xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_cq.c (revision d06955f9bdb1416d9196043ed781f9b36dae9adc)
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)
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 g;
171 
172 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
173 		struct mlx5_core_srq *msrq = NULL;
174 
175 		if (qp->ibqp.xrcd) {
176 			msrq = mlx5_core_get_srq(dev->mdev,
177 						 be32_to_cpu(cqe->srqn));
178 			srq = to_mibsrq(msrq);
179 		} else {
180 			srq = to_msrq(qp->ibqp.srq);
181 		}
182 		if (srq) {
183 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
184 			wc->wr_id = srq->wrid[wqe_ctr];
185 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
186 			if (msrq && atomic_dec_and_test(&msrq->refcount))
187 				complete(&msrq->free);
188 		}
189 	} else {
190 		wq	  = &qp->rq;
191 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
192 		++wq->tail;
193 	}
194 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
195 
196 	switch (cqe->op_own >> 4) {
197 	case MLX5_CQE_RESP_WR_IMM:
198 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
199 		wc->wc_flags	= IB_WC_WITH_IMM;
200 		wc->ex.imm_data = cqe->imm_inval_pkey;
201 		break;
202 	case MLX5_CQE_RESP_SEND:
203 		wc->opcode   = IB_WC_RECV;
204 		wc->wc_flags = IB_WC_IP_CSUM_OK;
205 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
206 			       (cqe->hds_ip_ext & CQE_L4_OK))))
207 			wc->wc_flags = 0;
208 		break;
209 	case MLX5_CQE_RESP_SEND_IMM:
210 		wc->opcode	= IB_WC_RECV;
211 		wc->wc_flags	= IB_WC_WITH_IMM;
212 		wc->ex.imm_data = cqe->imm_inval_pkey;
213 		break;
214 	case MLX5_CQE_RESP_SEND_INV:
215 		wc->opcode	= IB_WC_RECV;
216 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
217 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
218 		break;
219 	}
220 	wc->slid	   = be16_to_cpu(cqe->slid);
221 	wc->sl		   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
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 		return;
237 
238 	switch (wc->sl & 0x3) {
239 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
240 		wc->network_hdr_type = RDMA_NETWORK_IB;
241 		break;
242 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
243 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
244 		break;
245 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
246 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
247 		break;
248 	}
249 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
250 }
251 
252 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
253 {
254 	__be32 *p = (__be32 *)cqe;
255 	int i;
256 
257 	mlx5_ib_warn(dev, "dump error cqe\n");
258 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
259 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
260 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
261 			be32_to_cpu(p[3]));
262 }
263 
264 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
265 				  struct mlx5_err_cqe *cqe,
266 				  struct ib_wc *wc)
267 {
268 	int dump = 1;
269 
270 	switch (cqe->syndrome) {
271 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
272 		wc->status = IB_WC_LOC_LEN_ERR;
273 		break;
274 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
275 		wc->status = IB_WC_LOC_QP_OP_ERR;
276 		break;
277 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
278 		wc->status = IB_WC_LOC_PROT_ERR;
279 		break;
280 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
281 		dump = 0;
282 		wc->status = IB_WC_WR_FLUSH_ERR;
283 		break;
284 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
285 		wc->status = IB_WC_MW_BIND_ERR;
286 		break;
287 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
288 		wc->status = IB_WC_BAD_RESP_ERR;
289 		break;
290 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
291 		wc->status = IB_WC_LOC_ACCESS_ERR;
292 		break;
293 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
294 		wc->status = IB_WC_REM_INV_REQ_ERR;
