xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision b8265621)
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 		/* fall through */
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 		/* fall through */
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_cmd_get_srq(dev, be32_to_cpu(cqe->srqn));
183 			srq = to_mibsrq(msrq);
184 		} else {
185 			srq = to_msrq(qp->ibqp.srq);
186 		}
187 		if (srq) {
188 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
189 			wc->wr_id = srq->wrid[wqe_ctr];
190 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
191 			if (msrq)
192 				mlx5_core_res_put(&msrq->common);
193 		}
194 	} else {
195 		wq	  = &qp->rq;
196 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
197 		++wq->tail;
198 	}
199 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
200 
201 	switch (get_cqe_opcode(cqe)) {
202 	case MLX5_CQE_RESP_WR_IMM:
203 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
204 		wc->wc_flags	= IB_WC_WITH_IMM;
205 		wc->ex.imm_data = cqe->immediate;
206 		break;
207 	case MLX5_CQE_RESP_SEND:
208 		wc->opcode   = IB_WC_RECV;
209 		wc->wc_flags = IB_WC_IP_CSUM_OK;
210 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
211 			       (cqe->hds_ip_ext & CQE_L4_OK))))
212 			wc->wc_flags = 0;
213 		break;
214 	case MLX5_CQE_RESP_SEND_IMM:
215 		wc->opcode	= IB_WC_RECV;
216 		wc->wc_flags	= IB_WC_WITH_IMM;
217 		wc->ex.imm_data = cqe->immediate;
218 		break;
219 	case MLX5_CQE_RESP_SEND_INV:
220 		wc->opcode	= IB_WC_RECV;
221 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
222 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->inval_rkey);
223 		break;
224 	}
225 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
226 	wc->dlid_path_bits = cqe->ml_path;
227 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
228 	wc->wc_flags |= g ? IB_WC_GRH : 0;
229 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
230 		u16 pkey = be32_to_cpu(cqe->pkey) & 0xffff;
231 
232 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
233 				    &wc->pkey_index);
234 	} else {
235 		wc->pkey_index = 0;
236 	}
237 
238 	if (ll != IB_LINK_LAYER_ETHERNET) {
239 		wc->slid = be16_to_cpu(cqe->slid);
240 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
241 		return;
242 	}
243 
244 	wc->slid = 0;
245 	vlan_present = cqe->l4_l3_hdr_type & 0x1;
246 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
247 	if (vlan_present) {
248 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
249 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
250 		wc->wc_flags |= IB_WC_WITH_VLAN;
251 	} else {
252 		wc->sl = 0;
253 	}
254 
255 	switch (roce_packet_type) {
256 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
257 		wc->network_hdr_type = RDMA_NETWORK_IB;
258 		break;
259 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
260 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
261 		break;
262 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
263 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
264 		break;
265 	}
266 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
267 }
268 
269 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
270 {
271 	mlx5_ib_warn(dev, "dump error cqe\n");
272 	mlx5_dump_err_cqe(dev->mdev, cqe);
273 }
274 
275 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
276 				  struct mlx5_err_cqe *cqe,
277 				  struct ib_wc *wc)
278 {
279 	int dump = 1;
280 
281 	switch (cqe->syndrome) {
282 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
283 		wc->status = IB_WC_LOC_LEN_ERR;
284 		break;
285 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
286 		wc->status = IB_WC_LOC_QP_OP_ERR;
287 		break;
288 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
289 		wc->status = IB_WC_LOC_PROT_ERR;
290 		break;
291 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
292 		dump = 0;
293 		wc->status = IB_WC_WR_FLUSH_ERR;
294 		break;
295 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
296 		wc->status = IB_WC_MW_BIND_ERR;
297 		break;
298 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
299 		wc->status = IB_WC_BAD_RESP_ERR;
300 		break;
301 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
302 		wc->status = IB_WC_LOC_ACCESS_ERR;
303 		break;
304 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
305 		wc->status = IB_WC_REM_INV_REQ_ERR;
306 		break;
