xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision 4f6cce39)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. 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 <rdma/ib_user_verbs.h>
36 #include <rdma/ib_cache.h>
37 #include "mlx5_ib.h"
38 
39 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq)
40 {
41 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
42 
43 	ibcq->comp_handler(ibcq, ibcq->cq_context);
44 }
45 
46 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
47 {
48 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
49 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
50 	struct ib_cq *ibcq = &cq->ibcq;
51 	struct ib_event event;
52 
53 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
54 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
55 			     type, mcq->cqn);
56 		return;
57 	}
58 
59 	if (ibcq->event_handler) {
60 		event.device     = &dev->ib_dev;
61 		event.event      = IB_EVENT_CQ_ERR;
62 		event.element.cq = ibcq;
63 		ibcq->event_handler(&event, ibcq->cq_context);
64 	}
65 }
66 
67 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size)
68 {
69 	return mlx5_buf_offset(&buf->buf, n * size);
70 }
71 
72 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
73 {
74 	return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz);
75 }
76 
77 static u8 sw_ownership_bit(int n, int nent)
78 {
79 	return (n & nent) ? 1 : 0;
80 }
81 
82 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
83 {
84 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
85 	struct mlx5_cqe64 *cqe64;
86 
87 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
88 
89 	if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
90 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
91 		return cqe;
92 	} else {
93 		return NULL;
94 	}
95 }
96 
97 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
98 {
99 	return get_sw_cqe(cq, cq->mcq.cons_index);
100 }
101 
102 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
103 {
104 	switch (wq->wr_data[idx]) {
105 	case MLX5_IB_WR_UMR:
106 		return 0;
107 
108 	case IB_WR_LOCAL_INV:
109 		return IB_WC_LOCAL_INV;
110 
111 	case IB_WR_REG_MR:
112 		return IB_WC_REG_MR;
113 
114 	default:
115 		pr_warn("unknown completion status\n");
116 		return 0;
117 	}
118 }
119 
120 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
121 			    struct mlx5_ib_wq *wq, int idx)
122 {
123 	wc->wc_flags = 0;
124 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
125 	case MLX5_OPCODE_RDMA_WRITE_IMM:
126 		wc->wc_flags |= IB_WC_WITH_IMM;
127 	case MLX5_OPCODE_RDMA_WRITE:
128 		wc->opcode    = IB_WC_RDMA_WRITE;
129 		break;
130 	case MLX5_OPCODE_SEND_IMM:
131 		wc->wc_flags |= IB_WC_WITH_IMM;
132 	case MLX5_OPCODE_SEND:
133 	case MLX5_OPCODE_SEND_INVAL:
134 		wc->opcode    = IB_WC_SEND;
135 		break;
136 	case MLX5_OPCODE_RDMA_READ:
137 		wc->opcode    = IB_WC_RDMA_READ;
138 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
139 		break;
140 	case MLX5_OPCODE_ATOMIC_CS:
141 		wc->opcode    = IB_WC_COMP_SWAP;
142 		wc->byte_len  = 8;
143 		break;
144 	case MLX5_OPCODE_ATOMIC_FA:
145 		wc->opcode    = IB_WC_FETCH_ADD;
146 		wc->byte_len  = 8;
147 		break;
148 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
149 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
150 		wc->byte_len  = 8;
151 		break;
152 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
153 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
154 		wc->byte_len  = 8;
155 		break;
156 	case MLX5_OPCODE_UMR:
157 		wc->opcode = get_umr_comp(wq, idx);
158 		break;
159 	}
160 }
161 
162 enum {
163 	MLX5_GRH_IN_BUFFER = 1,
164 	MLX5_GRH_IN_CQE	   = 2,
165 };
166 
167 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
168 			     struct mlx5_ib_qp *qp)
169 {
170 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
171 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
172 	struct mlx5_ib_srq *srq;
173 	struct mlx5_ib_wq *wq;
174 	u16 wqe_ctr;
175 	u8 g;
176 
177 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
178 		struct mlx5_core_srq *msrq = NULL;
179 
180 		if (qp->ibqp.xrcd) {
181 			msrq = mlx5_core_get_srq(dev->mdev,
182 						 be32_to_cpu(cqe->srqn));
183 			srq = to_mibsrq(msrq);
184 		} else {
185 			srq = to_msrq(qp->ibqp.srq);
186 		}
187 		if (srq) {
188 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
189 			wc->wr_id = srq->wrid[wqe_ctr];
190 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
191 			if (msrq && atomic_dec_and_test(&msrq->refcount))
192 				complete(&msrq->free);
193 		}
194 	} else {
195 		wq	  = &qp->rq;
196 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
197 		++wq->tail;
198 	}
199 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
200 
201 	switch (cqe->op_own >> 4) {
202 	case MLX5_CQE_RESP_WR_IMM:
203 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
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:
208 		wc->opcode   = IB_WC_RECV;
