xref: /freebsd/contrib/ofed/libmlx4/cq.c (revision 5f4c09dd85bff675e0ca63c55ea3c517e0fddfcc)
1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
4  * Copyright (c) 2006, 2007 Cisco Systems.  All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <config.h>
36 
37 #include <stdio.h>
38 #include <stdlib.h>
39 #include <pthread.h>
40 #include <string.h>
41 
42 #include <infiniband/opcode.h>
43 
44 #include "mlx4.h"
45 #include "doorbell.h"
46 
47 enum {
48 	MLX4_CQ_DOORBELL			= 0x20
49 };
50 
51 enum {
52 	CQ_OK					=  0,
53 	CQ_EMPTY				= -1,
54 	CQ_POLL_ERR				= -2
55 };
56 
57 #define MLX4_CQ_DB_REQ_NOT_SOL			(1 << 24)
58 #define MLX4_CQ_DB_REQ_NOT			(2 << 24)
59 
60 enum {
61 	MLX4_CQE_VLAN_PRESENT_MASK		= 1 << 29,
62 	MLX4_CQE_QPN_MASK			= 0xffffff,
63 };
64 
65 enum {
66 	MLX4_CQE_OWNER_MASK			= 0x80,
67 	MLX4_CQE_IS_SEND_MASK			= 0x40,
68 	MLX4_CQE_OPCODE_MASK			= 0x1f
69 };
70 
71 enum {
72 	MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR		= 0x01,
73 	MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR		= 0x02,
74 	MLX4_CQE_SYNDROME_LOCAL_PROT_ERR		= 0x04,
75 	MLX4_CQE_SYNDROME_WR_FLUSH_ERR			= 0x05,
76 	MLX4_CQE_SYNDROME_MW_BIND_ERR			= 0x06,
77 	MLX4_CQE_SYNDROME_BAD_RESP_ERR			= 0x10,
78 	MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR		= 0x11,
79 	MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR		= 0x12,
80 	MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR		= 0x13,
81 	MLX4_CQE_SYNDROME_REMOTE_OP_ERR			= 0x14,
82 	MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR	= 0x15,
83 	MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR		= 0x16,
84 	MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR		= 0x22,
85 };
86 
87 struct mlx4_err_cqe {
88 	uint32_t	vlan_my_qpn;
89 	uint32_t	reserved1[5];
90 	uint16_t	wqe_index;
91 	uint8_t		vendor_err;
92 	uint8_t		syndrome;
93 	uint8_t		reserved2[3];
94 	uint8_t		owner_sr_opcode;
95 };
96 
97 static struct mlx4_cqe *get_cqe(struct mlx4_cq *cq, int entry)
98 {
99 	return cq->buf.buf + entry * cq->cqe_size;
100 }
101 
102 static void *get_sw_cqe(struct mlx4_cq *cq, int n)
103 {
104 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibv_cq.cqe);
105 	struct mlx4_cqe *tcqe = cq->cqe_size == 64 ? cqe + 1 : cqe;
106 
107 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
108 		!!(n & (cq->ibv_cq.cqe + 1))) ? NULL : cqe;
109 }
110 
111 static struct mlx4_cqe *next_cqe_sw(struct mlx4_cq *cq)
112 {
113 	return get_sw_cqe(cq, cq->cons_index);
114 }
115 
116 static enum ibv_wc_status mlx4_handle_error_cqe(struct mlx4_err_cqe *cqe)
117 {
118 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR)
119 		printf(PFX "local QP operation err "
120 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
121 		       "opcode = %02x)\n",
122 		       htobe32(cqe->vlan_my_qpn), htobe32(cqe->wqe_index),
123 		       cqe->vendor_err,
124 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
125 
126 	switch (cqe->syndrome) {
127 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
