xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_cq.c (revision 5eb61f6c6549f134a4f3bed4c164345d4f616bad)
1 /*-
2  * Copyright (c) 2013-2015, Mellanox Technologies, Ltd.  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  * $FreeBSD$
26  */
27 
28 #include <linux/kref.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/ib_user_verbs.h>
31 #include <rdma/ib_cache.h>
32 #include "mlx5_ib.h"
33 
34 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe __unused)
35 {
36 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
37 
38 	ibcq->comp_handler(ibcq, ibcq->cq_context);
39 }
40 
41 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, int type)
42 {
43 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
44 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
45 	struct ib_cq *ibcq = &cq->ibcq;
46 	struct ib_event event;
47 
48 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
49 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
50 			     type, mcq->cqn);
51 		return;
52 	}
53 
54 	if (ibcq->event_handler) {
55 		event.device     = &dev->ib_dev;
56 		event.event      = IB_EVENT_CQ_ERR;
57 		event.element.cq = ibcq;
58 		ibcq->event_handler(&event, ibcq->cq_context);
59 	}
60 }
61 
62 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
63 {
64 	return mlx5_buf_offset(&buf->buf, n * size);
65 }
66 
67 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
68 {
69 	return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
70 }
71 
72 static u8 sw_ownership_bit(int n, int nent)
73 {
74 	return (n & nent) ? 1 : 0;
75 }
76 
77 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
78 {
79 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
80 	struct mlx5_cqe64 *cqe64;
81 
82 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
83 
84 	if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
85 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
86 		return cqe;
87 	} else {
88 		return NULL;
89 	}
90 }
91 
92 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
93 {
94 	return get_sw_cqe(cq, cq->mcq.cons_index);
95 }
96 
97 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
98 {
99 	switch (wq->wr_data[idx]) {
100 	case MLX5_IB_WR_UMR:
101 		return 0;
102 
103 	case IB_WR_LOCAL_INV:
104 		return IB_WC_LOCAL_INV;
105 
106 	case IB_WR_REG_MR:
107 		return IB_WC_REG_MR;
108 
109 	default:
110 		pr_warn("unknown completion status\n");
111 		return 0;
112 	}
113 }
114 
115 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
116 			    struct mlx5_ib_wq *wq, int idx)
117 {
118 	wc->wc_flags = 0;
119 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
120 	case MLX5_OPCODE_RDMA_WRITE_IMM:
121 		wc->wc_flags |= IB_WC_WITH_IMM;
122 	case MLX5_OPCODE_RDMA_WRITE:
123 		wc->opcode    = IB_WC_RDMA_WRITE;
124 		break;
125 	case MLX5_OPCODE_SEND_IMM:
126 		wc->wc_flags |= IB_WC_WITH_IMM;
127 	case MLX5_OPCODE_SEND:
128 	case MLX5_OPCODE_SEND_INVAL:
129 		wc->opcode    = IB_WC_SEND;
130 		break;
131 	case MLX5_OPCODE_RDMA_READ:
132 		wc->opcode    = IB_WC_RDMA_READ;
133 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
134 		break;
135 	case MLX5_OPCODE_ATOMIC_CS:
136 		wc->opcode    = IB_WC_COMP_SWAP;
137 		wc->byte_len  = 8;
138 		break;
139 	case MLX5_OPCODE_ATOMIC_FA:
140 		wc->opcode    = IB_WC_FETCH_ADD;
141 		wc->byte_len  = 8;
142 		break;
143 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
144 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
145 		wc->byte_len  = 8;
146 		break;
147 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
148 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
149 		wc->byte_len  = 8;
150 		break;
151 	case MLX5_OPCODE_UMR:
152 		wc->opcode = get_umr_comp(wq, idx);
153 		break;
154 	}
155 }
156 
157 enum {
158 	MLX5_GRH_IN_BUFFER = 1,
159 	MLX5_GRH_IN_CQE	   = 2,
160 };
161 
162 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
163 			     struct mlx5_ib_qp *qp)
164 {
165 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
166 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
167 	struct mlx5_ib_srq *srq;
168 	struct mlx5_ib_wq *wq;
169 	u16 wqe_ctr;
170 	u8  roce_packet_type;
171 	bool vlan_present;
172 	u8 g;
173 
174 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
175 		struct mlx5_core_srq *msrq = NULL;
176 
177 		if (qp->ibqp.xrcd) {
178 			msrq = mlx5_core_get_srq(dev->mdev,
179 						 be32_to_cpu(cqe->srqn));
180 			srq = to_mibsrq(msrq);
181 		} else {
182 			srq = to_msrq(qp->ibqp.srq);
183 		}
184 		if (srq) {
185 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
186 			wc->wr_id = srq->wrid[wqe_ctr];
187 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
188 			if (msrq && atomic_dec_and_test(&msrq->refcount))
189 				complete(&msrq->free);
190 		}
191 	} else {
192 		wq	  = &qp->rq;
193 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
194 		++wq->tail;
195 	}
196 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
