xref: /linux/drivers/infiniband/hw/mlx5/cq.c (revision c4ee0af3fa0dc65f690fc908f02b8355f9576ea0)
1 /*
2  * Copyright (c) 2013, Mellanox Technologies inc.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/kref.h>
34 #include <rdma/ib_umem.h>
35 #include "mlx5_ib.h"
36 #include "user.h"
37 
38 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq)
39 {
40 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
41 
42 	ibcq->comp_handler(ibcq, ibcq->cq_context);
43 }
44 
45 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
46 {
47 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
48 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
49 	struct ib_cq *ibcq = &cq->ibcq;
50 	struct ib_event event;
51 
52 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
53 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
54 			     type, mcq->cqn);
55 		return;
56 	}
57 
58 	if (ibcq->event_handler) {
59 		event.device     = &dev->ib_dev;
60 		event.event      = IB_EVENT_CQ_ERR;
61 		event.element.cq = ibcq;
62 		ibcq->event_handler(&event, ibcq->cq_context);
63 	}
64 }
65 
66 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
67 {
68 	return mlx5_buf_offset(&buf->buf, n * size);
69 }
70 
71 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
72 {
73 	return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
74 }
75 
76 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
77 {
78 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
79 	struct mlx5_cqe64 *cqe64;
80 
81 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
82 	return ((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^
83 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
84 }
85 
86 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
87 {
88 	return get_sw_cqe(cq, cq->mcq.cons_index);
89 }
90 
91 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
92 {
93 	switch (wq->wr_data[idx]) {
94 	case MLX5_IB_WR_UMR:
95 		return 0;
96 
97 	case IB_WR_LOCAL_INV:
98 		return IB_WC_LOCAL_INV;
99 
100 	case IB_WR_FAST_REG_MR:
101 		return IB_WC_FAST_REG_MR;
102 
103 	default:
104 		pr_warn("unknown completion status\n");
105 		return 0;
106 	}
107 }
108 
109 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
110 			    struct mlx5_ib_wq *wq, int idx)
111 {
112 	wc->wc_flags = 0;
113 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
114 	case MLX5_OPCODE_RDMA_WRITE_IMM:
115 		wc->wc_flags |= IB_WC_WITH_IMM;
116 	case MLX5_OPCODE_RDMA_WRITE:
117 		wc->opcode    = IB_WC_RDMA_WRITE;
118 		break;
119 	case MLX5_OPCODE_SEND_IMM:
120 		wc->wc_flags |= IB_WC_WITH_IMM;
121 	case MLX5_OPCODE_SEND:
122 	case MLX5_OPCODE_SEND_INVAL:
123 		wc->opcode    = IB_WC_SEND;
124 		break;
125 	case MLX5_OPCODE_RDMA_READ:
126 		wc->opcode    = IB_WC_RDMA_READ;
127 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
128 		break;
129 	case MLX5_OPCODE_ATOMIC_CS:
130 		wc->opcode    = IB_WC_COMP_SWAP;
131 		wc->byte_len  = 8;
132 		break;
133 	case MLX5_OPCODE_ATOMIC_FA:
134 		wc->opcode    = IB_WC_FETCH_ADD;
135 		wc->byte_len  = 8;
136 		break;
137 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
138 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
139 		wc->byte_len  = 8;
140 		break;
141 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
142 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
143 		wc->byte_len  = 8;
144 		break;
145 	case MLX5_OPCODE_BIND_MW:
146 		wc->opcode    = IB_WC_BIND_MW;
147 		break;
148 	case MLX5_OPCODE_UMR:
149 		wc->opcode = get_umr_comp(wq, idx);
150 		break;
151 	}
152 }
153 
154 enum {
155 	MLX5_GRH_IN_BUFFER = 1,
156 	MLX5_GRH_IN_CQE	   = 2,
157 };
158 
159 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
160 			     struct mlx5_ib_qp *qp)
161 {
162 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
163 	struct mlx5_ib_srq *srq;
164 	struct mlx5_ib_wq *wq;
