xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_cq.c (revision 3c4ba5f55438f7afd4f4b0b56f88f2bb505fd6a6)
1 /*-
2  * Copyright (c) 2013-2020, Mellanox Technologies.  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 "opt_rss.h"
29 #include "opt_ratelimit.h"
30 
31 #include <linux/kref.h>
32 #include <rdma/ib_umem.h>
33 #include <rdma/ib_user_verbs.h>
34 #include <rdma/ib_cache.h>
35 #include <rdma/uverbs_ioctl.h>
36 #include <dev/mlx5/mlx5_ib/mlx5_ib.h>
37 
38 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe __unused)
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, int 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 u8 sw_ownership_bit(int n, int nent)
77 {
78 	return (n & nent) ? 1 : 0;
79 }
80 
81 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
82 {
83 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
84 	struct mlx5_cqe64 *cqe64;
85 
86 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
87 
88 	if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) &&
89 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
90 		return cqe;
91 	} else {
92 		return NULL;
93 	}
94 }
95 
96 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
97 {
98 	return get_sw_cqe(cq, cq->mcq.cons_index);
99 }
100 
101 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
102 {
103 	switch (wq->wr_data[idx]) {
104 	case MLX5_IB_WR_UMR:
105 		return 0;
106 
107 	case IB_WR_LOCAL_INV:
108 		return IB_WC_LOCAL_INV;
109 
110 	case IB_WR_REG_MR:
111 		return IB_WC_REG_MR;
112 
113 	default:
114 		pr_warn("unknown completion status\n");
115 		return 0;
116 	}
117 }
118 
119 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
120 			    struct mlx5_ib_wq *wq, int idx)
121 {
122 	wc->wc_flags = 0;
123 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
124 	case MLX5_OPCODE_RDMA_WRITE_IMM:
125 		wc->wc_flags |= IB_WC_WITH_IMM;
126 	case MLX5_OPCODE_RDMA_WRITE:
127 		wc->opcode    = IB_WC_RDMA_WRITE;
128 		break;
129 	case MLX5_OPCODE_SEND_IMM:
130 		wc->wc_flags |= IB_WC_WITH_IMM;
131 	case MLX5_OPCODE_SEND:
132 	case MLX5_OPCODE_SEND_INVAL:
133 		wc->opcode    = IB_WC_SEND;
134 		break;
135 	case MLX5_OPCODE_RDMA_READ:
136 		wc->opcode    = IB_WC_RDMA_READ;
137 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
138 		break;
139 	case MLX5_OPCODE_ATOMIC_CS:
140 		wc->opcode    = IB_WC_COMP_SWAP;
141 		wc->byte_len  = 8;
142 		break;
143 	case MLX5_OPCODE_ATOMIC_FA:
144 		wc->opcode    = IB_WC_FETCH_ADD;
145 		wc->byte_len  = 8;
146 		break;
147 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
148 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
149 		wc->byte_len  = 8;
150 		break;
151 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
152 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
153 		wc->byte_len  = 8;
154 		break;
155 	case MLX5_OPCODE_UMR:
156 		wc->opcode = get_umr_comp(wq, idx);
157 		break;
158 	}
159 }
160 
161 enum {
162 	MLX5_GRH_IN_BUFFER = 1,
163 	MLX5_GRH_IN_CQE	   = 2,
164 };
165 
166 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
167 			     struct mlx5_ib_qp *qp)
168 {
169 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
170 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
171 	struct mlx5_ib_srq *srq;
172 	struct mlx5_ib_wq *wq;
173 	u16 wqe_ctr;
174 	u8  roce_packet_type;
175 	bool vlan_present;
176 	u8 g;
177 
178 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
179 		struct mlx5_core_srq *msrq = NULL;
180 
181 		if (qp->ibqp.xrcd) {
182 			msrq = mlx5_core_get_srq(dev->mdev,
183 						 be32_to_cpu(cqe->srqn));
184 			srq = to_mibsrq(msrq);
185 		} else {
186 			srq = to_msrq(qp->ibqp.srq);
187 		}
188 		if (srq) {
189 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
190 			wc->wr_id = srq->wrid[wqe_ctr];
191 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
192 			if (msrq && atomic_dec_and_test(&msrq->refcount))
193 				complete(&msrq->free);
194 		}
195 	} else {
196 		wq	  = &qp->rq;
197 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
198 		++wq->tail;
199 	}
200 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
201 
202 	switch (cqe->op_own >> 4) {
203 	case MLX5_CQE_RESP_WR_IMM:
204 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
205 		wc->wc_flags	= IB_WC_WITH_IMM;
206 		wc->ex.imm_data = cqe->imm_inval_pkey;
207 		break;
208 	case MLX5_CQE_RESP_SEND:
209 		wc->opcode   = IB_WC_RECV;
210 		wc->wc_flags = IB_WC_IP_CSUM_OK;
211 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
212 			       (cqe->hds_ip_ext & CQE_L4_OK))))
213 			wc->wc_flags = 0;
214 		break;
215 	case MLX5_CQE_RESP_SEND_IMM:
216 		wc->opcode	= IB_WC_RECV;
