xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision 5abbeebd)
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 		     struct ib_wc *wc, const char *level)
272 {
273 	mlx5_ib_log(level, dev, "WC error: %d, Message: %s\n", wc->status,
274 		    ib_wc_status_msg(wc->status));
275 	print_hex_dump(level, "cqe_dump: ", DUMP_PREFIX_OFFSET, 16, 1,
276 		       cqe, sizeof(*cqe), false);
277 }
278 
279 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
280 				  struct mlx5_err_cqe *cqe,
281 				  struct ib_wc *wc)
282 {
283 	const char *dump = KERN_WARNING;
284 
285 	switch (cqe->syndrome) {
286 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
287 		wc->status = IB_WC_LOC_LEN_ERR;
288 		break;
289 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
290 		wc->status = IB_WC_LOC_QP_OP_ERR;
291 		break;
292 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
293 		dump = KERN_DEBUG;
294 		wc->status = IB_WC_LOC_PROT_ERR;
295 		break;
296 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
297 		dump = NULL;
298 		wc->status = IB_WC_WR_FLUSH_ERR;
299 		break;
300 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
301 		wc->status = IB_WC_MW_BIND_ERR;
302 		break;
303 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
304 		wc->status = IB_WC_BAD_RESP_ERR;
305 		break;
306 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
307 		wc->status = IB_WC_LOC_ACCESS_ERR;
308 		break;
309 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
310 		wc->status = IB_WC_REM_INV_REQ_ERR;
311 		break;
312 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
313 		dump = KERN_DEBUG;
314 		wc->status = IB_WC_REM_ACCESS_ERR;
315 		break;
316 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
317 		dump = KERN_DEBUG;
318 		wc->status = IB_WC_REM_OP_ERR;
319 		break;
320 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
321 		dump = NULL;
322 		wc->status = IB_WC_RETRY_EXC_ERR;
323 		break;
324 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
325 		dump = NULL;
326 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
327 		break;
328 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
329 		wc->status = IB_WC_REM_ABORT_ERR;
330 		break;
331 	default:
332 		wc->status = IB_WC_GENERAL_ERR;
333 		break;
334 	}
335 
336 	wc->vendor_err = cqe->vendor_err_synd;
337 	if (dump)
338 		dump_cqe(dev, cqe, wc, dump);
339 }
340 
341 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
342 			   u16 tail, u16 head)
343 {
344 	u16 idx;
345 
346 	do {
347 		idx = tail & (qp->sq.wqe_cnt - 1);
348 		if (idx == head)
349 			break;
350 
351 		tail = qp->sq.w_list[idx].next;
352 	} while (1);
353 	tail = qp->sq.w_list[idx].next;
354 	qp->sq.last_poll = tail;
355 }
356 
357 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
358 {
359 	mlx5_frag_buf_free(dev->mdev, &buf->frag_buf);
360 }
361 
362 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
363 			     struct ib_sig_err *item)
364 {
365 	u16 syndrome = be16_to_cpu(cqe->syndrome);
366 
367 #define GUARD_ERR   (1 << 13)
368 #define APPTAG_ERR  (1 << 12)
369 #define REFTAG_ERR  (1 << 11)
370 
371 	if (syndrome & GUARD_ERR) {
372 		item->err_type = IB_SIG_BAD_GUARD;
373 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
374 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
375 	} else
376 	if (syndrome & REFTAG_ERR) {
377 		item->err_type = IB_SIG_BAD_REFTAG;
378 		item->expected = be32_to_cpu(cqe->expected_reftag);
379 		item->actual = be32_to_cpu(cqe->actual_reftag);
380 	} else
381 	if (syndrome & APPTAG_ERR) {
382 		item->err_type = IB_SIG_BAD_APPTAG;
383 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
384 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
385 	} else {
386 		pr_err("Got signature completion error with bad syndrome %04x\n",
387 		       syndrome);
388 	}
389 