307 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
308 		wc->status = IB_WC_REM_ACCESS_ERR;
309 		break;
310 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
311 		wc->status = IB_WC_REM_OP_ERR;
312 		break;
313 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
314 		wc->status = IB_WC_RETRY_EXC_ERR;
315 		dump = 0;
316 		break;
317 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
318 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
319 		dump = 0;
320 		break;
321 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
322 		wc->status = IB_WC_REM_ABORT_ERR;
323 		break;
324 	default:
325 		wc->status = IB_WC_GENERAL_ERR;
326 		break;
327 	}
328 
329 	wc->vendor_err = cqe->vendor_err_synd;
330 	if (dump)
331 		dump_cqe(dev, cqe);
332 }
333 
334 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
335 			   u16 tail, u16 head)
336 {
337 	u16 idx;
338 
339 	do {
340 		idx = tail & (qp->sq.wqe_cnt - 1);
341 		if (idx == head)
342 			break;
343 
344 		tail = qp->sq.w_list[idx].next;
345 	} while (1);
346 	tail = qp->sq.w_list[idx].next;
347 	qp->sq.last_poll = tail;
348 }
349 
350 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
351 {
352 	mlx5_frag_buf_free(dev->mdev, &buf->frag_buf);
353 }
354 
355 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
356 			     struct ib_sig_err *item)
357 {
358 	u16 syndrome = be16_to_cpu(cqe->syndrome);
359 
360 #define GUARD_ERR   (1 << 13)
361 #define APPTAG_ERR  (1 << 12)
362 #define REFTAG_ERR  (1 << 11)
363 
364 	if (syndrome & GUARD_ERR) {
365 		item->err_type = IB_SIG_BAD_GUARD;
366 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
367 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
368 	} else
369 	if (syndrome & REFTAG_ERR) {
370 		item->err_type = IB_SIG_BAD_REFTAG;
371 		item->expected = be32_to_cpu(cqe->expected_reftag);
372 		item->actual = be32_to_cpu(cqe->actual_reftag);
373 	} else
374 	if (syndrome & APPTAG_ERR) {
375 		item->err_type = IB_SIG_BAD_APPTAG;
376 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
377 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
378 	} else {
379 		pr_err("Got signature completion error with bad syndrome %04x\n",
380 		       syndrome);
381 	}
382 
383 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
384 	item->key = be32_to_cpu(cqe->mkey);
385 }
386 
387 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
388 		    int *npolled, bool is_send)
389 {
390 	struct mlx5_ib_wq *wq;
391 	unsigned int cur;
392 	int np;
393 	int i;
394 
395 	wq = (is_send) ? &qp->sq : &qp->rq;
396 	cur = wq->head - wq->tail;
397 	np = *npolled;
398 
399 	if (cur == 0)
400 		return;
401 
402 	for (i = 0;  i < cur && np < num_entries; i++) {
403 		unsigned int idx;
404 
405 		idx = (is_send) ? wq->last_poll : wq->tail;
406 		idx &= (wq->wqe_cnt - 1);
407 		wc->wr_id = wq->wrid[idx];
408 		wc->status = IB_WC_WR_FLUSH_ERR;
409 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
410 		wq->tail++;
411 		if (is_send)
412 			wq->last_poll = wq->w_list[idx].next;
413 		np++;
414 		wc->qp = &qp->ibqp;
415 		wc++;
416 	}
417 	*npolled = np;
418 }
419 
420 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
421 				 struct ib_wc *wc, int *npolled)
422 {
423 	struct mlx5_ib_qp *qp;
424 
425 	*npolled = 0;
426 	/* Find uncompleted WQEs belonging to that cq and return mmics ones */
427 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
428 		sw_comp(qp, num_entries, wc + *npolled, npolled, true);
429 		if (*npolled >= num_entries)
430 			return;
431 	}
432 
433 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
434 		sw_comp(qp, num_entries, wc + *npolled, npolled, false);
435 		if (*npolled >= num_entries)
436 			return;
437 	}
438 }
439 
440 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
441 			 struct mlx5_ib_qp **cur_qp,
442 			 struct ib_wc *wc)
443 {
444 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
445 	struct mlx5_err_cqe *err_cqe;
446 	struct mlx5_cqe64 *cqe64;
447 	struct mlx5_core_qp *mqp;
448 	struct mlx5_ib_wq *wq;
449 	uint8_t opcode;
450 	uint32_t qpn;
451 	u16 wqe_ctr;
452 	void *cqe;
453 	int idx;
454 
455 repoll:
456 	cqe = next_cqe_sw(cq);
457 	if (!cqe)
458 		return -EAGAIN;
459 
460 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
461 
462 	++cq->mcq.cons_index;
463 
464 	/* Make sure we read CQ entry contents after we've checked the
465 	 * ownership bit.