295 		break;
296 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
297 		wc->status = IB_WC_REM_ACCESS_ERR;
298 		break;
299 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
300 		wc->status = IB_WC_REM_OP_ERR;
301 		break;
302 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
303 		wc->status = IB_WC_RETRY_EXC_ERR;
304 		dump = 0;
305 		break;
306 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
307 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
308 		dump = 0;
309 		break;
310 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
311 		wc->status = IB_WC_REM_ABORT_ERR;
312 		break;
313 	default:
314 		wc->status = IB_WC_GENERAL_ERR;
315 		break;
316 	}
317 
318 	wc->vendor_err = cqe->vendor_err_synd;
319 	if (dump)
320 		dump_cqe(dev, cqe);
321 }
322 
323 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
324 {
325 	/* TBD: waiting decision
326 	*/
327 	return 0;
328 }
329 
330 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
331 {
332 	struct mlx5_wqe_data_seg *dpseg;
333 	void *addr;
334 
335 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
336 		sizeof(struct mlx5_wqe_raddr_seg) +
337 		sizeof(struct mlx5_wqe_atomic_seg);
338 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
339 	return addr;
340 }
341 
342 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
343 			  uint16_t idx)
344 {
345 	void *addr;
346 	int byte_count;
347 	int i;
348 
349 	if (!is_atomic_response(qp, idx))
350 		return;
351 
352 	byte_count = be32_to_cpu(cqe64->byte_cnt);
353 	addr = mlx5_get_atomic_laddr(qp, idx);
354 
355 	if (byte_count == 4) {
356 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
357 	} else {
358 		for (i = 0; i < byte_count; i += 8) {
359 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
360 			addr += 8;
361 		}
362 	}
363 
364 	return;
365 }
366 
367 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
368 			   u16 tail, u16 head)
369 {
370 	u16 idx;
371 
372 	do {
373 		idx = tail & (qp->sq.wqe_cnt - 1);
374 		handle_atomic(qp, cqe64, idx);
375 		if (idx == head)
376 			break;
377 
378 		tail = qp->sq.w_list[idx].next;
379 	} while (1);
380 	tail = qp->sq.w_list[idx].next;
381 	qp->sq.last_poll = tail;
382 }
383 
384 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
385 {
386 	mlx5_buf_free(dev->mdev, &buf->buf);
387 }
388 
389 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
390 			     struct ib_sig_err *item)
391 {
392 	u16 syndrome = be16_to_cpu(cqe->syndrome);
393 
394 #define GUARD_ERR   (1 << 13)
395 #define APPTAG_ERR  (1 << 12)
396 #define REFTAG_ERR  (1 << 11)
397 
398 	if (syndrome & GUARD_ERR) {
399 		item->err_type = IB_SIG_BAD_GUARD;
400 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
401 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
402 	} else
403 	if (syndrome & REFTAG_ERR) {
404 		item->err_type = IB_SIG_BAD_REFTAG;
405 		item->expected = be32_to_cpu(cqe->expected_reftag);
406 		item->actual = be32_to_cpu(cqe->actual_reftag);
407 	} else
408 	if (syndrome & APPTAG_ERR) {
409 		item->err_type = IB_SIG_BAD_APPTAG;
410 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
411 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
412 	} else {
413 		pr_err("Got signature completion error with bad syndrome %04x\n",
414 		       syndrome);
415 	}
416 
417 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
418 	item->key = be32_to_cpu(cqe->mkey);
419 }
420 
421 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
422 			 struct ib_wc *wc, int *npolled)
423 {
424 	struct mlx5_ib_wq *wq;
425 	unsigned int cur;
426 	unsigned int idx;
427 	int np;
428 	int i;
429 
430 	wq = &qp->sq;
431 	cur = wq->head - wq->tail;
432 	np = *npolled;
433 
434 	if (cur == 0)
435 		return;
436 
437 	for (i = 0;  i < cur && np < num_entries; i++) {
438 		idx = wq->last_poll & (wq->wqe_cnt - 1);
439 		wc->wr_id = wq->wrid[idx];
440 		wc->status = IB_WC_WR_FLUSH_ERR;
