_wc, *next;
651 	int npolled = 0;
652 
653 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
654 		if (npolled >= num_entries)
655 			break;
656 
657 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
658 			    cq->mcq.cqn);
659 
660 		wc[npolled++] = soft_wc->wc;
661 		list_del(&soft_wc->list);
662 		kfree(soft_wc);
663 	}
664 
665 	return npolled;
666 }
667 
668 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
669 {
670 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
671 	struct mlx5_ib_qp *cur_qp = NULL;
672 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
673 	struct mlx5_core_dev *mdev = dev->mdev;
674 	unsigned long flags;
675 	int soft_polled = 0;
676 	int npolled;
677 
678 	spin_lock_irqsave(&cq->lock, flags);
679 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)) {
680 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
681 		goto out;
682 	}
683 
684 	if (unlikely(!list_empty(&cq->wc_list)))
685 		soft_polled = poll_soft_wc(cq, num_entries, wc);
686 
687 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
688 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
689 			break;
690 	}
691 
692 	if (npolled)
693 		mlx5_cq_set_ci(&cq->mcq);
694 out:
695 	spin_unlock_irqrestore(&cq->lock, flags);
696 
697 	return soft_polled + npolled;
698 }
699 
700 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
701 {
702 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
703 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
704 	void __iomem *uar_page = mdev->priv.uar->map;
705 	unsigned long irq_flags;
706 	int ret = 0;
707 
708 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
709 		return -1;
710 
711 	spin_lock_irqsave(&cq->lock, irq_flags);
712 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
713 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
714 
715 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
716 		ret = 1;
717 	spin_unlock_irqrestore(&cq->lock, irq_flags);
718 
719 	mlx5_cq_arm(&cq->mcq,
720 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
721 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
722 		    uar_page,
723 		    MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
724 		    cq->mcq.cons_index);
725 
726 	return ret;
727 }
728 
729 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
730 			int nent, int cqe_size)
731 {
732 	int err;
733 
734 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
735 	    2 * PAGE_SIZE, &buf->buf);
736 	if (err)
737 		return err;
738 
739 	buf->cqe_size = cqe_size;
740 	buf->nent = nent;
741 
742 	return 0;
743 }
744 
745 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
746 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
747 			  int *cqe_size, int *index, int *inlen)
748 {
749 	struct mlx5_ib_create_cq ucmd = {};
750 	size_t ucmdlen;
751 	int page_shift;
752 	__be64 *pas;
753 	int npages;
754 	int ncont;
755 	void *cqc;
756 	int err;
757 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
758 		udata, struct mlx5_ib_ucontext, ibucontext);
759 
760 	ucmdlen = min(udata->inlen, sizeof(ucmd));
761 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
762 		return -EINVAL;
763 
764 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
765 		return -EFAULT;
766 
767 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX)))
768 		return -EINVAL;
769 
770 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
771 		return -EINVAL;
772 
773 	*cqe_size = ucmd.cqe_size;
774 
775 	cq->buf.umem = ib_umem_get(&context->ibucontext, ucmd.buf_addr,
776 				   entries * ucmd.cqe_size,
777 				   IB_ACCESS_LOCAL_WRITE, 1);
778 	if (IS_ERR(cq->buf.umem)) {
779 		err = PTR_ERR(cq->buf.umem);
780 		return err;
781 	}
782 
783 	err = mlx5_ib_db_map_user(context, ucmd.db_addr,
784 				  &cq->db);
785 	if (err)
786 		goto err_umem;
787 
788 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
789 			   &ncont, NULL);
790 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
791 		    (long long)ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
792 
793 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