466 	 */
467 	rmb();
468 
469 	opcode = get_cqe_opcode(cqe64);
470 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
471 		if (likely(cq->resize_buf)) {
472 			free_cq_buf(dev, &cq->buf);
473 			cq->buf = *cq->resize_buf;
474 			kfree(cq->resize_buf);
475 			cq->resize_buf = NULL;
476 			goto repoll;
477 		} else {
478 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
479 		}
480 	}
481 
482 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
483 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
484 		/* We do not have to take the QP table lock here,
485 		 * because CQs will be locked while QPs are removed
486 		 * from the table.
487 		 */
488 		mqp = radix_tree_lookup(&dev->qp_table.tree, qpn);
489 		*cur_qp = to_mibqp(mqp);
490 	}
491 
492 	wc->qp  = &(*cur_qp)->ibqp;
493 	switch (opcode) {
494 	case MLX5_CQE_REQ:
495 		wq = &(*cur_qp)->sq;
496 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
497 		idx = wqe_ctr & (wq->wqe_cnt - 1);
498 		handle_good_req(wc, cqe64, wq, idx);
499 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
500 		wc->wr_id = wq->wrid[idx];
501 		wq->tail = wq->wqe_head[idx] + 1;
502 		wc->status = IB_WC_SUCCESS;
503 		break;
504 	case MLX5_CQE_RESP_WR_IMM:
505 	case MLX5_CQE_RESP_SEND:
506 	case MLX5_CQE_RESP_SEND_IMM:
507 	case MLX5_CQE_RESP_SEND_INV:
508 		handle_responder(wc, cqe64, *cur_qp);
509 		wc->status = IB_WC_SUCCESS;
510 		break;
511 	case MLX5_CQE_RESIZE_CQ:
512 		break;
513 	case MLX5_CQE_REQ_ERR:
514 	case MLX5_CQE_RESP_ERR:
515 		err_cqe = (struct mlx5_err_cqe *)cqe64;
516 		mlx5_handle_error_cqe(dev, err_cqe, wc);
517 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
518 			    opcode == MLX5_CQE_REQ_ERR ?
519 			    "Requestor" : "Responder", cq->mcq.cqn);
520 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
521 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
522 		if (opcode == MLX5_CQE_REQ_ERR) {
523 			wq = &(*cur_qp)->sq;
524 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
525 			idx = wqe_ctr & (wq->wqe_cnt - 1);
526 			wc->wr_id = wq->wrid[idx];
527 			wq->tail = wq->wqe_head[idx] + 1;
528 		} else {
529 			struct mlx5_ib_srq *srq;
530 
531 			if ((*cur_qp)->ibqp.srq) {
532 				srq = to_msrq((*cur_qp)->ibqp.srq);
533 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
534 				wc->wr_id = srq->wrid[wqe_ctr];
535 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
536 			} else {
537 				wq = &(*cur_qp)->rq;
538 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
539 				++wq->tail;
540 			}
541 		}
542 		break;
543 	case MLX5_CQE_SIG_ERR: {
544 		struct mlx5_sig_err_cqe *sig_err_cqe =
545 			(struct mlx5_sig_err_cqe *)cqe64;
546 		struct mlx5_core_sig_ctx *sig;
547 
548 		xa_lock(&dev->sig_mrs);
549 		sig = xa_load(&dev->sig_mrs,
550 				mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
551 		get_sig_err_item(sig_err_cqe, &sig->err_item);
552 		sig->sig_err_exists = true;
553 		sig->sigerr_count++;
554 
555 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
556 			     cq->mcq.cqn, sig->err_item.key,
557 			     sig->err_item.err_type,
558 			     sig->err_item.sig_err_offset,
559 			     sig->err_item.expected,
560 			     sig->err_item.actual);
561 
562 		xa_unlock(&dev->sig_mrs);
563 		goto repoll;
564 	}
565 	}
566 
567 	return 0;
568 }
569 
570 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
571 			struct ib_wc *wc, bool is_fatal_err)
572 {
573 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
574 	struct mlx5_ib_wc *soft_wc, *next;
575 	int npolled = 0;
576 
577 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
578 		if (npolled >= num_entries)
579 			break;
580 
581 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
582 			    cq->mcq.cqn);
583 
584 		if (unlikely(is_fatal_err)) {
585 			soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
586 			soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
587 		}
588 		wc[npolled++] = soft_wc->wc;
589 		list_del(&soft_wc->list);
590 		kfree(soft_wc);
591 	}
592 
593 	return npolled;
594 }
595 
596 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
597 {
598 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
599 	struct mlx5_ib_qp *cur_qp = NULL;
600 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
601 	struct mlx5_core_dev *mdev = dev->mdev;
602 	unsigned long flags;
603 	int soft_polled = 0;
604 	int npolled;
605 
606 	spin_lock_irqsave(&cq->lock, flags);
607 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
608 		/* make sure no soft wqe's are waiting */
609 		if (unlikely(!list_empty(&cq->wc_list)))
610 			soft_polled = poll_soft_wc(cq, num_entries, wc, true);