441 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
442 		wq->tail++;
443 		np++;
444 		wc->qp = &qp->ibqp;
445 		wc++;
446 		wq->last_poll = wq->w_list[idx].next;
447 	}
448 	*npolled = np;
449 }
450 
451 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
452 			 struct ib_wc *wc, int *npolled)
453 {
454 	struct mlx5_ib_wq *wq;
455 	unsigned int cur;
456 	int np;
457 	int i;
458 
459 	wq = &qp->rq;
460 	cur = wq->head - wq->tail;
461 	np = *npolled;
462 
463 	if (cur == 0)
464 		return;
465 
466 	for (i = 0;  i < cur && np < num_entries; i++) {
467 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
468 		wc->status = IB_WC_WR_FLUSH_ERR;
469 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
470 		wq->tail++;
471 		np++;
472 		wc->qp = &qp->ibqp;
473 		wc++;
474 	}
475 	*npolled = np;
476 }
477 
478 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
479 				 struct ib_wc *wc, int *npolled)
480 {
481 	struct mlx5_ib_qp *qp;
482 
483 	*npolled = 0;
484 	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
485 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
486 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
487 		if (*npolled >= num_entries)
488 			return;
489 	}
490 
491 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
492 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
493 		if (*npolled >= num_entries)
494 			return;
495 	}
496 }
497 
498 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
499 			 struct mlx5_ib_qp **cur_qp,
500 			 struct ib_wc *wc)
501 {
502 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
503 	struct mlx5_err_cqe *err_cqe;
504 	struct mlx5_cqe64 *cqe64;
505 	struct mlx5_core_qp *mqp;
506 	struct mlx5_ib_wq *wq;
507 	struct mlx5_sig_err_cqe *sig_err_cqe;
508 	struct mlx5_core_mr *mmkey;
509 	struct mlx5_ib_mr *mr;
510 	unsigned long flags;
511 	uint8_t opcode;
512 	uint32_t qpn;
513 	u16 wqe_ctr;
514 	void *cqe;
515 	int idx;
516 
517 repoll:
518 	cqe = next_cqe_sw(cq);
519 	if (!cqe)
520 		return -EAGAIN;
521 
522 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
523 
524 	++cq->mcq.cons_index;
525 
526 	/* Make sure we read CQ entry contents after we've checked the
527 	 * ownership bit.
528 	 */
529 	rmb();
530 
531 	opcode = cqe64->op_own >> 4;
532 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
533 		if (likely(cq->resize_buf)) {
534 			free_cq_buf(dev, &cq->buf);
535 			cq->buf = *cq->resize_buf;
536 			kfree(cq->resize_buf);
537 			cq->resize_buf = NULL;
538 			goto repoll;
539 		} else {
540 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
541 		}
542 	}
543 
544 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
545 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
546 		/* We do not have to take the QP table lock here,
547 		 * because CQs will be locked while QPs are removed
548 		 * from the table.
549 		 */
550 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
551 		*cur_qp = to_mibqp(mqp);
552 	}
553 
554 	wc->qp  = &(*cur_qp)->ibqp;
555 	switch (opcode) {
556 	case MLX5_CQE_REQ:
557 		wq = &(*cur_qp)->sq;
558 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
559 		idx = wqe_ctr & (wq->wqe_cnt - 1);
560 		handle_good_req(wc, cqe64, wq, idx);
561 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
562 		wc->wr_id = wq->wrid[idx];
563 		wq->tail = wq->wqe_head[idx] + 1;
564 		wc->status = IB_WC_SUCCESS;
565 		break;
566 	case MLX5_CQE_RESP_WR_IMM:
567 	case MLX5_CQE_RESP_SEND:
568 	case MLX5_CQE_RESP_SEND_IMM:
569 	case MLX5_CQE_RESP_SEND_INV:
570 		handle_responder(wc, cqe64, *cur_qp);
571 		wc->status = IB_WC_SUCCESS;
572 		break;
573 	case MLX5_CQE_RESIZE_CQ:
574 		break;
575 	case MLX5_CQE_REQ_ERR:
576 	case MLX5_CQE_RESP_ERR:
577 		err_cqe = (struct mlx5_err_cqe *)cqe64;
578 		mlx5_handle_error_cqe(dev, err_cqe, wc);
579 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
580 			    opcode == MLX5_CQE_REQ_ERR ?