197 
198 	switch (cqe->op_own >> 4) {
199 	case MLX5_CQE_RESP_WR_IMM:
200 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
201 		wc->wc_flags	= IB_WC_WITH_IMM;
202 		wc->ex.imm_data = cqe->imm_inval_pkey;
203 		break;
204 	case MLX5_CQE_RESP_SEND:
205 		wc->opcode   = IB_WC_RECV;
206 		wc->wc_flags = IB_WC_IP_CSUM_OK;
207 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
208 			       (cqe->hds_ip_ext & CQE_L4_OK))))
209 			wc->wc_flags = 0;
210 		break;
211 	case MLX5_CQE_RESP_SEND_IMM:
212 		wc->opcode	= IB_WC_RECV;
213 		wc->wc_flags	= IB_WC_WITH_IMM;
214 		wc->ex.imm_data = cqe->imm_inval_pkey;
215 		break;
216 	case MLX5_CQE_RESP_SEND_INV:
217 		wc->opcode	= IB_WC_RECV;
218 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
219 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
220 		break;
221 	}
222 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
223 	wc->dlid_path_bits = cqe->ml_path;
224 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
225 	wc->wc_flags |= g ? IB_WC_GRH : 0;
226 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
227 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
228 
229 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
230 				    &wc->pkey_index);
231 	} else {
232 		wc->pkey_index = 0;
233 	}
234 
235 	if (ll != IB_LINK_LAYER_ETHERNET) {
236 		wc->slid = be16_to_cpu(cqe->slid);
237 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
238 		return;
239 	}
240 
241 	wc->slid = 0;
242 	vlan_present = cqe_has_vlan(cqe);
243 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
244 	if (vlan_present) {
245 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
246 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
247 		wc->wc_flags |= IB_WC_WITH_VLAN;
248 	} else {
249 		wc->sl = 0;
250 	}
251 
252 	switch (roce_packet_type) {
253 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
254 		wc->network_hdr_type = RDMA_NETWORK_IB;
255 		break;
256 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
257 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
258 		break;
259 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
260 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
261 		break;
262 	}
263 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
264 }
265 
266 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
267 {
268 	__be32 *p = (__be32 *)cqe;
269 	int i;
270 
271 	mlx5_ib_warn(dev, "dump error cqe\n");
272 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
273 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
274 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
275 			be32_to_cpu(p[3]));
276 }
277 
278 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
279 				  struct mlx5_err_cqe *cqe,
280 				  struct ib_wc *wc)
281 {
282 	int dump = 1;
283 
284 	switch (cqe->syndrome) {
285 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
286 		wc->status = IB_WC_LOC_LEN_ERR;
287 		break;
288 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
289 		wc->status = IB_WC_LOC_QP_OP_ERR;
290 		break;
291 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
292 		wc->status = IB_WC_LOC_PROT_ERR;
293 		break;
294 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
295 		dump = 0;
296 		wc->status = IB_WC_WR_FLUSH_ERR;
297 		break;
298 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
299 		wc->status = IB_WC_MW_BIND_ERR;
300 		break;
301 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
302 		wc->status = IB_WC_BAD_RESP_ERR;
303 		break;
304 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
305 		wc->status = IB_WC_LOC_ACCESS_ERR;
306 		break;
307 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
308 		wc->status = IB_WC_REM_INV_REQ_ERR;
309 		break;
310 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
311 		wc->status = IB_WC_REM_ACCESS_ERR;
312 		break;
313 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
314 		wc->status = IB_WC_REM_OP_ERR;
315 		break;
316 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
317 		wc->status = IB_WC_RETRY_EXC_ERR;
318 		dump = 0;
319 		break;
320 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
321 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
322 		dump = 0;
323 		break;
324 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
325 		wc->status = IB_WC_REM_ABORT_ERR;
326 		break;
327 	default:
328 		wc->status = IB_WC_GENERAL_ERR;
329 		break;
330 	}
331 
332 	wc->vendor_err = cqe->vendor_err_synd;
333 	if (dump)
334 		dump_cqe(dev, cqe);
335 }
336 
337 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
338 {
339 	/* TBD: waiting decision
340 	*/
341 	return 0;
342 }
343 
344 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
345 {
346 	struct mlx5_wqe_data_seg *dpseg;
347 	void *addr;
348 
349 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
350 		sizeof(struct mlx5_wqe_raddr_seg) +
351 		sizeof(struct mlx5_wqe_atomic_seg);
