xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision aac28965)
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 #include "srq.h"
39 #include "qp.h"
40 
41 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe)
42 {
43 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
44 
45 	ibcq->comp_handler(ibcq, ibcq->cq_context);
46 }
47 
48 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
49 {
50 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
51 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
52 	struct ib_cq *ibcq = &cq->ibcq;
53 	struct ib_event event;
54 
55 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
56 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
57 			     type, mcq->cqn);
58 		return;
59 	}
60 
61 	if (ibcq->event_handler) {
62 		event.device     = &dev->ib_dev;
63 		event.event      = IB_EVENT_CQ_ERR;
64 		event.element.cq = ibcq;
65 		ibcq->event_handler(&event, ibcq->cq_context);
66 	}
67 }
68 
69 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
70 {
71 	return mlx5_frag_buf_get_wqe(&cq->buf.fbc, n);
72 }
73 
74 static u8 sw_ownership_bit(int n, int nent)
75 {
76 	return (n & nent) ? 1 : 0;
77 }
78 
79 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
80 {
81 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
82 	struct mlx5_cqe64 *cqe64;
83 
84 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
85 
86 	if (likely(get_cqe_opcode(cqe64) != MLX5_CQE_INVALID) &&
87 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
88 		return cqe;
89 	} else {
90 		return NULL;
91 	}
92 }
93 
94 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
95 {
96 	return get_sw_cqe(cq, cq->mcq.cons_index);
97 }
98 
99 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
100 {
101 	switch (wq->wr_data[idx]) {
102 	case MLX5_IB_WR_UMR:
103 		return 0;
104 
105 	case IB_WR_LOCAL_INV:
106 		return IB_WC_LOCAL_INV;
107 
108 	case IB_WR_REG_MR:
109 		return IB_WC_REG_MR;
110 
111 	default:
112 		pr_warn("unknown completion status\n");
113 		return 0;
114 	}
115 }
116 
117 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
118 			    struct mlx5_ib_wq *wq, int idx)
119 {
120 	wc->wc_flags = 0;
121 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
122 	case MLX5_OPCODE_RDMA_WRITE_IMM:
123 		wc->wc_flags |= IB_WC_WITH_IMM;
124 		fallthrough;
125 	case MLX5_OPCODE_RDMA_WRITE:
126 		wc->opcode    = IB_WC_RDMA_WRITE;
127 		break;
128 	case MLX5_OPCODE_SEND_IMM:
129 		wc->wc_flags |= IB_WC_WITH_IMM;
130 		fallthrough;
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 = NULL;
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_cmd_get_srq(dev, be32_to_cpu(cqe->srqn));
183 			if (msrq)
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)
193 				mlx5_core_res_put(&msrq->common);
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 (get_cqe_opcode(cqe)) {
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->immediate;
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->immediate;
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->inval_rkey);
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 (is_qp1(qp->type)) {
231 		u16 pkey = be32_to_cpu(cqe->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->l4_l3_hdr_type & 0x1;
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_ROCE_V1;
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 	mlx5_ib_warn(dev, "dump error cqe\n");
273 	mlx5_dump_err_cqe(dev->mdev, cqe);
274 }
275 
276 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
277 				  struct mlx5_err_cqe *cqe,
278 				  struct ib_wc *wc)
279 {
280 	int dump = 1;
281 
282 	switch (cqe->syndrome) {
283 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
284 		wc->status = IB_WC_LOC_LEN_ERR;
285 		break;
286 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
287 		wc->status = IB_WC_LOC_QP_OP_ERR;
288 		break;
289 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
290 		wc->status = IB_WC_LOC_PROT_ERR;
291 		break;
292 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
293 		dump = 0;
294 		wc->status = IB_WC_WR_FLUSH_ERR;
295 		break;
296 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
297 		wc->status = IB_WC_MW_BIND_ERR;
298 		break;
299 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
300 		wc->status = IB_WC_BAD_RESP_ERR;
301 		break;
302 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
303 		wc->status = IB_WC_LOC_ACCESS_ERR;
304 		break;
305 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
306 		wc->status = IB_WC_REM_INV_REQ_ERR;
