xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision a34a3ed7)
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 		/* fall through */
128 	case MLX5_OPCODE_RDMA_WRITE:
129 		wc->opcode    = IB_WC_RDMA_WRITE;
130 		break;
131 	case MLX5_OPCODE_SEND_IMM:
132 		wc->wc_flags |= IB_WC_WITH_IMM;
133 		/* fall through */
134 	case MLX5_OPCODE_SEND:
135 	case MLX5_OPCODE_SEND_INVAL:
136 		wc->opcode    = IB_WC_SEND;
137 		break;
138 	case MLX5_OPCODE_RDMA_READ:
139 		wc->opcode    = IB_WC_RDMA_READ;
140 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
141 		break;
142 	case MLX5_OPCODE_ATOMIC_CS:
143 		wc->opcode    = IB_WC_COMP_SWAP;
144 		wc->byte_len  = 8;
145 		break;
146 	case MLX5_OPCODE_ATOMIC_FA:
147 		wc->opcode    = IB_WC_FETCH_ADD;
148 		wc->byte_len  = 8;
149 		break;
150 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
151 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
152 		wc->byte_len  = 8;
153 		break;
154 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
155 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
156 		wc->byte_len  = 8;
157 		break;
158 	case MLX5_OPCODE_UMR:
159 		wc->opcode = get_umr_comp(wq, idx);
160 		break;
161 	}
162 }
163 
164 enum {
165 	MLX5_GRH_IN_BUFFER = 1,
166 	MLX5_GRH_IN_CQE	   = 2,
167 };
168 
169 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
170 			     struct mlx5_ib_qp *qp)
171 {
172 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
173 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
174 	struct mlx5_ib_srq *srq;
175 	struct mlx5_ib_wq *wq;
176 	u16 wqe_ctr;
177 	u8  roce_packet_type;
178 	bool vlan_present;
179 	u8 g;
180 
181 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
182 		struct mlx5_core_srq *msrq = NULL;
183 
184 		if (qp->ibqp.xrcd) {
185 			msrq = mlx5_core_get_srq(dev->mdev,
186 						 be32_to_cpu(cqe->srqn));
187 			srq = to_mibsrq(msrq);
188 		} else {
189 			srq = to_msrq(qp->ibqp.srq);
190 		}
191 		if (srq) {
192 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
193 			wc->wr_id = srq->wrid[wqe_ctr];
194 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
195 			if (msrq && atomic_dec_and_test(&msrq->refcount))
196 				complete(&msrq->free);
197 		}
198 	} else {
199 		wq	  = &qp->rq;
200 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
201 		++wq->tail;
202 	}
203 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
204 
205 	switch (cqe->op_own >> 4) {
206 	case MLX5_CQE_RESP_WR_IMM:
207 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
208 		wc->wc_flags	= IB_WC_WITH_IMM;
209 		wc->ex.imm_data = cqe->imm_inval_pkey;
210 		break;
211 	case MLX5_CQE_RESP_SEND:
212 		wc->opcode   = IB_WC_RECV;
213 		wc->wc_flags = IB_WC_IP_CSUM_OK;
214 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
215 			       (cqe->hds_ip_ext & CQE_L4_OK))))
216 			wc->wc_flags = 0;
217 		break;
218 	case MLX5_CQE_RESP_SEND_IMM:
219 		wc->opcode	= IB_WC_RECV;
220 		wc->wc_flags	= IB_WC_WITH_IMM;
221 		wc->ex.imm_data = cqe->imm_inval_pkey;
222 		break;
223 	case MLX5_CQE_RESP_SEND_INV:
224 		wc->opcode	= IB_WC_RECV;
225 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
226 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
227 		break;
228 	}
229 	wc->slid	   = be16_to_cpu(cqe->slid);
230 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
231 	wc->dlid_path_bits = cqe->ml_path;
232 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
233 	wc->wc_flags |= g ? IB_WC_GRH : 0;
234 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
235 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
236 
237 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
238 				    &wc->pkey_index);
239 	} else {
240 		wc->pkey_index = 0;
241 	}
242 
243 	if (ll != IB_LINK_LAYER_ETHERNET) {
244 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
245 		return;
246 	}
247 
248 	vlan_present = cqe->l4_l3_hdr_type & 0x1;
249 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
250 	if (vlan_present) {
251 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
252 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
253 		wc->wc_flags |= IB_WC_WITH_VLAN;
254 	} else {
255 		wc->sl = 0;
256 	}
257 
258 	switch (roce_packet_type) {
259 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
260 		wc->network_hdr_type = RDMA_NETWORK_IB;
261 		break;
262 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
263 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
264 		break;
265 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
266 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
267 		break;
268 	}
269 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