352 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
353 	return addr;
354 }
355 
356 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
357 			  uint16_t idx)
358 {
359 	void *addr;
360 	int byte_count;
361 	int i;
362 
363 	if (!is_atomic_response(qp, idx))
364 		return;
365 
366 	byte_count = be32_to_cpu(cqe64->byte_cnt);
367 	addr = mlx5_get_atomic_laddr(qp, idx);
368 
369 	if (byte_count == 4) {
370 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
371 	} else {
372 		for (i = 0; i < byte_count; i += 8) {
373 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
374 			addr += 8;
375 		}
376 	}
377 
378 	return;
379 }
380 
381 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
382 			   u16 tail, u16 head)
383 {
384 	u16 idx;
385 
386 	do {
387 		idx = tail & (qp->sq.wqe_cnt - 1);
388 		handle_atomic(qp, cqe64, idx);
389 		if (idx == head)
390 			break;
391 
392 		tail = qp->sq.w_list[idx].next;
393 	} while (1);
394 	tail = qp->sq.w_list[idx].next;
395 	qp->sq.last_poll = tail;
396 }
397 
398 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
399 {
400 	mlx5_buf_free(dev->mdev, &buf->buf);
401 }
402 
403 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
404 			     struct ib_sig_err *item)
405 {
406 	u16 syndrome = be16_to_cpu(cqe->syndrome);
407 
408 #define GUARD_ERR   (1 << 13)
409 #define APPTAG_ERR  (1 << 12)
410 #define REFTAG_ERR  (1 << 11)
411 
412 	if (syndrome & GUARD_ERR) {
413 		item->err_type = IB_SIG_BAD_GUARD;
414 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
415 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
416 	} else
417 	if (syndrome & REFTAG_ERR) {
418 		item->err_type = IB_SIG_BAD_REFTAG;
419 		item->expected = be32_to_cpu(cqe->expected_reftag);
420 		item->actual = be32_to_cpu(cqe->actual_reftag);
421 	} else
422 	if (syndrome & APPTAG_ERR) {
423 		item->err_type = IB_SIG_BAD_APPTAG;
424 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
425 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
426 	} else {
427 		pr_err("Got signature completion error with bad syndrome %04x\n",
428 		       syndrome);
429 	}
430 
431 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
432 	item->key = be32_to_cpu(cqe->mkey);
433 }
434 
435 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
436 			 struct ib_wc *wc, int *npolled)
437 {
438 	struct mlx5_ib_wq *wq;
439 	unsigned int cur;
440 	unsigned int idx;
441 	int np;
442 	int i;
443 
444 	wq = &qp->sq;
445 	cur = wq->head - wq->tail;
446 	np = *npolled;
447 
448 	if (cur == 0)
449 		return;
450 
451 	for (i = 0;  i < cur && np < num_entries; i++) {
452 		idx = wq->last_poll & (wq->wqe_cnt - 1);
453 		wc->wr_id = wq->wrid[idx];
454 		wc->status = IB_WC_WR_FLUSH_ERR;
455 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
456 		wq->tail++;
457 		np++;
458 		wc->qp = &qp->ibqp;
459 		wc++;
460 		wq->last_poll = wq->w_list[idx].next;
461 	}
462 	*npolled = np;
463 }
464 
465 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
466 			 struct ib_wc *wc, int *npolled)
467 {
468 	struct mlx5_ib_wq *wq;
469 	unsigned int cur;
470 	int np;
471 	int i;
472 
473 	wq = &qp->rq;
474 	cur = wq->head - wq->tail;
475 	np = *npolled;
476 
477 	if (cur == 0)
478 		return;
479 
480 	for (i = 0;  i < cur && np < num_entries; i++) {
481 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
482 		wc->status = IB_WC_WR_FLUSH_ERR;
483 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
484 		wq->tail++;
485 		np++;
486 		wc->qp = &qp->ibqp;
487 		wc++;
488 	}
489 	*npolled = np;
490 }
491 
492 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
493 				 struct ib_wc *wc, int *npolled)
494 {
495 	struct mlx5_ib_qp *qp;
496 
497 	*npolled = 0;
498 	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
499 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
500 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
501 		if (*npolled >= num_entries)
502 			return;
503 	}
504 
505 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
506 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
507 		if (*npolled >= num_entries)
508 			return;
509 	}
510 }
511 
512 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
513 			 struct mlx5_ib_qp **cur_qp,
514 			 struct ib_wc *wc)
515 {
516 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
517 	struct mlx5_err_cqe *err_cqe;
518 	struct mlx5_cqe64 *cqe64;
519 	struct mlx5_core_qp *mqp;
520 	struct mlx5_ib_wq *wq;
521 	struct mlx5_sig_err_cqe *sig_err_cqe;
522 	struct mlx5_core_mr *mmkey;
523 	struct mlx5_ib_mr *mr;
524 	unsigned long flags;
525 	uint8_t opcode;
526 	uint32_t qpn;
527 	u16 wqe_ctr;
528 	void *cqe;
529 	int idx;
530 
531 repoll:
532 	cqe = next_cqe_sw(cq);
533 	if (!cqe)
534 		return -EAGAIN;
535 
536 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
537 
538 	++cq->mcq.cons_index;
539 
540 	/* Make sure we read CQ entry contents after we've checked the
541 	 * ownership bit.