217 		wc->wc_flags	= IB_WC_WITH_IMM;
218 		wc->ex.imm_data = cqe->imm_inval_pkey;
219 		break;
220 	case MLX5_CQE_RESP_SEND_INV:
221 		wc->opcode	= IB_WC_RECV;
222 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
223 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
224 		break;
225 	}
226 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
227 	wc->dlid_path_bits = cqe->ml_path;
228 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
229 	wc->wc_flags |= g ? IB_WC_GRH : 0;
230 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
231 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
232 
233 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
234 				    &wc->pkey_index);
235 	} else {
236 		wc->pkey_index = 0;
237 	}
238 
239 	if (ll != IB_LINK_LAYER_ETHERNET) {
240 		wc->slid = be16_to_cpu(cqe->slid);
241 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
242 		return;
243 	}
244 
245 	wc->slid = 0;
246 	vlan_present = cqe_has_vlan(cqe);
247 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
248 	if (vlan_present) {
249 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
250 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
251 		wc->wc_flags |= IB_WC_WITH_VLAN;
252 	} else {
253 		wc->sl = 0;
254 	}
255 
256 	switch (roce_packet_type) {
257 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
258 		wc->network_hdr_type = RDMA_NETWORK_IB;
259 		break;
260 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
261 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
262 		break;
263 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
264 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
265 		break;
266 	}
267 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
268 }
269 
270 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
271 {
272 	__be32 *p = (__be32 *)cqe;
273 	int i;
274 
275 	mlx5_ib_warn(dev, "dump error cqe\n");
276 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
277 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
278 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
279 			be32_to_cpu(p[3]));
280 }
281 
282 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
283 				  struct mlx5_err_cqe *cqe,
284 				  struct ib_wc *wc)
285 {
286 	int dump = 1;
287 
288 	switch (cqe->syndrome) {
289 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
290 		wc->status = IB_WC_LOC_LEN_ERR;
291 		break;
292 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
293 		wc->status = IB_WC_LOC_QP_OP_ERR;
294 		break;
295 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
296 		wc->status = IB_WC_LOC_PROT_ERR;
297 		break;
298 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
299 		dump = 0;
300 		wc->status = IB_WC_WR_FLUSH_ERR;
301 		break;
302 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
303 		wc->status = IB_WC_MW_BIND_ERR;
304 		break;
305 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
306 		wc->status = IB_WC_BAD_RESP_ERR;
307 		break;
308 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
309 		wc->status = IB_WC_LOC_ACCESS_ERR;
310 		break;
311 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
312 		wc->status = IB_WC_REM_INV_REQ_ERR;
313 		break;
314 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
315 		wc->status = IB_WC_REM_ACCESS_ERR;
316 		break;
317 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
318 		wc->status = IB_WC_REM_OP_ERR;
319 		break;
320 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
321 		wc->status = IB_WC_RETRY_EXC_ERR;
322 		dump = 0;
323 		break;
324 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
325 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
326 		dump = 0;
327 		break;
328 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
329 		wc->status = IB_WC_REM_ABORT_ERR;
330 		break;
331 	default:
332 		wc->status = IB_WC_GENERAL_ERR;
333 		break;
334 	}
335 
336 	wc->vendor_err = cqe->vendor_err_synd;
337 	if (dump)
338 		dump_cqe(dev, cqe);
339 }
340 
341 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
342 {
343 	/* TBD: waiting decision
344 	*/
345 	return 0;
346 }
347 
348 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
349 {
350 	struct mlx5_wqe_data_seg *dpseg;
351 	void *addr;
352 
353 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
354 		sizeof(struct mlx5_wqe_raddr_seg) +
355 		sizeof(struct mlx5_wqe_atomic_seg);
356 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
357 	return addr;
358 }
359 
360 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
361 			  uint16_t idx)
362 {
363 	void *addr;
364 	int byte_count;
365 	int i;
366 
367 	if (!is_atomic_response(qp, idx))
368 		return;
369 