270 }
271 
272 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
273 {
274 	__be32 *p = (__be32 *)cqe;
275 	int i;
276 
277 	mlx5_ib_warn(dev, "dump error cqe\n");
278 	for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4)
279 		pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]),
280 			be32_to_cpu(p[1]), be32_to_cpu(p[2]),
281 			be32_to_cpu(p[3]));
282 }
283 
284 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
285 				  struct mlx5_err_cqe *cqe,
286 				  struct ib_wc *wc)
287 {
288 	int dump = 1;
289 
290 	switch (cqe->syndrome) {
291 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
292 		wc->status = IB_WC_LOC_LEN_ERR;
293 		break;
294 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
295 		wc->status = IB_WC_LOC_QP_OP_ERR;
296 		break;
297 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
298 		wc->status = IB_WC_LOC_PROT_ERR;
299 		break;
300 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
301 		dump = 0;
302 		wc->status = IB_WC_WR_FLUSH_ERR;
303 		break;
304 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
305 		wc->status = IB_WC_MW_BIND_ERR;
306 		break;
307 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
308 		wc->status = IB_WC_BAD_RESP_ERR;
309 		break;
310 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
311 		wc->status = IB_WC_LOC_ACCESS_ERR;
312 		break;
313 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
314 		wc->status = IB_WC_REM_INV_REQ_ERR;
315 		break;
316 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
317 		wc->status = IB_WC_REM_ACCESS_ERR;
318 		break;
319 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
320 		wc->status = IB_WC_REM_OP_ERR;
321 		break;
322 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
323 		wc->status = IB_WC_RETRY_EXC_ERR;
324 		dump = 0;
325 		break;
326 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
327 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
328 		dump = 0;
329 		break;
330 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
331 		wc->status = IB_WC_REM_ABORT_ERR;
332 		break;
333 	default:
334 		wc->status = IB_WC_GENERAL_ERR;
335 		break;
336 	}
337 
338 	wc->vendor_err = cqe->vendor_err_synd;
339 	if (dump)
340 		dump_cqe(dev, cqe);
341 }
342 
343 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx)
344 {
345 	/* TBD: waiting decision
346 	*/
347 	return 0;
348 }
349 
350 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx)
351 {
352 	struct mlx5_wqe_data_seg *dpseg;
353 	void *addr;
354 
355 	dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) +
356 		sizeof(struct mlx5_wqe_raddr_seg) +
357 		sizeof(struct mlx5_wqe_atomic_seg);
358 	addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr);
359 	return addr;
360 }
361 
362 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
363 			  uint16_t idx)
364 {
365 	void *addr;
366 	int byte_count;
367 	int i;
368 
369 	if (!is_atomic_response(qp, idx))
370 		return;
371 
372 	byte_count = be32_to_cpu(cqe64->byte_cnt);
373 	addr = mlx5_get_atomic_laddr(qp, idx);
374 
375 	if (byte_count == 4) {
376 		*(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr));
377 	} else {
378 		for (i = 0; i < byte_count; i += 8) {
379 			*(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr));
380 			addr += 8;
381 		}
382 	}
383 
384 	return;
385 }
386 
387 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
388 			   u16 tail, u16 head)
389 {
390 	u16 idx;
391 
392 	do {
393 		idx = tail & (qp->sq.wqe_cnt - 1);
394 		handle_atomic(qp, cqe64, idx);
395 		if (idx == head)
396 			break;
397 
398 		tail = qp->sq.w_list[idx].next;
399 	} while (1);
400 	tail = qp->sq.w_list[idx].next;
401 	qp->sq.last_poll = tail;
402 }
403 
404 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
405 {
406 	mlx5_buf_free(dev->mdev, &buf->buf);
407 }
408 
409 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
410 			     struct ib_sig_err *item)
411 {
412 	u16 syndrome = be16_to_cpu(cqe->syndrome);
413 
414 #define GUARD_ERR   (1 << 13)
415 #define APPTAG_ERR  (1 << 12)
416 #define REFTAG_ERR  (1 << 11)
417 
418 	if (syndrome & GUARD_ERR) {
419 		item->err_type = IB_SIG_BAD_GUARD;
420 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
421 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
422 	} else
423 	if (syndrome & REFTAG_ERR) {
424 		item->err_type = IB_SIG_BAD_REFTAG;
425 		item->expected = be32_to_cpu(cqe->expected_reftag);
426 		item->actual = be32_to_cpu(cqe->actual_reftag);
427 	} else
428 	if (syndrome & APPTAG_ERR) {
429 		item->err_type = IB_SIG_BAD_APPTAG;
430 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
431 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