611 
612 		mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
613 				     wc + soft_polled, &npolled);
614 		goto out;
615 	}
616 
617 	if (unlikely(!list_empty(&cq->wc_list)))
618 		soft_polled = poll_soft_wc(cq, num_entries, wc, false);
619 
620 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
621 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
622 			break;
623 	}
624 
625 	if (npolled)
626 		mlx5_cq_set_ci(&cq->mcq);
627 out:
628 	spin_unlock_irqrestore(&cq->lock, flags);
629 
630 	return soft_polled + npolled;
631 }
632 
633 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
634 {
635 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
636 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
637 	void __iomem *uar_page = mdev->priv.uar->map;
638 	unsigned long irq_flags;
639 	int ret = 0;
640 
641 	spin_lock_irqsave(&cq->lock, irq_flags);
642 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
643 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
644 
645 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
646 		ret = 1;
647 	spin_unlock_irqrestore(&cq->lock, irq_flags);
648 
649 	mlx5_cq_arm(&cq->mcq,
650 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
651 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
652 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
653 
654 	return ret;
655 }
656 
657 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,
658 			     struct mlx5_ib_cq_buf *buf,
659 			     int nent,
660 			     int cqe_size)
661 {
662 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
663 	u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0);
664 	u8 log_wq_sz     = ilog2(cqe_size);
665 	int err;
666 
667 	err = mlx5_frag_buf_alloc_node(dev->mdev,
668 				       nent * cqe_size,
669 				       frag_buf,
670 				       dev->mdev->priv.numa_node);
671 	if (err)
672 		return err;
673 
674 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
675 
676 	buf->cqe_size = cqe_size;
677 	buf->nent = nent;
678 
679 	return 0;
680 }
681 
682 enum {
683 	MLX5_CQE_RES_FORMAT_HASH = 0,
684 	MLX5_CQE_RES_FORMAT_CSUM = 1,
685 	MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3,
686 };
687 
688 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
689 {
690 	switch (format) {
691 	case MLX5_IB_CQE_RES_FORMAT_HASH:
692 		return MLX5_CQE_RES_FORMAT_HASH;
693 	case MLX5_IB_CQE_RES_FORMAT_CSUM:
694 		return MLX5_CQE_RES_FORMAT_CSUM;
695 	case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX:
696 		if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
697 			return MLX5_CQE_RES_FORMAT_CSUM_STRIDX;
698 		return -EOPNOTSUPP;
699 	default:
700 		return -EINVAL;
701 	}
702 }
703 
704 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
705 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
706 			  int *cqe_size, int *index, int *inlen)
707 {
708 	struct mlx5_ib_create_cq ucmd = {};
709 	size_t ucmdlen;
710 	int page_shift;
711 	__be64 *pas;
712 	int npages;
713 	int ncont;
714 	void *cqc;
715 	int err;
716 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
717 		udata, struct mlx5_ib_ucontext, ibucontext);
718 
719 	ucmdlen = min(udata->inlen, sizeof(ucmd));
720 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
721 		return -EINVAL;
722 
723 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
724 		return -EFAULT;
725 
726 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD |
727 			    MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX)))
728 		return -EINVAL;
729 
730 	if ((ucmd.cqe_size != 64 && ucmd.cqe_size != 128) ||
731 	    ucmd.reserved0 || ucmd.reserved1)
732 		return -EINVAL;
733 
734 	*cqe_size = ucmd.cqe_size;
735 
736 	cq->buf.umem =
737 		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
738 			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
739 	if (IS_ERR(cq->buf.umem)) {
740 		err = PTR_ERR(cq->buf.umem);
741 		return err;
742 	}
743 
744 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &cq->db);
745 	if (err)
746 		goto err_umem;
747 
748 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
749 			   &ncont, NULL);
750 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
751 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
752 
753 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
754 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
755 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
756 	if (!*cqb) {
757 		err = -ENOMEM;
758 		goto err_db;
759 	}
760 
761 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
762 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
763 
764 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