370 	byte_count = be32_to_cpu(cqe64->byte_cnt);
371 	addr = mlx5_get_atomic_laddr(qp, idx);
372 
373 	if (byte_count == 4) {
374 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
375 	} else {
376 		for (i = 0; i < byte_count; i += 8) {
377 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
378 			addr += 8;
379 		}
380 	}
381 
382 	return;
383 }
384 
385 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
386 			   u16 tail, u16 head)
387 {
388 	u16 idx;
389 
390 	do {
391 		idx = tail & (qp->sq.wqe_cnt - 1);
392 		handle_atomic(qp, cqe64, idx);
393 		if (idx == head)
394 			break;
395 
396 		tail = qp->sq.w_list[idx].next;
397 	} while (1);
398 	tail = qp->sq.w_list[idx].next;
399 	qp->sq.last_poll = tail;
400 }
401 
402 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
403 {
404 	mlx5_buf_free(dev->mdev, &buf->buf);
405 }
406 
407 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
408 			     struct ib_sig_err *item)
409 {
410 	u16 syndrome = be16_to_cpu(cqe->syndrome);
411 
412 #define GUARD_ERR   (1 << 13)
413 #define APPTAG_ERR  (1 << 12)
414 #define REFTAG_ERR  (1 << 11)
415 
416 	if (syndrome & GUARD_ERR) {
417 		item->err_type = IB_SIG_BAD_GUARD;
418 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
419 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
420 	} else
421 	if (syndrome & REFTAG_ERR) {
422 		item->err_type = IB_SIG_BAD_REFTAG;
423 		item->expected = be32_to_cpu(cqe->expected_reftag);
424 		item->actual = be32_to_cpu(cqe->actual_reftag);
425 	} else
426 	if (syndrome & APPTAG_ERR) {
427 		item->err_type = IB_SIG_BAD_APPTAG;
428 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
429 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
430 	} else {
431 		pr_err("Got signature completion error with bad syndrome %04x\n",
432 		       syndrome);
433 	}
434 
435 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
436 	item->key = be32_to_cpu(cqe->mkey);
437 }
438 
439 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
440 			 struct ib_wc *wc, int *npolled)
441 {
442 	struct mlx5_ib_wq *wq;
443 	unsigned int cur;
444 	unsigned int idx;
445 	int np;
446 	int i;
447 
448 	wq = &qp->sq;
449 	cur = wq->head - wq->tail;
450 	np = *npolled;
451 
452 	if (cur == 0)
453 		return;
454 
455 	for (i = 0;  i < cur && np < num_entries; i++) {
456 		idx = wq->last_poll & (wq->wqe_cnt - 1);
457 		wc->wr_id = wq->wrid[idx];
458 		wc->status = IB_WC_WR_FLUSH_ERR;
459 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
460 		wq->tail++;
461 		np++;
462 		wc->qp = &qp->ibqp;
463 		wc++;
464 		wq->last_poll = wq->w_list[idx].next;
465 	}
466 	*npolled = np;
467 }
468 
469 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
470 			 struct ib_wc *wc, int *npolled)
471 {
472 	struct mlx5_ib_wq *wq;
473 	unsigned int cur;
474 	int np;
475 	int i;
476 
477 	wq = &qp->rq;
478 	cur = wq->head - wq->tail;
479 	np = *npolled;
480 
481 	if (cur == 0)
482 		return;
483 
484 	for (i = 0;  i < cur && np < num_entries; i++) {
485 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
486 		wc->status = IB_WC_WR_FLUSH_ERR;
487 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
488 		wq->tail++;
489 		np++;
490 		wc->qp = &qp->ibqp;
491 		wc++;
492 	}
493 	*npolled = np;
494 }
495 
496 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
497 				 struct ib_wc *wc, int *npolled)
498 {
499 	struct mlx5_ib_qp *qp;
500 
501 	*npolled = 0;
502 	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
503 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
504 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
505 		if (*npolled >= num_entries)
506 			return;
507 	}
508 
509 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
510 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
511 		if (*npolled >= num_entries)
512 			return;
513 	}
514 }
515 
516 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
517 			 struct mlx5_ib_qp **cur_qp,
518 			 struct ib_wc *wc)
519 {
520 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
521 	struct mlx5_err_cqe *err_cqe;
522 	struct mlx5_cqe64 *cqe64;
523 	struct mlx5_core_qp *mqp;
524 	struct mlx5_ib_wq *wq;
525 	struct mlx5_sig_err_cqe *sig_err_cqe;
526 	struct mlx5_core_mkey *mmkey;
527 	struct mlx5_ib_mr *mr;
528 	unsigned long flags;
529 	uint8_t opcode;
530 	uint32_t qpn;
531 	u16 wqe_ctr;
532 	void *cqe;
533 	int idx;
534 
535 repoll:
536 	cqe = next_cqe_sw(cq);
537 	if (!cqe)
538 		return -EAGAIN;
539 
540 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
541 
542 	++cq->mcq.cons_index;
543 
544 	/* Make sure we read CQ entry contents after we've checked the
545 	 * ownership bit.