209 		wc->wc_flags = IB_WC_IP_CSUM_OK;
210 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
211 			       (cqe->hds_ip_ext & CQE_L4_OK))))
212 			wc->wc_flags = 0;
213 		break;
214 	case MLX5_CQE_RESP_SEND_IMM:
215 		wc->opcode	= IB_WC_RECV;
216 		wc->wc_flags	= IB_WC_WITH_IMM;
217 		wc->ex.imm_data = cqe->imm_inval_pkey;
218 		break;
219 	case MLX5_CQE_RESP_SEND_INV:
220 		wc->opcode	= IB_WC_RECV;
221 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
222 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
223 		break;
224 	}
225 	wc->slid	   = be16_to_cpu(cqe->slid);
226 	wc->sl		   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
227 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
228 	wc->dlid_path_bits = cqe->ml_path;
229 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
230 	wc->wc_flags |= g ? IB_WC_GRH : 0;
231 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
232 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
233 
234 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
235 				    &wc->pkey_index);
236 	} else {
237 		wc->pkey_index = 0;
238 	}
239 
240 	if (ll != IB_LINK_LAYER_ETHERNET)
241 		return;
242 
243 	switch (wc->sl & 0x3) {
244 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
245 		wc->network_hdr_type = RDMA_NETWORK_IB;
246 		break;
247 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
248 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
249 		break;
250 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
251 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
252 		break;
253 	}
254 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
255 }
256 
257 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
258 {
259 	__be32 *p = (__be32 *)cqe;
260 	int i;
261 
262 	mlx5_ib_warn(dev, "dump error cqe\n");
263 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
264 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
265 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
266 			be32_to_cpu(p[3]));
267 }
268 
269 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
270 				  struct mlx5_err_cqe *cqe,
271 				  struct ib_wc *wc)
272 {
273 	int dump = 1;
274 
275 	switch (cqe->syndrome) {
276 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
277 		wc->status = IB_WC_LOC_LEN_ERR;
278 		break;
279 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
280 		wc->status = IB_WC_LOC_QP_OP_ERR;
281 		break;
282 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
283 		wc->status = IB_WC_LOC_PROT_ERR;
284 		break;
285 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
286 		dump = 0;
287 		wc->status = IB_WC_WR_FLUSH_ERR;
288 		break;
289 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
290 		wc->status = IB_WC_MW_BIND_ERR;
291 		break;
292 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
293 		wc->status = IB_WC_BAD_RESP_ERR;
294 		break;
295 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
296 		wc->status = IB_WC_LOC_ACCESS_ERR;
297 		break;
298 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
299 		wc->status = IB_WC_REM_INV_REQ_ERR;
300 		break;
301 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
302 		wc->status = IB_WC_REM_ACCESS_ERR;
303 		break;
304 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
305 		wc->status = IB_WC_REM_OP_ERR;
306 		break;
307 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
308 		wc->status = IB_WC_RETRY_EXC_ERR;
309 		dump = 0;
310 		break;
311 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
312 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
313 		dump = 0;
314 		break;
315 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
316 		wc->status = IB_WC_REM_ABORT_ERR;
317 		break;
318 	default:
319 		wc->status = IB_WC_GENERAL_ERR;
320 		break;
321 	}
322 
323 	wc->vendor_err = cqe->vendor_err_synd;
324 	if (dump)
325 		dump_cqe(dev, cqe);
326 }
327 
328 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
329 {
330 	/* TBD: waiting decision
331 	*/
332 	return 0;
333 }
334 
335 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
336 {
337 	struct mlx5_wqe_data_seg *dpseg;
338 	void *addr;
339 
340 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
341 		sizeof(struct mlx5_wqe_raddr_seg) +
342 		sizeof(struct mlx5_wqe_atomic_seg);
343 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
344 	return addr;
345 }
346 
347 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
348 			  uint16_t idx)
349 {
350 	void *addr;
351 	int byte_count;
352 	int i;
353 
354 	if (!is_atomic_response(qp, idx))
355 		return;
356 
357 	byte_count = be32_to_cpu(cqe64->byte_cnt);
358 	addr = mlx5_get_atomic_laddr(qp, idx);
359 
360 	if (byte_count == 4) {
361 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
362 	} else {
363 		for (i = 0; i < byte_count; i += 8) {
364 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
365 			addr += 8;
366 		}
367 	}
368 
369 	return;