165 	u16 wqe_ctr;
166 	u8 g;
167 
168 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
169 		struct mlx5_core_srq *msrq = NULL;
170 
171 		if (qp->ibqp.xrcd) {
172 			msrq = mlx5_core_get_srq(&dev->mdev,
173 						 be32_to_cpu(cqe->srqn));
174 			srq = to_mibsrq(msrq);
175 		} else {
176 			srq = to_msrq(qp->ibqp.srq);
177 		}
178 		if (srq) {
179 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
180 			wc->wr_id = srq->wrid[wqe_ctr];
181 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
182 			if (msrq && atomic_dec_and_test(&msrq->refcount))
183 				complete(&msrq->free);
184 		}
185 	} else {
186 		wq	  = &qp->rq;
187 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
188 		++wq->tail;
189 	}
190 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
191 
192 	switch (cqe->op_own >> 4) {
193 	case MLX5_CQE_RESP_WR_IMM:
194 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
195 		wc->wc_flags	= IB_WC_WITH_IMM;
196 		wc->ex.imm_data = cqe->imm_inval_pkey;
197 		break;
198 	case MLX5_CQE_RESP_SEND:
199 		wc->opcode   = IB_WC_RECV;
200 		wc->wc_flags = 0;
201 		break;
202 	case MLX5_CQE_RESP_SEND_IMM:
203 		wc->opcode	= IB_WC_RECV;
204 		wc->wc_flags	= IB_WC_WITH_IMM;
205 		wc->ex.imm_data = cqe->imm_inval_pkey;
206 		break;
207 	case MLX5_CQE_RESP_SEND_INV:
208 		wc->opcode	= IB_WC_RECV;
209 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
210 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
211 		break;
212 	}
213 	wc->slid	   = be16_to_cpu(cqe->slid);
214 	wc->sl		   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
215 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
216 	wc->dlid_path_bits = cqe->ml_path;
217 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
218 	wc->wc_flags |= g ? IB_WC_GRH : 0;
219 	wc->pkey_index     = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
220 }
221 
222 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
223 {
224 	__be32 *p = (__be32 *)cqe;
225 	int i;
226 
227 	mlx5_ib_warn(dev, "dump error cqe\n");
228 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
229 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
230 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
231 			be32_to_cpu(p[3]));
232 }
233 
234 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
235 				  struct mlx5_err_cqe *cqe,
236 				  struct ib_wc *wc)
237 {
238 	int dump = 1;
239 
240 	switch (cqe->syndrome) {
241 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
242 		wc->status = IB_WC_LOC_LEN_ERR;
243 		break;
244 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
245 		wc->status = IB_WC_LOC_QP_OP_ERR;
246 		break;
247 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
248 		wc->status = IB_WC_LOC_PROT_ERR;
249 		break;
250 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
251 		dump = 0;
252 		wc->status = IB_WC_WR_FLUSH_ERR;
253 		break;
254 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
255 		wc->status = IB_WC_MW_BIND_ERR;
256 		break;
257 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
258 		wc->status = IB_WC_BAD_RESP_ERR;
259 		break;
260 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
261 		wc->status = IB_WC_LOC_ACCESS_ERR;
262 		break;
263 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
264 		wc->status = IB_WC_REM_INV_REQ_ERR;
265 		break;
266 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
267 		wc->status = IB_WC_REM_ACCESS_ERR;
268 		break;
269 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
270 		wc->status = IB_WC_REM_OP_ERR;
271 		break;
272 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
273 		wc->status = IB_WC_RETRY_EXC_ERR;
274 		dump = 0;
275 		break;
276 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
277 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
278 		dump = 0;
279 		break;
280 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