546 	 */
547 	rmb();
548 
549 	opcode = cqe64->op_own >> 4;
550 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
551 		if (likely(cq->resize_buf)) {
552 			free_cq_buf(dev, &cq->buf);
553 			cq->buf = *cq->resize_buf;
554 			kfree(cq->resize_buf);
555 			cq->resize_buf = NULL;
556 			goto repoll;
557 		} else {
558 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
559 		}
560 	}
561 
562 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
563 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
564 		/* We do not have to take the QP table lock here,
565 		 * because CQs will be locked while QPs are removed
566 		 * from the table.
567 		 */
568 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
569 		*cur_qp = to_mibqp(mqp);
570 	}
571 
572 	wc->qp  = &(*cur_qp)->ibqp;
573 	switch (opcode) {
574 	case MLX5_CQE_REQ:
575 		wq = &(*cur_qp)->sq;
576 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
577 		idx = wqe_ctr & (wq->wqe_cnt - 1);
578 		handle_good_req(wc, cqe64, wq, idx);
579 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
580 		wc->wr_id = wq->wrid[idx];
581 		wq->tail = wq->wqe_head[idx] + 1;
582 		wc->status = IB_WC_SUCCESS;
583 		break;
584 	case MLX5_CQE_RESP_WR_IMM:
585 	case MLX5_CQE_RESP_SEND:
586 	case MLX5_CQE_RESP_SEND_IMM:
587 	case MLX5_CQE_RESP_SEND_INV:
588 		handle_responder(wc, cqe64, *cur_qp);
589 		wc->status = IB_WC_SUCCESS;
590 		break;
591 	case MLX5_CQE_RESIZE_CQ:
592 		break;
593 	case MLX5_CQE_REQ_ERR:
594 	case MLX5_CQE_RESP_ERR:
595 		err_cqe = (struct mlx5_err_cqe *)cqe64;
596 		mlx5_handle_error_cqe(dev, err_cqe, wc);
597 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
598 			    opcode == MLX5_CQE_REQ_ERR ?
599 			    "Requestor" : "Responder", cq->mcq.cqn);
600 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
601 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
602 		if (opcode == MLX5_CQE_REQ_ERR) {
603 			wq = &(*cur_qp)->sq;
604 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
605 			idx = wqe_ctr & (wq->wqe_cnt - 1);
606 			wc->wr_id = wq->wrid[idx];
607 			wq->tail = wq->wqe_head[idx] + 1;
608 		} else {
609 			struct mlx5_ib_srq *srq;
610 
611 			if ((*cur_qp)->ibqp.srq) {
612 				srq = to_msrq((*cur_qp)->ibqp.srq);
613 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
614 				wc->wr_id = srq->wrid[wqe_ctr];
615 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
616 			} else {
617 				wq = &(*cur_qp)->rq;
618 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
619 				++wq->tail;
620 			}
621 		}
622 		break;
623 	case MLX5_CQE_SIG_ERR:
624 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
625 
626 		spin_lock_irqsave(&dev->mdev->priv.mr_table.lock, flags);
627 		mmkey = __mlx5_mr_lookup(dev->mdev,
628 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
629 		mr = to_mibmr(mmkey);
630 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
631 		mr->sig->sig_err_exists = true;
632 		mr->sig->sigerr_count++;
633 
634 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
635 			     cq->mcq.cqn, mr->sig->err_item.key,
636 			     mr->sig->err_item.err_type,
637 			     (long long)mr->sig->err_item.sig_err_offset,
638 			     mr->sig->err_item.expected,
639 			     mr->sig->err_item.actual);
640 
641 		spin_unlock_irqrestore(&dev->mdev->priv.mr_table.lock, flags);
642 		goto repoll;
643 	}
644 
645 	return 0;
646 }
647 
648 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
649 			struct ib_wc *wc)
650 {
651 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
652 	struct mlx5_ib_wc *soft_wc, *next;
653 	int npolled = 0;
654 
655 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
656 		if (npolled >= num_entries)
657 			break;
658 
659 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
660 			    cq->mcq.cqn);
661 
662 		wc[npolled++] = soft_wc->wc;
663 		list_del(&soft_wc->list);
664 		kfree(soft_wc);
665 	}
666 
667 	return npolled;
668 }
669 
670 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
671 {
672 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
673 	struct mlx5_ib_qp *cur_qp = NULL;
674 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
675 	struct mlx5_core_dev *mdev = dev->mdev;
676 	unsigned long flags;
677 	int soft_polled = 0;
678 	int npolled;
679 
680 	spin_lock_irqsave(&cq->lock, flags);
681 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)) {
682 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
683 		goto out;
684 	}
685 
686 	if (unlikely(!list_empty(&cq->wc_list)))
687 		soft_polled = poll_soft_wc(cq, num_entries, wc);
688 
689 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
690 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
691 			break;
692 	}
693 
694 	if (npolled)
695 		mlx5_cq_set_ci(&cq->mcq);
696 out:
697 	spin_unlock_irqrestore(&cq->lock, flags);
698 
699 	return soft_polled + npolled;
700 }
701 
702 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