542 	 */
543 	rmb();
544 
545 	opcode = cqe64->op_own >> 4;
546 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
547 		if (likely(cq->resize_buf)) {
548 			free_cq_buf(dev, &cq->buf);
549 			cq->buf = *cq->resize_buf;
550 			kfree(cq->resize_buf);
551 			cq->resize_buf = NULL;
552 			goto repoll;
553 		} else {
554 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
555 		}
556 	}
557 
558 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
559 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
560 		/* We do not have to take the QP table lock here,
561 		 * because CQs will be locked while QPs are removed
562 		 * from the table.
563 		 */
564 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
565 		*cur_qp = to_mibqp(mqp);
566 	}
567 
568 	wc->qp  = &(*cur_qp)->ibqp;
569 	switch (opcode) {
570 	case MLX5_CQE_REQ:
571 		wq = &(*cur_qp)->sq;
572 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
573 		idx = wqe_ctr & (wq->wqe_cnt - 1);
574 		handle_good_req(wc, cqe64, wq, idx);
575 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
576 		wc->wr_id = wq->wrid[idx];
577 		wq->tail = wq->wqe_head[idx] + 1;
578 		wc->status = IB_WC_SUCCESS;
579 		break;
580 	case MLX5_CQE_RESP_WR_IMM:
581 	case MLX5_CQE_RESP_SEND:
582 	case MLX5_CQE_RESP_SEND_IMM:
583 	case MLX5_CQE_RESP_SEND_INV:
584 		handle_responder(wc, cqe64, *cur_qp);
585 		wc->status = IB_WC_SUCCESS;
586 		break;
587 	case MLX5_CQE_RESIZE_CQ:
588 		break;
589 	case MLX5_CQE_REQ_ERR:
590 	case MLX5_CQE_RESP_ERR:
591 		err_cqe = (struct mlx5_err_cqe *)cqe64;
592 		mlx5_handle_error_cqe(dev, err_cqe, wc);
593 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
594 			    opcode == MLX5_CQE_REQ_ERR ?
595 			    "Requestor" : "Responder", cq->mcq.cqn);
596 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
597 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
598 		if (opcode == MLX5_CQE_REQ_ERR) {
599 			wq = &(*cur_qp)->sq;
600 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
601 			idx = wqe_ctr & (wq->wqe_cnt - 1);
602 			wc->wr_id = wq->wrid[idx];
603 			wq->tail = wq->wqe_head[idx] + 1;
604 		} else {
605 			struct mlx5_ib_srq *srq;
606 
607 			if ((*cur_qp)->ibqp.srq) {
608 				srq = to_msrq((*cur_qp)->ibqp.srq);
609 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
610 				wc->wr_id = srq->wrid[wqe_ctr];
611 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
612 			} else {
613 				wq = &(*cur_qp)->rq;
614 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
615 				++wq->tail;
616 			}
617 		}
618 		break;
619 	case MLX5_CQE_SIG_ERR:
620 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
621 
622 		spin_lock_irqsave(&dev->mdev->priv.mr_table.lock, flags);
623 		mmkey = __mlx5_mr_lookup(dev->mdev,
624 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
625 		mr = to_mibmr(mmkey);
626 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
627 		mr->sig->sig_err_exists = true;
628 		mr->sig->sigerr_count++;
629 
630 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
631 			     cq->mcq.cqn, mr->sig->err_item.key,
632 			     mr->sig->err_item.err_type,
633 			     (long long)mr->sig->err_item.sig_err_offset,
634 			     mr->sig->err_item.expected,
635 			     mr->sig->err_item.actual);
636 
637 		spin_unlock_irqrestore(&dev->mdev->priv.mr_table.lock, flags);
638 		goto repoll;
639 	}
640 
641 	return 0;
642 }
643 
644 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
645 			struct ib_wc *wc)
646 {
647 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
648 	struct mlx5_ib_wc *soft_wc, *next;
649 	int npolled = 0;
650 
651 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
652 		if (npolled >= num_entries)
653 			break;
654 
655 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
656 			    cq->mcq.cqn);
657 
658 		wc[npolled++] = soft_wc->wc;
659 		list_del(&soft_wc->list);
660 		kfree(soft_wc);
661 	}
662 
663 	return npolled;
664 }
665 
666 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
667 {
668 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
669 	struct mlx5_ib_qp *cur_qp = NULL;
670 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
671 	struct mlx5_core_dev *mdev = dev->mdev;
672 	unsigned long flags;
673 	int soft_polled = 0;
674 	int npolled;
675 
676 	spin_lock_irqsave(&cq->lock, flags);
677 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)) {
678 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
679 		goto out;
680 	}
681 
682 	if (unlikely(!list_empty(&cq->wc_list)))
683 		soft_polled = poll_soft_wc(cq, num_entries, wc);
684 
685 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
686 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
687 			break;
688 	}
689 
690 	if (npolled)
691 		mlx5_cq_set_ci(&cq->mcq);
692 out:
693 	spin_unlock_irqrestore(&cq->lock, flags);
694 
695 	return soft_polled + npolled;
696 }
697 
698 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
699 {
700 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