370 }
371 
372 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
373 			   u16 tail, u16 head)
374 {
375 	u16 idx;
376 
377 	do {
378 		idx = tail & (qp->sq.wqe_cnt - 1);
379 		handle_atomic(qp, cqe64, idx);
380 		if (idx == head)
381 			break;
382 
383 		tail = qp->sq.w_list[idx].next;
384 	} while (1);
385 	tail = qp->sq.w_list[idx].next;
386 	qp->sq.last_poll = tail;
387 }
388 
389 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
390 {
391 	mlx5_buf_free(dev->mdev, &buf->buf);
392 }
393 
394 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
395 			     struct ib_sig_err *item)
396 {
397 	u16 syndrome = be16_to_cpu(cqe->syndrome);
398 
399 #define GUARD_ERR   (1 << 13)
400 #define APPTAG_ERR  (1 << 12)
401 #define REFTAG_ERR  (1 << 11)
402 
403 	if (syndrome & GUARD_ERR) {
404 		item->err_type = IB_SIG_BAD_GUARD;
405 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
406 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
407 	} else
408 	if (syndrome & REFTAG_ERR) {
409 		item->err_type = IB_SIG_BAD_REFTAG;
410 		item->expected = be32_to_cpu(cqe->expected_reftag);
411 		item->actual = be32_to_cpu(cqe->actual_reftag);
412 	} else
413 	if (syndrome & APPTAG_ERR) {
414 		item->err_type = IB_SIG_BAD_APPTAG;
415 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
416 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
417 	} else {
418 		pr_err("Got signature completion error with bad syndrome %04x\n",
419 		       syndrome);
420 	}
421 
422 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
423 	item->key = be32_to_cpu(cqe->mkey);
424 }
425 
426 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
427 			 struct ib_wc *wc, int *npolled)
428 {
429 	struct mlx5_ib_wq *wq;
430 	unsigned int cur;
431 	unsigned int idx;
432 	int np;
433 	int i;
434 
435 	wq = &qp->sq;
436 	cur = wq->head - wq->tail;
437 	np = *npolled;
438 
439 	if (cur == 0)
440 		return;
441 
442 	for (i = 0;  i < cur && np < num_entries; i++) {
443 		idx = wq->last_poll & (wq->wqe_cnt - 1);
444 		wc->wr_id = wq->wrid[idx];
445 		wc->status = IB_WC_WR_FLUSH_ERR;
446 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
447 		wq->tail++;
448 		np++;
449 		wc->qp = &qp->ibqp;
450 		wc++;
451 		wq->last_poll = wq->w_list[idx].next;
452 	}
453 	*npolled = np;
454 }
455 
456 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
457 			 struct ib_wc *wc, int *npolled)
458 {
459 	struct mlx5_ib_wq *wq;
460 	unsigned int cur;
461 	int np;
462 	int i;
463 
464 	wq = &qp->rq;
465 	cur = wq->head - wq->tail;
466 	np = *npolled;
467 
468 	if (cur == 0)
469 		return;
470 
471 	for (i = 0;  i < cur && np < num_entries; i++) {
472 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
473 		wc->status = IB_WC_WR_FLUSH_ERR;
474 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
475 		wq->tail++;
476 		np++;
477 		wc->qp = &qp->ibqp;
478 		wc++;
479 	}
480 	*npolled = np;
481 }
482 
483 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
484 				 struct ib_wc *wc, int *npolled)
485 {
486 	struct mlx5_ib_qp *qp;
487 
488 	*npolled = 0;
489 	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
490 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
491 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
492 		if (*npolled >= num_entries)
493 			return;
494 	}
495 
496 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
497 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
498 		if (*npolled >= num_entries)
499 			return;
500 	}
501 }
502 
503 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
504 			 struct mlx5_ib_qp **cur_qp,
505 			 struct ib_wc *wc)
506 {
507 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
508 	struct mlx5_err_cqe *err_cqe;
509 	struct mlx5_cqe64 *cqe64;
510 	struct mlx5_core_qp *mqp;
511 	struct mlx5_ib_wq *wq;
512 	struct mlx5_sig_err_cqe *sig_err_cqe;
513 	struct mlx5_core_mkey *mmkey;
514 	struct mlx5_ib_mr *mr;
515 	uint8_t opcode;
516 	uint32_t qpn;
517 	u16 wqe_ctr;
518 	void *cqe;
519 	int idx;
520 
521 repoll:
522 	cqe = next_cqe_sw(cq);
523 	if (!cqe)
524 		return -EAGAIN;
525 
526 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
527 
528 	++cq->mcq.cons_index;
529 
530 	/* Make sure we read CQ entry contents after we've checked the
531 	 * ownership bit.
532 	 */
533 	rmb();
534 
535 	opcode = cqe64->op_own >> 4;
536 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
537 		if (likely(cq->resize_buf)) {
538 			free_cq_buf(dev, &cq->buf);
539 			cq->buf = *cq->resize_buf;
540 			kfree(cq->resize_buf);
541 			cq->resize_buf = NULL;
542 			goto repoll;
543 		} else {
544 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
545 		}
546 	}
547 
548 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
549 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
550 		/* We do not have to take the QP table lock here,
551 		 * because CQs will be locked while QPs are removed
552 		 * from the table.