128 		return IBV_WC_LOC_LEN_ERR;
129 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
130 		return IBV_WC_LOC_QP_OP_ERR;
131 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
132 		return IBV_WC_LOC_PROT_ERR;
133 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
134 		return IBV_WC_WR_FLUSH_ERR;
135 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
136 		return IBV_WC_MW_BIND_ERR;
137 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
138 		return IBV_WC_BAD_RESP_ERR;
139 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
140 		return IBV_WC_LOC_ACCESS_ERR;
141 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
142 		return IBV_WC_REM_INV_REQ_ERR;
143 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
144 		return IBV_WC_REM_ACCESS_ERR;
145 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
146 		return IBV_WC_REM_OP_ERR;
147 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
148 		return IBV_WC_RETRY_EXC_ERR;
149 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
150 		return IBV_WC_RNR_RETRY_EXC_ERR;
151 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
152 		return IBV_WC_REM_ABORT_ERR;
153 	default:
154 		return IBV_WC_GENERAL_ERR;
155 	}
156 }
157 
158 static inline void handle_good_req(struct ibv_wc *wc, struct mlx4_cqe *cqe)
159 {
160 	wc->wc_flags = 0;
161 	switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
162 	case MLX4_OPCODE_RDMA_WRITE_IMM:
163 		wc->wc_flags |= IBV_WC_WITH_IMM;
164 		SWITCH_FALLTHROUGH;
165 	case MLX4_OPCODE_RDMA_WRITE:
166 		wc->opcode    = IBV_WC_RDMA_WRITE;
167 		break;
168 	case MLX4_OPCODE_SEND_IMM:
169 		wc->wc_flags |= IBV_WC_WITH_IMM;
170 		SWITCH_FALLTHROUGH;
171 	case MLX4_OPCODE_SEND:
172 	case MLX4_OPCODE_SEND_INVAL:
173 		wc->opcode    = IBV_WC_SEND;
174 		break;
175 	case MLX4_OPCODE_RDMA_READ:
176 		wc->opcode    = IBV_WC_RDMA_READ;
177 		wc->byte_len  = be32toh(cqe->byte_cnt);
178 		break;
179 	case MLX4_OPCODE_ATOMIC_CS:
180 		wc->opcode    = IBV_WC_COMP_SWAP;
181 		wc->byte_len  = 8;
182 		break;
183 	case MLX4_OPCODE_ATOMIC_FA:
184 		wc->opcode    = IBV_WC_FETCH_ADD;
185 		wc->byte_len  = 8;
186 		break;
187 	case MLX4_OPCODE_LOCAL_INVAL:
188 		wc->opcode    = IBV_WC_LOCAL_INV;
189 		break;
190 	case MLX4_OPCODE_BIND_MW:
191 		wc->opcode    = IBV_WC_BIND_MW;
192 		break;
193 	default:
194 		/* assume it's a send completion */
195 		wc->opcode    = IBV_WC_SEND;
196 		break;
197 	}
198 }
199 
200 static inline int mlx4_get_next_cqe(struct mlx4_cq *cq,
201 				    struct mlx4_cqe **pcqe)
202 				    ALWAYS_INLINE;
203 static inline int mlx4_get_next_cqe(struct mlx4_cq *cq,
204 				    struct mlx4_cqe **pcqe)
205 {
206 	struct mlx4_cqe *cqe;
207 
208 	cqe = next_cqe_sw(cq);
209 	if (!cqe)
210 		return CQ_EMPTY;
211 
212 	if (cq->cqe_size == 64)
213 		++cqe;
214 
215 	++cq->cons_index;
216 
217 	VALGRIND_MAKE_MEM_DEFINED(cqe, sizeof *cqe);
218 
219 	/*
220 	 * Make sure we read CQ entry contents after we've checked the
221 	 * ownership bit.