307 		break;
308 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
309 		wc->status = IB_WC_REM_ACCESS_ERR;
310 		break;
311 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
312 		wc->status = IB_WC_REM_OP_ERR;
313 		break;
314 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
315 		wc->status = IB_WC_RETRY_EXC_ERR;
316 		dump = 0;
317 		break;
318 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
319 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
320 		dump = 0;
321 		break;
322 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
323 		wc->status = IB_WC_REM_ABORT_ERR;
324 		break;
325 	default:
326 		wc->status = IB_WC_GENERAL_ERR;
327 		break;
328 	}
329 
330 	wc->vendor_err = cqe->vendor_err_synd;
331 	if (dump) {
332 		mlx5_ib_warn(dev, "WC error: %d, Message: %s\n", wc->status,
333 			     ib_wc_status_msg(wc->status));
334 		dump_cqe(dev, cqe);
335 	}
336 }
337 
338 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
339 			   u16 tail, u16 head)
340 {
341 	u16 idx;
342 
343 	do {
344 		idx = tail & (qp->sq.wqe_cnt - 1);
345 		if (idx == head)
346 			break;
347 
348 		tail = qp->sq.w_list[idx].next;
349 	} while (1);
350 	tail = qp->sq.w_list[idx].next;
351 	qp->sq.last_poll = tail;
352 }
353 
354 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
355 {
356 	mlx5_frag_buf_free(dev->mdev, &buf->frag_buf);
357 }
358 
359 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
360 			     struct ib_sig_err *item)
361 {
362 	u16 syndrome = be16_to_cpu(cqe->syndrome);
363 
364 #define GUARD_ERR   (1 << 13)
365 #define APPTAG_ERR  (1 << 12)
366 #define REFTAG_ERR  (1 << 11)
367 
368 	if (syndrome & GUARD_ERR) {
369 		item->err_type = IB_SIG_BAD_GUARD;
370 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
371 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
372 	} else
373 	if (syndrome & REFTAG_ERR) {
374 		item->err_type = IB_SIG_BAD_REFTAG;
375 		item->expected = be32_to_cpu(cqe->expected_reftag);
376 		item->actual = be32_to_cpu(cqe->actual_reftag);
377 	} else
378 	if (syndrome & APPTAG_ERR) {
379 		item->err_type = IB_SIG_BAD_APPTAG;
380 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
381 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
382 	} else {
383 		pr_err("Got signature completion error with bad syndrome %04x\n",
384 		       syndrome);
385 	}
386 
387 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
388 	item->key = be32_to_cpu(cqe->mkey);
389 }
390 
391 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
392 		    int *npolled, bool is_send)
393 {
394 	struct mlx5_ib_wq *wq;
395 	unsigned int cur;
396 	int np;
397 	int i;
398 
399 	wq = (is_send) ? &qp->sq : &qp->rq;
400 	cur = wq->head - wq->tail;
401 	np = *npolled;
402 
403 	if (cur == 0)
404 		return;
405 
406 	for (i = 0;  i < cur && np < num_entries; i++) {
407 		unsigned int idx;
408 
409 		idx = (is_send) ? wq->last_poll : wq->tail;
410 		idx &= (wq->wqe_cnt - 1);
411 		wc->wr_id = wq->wrid[idx];
412 		wc->status = IB_WC_WR_FLUSH_ERR;
413 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
414 		wq->tail++;
415 		if (is_send)
416 			wq->last_poll = wq->w_list[idx].next;
417 		np++;
418 		wc->qp = &qp->ibqp;
419 		wc++;
420 	}
421 	*npolled = np;
422 }
423 
424 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
425 				 struct ib_wc *wc, int *npolled)
426 {
427 	struct mlx5_ib_qp *qp;
428 
429 	*npolled = 0;
430 	/* Find uncompleted WQEs belonging to that cq and return mmics ones */
431 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
432 		sw_comp(qp, num_entries, wc + *npolled, npolled, true);
433 		if (*npolled >= num_entries)
434 			return;
435 	}
436 
437 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
438 		sw_comp(qp, num_entries, wc + *npolled, npolled, false);
439 		if (*npolled >= num_entries)
440 			return;
441 	}
442 }
443 
444 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
445 			 struct mlx5_ib_qp **cur_qp,
446 			 struct ib_wc *wc)
447 {
448 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
449 	struct mlx5_err_cqe *err_cqe;
450 	struct mlx5_cqe64 *cqe64;
451 	struct mlx5_core_qp *mqp;
452 	struct mlx5_ib_wq *wq;
453 	uint8_t opcode;
454 	uint32_t qpn;
455 	u16 wqe_ctr;
456 	void *cqe;
457 	int idx;
458 
459 repoll:
460 	cqe = next_cqe_sw(cq);
461 	if (!cqe)
462 		return -EAGAIN;
463 
464 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
465 
466 	++cq->mcq.cons_index;
467 
468 	/* Make sure we read CQ entry contents after we've checked the
469 	 * ownership bit.