432 	} else {
433 		pr_err("Got signature completion error with bad syndrome %04x\n",
434 		       syndrome);
435 	}
436 
437 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
438 	item->key = be32_to_cpu(cqe->mkey);
439 }
440 
441 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
442 			 struct ib_wc *wc, int *npolled)
443 {
444 	struct mlx5_ib_wq *wq;
445 	unsigned int cur;
446 	unsigned int idx;
447 	int np;
448 	int i;
449 
450 	wq = &qp->sq;
451 	cur = wq->head - wq->tail;
452 	np = *npolled;
453 
454 	if (cur == 0)
455 		return;
456 
457 	for (i = 0;  i < cur && np < num_entries; i++) {
458 		idx = wq->last_poll & (wq->wqe_cnt - 1);
459 		wc->wr_id = wq->wrid[idx];
460 		wc->status = IB_WC_WR_FLUSH_ERR;
461 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
462 		wq->tail++;
463 		np++;
464 		wc->qp = &qp->ibqp;
465 		wc++;
466 		wq->last_poll = wq->w_list[idx].next;
467 	}
468 	*npolled = np;
469 }
470 
471 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
472 			 struct ib_wc *wc, int *npolled)
473 {
474 	struct mlx5_ib_wq *wq;
475 	unsigned int cur;
476 	int np;
477 	int i;
478 
479 	wq = &qp->rq;
480 	cur = wq->head - wq->tail;
481 	np = *npolled;
482 
483 	if (cur == 0)
484 		return;
485 
486 	for (i = 0;  i < cur && np < num_entries; i++) {
487 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
488 		wc->status = IB_WC_WR_FLUSH_ERR;
489 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
490 		wq->tail++;
491 		np++;
492 		wc->qp = &qp->ibqp;
493 		wc++;
494 	}
495 	*npolled = np;
496 }
497 
498 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
499 				 struct ib_wc *wc, int *npolled)
500 {
501 	struct mlx5_ib_qp *qp;
502 
503 	*npolled = 0;
504 	/* Find uncompleted WQEs belonging to that cq and return mmics ones */
505 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
506 		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
507 		if (*npolled >= num_entries)
508 			return;
509 	}
510 
511 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
512 		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
513 		if (*npolled >= num_entries)
514 			return;
515 	}
516 }
517 
518 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
519 			 struct mlx5_ib_qp **cur_qp,
520 			 struct ib_wc *wc)
521 {
522 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
523 	struct mlx5_err_cqe *err_cqe;
524 	struct mlx5_cqe64 *cqe64;
525 	struct mlx5_core_qp *mqp;
526 	struct mlx5_ib_wq *wq;
527 	struct mlx5_sig_err_cqe *sig_err_cqe;
528 	struct mlx5_core_mkey *mmkey;
529 	struct mlx5_ib_mr *mr;
530 	uint8_t opcode;
531 	uint32_t qpn;
532 	u16 wqe_ctr;
533 	void *cqe;
534 	int idx;
535 
536 repoll:
537 	cqe = next_cqe_sw(cq);
538 	if (!cqe)
539 		return -EAGAIN;
540 
541 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
542 
543 	++cq->mcq.cons_index;
544 
545 	/* Make sure we read CQ entry contents after we've checked the
546 	 * ownership bit.
547 	 */
548 	rmb();
549 
550 	opcode = cqe64->op_own >> 4;
551 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
552 		if (likely(cq->resize_buf)) {
553 			free_cq_buf(dev, &cq->buf);
554 			cq->buf = *cq->resize_buf;
555 			kfree(cq->resize_buf);
556 			cq->resize_buf = NULL;
557 			goto repoll;
558 		} else {
559 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
560 		}
561 	}
562 
563 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
564 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
565 		/* We do not have to take the QP table lock here,
566 		 * because CQs will be locked while QPs are removed
567 		 * from the table.
568 		 */
569 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
570 		*cur_qp = to_mibqp(mqp);
571 	}
572 
573 	wc->qp  = &(*cur_qp)->ibqp;
574 	switch (opcode) {
575 	case MLX5_CQE_REQ:
576 		wq = &(*cur_qp)->sq;
577 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
578 		idx = wqe_ctr & (wq->wqe_cnt - 1);
579 		handle_good_req(wc, cqe64, wq, idx);
580 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
581 		wc->wr_id = wq->wrid[idx];
582 		wq->tail = wq->wqe_head[idx] + 1;
583 		wc->status = IB_WC_SUCCESS;
584 		break;
585 	case MLX5_CQE_RESP_WR_IMM:
586 	case MLX5_CQE_RESP_SEND:
587 	case MLX5_CQE_RESP_SEND_IMM:
588 	case MLX5_CQE_RESP_SEND_INV:
589 		handle_responder(wc, cqe64, *cur_qp);
590 		wc->status = IB_WC_SUCCESS;
591 		break;
592 	case MLX5_CQE_RESIZE_CQ:
593 		break;
594 	case MLX5_CQE_REQ_ERR:
595 	case MLX5_CQE_RESP_ERR:
596 		err_cqe = (struct mlx5_err_cqe *)cqe64;
597 		mlx5_handle_error_cqe(dev, err_cqe, wc);
598 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
599 			    opcode == MLX5_CQE_REQ_ERR ?