553 		 */
554 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
555 		*cur_qp = to_mibqp(mqp);
556 	}
557 
558 	wc->qp  = &(*cur_qp)->ibqp;
559 	switch (opcode) {
560 	case MLX5_CQE_REQ:
561 		wq = &(*cur_qp)->sq;
562 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
563 		idx = wqe_ctr & (wq->wqe_cnt - 1);
564 		handle_good_req(wc, cqe64, wq, idx);
565 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
566 		wc->wr_id = wq->wrid[idx];
567 		wq->tail = wq->wqe_head[idx] + 1;
568 		wc->status = IB_WC_SUCCESS;
569 		break;
570 	case MLX5_CQE_RESP_WR_IMM:
571 	case MLX5_CQE_RESP_SEND:
572 	case MLX5_CQE_RESP_SEND_IMM:
573 	case MLX5_CQE_RESP_SEND_INV:
574 		handle_responder(wc, cqe64, *cur_qp);
575 		wc->status = IB_WC_SUCCESS;
576 		break;
577 	case MLX5_CQE_RESIZE_CQ:
578 		break;
579 	case MLX5_CQE_REQ_ERR:
580 	case MLX5_CQE_RESP_ERR:
581 		err_cqe = (struct mlx5_err_cqe *)cqe64;
582 		mlx5_handle_error_cqe(dev, err_cqe, wc);
583 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
584 			    opcode == MLX5_CQE_REQ_ERR ?
585 			    "Requestor" : "Responder", cq->mcq.cqn);
586 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
587 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
588 		if (opcode == MLX5_CQE_REQ_ERR) {
589 			wq = &(*cur_qp)->sq;
590 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
591 			idx = wqe_ctr & (wq->wqe_cnt - 1);
592 			wc->wr_id = wq->wrid[idx];
593 			wq->tail = wq->wqe_head[idx] + 1;
594 		} else {
595 			struct mlx5_ib_srq *srq;
596 
597 			if ((*cur_qp)->ibqp.srq) {
598 				srq = to_msrq((*cur_qp)->ibqp.srq);
599 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
600 				wc->wr_id = srq->wrid[wqe_ctr];
601 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
602 			} else {
603 				wq = &(*cur_qp)->rq;
604 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
605 				++wq->tail;
606 			}
607 		}
608 		break;
609 	case MLX5_CQE_SIG_ERR:
610 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
611 
612 		read_lock(&dev->mdev->priv.mkey_table.lock);
613 		mmkey = __mlx5_mr_lookup(dev->mdev,
614 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
615 		mr = to_mibmr(mmkey);
616 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
617 		mr->sig->sig_err_exists = true;
618 		mr->sig->sigerr_count++;
619 
620 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
621 			     cq->mcq.cqn, mr->sig->err_item.key,
622 			     mr->sig->err_item.err_type,
623 			     mr->sig->err_item.sig_err_offset,
624 			     mr->sig->err_item.expected,
625 			     mr->sig->err_item.actual);
626 
627 		read_unlock(&dev->mdev->priv.mkey_table.lock);
628 		goto repoll;
629 	}
630 
631 	return 0;
632 }
633 
634 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
635 			struct ib_wc *wc)
636 {
637 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
638 	struct mlx5_ib_wc *soft_wc, *next;
639 	int npolled = 0;
640 
641 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
642 		if (npolled >= num_entries)
643 			break;
644 
645 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
646 			    cq->mcq.cqn);
647 
648 		wc[npolled++] = soft_wc->wc;
649 		list_del(&soft_wc->list);
650 		kfree(soft_wc);
651 	}
652 
653 	return npolled;
654 }
655 
656 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
657 {
658 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
659 	struct mlx5_ib_qp *cur_qp = NULL;
660 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
661 	struct mlx5_core_dev *mdev = dev->mdev;
662 	unsigned long flags;
663 	int soft_polled = 0;
664 	int npolled;
665 
666 	spin_lock_irqsave(&cq->lock, flags);
667 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
668 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
669 		goto out;
670 	}
671 
672 	if (unlikely(!list_empty(&cq->wc_list)))
673 		soft_polled = poll_soft_wc(cq, num_entries, wc);
674 
675 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
676 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
677 			break;
678 	}
679 
680 	if (npolled)
681 		mlx5_cq_set_ci(&cq->mcq);
682 out:
683 	spin_unlock_irqrestore(&cq->lock, flags);
684 
685 	return soft_polled + npolled;
686 }
687 
688 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
689 {
690 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
691 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
692 	void __iomem *uar_page = mdev->priv.uar->map;
693 	unsigned long irq_flags;
694 	int ret = 0;
695 
696 	spin_lock_irqsave(&cq->lock, irq_flags);
697 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
698 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
699 
700 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
701 		ret = 1;
702 	spin_unlock_irqrestore(&cq->lock, irq_flags);
703 
704 	mlx5_cq_arm(&cq->mcq,
705 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
706 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