701 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
702 	void __iomem *uar_page = mdev->priv.uar->map;
703 	unsigned long irq_flags;
704 	int ret = 0;
705 
706 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
707 		return -1;
708 
709 	spin_lock_irqsave(&cq->lock, irq_flags);
710 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
711 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
712 
713 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
714 		ret = 1;
715 	spin_unlock_irqrestore(&cq->lock, irq_flags);
716 
717 	mlx5_cq_arm(&cq->mcq,
718 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
719 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
720 		    uar_page,
721 		    MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
722 		    cq->mcq.cons_index);
723 
724 	return ret;
725 }
726 
727 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
728 			int nent, int cqe_size)
729 {
730 	int err;
731 
732 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
733 	    2 * PAGE_SIZE, &buf->buf);
734 	if (err)
735 		return err;
736 
737 	buf->cqe_size = cqe_size;
738 	buf->nent = nent;
739 
740 	return 0;
741 }
742 
743 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
744 			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
745 			  int entries, u32 **cqb,
746 			  int *cqe_size, int *index, int *inlen)
747 {
748 	struct mlx5_ib_create_cq ucmd;
749 	size_t ucmdlen;
750 	int page_shift;
751 	__be64 *pas;
752 	int npages;
753 	int ncont;
754 	void *cqc;
755 	int err;
756 
757 	ucmdlen = min(udata->inlen, sizeof(ucmd));
758 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
759 		return -EINVAL;
760 
761 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
762 		return -EFAULT;
763 
764 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX)))
765 		return -EINVAL;
766 
767 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
768 		return -EINVAL;
769 
770 	*cqe_size = ucmd.cqe_size;
771 
772 	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
773 				   entries * ucmd.cqe_size,
774 				   IB_ACCESS_LOCAL_WRITE, 1);
775 	if (IS_ERR(cq->buf.umem)) {
776 		err = PTR_ERR(cq->buf.umem);
777 		return err;
778 	}
779 
780 	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
781 				  &cq->db);
782 	if (err)
783 		goto err_umem;
784 
785 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
786 			   &ncont, NULL);
787 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
788 		    (long long)ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
789 
790 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
791 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
792 	*cqb = mlx5_vzalloc(*inlen);
793 	if (!*cqb) {
794 		err = -ENOMEM;
795 		goto err_db;
796 	}
797 
798 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
799 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
800 
801 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
802 	MLX5_SET(cqc, cqc, log_page_size,
803 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
804 
805 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
806 		*index = ucmd.uar_page_index;
807 	} else if (to_mucontext(context)->bfregi.lib_uar_dyn) {
808 		err = -EINVAL;
809 		goto err_cqb;
810 	} else {
811 		*index = to_mucontext(context)->bfregi.sys_pages[0];
812 	}
813 
814 	return 0;
815 
816 err_cqb:
817 	kvfree(*cqb);
818 
819 err_db:
820 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
821 
822 err_umem:
823 	ib_umem_release(cq->buf.umem);
824 	return err;
825 }
826 
827 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
828 {
829 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
830 	ib_umem_release(cq->buf.umem);
831 }
832 
833 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
834 {
835 	int i;
836 	void *cqe;
837 	struct mlx5_cqe64 *cqe64;
838 
839 	for (i = 0; i < buf->nent; i++) {
840 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
841 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
842 		cqe64->op_own = MLX5_CQE_INVALID << 4;
843 	}
844 }
845 
846 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
847 			    int entries, int cqe_size,
848 			    u32 **cqb, int *index, int *inlen)
849 {
850 	__be64 *pas;
851 	void *cqc;
852 	int err;
853 
854 	err = mlx5_db_alloc(dev->mdev, &cq->db);
855 	if (err)
856 		return err;
857 
858 	cq->mcq.set_ci_db  = cq->db.db;
859 	cq->mcq.arm_db     = cq->db.db + 1;
860 	cq->mcq.cqe_sz = cqe_size;
861 
862 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
863 	if (err)
864 		goto err_db;
865 
866 	init_cq_buf(cq, &cq->buf);
867 
868 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
869 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
870 	*cqb = mlx5_vzalloc(*inlen);
871 	if (!*cqb) {
872 		err = -ENOMEM;
873 		goto err_buf;
874 	}
875 
876 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
877 	mlx5_fill_page_array(&cq->buf.buf, pas);
878 
879 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