222 	 */
223 	udma_from_device_barrier();
224 
225 	*pcqe = cqe;
226 
227 	return CQ_OK;
228 }
229 
230 static inline int mlx4_parse_cqe(struct mlx4_cq *cq,
231 					struct mlx4_cqe *cqe,
232 					struct mlx4_qp **cur_qp,
233 					struct ibv_wc *wc, int lazy)
234 					ALWAYS_INLINE;
235 static inline int mlx4_parse_cqe(struct mlx4_cq *cq,
236 					struct mlx4_cqe *cqe,
237 					struct mlx4_qp **cur_qp,
238 					struct ibv_wc *wc, int lazy)
239 {
240 	struct mlx4_wq *wq;
241 	struct mlx4_srq *srq;
242 	uint32_t qpn;
243 	uint32_t g_mlpath_rqpn;
244 	uint64_t *pwr_id;
245 	uint16_t wqe_index;
246 	struct mlx4_err_cqe *ecqe;
247 	struct mlx4_context *mctx;
248 	int is_error;
249 	int is_send;
250 	enum ibv_wc_status *pstatus;
251 
252 	mctx = to_mctx(cq->ibv_cq.context);
253 	qpn = be32toh(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK;
254 	if (lazy) {
255 		cq->cqe = cqe;
256 		cq->flags &= (~MLX4_CQ_FLAGS_RX_CSUM_VALID);
257 	} else
258 		wc->qp_num = qpn;
259 
260 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
261 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
262 		MLX4_CQE_OPCODE_ERROR;
263 
264 	if ((qpn & MLX4_XRC_QPN_BIT) && !is_send) {
265 		/*
266 		 * We do not have to take the XSRQ table lock here,
267 		 * because CQs will be locked while SRQs are removed
268 		 * from the table.
269 		 */
270 		srq = mlx4_find_xsrq(&mctx->xsrq_table,
271 				     be32toh(cqe->g_mlpath_rqpn) & MLX4_CQE_QPN_MASK);
272 		if (!srq)
273 			return CQ_POLL_ERR;
274 	} else {
275 		if (!*cur_qp || (qpn != (*cur_qp)->verbs_qp.qp.qp_num)) {
276 			/*
277 			 * We do not have to take the QP table lock here,
278 			 * because CQs will be locked while QPs are removed
279 			 * from the table.
280 			 */
281 			*cur_qp = mlx4_find_qp(mctx, qpn);
282 			if (!*cur_qp)
283 				return CQ_POLL_ERR;
284 		}
285 		srq = ((*cur_qp)->verbs_qp.qp.srq) ? to_msrq((*cur_qp)->verbs_qp.qp.srq) : NULL;
286 	}
287 
288 	pwr_id = lazy ? &cq->ibv_cq.wr_id : &wc->wr_id;
289 	if (is_send) {
290 		wq = &(*cur_qp)->sq;
291 		wqe_index = be16toh(cqe->wqe_index);
292 		wq->tail += (uint16_t) (wqe_index - (uint16_t) wq->tail);
293 		*pwr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
294 		++wq->tail;
295 	} else if (srq) {
296 		wqe_index = be16toh(cqe->wqe_index);
297 		*pwr_id = srq->wrid[wqe_index];
298 		mlx4_free_srq_wqe(srq, wqe_index);
299 	} else {
300 		wq = &(*cur_qp)->rq;
301 		*pwr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
302 		++wq->tail;
303 	}
304 
305 	pstatus = lazy ? &cq->ibv_cq.status : &wc->status;
306 	if (is_error) {
307 		ecqe = (struct mlx4_err_cqe *)cqe;
308 		*pstatus = mlx4_handle_error_cqe(ecqe);
309 		if (!lazy)
310 			wc->vendor_err = ecqe->vendor_err;
311 		return CQ_OK;
312 	}
313 
314 	*pstatus = IBV_WC_SUCCESS;
315 	if (lazy) {
316 		if (!is_send)
317 			if ((*cur_qp) && ((*cur_qp)->qp_cap_cache & MLX4_RX_CSUM_VALID))
318 				cq->flags |= MLX4_CQ_FLAGS_RX_CSUM_VALID;
319 	} else if (is_send) {
320 		handle_good_req(wc, cqe);
321 	} else {
322 		wc->byte_len = be32toh(cqe->byte_cnt);
323 
324 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
325 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
326 			wc->opcode   = IBV_WC_RECV_RDMA_WITH_IMM;
327 			wc->wc_flags = IBV_WC_WITH_IMM;
328 			wc->imm_data = cqe->immed_rss_invalid;
329 			break;
330 		case MLX4_RECV_OPCODE_SEND_INVAL:
331 			wc->opcode   = IBV_WC_RECV;
332 			wc->wc_flags |= IBV_WC_WITH_INV;
333 			wc->imm_data = be32toh(cqe->immed_rss_invalid);
334 			break;
335 		case MLX4_RECV_OPCODE_SEND:
336 			wc->opcode   = IBV_WC_RECV;
337 			wc->wc_flags = 0;
338 			break;
339 		case MLX4_RECV_OPCODE_SEND_IMM:
340 			wc->opcode   = IBV_WC_RECV;
341 			wc->wc_flags = IBV_WC_WITH_IMM;
342 			wc->imm_data = cqe->immed_rss_invalid;
343 			break;
344 		}
345 
346 		wc->slid	   = be16toh(cqe->rlid);
347 		g_mlpath_rqpn	   = be32toh(cqe->g_mlpath_rqpn);
348 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
349 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
350 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IBV_WC_GRH : 0;
351 		wc->pkey_index     = be32toh(cqe->immed_rss_invalid) & 0x7f;
352 		/* When working with xrc srqs, don't have qp to check link layer.