707 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
708 
709 	return ret;
710 }
711 
712 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
713 			int nent, int cqe_size)
714 {
715 	int err;
716 
717 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size, &buf->buf);
718 	if (err)
719 		return err;
720 
721 	buf->cqe_size = cqe_size;
722 	buf->nent = nent;
723 
724 	return 0;
725 }
726 
727 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
728 			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
729 			  int entries, u32 **cqb,
730 			  int *cqe_size, int *index, int *inlen)
731 {
732 	struct mlx5_ib_create_cq ucmd = {};
733 	size_t ucmdlen;
734 	int page_shift;
735 	__be64 *pas;
736 	int npages;
737 	int ncont;
738 	void *cqc;
739 	int err;
740 
741 	ucmdlen =
742 		(udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
743 		 sizeof(ucmd)) ? (sizeof(ucmd) -
744 				  sizeof(ucmd.reserved)) : sizeof(ucmd);
745 
746 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
747 		return -EFAULT;
748 
749 	if (ucmdlen == sizeof(ucmd) &&
750 	    ucmd.reserved != 0)
751 		return -EINVAL;
752 
753 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
754 		return -EINVAL;
755 
756 	*cqe_size = ucmd.cqe_size;
757 
758 	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
759 				   entries * ucmd.cqe_size,
760 				   IB_ACCESS_LOCAL_WRITE, 1);
761 	if (IS_ERR(cq->buf.umem)) {
762 		err = PTR_ERR(cq->buf.umem);
763 		return err;
764 	}
765 
766 	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
767 				  &cq->db);
768 	if (err)
769 		goto err_umem;
770 
771 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
772 			   &ncont, NULL);
773 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
774 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
775 
776 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
777 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
778 	*cqb = mlx5_vzalloc(*inlen);
779 	if (!*cqb) {
780 		err = -ENOMEM;
781 		goto err_db;
782 	}
783 
784 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
785 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
786 
787 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
788 	MLX5_SET(cqc, cqc, log_page_size,
789 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
790 
791 	*index = to_mucontext(context)->bfregi.sys_pages[0];
792 
793 	if (ucmd.cqe_comp_en == 1) {
794 		if (unlikely((*cqe_size != 64) ||
795 			     !MLX5_CAP_GEN(dev->mdev, cqe_compression))) {
796 			err = -EOPNOTSUPP;
797 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
798 				     *cqe_size);
799 			goto err_cqb;
800 		}
801 
802 		if (unlikely(!ucmd.cqe_comp_res_format ||
803 			     !(ucmd.cqe_comp_res_format <
804 			       MLX5_IB_CQE_RES_RESERVED) ||
805 			     (ucmd.cqe_comp_res_format &
806 			      (ucmd.cqe_comp_res_format - 1)))) {
807 			err = -EOPNOTSUPP;
808 			mlx5_ib_warn(dev, "CQE compression res format %d is not supported!\n",
809 				     ucmd.cqe_comp_res_format);
810 			goto err_cqb;
811 		}
812 
813 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
814 		MLX5_SET(cqc, cqc, mini_cqe_res_format,
815 			 ilog2(ucmd.cqe_comp_res_format));
816 	}
817 
818 	return 0;
819 
820 err_cqb:
821 	kfree(cqb);
822 
823 err_db:
824 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
825 
826 err_umem:
827 	ib_umem_release(cq->buf.umem);
828 	return err;
829 }
830 
831 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
832 {
833 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
834 	ib_umem_release(cq->buf.umem);
835 }
836 
837 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
838 {
839 	int i;
840 	void *cqe;
841 	struct mlx5_cqe64 *cqe64;
842 
843 	for (i = 0; i < buf->nent; i++) {
844 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
845 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
846 		cqe64->op_own = MLX5_CQE_INVALID << 4;
847 	}
848 }
849 
850 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
851 			    int entries, int cqe_size,
852 			    u32 **cqb, int *index, int *inlen)
853 {
854 	__be64 *pas;
855 	void *cqc;
856 	int err;
857 
858 	err = mlx5_db_alloc(dev->mdev, &cq->db);
859 	if (err)
860 		return err;
861 
862 	cq->mcq.set_ci_db  = cq->db.db;
863 	cq->mcq.arm_db     = cq->db.db + 1;
864 	cq->mcq.cqe_sz = cqe_size;
865 
866 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
867 	if (err)
868 		goto err_db;
869 
870 	init_cq_buf(cq, &cq->buf);
871 
872 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
873 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
874 	*cqb = mlx5_vzalloc(*inlen);
875 	if (!*cqb) {
876 		err = -ENOMEM;
877 		goto err_buf;
878 	}
879 
880 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
881 	mlx5_fill_page_array(&cq->buf.buf, pas);
882 
883 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
884 	MLX5_SET(cqc, cqc, log_page_size,
885 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