794 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
795 	*cqb = mlx5_vzalloc(*inlen);
796 	if (!*cqb) {
797 		err = -ENOMEM;
798 		goto err_db;
799 	}
800 
801 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
802 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
803 
804 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
805 	MLX5_SET(cqc, cqc, log_page_size,
806 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
807 
808 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
809 		*index = ucmd.uar_page_index;
810 	} else if (context->bfregi.lib_uar_dyn) {
811 		err = -EINVAL;
812 		goto err_cqb;
813 	} else {
814 		*index = context->bfregi.sys_pages[0];
815 	}
816 
817 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
818 	return 0;
819 
820 err_cqb:
821 	kvfree(*cqb);
822 
823 err_db:
824 	mlx5_ib_db_unmap_user(context, &cq->db);
825 
826 err_umem:
827 	ib_umem_release(cq->buf.umem);
828 	return err;
829 }
830 
831 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
832 {
833 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
834 		udata, struct mlx5_ib_ucontext, ibucontext);
835 
836 	mlx5_ib_db_unmap_user(context, &cq->db);
837 	ib_umem_release(cq->buf.umem);
838 }
839 
840 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
841 {
842 	int i;
843 	void *cqe;
844 	struct mlx5_cqe64 *cqe64;
845 
846 	for (i = 0; i < buf->nent; i++) {
847 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
848 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
849 		cqe64->op_own = MLX5_CQE_INVALID << 4;
850 	}
851 }
852 
853 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
854 			    int entries, int cqe_size,
855 			    u32 **cqb, int *index, int *inlen)
856 {
857 	__be64 *pas;
858 	void *cqc;
859 	int err;
860 
861 	err = mlx5_db_alloc(dev->mdev, &cq->db);
862 	if (err)
863 		return err;
864 
865 	cq->mcq.set_ci_db  = cq->db.db;
866 	cq->mcq.arm_db     = cq->db.db + 1;
867 	cq->mcq.cqe_sz = cqe_size;
868 
869 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
870 	if (err)
871 		goto err_db;
872 
873 	init_cq_buf(cq, &cq->buf);
874 
875 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
876 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
877 	*cqb = mlx5_vzalloc(*inlen);
878 	if (!*cqb) {
879 		err = -ENOMEM;
880 		goto err_buf;
881 	}
882 
883 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
884 	mlx5_fill_page_array(&cq->buf.buf, pas);
885 
886 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
887 	MLX5_SET(cqc, cqc, log_page_size,
888 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
889 
890 	*index = dev->mdev->priv.uar->index;
891 
892 	return 0;
893 
894 err_buf:
895 	free_cq_buf(dev, &cq->buf);
896 
897 err_db:
898 	mlx5_db_free(dev->mdev, &cq->db);
899 	return err;
900 }
901 
902 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
903 {
904 	free_cq_buf(dev, &cq->buf);
905 	mlx5_db_free(dev->mdev, &cq->db);
906 }
907 
908 static void notify_soft_wc_handler(struct work_struct *work)
909 {
910 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
911 					     notify_work);
912 
913 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
914 }
915 
916 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
917 		      struct ib_udata *udata)
918 {
919 	struct ib_device *ibdev = ibcq->device;
920 	int entries = attr->cqe;
921 	int vector = attr->comp_vector;
922 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
923 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
924 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
925 	int uninitialized_var(index);
926 	int uninitialized_var(inlen);
927 	u32 *cqb = NULL;
928 	void *cqc;
929 	int cqe_size;
930 	unsigned int irqn;
931 	int eqn;
932 	int err;
933 
934 	if (entries < 0 ||
935 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
936 		return -EINVAL;
937 
938 	if (check_cq_create_flags(attr->flags))
939 		return -EOPNOTSUPP;
940 
941 	entries = roundup_pow_of_two(entries + 1);
942 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
943 		return -EINVAL;
944 
945 	cq->ibcq.cqe = entries - 1;
946 	mutex_init(&cq->resize_mutex);