600 			    "Requestor" : "Responder", cq->mcq.cqn);
601 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
602 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
603 		if (opcode == MLX5_CQE_REQ_ERR) {
604 			wq = &(*cur_qp)->sq;
605 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
606 			idx = wqe_ctr & (wq->wqe_cnt - 1);
607 			wc->wr_id = wq->wrid[idx];
608 			wq->tail = wq->wqe_head[idx] + 1;
609 		} else {
610 			struct mlx5_ib_srq *srq;
611 
612 			if ((*cur_qp)->ibqp.srq) {
613 				srq = to_msrq((*cur_qp)->ibqp.srq);
614 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
615 				wc->wr_id = srq->wrid[wqe_ctr];
616 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
617 			} else {
618 				wq = &(*cur_qp)->rq;
619 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
620 				++wq->tail;
621 			}
622 		}
623 		break;
624 	case MLX5_CQE_SIG_ERR:
625 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
626 
627 		read_lock(&dev->mdev->priv.mkey_table.lock);
628 		mmkey = __mlx5_mr_lookup(dev->mdev,
629 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
630 		mr = to_mibmr(mmkey);
631 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
632 		mr->sig->sig_err_exists = true;
633 		mr->sig->sigerr_count++;
634 
635 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
636 			     cq->mcq.cqn, mr->sig->err_item.key,
637 			     mr->sig->err_item.err_type,
638 			     mr->sig->err_item.sig_err_offset,
639 			     mr->sig->err_item.expected,
640 			     mr->sig->err_item.actual);
641 
642 		read_unlock(&dev->mdev->priv.mkey_table.lock);
643 		goto repoll;
644 	}
645 
646 	return 0;
647 }
648 
649 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
650 			struct ib_wc *wc)
651 {
652 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
653 	struct mlx5_ib_wc *soft_wc, *next;
654 	int npolled = 0;
655 
656 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
657 		if (npolled >= num_entries)
658 			break;
659 
660 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
661 			    cq->mcq.cqn);
662 
663 		wc[npolled++] = soft_wc->wc;
664 		list_del(&soft_wc->list);
665 		kfree(soft_wc);
666 	}
667 
668 	return npolled;
669 }
670 
671 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
672 {
673 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
674 	struct mlx5_ib_qp *cur_qp = NULL;
675 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
676 	struct mlx5_core_dev *mdev = dev->mdev;
677 	unsigned long flags;
678 	int soft_polled = 0;
679 	int npolled;
680 
681 	spin_lock_irqsave(&cq->lock, flags);
682 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
683 		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
684 		goto out;
685 	}
686 
687 	if (unlikely(!list_empty(&cq->wc_list)))
688 		soft_polled = poll_soft_wc(cq, num_entries, wc);
689 
690 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
691 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
692 			break;
693 	}
694 
695 	if (npolled)
696 		mlx5_cq_set_ci(&cq->mcq);
697 out:
698 	spin_unlock_irqrestore(&cq->lock, flags);
699 
700 	return soft_polled + npolled;
701 }
702 
703 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
704 {
705 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
706 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
707 	void __iomem *uar_page = mdev->priv.uar->map;
708 	unsigned long irq_flags;
709 	int ret = 0;
710 
711 	spin_lock_irqsave(&cq->lock, irq_flags);
712 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
713 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
714 
715 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
716 		ret = 1;
717 	spin_unlock_irqrestore(&cq->lock, irq_flags);
718 
719 	mlx5_cq_arm(&cq->mcq,
720 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
721 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
722 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
723 
724 	return ret;
725 }
726 
727 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf,
728 			int nent, int cqe_size)
729 {
730 	int err;
731 
732 	err = mlx5_buf_alloc(dev->mdev, nent * cqe_size, &buf->buf);
733 	if (err)
734 		return err;
735 
736 	buf->cqe_size = cqe_size;
737 	buf->nent = nent;
738 
739 	return 0;
740 }
741 
742 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
743 			  struct ib_ucontext *context, struct mlx5_ib_cq *cq,
744 			  int entries, u32 **cqb,
745 			  int *cqe_size, int *index, int *inlen)
746 {
747 	struct mlx5_ib_create_cq ucmd = {};
748 	size_t ucmdlen;
749 	int page_shift;
750 	__be64 *pas;
751 	int npages;
752 	int ncont;
753 	void *cqc;
754 	int err;
755 
756 	ucmdlen = udata->inlen < sizeof(ucmd) ?