886 
887 	*index = dev->mdev->priv.uar->index;
888 
889 	return 0;
890 
891 err_buf:
892 	free_cq_buf(dev, &cq->buf);
893 
894 err_db:
895 	mlx5_db_free(dev->mdev, &cq->db);
896 	return err;
897 }
898 
899 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
900 {
901 	free_cq_buf(dev, &cq->buf);
902 	mlx5_db_free(dev->mdev, &cq->db);
903 }
904 
905 static void notify_soft_wc_handler(struct work_struct *work)
906 {
907 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
908 					     notify_work);
909 
910 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
911 }
912 
913 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
914 				const struct ib_cq_init_attr *attr,
915 				struct ib_ucontext *context,
916 				struct ib_udata *udata)
917 {
918 	int entries = attr->cqe;
919 	int vector = attr->comp_vector;
920 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
921 	struct mlx5_ib_cq *cq;
922 	int uninitialized_var(index);
923 	int uninitialized_var(inlen);
924 	u32 *cqb = NULL;
925 	void *cqc;
926 	int cqe_size;
927 	unsigned int irqn;
928 	int eqn;
929 	int err;
930 
931 	if (entries < 0 ||
932 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
933 		return ERR_PTR(-EINVAL);
934 
935 	if (check_cq_create_flags(attr->flags))
936 		return ERR_PTR(-EOPNOTSUPP);
937 
938 	entries = roundup_pow_of_two(entries + 1);
939 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
940 		return ERR_PTR(-EINVAL);
941 
942 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
943 	if (!cq)
944 		return ERR_PTR(-ENOMEM);
945 
946 	cq->ibcq.cqe = entries - 1;
947 	mutex_init(&cq->resize_mutex);
948 	spin_lock_init(&cq->lock);
949 	cq->resize_buf = NULL;
950 	cq->resize_umem = NULL;
951 	cq->create_flags = attr->flags;
952 	INIT_LIST_HEAD(&cq->list_send_qp);
953 	INIT_LIST_HEAD(&cq->list_recv_qp);
954 
955 	if (context) {
956 		err = create_cq_user(dev, udata, context, cq, entries,
957 				     &cqb, &cqe_size, &index, &inlen);
958 		if (err)
959 			goto err_create;
960 	} else {
961 		cqe_size = cache_line_size() == 128 ? 128 : 64;
962 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
963 				       &index, &inlen);
964 		if (err)
965 			goto err_create;
966 
967 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
968 	}
969 
970 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
971 	if (err)
972 		goto err_cqb;
973 
974 	cq->cqe_size = cqe_size;
975 
976 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
977 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
978 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
979 	MLX5_SET(cqc, cqc, uar_page, index);
980 	MLX5_SET(cqc, cqc, c_eqn, eqn);
981 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
982 	if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
983 		MLX5_SET(cqc, cqc, oi, 1);
984 
985 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
986 	if (err)
987 		goto err_cqb;
988 
989 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
990 	cq->mcq.irqn = irqn;
991 	if (context)
992 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
993 	else
994 		cq->mcq.comp  = mlx5_ib_cq_comp;
995 	cq->mcq.event = mlx5_ib_cq_event;
996 
997 	INIT_LIST_HEAD(&cq->wc_list);
998 
999 	if (context)
1000 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
1001 			err = -EFAULT;
1002 			goto err_cmd;
1003 		}
1004 
1005 
1006 	kvfree(cqb);
1007 	return &cq->ibcq;
1008 
1009 err_cmd:
1010 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1011 
1012 err_cqb:
1013 	kvfree(cqb);
1014 	if (context)
1015 		destroy_cq_user(cq, context);
1016 	else
1017 		destroy_cq_kernel(dev, cq);
1018 
1019 err_create:
1020 	kfree(cq);
1021 
1022 	return ERR_PTR(err);
1023 }
1024 
1025 
1026 int mlx5_ib_destroy_cq(struct ib_cq *cq)
1027 {
1028 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1029 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1030 	struct ib_ucontext *context = NULL;
1031 
1032 	if (cq->uobject)
1033 		context = cq->uobject->context;
1034 
1035 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1036 	if (context)
1037 		destroy_cq_user(mcq, context);
1038 	else
1039 		destroy_cq_kernel(dev, mcq);
1040 
1041 	kfree(mcq);
1042 
1043 	return 0;
1044 }
1045 
1046 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1047 {
1048 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1049 }
1050 
1051 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1052 {
1053 	struct mlx5_cqe64 *cqe64, *dest64;
1054 	void *cqe, *dest;
1055 	u32 prod_index;
1056 	int nfreed = 0;
1057 	u8 owner_bit;
1058 
1059 	if (!cq)
1060 		return;
1061 
1062 	/* First we need to find the current producer index, so we
1063 	 * know where to start cleaning from.  It doesn't matter if HW
1064 	 * adds new entries after this loop -- the QP we're worried
1065 	 * about is already in RESET, so the new entries won't come
1066 	 * from our QP and therefore don't need to be checked.