581 			    "Requestor" : "Responder", cq->mcq.cqn);
582 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
583 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
584 		if (opcode == MLX5_CQE_REQ_ERR) {
585 			wq = &(*cur_qp)->sq;
586 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
587 			idx = wqe_ctr & (wq->wqe_cnt - 1);
588 			wc->wr_id = wq->wrid[idx];
589 			wq->tail = wq->wqe_head[idx] + 1;
590 		} else {
591 			struct mlx5_ib_srq *srq;
592 
593 			if ((*cur_qp)->ibqp.srq) {
594 				srq = to_msrq((*cur_qp)->ibqp.srq);
595 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
596 				wc->wr_id = srq->wrid[wqe_ctr];
597 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
598 			} else {
599 				wq = &(*cur_qp)->rq;
600 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
601 				++wq->tail;
602 			}
603 		}
604 		break;
605 	case MLX5_CQE_SIG_ERR:
606 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
607 
608 		spin_lock_irqsave(&dev->mdev->priv.mr_table.lock, flags);
609 		mmkey = __mlx5_mr_lookup(dev->mdev,
610 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
611 		mr = to_mibmr(mmkey);
612 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
613 		mr->sig->sig_err_exists = true;
614 		mr->sig->sigerr_count++;
615 
616 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
617 			     cq->mcq.cqn, mr->sig->err_item.key,
618 			     mr->sig->err_item.err_type,
619 			     (long long)mr->sig->err_item.sig_err_offset,
620 			     mr->sig->err_item.expected,
621 			     mr->sig->err_item.actual);
622 
623 		spin_unlock_irqrestore(&dev->mdev->priv.mr_table.lock, flags);
624 		goto repoll;
625 	}
626 
627 	return 0;
628 }
629 
630 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
631 			struct ib_wc *wc)
632 {
633 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
634 	struct mlx5_ib_wc *soft_wc, *next;
635 	int npolled = 0;
636 
637 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
638 		if (npolled >= num_entries)
639 			break;
640 
641 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
642 			    cq->mcq.cqn);
643 
644 		wc[npolled++] = soft_wc->wc;
645 		list_del(&soft_wc->list);
646 		kfree(soft_wc);
647 	}
648 
649 	return npolled;
650 }
651 
652 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
653 {
654 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
655 	struct mlx5_ib_qp *cur_qp = NULL;
656 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
657 	struct mlx5_core_dev *mdev = dev->mdev;
658 	unsigned long flags;
659 	int soft_polled = 0;
660 	int npolled;
661 
662 	spin_lock_irqsave(&cq->lock, flags);
663 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
664 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
665 		goto out;
666 	}
667 
668 	if (unlikely(!list_empty(&cq->wc_list)))
669 		soft_polled = poll_soft_wc(cq, num_entries, wc);
670 
671 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
672 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
673 			break;
674 	}
675 
676 	if (npolled)
677 		mlx5_cq_set_ci(&cq->mcq);
678 out:
679 	spin_unlock_irqrestore(&cq->lock, flags);
680 
681 	return soft_polled + npolled;
682 }
683 
684 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
685 {
686 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
687 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
688 	void __iomem *uar_page = mdev->priv.uuari.uars[0].map;
689 	unsigned long irq_flags;
690 	int ret = 0;
691 
692 	spin_lock_irqsave(&cq->lock, irq_flags);
693 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
694 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
695 
696 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
697 		ret = 1;
698 	spin_unlock_irqrestore(&cq->lock, irq_flags);
699 
700 	mlx5_cq_arm(&cq->mcq,
701 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
702 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
703 		    uar_page,
704 		    MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
705 		    to_mcq(ibcq)->mcq.cons_index);
706 
707 	return ret;
708 }
709 
710 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
711 			int nent, int cqe_size)
712 {
713 	int err;
714 
715 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
716 	    2 * PAGE_SIZE, &buf->buf);
717 	if (err)
718 		return err;
719 
720 	buf->cqe_size = cqe_size;
721 	buf->nent = nent;
722 
723 	return 0;
724 }
725 
726 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
727 			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
728 			  int entries, u32 **cqb,
729 			  int *cqe_size, int *index, int *inlen)
730 {
731 	struct mlx5_ib_create_cq ucmd;
732 	size_t ucmdlen;
733 	int page_shift;
734 	__be64 *pas;
735 	int npages;