757 		  (sizeof(ucmd) - sizeof(ucmd.flags)) : sizeof(ucmd);
758 
759 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
760 		return -EFAULT;
761 
762 	if (ucmdlen == sizeof(ucmd) &&
763 	    (ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD)))
764 		return -EINVAL;
765 
766 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
767 		return -EINVAL;
768 
769 	*cqe_size = ucmd.cqe_size;
770 
771 	cq->buf.umem = ib_umem_get(context, ucmd.buf_addr,
772 				   entries * ucmd.cqe_size,
773 				   IB_ACCESS_LOCAL_WRITE, 1);
774 	if (IS_ERR(cq->buf.umem)) {
775 		err = PTR_ERR(cq->buf.umem);
776 		return err;
777 	}
778 
779 	err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
780 				  &cq->db);
781 	if (err)
782 		goto err_umem;
783 
784 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
785 			   &ncont, NULL);
786 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
787 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
788 
789 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
790 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
791 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
792 	if (!*cqb) {
793 		err = -ENOMEM;
794 		goto err_db;
795 	}
796 
797 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
798 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
799 
800 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
801 	MLX5_SET(cqc, cqc, log_page_size,
802 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
803 
804 	*index = to_mucontext(context)->bfregi.sys_pages[0];
805 
806 	if (ucmd.cqe_comp_en == 1) {
807 		if (!((*cqe_size == 128 &&
808 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
809 		      (*cqe_size == 64  &&
810 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
811 			err = -EOPNOTSUPP;
812 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
813 				     *cqe_size);
814 			goto err_cqb;
815 		}
816 
817 		if (unlikely(!ucmd.cqe_comp_res_format ||
818 			     !(ucmd.cqe_comp_res_format <
819 			       MLX5_IB_CQE_RES_RESERVED) ||
820 			     (ucmd.cqe_comp_res_format &
821 			      (ucmd.cqe_comp_res_format - 1)))) {
822 			err = -EOPNOTSUPP;
823 			mlx5_ib_warn(dev, "CQE compression res format %d is not supported!\n",
824 				     ucmd.cqe_comp_res_format);
825 			goto err_cqb;
826 		}
827 
828 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
829 		MLX5_SET(cqc, cqc, mini_cqe_res_format,
830 			 ilog2(ucmd.cqe_comp_res_format));
831 	}
832 
833 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
834 		if (*cqe_size != 128 ||
835 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
836 			err = -EOPNOTSUPP;
837 			mlx5_ib_warn(dev,
838 				     "CQE padding is not supported for CQE size of %dB!\n",
839 				     *cqe_size);
840 			goto err_cqb;
841 		}
842 
843 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
844 	}
845 
846 	return 0;
847 
848 err_cqb:
849 	kfree(*cqb);
850 
851 err_db:
852 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
853 
854 err_umem:
855 	ib_umem_release(cq->buf.umem);
856 	return err;
857 }
858 
859 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context)
860 {
861 	mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db);
862 	ib_umem_release(cq->buf.umem);
863 }
864 
865 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf)
866 {
867 	int i;
868 	void *cqe;
869 	struct mlx5_cqe64 *cqe64;
870 
871 	for (i = 0; i < buf->nent; i++) {
872 		cqe = get_cqe_from_buf(buf, i, buf->cqe_size);
873 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
874 		cqe64->op_own = MLX5_CQE_INVALID << 4;
875 	}
876 }
877 
878 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
879 			    int entries, int cqe_size,
880 			    u32 **cqb, int *index, int *inlen)
881 {
882 	__be64 *pas;
883 	void *cqc;
884 	int err;
885 
886 	err = mlx5_db_alloc(dev->mdev, &cq->db);
887 	if (err)
888 		return err;
889 
890 	cq->mcq.set_ci_db  = cq->db.db;
891 	cq->mcq.arm_db     = cq->db.db + 1;
892 	cq->mcq.cqe_sz = cqe_size;
893 
894 	err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size);
895 	if (err)
896 		goto err_db;
897 
898 	init_cq_buf(cq, &cq->buf);
899 
900 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
901 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages;
902 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
903 	if (!*cqb) {
904 		err = -ENOMEM;
905 		goto err_buf;
906 	}
907 
908 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
909 	mlx5_fill_page_array(&cq->buf.buf, pas);
910 
911 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
912 	MLX5_SET(cqc, cqc, log_page_size,
913 		 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
914 
915 	*index = dev->mdev->priv.uar->index;
916 
917 	return 0;
918 
919 err_buf:
920 	free_cq_buf(dev, &cq->buf);
921 
922 err_db:
923 	mlx5_db_free(dev->mdev, &cq->db);
924 	return err;