1067 	 */
1068 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1069 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1070 			break;
1071 
1072 	/* Now sweep backwards through the CQ, removing CQ entries
1073 	 * that match our QP by copying older entries on top of them.
1074 	 */
1075 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1076 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1077 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1078 		if (is_equal_rsn(cqe64, rsn)) {
1079 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1080 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1081 			++nfreed;
1082 		} else if (nfreed) {
1083 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1084 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1085 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1086 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1087 			dest64->op_own = owner_bit |
1088 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1089 		}
1090 	}
1091 
1092 	if (nfreed) {
1093 		cq->mcq.cons_index += nfreed;
1094 		/* Make sure update of buffer contents is done before
1095 		 * updating consumer index.
1096 		 */
1097 		wmb();
1098 		mlx5_cq_set_ci(&cq->mcq);
1099 	}
1100 }
1101 
1102 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1103 {
1104 	if (!cq)
1105 		return;
1106 
1107 	spin_lock_irq(&cq->lock);
1108 	__mlx5_ib_cq_clean(cq, qpn, srq);
1109 	spin_unlock_irq(&cq->lock);
1110 }
1111 
1112 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1113 {
1114 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1115 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1116 	int err;
1117 
1118 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1119 		return -ENOSYS;
1120 
1121 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1122 					     cq_period, cq_count);
1123 	if (err)
1124 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1125 
1126 	return err;
1127 }
1128 
1129 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1130 		       int entries, struct ib_udata *udata, int *npas,
1131 		       int *page_shift, int *cqe_size)
1132 {
1133 	struct mlx5_ib_resize_cq ucmd;
1134 	struct ib_umem *umem;
1135 	int err;
1136 	int npages;
1137 	struct ib_ucontext *context = cq->buf.umem->context;
1138 
1139 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1140 	if (err)
1141 		return err;
1142 
1143 	if (ucmd.reserved0 || ucmd.reserved1)
1144 		return -EINVAL;
1145 
1146 	umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
1147 			   IB_ACCESS_LOCAL_WRITE, 1);
1148 	if (IS_ERR(umem)) {
1149 		err = PTR_ERR(umem);
1150 		return err;
1151 	}
1152 
1153 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1154 			   npas, NULL);
1155 
1156 	cq->resize_umem = umem;
1157 	*cqe_size = ucmd.cqe_size;
1158 
1159 	return 0;
1160 }
1161 
1162 static void un_resize_user(struct mlx5_ib_cq *cq)
1163 {
1164 	ib_umem_release(cq->resize_umem);
1165 }
1166 
1167 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1168 			 int entries, int cqe_size)
1169 {
1170 	int err;
1171 
1172 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1173 	if (!cq->resize_buf)
1174 		return -ENOMEM;
1175 
1176 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1177 	if (err)
1178 		goto ex;
1179 
1180 	init_cq_buf(cq, cq->resize_buf);
1181 
1182 	return 0;
1183 
1184 ex:
1185 	kfree(cq->resize_buf);
1186 	return err;
1187 }
1188 
1189 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1190 {
1191 	free_cq_buf(dev, cq->resize_buf);
1192 	cq->resize_buf = NULL;
1193 }
1194 
1195 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1196 {
1197 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1198 	struct mlx5_cqe64 *scqe64;
1199 	struct mlx5_cqe64 *dcqe64;
1200 	void *start_cqe;
1201 	void *scqe;
1202 	void *dcqe;
1203 	int ssize;
1204 	int dsize;
1205 	int i;
1206 	u8 sw_own;
1207 
1208 	ssize = cq->buf.cqe_size;
1209 	dsize = cq->resize_buf->cqe_size;
1210 	if (ssize != dsize) {
1211 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1212 		return -EINVAL;
1213 	}
1214 
1215 	i = cq->mcq.cons_index;
1216 	scqe = get_sw_cqe(cq, i);
1217 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1218 	start_cqe = scqe;
1219 	if (!scqe) {
1220 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1221 		return -EINVAL;
1222 	}
1223 
1224 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1225 		dcqe = get_cqe_from_buf(cq->resize_buf,
1226 					(i + 1) & (cq->resize_buf->nent),
1227 					dsize);
1228 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1229 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1230 		memcpy(dcqe, scqe, dsize);
1231 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1232 
1233 		++i;
1234 		scqe = get_sw_cqe(cq, i);
1235 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1236 		if (!scqe) {