390 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
391 	item->key = be32_to_cpu(cqe->mkey);
392 }
393 
394 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
395 		    int *npolled, bool is_send)
396 {
397 	struct mlx5_ib_wq *wq;
398 	unsigned int cur;
399 	int np;
400 	int i;
401 
402 	wq = (is_send) ? &qp->sq : &qp->rq;
403 	cur = wq->head - wq->tail;
404 	np = *npolled;
405 
406 	if (cur == 0)
407 		return;
408 
409 	for (i = 0;  i < cur && np < num_entries; i++) {
410 		unsigned int idx;
411 
412 		idx = (is_send) ? wq->last_poll : wq->tail;
413 		idx &= (wq->wqe_cnt - 1);
414 		wc->wr_id = wq->wrid[idx];
415 		wc->status = IB_WC_WR_FLUSH_ERR;
416 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
417 		wq->tail++;
418 		if (is_send)
419 			wq->last_poll = wq->w_list[idx].next;
420 		np++;
421 		wc->qp = &qp->ibqp;
422 		wc++;
423 	}
424 	*npolled = np;
425 }
426 
427 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
428 				 struct ib_wc *wc, int *npolled)
429 {
430 	struct mlx5_ib_qp *qp;
431 
432 	*npolled = 0;
433 	/* Find uncompleted WQEs belonging to that cq and return mmics ones */
434 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
435 		sw_comp(qp, num_entries, wc + *npolled, npolled, true);
436 		if (*npolled >= num_entries)
437 			return;
438 	}
439 
440 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
441 		sw_comp(qp, num_entries, wc + *npolled, npolled, false);
442 		if (*npolled >= num_entries)
443 			return;
444 	}
445 }
446 
447 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
448 			 struct mlx5_ib_qp **cur_qp,
449 			 struct ib_wc *wc)
450 {
451 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
452 	struct mlx5_err_cqe *err_cqe;
453 	struct mlx5_cqe64 *cqe64;
454 	struct mlx5_core_qp *mqp;
455 	struct mlx5_ib_wq *wq;
456 	uint8_t opcode;
457 	uint32_t qpn;
458 	u16 wqe_ctr;
459 	void *cqe;
460 	int idx;
461 
462 repoll:
463 	cqe = next_cqe_sw(cq);
464 	if (!cqe)
465 		return -EAGAIN;
466 
467 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
468 
469 	++cq->mcq.cons_index;
470 
471 	/* Make sure we read CQ entry contents after we've checked the
472 	 * ownership bit.
473 	 */
474 	rmb();
475 
476 	opcode = get_cqe_opcode(cqe64);
477 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
478 		if (likely(cq->resize_buf)) {
479 			free_cq_buf(dev, &cq->buf);
480 			cq->buf = *cq->resize_buf;
481 			kfree(cq->resize_buf);
482 			cq->resize_buf = NULL;
483 			goto repoll;
484 		} else {
485 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
486 		}
487 	}
488 
489 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
490 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
491 		/* We do not have to take the QP table lock here,
492 		 * because CQs will be locked while QPs are removed
493 		 * from the table.
494 		 */
495 		mqp = radix_tree_lookup(&dev->qp_table.tree, qpn);
496 		*cur_qp = to_mibqp(mqp);
497 	}
498 
499 	wc->qp  = &(*cur_qp)->ibqp;
500 	switch (opcode) {
501 	case MLX5_CQE_REQ:
502 		wq = &(*cur_qp)->sq;
503 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
504 		idx = wqe_ctr & (wq->wqe_cnt - 1);
505 		handle_good_req(wc, cqe64, wq, idx);
506 		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
507 		wc->wr_id = wq->wrid[idx];
508 		wq->tail = wq->wqe_head[idx] + 1;
509 		wc->status = IB_WC_SUCCESS;
510 		break;
511 	case MLX5_CQE_RESP_WR_IMM:
512 	case MLX5_CQE_RESP_SEND:
513 	case MLX5_CQE_RESP_SEND_IMM:
514 	case MLX5_CQE_RESP_SEND_INV:
515 		handle_responder(wc, cqe64, *cur_qp);
516 		wc->status = IB_WC_SUCCESS;
517 		break;
518 	case MLX5_CQE_RESIZE_CQ:
519 		break;
520 	case MLX5_CQE_REQ_ERR:
521 	case MLX5_CQE_RESP_ERR:
522 		err_cqe = (struct mlx5_err_cqe *)cqe64;
523 		mlx5_handle_error_cqe(dev, err_cqe, wc);
524 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
525 			    opcode == MLX5_CQE_REQ_ERR ?