470 	 */
471 	rmb();
472 
473 	opcode = get_cqe_opcode(cqe64);
474 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
475 		if (likely(cq->resize_buf)) {
476 			free_cq_buf(dev, &cq->buf);
477 			cq->buf = *cq->resize_buf;
478 			kfree(cq->resize_buf);
479 			cq->resize_buf = NULL;
480 			goto repoll;
481 		} else {
482 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
483 		}
484 	}
485 
486 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
487 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
488 		/* We do not have to take the QP table lock here,
489 		 * because CQs will be locked while QPs are removed
490 		 * from the table.
491 		 */
492 		mqp = radix_tree_lookup(&dev->qp_table.tree, qpn);
493 		*cur_qp = to_mibqp(mqp);
494 	}
495 
496 	wc->qp  = &(*cur_qp)->ibqp;
497 	switch (opcode) {
498 	case MLX5_CQE_REQ:
499 		wq = &(*cur_qp)->sq;
500 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
501 		idx = wqe_ctr & (wq->wqe_cnt - 1);
502 		handle_good_req(wc, cqe64, wq, idx);
503 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
504 		wc->wr_id = wq->wrid[idx];
505 		wq->tail = wq->wqe_head[idx] + 1;
506 		wc->status = IB_WC_SUCCESS;
507 		break;
508 	case MLX5_CQE_RESP_WR_IMM:
509 	case MLX5_CQE_RESP_SEND:
510 	case MLX5_CQE_RESP_SEND_IMM:
511 	case MLX5_CQE_RESP_SEND_INV:
512 		handle_responder(wc, cqe64, *cur_qp);
513 		wc->status = IB_WC_SUCCESS;
514 		break;
515 	case MLX5_CQE_RESIZE_CQ:
516 		break;
517 	case MLX5_CQE_REQ_ERR:
518 	case MLX5_CQE_RESP_ERR:
519 		err_cqe = (struct mlx5_err_cqe *)cqe64;
520 		mlx5_handle_error_cqe(dev, err_cqe, wc);
521 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
522 			    opcode == MLX5_CQE_REQ_ERR ?
523 			    "Requestor" : "Responder", cq->mcq.cqn);
524 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
525 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
526 		if (opcode == MLX5_CQE_REQ_ERR) {
527 			wq = &(*cur_qp)->sq;
528 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
529 			idx = wqe_ctr & (wq->wqe_cnt - 1);
530 			wc->wr_id = wq->wrid[idx];
531 			wq->tail = wq->wqe_head[idx] + 1;
532 		} else {
533 			struct mlx5_ib_srq *srq;
534 
535 			if ((*cur_qp)->ibqp.srq) {
536 				srq = to_msrq((*cur_qp)->ibqp.srq);
537 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
538 				wc->wr_id = srq->wrid[wqe_ctr];
539 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
540 			} else {
541 				wq = &(*cur_qp)->rq;
542 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
543 				++wq->tail;
544 			}
545 		}
546 		break;
547 	case MLX5_CQE_SIG_ERR: {
548 		struct mlx5_sig_err_cqe *sig_err_cqe =
549 			(struct mlx5_sig_err_cqe *)cqe64;
550 		struct mlx5_core_sig_ctx *sig;
551 
552 		xa_lock(&dev->sig_mrs);
553 		sig = xa_load(&dev->sig_mrs,
554 				mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
555 		get_sig_err_item(sig_err_cqe, &sig->err_item);
556 		sig->sig_err_exists = true;
557 		sig->sigerr_count++;
558 
559 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
560 			     cq->mcq.cqn, sig->err_item.key,
561 			     sig->err_item.err_type,
562 			     sig->err_item.sig_err_offset,
563 			     sig->err_item.expected,
564 			     sig->err_item.actual);
565 
566 		xa_unlock(&dev->sig_mrs);
567 		goto repoll;
568 	}
569 	}
570 
571 	return 0;
572 }
573 
574 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
575 			struct ib_wc *wc, bool is_fatal_err)
576 {
577 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
578 	struct mlx5_ib_wc *soft_wc, *next;
579 	int npolled = 0;
580 
581 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
582 		if (npolled >= num_entries)
583 			break;
584 
585 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
586 			    cq->mcq.cqn);
587 
588 		if (unlikely(is_fatal_err)) {
589 			soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
590 			soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
591 		}
592 		wc[npolled++] = soft_wc->wc;
593 		list_del(&soft_wc->list);
594 		kfree(soft_wc);
595 	}
596 
597 	return npolled;
598 }
599 
600 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
601 {
602 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
603 	struct mlx5_ib_qp *cur_qp = NULL;
604 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
605 	struct mlx5_core_dev *mdev = dev->mdev;
606 	unsigned long flags;
607 	int soft_polled = 0;
608 	int npolled;
609 
610 	spin_lock_irqsave(&cq->lock, flags);
611 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
612 		/* make sure no soft wqe's are waiting */
613 		if (unlikely(!list_empty(&cq->wc_list)))
614 			soft_polled = poll_soft_wc(cq, num_entries, wc, true);
615 
616 		mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
617 				     wc + soft_polled, &npolled);
618 		goto out;
619 	}
620 
621 	if (unlikely(!list_empty(&cq->wc_list)))