281 		wc->status = IB_WC_REM_ABORT_ERR;
282 		break;
283 	default:
284 		wc->status = IB_WC_GENERAL_ERR;
285 		break;
286 	}
287 
288 	wc->vendor_err = cqe->vendor_err_synd;
289 	if (dump)
290 		dump_cqe(dev, cqe);
291 }
292 
293 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
294 {
295 	/* TBD: waiting decision
296 	*/
297 	return 0;
298 }
299 
300 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
301 {
302 	struct mlx5_wqe_data_seg *dpseg;
303 	void *addr;
304 
305 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
306 		sizeof(struct mlx5_wqe_raddr_seg) +
307 		sizeof(struct mlx5_wqe_atomic_seg);
308 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
309 	return addr;
310 }
311 
312 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
313 			  uint16_t idx)
314 {
315 	void *addr;
316 	int byte_count;
317 	int i;
318 
319 	if (!is_atomic_response(qp, idx))
320 		return;
321 
322 	byte_count = be32_to_cpu(cqe64->byte_cnt);
323 	addr = mlx5_get_atomic_laddr(qp, idx);
324 
325 	if (byte_count == 4) {
326 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
327 	} else {
328 		for (i = 0; i < byte_count; i += 8) {
329 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
330 			addr += 8;
331 		}
332 	}
333 
334 	return;
335 }
336 
337 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
338 			   u16 tail, u16 head)
339 {
340 	int idx;
341 
342 	do {
343 		idx = tail & (qp->sq.wqe_cnt - 1);
344 		handle_atomic(qp, cqe64, idx);
345 		if (idx == head)
346 			break;
347 
348 		tail = qp->sq.w_list[idx].next;
349 	} while (1);
350 	tail = qp->sq.w_list[idx].next;
351 	qp->sq.last_poll = tail;
352 }
353 
354 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
355 			 struct mlx5_ib_qp **cur_qp,
356 			 struct ib_wc *wc)
357 {
358 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
359 	struct mlx5_err_cqe *err_cqe;
360 	struct mlx5_cqe64 *cqe64;
361 	struct mlx5_core_qp *mqp;
362 	struct mlx5_ib_wq *wq;
363 	uint8_t opcode;
364 	uint32_t qpn;
365 	u16 wqe_ctr;
366 	void *cqe;
367 	int idx;
368 
369 	cqe = next_cqe_sw(cq);
370 	if (!cqe)
371 		return -EAGAIN;
372 
373 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
374 
375 	++cq->mcq.cons_index;
376 
377 	/* Make sure we read CQ entry contents after we've checked the
378 	 * ownership bit.
379 	 */
380 	rmb();
381 
382 	/* TBD: resize CQ */
383 
384 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
385 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
386 		/* We do not have to take the QP table lock here,
387 		 * because CQs will be locked while QPs are removed
388 		 * from the table.
389 		 */
390 		mqp = __mlx5_qp_lookup(&dev->mdev, qpn);
391 		if (unlikely(!mqp)) {
392 			mlx5_ib_warn(dev, "CQE@CQ %06x for unknown QPN %6x\n",
393 				     cq->mcq.cqn, qpn);
394 			return -EINVAL;
395 		}
396 
397 		*cur_qp = to_mibqp(mqp);
398 	}
399 
400 	wc->qp  = &(*cur_qp)->ibqp;
401 	opcode = cqe64->op_own >> 4;
402 	switch (opcode) {
403 	case MLX5_CQE_REQ:
404 		wq = &(*cur_qp)->sq;
405 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
406 		idx = wqe_ctr & (wq->wqe_cnt - 1);
407 		handle_good_req(wc, cqe64, wq, idx);
408 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
409 		wc->wr_id = wq->wrid[idx];
410 		wq->tail = wq->wqe_head[idx] + 1;
411 		wc->status = IB_WC_SUCCESS;
412 		break;
413 	case MLX5_CQE_RESP_WR_IMM:
414 	case MLX5_CQE_RESP_SEND:
415 	case MLX5_CQE_RESP_SEND_IMM:
416 	case MLX5_CQE_RESP_SEND_INV:
417 		handle_responder(wc, cqe64, *cur_qp);
418 		wc->status = IB_WC_SUCCESS;
419 		break;
420 	case MLX5_CQE_RESIZE_CQ:
421 		break;
422 	case MLX5_CQE_REQ_ERR:
423 	case MLX5_CQE_RESP_ERR:
424 		err_cqe = (struct mlx5_err_cqe *)cqe64;
425 		mlx5_handle_error_cqe(dev, err_cqe, wc);
426 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
427 			    opcode == MLX5_CQE_REQ_ERR ?