1237 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1238 			return -EINVAL;
1239 		}
1240 
1241 		if (scqe == start_cqe) {
1242 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1243 				cq->mcq.cqn);
1244 			return -ENOMEM;
1245 		}
1246 	}
1247 	++cq->mcq.cons_index;
1248 	return 0;
1249 }
1250 
1251 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1252 {
1253 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1254 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1255 	void *cqc;
1256 	u32 *in;
1257 	int err;
1258 	int npas;
1259 	__be64 *pas;
1260 	int page_shift;
1261 	int inlen;
1262 	int uninitialized_var(cqe_size);
1263 	unsigned long flags;
1264 
1265 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1266 		pr_info("Firmware does not support resize CQ\n");
1267 		return -ENOSYS;
1268 	}
1269 
1270 	if (entries < 1 ||
1271 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1272 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1273 			     entries,
1274 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1275 		return -EINVAL;
1276 	}
1277 
1278 	entries = roundup_pow_of_two(entries + 1);
1279 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1280 		return -EINVAL;
1281 
1282 	if (entries == ibcq->cqe + 1)
1283 		return 0;
1284 
1285 	mutex_lock(&cq->resize_mutex);
1286 	if (udata) {
1287 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1288 				  &cqe_size);
1289 	} else {
1290 		cqe_size = 64;
1291 		err = resize_kernel(dev, cq, entries, cqe_size);
1292 		if (!err) {
1293 			npas = cq->resize_buf->buf.npages;
1294 			page_shift = cq->resize_buf->buf.page_shift;
1295 		}
1296 	}
1297 
1298 	if (err)
1299 		goto ex;
1300 
1301 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1302 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1303 
1304 	in = mlx5_vzalloc(inlen);
1305 	if (!in) {
1306 		err = -ENOMEM;
1307 		goto ex_resize;
1308 	}
1309 
1310 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1311 	if (udata)
1312 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1313 				     pas, 0);
1314 	else
1315 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1316 
1317 	MLX5_SET(modify_cq_in, in,
1318 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1319 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1320 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1321 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1322 
1323 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1324 
1325 	MLX5_SET(cqc, cqc, log_page_size,
1326 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1327 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1328 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1329 
1330 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1331 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1332 
1333 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1334 	if (err)
1335 		goto ex_alloc;
1336 
1337 	if (udata) {
1338 		cq->ibcq.cqe = entries - 1;
1339 		ib_umem_release(cq->buf.umem);
1340 		cq->buf.umem = cq->resize_umem;
1341 		cq->resize_umem = NULL;
1342 	} else {
1343 		struct mlx5_ib_cq_buf tbuf;
1344 		int resized = 0;
1345 
1346 		spin_lock_irqsave(&cq->lock, flags);
1347 		if (cq->resize_buf) {
1348 			err = copy_resize_cqes(cq);
1349 			if (!err) {
1350 				tbuf = cq->buf;
1351 				cq->buf = *cq->resize_buf;
1352 				kfree(cq->resize_buf);
1353 				cq->resize_buf = NULL;
1354 				resized = 1;
1355 			}
1356 		}
1357 		cq->ibcq.cqe = entries - 1;
1358 		spin_unlock_irqrestore(&cq->lock, flags);
1359 		if (resized)
1360 			free_cq_buf(dev, &tbuf);
1361 	}
1362 	mutex_unlock(&cq->resize_mutex);
1363 
1364 	kvfree(in);
1365 	return 0;
1366 
1367 ex_alloc:
1368 	kvfree(in);
1369 
1370 ex_resize:
1371 	if (udata)
1372 		un_resize_user(cq);
1373 	else
1374 		un_resize_kernel(dev, cq);
1375 ex:
1376 	mutex_unlock(&cq->resize_mutex);
1377 	return err;
1378 }
1379 
1380 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1381 {
1382 	struct mlx5_ib_cq *cq;
1383 
1384 	if (!ibcq)
1385 		return 128;
1386 
1387 	cq = to_mcq(ibcq);
1388 	return cq->cqe_size;
1389 }
1390 
1391 /* Called from atomic context */
1392 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1393 {
1394 	struct mlx5_ib_wc *soft_wc;
1395 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1396 	unsigned long flags;
1397 
1398 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1399 	if (!soft_wc)
1400 		return -ENOMEM;
1401 
1402 	soft_wc->wc = *wc;
1403 	spin_lock_irqsave(&cq->lock, flags);
1404 	list_add_tail(&soft_wc->list, &cq->wc_list);
1405 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1406 	    wc->status != IB_WC_SUCCESS) {
1407 		cq->notify_flags = 0;
1408 		schedule_work(&cq->notify_work);
1409 	}
1410 	spin_unlock_irqrestore(&cq->lock, flags);
1411 
1412 	return 0;
1413 }
1414