947 	spin_lock_init(&cq->lock);
948 	cq->resize_buf = NULL;
949 	cq->resize_umem = NULL;
950 	cq->create_flags = attr->flags;
951 	INIT_LIST_HEAD(&cq->list_send_qp);
952 	INIT_LIST_HEAD(&cq->list_recv_qp);
953 
954 	if (udata) {
955 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
956 				     &index, &inlen);
957 		if (err)
958 			return err;
959 	} else {
960 		cqe_size = cache_line_size() == 128 ? 128 : 64;
961 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
962 				       &index, &inlen);
963 		if (err)
964 			return err;
965 
966 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
967 	}
968 
969 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
970 	if (err)
971 		goto err_cqb;
972 
973 	cq->cqe_size = cqe_size;
974 
975 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
976 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
977 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
978 	MLX5_SET(cqc, cqc, uar_page, index);
979 	MLX5_SET(cqc, cqc, c_eqn, eqn);
980 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
981 	if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
982 		MLX5_SET(cqc, cqc, oi, 1);
983 
984 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
985 	if (err)
986 		goto err_cqb;
987 
988 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
989 	cq->mcq.irqn = irqn;
990 	cq->mcq.comp  = mlx5_ib_cq_comp;
991 	cq->mcq.event = mlx5_ib_cq_event;
992 
993 	INIT_LIST_HEAD(&cq->wc_list);
994 
995 	if (udata)
996 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
997 			err = -EFAULT;
998 			goto err_cmd;
999 		}
1000 
1001 
1002 	kvfree(cqb);
1003 	return 0;
1004 
1005 err_cmd:
1006 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1007 
1008 err_cqb:
1009 	kvfree(cqb);
1010 	if (udata)
1011 		destroy_cq_user(cq, udata);
1012 	else
1013 		destroy_cq_kernel(dev, cq);
1014 	return err;
1015 }
1016 
1017 void mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1018 {
1019 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1020 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1021 
1022 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1023 	if (udata)
1024 		destroy_cq_user(mcq, udata);
1025 	else
1026 		destroy_cq_kernel(dev, mcq);
1027 }
1028 
1029 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1030 {
1031 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1032 }
1033 
1034 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1035 {
1036 	struct mlx5_cqe64 *cqe64, *dest64;
1037 	void *cqe, *dest;
1038 	u32 prod_index;
1039 	int nfreed = 0;
1040 	u8 owner_bit;
1041 
1042 	if (!cq)
1043 		return;
1044 
1045 	/* First we need to find the current producer index, so we
1046 	 * know where to start cleaning from.  It doesn't matter if HW
1047 	 * adds new entries after this loop -- the QP we're worried
1048 	 * about is already in RESET, so the new entries won't come
1049 	 * from our QP and therefore don't need to be checked.
1050 	 */
1051 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1052 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1053 			break;
1054 
1055 	/* Now sweep backwards through the CQ, removing CQ entries
1056 	 * that match our QP by copying older entries on top of them.
1057 	 */
1058 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1059 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1060 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1061 		if (is_equal_rsn(cqe64, rsn)) {
1062 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1063 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1064 			++nfreed;
1065 		} else if (nfreed) {
1066 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1067 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1068 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1069 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1070 			dest64->op_own = owner_bit |
1071 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1072 		}
1073 	}
1074 
1075 	if (nfreed) {
1076 		cq->mcq.cons_index += nfreed;
1077 		/* Make sure update of buffer contents is done before
1078 		 * updating consumer index.