880 	MLX5_SET(cqc, cqc, log_page_size,
881 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
882 
883 	*index = dev->mdev->priv.uar->index;
884 
885 	return 0;
886 
887 err_buf:
888 	free_cq_buf(dev, &cq->buf);
889 
890 err_db:
891 	mlx5_db_free(dev->mdev, &cq->db);
892 	return err;
893 }
894 
895 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
896 {
897 	free_cq_buf(dev, &cq->buf);
898 	mlx5_db_free(dev->mdev, &cq->db);
899 }
900 
901 static void notify_soft_wc_handler(struct work_struct *work)
902 {
903 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
904 					     notify_work);
905 
906 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
907 }
908 
909 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
910 				const struct ib_cq_init_attr *attr,
911 				struct ib_ucontext *context,
912 				struct ib_udata *udata)
913 {
914 	int entries = attr->cqe;
915 	int vector = attr->comp_vector;
916 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
917 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
918 	struct mlx5_ib_cq *cq;
919 	int uninitialized_var(index);
920 	int uninitialized_var(inlen);
921 	u32 *cqb = NULL;
922 	void *cqc;
923 	int cqe_size;
924 	unsigned int irqn;
925 	int eqn;
926 	int err;
927 
928 	if (entries < 0 ||
929 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
930 		return ERR_PTR(-EINVAL);
931 
932 	if (check_cq_create_flags(attr->flags))
933 		return ERR_PTR(-EOPNOTSUPP);
934 
935 	entries = roundup_pow_of_two(entries + 1);
936 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
937 		return ERR_PTR(-EINVAL);
938 
939 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
940 	if (!cq)
941 		return ERR_PTR(-ENOMEM);
942 
943 	cq->ibcq.cqe = entries - 1;
944 	mutex_init(&cq->resize_mutex);
945 	spin_lock_init(&cq->lock);
946 	cq->resize_buf = NULL;
947 	cq->resize_umem = NULL;
948 	cq->create_flags = attr->flags;
949 	INIT_LIST_HEAD(&cq->list_send_qp);
950 	INIT_LIST_HEAD(&cq->list_recv_qp);
951 
952 	if (context) {
953 		err = create_cq_user(dev, udata, context, cq, entries,
954 				     &cqb, &cqe_size, &index, &inlen);
955 		if (err)
956 			goto err_create;
957 	} else {
958 		cqe_size = cache_line_size() == 128 ? 128 : 64;
959 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
960 				       &index, &inlen);
961 		if (err)
962 			goto err_create;
963 
964 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
965 	}
966 
967 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
968 	if (err)
969 		goto err_cqb;
970 
971 	cq->cqe_size = cqe_size;
972 
973 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
974 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
975 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
976 	MLX5_SET(cqc, cqc, uar_page, index);
977 	MLX5_SET(cqc, cqc, c_eqn, eqn);
978 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
979 	if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
980 		MLX5_SET(cqc, cqc, oi, 1);
981 
982 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
983 	if (err)
984 		goto err_cqb;
985 
986 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
987 	cq->mcq.irqn = irqn;
988 	cq->mcq.comp  = mlx5_ib_cq_comp;
989 	cq->mcq.event = mlx5_ib_cq_event;
990 
991 	INIT_LIST_HEAD(&cq->wc_list);
992 
993 	if (context)
994 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
995 			err = -EFAULT;
996 			goto err_cmd;
997 		}
998 
999 
1000 	kvfree(cqb);
1001 	return &cq->ibcq;
1002 
1003 err_cmd:
1004 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1005 
1006 err_cqb:
1007 	kvfree(cqb);
1008 	if (context)
1009 		destroy_cq_user(cq, context);
1010 	else
1011 		destroy_cq_kernel(dev, cq);
1012 
1013 err_create:
1014 	kfree(cq);
1015 
1016 	return ERR_PTR(err);
1017 }
1018 
1019 
1020 int mlx5_ib_destroy_cq(struct ib_cq *cq)
1021 {
1022 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1023 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1024 	struct ib_ucontext *context = NULL;
1025 
1026 	if (cq->uobject)
1027 		context = cq->uobject->context;
1028 
1029 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1030 	if (context)
1031 		destroy_cq_user(mcq, context);
1032 	else
1033 		destroy_cq_kernel(dev, mcq);
1034 
1035 	kfree(mcq);
1036 
1037 	return 0;
1038 }
1039 
1040 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1041 {
1042 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1043 }
1044 
1045 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1046 {
1047 	struct mlx5_cqe64 *cqe64, *dest64;
1048 	void *cqe, *dest;
1049 	u32 prod_index;
1050 	int nfreed = 0;
1051 	u8 owner_bit;
1052 
1053 	if (!cq)
1054 		return;
1055 
1056 	/* First we need to find the current producer index, so we
1057 	 * know where to start cleaning from.  It doesn't matter if HW
1058 	 * adds new entries after this loop -- the QP we're worried
1059 	 * about is already in RESET, so the new entries won't come
1060 	 * from our QP and therefore don't need to be checked.