765 	MLX5_SET(cqc, cqc, log_page_size,
766 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
767 
768 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
769 		*index = ucmd.uar_page_index;
770 	} else if (context->bfregi.lib_uar_dyn) {
771 		err = -EINVAL;
772 		goto err_cqb;
773 	} else {
774 		*index = context->bfregi.sys_pages[0];
775 	}
776 
777 	if (ucmd.cqe_comp_en == 1) {
778 		int mini_cqe_format;
779 
780 		if (!((*cqe_size == 128 &&
781 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
782 		      (*cqe_size == 64  &&
783 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
784 			err = -EOPNOTSUPP;
785 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
786 				     *cqe_size);
787 			goto err_cqb;
788 		}
789 
790 		mini_cqe_format =
791 			mini_cqe_res_format_to_hw(dev,
792 						  ucmd.cqe_comp_res_format);
793 		if (mini_cqe_format < 0) {
794 			err = mini_cqe_format;
795 			mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n",
796 				    ucmd.cqe_comp_res_format, err);
797 			goto err_cqb;
798 		}
799 
800 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
801 		MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format);
802 	}
803 
804 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
805 		if (*cqe_size != 128 ||
806 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
807 			err = -EOPNOTSUPP;
808 			mlx5_ib_warn(dev,
809 				     "CQE padding is not supported for CQE size of %dB!\n",
810 				     *cqe_size);
811 			goto err_cqb;
812 		}
813 
814 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
815 	}
816 
817 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
818 	return 0;
819 
820 err_cqb:
821 	kvfree(*cqb);
822 
823 err_db:
824 	mlx5_ib_db_unmap_user(context, &cq->db);
825 
826 err_umem:
827 	ib_umem_release(cq->buf.umem);
828 	return err;
829 }
830 
831 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
832 {
833 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
834 		udata, struct mlx5_ib_ucontext, ibucontext);
835 
836 	mlx5_ib_db_unmap_user(context, &cq->db);
837 	ib_umem_release(cq->buf.umem);
838 }
839 
840 static void init_cq_frag_buf(struct mlx5_ib_cq *cq,
841 			     struct mlx5_ib_cq_buf *buf)
842 {
843 	int i;
844 	void *cqe;
845 	struct mlx5_cqe64 *cqe64;
846 
847 	for (i = 0; i < buf->nent; i++) {
848 		cqe = get_cqe(cq, i);
849 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
850 		cqe64->op_own = MLX5_CQE_INVALID << 4;
851 	}
852 }
853 
854 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
855 			    int entries, int cqe_size,
856 			    u32 **cqb, int *index, int *inlen)
857 {
858 	__be64 *pas;
859 	void *cqc;
860 	int err;
861 
862 	err = mlx5_db_alloc(dev->mdev, &cq->db);
863 	if (err)
864 		return err;
865 
866 	cq->mcq.set_ci_db  = cq->db.db;
867 	cq->mcq.arm_db     = cq->db.db + 1;
868 	cq->mcq.cqe_sz = cqe_size;
869 
870 	err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size);
871 	if (err)
872 		goto err_db;
873 
874 	init_cq_frag_buf(cq, &cq->buf);
875 
876 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
877 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) *
878 		 cq->buf.frag_buf.npages;
879 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
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_frag_array(&cq->buf.frag_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.frag_buf.page_shift -
891 		 MLX5_ADAPTER_PAGE_SHIFT);
892 
893 	*index = dev->mdev->priv.uar->index;
894 
895 	return 0;
896 
897 err_buf:
898 	free_cq_buf(dev, &cq->buf);
899 
900 err_db:
901 	mlx5_db_free(dev->mdev, &cq->db);
902 	return err;
903 }
904 
905 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
906 {
907 	free_cq_buf(dev, &cq->buf);
908 	mlx5_db_free(dev->mdev, &cq->db);
909 }
910 
911 static void notify_soft_wc_handler(struct work_struct *work)
912 {
913 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
914 					     notify_work);
915 
916 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
917 }
918 
919 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
920 		      struct ib_udata *udata)
921 {
922 	struct ib_device *ibdev = ibcq->device;
923 	int entries = attr->cqe;
924 	int vector = attr->comp_vector;
925 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
926 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
927 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
928 	int uninitialized_var(index);
929 	int uninitialized_var(inlen);
930 	u32 *cqb = NULL;
931 	void *cqc;
932 	int cqe_size;
933 	unsigned int irqn;
934 	int eqn;
935 	int err;
936 
937 	if (entries < 0 ||
938 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