353 		* Using IB SL, should consider Roce. (TBD)
354 		*/
355 		if ((*cur_qp) && (*cur_qp)->link_layer == IBV_LINK_LAYER_ETHERNET)
356 			wc->sl	   = be16toh(cqe->sl_vid) >> 13;
357 		else
358 			wc->sl	   = be16toh(cqe->sl_vid) >> 12;
359 
360 		if ((*cur_qp) && ((*cur_qp)->qp_cap_cache & MLX4_RX_CSUM_VALID)) {
361 			wc->wc_flags |= ((cqe->status & htobe32(MLX4_CQE_STATUS_IPV4_CSUM_OK)) ==
362 				 htobe32(MLX4_CQE_STATUS_IPV4_CSUM_OK)) <<
363 				IBV_WC_IP_CSUM_OK_SHIFT;
364 		}
365 	}
366 
367 	return CQ_OK;
368 }
369 
370 static inline int mlx4_parse_lazy_cqe(struct mlx4_cq *cq,
371 				      struct mlx4_cqe *cqe)
372 				      ALWAYS_INLINE;
373 static inline int mlx4_parse_lazy_cqe(struct mlx4_cq *cq,
374 				      struct mlx4_cqe *cqe)
375 {
376 	return mlx4_parse_cqe(cq, cqe, &cq->cur_qp, NULL, 1);
377 }
378 
379 static inline int mlx4_poll_one(struct mlx4_cq *cq,
380 			 struct mlx4_qp **cur_qp,
381 			 struct ibv_wc *wc)
382 			 ALWAYS_INLINE;
383 static inline int mlx4_poll_one(struct mlx4_cq *cq,
384 			 struct mlx4_qp **cur_qp,
385 			 struct ibv_wc *wc)
386 {
387 	struct mlx4_cqe *cqe;
388 	int err;
389 
390 	err = mlx4_get_next_cqe(cq, &cqe);
391 	if (err == CQ_EMPTY)
392 		return err;
393 
394 	return mlx4_parse_cqe(cq, cqe, cur_qp, wc, 0);
395 }
396 
397 int mlx4_poll_cq(struct ibv_cq *ibcq, int ne, struct ibv_wc *wc)
398 {
399 	struct mlx4_cq *cq = to_mcq(ibcq);
400 	struct mlx4_qp *qp = NULL;
401 	int npolled;
402 	int err = CQ_OK;
403 
404 	pthread_spin_lock(&cq->lock);
405 
406 	for (npolled = 0; npolled < ne; ++npolled) {
407 		err = mlx4_poll_one(cq, &qp, wc + npolled);
408 		if (err != CQ_OK)
409 			break;
410 	}
411 
412 	if (npolled || err == CQ_POLL_ERR)
413 		mlx4_update_cons_index(cq);
414 
415 	pthread_spin_unlock(&cq->lock);
416 
417 	return err == CQ_POLL_ERR ? err : npolled;
418 }
419 
420 static inline void _mlx4_end_poll(struct ibv_cq_ex *ibcq, int lock)
421 				  ALWAYS_INLINE;
422 static inline void _mlx4_end_poll(struct ibv_cq_ex *ibcq, int lock)
423 {
424 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
425 
426 	mlx4_update_cons_index(cq);
427 
428 	if (lock)
429 		pthread_spin_unlock(&cq->lock);
430 }
431 
432 static inline int _mlx4_start_poll(struct ibv_cq_ex *ibcq,
433 				   struct ibv_poll_cq_attr *attr,
434 				   int lock)
435 				   ALWAYS_INLINE;
436 static inline int _mlx4_start_poll(struct ibv_cq_ex *ibcq,
437 				   struct ibv_poll_cq_attr *attr,
438 				   int lock)
439 {
440 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
441 	struct mlx4_cqe *cqe;
442 	int err;
443 
444 	if (unlikely(attr->comp_mask))
445 		return EINVAL;
446 
447 	if (lock)
448 		pthread_spin_lock(&cq->lock);
449 
450 	cq->cur_qp = NULL;
451 
452 	err = mlx4_get_next_cqe(cq, &cqe);
453 	if (err == CQ_EMPTY) {
454 		if (lock)
455 			pthread_spin_unlock(&cq->lock);
456 		return ENOENT;