925 }
926 
927 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
928 {
929 	free_cq_buf(dev, &cq->buf);
930 	mlx5_db_free(dev->mdev, &cq->db);
931 }
932 
933 static void notify_soft_wc_handler(struct work_struct *work)
934 {
935 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
936 					     notify_work);
937 
938 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
939 }
940 
941 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
942 				const struct ib_cq_init_attr *attr,
943 				struct ib_ucontext *context,
944 				struct ib_udata *udata)
945 {
946 	int entries = attr->cqe;
947 	int vector = attr->comp_vector;
948 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
949 	struct mlx5_ib_cq *cq;
950 	int uninitialized_var(index);
951 	int uninitialized_var(inlen);
952 	u32 *cqb = NULL;
953 	void *cqc;
954 	int cqe_size;
955 	unsigned int irqn;
956 	int eqn;
957 	int err;
958 
959 	if (entries < 0 ||
960 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
961 		return ERR_PTR(-EINVAL);
962 
963 	if (check_cq_create_flags(attr->flags))
964 		return ERR_PTR(-EOPNOTSUPP);
965 
966 	entries = roundup_pow_of_two(entries + 1);
967 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
968 		return ERR_PTR(-EINVAL);
969 
970 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
971 	if (!cq)
972 		return ERR_PTR(-ENOMEM);
973 
974 	cq->ibcq.cqe = entries - 1;
975 	mutex_init(&cq->resize_mutex);
976 	spin_lock_init(&cq->lock);
977 	cq->resize_buf = NULL;
978 	cq->resize_umem = NULL;
979 	cq->create_flags = attr->flags;
980 	INIT_LIST_HEAD(&cq->list_send_qp);
981 	INIT_LIST_HEAD(&cq->list_recv_qp);
982 
983 	if (context) {
984 		err = create_cq_user(dev, udata, context, cq, entries,
985 				     &cqb, &cqe_size, &index, &inlen);
986 		if (err)
987 			goto err_create;
988 	} else {
989 		cqe_size = cache_line_size() == 128 ? 128 : 64;
990 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
991 				       &index, &inlen);
992 		if (err)
993 			goto err_create;
994 
995 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
996 	}
997 
998 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
999 	if (err)
1000 		goto err_cqb;
1001 
1002 	cq->cqe_size = cqe_size;
1003 
1004 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
1005 	MLX5_SET(cqc, cqc, cqe_sz,
1006 		 cqe_sz_to_mlx_sz(cqe_size,
1007 				  cq->private_flags &
1008 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1009 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1010 	MLX5_SET(cqc, cqc, uar_page, index);
1011 	MLX5_SET(cqc, cqc, c_eqn, eqn);
1012 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
1013 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
1014 		MLX5_SET(cqc, cqc, oi, 1);
1015 
1016 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
1017 	if (err)
1018 		goto err_cqb;
1019 
1020 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
1021 	cq->mcq.irqn = irqn;
1022 	if (context)
1023 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
1024 	else
1025 		cq->mcq.comp  = mlx5_ib_cq_comp;
1026 	cq->mcq.event = mlx5_ib_cq_event;
1027 
1028 	INIT_LIST_HEAD(&cq->wc_list);
1029 
1030 	if (context)
1031 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
1032 			err = -EFAULT;
1033 			goto err_cmd;
1034 		}
1035 
1036 
1037 	kvfree(cqb);
1038 	return &cq->ibcq;
1039 
1040 err_cmd:
1041 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1042 
1043 err_cqb:
1044 	kvfree(cqb);
1045 	if (context)
1046 		destroy_cq_user(cq, context);
1047 	else
1048 		destroy_cq_kernel(dev, cq);
1049 
1050 err_create:
1051 	kfree(cq);
1052 
1053 	return ERR_PTR(err);
1054 }
1055 
1056 
1057 int mlx5_ib_destroy_cq(struct ib_cq *cq)
1058 {
1059 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1060 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1061 	struct ib_ucontext *context = NULL;
1062 
1063 	if (cq->uobject)
1064 		context = cq->uobject->context;
1065 
1066 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1067 	if (context)
1068 		destroy_cq_user(mcq, context);
1069 	else
1070 		destroy_cq_kernel(dev, mcq);
1071 
1072 	kfree(mcq);
1073 
1074 	return 0;
1075 }
1076 
1077 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1078 {
1079 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1080 }
1081 
1082 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1083 {
1084 	struct mlx5_cqe64 *cqe64, *dest64;
1085 	void *cqe, *dest;
1086 	u32 prod_index;
1087 	int nfreed = 0;
1088 	u8 owner_bit;
1089 
1090 	if (!cq)
1091 		return;
1092 
1093 	/* First we need to find the current producer index, so we
1094 	 * know where to start cleaning from.  It doesn't matter if HW
1095 	 * adds new entries after this loop -- the QP we're worried
1096 	 * about is already in RESET, so the new entries won't come
1097 	 * from our QP and therefore don't need to be checked.