703 {
704 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
705 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
706 	void __iomem *uar_page = mdev->priv.uar->map;
707 	unsigned long irq_flags;
708 	int ret = 0;
709 
710 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR))
711 		return -1;
712 
713 	spin_lock_irqsave(&cq->lock, irq_flags);
714 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
715 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
716 
717 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
718 		ret = 1;
719 	spin_unlock_irqrestore(&cq->lock, irq_flags);
720 
721 	mlx5_cq_arm(&cq->mcq,
722 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
723 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
724 		    uar_page,
725 		    MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock),
726 		    cq->mcq.cons_index);
727 
728 	return ret;
729 }
730 
731 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
732 			int nent, int cqe_size)
733 {
734 	int err;
735 
736 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size,
737 	    2 * PAGE_SIZE, &buf->buf);
738 	if (err)
739 		return err;
740 
741 	buf->cqe_size = cqe_size;
742 	buf->nent = nent;
743 
744 	return 0;
745 }
746 
747 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
748 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
749 			  int *cqe_size, int *index, int *inlen)
750 {
751 	struct mlx5_ib_create_cq ucmd = {};
752 	size_t ucmdlen;
753 	int page_shift;
754 	__be64 *pas;
755 	int npages;
756 	int ncont;
757 	void *cqc;
758 	int err;
759 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
760 		udata, struct mlx5_ib_ucontext, ibucontext);
761 
762 	ucmdlen = min(udata->inlen, sizeof(ucmd));
763 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
764 		return -EINVAL;
765 
766 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
767 		return -EFAULT;
768 
769 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX)))
770 		return -EINVAL;
771 
772 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
773 		return -EINVAL;
774 
775 	*cqe_size = ucmd.cqe_size;
776 
777 	cq->buf.umem = ib_umem_get(&context->ibucontext, ucmd.buf_addr,
778 				   entries * ucmd.cqe_size,
779 				   IB_ACCESS_LOCAL_WRITE, 1);
780 	if (IS_ERR(cq->buf.umem)) {
781 		err = PTR_ERR(cq->buf.umem);
782 		return err;
783 	}
784 
785 	err = mlx5_ib_db_map_user(context, ucmd.db_addr,
786 				  &cq->db);
787 	if (err)
788 		goto err_umem;
789 
790 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
791 			   &ncont, NULL);
792 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
793 		    (long long)ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
794 
795 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
796 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
797 	*cqb = mlx5_vzalloc(*inlen);
798 	if (!*cqb) {
799 		err = -ENOMEM;
800 		goto err_db;
801 	}
802 
803 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
804 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
805 
806 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
807 	MLX5_SET(cqc, cqc, log_page_size,
808 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
809 
810 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
811 		*index = ucmd.uar_page_index;
812 	} else if (context->bfregi.lib_uar_dyn) {
813 		err = -EINVAL;
814 		goto err_cqb;
815 	} else {
816 		*index = context->bfregi.sys_pages[0];
817 	}
818 
819 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
820 	return 0;
821 
822 err_cqb:
823 	kvfree(*cqb);
824 
825 err_db:
826 	mlx5_ib_db_unmap_user(context, &cq->db);
827 
828 err_umem:
829 	ib_umem_release(cq->buf.umem);
830 	return err;
831 }
832 
833 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
834 {
835 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
836 		udata, struct mlx5_ib_ucontext, ibucontext);
837 
838 	mlx5_ib_db_unmap_user(context, &cq->db);
839 	ib_umem_release(cq->buf.umem);
840 }
841 
842 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
843 {
844 	int i;
845 	void *cqe;
846 	struct mlx5_cqe64 *cqe64;
847 
848 	for (i = 0; i < buf->nent; i++) {
849 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
850 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
851 		cqe64->op_own = MLX5_CQE_INVALID << 4;
852 	}
853 }
854 
855 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
856 			    int entries, int cqe_size,
857 			    u32 **cqb, int *index, int *inlen)
858 {
859 	__be64 *pas;
860 	void *cqc;
861 	int err;
862 
863 	err = mlx5_db_alloc(dev->mdev, &cq->db);
864 	if (err)
865 		return err;
866 
867 	cq->mcq.set_ci_db  = cq->db.db;
868 	cq->mcq.arm_db     = cq->db.db + 1;
869 	cq->mcq.cqe_sz = cqe_size;
870 
871 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
872 	if (err)
873 		goto err_db;
874 
875 	init_cq_buf(cq, &cq->buf);
876 
877 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