622 		soft_polled = poll_soft_wc(cq, num_entries, wc, false);
623 
624 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
625 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
626 			break;
627 	}
628 
629 	if (npolled)
630 		mlx5_cq_set_ci(&cq->mcq);
631 out:
632 	spin_unlock_irqrestore(&cq->lock, flags);
633 
634 	return soft_polled + npolled;
635 }
636 
637 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
638 {
639 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
640 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
641 	void __iomem *uar_page = mdev->priv.uar->map;
642 	unsigned long irq_flags;
643 	int ret = 0;
644 
645 	spin_lock_irqsave(&cq->lock, irq_flags);
646 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
647 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
648 
649 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
650 		ret = 1;
651 	spin_unlock_irqrestore(&cq->lock, irq_flags);
652 
653 	mlx5_cq_arm(&cq->mcq,
654 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
655 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
656 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
657 
658 	return ret;
659 }
660 
661 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,
662 			     struct mlx5_ib_cq_buf *buf,
663 			     int nent,
664 			     int cqe_size)
665 {
666 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
667 	u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0);
668 	u8 log_wq_sz     = ilog2(cqe_size);
669 	int err;
670 
671 	err = mlx5_frag_buf_alloc_node(dev->mdev,
672 				       nent * cqe_size,
673 				       frag_buf,
674 				       dev->mdev->priv.numa_node);
675 	if (err)
676 		return err;
677 
678 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
679 
680 	buf->cqe_size = cqe_size;
681 	buf->nent = nent;
682 
683 	return 0;
684 }
685 
686 enum {
687 	MLX5_CQE_RES_FORMAT_HASH = 0,
688 	MLX5_CQE_RES_FORMAT_CSUM = 1,
689 	MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3,
690 };
691 
692 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
693 {
694 	switch (format) {
695 	case MLX5_IB_CQE_RES_FORMAT_HASH:
696 		return MLX5_CQE_RES_FORMAT_HASH;
697 	case MLX5_IB_CQE_RES_FORMAT_CSUM:
698 		return MLX5_CQE_RES_FORMAT_CSUM;
699 	case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX:
700 		if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
701 			return MLX5_CQE_RES_FORMAT_CSUM_STRIDX;
702 		return -EOPNOTSUPP;
703 	default:
704 		return -EINVAL;
705 	}
706 }
707 
708 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
709 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
710 			  int *cqe_size, int *index, int *inlen)
711 {
712 	struct mlx5_ib_create_cq ucmd = {};
713 	unsigned long page_size;
714 	unsigned int page_offset_quantized;
715 	size_t ucmdlen;
716 	__be64 *pas;
717 	int ncont;
718 	void *cqc;
719 	int err;
720 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
721 		udata, struct mlx5_ib_ucontext, ibucontext);
722 
723 	ucmdlen = min(udata->inlen, sizeof(ucmd));
724 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
725 		return -EINVAL;
726 
727 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
728 		return -EFAULT;
729 
730 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD |
731 			    MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX |
732 			    MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)))
733 		return -EINVAL;
734 
735 	if ((ucmd.cqe_size != 64 && ucmd.cqe_size != 128) ||
736 	    ucmd.reserved0 || ucmd.reserved1)
737 		return -EINVAL;
738 
739 	*cqe_size = ucmd.cqe_size;
740 
741 	cq->buf.umem =
742 		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
743 			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
744 	if (IS_ERR(cq->buf.umem)) {
745 		err = PTR_ERR(cq->buf.umem);
746 		return err;
747 	}
748 
749 	page_size = mlx5_umem_find_best_cq_quantized_pgoff(
750 		cq->buf.umem, cqc, log_page_size, MLX5_ADAPTER_PAGE_SHIFT,
751 		page_offset, 64, &page_offset_quantized);
752 	if (!page_size) {
753 		err = -EINVAL;
754 		goto err_umem;
755 	}
756 
757 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &cq->db);
758 	if (err)
759 		goto err_umem;
760 
761 	ncont = ib_umem_num_dma_blocks(cq->buf.umem, page_size);
762 	mlx5_ib_dbg(
763 		dev,
764 		"addr 0x%llx, size %u, npages %zu, page_size %lu, ncont %d\n",
765 		ucmd.buf_addr, entries * ucmd.cqe_size,
766 		ib_umem_num_pages(cq->buf.umem), page_size, ncont);
767 
768 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
769 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
770 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
771 	if (!*cqb) {
772 		err = -ENOMEM;
773 		goto err_db;
774 	}
775 
776 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
777 	mlx5_ib_populate_pas(cq->buf.umem, page_size, pas, 0);
778 