457 	}
458 
459 	err = mlx4_parse_lazy_cqe(cq, cqe);
460 	if (lock && err)
461 		pthread_spin_unlock(&cq->lock);
462 
463 	return err;
464 }
465 
466 static int mlx4_next_poll(struct ibv_cq_ex *ibcq)
467 {
468 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
469 	struct mlx4_cqe *cqe;
470 	int err;
471 
472 	err = mlx4_get_next_cqe(cq, &cqe);
473 	if (err == CQ_EMPTY)
474 		return ENOENT;
475 
476 	return mlx4_parse_lazy_cqe(cq, cqe);
477 }
478 
479 static void mlx4_end_poll(struct ibv_cq_ex *ibcq)
480 {
481 	_mlx4_end_poll(ibcq, 0);
482 }
483 
484 static void mlx4_end_poll_lock(struct ibv_cq_ex *ibcq)
485 {
486 	_mlx4_end_poll(ibcq, 1);
487 }
488 
489 static int mlx4_start_poll(struct ibv_cq_ex *ibcq,
490 		    struct ibv_poll_cq_attr *attr)
491 {
492 	return _mlx4_start_poll(ibcq, attr, 0);
493 }
494 
495 static int mlx4_start_poll_lock(struct ibv_cq_ex *ibcq,
496 			 struct ibv_poll_cq_attr *attr)
497 {
498 	return _mlx4_start_poll(ibcq, attr, 1);
499 }
500 
501 static enum ibv_wc_opcode mlx4_cq_read_wc_opcode(struct ibv_cq_ex *ibcq)
502 {
503 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
504 
505 	if (cq->cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK) {
506 		switch (cq->cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
507 		case MLX4_OPCODE_RDMA_WRITE_IMM:
508 		case MLX4_OPCODE_RDMA_WRITE:
509 			return IBV_WC_RDMA_WRITE;
510 		case MLX4_OPCODE_SEND_INVAL:
511 		case MLX4_OPCODE_SEND_IMM:
512 		case MLX4_OPCODE_SEND:
513 			return IBV_WC_SEND;
514 		case MLX4_OPCODE_RDMA_READ:
515 			return IBV_WC_RDMA_READ;
516 		case MLX4_OPCODE_ATOMIC_CS:
517 			return IBV_WC_COMP_SWAP;
518 		case MLX4_OPCODE_ATOMIC_FA:
519 			return IBV_WC_FETCH_ADD;
520 		case MLX4_OPCODE_LOCAL_INVAL:
521 			return IBV_WC_LOCAL_INV;
522 		case MLX4_OPCODE_BIND_MW:
523 			return IBV_WC_BIND_MW;
524 		}
525 	} else {
526 		switch (cq->cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
527 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
528 			return IBV_WC_RECV_RDMA_WITH_IMM;
529 		case MLX4_RECV_OPCODE_SEND_INVAL:
530 		case MLX4_RECV_OPCODE_SEND_IMM:
531 		case MLX4_RECV_OPCODE_SEND:
532 			return IBV_WC_RECV;
533 		}
534 	}
535 
536 	return 0;
537 }
538 
539 static uint32_t mlx4_cq_read_wc_qp_num(struct ibv_cq_ex *ibcq)
540 {
541 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
542 
543 	return be32toh(cq->cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK;
544 }
545 
546 static int mlx4_cq_read_wc_flags(struct ibv_cq_ex *ibcq)
547 {
548 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
549 	int is_send  = cq->cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
550 	int wc_flags = 0;
551 
552 	if (is_send) {
553 		switch (cq->cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
554 		case MLX4_OPCODE_RDMA_WRITE_IMM:
555 		case MLX4_OPCODE_SEND_IMM:
556 			wc_flags |= IBV_WC_WITH_IMM;