526 			    "Requestor" : "Responder", cq->mcq.cqn);
527 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
528 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
529 		if (wc->status != IB_WC_WR_FLUSH_ERR &&
530 		    (*cur_qp)->type == MLX5_IB_QPT_REG_UMR)
531 			dev->umrc.state = MLX5_UMR_STATE_RECOVER;
532 
533 		if (opcode == MLX5_CQE_REQ_ERR) {
534 			wq = &(*cur_qp)->sq;
535 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
536 			idx = wqe_ctr & (wq->wqe_cnt - 1);
537 			wc->wr_id = wq->wrid[idx];
538 			wq->tail = wq->wqe_head[idx] + 1;
539 		} else {
540 			struct mlx5_ib_srq *srq;
541 
542 			if ((*cur_qp)->ibqp.srq) {
543 				srq = to_msrq((*cur_qp)->ibqp.srq);
544 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
545 				wc->wr_id = srq->wrid[wqe_ctr];
546 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
547 			} else {
548 				wq = &(*cur_qp)->rq;
549 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
550 				++wq->tail;
551 			}
552 		}
553 		break;
554 	case MLX5_CQE_SIG_ERR: {
555 		struct mlx5_sig_err_cqe *sig_err_cqe =
556 			(struct mlx5_sig_err_cqe *)cqe64;
557 		struct mlx5_core_sig_ctx *sig;
558 
559 		xa_lock(&dev->sig_mrs);
560 		sig = xa_load(&dev->sig_mrs,
561 				mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
562 		get_sig_err_item(sig_err_cqe, &sig->err_item);
563 		sig->sig_err_exists = true;
564 		sig->sigerr_count++;
565 
566 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
567 			     cq->mcq.cqn, sig->err_item.key,
568 			     sig->err_item.err_type,
569 			     sig->err_item.sig_err_offset,
570 			     sig->err_item.expected,
571 			     sig->err_item.actual);
572 
573 		xa_unlock(&dev->sig_mrs);
574 		goto repoll;
575 	}
576 	}
577 
578 	return 0;
579 }
580 
581 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
582 			struct ib_wc *wc, bool is_fatal_err)
583 {
584 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
585 	struct mlx5_ib_wc *soft_wc, *next;
586 	int npolled = 0;
587 
588 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
589 		if (npolled >= num_entries)
590 			break;
591 
592 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
593 			    cq->mcq.cqn);
594 
595 		if (unlikely(is_fatal_err)) {
596 			soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
597 			soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
598 		}
599 		wc[npolled++] = soft_wc->wc;
600 		list_del(&soft_wc->list);
601 		kfree(soft_wc);
602 	}
603 
604 	return npolled;
605 }
606 
607 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
608 {
609 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
610 	struct mlx5_ib_qp *cur_qp = NULL;
611 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
612 	struct mlx5_core_dev *mdev = dev->mdev;
613 	unsigned long flags;
614 	int soft_polled = 0;
615 	int npolled;
616 
617 	spin_lock_irqsave(&cq->lock, flags);
618 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
619 		/* make sure no soft wqe's are waiting */
620 		if (unlikely(!list_empty(&cq->wc_list)))
621 			soft_polled = poll_soft_wc(cq, num_entries, wc, true);
622 
623 		mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
624 				     wc + soft_polled, &npolled);
625 		goto out;
626 	}
627 
628 	if (unlikely(!list_empty(&cq->wc_list)))
629 		soft_polled = poll_soft_wc(cq, num_entries, wc, false);
630 
631 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
632 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
633 			break;
634 	}
635 
636 	if (npolled)
637 		mlx5_cq_set_ci(&cq->mcq);
638 out:
639 	spin_unlock_irqrestore(&cq->lock, flags);
640 
641 	return soft_polled + npolled;
642 }
643 
644 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
645 {
646 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
647 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