736 	int ncont;
737 	void *cqc;
738 	int err;
739 
740 	ucmdlen =
741 		(udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
742 		 sizeof(ucmd)) ? (sizeof(ucmd) -
743 				  sizeof(ucmd.reserved)) : sizeof(ucmd);
744 
745 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
746 		return -EFAULT;
747 
748 	if (ucmdlen == sizeof(ucmd) &&
749 	    ucmd.reserved != 0)
750 		return -EINVAL;
751 
752 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
753 		return -EINVAL;
754 
755 	*cqe_size = ucmd.cqe_size;
756 
757 	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
758 				   entries * ucmd.cqe_size,
759 				   IB_ACCESS_LOCAL_WRITE, 1);
760 	if (IS_ERR(cq->buf.umem)) {
761 		err = PTR_ERR(cq->buf.umem);
762 		return err;
763 	}
764 
765 	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
766 				  &cq->db);
767 	if (err)
768 		goto err_umem;
769 
770 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
771 			   &ncont, NULL);
772 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
773 		    (long long)ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
774 
775 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
776 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
777 	*cqb = mlx5_vzalloc(*inlen);
778 	if (!*cqb) {
779 		err = -ENOMEM;
780 		goto err_db;
781 	}
782 
783 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
784 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
785 
786 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
787 	MLX5_SET(cqc, cqc, log_page_size,
788 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
789 
790 	*index = to_mucontext(context)->uuari.uars[0].index;
791 
792 	return 0;
793 
794 err_db:
795 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
796 
797 err_umem:
798 	ib_umem_release(cq->buf.umem);
799 	return err;
800 }
801 
802 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
803 {
804 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
805 	ib_umem_release(cq->buf.umem);
806 }
807 
808 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
809 {
810 	int i;
811 	void *cqe;
812 	struct mlx5_cqe64 *cqe64;
813 
814 	for (i = 0; i < buf->nent; i++) {
815 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
816 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
817 		cqe64->op_own = MLX5_CQE_INVALID << 4;
818 	}
819 }
820 
821 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
822 			    int entries, int cqe_size,
823 			    u32 **cqb, int *index, int *inlen)
824 {
825 	__be64 *pas;
826 	void *cqc;
827 	int err;
828 
829 	err = mlx5_db_alloc(dev->mdev, &cq->db);
830 	if (err)
831 		return err;
832 
833 	cq->mcq.set_ci_db  = cq->db.db;
834 	cq->mcq.arm_db     = cq->db.db + 1;
835 	cq->mcq.cqe_sz = cqe_size;
836 
837 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
838 	if (err)
839 		goto err_db;
840 
841 	init_cq_buf(cq, &cq->buf);
842 
843 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
844 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
845 	*cqb = mlx5_vzalloc(*inlen);
846 	if (!*cqb) {
847 		err = -ENOMEM;
848 		goto err_buf;
849 	}
850 
851 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
852 	mlx5_fill_page_array(&cq->buf.buf, pas);
853 
854 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
855 	MLX5_SET(cqc, cqc, log_page_size,
856 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
857 
858 	*index = dev->mdev->priv.uuari.uars[0].index;
859 
860 	return 0;
861 
862 err_buf:
863 	free_cq_buf(dev, &cq->buf);
864 
865 err_db:
866 	mlx5_db_free(dev->mdev, &cq->db);
867 	return err;
868 }
869 
870 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
871 {
872 	free_cq_buf(dev, &cq->buf);
873 	mlx5_db_free(dev->mdev, &cq->db);
874 }
875 
876 static void notify_soft_wc_handler(struct work_struct *work)
877 {
878 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
879 					     notify_work);
880 
881 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
882 }
883 
884 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
885 				const struct ib_cq_init_attr *attr,
886 				struct ib_ucontext *context,
887 				struct ib_udata *udata)
888 {
889 	int entries = attr->cqe;
890 	int vector = attr->comp_vector;
891 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
892 	struct mlx5_ib_cq *cq;
893 	int uninitialized_var(index);
894 	int uninitialized_var(inlen);
895 	u32 *cqb = NULL;
896 	void *cqc;
897 	int cqe_size;
898 	unsigned int irqn;
899 	int eqn;
900 	int err;
901 
902 	if (entries < 0 ||
903 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
904 		return ERR_PTR(-EINVAL);
905 
906 	if (check_cq_create_flags(attr->flags))
907 		return ERR_PTR(-EOPNOTSUPP);
908 