557 			break;
558 		}
559 	} else {
560 		if (cq->flags & MLX4_CQ_FLAGS_RX_CSUM_VALID)
561 			wc_flags |= ((cq->cqe->status &
562 				htobe32(MLX4_CQE_STATUS_IPV4_CSUM_OK)) ==
563 				htobe32(MLX4_CQE_STATUS_IPV4_CSUM_OK)) <<
564 				IBV_WC_IP_CSUM_OK_SHIFT;
565 
566 		switch (cq->cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
567 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
568 		case MLX4_RECV_OPCODE_SEND_IMM:
569 			wc_flags |= IBV_WC_WITH_IMM;
570 			break;
571 		case MLX4_RECV_OPCODE_SEND_INVAL:
572 			wc_flags |= IBV_WC_WITH_INV;
573 			break;
574 		}
575 		wc_flags |= (be32toh(cq->cqe->g_mlpath_rqpn) & 0x80000000) ? IBV_WC_GRH : 0;
576 	}
577 
578 	return wc_flags;
579 }
580 
581 static uint32_t mlx4_cq_read_wc_byte_len(struct ibv_cq_ex *ibcq)
582 {
583 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
584 
585 	return be32toh(cq->cqe->byte_cnt);
586 }
587 
588 static uint32_t mlx4_cq_read_wc_vendor_err(struct ibv_cq_ex *ibcq)
589 {
590 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
591 	struct mlx4_err_cqe *ecqe = (struct mlx4_err_cqe *)cq->cqe;
592 
593 	return ecqe->vendor_err;
594 }
595 
596 static uint32_t mlx4_cq_read_wc_imm_data(struct ibv_cq_ex *ibcq)
597 {
598 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
599 
600 	switch (cq->cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
601 	case MLX4_RECV_OPCODE_SEND_INVAL:
602 		return be32toh(cq->cqe->immed_rss_invalid);
603 	default:
604 		return cq->cqe->immed_rss_invalid;
605 	}
606 }
607 
608 static uint32_t mlx4_cq_read_wc_slid(struct ibv_cq_ex *ibcq)
609 {
610 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
611 
612 	return (uint32_t)be16toh(cq->cqe->rlid);
613 }
614 
615 static uint8_t mlx4_cq_read_wc_sl(struct ibv_cq_ex *ibcq)
616 {
617 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
618 
619 	if ((cq->cur_qp) && (cq->cur_qp->link_layer == IBV_LINK_LAYER_ETHERNET))
620 		return be16toh(cq->cqe->sl_vid) >> 13;
621 	else
622 		return be16toh(cq->cqe->sl_vid) >> 12;
623 }
624 
625 static uint32_t mlx4_cq_read_wc_src_qp(struct ibv_cq_ex *ibcq)
626 {
627 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
628 
629 	return be32toh(cq->cqe->g_mlpath_rqpn) & 0xffffff;
630 }
631 
632 static uint8_t mlx4_cq_read_wc_dlid_path_bits(struct ibv_cq_ex *ibcq)
633 {
634 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
635 
636 	return (be32toh(cq->cqe->g_mlpath_rqpn) >> 24) & 0x7f;
637 }
638 
639 static uint64_t mlx4_cq_read_wc_completion_ts(struct ibv_cq_ex *ibcq)
640 {
641 	struct mlx4_cq *cq = to_mcq(ibv_cq_ex_to_cq(ibcq));
642 
643 	return ((uint64_t)be32toh(cq->cqe->ts_47_16) << 16) |
644 			       (cq->cqe->ts_15_8   <<  8) |
645 			       (cq->cqe->ts_7_0);
646 }
647 
648 void mlx4_cq_fill_pfns(struct mlx4_cq *cq, const struct ibv_cq_init_attr_ex *cq_attr)
649 {
650 
651 	if (cq->flags & MLX4_CQ_FLAGS_SINGLE_THREADED) {