779 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
780 	MLX5_SET(cqc, cqc, log_page_size,
781 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
782 	MLX5_SET(cqc, cqc, page_offset, page_offset_quantized);
783 
784 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
785 		*index = ucmd.uar_page_index;
786 	} else if (context->bfregi.lib_uar_dyn) {
787 		err = -EINVAL;
788 		goto err_cqb;
789 	} else {
790 		*index = context->bfregi.sys_pages[0];
791 	}
792 
793 	if (ucmd.cqe_comp_en == 1) {
794 		int mini_cqe_format;
795 
796 		if (!((*cqe_size == 128 &&
797 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
798 		      (*cqe_size == 64  &&
799 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
800 			err = -EOPNOTSUPP;
801 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
802 				     *cqe_size);
803 			goto err_cqb;
804 		}
805 
806 		mini_cqe_format =
807 			mini_cqe_res_format_to_hw(dev,
808 						  ucmd.cqe_comp_res_format);
809 		if (mini_cqe_format < 0) {
810 			err = mini_cqe_format;
811 			mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n",
812 				    ucmd.cqe_comp_res_format, err);
813 			goto err_cqb;
814 		}
815 
816 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
817 		MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format);
818 	}
819 
820 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
821 		if (*cqe_size != 128 ||
822 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
823 			err = -EOPNOTSUPP;
824 			mlx5_ib_warn(dev,
825 				     "CQE padding is not supported for CQE size of %dB!\n",
826 				     *cqe_size);
827 			goto err_cqb;
828 		}
829 
830 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
831 	}
832 
833 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)
834 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS;
835 
836 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
837 	return 0;
838 
839 err_cqb:
840 	kvfree(*cqb);
841 
842 err_db:
843 	mlx5_ib_db_unmap_user(context, &cq->db);
844 
845 err_umem:
846 	ib_umem_release(cq->buf.umem);
847 	return err;
848 }
849 
850 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
851 {
852 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
853 		udata, struct mlx5_ib_ucontext, ibucontext);
854 
855 	mlx5_ib_db_unmap_user(context, &cq->db);
856 	ib_umem_release(cq->buf.umem);
857 }
858 
859 static void init_cq_frag_buf(struct mlx5_ib_cq_buf *buf)
860 {
861 	int i;
862 	void *cqe;
863 	struct mlx5_cqe64 *cqe64;
864 
865 	for (i = 0; i < buf->nent; i++) {
866 		cqe = mlx5_frag_buf_get_wqe(&buf->fbc, i);
867 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
868 		cqe64->op_own = MLX5_CQE_INVALID << 4;
869 	}
870 }
871 
872 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
873 			    int entries, int cqe_size,
874 			    u32 **cqb, int *index, int *inlen)
875 {
876 	__be64 *pas;
877 	void *cqc;
878 	int err;
879 
880 	err = mlx5_db_alloc(dev->mdev, &cq->db);
881 	if (err)
882 		return err;
883 
884 	cq->mcq.set_ci_db  = cq->db.db;
885 	cq->mcq.arm_db     = cq->db.db + 1;
886 	cq->mcq.cqe_sz = cqe_size;
887 
888 	err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size);
889 	if (err)
890 		goto err_db;
891 
892 	init_cq_frag_buf(&cq->buf);
893 
894 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
895 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) *
896 		 cq->buf.frag_buf.npages;
897 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
898 	if (!*cqb) {
899 		err = -ENOMEM;
900 		goto err_buf;
901 	}
902 
903 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
904 	mlx5_fill_page_frag_array(&cq->buf.frag_buf, pas);
905 
906 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
907 	MLX5_SET(cqc, cqc, log_page_size,
908 		 cq->buf.frag_buf.page_shift -
909 		 MLX5_ADAPTER_PAGE_SHIFT);
910 
911 	*index = dev->mdev->priv.uar->index;
912 
913 	return 0;
914 
915 err_buf:
916 	free_cq_buf(dev, &cq->buf);
917 
918 err_db:
919 	mlx5_db_free(dev->mdev, &cq->db);
920 	return err;
921 }
922 
923 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
924 {
925 	free_cq_buf(dev, &cq->buf);
926 	mlx5_db_free(dev->mdev, &cq->db);
927 }
928 
929 static void notify_soft_wc_handler(struct work_struct *work)
930 {
931 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
932 					     notify_work);
933 
934 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
935 }
936 
937 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
938 		      struct ib_udata *udata)
939 {
940 	struct ib_device *ibdev = ibcq->device;
941 	int entries = attr->cqe;
942 	int vector = attr->comp_vector;
943 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
944 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
945 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
946 	int index;
947 	int inlen;
948 	u32 *cqb = NULL;
949 	void *cqc;