428 			    "Requestor" : "Responder", cq->mcq.cqn);
429 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
430 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
431 		if (opcode == MLX5_CQE_REQ_ERR) {
432 			wq = &(*cur_qp)->sq;
433 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
434 			idx = wqe_ctr & (wq->wqe_cnt - 1);
435 			wc->wr_id = wq->wrid[idx];
436 			wq->tail = wq->wqe_head[idx] + 1;
437 		} else {
438 			struct mlx5_ib_srq *srq;
439 
440 			if ((*cur_qp)->ibqp.srq) {
441 				srq = to_msrq((*cur_qp)->ibqp.srq);
442 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
443 				wc->wr_id = srq->wrid[wqe_ctr];
444 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
445 			} else {
446 				wq = &(*cur_qp)->rq;
447 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
448 				++wq->tail;
449 			}
450 		}
451 		break;
452 	}
453 
454 	return 0;
455 }
456 
457 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
458 {
459 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
460 	struct mlx5_ib_qp *cur_qp = NULL;
461 	unsigned long flags;
462 	int npolled;
463 	int err = 0;
464 
465 	spin_lock_irqsave(&cq->lock, flags);
466 
467 	for (npolled = 0; npolled < num_entries; npolled++) {
468 		err = mlx5_poll_one(cq, &cur_qp, wc + npolled);
469 		if (err)
470 			break;
471 	}
472 
473 	if (npolled)
474 		mlx5_cq_set_ci(&cq->mcq);
475 
476 	spin_unlock_irqrestore(&cq->lock, flags);
477 
478 	if (err == 0 || err == -EAGAIN)
479 		return npolled;
480 	else
481 		return err;
482 }
483 
484 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
485 {
486 	mlx5_cq_arm(&to_mcq(ibcq)->mcq,
487 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
488 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
489 		    to_mdev(ibcq->device)->mdev.priv.uuari.uars[0].map,
490 		    MLX5_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->mdev.priv.cq_uar_lock));
491 
492 	return 0;
493 }
494 
495 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
496 			int nent, int cqe_size)
497 {
498 	int err;
499 
500 	err = mlx5_buf_alloc(&dev->mdev, nent * cqe_size,
501 			     PAGE_SIZE * 2, &buf->buf);
502 	if (err)
503 		return err;
504 
505 	buf->cqe_size = cqe_size;
506 
507 	return 0;
508 }
509 
510 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
511 {
512 	mlx5_buf_free(&dev->mdev, &buf->buf);
513 }
514 
515 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
516 			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
517 			  int entries, struct mlx5_create_cq_mbox_in **cqb,
518 			  int *cqe_size, int *index, int *inlen)
519 {
520 	struct mlx5_ib_create_cq ucmd;
521 	int page_shift;
522 	int npages;
523 	int ncont;
524 	int err;
525 
526 	if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
527 		return -EFAULT;
528 
529 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
530 		return -EINVAL;
531 
532 	*cqe_size = ucmd.cqe_size;
533 
534 	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
535 				   entries * ucmd.cqe_size,
536 				   IB_ACCESS_LOCAL_WRITE, 1);
537 	if (IS_ERR(cq->buf.umem)) {
538 		err = PTR_ERR(cq->buf.umem);
539 		return err;
540 	}
541 
542 	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
543 				  &cq->db);
544 	if (err)
545 		goto err_umem;
546 
547 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, &npages, &page_shift,
548 			   &ncont, NULL);
549 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
550 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
551 
552 	*inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * ncont;
553 	*cqb = mlx5_vzalloc(*inlen);
554 	if (!*cqb) {
555 		err = -ENOMEM;
556 		goto err_db;
557 	}
558 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, (*cqb)->pas, 0);