878 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
879 	*cqb = mlx5_vzalloc(*inlen);
880 	if (!*cqb) {
881 		err = -ENOMEM;
882 		goto err_buf;
883 	}
884 
885 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
886 	mlx5_fill_page_array(&cq->buf.buf, pas);
887 
888 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
889 	MLX5_SET(cqc, cqc, log_page_size,
890 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
891 
892 	*index = dev->mdev->priv.uar->index;
893 
894 	return 0;
895 
896 err_buf:
897 	free_cq_buf(dev, &cq->buf);
898 
899 err_db:
900 	mlx5_db_free(dev->mdev, &cq->db);
901 	return err;
902 }
903 
904 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
905 {
906 	free_cq_buf(dev, &cq->buf);
907 	mlx5_db_free(dev->mdev, &cq->db);
908 }
909 
910 static void notify_soft_wc_handler(struct work_struct *work)
911 {
912 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
913 					     notify_work);
914 
915 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
916 }
917 
918 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
919 		      struct ib_udata *udata)
920 {
921 	struct ib_device *ibdev = ibcq->device;
922 	int entries = attr->cqe;
923 	int vector = attr->comp_vector;
924 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
925 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
926 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
927 	int uninitialized_var(index);
928 	int uninitialized_var(inlen);
929 	u32 *cqb = NULL;
930 	void *cqc;
931 	int cqe_size;
932 	unsigned int irqn;
933 	int eqn;
934 	int err;
935 
936 	if (entries < 0 ||
937 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
938 		return -EINVAL;
939 
940 	if (check_cq_create_flags(attr->flags))
941 		return -EOPNOTSUPP;
942 
943 	entries = roundup_pow_of_two(entries + 1);
944 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
945 		return -EINVAL;
946 
947 	cq->ibcq.cqe = entries - 1;
948 	mutex_init(&cq->resize_mutex);
949 	spin_lock_init(&cq->lock);
950 	cq->resize_buf = NULL;
951 	cq->resize_umem = NULL;
952 	cq->create_flags = attr->flags;
953 	INIT_LIST_HEAD(&cq->list_send_qp);
954 	INIT_LIST_HEAD(&cq->list_recv_qp);
955 
956 	if (udata) {
957 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
958 				     &index, &inlen);
959 		if (err)
960 			return err;
961 	} else {
962 		cqe_size = cache_line_size() == 128 ? 128 : 64;
963 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
964 				       &index, &inlen);
965 		if (err)
966 			return err;
967 
968 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
969 	}
970 
971 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
972 	if (err)
973 		goto err_cqb;
974 
975 	cq->cqe_size = cqe_size;
976 
977 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
978 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
979 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
980 	MLX5_SET(cqc, cqc, uar_page, index);
981 	MLX5_SET(cqc, cqc, c_eqn, eqn);
982 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
983 	if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN)
984 		MLX5_SET(cqc, cqc, oi, 1);
985 
986 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
987 	if (err)
988 		goto err_cqb;
989 
990 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
991 	cq->mcq.irqn = irqn;
992 	cq->mcq.comp  = mlx5_ib_cq_comp;
993 	cq->mcq.event = mlx5_ib_cq_event;
994 
995 	INIT_LIST_HEAD(&cq->wc_list);
996 
997 	if (udata)
998 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
999 			err = -EFAULT;
1000 			goto err_cmd;
1001 		}
1002 
1003 
1004 	kvfree(cqb);
1005 	return 0;
1006 
1007 err_cmd:
1008 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1009 
1010 err_cqb:
1011 	kvfree(cqb);
1012 	if (udata)
1013 		destroy_cq_user(cq, udata);
1014 	else
1015 		destroy_cq_kernel(dev, cq);
1016 	return err;
1017 }
1018 
1019 void mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1020 {
1021 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1022 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1023 
1024 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1025 	if (udata)
1026 		destroy_cq_user(mcq, udata);
1027 	else
1028 		destroy_cq_kernel(dev, mcq);
1029 }
1030 
1031 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1032 {
1033 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1034 }
1035 
1036 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1037 {
1038 	struct mlx5_cqe64 *cqe64, *dest64;
1039 	void *cqe, *dest;
1040 	u32 prod_index;
1041 	int nfreed = 0;
1042 	u8 owner_bit;
1043 
1044 	if (!cq)
1045 		return;
1046 
1047 	/* First we need to find the current producer index, so we
1048 	 * know where to start cleaning from.  It doesn't matter if HW
1049 	 * adds new entries after this loop -- the QP we're worried
1050 	 * about is already in RESET, so the new entries won't come
1051 	 * from our QP and therefore don't need to be checked.