950 	int cqe_size;
951 	int eqn;
952 	int err;
953 
954 	if (entries < 0 ||
955 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
956 		return -EINVAL;
957 
958 	if (check_cq_create_flags(attr->flags))
959 		return -EOPNOTSUPP;
960 
961 	entries = roundup_pow_of_two(entries + 1);
962 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
963 		return -EINVAL;
964 
965 	cq->ibcq.cqe = entries - 1;
966 	mutex_init(&cq->resize_mutex);
967 	spin_lock_init(&cq->lock);
968 	cq->resize_buf = NULL;
969 	cq->resize_umem = NULL;
970 	cq->create_flags = attr->flags;
971 	INIT_LIST_HEAD(&cq->list_send_qp);
972 	INIT_LIST_HEAD(&cq->list_recv_qp);
973 
974 	if (udata) {
975 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
976 				     &index, &inlen);
977 		if (err)
978 			return err;
979 	} else {
980 		cqe_size = cache_line_size() == 128 ? 128 : 64;
981 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
982 				       &index, &inlen);
983 		if (err)
984 			return err;
985 
986 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
987 	}
988 
989 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn);
990 	if (err)
991 		goto err_cqb;
992 
993 	cq->cqe_size = cqe_size;
994 
995 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
996 	MLX5_SET(cqc, cqc, cqe_sz,
997 		 cqe_sz_to_mlx_sz(cqe_size,
998 				  cq->private_flags &
999 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1000 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1001 	MLX5_SET(cqc, cqc, uar_page, index);
1002 	MLX5_SET(cqc, cqc, c_eqn_or_apu_element, eqn);
1003 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
1004 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
1005 		MLX5_SET(cqc, cqc, oi, 1);
1006 
1007 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
1008 	if (err)
1009 		goto err_cqb;
1010 
1011 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
1012 	if (udata)
1013 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
1014 	else
1015 		cq->mcq.comp  = mlx5_ib_cq_comp;
1016 	cq->mcq.event = mlx5_ib_cq_event;
1017 
1018 	INIT_LIST_HEAD(&cq->wc_list);
1019 
1020 	if (udata)
1021 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
1022 			err = -EFAULT;
1023 			goto err_cmd;
1024 		}
1025 
1026 
1027 	kvfree(cqb);
1028 	return 0;
1029 
1030 err_cmd:
1031 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1032 
1033 err_cqb:
1034 	kvfree(cqb);
1035 	if (udata)
1036 		destroy_cq_user(cq, udata);
1037 	else
1038 		destroy_cq_kernel(dev, cq);
1039 	return err;
1040 }
1041 
1042 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1043 {
1044 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1045 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1046 	int ret;
1047 
1048 	ret = mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1049 	if (ret)
1050 		return ret;
1051 
1052 	if (udata)
1053 		destroy_cq_user(mcq, udata);
1054 	else
1055 		destroy_cq_kernel(dev, mcq);
1056 	return 0;
1057 }
1058 
1059 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1060 {
1061 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1062 }
1063 
1064 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1065 {
1066 	struct mlx5_cqe64 *cqe64, *dest64;
1067 	void *cqe, *dest;
1068 	u32 prod_index;
1069 	int nfreed = 0;
1070 	u8 owner_bit;
1071 
1072 	if (!cq)
1073 		return;
1074 
1075 	/* First we need to find the current producer index, so we
1076 	 * know where to start cleaning from.  It doesn't matter if HW
1077 	 * adds new entries after this loop -- the QP we're worried
1078 	 * about is already in RESET, so the new entries won't come
1079 	 * from our QP and therefore don't need to be checked.
1080 	 */
1081 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1082 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1083 			break;
1084 
1085 	/* Now sweep backwards through the CQ, removing CQ entries
1086 	 * that match our QP by copying older entries on top of them.
1087 	 */
1088 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1089 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1090 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1091 		if (is_equal_rsn(cqe64, rsn)) {
1092 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1093 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1094 			++nfreed;
1095 		} else if (nfreed) {
1096 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1097 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1098 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1099 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1100 			dest64->op_own = owner_bit |
1101 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1102 		}
1103 	}
1104 
1105 	if (nfreed) {
1106 		cq->mcq.cons_index += nfreed;
1107 		/* Make sure update of buffer contents is done before
1108 		 * updating consumer index.