909 	entries = roundup_pow_of_two(entries + 1);
910 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
911 		return ERR_PTR(-EINVAL);
912 
913 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
914 	if (!cq)
915 		return ERR_PTR(-ENOMEM);
916 
917 	cq->ibcq.cqe = entries - 1;
918 	mutex_init(&cq->resize_mutex);
919 	spin_lock_init(&cq->lock);
920 	cq->resize_buf = NULL;
921 	cq->resize_umem = NULL;
922 	cq->create_flags = attr->flags;
923 	INIT_LIST_HEAD(&cq->list_send_qp);
924 	INIT_LIST_HEAD(&cq->list_recv_qp);
925 
926 	if (context) {
927 		err = create_cq_user(dev, udata, context, cq, entries,
928 				     &cqb, &cqe_size, &index, &inlen);
929 		if (err)
930 			goto err_create;
931 	} else {
932 		cqe_size = cache_line_size() == 128 ? 128 : 64;
933 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
934 				       &index, &inlen);
935 		if (err)
936 			goto err_create;
937 
938 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
939 	}
940 
941 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
942 	if (err)
943 		goto err_cqb;
944 
945 	cq->cqe_size = cqe_size;
946 
947 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
948 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
949 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
950 	MLX5_SET(cqc, cqc, uar_page, index);
951 	MLX5_SET(cqc, cqc, c_eqn, eqn);
952 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
953 	if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
954 		MLX5_SET(cqc, cqc, oi, 1);
955 
956 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq,
957 	    (struct mlx5_create_cq_mbox_in *)cqb, inlen);
958 	if (err)
959 		goto err_cqb;
960 
961 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
962 	cq->mcq.irqn = irqn;
963 	cq->mcq.comp  = mlx5_ib_cq_comp;
964 	cq->mcq.event = mlx5_ib_cq_event;
965 
966 	INIT_LIST_HEAD(&cq->wc_list);
967 
968 	if (context)
969 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
970 			err = -EFAULT;
971 			goto err_cmd;
972 		}
973 
974 
975 	kvfree(cqb);
976 	return &cq->ibcq;
977 
978 err_cmd:
979 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
980 
981 err_cqb:
982 	kvfree(cqb);
983 	if (context)
984 		destroy_cq_user(cq, context);
985 	else
986 		destroy_cq_kernel(dev, cq);
987 
988 err_create:
989 	kfree(cq);
990 
991 	return ERR_PTR(err);
992 }
993 
994 
995 int mlx5_ib_destroy_cq(struct ib_cq *cq)
996 {
997 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
998 	struct mlx5_ib_cq *mcq = to_mcq(cq);
999 	struct ib_ucontext *context = NULL;
1000 
1001 	if (cq->uobject)
1002 		context = cq->uobject->context;
1003 
1004 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1005 	if (context)
1006 		destroy_cq_user(mcq, context);
1007 	else
1008 		destroy_cq_kernel(dev, mcq);
1009 
1010 	kfree(mcq);
1011 
1012 	return 0;
1013 }
1014 
1015 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1016 {
1017 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1018 }
1019 
1020 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1021 {
1022 	struct mlx5_cqe64 *cqe64, *dest64;
1023 	void *cqe, *dest;
1024 	u32 prod_index;
1025 	int nfreed = 0;
1026 	u8 owner_bit;
1027 
1028 	if (!cq)
1029 		return;
1030 
1031 	/* First we need to find the current producer index, so we
1032 	 * know where to start cleaning from.  It doesn't matter if HW
1033 	 * adds new entries after this loop -- the QP we're worried
1034 	 * about is already in RESET, so the new entries won't come
1035 	 * from our QP and therefore don't need to be checked.
1036 	 */
1037 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1038 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1039 			break;
1040 
1041 	/* Now sweep backwards through the CQ, removing CQ entries
1042 	 * that match our QP by copying older entries on top of them.
1043 	 */
1044 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1045 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1046 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1047 		if (is_equal_rsn(cqe64, rsn)) {
1048 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1049 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1050 			++nfreed;
1051 		} else if (nfreed) {
1052 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1053 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1054 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1055 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1056 			dest64->op_own = owner_bit |
1057 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1058 		}
1059 	}
1060 
1061 	if (nfreed) {
1062 		cq->mcq.cons_index += nfreed;
1063 		/* Make sure update of buffer contents is done before
1064 		 * updating consumer index.