1098 	 */
1099 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1100 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1101 			break;
1102 
1103 	/* Now sweep backwards through the CQ, removing CQ entries
1104 	 * that match our QP by copying older entries on top of them.
1105 	 */
1106 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1107 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1108 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1109 		if (is_equal_rsn(cqe64, rsn)) {
1110 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1111 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1112 			++nfreed;
1113 		} else if (nfreed) {
1114 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1115 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1116 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1117 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1118 			dest64->op_own = owner_bit |
1119 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1120 		}
1121 	}
1122 
1123 	if (nfreed) {
1124 		cq->mcq.cons_index += nfreed;
1125 		/* Make sure update of buffer contents is done before
1126 		 * updating consumer index.
1127 		 */
1128 		wmb();
1129 		mlx5_cq_set_ci(&cq->mcq);
1130 	}
1131 }
1132 
1133 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1134 {
1135 	if (!cq)
1136 		return;
1137 
1138 	spin_lock_irq(&cq->lock);
1139 	__mlx5_ib_cq_clean(cq, qpn, srq);
1140 	spin_unlock_irq(&cq->lock);
1141 }
1142 
1143 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1144 {
1145 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1146 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1147 	int err;
1148 
1149 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1150 		return -ENOSYS;
1151 
1152 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1153 		return -EINVAL;
1154 
1155 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1156 					     cq_period, cq_count);
1157 	if (err)
1158 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1159 
1160 	return err;
1161 }
1162 
1163 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1164 		       int entries, struct ib_udata *udata, int *npas,
1165 		       int *page_shift, int *cqe_size)
1166 {
1167 	struct mlx5_ib_resize_cq ucmd;
1168 	struct ib_umem *umem;
1169 	int err;
1170 	int npages;
1171 	struct ib_ucontext *context = cq->buf.umem->context;
1172 
1173 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1174 	if (err)
1175 		return err;
1176 
1177 	if (ucmd.reserved0 || ucmd.reserved1)
1178 		return -EINVAL;
1179 
1180 	umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size,
1181 			   IB_ACCESS_LOCAL_WRITE, 1);
1182 	if (IS_ERR(umem)) {
1183 		err = PTR_ERR(umem);
1184 		return err;
1185 	}
1186 
1187 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1188 			   npas, NULL);
1189 
1190 	cq->resize_umem = umem;
1191 	*cqe_size = ucmd.cqe_size;
1192 
1193 	return 0;
1194 }
1195 
1196 static void un_resize_user(struct mlx5_ib_cq *cq)
1197 {
1198 	ib_umem_release(cq->resize_umem);
1199 }
1200 
1201 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1202 			 int entries, int cqe_size)
1203 {
1204 	int err;
1205 
1206 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1207 	if (!cq->resize_buf)
1208 		return -ENOMEM;
1209 
1210 	err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size);
1211 	if (err)
1212 		goto ex;
1213 
1214 	init_cq_buf(cq, cq->resize_buf);
1215 
1216 	return 0;
1217 
1218 ex:
1219 	kfree(cq->resize_buf);
1220 	return err;
1221 }
1222 
1223 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1224 {
1225 	free_cq_buf(dev, cq->resize_buf);
1226 	cq->resize_buf = NULL;
1227 }
1228 
1229 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1230 {
1231 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1232 	struct mlx5_cqe64 *scqe64;
1233 	struct mlx5_cqe64 *dcqe64;
1234 	void *start_cqe;
1235 	void *scqe;
1236 	void *dcqe;
1237 	int ssize;
1238 	int dsize;
1239 	int i;
1240 	u8 sw_own;
1241 
1242 	ssize = cq->buf.cqe_size;
1243 	dsize = cq->resize_buf->cqe_size;
1244 	if (ssize != dsize) {
1245 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1246 		return -EINVAL;
1247 	}
1248 
1249 	i = cq->mcq.cons_index;
1250 	scqe = get_sw_cqe(cq, i);
1251 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1252 	start_cqe = scqe;
1253 	if (!scqe) {
1254 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1255 		return -EINVAL;
1256 	}
1257 
1258 	while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) {
1259 		dcqe = get_cqe_from_buf(cq->resize_buf,
1260 					(i + 1) & (cq->resize_buf->nent),
1261 					dsize);
1262 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1263 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1264 		memcpy(dcqe, scqe, dsize);
1265 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1266 
1267 		++i;
1268 		scqe = get_sw_cqe(cq, i);
1269 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1270 		if (!scqe) {