648 	void __iomem *uar_page = mdev->priv.uar->map;
649 	unsigned long irq_flags;
650 	int ret = 0;
651 
652 	spin_lock_irqsave(&cq->lock, irq_flags);
653 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
654 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
655 
656 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
657 		ret = 1;
658 	spin_unlock_irqrestore(&cq->lock, irq_flags);
659 
660 	mlx5_cq_arm(&cq->mcq,
661 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
662 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
663 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
664 
665 	return ret;
666 }
667 
668 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,
669 			     struct mlx5_ib_cq_buf *buf,
670 			     int nent,
671 			     int cqe_size)
672 {
673 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
674 	u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0);
675 	u8 log_wq_sz     = ilog2(cqe_size);
676 	int err;
677 
678 	err = mlx5_frag_buf_alloc_node(dev->mdev,
679 				       nent * cqe_size,
680 				       frag_buf,
681 				       dev->mdev->priv.numa_node);
682 	if (err)
683 		return err;
684 
685 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
686 
687 	buf->cqe_size = cqe_size;
688 	buf->nent = nent;
689 
690 	return 0;
691 }
692 
693 enum {
694 	MLX5_CQE_RES_FORMAT_HASH = 0,
695 	MLX5_CQE_RES_FORMAT_CSUM = 1,
696 	MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3,
697 };
698 
699 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
700 {
701 	switch (format) {
702 	case MLX5_IB_CQE_RES_FORMAT_HASH:
703 		return MLX5_CQE_RES_FORMAT_HASH;
704 	case MLX5_IB_CQE_RES_FORMAT_CSUM:
705 		return MLX5_CQE_RES_FORMAT_CSUM;
706 	case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX:
707 		if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
708 			return MLX5_CQE_RES_FORMAT_CSUM_STRIDX;
709 		return -EOPNOTSUPP;
710 	default:
711 		return -EINVAL;
712 	}
713 }
714 
715 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
716 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
717 			  int *cqe_size, int *index, int *inlen)
718 {
719 	struct mlx5_ib_create_cq ucmd = {};
720 	unsigned long page_size;
721 	unsigned int page_offset_quantized;
722 	size_t ucmdlen;
723 	__be64 *pas;
724 	int ncont;
725 	void *cqc;
726 	int err;
727 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
728 		udata, struct mlx5_ib_ucontext, ibucontext);
729 
730 	ucmdlen = min(udata->inlen, sizeof(ucmd));
731 	if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags))
732 		return -EINVAL;
733 
734 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
735 		return -EFAULT;
736 
737 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD |
738 			    MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX |
739 			    MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)))
740 		return -EINVAL;
741 
742 	if ((ucmd.cqe_size != 64 && ucmd.cqe_size != 128) ||
743 	    ucmd.reserved0 || ucmd.reserved1)
744 		return -EINVAL;
745 
746 	*cqe_size = ucmd.cqe_size;
747 
748 	cq->buf.umem =
749 		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
750 			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
751 	if (IS_ERR(cq->buf.umem)) {
752 		err = PTR_ERR(cq->buf.umem);
753 		return err;
754 	}
755 
756 	page_size = mlx5_umem_find_best_cq_quantized_pgoff(
757 		cq->buf.umem, cqc, log_page_size, MLX5_ADAPTER_PAGE_SHIFT,
758 		page_offset, 64, &page_offset_quantized);
759 	if (!page_size) {
760 		err = -EINVAL;
761 		goto err_umem;
762 	}
763 
764 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &cq->db);
765 	if (err)
766 		goto err_umem;
767 
768 	ncont = ib_umem_num_dma_blocks(cq->buf.umem, page_size);
769 	mlx5_ib_dbg(
770 		dev,