559 	(*cqb)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
560 
561 	*index = to_mucontext(context)->uuari.uars[0].index;
562 
563 	return 0;
564 
565 err_db:
566 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
567 
568 err_umem:
569 	ib_umem_release(cq->buf.umem);
570 	return err;
571 }
572 
573 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
574 {
575 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
576 	ib_umem_release(cq->buf.umem);
577 }
578 
579 static void init_cq_buf(struct mlx5_ib_cq *cq, int nent)
580 {
581 	int i;
582 	void *cqe;
583 	struct mlx5_cqe64 *cqe64;
584 
585 	for (i = 0; i < nent; i++) {
586 		cqe = get_cqe(cq, i);
587 		cqe64 = (cq->buf.cqe_size == 64) ? cqe : cqe + 64;
588 		cqe64->op_own = 0xf1;
589 	}
590 }
591 
592 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
593 			    int entries, int cqe_size,
594 			    struct mlx5_create_cq_mbox_in **cqb,
595 			    int *index, int *inlen)
596 {
597 	int err;
598 
599 	err = mlx5_db_alloc(&dev->mdev, &cq->db);
600 	if (err)
601 		return err;
602 
603 	cq->mcq.set_ci_db  = cq->db.db;
604 	cq->mcq.arm_db     = cq->db.db + 1;
605 	*cq->mcq.set_ci_db = 0;
606 	*cq->mcq.arm_db    = 0;
607 	cq->mcq.cqe_sz = cqe_size;
608 
609 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
610 	if (err)
611 		goto err_db;
612 
613 	init_cq_buf(cq, entries);
614 
615 	*inlen = sizeof(**cqb) + sizeof(*(*cqb)->pas) * cq->buf.buf.npages;
616 	*cqb = mlx5_vzalloc(*inlen);
617 	if (!*cqb) {
618 		err = -ENOMEM;
619 		goto err_buf;
620 	}
621 	mlx5_fill_page_array(&cq->buf.buf, (*cqb)->pas);
622 
623 	(*cqb)->ctx.log_pg_sz = cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
624 	*index = dev->mdev.priv.uuari.uars[0].index;
625 
626 	return 0;
627 
628 err_buf:
629 	free_cq_buf(dev, &cq->buf);
630 
631 err_db:
632 	mlx5_db_free(&dev->mdev, &cq->db);
633 	return err;
634 }
635 
636 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
637 {
638 	free_cq_buf(dev, &cq->buf);
639 	mlx5_db_free(&dev->mdev, &cq->db);
640 }
641 
642 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev, int entries,
643 				int vector, struct ib_ucontext *context,
644 				struct ib_udata *udata)
645 {
646 	struct mlx5_create_cq_mbox_in *cqb = NULL;
647 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
648 	struct mlx5_ib_cq *cq;
649 	int uninitialized_var(index);
650 	int uninitialized_var(inlen);
651 	int cqe_size;
652 	int irqn;
653 	int eqn;
654 	int err;
655 
656 	if (entries < 0)
657 		return ERR_PTR(-EINVAL);
658 
659 	entries = roundup_pow_of_two(entries + 1);
660 	if (entries > dev->mdev.caps.max_cqes)
661 		return ERR_PTR(-EINVAL);
662 
663 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
664 	if (!cq)
665 		return ERR_PTR(-ENOMEM);
666 
667 	cq->ibcq.cqe = entries - 1;
668 	mutex_init(&cq->resize_mutex);
669 	spin_lock_init(&cq->lock);
670 	cq->resize_buf = NULL;
671 	cq->resize_umem = NULL;
672 
673 	if (context) {
674 		err = create_cq_user(dev, udata, context, cq, entries,
675 				     &cqb, &cqe_size, &index, &inlen);
676 		if (err)
677 			goto err_create;
678 	} else {
679 		/* for now choose 64 bytes till we have a proper interface */
680 		cqe_size = 64;
681 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
682 				       &index, &inlen);
683 		if (err)
684 			goto err_create;
685 	}
686 
687 	cq->cqe_size = cqe_size;
688 	cqb->ctx.cqe_sz_flags = cqe_sz_to_mlx_sz(cqe_size) << 5;
689 	cqb->ctx.log_sz_usr_page = cpu_to_be32((ilog2(entries) << 24) | index);
690 	err = mlx5_vector2eqn(dev, vector, &eqn, &irqn);
691 	if (err)
692 		goto err_cqb;
693 
694 	cqb->ctx.c_eqn = cpu_to_be16(eqn);
695 	cqb->ctx.db_record_addr = cpu_to_be64(cq->db.dma);