1061 	 */
1062 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1063 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1064 			break;
1065 
1066 	/* Now sweep backwards through the CQ, removing CQ entries
1067 	 * that match our QP by copying older entries on top of them.
1068 	 */
1069 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1070 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1071 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1072 		if (is_equal_rsn(cqe64, rsn)) {
1073 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1074 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1075 			++nfreed;
1076 		} else if (nfreed) {
1077 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1078 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1079 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1080 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1081 			dest64->op_own = owner_bit |
1082 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1083 		}
1084 	}
1085 
1086 	if (nfreed) {
1087 		cq->mcq.cons_index += nfreed;
1088 		/* Make sure update of buffer contents is done before
1089 		 * updating consumer index.
1090 		 */
1091 		wmb();
1092 		mlx5_cq_set_ci(&cq->mcq);
1093 	}
1094 }
1095 
1096 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1097 {
1098 	if (!cq)
1099 		return;
1100 
1101 	spin_lock_irq(&cq->lock);
1102 	__mlx5_ib_cq_clean(cq, qpn, srq);
1103 	spin_unlock_irq(&cq->lock);
1104 }
1105 
1106 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1107 {
1108 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1109 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1110 	int err;
1111 
1112 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1113 		return -ENOSYS;
1114 
1115 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1116 					     cq_period, cq_count);
1117 	if (err)
1118 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1119 
1120 	return err;
1121 }
1122 
1123 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1124 		       int entries, struct ib_udata *udata, int *npas,
1125 		       int *page_shift, int *cqe_size)
1126 {
1127 	struct mlx5_ib_resize_cq ucmd;
1128 	struct ib_umem *umem;
1129 	int err;
1130 	int npages;
1131 	struct ib_ucontext *context = cq->buf.umem->context;
1132 
1133 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1134 	if (err)
1135 		return err;
1136 
1137 	if (ucmd.reserved0 || ucmd.reserved1)
1138 		return -EINVAL;
1139 
1140 	/* check multiplication overflow */
1141 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1142 		return -EINVAL;
1143 
1144 	umem = ib_umem_get(context, ucmd.buf_addr,
1145 			   (size_t)ucmd.cqe_size * entries,
1146 			   IB_ACCESS_LOCAL_WRITE, 1);
1147 	if (IS_ERR(umem)) {
1148 		err = PTR_ERR(umem);
1149 		return err;
1150 	}
1151 
1152 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1153 			   npas, NULL);
1154 
1155 	cq->resize_umem = umem;
1156 	*cqe_size = ucmd.cqe_size;
1157 
1158 	return 0;
1159 }
1160 
1161 static void un_resize_user(struct mlx5_ib_cq *cq)
1162 {
1163 	ib_umem_release(cq->resize_umem);
1164 }
1165 
1166 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1167 			 int entries, int cqe_size)
1168 {
1169 	int err;
1170 
1171 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1172 	if (!cq->resize_buf)
1173 		return -ENOMEM;
1174 
1175 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1176 	if (err)
1177 		goto ex;
1178 
1179 	init_cq_buf(cq, cq->resize_buf);
1180 
1181 	return 0;
1182 
1183 ex:
1184 	kfree(cq->resize_buf);
1185 	return err;
1186 }
1187 
1188 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1189 {
1190 	free_cq_buf(dev, cq->resize_buf);
1191 	cq->resize_buf = NULL;
1192 }
1193 
1194 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1195 {
1196 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1197 	struct mlx5_cqe64 *scqe64;
1198 	struct mlx5_cqe64 *dcqe64;
1199 	void *start_cqe;
1200 	void *scqe;
1201 	void *dcqe;
1202 	int ssize;
1203 	int dsize;
1204 	int i;
1205 	u8 sw_own;
1206 
1207 	ssize = cq->buf.cqe_size;
1208 	dsize = cq->resize_buf->cqe_size;
1209 	if (ssize != dsize) {
1210 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1211 		return -EINVAL;
1212 	}
1213 
1214 	i = cq->mcq.cons_index;
1215 	scqe = get_sw_cqe(cq, i);
1216 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1217 	start_cqe = scqe;
1218 	if (!scqe) {
1219 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1220 		return -EINVAL;
1221 	}
1222 
1223 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1224 		dcqe = get_cqe_from_buf(cq->resize_buf,
1225 					(i + 1) & (cq->resize_buf->nent),
1226 					dsize);
1227 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1228 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1229 		memcpy(dcqe, scqe, dsize);
1230 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1231 
1232 		++i;
1233 		scqe = get_sw_cqe(cq, i);