939 		return -EINVAL;
940 
941 	if (check_cq_create_flags(attr->flags))
942 		return -EOPNOTSUPP;
943 
944 	entries = roundup_pow_of_two(entries + 1);
945 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
946 		return -EINVAL;
947 
948 	cq->ibcq.cqe = entries - 1;
949 	mutex_init(&cq->resize_mutex);
950 	spin_lock_init(&cq->lock);
951 	cq->resize_buf = NULL;
952 	cq->resize_umem = NULL;
953 	cq->create_flags = attr->flags;
954 	INIT_LIST_HEAD(&cq->list_send_qp);
955 	INIT_LIST_HEAD(&cq->list_recv_qp);
956 
957 	if (udata) {
958 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
959 				     &index, &inlen);
960 		if (err)
961 			return err;
962 	} else {
963 		cqe_size = cache_line_size() == 128 ? 128 : 64;
964 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
965 				       &index, &inlen);
966 		if (err)
967 			return err;
968 
969 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
970 	}
971 
972 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
973 	if (err)
974 		goto err_cqb;
975 
976 	cq->cqe_size = cqe_size;
977 
978 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
979 	MLX5_SET(cqc, cqc, cqe_sz,
980 		 cqe_sz_to_mlx_sz(cqe_size,
981 				  cq->private_flags &
982 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
983 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
984 	MLX5_SET(cqc, cqc, uar_page, index);
985 	MLX5_SET(cqc, cqc, c_eqn, eqn);
986 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
987 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
988 		MLX5_SET(cqc, cqc, oi, 1);
989 
990 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
991 	if (err)
992 		goto err_cqb;
993 
994 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
995 	cq->mcq.irqn = irqn;
996 	if (udata)
997 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
998 	else
999 		cq->mcq.comp  = mlx5_ib_cq_comp;
1000 	cq->mcq.event = mlx5_ib_cq_event;
1001 
1002 	INIT_LIST_HEAD(&cq->wc_list);
1003 
1004 	if (udata)
1005 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
1006 			err = -EFAULT;
1007 			goto err_cmd;
1008 		}
1009 
1010 
1011 	kvfree(cqb);
1012 	return 0;
1013 
1014 err_cmd:
1015 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1016 
1017 err_cqb:
1018 	kvfree(cqb);
1019 	if (udata)
1020 		destroy_cq_user(cq, udata);
1021 	else
1022 		destroy_cq_kernel(dev, cq);
1023 	return err;
1024 }
1025 
1026 void mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1027 {
1028 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1029 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1030 
1031 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1032 	if (udata)
1033 		destroy_cq_user(mcq, udata);
1034 	else
1035 		destroy_cq_kernel(dev, mcq);
1036 }
1037 
1038 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1039 {
1040 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1041 }
1042 
1043 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1044 {
1045 	struct mlx5_cqe64 *cqe64, *dest64;
1046 	void *cqe, *dest;
1047 	u32 prod_index;
1048 	int nfreed = 0;
1049 	u8 owner_bit;
1050 
1051 	if (!cq)
1052 		return;
1053 
1054 	/* First we need to find the current producer index, so we
1055 	 * know where to start cleaning from.  It doesn't matter if HW
1056 	 * adds new entries after this loop -- the QP we're worried
1057 	 * about is already in RESET, so the new entries won't come
1058 	 * from our QP and therefore don't need to be checked.
1059 	 */
1060 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1061 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1062 			break;
1063 
1064 	/* Now sweep backwards through the CQ, removing CQ entries
1065 	 * that match our QP by copying older entries on top of them.
1066 	 */
1067 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1068 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1069 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1070 		if (is_equal_rsn(cqe64, rsn)) {
1071 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1072 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1073 			++nfreed;
1074 		} else if (nfreed) {
1075 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1076 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1077 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1078 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1079 			dest64->op_own = owner_bit |
1080 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1081 		}
1082 	}
1083 
1084 	if (nfreed) {
1085 		cq->mcq.cons_index += nfreed;
1086 		/* Make sure update of buffer contents is done before
1087 		 * updating consumer index.