1065 		 */
1066 		wmb();
1067 		mlx5_cq_set_ci(&cq->mcq);
1068 	}
1069 }
1070 
1071 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1072 {
1073 	if (!cq)
1074 		return;
1075 
1076 	spin_lock_irq(&cq->lock);
1077 	__mlx5_ib_cq_clean(cq, qpn, srq);
1078 	spin_unlock_irq(&cq->lock);
1079 }
1080 
1081 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1082 {
1083 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1084 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1085 	int err;
1086 
1087 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1088 		return -ENOSYS;
1089 
1090 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1091 					     cq_period, cq_count);
1092 	if (err)
1093 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1094 
1095 	return err;
1096 }
1097 
1098 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1099 		       int entries, struct ib_udata *udata, int *npas,
1100 		       int *page_shift, int *cqe_size)
1101 {
1102 	struct mlx5_ib_resize_cq ucmd;
1103 	struct ib_umem *umem;
1104 	int err;
1105 	int npages;
1106 	struct ib_ucontext *context = cq->buf.umem->context;
1107 
1108 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1109 	if (err)
1110 		return err;
1111 
1112 	if (ucmd.reserved0 || ucmd.reserved1)
1113 		return -EINVAL;
1114 
1115 	umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
1116 			   IB_ACCESS_LOCAL_WRITE, 1);
1117 	if (IS_ERR(umem)) {
1118 		err = PTR_ERR(umem);
1119 		return err;
1120 	}
1121 
1122 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, &npages, page_shift,
1123 			   npas, NULL);
1124 
1125 	cq->resize_umem = umem;
1126 	*cqe_size = ucmd.cqe_size;
1127 
1128 	return 0;
1129 }
1130 
1131 static void un_resize_user(struct mlx5_ib_cq *cq)
1132 {
1133 	ib_umem_release(cq->resize_umem);
1134 }
1135 
1136 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1137 			 int entries, int cqe_size)
1138 {
1139 	int err;
1140 
1141 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1142 	if (!cq->resize_buf)
1143 		return -ENOMEM;
1144 
1145 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1146 	if (err)
1147 		goto ex;
1148 
1149 	init_cq_buf(cq, cq->resize_buf);
1150 
1151 	return 0;
1152 
1153 ex:
1154 	kfree(cq->resize_buf);
1155 	return err;
1156 }
1157 
1158 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1159 {
1160 	free_cq_buf(dev, cq->resize_buf);
1161 	cq->resize_buf = NULL;
1162 }
1163 
1164 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1165 {
1166 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1167 	struct mlx5_cqe64 *scqe64;
1168 	struct mlx5_cqe64 *dcqe64;
1169 	void *start_cqe;
1170 	void *scqe;
1171 	void *dcqe;
1172 	int ssize;
1173 	int dsize;
1174 	int i;
1175 	u8 sw_own;
1176 
1177 	ssize = cq->buf.cqe_size;
1178 	dsize = cq->resize_buf->cqe_size;
1179 	if (ssize != dsize) {
1180 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1181 		return -EINVAL;
1182 	}
1183 
1184 	i = cq->mcq.cons_index;
1185 	scqe = get_sw_cqe(cq, i);
1186 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1187 	start_cqe = scqe;
1188 	if (!scqe) {
1189 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1190 		return -EINVAL;
1191 	}
1192 
1193 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1194 		dcqe = get_cqe_from_buf(cq->resize_buf,
1195 					(i + 1) & (cq->resize_buf->nent),
1196 					dsize);
1197 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1198 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1199 		memcpy(dcqe, scqe, dsize);
1200 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1201 
1202 		++i;
1203 		scqe = get_sw_cqe(cq, i);
1204 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1205 		if (!scqe) {
1206 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1207 			return -EINVAL;
1208 		}
1209 
1210 		if (scqe == start_cqe) {
1211 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1212 				cq->mcq.cqn);
1213 			return -ENOMEM;
1214 		}
1215 	}
1216 	++cq->mcq.cons_index;
1217 	return 0;
1218 }
1219 
1220 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1221 {
1222 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1223 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1224 	void *cqc;
1225 	u32 *in;
1226 	int err;
1227 	int npas;
1228 	__be64 *pas;
1229 	int page_shift;