1109 		 */
1110 		wmb();
1111 		mlx5_cq_set_ci(&cq->mcq);
1112 	}
1113 }
1114 
1115 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1116 {
1117 	if (!cq)
1118 		return;
1119 
1120 	spin_lock_irq(&cq->lock);
1121 	__mlx5_ib_cq_clean(cq, qpn, srq);
1122 	spin_unlock_irq(&cq->lock);
1123 }
1124 
1125 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1126 {
1127 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1128 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1129 	int err;
1130 
1131 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1132 		return -EOPNOTSUPP;
1133 
1134 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1135 		return -EINVAL;
1136 
1137 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1138 					     cq_period, cq_count);
1139 	if (err)
1140 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1141 
1142 	return err;
1143 }
1144 
1145 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1146 		       int entries, struct ib_udata *udata,
1147 		       int *cqe_size)
1148 {
1149 	struct mlx5_ib_resize_cq ucmd;
1150 	struct ib_umem *umem;
1151 	int err;
1152 
1153 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1154 	if (err)
1155 		return err;
1156 
1157 	if (ucmd.reserved0 || ucmd.reserved1)
1158 		return -EINVAL;
1159 
1160 	/* check multiplication overflow */
1161 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1162 		return -EINVAL;
1163 
1164 	umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
1165 			   (size_t)ucmd.cqe_size * entries,
1166 			   IB_ACCESS_LOCAL_WRITE);
1167 	if (IS_ERR(umem)) {
1168 		err = PTR_ERR(umem);
1169 		return err;
1170 	}
1171 
1172 	cq->resize_umem = umem;
1173 	*cqe_size = ucmd.cqe_size;
1174 
1175 	return 0;
1176 }
1177 
1178 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1179 			 int entries, int cqe_size)
1180 {
1181 	int err;
1182 
1183 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1184 	if (!cq->resize_buf)
1185 		return -ENOMEM;
1186 
1187 	err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size);
1188 	if (err)
1189 		goto ex;
1190 
1191 	init_cq_frag_buf(cq->resize_buf);
1192 
1193 	return 0;
1194 
1195 ex:
1196 	kfree(cq->resize_buf);
1197 	return err;
1198 }
1199 
1200 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1201 {
1202 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1203 	struct mlx5_cqe64 *scqe64;
1204 	struct mlx5_cqe64 *dcqe64;
1205 	void *start_cqe;
1206 	void *scqe;
1207 	void *dcqe;
1208 	int ssize;
1209 	int dsize;
1210 	int i;
1211 	u8 sw_own;
1212 
1213 	ssize = cq->buf.cqe_size;
1214 	dsize = cq->resize_buf->cqe_size;
1215 	if (ssize != dsize) {
1216 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1217 		return -EINVAL;
1218 	}
1219 
1220 	i = cq->mcq.cons_index;
1221 	scqe = get_sw_cqe(cq, i);
1222 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1223 	start_cqe = scqe;
1224 	if (!scqe) {
1225 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1226 		return -EINVAL;
1227 	}
1228 
1229 	while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) {
1230 		dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc,
1231 					     (i + 1) & cq->resize_buf->nent);
1232 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1233 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1234 		memcpy(dcqe, scqe, dsize);
1235 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1236 
1237 		++i;
1238 		scqe = get_sw_cqe(cq, i);
1239 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1240 		if (!scqe) {
1241 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1242 			return -EINVAL;
1243 		}
1244 
1245 		if (scqe == start_cqe) {
1246 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1247 				cq->mcq.cqn);
1248 			return -ENOMEM;
1249 		}
1250 	}
1251 	++cq->mcq.cons_index;
1252 	return 0;
1253 }
1254 
1255 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1256 {
1257 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1258 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1259 	void *cqc;
1260 	u32 *in;
1261 	int err;
1262 	int npas;
1263 	__be64 *pas;
1264 	unsigned int page_offset_quantized = 0;
1265 	unsigned int page_shift;
1266 	int inlen;
1267 	int cqe_size;
1268 	unsigned long flags;
1269 
1270 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1271 		pr_info("Firmware does not support resize CQ\n");
1272 		return -ENOSYS;
1273 	}
1274 
1275 	if (entries < 1 ||
1276 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1277 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1278 			     entries,