1052 	 */
1053 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1054 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1055 			break;
1056 
1057 	/* Now sweep backwards through the CQ, removing CQ entries
1058 	 * that match our QP by copying older entries on top of them.
1059 	 */
1060 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1061 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1062 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1063 		if (is_equal_rsn(cqe64, rsn)) {
1064 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1065 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1066 			++nfreed;
1067 		} else if (nfreed) {
1068 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1069 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1070 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1071 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1072 			dest64->op_own = owner_bit |
1073 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1074 		}
1075 	}
1076 
1077 	if (nfreed) {
1078 		cq->mcq.cons_index += nfreed;
1079 		/* Make sure update of buffer contents is done before
1080 		 * updating consumer index.
1081 		 */
1082 		wmb();
1083 		mlx5_cq_set_ci(&cq->mcq);
1084 	}
1085 }
1086 
1087 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1088 {
1089 	if (!cq)
1090 		return;
1091 
1092 	spin_lock_irq(&cq->lock);
1093 	__mlx5_ib_cq_clean(cq, qpn, srq);
1094 	spin_unlock_irq(&cq->lock);
1095 }
1096 
1097 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1098 {
1099 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1100 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1101 	int err;
1102 
1103 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1104 		return -ENOSYS;
1105 
1106 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1107 					     cq_period, cq_count);
1108 	if (err)
1109 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1110 
1111 	return err;
1112 }
1113 
1114 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1115 		       int entries, struct ib_udata *udata, int *npas,
1116 		       int *page_shift, int *cqe_size)
1117 {
1118 	struct mlx5_ib_resize_cq ucmd;
1119 	struct ib_umem *umem;
1120 	int err;
1121 	int npages;
1122 	struct ib_ucontext *context = cq->buf.umem->context;
1123 
1124 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1125 	if (err)
1126 		return err;
1127 
1128 	if (ucmd.reserved0 || ucmd.reserved1)
1129 		return -EINVAL;
1130 
1131 	/* check multiplication overflow */
1132 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1133 		return -EINVAL;
1134 
1135 	umem = ib_umem_get(context, ucmd.buf_addr,
1136 			   (size_t)ucmd.cqe_size * entries,
1137 			   IB_ACCESS_LOCAL_WRITE, 1);
1138 	if (IS_ERR(umem)) {
1139 		err = PTR_ERR(umem);
1140 		return err;
1141 	}
1142 
1143 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1144 			   npas, NULL);
1145 
1146 	cq->resize_umem = umem;
1147 	*cqe_size = ucmd.cqe_size;
1148 
1149 	return 0;
1150 }
1151 
1152 static void un_resize_user(struct mlx5_ib_cq *cq)
1153 {
1154 	ib_umem_release(cq->resize_umem);
1155 }
1156 
1157 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1158 			 int entries, int cqe_size)
1159 {
1160 	int err;
1161 
1162 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1163 	if (!cq->resize_buf)
1164 		return -ENOMEM;
1165 
1166 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1167 	if (err)
1168 		goto ex;
1169 
1170 	init_cq_buf(cq, cq->resize_buf);
1171 
1172 	return 0;
1173 
1174 ex:
1175 	kfree(cq->resize_buf);
1176 	return err;
1177 }
1178 
1179 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1180 {
1181 	free_cq_buf(dev, cq->resize_buf);
1182 	cq->resize_buf = NULL;
1183 }
1184 
1185 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1186 {
1187 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1188 	struct mlx5_cqe64 *scqe64;
1189 	struct mlx5_cqe64 *dcqe64;
1190 	void *start_cqe;
1191 	void *scqe;
1192 	void *dcqe;
1193 	int ssize;
1194 	int dsize;
1195 	int i;
1196 	u8 sw_own;
1197 
1198 	ssize = cq->buf.cqe_size;
1199 	dsize = cq->resize_buf->cqe_size;
1200 	if (ssize != dsize) {
1201 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1202 		return -EINVAL;
1203 	}
1204 
1205 	i = cq->mcq.cons_index;
1206 	scqe = get_sw_cqe(cq, i);
1207 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1208 	start_cqe = scqe;
1209 	if (!scqe) {
1210 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1211 		return -EINVAL;
1212 	}
1213 
1214 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1215 		dcqe = get_cqe_from_buf(cq->resize_buf,
1216 					(i + 1) & (cq->resize_buf->nent),
1217 					dsize);
1218 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1219 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1220 		memcpy(dcqe, scqe, dsize);
1221 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1222 
1223 		++i;
1224 		scqe = get_sw_cqe(cq, i);