1079 		 */
1080 		wmb();
1081 		mlx5_cq_set_ci(&cq->mcq);
1082 	}
1083 }
1084 
1085 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1086 {
1087 	if (!cq)
1088 		return;
1089 
1090 	spin_lock_irq(&cq->lock);
1091 	__mlx5_ib_cq_clean(cq, qpn, srq);
1092 	spin_unlock_irq(&cq->lock);
1093 }
1094 
1095 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1096 {
1097 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1098 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1099 	int err;
1100 
1101 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1102 		return -ENOSYS;
1103 
1104 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1105 					     cq_period, cq_count);
1106 	if (err)
1107 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1108 
1109 	return err;
1110 }
1111 
1112 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1113 		       int entries, struct ib_udata *udata, int *npas,
1114 		       int *page_shift, int *cqe_size)
1115 {
1116 	struct mlx5_ib_resize_cq ucmd;
1117 	struct ib_umem *umem;
1118 	int err;
1119 	int npages;
1120 	struct ib_ucontext *context = cq->buf.umem->context;
1121 
1122 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1123 	if (err)
1124 		return err;
1125 
1126 	if (ucmd.reserved0 || ucmd.reserved1)
1127 		return -EINVAL;
1128 
1129 	/* check multiplication overflow */
1130 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1131 		return -EINVAL;
1132 
1133 	umem = ib_umem_get(context, ucmd.buf_addr,
1134 			   (size_t)ucmd.cqe_size * entries,
1135 			   IB_ACCESS_LOCAL_WRITE, 1);
1136 	if (IS_ERR(umem)) {
1137 		err = PTR_ERR(umem);
1138 		return err;
1139 	}
1140 
1141 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1142 			   npas, NULL);
1143 
1144 	cq->resize_umem = umem;
1145 	*cqe_size = ucmd.cqe_size;
1146 
1147 	return 0;
1148 }
1149 
1150 static void un_resize_user(struct mlx5_ib_cq *cq)
1151 {
1152 	ib_umem_release(cq->resize_umem);
1153 }
1154 
1155 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1156 			 int entries, int cqe_size)
1157 {
1158 	int err;
1159 
1160 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1161 	if (!cq->resize_buf)
1162 		return -ENOMEM;
1163 
1164 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1165 	if (err)
1166 		goto ex;
1167 
1168 	init_cq_buf(cq, cq->resize_buf);
1169 
1170 	return 0;
1171 
1172 ex:
1173 	kfree(cq->resize_buf);
1174 	return err;
1175 }
1176 
1177 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1178 {
1179 	free_cq_buf(dev, cq->resize_buf);
1180 	cq->resize_buf = NULL;
1181 }
1182 
1183 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1184 {
1185 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1186 	struct mlx5_cqe64 *scqe64;
1187 	struct mlx5_cqe64 *dcqe64;
1188 	void *start_cqe;
1189 	void *scqe;
1190 	void *dcqe;
1191 	int ssize;
1192 	int dsize;
1193 	int i;
1194 	u8 sw_own;
1195 
1196 	ssize = cq->buf.cqe_size;
1197 	dsize = cq->resize_buf->cqe_size;
1198 	if (ssize != dsize) {
1199 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1200 		return -EINVAL;
1201 	}
1202 
1203 	i = cq->mcq.cons_index;
1204 	scqe = get_sw_cqe(cq, i);
1205 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1206 	start_cqe = scqe;
1207 	if (!scqe) {
1208 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1209 		return -EINVAL;
1210 	}
1211 
1212 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1213 		dcqe = get_cqe_from_buf(cq->resize_buf,
1214 					(i + 1) & (cq->resize_buf->nent),
1215 					dsize);
1216 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1217 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1218 		memcpy(dcqe, scqe, dsize);
1219 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1220 
1221 		++i;
1222 		scqe = get_sw_cqe(cq, i);
1223 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1224 		if (!scqe) {
1225 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1226 			return -EINVAL;
1227 		}
1228 
1229 		if (scqe == start_cqe) {
1230 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1231 				cq->mcq.cqn);
1232 			return -ENOMEM;
1233 		}
1234 	}
1235 	++cq->mcq.cons_index;
1236 	return 0;
1237 }
1238 
1239 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1240 {
1241 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1242 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1243 	void *cqc;