1279 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1280 		return -EINVAL;
1281 	}
1282 
1283 	entries = roundup_pow_of_two(entries + 1);
1284 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1285 		return -EINVAL;
1286 
1287 	if (entries == ibcq->cqe + 1)
1288 		return 0;
1289 
1290 	mutex_lock(&cq->resize_mutex);
1291 	if (udata) {
1292 		unsigned long page_size;
1293 
1294 		err = resize_user(dev, cq, entries, udata, &cqe_size);
1295 		if (err)
1296 			goto ex;
1297 
1298 		page_size = mlx5_umem_find_best_cq_quantized_pgoff(
1299 			cq->resize_umem, cqc, log_page_size,
1300 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
1301 			&page_offset_quantized);
1302 		if (!page_size) {
1303 			err = -EINVAL;
1304 			goto ex_resize;
1305 		}
1306 		npas = ib_umem_num_dma_blocks(cq->resize_umem, page_size);
1307 		page_shift = order_base_2(page_size);
1308 	} else {
1309 		struct mlx5_frag_buf *frag_buf;
1310 
1311 		cqe_size = 64;
1312 		err = resize_kernel(dev, cq, entries, cqe_size);
1313 		if (err)
1314 			goto ex;
1315 		frag_buf = &cq->resize_buf->frag_buf;
1316 		npas = frag_buf->npages;
1317 		page_shift = frag_buf->page_shift;
1318 	}
1319 
1320 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1321 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1322 
1323 	in = kvzalloc(inlen, GFP_KERNEL);
1324 	if (!in) {
1325 		err = -ENOMEM;
1326 		goto ex_resize;
1327 	}
1328 
1329 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1330 	if (udata)
1331 		mlx5_ib_populate_pas(cq->resize_umem, 1UL << page_shift, pas,
1332 				     0);
1333 	else
1334 		mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas);
1335 
1336 	MLX5_SET(modify_cq_in, in,
1337 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1338 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1339 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1340 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1341 
1342 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1343 
1344 	MLX5_SET(cqc, cqc, log_page_size,
1345 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1346 	MLX5_SET(cqc, cqc, page_offset, page_offset_quantized);
1347 	MLX5_SET(cqc, cqc, cqe_sz,
1348 		 cqe_sz_to_mlx_sz(cqe_size,
1349 				  cq->private_flags &
1350 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1351 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1352 
1353 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1354 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1355 
1356 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1357 	if (err)
1358 		goto ex_alloc;
1359 
1360 	if (udata) {
1361 		cq->ibcq.cqe = entries - 1;
1362 		ib_umem_release(cq->buf.umem);
1363 		cq->buf.umem = cq->resize_umem;
1364 		cq->resize_umem = NULL;
1365 	} else {
1366 		struct mlx5_ib_cq_buf tbuf;
1367 		int resized = 0;
1368 
1369 		spin_lock_irqsave(&cq->lock, flags);
1370 		if (cq->resize_buf) {
1371 			err = copy_resize_cqes(cq);
1372 			if (!err) {
1373 				tbuf = cq->buf;
1374 				cq->buf = *cq->resize_buf;
1375 				kfree(cq->resize_buf);
1376 				cq->resize_buf = NULL;
1377 				resized = 1;
1378 			}
1379 		}
1380 		cq->ibcq.cqe = entries - 1;
1381 		spin_unlock_irqrestore(&cq->lock, flags);
1382 		if (resized)
1383 			free_cq_buf(dev, &tbuf);
1384 	}
1385 	mutex_unlock(&cq->resize_mutex);
1386 
1387 	kvfree(in);
1388 	return 0;
1389 
1390 ex_alloc:
1391 	kvfree(in);
1392 
1393 ex_resize:
1394 	ib_umem_release(cq->resize_umem);
1395 	if (!udata) {
1396 		free_cq_buf(dev, cq->resize_buf);
1397 		cq->resize_buf = NULL;
1398 	}
1399 ex:
1400 	mutex_unlock(&cq->resize_mutex);
1401 	return err;
1402 }
1403 
1404 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq)
1405 {
1406 	struct mlx5_ib_cq *cq;
1407 
1408 	if (!ibcq)
1409 		return 128;
1410 
1411 	cq = to_mcq(ibcq);
1412 	return cq->cqe_size;
1413 }
1414 
1415 /* Called from atomic context */
1416 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1417 {
1418 	struct mlx5_ib_wc *soft_wc;
1419 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1420 	unsigned long flags;
1421 
1422 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1423 	if (!soft_wc)
1424 		return -ENOMEM;
1425 
1426 	soft_wc->wc = *wc;
1427 	spin_lock_irqsave(&cq->lock, flags);
1428 	list_add_tail(&soft_wc->list, &cq->wc_list);
1429 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1430 	    wc->status != IB_WC_SUCCESS) {
1431 		cq->notify_flags = 0;
1432 		schedule_work(&cq->notify_work);
1433 	}
1434 	spin_unlock_irqrestore(&cq->lock, flags);
1435 
1436 	return 0;
1437 }
1438