1225 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1226 		if (!scqe) {
1227 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1228 			return -EINVAL;
1229 		}
1230 
1231 		if (scqe == start_cqe) {
1232 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1233 				cq->mcq.cqn);
1234 			return -ENOMEM;
1235 		}
1236 	}
1237 	++cq->mcq.cons_index;
1238 	return 0;
1239 }
1240 
1241 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1242 {
1243 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1244 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1245 	void *cqc;
1246 	u32 *in;
1247 	int err;
1248 	int npas;
1249 	__be64 *pas;
1250 	int page_shift;
1251 	int inlen;
1252 	int uninitialized_var(cqe_size);
1253 	unsigned long flags;
1254 
1255 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1256 		pr_info("Firmware does not support resize CQ\n");
1257 		return -ENOSYS;
1258 	}
1259 
1260 	if (entries < 1 ||
1261 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1262 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1263 			     entries,
1264 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1265 		return -EINVAL;
1266 	}
1267 
1268 	entries = roundup_pow_of_two(entries + 1);
1269 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1270 		return -EINVAL;
1271 
1272 	if (entries == ibcq->cqe + 1)
1273 		return 0;
1274 
1275 	mutex_lock(&cq->resize_mutex);
1276 	if (udata) {
1277 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1278 				  &cqe_size);
1279 	} else {
1280 		cqe_size = 64;
1281 		err = resize_kernel(dev, cq, entries, cqe_size);
1282 		if (!err) {
1283 			npas = cq->resize_buf->buf.npages;
1284 			page_shift = cq->resize_buf->buf.page_shift;
1285 		}
1286 	}
1287 
1288 	if (err)
1289 		goto ex;
1290 
1291 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1292 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1293 
1294 	in = mlx5_vzalloc(inlen);
1295 	if (!in) {
1296 		err = -ENOMEM;
1297 		goto ex_resize;
1298 	}
1299 
1300 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1301 	if (udata)
1302 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1303 				     pas, 0);
1304 	else
1305 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1306 
1307 	MLX5_SET(modify_cq_in, in,
1308 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1309 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1310 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1311 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1312 
1313 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1314 
1315 	MLX5_SET(cqc, cqc, log_page_size,
1316 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1317 	MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size));
1318 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1319 
1320 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1321 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1322 
1323 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1324 	if (err)
1325 		goto ex_alloc;
1326 
1327 	if (udata) {
1328 		cq->ibcq.cqe = entries - 1;
1329 		ib_umem_release(cq->buf.umem);
1330 		cq->buf.umem = cq->resize_umem;
1331 		cq->resize_umem = NULL;
1332 	} else {
1333 		struct mlx5_ib_cq_buf tbuf;
1334 		int resized = 0;
1335 
1336 		spin_lock_irqsave(&cq->lock, flags);
1337 		if (cq->resize_buf) {
1338 			err = copy_resize_cqes(cq);
1339 			if (!err) {
1340 				tbuf = cq->buf;
1341 				cq->buf = *cq->resize_buf;
1342 				kfree(cq->resize_buf);
1343 				cq->resize_buf = NULL;
1344 				resized = 1;
1345 			}
1346 		}
1347 		cq->ibcq.cqe = entries - 1;
1348 		spin_unlock_irqrestore(&cq->lock, flags);
1349 		if (resized)
1350 			free_cq_buf(dev, &tbuf);
1351 	}
1352 	mutex_unlock(&cq->resize_mutex);
1353 
1354 	kvfree(in);
1355 	return 0;
1356 
1357 ex_alloc:
1358 	kvfree(in);
1359 
1360 ex_resize:
1361 	if (udata)
1362 		un_resize_user(cq);
1363 	else
1364 		un_resize_kernel(dev, cq);
1365 ex:
1366 	mutex_unlock(&cq->resize_mutex);
1367 	return err;
1368 }
1369 
1370 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1371 {
1372 	struct mlx5_ib_cq *cq;
1373 
1374 	if (!ibcq)
1375 		return 128;
1376 
1377 	cq = to_mcq(ibcq);
1378 	return cq->cqe_size;
1379 }
1380 
1381 /* Called from atomic context */
1382 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1383 {
1384 	struct mlx5_ib_wc *soft_wc;
1385 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1386 	unsigned long flags;
1387 
1388 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1389 	if (!soft_wc)
1390 		return -ENOMEM;
1391 
1392 	soft_wc->wc = *wc;
1393 	spin_lock_irqsave(&cq->lock, flags);
1394 	list_add_tail(&soft_wc->list, &cq->wc_list);
1395 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1396 	    wc->status != IB_WC_SUCCESS) {
1397 		cq->notify_flags = 0;
1398 		schedule_work(&cq->notify_work);
1399 	}
1400 	spin_unlock_irqrestore(&cq->lock, flags);
1401 
1402 	return 0;
1403 }
1404