1271 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1272 			return -EINVAL;
1273 		}
1274 
1275 		if (scqe == start_cqe) {
1276 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1277 				cq->mcq.cqn);
1278 			return -ENOMEM;
1279 		}
1280 	}
1281 	++cq->mcq.cons_index;
1282 	return 0;
1283 }
1284 
1285 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1286 {
1287 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1288 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1289 	void *cqc;
1290 	u32 *in;
1291 	int err;
1292 	int npas;
1293 	__be64 *pas;
1294 	int page_shift;
1295 	int inlen;
1296 	int uninitialized_var(cqe_size);
1297 	unsigned long flags;
1298 
1299 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1300 		pr_info("Firmware does not support resize CQ\n");
1301 		return -ENOSYS;
1302 	}
1303 
1304 	if (entries < 1 ||
1305 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1306 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1307 			     entries,
1308 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1309 		return -EINVAL;
1310 	}
1311 
1312 	entries = roundup_pow_of_two(entries + 1);
1313 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1314 		return -EINVAL;
1315 
1316 	if (entries == ibcq->cqe + 1)
1317 		return 0;
1318 
1319 	mutex_lock(&cq->resize_mutex);
1320 	if (udata) {
1321 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1322 				  &cqe_size);
1323 	} else {
1324 		cqe_size = 64;
1325 		err = resize_kernel(dev, cq, entries, cqe_size);
1326 		if (!err) {
1327 			npas = cq->resize_buf->buf.npages;
1328 			page_shift = cq->resize_buf->buf.page_shift;
1329 		}
1330 	}
1331 
1332 	if (err)
1333 		goto ex;
1334 
1335 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1336 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1337 
1338 	in = kvzalloc(inlen, GFP_KERNEL);
1339 	if (!in) {
1340 		err = -ENOMEM;
1341 		goto ex_resize;
1342 	}
1343 
1344 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1345 	if (udata)
1346 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1347 				     pas, 0);
1348 	else
1349 		mlx5_fill_page_array(&cq->resize_buf->buf, pas);
1350 
1351 	MLX5_SET(modify_cq_in, in,
1352 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1353 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1354 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1355 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1356 
1357 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1358 
1359 	MLX5_SET(cqc, cqc, log_page_size,
1360 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1361 	MLX5_SET(cqc, cqc, cqe_sz,
1362 		 cqe_sz_to_mlx_sz(cqe_size,
1363 				  cq->private_flags &
1364 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1365 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1366 
1367 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1368 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1369 
1370 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1371 	if (err)
1372 		goto ex_alloc;
1373 
1374 	if (udata) {
1375 		cq->ibcq.cqe = entries - 1;
1376 		ib_umem_release(cq->buf.umem);
1377 		cq->buf.umem = cq->resize_umem;
1378 		cq->resize_umem = NULL;
1379 	} else {
1380 		struct mlx5_ib_cq_buf tbuf;
1381 		int resized = 0;
1382 
1383 		spin_lock_irqsave(&cq->lock, flags);
1384 		if (cq->resize_buf) {
1385 			err = copy_resize_cqes(cq);
1386 			if (!err) {
1387 				tbuf = cq->buf;
1388 				cq->buf = *cq->resize_buf;
1389 				kfree(cq->resize_buf);
1390 				cq->resize_buf = NULL;
1391 				resized = 1;
1392 			}
1393 		}
1394 		cq->ibcq.cqe = entries - 1;
1395 		spin_unlock_irqrestore(&cq->lock, flags);
1396 		if (resized)
1397 			free_cq_buf(dev, &tbuf);
1398 	}
1399 	mutex_unlock(&cq->resize_mutex);
1400 
1401 	kvfree(in);
1402 	return 0;
1403 
1404 ex_alloc:
1405 	kvfree(in);
1406 
1407 ex_resize:
1408 	if (udata)
1409 		un_resize_user(cq);
1410 	else
1411 		un_resize_kernel(dev, cq);
1412 ex:
1413 	mutex_unlock(&cq->resize_mutex);
1414 	return err;
1415 }
1416 
1417 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq)
1418 {
1419 	struct mlx5_ib_cq *cq;
1420 
1421 	if (!ibcq)
1422 		return 128;
1423 
1424 	cq = to_mcq(ibcq);
1425 	return cq->cqe_size;
1426 }
1427 
1428 /* Called from atomic context */
1429 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1430 {
1431 	struct mlx5_ib_wc *soft_wc;
1432 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1433 	unsigned long flags;
1434 
1435 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1436 	if (!soft_wc)
1437 		return -ENOMEM;
1438 
1439 	soft_wc->wc = *wc;
1440 	spin_lock_irqsave(&cq->lock, flags);
1441 	list_add_tail(&soft_wc->list, &cq->wc_list);
1442 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1443 	    wc->status != IB_WC_SUCCESS) {
1444 		cq->notify_flags = 0;
1445 		schedule_work(&cq->notify_work);
1446 	}
1447 	spin_unlock_irqrestore(&cq->lock, flags);
1448 
1449 	return 0;
1450 }
1451