1234 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1235 		if (!scqe) {
1236 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1237 			return -EINVAL;
1238 		}
1239 
1240 		if (scqe == start_cqe) {
1241 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1242 				cq->mcq.cqn);
1243 			return -ENOMEM;
1244 		}
1245 	}
1246 	++cq->mcq.cons_index;
1247 	return 0;
1248 }
1249 
1250 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1251 {
1252 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1253 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1254 	void *cqc;
1255 	u32 *in;
1256 	int err;
1257 	int npas;
1258 	__be64 *pas;
1259 	int page_shift;
1260 	int inlen;
1261 	int uninitialized_var(cqe_size);
1262 	unsigned long flags;
1263 
1264 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1265 		pr_info("Firmware does not support resize CQ\n");
1266 		return -ENOSYS;
1267 	}
1268 
1269 	if (entries < 1 ||
1270 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1271 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1272 			     entries,
1273 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1274 		return -EINVAL;
1275 	}
1276 
1277 	entries = roundup_pow_of_two(entries + 1);
1278 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1279 		return -EINVAL;
1280 
1281 	if (entries == ibcq->cqe + 1)
1282 		return 0;
1283 
1284 	mutex_lock(&cq->resize_mutex);
1285 	if (udata) {
1286 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1287 				  &cqe_size);
1288 	} else {
1289 		cqe_size = 64;
1290 		err = resize_kernel(dev, cq, entries, cqe_size);
1291 		if (!err) {
1292 			npas = cq->resize_buf->buf.npages;
1293 			page_shift = cq->resize_buf->buf.page_shift;
1294 		}
1295 	}
1296 
1297 	if (err)
1298 		goto ex;
1299 
1300 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1301 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1302 
1303 	in = mlx5_vzalloc(inlen);
1304 	if (!in) {
1305 		err = -ENOMEM;
1306 		goto ex_resize;
1307 	}
1308 
1309 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1310 	if (udata)
1311 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1312 				     pas, 0);
1313 	else
1314 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1315 
1316 	MLX5_SET(modify_cq_in, in,
1317 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1318 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1319 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1320 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1321 
1322 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1323 
1324 	MLX5_SET(cqc, cqc, log_page_size,
1325 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1326 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1327 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1328 
1329 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1330 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1331 
1332 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1333 	if (err)
1334 		goto ex_alloc;
1335 
1336 	if (udata) {
1337 		cq->ibcq.cqe = entries - 1;
1338 		ib_umem_release(cq->buf.umem);
1339 		cq->buf.umem = cq->resize_umem;
1340 		cq->resize_umem = NULL;
1341 	} else {
1342 		struct mlx5_ib_cq_buf tbuf;
1343 		int resized = 0;
1344 
1345 		spin_lock_irqsave(&cq->lock, flags);
1346 		if (cq->resize_buf) {
1347 			err = copy_resize_cqes(cq);
1348 			if (!err) {
1349 				tbuf = cq->buf;
1350 				cq->buf = *cq->resize_buf;
1351 				kfree(cq->resize_buf);
1352 				cq->resize_buf = NULL;
1353 				resized = 1;
1354 			}
1355 		}
1356 		cq->ibcq.cqe = entries - 1;
1357 		spin_unlock_irqrestore(&cq->lock, flags);
1358 		if (resized)
1359 			free_cq_buf(dev, &tbuf);
1360 	}
1361 	mutex_unlock(&cq->resize_mutex);
1362 
1363 	kvfree(in);
1364 	return 0;
1365 
1366 ex_alloc:
1367 	kvfree(in);
1368 
1369 ex_resize:
1370 	if (udata)
1371 		un_resize_user(cq);
1372 	else
1373 		un_resize_kernel(dev, cq);
1374 ex:
1375 	mutex_unlock(&cq->resize_mutex);
1376 	return err;
1377 }
1378 
1379 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1380 {
1381 	struct mlx5_ib_cq *cq;
1382 
1383 	if (!ibcq)
1384 		return 128;
1385 
1386 	cq = to_mcq(ibcq);
1387 	return cq->cqe_size;
1388 }
1389 
1390 /* Called from atomic context */
1391 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1392 {
1393 	struct mlx5_ib_wc *soft_wc;
1394 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1395 	unsigned long flags;
1396 
1397 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1398 	if (!soft_wc)
1399 		return -ENOMEM;
1400 
1401 	soft_wc->wc = *wc;
1402 	spin_lock_irqsave(&cq->lock, flags);
1403 	list_add_tail(&soft_wc->list, &cq->wc_list);
1404 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1405 	    wc->status != IB_WC_SUCCESS) {
1406 		cq->notify_flags = 0;
1407 		schedule_work(&cq->notify_work);
1408 	}
1409 	spin_unlock_irqrestore(&cq->lock, flags);
1410 
1411 	return 0;
1412 }
1413