1088 		 */
1089 		wmb();
1090 		mlx5_cq_set_ci(&cq->mcq);
1091 	}
1092 }
1093 
1094 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1095 {
1096 	if (!cq)
1097 		return;
1098 
1099 	spin_lock_irq(&cq->lock);
1100 	__mlx5_ib_cq_clean(cq, qpn, srq);
1101 	spin_unlock_irq(&cq->lock);
1102 }
1103 
1104 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1105 {
1106 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1107 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1108 	int err;
1109 
1110 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1111 		return -EOPNOTSUPP;
1112 
1113 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1114 		return -EINVAL;
1115 
1116 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1117 					     cq_period, cq_count);
1118 	if (err)
1119 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1120 
1121 	return err;
1122 }
1123 
1124 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1125 		       int entries, struct ib_udata *udata, int *npas,
1126 		       int *page_shift, int *cqe_size)
1127 {
1128 	struct mlx5_ib_resize_cq ucmd;
1129 	struct ib_umem *umem;
1130 	int err;
1131 	int npages;
1132 
1133 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1134 	if (err)
1135 		return err;
1136 
1137 	if (ucmd.reserved0 || ucmd.reserved1)
1138 		return -EINVAL;
1139 
1140 	/* check multiplication overflow */
1141 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1142 		return -EINVAL;
1143 
1144 	umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
1145 			   (size_t)ucmd.cqe_size * entries,
1146 			   IB_ACCESS_LOCAL_WRITE);
1147 	if (IS_ERR(umem)) {
1148 		err = PTR_ERR(umem);
1149 		return err;
1150 	}
1151 
1152 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1153 			   npas, NULL);
1154 
1155 	cq->resize_umem = umem;
1156 	*cqe_size = ucmd.cqe_size;
1157 
1158 	return 0;
1159 }
1160 
1161 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1162 			 int entries, int cqe_size)
1163 {
1164 	int err;
1165 
1166 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1167 	if (!cq->resize_buf)
1168 		return -ENOMEM;
1169 
1170 	err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size);
1171 	if (err)
1172 		goto ex;
1173 
1174 	init_cq_frag_buf(cq, cq->resize_buf);
1175 
1176 	return 0;
1177 
1178 ex:
1179 	kfree(cq->resize_buf);
1180 	return err;
1181 }
1182 
1183 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1184 {
1185 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1186 	struct mlx5_cqe64 *scqe64;
1187 	struct mlx5_cqe64 *dcqe64;
1188 	void *start_cqe;
1189 	void *scqe;
1190 	void *dcqe;
1191 	int ssize;
1192 	int dsize;
1193 	int i;
1194 	u8 sw_own;
1195 
1196 	ssize = cq->buf.cqe_size;
1197 	dsize = cq->resize_buf->cqe_size;
1198 	if (ssize != dsize) {
1199 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1200 		return -EINVAL;
1201 	}
1202 
1203 	i = cq->mcq.cons_index;
1204 	scqe = get_sw_cqe(cq, i);
1205 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1206 	start_cqe = scqe;
1207 	if (!scqe) {
1208 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1209 		return -EINVAL;
1210 	}
1211 
1212 	while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) {
1213 		dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc,
1214 					     (i + 1) & cq->resize_buf->nent);
1215 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1216 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1217 		memcpy(dcqe, scqe, dsize);
1218 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1219 
1220 		++i;
1221 		scqe = get_sw_cqe(cq, i);
1222 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1223 		if (!scqe) {
1224 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1225 			return -EINVAL;
1226 		}
1227 
1228 		if (scqe == start_cqe) {
1229 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1230 				cq->mcq.cqn);
1231 			return -ENOMEM;
1232 		}
1233 	}
1234 	++cq->mcq.cons_index;
1235 	return 0;
1236 }
1237 
1238 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1239 {
1240 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1241 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1242 	void *cqc;
1243 	u32 *in;
1244 	int err;
1245 	int npas;
1246 	__be64 *pas;
1247 	int page_shift;
1248 	int inlen;
1249 	int uninitialized_var(cqe_size);
1250 	unsigned long flags;
1251 