1244 	u32 *in;
1245 	int err;
1246 	int npas;
1247 	__be64 *pas;
1248 	int page_shift;
1249 	int inlen;
1250 	int uninitialized_var(cqe_size);
1251 	unsigned long flags;
1252 
1253 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1254 		pr_info("Firmware does not support resize CQ\n");
1255 		return -ENOSYS;
1256 	}
1257 
1258 	if (entries < 1 ||
1259 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1260 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1261 			     entries,
1262 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1263 		return -EINVAL;
1264 	}
1265 
1266 	entries = roundup_pow_of_two(entries + 1);
1267 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1268 		return -EINVAL;
1269 
1270 	if (entries == ibcq->cqe + 1)
1271 		return 0;
1272 
1273 	mutex_lock(&cq->resize_mutex);
1274 	if (udata) {
1275 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1276 				  &cqe_size);
1277 	} else {
1278 		cqe_size = 64;
1279 		err = resize_kernel(dev, cq, entries, cqe_size);
1280 		if (!err) {
1281 			npas = cq->resize_buf->buf.npages;
1282 			page_shift = cq->resize_buf->buf.page_shift;
1283 		}
1284 	}
1285 
1286 	if (err)
1287 		goto ex;
1288 
1289 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1290 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1291 
1292 	in = mlx5_vzalloc(inlen);
1293 	if (!in) {
1294 		err = -ENOMEM;
1295 		goto ex_resize;
1296 	}
1297 
1298 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1299 	if (udata)
1300 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1301 				     pas, 0);
1302 	else
1303 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1304 
1305 	MLX5_SET(modify_cq_in, in,
1306 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1307 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1308 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1309 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1310 
1311 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1312 
1313 	MLX5_SET(cqc, cqc, log_page_size,
1314 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1315 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1316 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1317 
1318 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1319 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1320 
1321 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1322 	if (err)
1323 		goto ex_alloc;
1324 
1325 	if (udata) {
1326 		cq->ibcq.cqe = entries - 1;
1327 		ib_umem_release(cq->buf.umem);
1328 		cq->buf.umem = cq->resize_umem;
1329 		cq->resize_umem = NULL;
1330 	} else {
1331 		struct mlx5_ib_cq_buf tbuf;
1332 		int resized = 0;
1333 
1334 		spin_lock_irqsave(&cq->lock, flags);
1335 		if (cq->resize_buf) {
1336 			err = copy_resize_cqes(cq);
1337 			if (!err) {
1338 				tbuf = cq->buf;
1339 				cq->buf = *cq->resize_buf;
1340 				kfree(cq->resize_buf);
1341 				cq->resize_buf = NULL;
1342 				resized = 1;
1343 			}
1344 		}
1345 		cq->ibcq.cqe = entries - 1;
1346 		spin_unlock_irqrestore(&cq->lock, flags);
1347 		if (resized)
1348 			free_cq_buf(dev, &tbuf);
1349 	}
1350 	mutex_unlock(&cq->resize_mutex);
1351 
1352 	kvfree(in);
1353 	return 0;
1354 
1355 ex_alloc:
1356 	kvfree(in);
1357 
1358 ex_resize:
1359 	if (udata)
1360 		un_resize_user(cq);
1361 	else
1362 		un_resize_kernel(dev, cq);
1363 ex:
1364 	mutex_unlock(&cq->resize_mutex);
1365 	return err;
1366 }
1367 
1368 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1369 {
1370 	struct mlx5_ib_cq *cq;
1371 
1372 	if (!ibcq)
1373 		return 128;
1374 
1375 	cq = to_mcq(ibcq);
1376 	return cq->cqe_size;
1377 }
1378 
1379 /* Called from atomic context */
1380 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1381 {
1382 	struct mlx5_ib_wc *soft_wc;
1383 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1384 	unsigned long flags;
1385 
1386 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1387 	if (!soft_wc)
1388 		return -ENOMEM;
1389 
1390 	soft_wc->wc = *wc;
1391 	spin_lock_irqsave(&cq->lock, flags);
1392 	list_add_tail(&soft_wc->list, &cq->wc_list);
1393 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1394 	    wc->status != IB_WC_SUCCESS) {
1395 		cq->notify_flags = 0;
1396 		schedule_work(&cq->notify_work);
1397 	}
1398 	spin_unlock_irqrestore(&cq->lock, flags);
1399 
1400 	return 0;
1401 }
1402