771 		"addr 0x%llx, size %u, npages %zu, page_size %lu, ncont %d\n",
772 		ucmd.buf_addr, entries * ucmd.cqe_size,
773 		ib_umem_num_pages(cq->buf.umem), page_size, ncont);
774 
775 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
776 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
777 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
778 	if (!*cqb) {
779 		err = -ENOMEM;
780 		goto err_db;
781 	}
782 
783 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
784 	mlx5_ib_populate_pas(cq->buf.umem, page_size, pas, 0);
785 
786 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
787 	MLX5_SET(cqc, cqc, log_page_size,
788 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
789 	MLX5_SET(cqc, cqc, page_offset, page_offset_quantized);
790 
791 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) {
792 		*index = ucmd.uar_page_index;
793 	} else if (context->bfregi.lib_uar_dyn) {
794 		err = -EINVAL;
795 		goto err_cqb;
796 	} else {
797 		*index = context->bfregi.sys_pages[0];
798 	}
799 
800 	if (ucmd.cqe_comp_en == 1) {
801 		int mini_cqe_format;
802 
803 		if (!((*cqe_size == 128 &&
804 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
805 		      (*cqe_size == 64  &&
806 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
807 			err = -EOPNOTSUPP;
808 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
809 				     *cqe_size);
810 			goto err_cqb;
811 		}
812 
813 		mini_cqe_format =
814 			mini_cqe_res_format_to_hw(dev,
815 						  ucmd.cqe_comp_res_format);
816 		if (mini_cqe_format < 0) {
817 			err = mini_cqe_format;
818 			mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n",
819 				    ucmd.cqe_comp_res_format, err);
820 			goto err_cqb;
821 		}
822 
823 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
824 		MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format);
825 	}
826 
827 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
828 		if (*cqe_size != 128 ||
829 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
830 			err = -EOPNOTSUPP;
831 			mlx5_ib_warn(dev,
832 				     "CQE padding is not supported for CQE size of %dB!\n",
833 				     *cqe_size);
834 			goto err_cqb;
835 		}
836 
837 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
838 	}
839 
840 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)
841 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS;
842 
843 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
844 	return 0;
845 
846 err_cqb:
847 	kvfree(*cqb);
848 
849 err_db:
850 	mlx5_ib_db_unmap_user(context, &cq->db);
851 
852 err_umem:
853 	ib_umem_release(cq->buf.umem);
854 	return err;
855 }
856 
857 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
858 {
859 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
860 		udata, struct mlx5_ib_ucontext, ibucontext);
861 
862 	mlx5_ib_db_unmap_user(context, &cq->db);
863 	ib_umem_release(cq->buf.umem);
864 }
865 
866 static void init_cq_frag_buf(struct mlx5_ib_cq_buf *buf)
867 {
868 	int i;
869 	void *cqe;
870 	struct mlx5_cqe64 *cqe64;
871 
872 	for (i = 0; i < buf->nent; i++) {
873 		cqe = mlx5_frag_buf_get_wqe(&buf->fbc, i);
874 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
875 		cqe64->op_own = MLX5_CQE_INVALID << 4;
876 	}
877 }
878 
879 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
880 			    int entries, int cqe_size,
881 			    u32 **cqb, int *index, int *inlen)
882 {
883 	__be64 *pas;
884 	void *cqc;
885 	int err;
886 
887 	err = mlx5_db_alloc(dev->mdev, &cq->db);
888 	if (err)
889 		return err;
890 
891 	cq->mcq.set_ci_db  = cq->db.db;
892 	cq->mcq.arm_db     = cq->db.db + 1;
893 	cq->mcq.cqe_sz = cqe_size;
894 
895 	err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size);
896 	if (err)
897 		goto err_db;
898 
899 	init_cq_frag_buf(&cq->buf);
900 
901 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