1230 	int inlen;
1231 	int uninitialized_var(cqe_size);
1232 	unsigned long flags;
1233 
1234 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1235 		pr_info("Firmware does not support resize CQ\n");
1236 		return -ENOSYS;
1237 	}
1238 
1239 	if (entries < 1 ||
1240 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1241 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1242 			     entries,
1243 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1244 		return -EINVAL;
1245 	}
1246 
1247 	entries = roundup_pow_of_two(entries + 1);
1248 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1249 		return -EINVAL;
1250 
1251 	if (entries == ibcq->cqe + 1)
1252 		return 0;
1253 
1254 	mutex_lock(&cq->resize_mutex);
1255 	if (udata) {
1256 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1257 				  &cqe_size);
1258 	} else {
1259 		cqe_size = 64;
1260 		err = resize_kernel(dev, cq, entries, cqe_size);
1261 		if (!err) {
1262 			npas = cq->resize_buf->buf.npages;
1263 			page_shift = cq->resize_buf->buf.page_shift;
1264 		}
1265 	}
1266 
1267 	if (err)
1268 		goto ex;
1269 
1270 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1271 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1272 
1273 	in = mlx5_vzalloc(inlen);
1274 	if (!in) {
1275 		err = -ENOMEM;
1276 		goto ex_resize;
1277 	}
1278 
1279 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1280 	if (udata)
1281 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1282 				     pas, 0);
1283 	else
1284 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1285 
1286 	MLX5_SET(modify_cq_in, in,
1287 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1288 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1289 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1290 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1291 
1292 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1293 
1294 	MLX5_SET(cqc, cqc, log_page_size,
1295 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1296 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1297 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1298 
1299 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1300 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1301 
1302 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq,
1303 	    (struct mlx5_modify_cq_mbox_in *)in, inlen);
1304 	if (err)
1305 		goto ex_alloc;
1306 
1307 	if (udata) {
1308 		cq->ibcq.cqe = entries - 1;
1309 		ib_umem_release(cq->buf.umem);
1310 		cq->buf.umem = cq->resize_umem;
1311 		cq->resize_umem = NULL;
1312 	} else {
1313 		struct mlx5_ib_cq_buf tbuf;
1314 		int resized = 0;
1315 
1316 		spin_lock_irqsave(&cq->lock, flags);
1317 		if (cq->resize_buf) {
1318 			err = copy_resize_cqes(cq);
1319 			if (!err) {
1320 				tbuf = cq->buf;
1321 				cq->buf = *cq->resize_buf;
1322 				kfree(cq->resize_buf);
1323 				cq->resize_buf = NULL;
1324 				resized = 1;
1325 			}
1326 		}
1327 		cq->ibcq.cqe = entries - 1;
1328 		spin_unlock_irqrestore(&cq->lock, flags);
1329 		if (resized)
1330 			free_cq_buf(dev, &tbuf);
1331 	}
1332 	mutex_unlock(&cq->resize_mutex);
1333 
1334 	kvfree(in);
1335 	return 0;
1336 
1337 ex_alloc:
1338 	kvfree(in);
1339 
1340 ex_resize:
1341 	if (udata)
1342 		un_resize_user(cq);
1343 	else
1344 		un_resize_kernel(dev, cq);
1345 ex:
1346 	mutex_unlock(&cq->resize_mutex);
1347 	return err;
1348 }
1349 
1350 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1351 {
1352 	struct mlx5_ib_cq *cq;
1353 
1354 	if (!ibcq)
1355 		return 128;
1356 
1357 	cq = to_mcq(ibcq);
1358 	return cq->cqe_size;
1359 }
1360 
1361 /* Called from atomic context */
1362 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1363 {
1364 	struct mlx5_ib_wc *soft_wc;
1365 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1366 	unsigned long flags;
1367 
1368 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1369 	if (!soft_wc)
1370 		return -ENOMEM;
1371 
1372 	soft_wc->wc = *wc;
1373 	spin_lock_irqsave(&cq->lock, flags);
1374 	list_add_tail(&soft_wc->list, &cq->wc_list);
1375 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1376 	    wc->status != IB_WC_SUCCESS) {
1377 		cq->notify_flags = 0;
1378 		schedule_work(&cq->notify_work);
1379 	}
1380 	spin_unlock_irqrestore(&cq->lock, flags);
1381 
1382 	return 0;
1383 }
1384