1252 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1253 		pr_info("Firmware does not support resize CQ\n");
1254 		return -ENOSYS;
1255 	}
1256 
1257 	if (entries < 1 ||
1258 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1259 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1260 			     entries,
1261 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1262 		return -EINVAL;
1263 	}
1264 
1265 	entries = roundup_pow_of_two(entries + 1);
1266 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1267 		return -EINVAL;
1268 
1269 	if (entries == ibcq->cqe + 1)
1270 		return 0;
1271 
1272 	mutex_lock(&cq->resize_mutex);
1273 	if (udata) {
1274 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1275 				  &cqe_size);
1276 	} else {
1277 		cqe_size = 64;
1278 		err = resize_kernel(dev, cq, entries, cqe_size);
1279 		if (!err) {
1280 			struct mlx5_frag_buf *frag_buf = &cq->resize_buf->frag_buf;
1281 
1282 			npas = frag_buf->npages;
1283 			page_shift = frag_buf->page_shift;
1284 		}
1285 	}
1286 
1287 	if (err)
1288 		goto ex;
1289 
1290 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1291 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1292 
1293 	in = kvzalloc(inlen, GFP_KERNEL);
1294 	if (!in) {
1295 		err = -ENOMEM;
1296 		goto ex_resize;
1297 	}
1298 
1299 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1300 	if (udata)
1301 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1302 				     pas, 0);
1303 	else
1304 		mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas);
1305 
1306 	MLX5_SET(modify_cq_in, in,
1307 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1308 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1309 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1310 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1311 
1312 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1313 
1314 	MLX5_SET(cqc, cqc, log_page_size,
1315 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1316 	MLX5_SET(cqc, cqc, cqe_sz,
1317 		 cqe_sz_to_mlx_sz(cqe_size,
1318 				  cq->private_flags &
1319 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1320 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1321 
1322 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1323 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1324 
1325 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1326 	if (err)
1327 		goto ex_alloc;
1328 
1329 	if (udata) {
1330 		cq->ibcq.cqe = entries - 1;
1331 		ib_umem_release(cq->buf.umem);
1332 		cq->buf.umem = cq->resize_umem;
1333 		cq->resize_umem = NULL;
1334 	} else {
1335 		struct mlx5_ib_cq_buf tbuf;
1336 		int resized = 0;
1337 
1338 		spin_lock_irqsave(&cq->lock, flags);
1339 		if (cq->resize_buf) {
1340 			err = copy_resize_cqes(cq);
1341 			if (!err) {
1342 				tbuf = cq->buf;
1343 				cq->buf = *cq->resize_buf;
1344 				kfree(cq->resize_buf);
1345 				cq->resize_buf = NULL;
1346 				resized = 1;
1347 			}
1348 		}
1349 		cq->ibcq.cqe = entries - 1;
1350 		spin_unlock_irqrestore(&cq->lock, flags);
1351 		if (resized)
1352 			free_cq_buf(dev, &tbuf);
1353 	}
1354 	mutex_unlock(&cq->resize_mutex);
1355 
1356 	kvfree(in);
1357 	return 0;
1358 
1359 ex_alloc:
1360 	kvfree(in);
1361 
1362 ex_resize:
1363 	ib_umem_release(cq->resize_umem);
1364 	if (!udata) {
1365 		free_cq_buf(dev, cq->resize_buf);
1366 		cq->resize_buf = NULL;
1367 	}
1368 ex:
1369 	mutex_unlock(&cq->resize_mutex);
1370 	return err;
1371 }
1372 
1373 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq)
1374 {
1375 	struct mlx5_ib_cq *cq;
1376 
1377 	if (!ibcq)
1378 		return 128;
1379 
1380 	cq = to_mcq(ibcq);
1381 	return cq->cqe_size;
1382 }
1383 
1384 /* Called from atomic context */
1385 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1386 {
1387 	struct mlx5_ib_wc *soft_wc;
1388 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1389 	unsigned long flags;
1390 
1391 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1392 	if (!soft_wc)
1393 		return -ENOMEM;
1394 
1395 	soft_wc->wc = *wc;
1396 	spin_lock_irqsave(&cq->lock, flags);
1397 	list_add_tail(&soft_wc->list, &cq->wc_list);
1398 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1399 	    wc->status != IB_WC_SUCCESS) {
1400 		cq->notify_flags = 0;
1401 		schedule_work(&cq->notify_work);
1402 	}
1403 	spin_unlock_irqrestore(&cq->lock, flags);
1404 
1405 	return 0;
1406 }
1407