696 
697 	err = mlx5_core_create_cq(&dev->mdev, &cq->mcq, cqb, inlen);
698 	if (err)
699 		goto err_cqb;
700 
701 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
702 	cq->mcq.irqn = irqn;
703 	cq->mcq.comp  = mlx5_ib_cq_comp;
704 	cq->mcq.event = mlx5_ib_cq_event;
705 
706 	if (context)
707 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
708 			err = -EFAULT;
709 			goto err_cmd;
710 		}
711 
712 
713 	mlx5_vfree(cqb);
714 	return &cq->ibcq;
715 
716 err_cmd:
717 	mlx5_core_destroy_cq(&dev->mdev, &cq->mcq);
718 
719 err_cqb:
720 	mlx5_vfree(cqb);
721 	if (context)
722 		destroy_cq_user(cq, context);
723 	else
724 		destroy_cq_kernel(dev, cq);
725 
726 err_create:
727 	kfree(cq);
728 
729 	return ERR_PTR(err);
730 }
731 
732 
733 int mlx5_ib_destroy_cq(struct ib_cq *cq)
734 {
735 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
736 	struct mlx5_ib_cq *mcq = to_mcq(cq);
737 	struct ib_ucontext *context = NULL;
738 
739 	if (cq->uobject)
740 		context = cq->uobject->context;
741 
742 	mlx5_core_destroy_cq(&dev->mdev, &mcq->mcq);
743 	if (context)
744 		destroy_cq_user(mcq, context);
745 	else
746 		destroy_cq_kernel(dev, mcq);
747 
748 	kfree(mcq);
749 
750 	return 0;
751 }
752 
753 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
754 {
755 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
756 }
757 
758 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
759 {
760 	struct mlx5_cqe64 *cqe64, *dest64;
761 	void *cqe, *dest;
762 	u32 prod_index;
763 	int nfreed = 0;
764 	u8 owner_bit;
765 
766 	if (!cq)
767 		return;
768 
769 	/* First we need to find the current producer index, so we
770 	 * know where to start cleaning from.  It doesn't matter if HW
771 	 * adds new entries after this loop -- the QP we're worried
772 	 * about is already in RESET, so the new entries won't come
773 	 * from our QP and therefore don't need to be checked.
774 	 */
775 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
776 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
777 			break;
778 
779 	/* Now sweep backwards through the CQ, removing CQ entries
780 	 * that match our QP by copying older entries on top of them.
781 	 */
782 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
783 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
784 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
785 		if (is_equal_rsn(cqe64, rsn)) {
786 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
787 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
788 			++nfreed;
789 		} else if (nfreed) {
790 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
791 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
792 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
793 			memcpy(dest, cqe, cq->mcq.cqe_sz);
794 			dest64->op_own = owner_bit |
795 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
796 		}
797 	}
798 
799 	if (nfreed) {
800 		cq->mcq.cons_index += nfreed;
801 		/* Make sure update of buffer contents is done before
802 		 * updating consumer index.
803 		 */
804 		wmb();
805 		mlx5_cq_set_ci(&cq->mcq);
806 	}
807 }
808 
809 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
810 {
811 	if (!cq)
812 		return;
813 
814 	spin_lock_irq(&cq->lock);
815 	__mlx5_ib_cq_clean(cq, qpn, srq);
816 	spin_unlock_irq(&cq->lock);
817 }
818 
819 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
820 {
821 	return -ENOSYS;
822 }
823 
824 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
825 {
826 	return -ENOSYS;
827 }
828 
829 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
830 {
831 	struct mlx5_ib_cq *cq;
832 
833 	if (!ibcq)
834 		return 128;
835 
836 	cq = to_mcq(ibcq);
837 	return cq->cqe_size;
838 }
839