652 		cq->ibv_cq.start_poll = mlx4_start_poll;
653 		cq->ibv_cq.end_poll = mlx4_end_poll;
654 	} else {
655 		cq->ibv_cq.start_poll = mlx4_start_poll_lock;
656 		cq->ibv_cq.end_poll = mlx4_end_poll_lock;
657 	}
658 	cq->ibv_cq.next_poll = mlx4_next_poll;
659 
660 	cq->ibv_cq.read_opcode = mlx4_cq_read_wc_opcode;
661 	cq->ibv_cq.read_vendor_err = mlx4_cq_read_wc_vendor_err;
662 	cq->ibv_cq.read_wc_flags = mlx4_cq_read_wc_flags;
663 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_BYTE_LEN)
664 		cq->ibv_cq.read_byte_len = mlx4_cq_read_wc_byte_len;
665 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_IMM)
666 		cq->ibv_cq.read_imm_data = mlx4_cq_read_wc_imm_data;
667 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_QP_NUM)
668 		cq->ibv_cq.read_qp_num = mlx4_cq_read_wc_qp_num;
669 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_SRC_QP)
670 		cq->ibv_cq.read_src_qp = mlx4_cq_read_wc_src_qp;
671 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_SLID)
672 		cq->ibv_cq.read_slid = mlx4_cq_read_wc_slid;
673 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_SL)
674 		cq->ibv_cq.read_sl = mlx4_cq_read_wc_sl;
675 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_DLID_PATH_BITS)
676 		cq->ibv_cq.read_dlid_path_bits = mlx4_cq_read_wc_dlid_path_bits;
677 	if (cq_attr->wc_flags & IBV_WC_EX_WITH_COMPLETION_TIMESTAMP)
678 		cq->ibv_cq.read_completion_ts = mlx4_cq_read_wc_completion_ts;
679 }
680 
681 int mlx4_arm_cq(struct ibv_cq *ibvcq, int solicited)
682 {
683 	struct mlx4_cq *cq = to_mcq(ibvcq);
684 	uint32_t doorbell[2];
685 	uint32_t sn;
686 	uint32_t ci;
687 	uint32_t cmd;
688 
689 	sn  = cq->arm_sn & 3;
690 	ci  = cq->cons_index & 0xffffff;
691 	cmd = solicited ? MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT;
692 
693 	*cq->arm_db = htobe32(sn << 28 | cmd | ci);
694 
695 	/*
696 	 * Make sure that the doorbell record in host memory is
697 	 * written before ringing the doorbell via PCI MMIO.
698 	 */
699 	udma_to_device_barrier();
700 
701 	doorbell[0] = htobe32(sn << 28 | cmd | cq->cqn);
702 	doorbell[1] = htobe32(ci);
703 
704 	mlx4_write64(doorbell, to_mctx(ibvcq->context), MLX4_CQ_DOORBELL);
705 
706 	return 0;
707 }
708 
709 void mlx4_cq_event(struct ibv_cq *cq)
710 {
711 	to_mcq(cq)->arm_sn++;
712 }
713 
714 void __mlx4_cq_clean(struct mlx4_cq *cq, uint32_t qpn, struct mlx4_srq *srq)
715 {
716 	struct mlx4_cqe *cqe, *dest;
717 	uint32_t prod_index;
718 	uint8_t owner_bit;
719 	int nfreed = 0;
720 	int cqe_inc = cq->cqe_size == 64 ? 1 : 0;
721 
722 	/*
723 	 * First we need to find the current producer index, so we
724 	 * know where to start cleaning from.  It doesn't matter if HW
725 	 * adds new entries after this loop -- the QP we're worried
726 	 * about is already in RESET, so the new entries won't come
727 	 * from our QP and therefore don't need to be checked.
728 	 */
729 	for (prod_index = cq->cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
730 		if (prod_index == cq->cons_index + cq->ibv_cq.cqe)
731 			break;
732 
733 	/*
734 	 * Now sweep backwards through the CQ, removing CQ entries
735 	 * that match our QP by copying older entries on top of them.
736 	 */
737 	while ((int) --prod_index - (int) cq->cons_index >= 0) {
738 		cqe = get_cqe(cq, prod_index & cq->ibv_cq.cqe);
739 		cqe += cqe_inc;
740 		if (srq && srq->ext_srq &&
741 		    (be32toh(cqe->g_mlpath_rqpn) & MLX4_CQE_QPN_MASK) == srq->verbs_srq.srq_num &&
742 		    !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK)) {
743 			mlx4_free_srq_wqe(srq, be16toh(cqe->wqe_index));
744 			++nfreed;
745 		} else if ((be32toh(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
746 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
747 				mlx4_free_srq_wqe(srq, be16toh(cqe->wqe_index));
748 			++nfreed;
749 		} else if (nfreed) {
750 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibv_cq.cqe);
751 			dest += cqe_inc;
752 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
753 			memcpy(dest, cqe, sizeof *cqe);
754 			dest->owner_sr_opcode = owner_bit |
755 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
756 		}
757 	}
758 
759 	if (nfreed) {
760 		cq->cons_index += nfreed;
761 		/*
762 		 * Make sure update of buffer contents is done before
763 		 * updating consumer index.
764 		 */
765 		udma_to_device_barrier();
766 		mlx4_update_cons_index(cq);
767 	}
768 }
769 
770 void mlx4_cq_clean(struct mlx4_cq *cq, uint32_t qpn, struct mlx4_srq *srq)
771 {
772 	pthread_spin_lock(&cq->lock);
773 	__mlx4_cq_clean(cq, qpn, srq);
774 	pthread_spin_unlock(&cq->lock);
775 }
776 
777 int mlx4_get_outstanding_cqes(struct mlx4_cq *cq)
778 {
779 	uint32_t i;
780 
781 	for (i = cq->cons_index; get_sw_cqe(cq, i); ++i)
782 		;
783 
784 	return i - cq->cons_index;
785 }
786 
787 void mlx4_cq_resize_copy_cqes(struct mlx4_cq *cq, void *buf, int old_cqe)
788 {
789 	struct mlx4_cqe *cqe;
790 	int i;
791 	int cqe_inc = cq->cqe_size == 64 ? 1 : 0;
792 
793 	i = cq->cons_index;
794 	cqe = get_cqe(cq, (i & old_cqe));
795 	cqe += cqe_inc;
796 
797 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
798 		cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
799 			(((i + 1) & (cq->ibv_cq.cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
800 		memcpy(buf + ((i + 1) & cq->ibv_cq.cqe) * cq->cqe_size,
801 		       cqe - cqe_inc, cq->cqe_size);
802 		++i;
803 		cqe = get_cqe(cq, (i & old_cqe));
804 		cqe += cqe_inc;
805 	}
806 
807 	++cq->cons_index;
808 }
809 
810 int mlx4_alloc_cq_buf(struct mlx4_device *dev, struct mlx4_buf *buf, int nent,
811 		      int entry_size)
812 {
813 	if (mlx4_alloc_buf(buf, align(nent * entry_size, dev->page_size),
814 			   dev->page_size))
815 		return -1;
816 	memset(buf->buf, 0, nent * entry_size);
817 
818 	return 0;
819 }
820