902 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) *
903 		 cq->buf.frag_buf.npages;
904 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
905 	if (!*cqb) {
906 		err = -ENOMEM;
907 		goto err_buf;
908 	}
909 
910 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
911 	mlx5_fill_page_frag_array(&cq->buf.frag_buf, pas);
912 
913 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
914 	MLX5_SET(cqc, cqc, log_page_size,
915 		 cq->buf.frag_buf.page_shift -
916 		 MLX5_ADAPTER_PAGE_SHIFT);
917 
918 	*index = dev->mdev->priv.uar->index;
919 
920 	return 0;
921 
922 err_buf:
923 	free_cq_buf(dev, &cq->buf);
924 
925 err_db:
926 	mlx5_db_free(dev->mdev, &cq->db);
927 	return err;
928 }
929 
930 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
931 {
932 	free_cq_buf(dev, &cq->buf);
933 	mlx5_db_free(dev->mdev, &cq->db);
934 }
935 
936 static void notify_soft_wc_handler(struct work_struct *work)
937 {
938 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
939 					     notify_work);
940 
941 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
942 }
943 
944 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
945 		      struct ib_udata *udata)
946 {
947 	struct ib_device *ibdev = ibcq->device;
948 	int entries = attr->cqe;
949 	int vector = attr->comp_vector;
950 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
951 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
952 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
953 	int index;
954 	int inlen;
955 	u32 *cqb = NULL;
956 	void *cqc;
957 	int cqe_size;
958 	int eqn;
959 	int err;
960 
961 	if (entries < 0 ||
962 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
963 		return -EINVAL;
964 
965 	if (check_cq_create_flags(attr->flags))
966 		return -EOPNOTSUPP;
967 
968 	entries = roundup_pow_of_two(entries + 1);
969 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
970 		return -EINVAL;
971 
972 	cq->ibcq.cqe = entries - 1;
973 	mutex_init(&cq->resize_mutex);
974 	spin_lock_init(&cq->lock);
975 	cq->resize_buf = NULL;
976 	cq->resize_umem = NULL;
977 	cq->create_flags = attr->flags;
978 	INIT_LIST_HEAD(&cq->list_send_qp);
979 	INIT_LIST_HEAD(&cq->list_recv_qp);
980 
981 	if (udata) {
982 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
983 				     &index, &inlen);
984 		if (err)
985 			return err;
986 	} else {
987 		cqe_size = cache_line_size() == 128 ? 128 : 64;
988 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
989 				       &index, &inlen);
990 		if (err)
991 			return err;
992 
993 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
994 	}
995 
996 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn);
997 	if (err)
998 		goto err_cqb;
999 
1000 	cq->cqe_size = cqe_size;
1001 
1002 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
1003 	MLX5_SET(cqc, cqc, cqe_sz,
1004 		 cqe_sz_to_mlx_sz(cqe_size,
1005 				  cq->private_flags &
1006 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1007 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1008 	MLX5_SET(cqc, cqc, uar_page, index);
1009 	MLX5_SET(cqc, cqc, c_eqn_or_apu_element, eqn);
1010 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
1011 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
1012 		MLX5_SET(cqc, cqc, oi, 1);
1013 
1014 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
1015 	if (err)
1016 		goto err_cqb;
1017 
1018 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
1019 	if (udata)
1020 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
1021 	else
1022 		cq->mcq.comp  = mlx5_ib_cq_comp;
1023 	cq->mcq.event = mlx5_ib_cq_event;
1024 
1025 	INIT_LIST_HEAD(&cq->wc_list);
1026 
1027 	if (udata)
1028 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
1029 			err = -EFAULT;
1030 			goto err_cmd;
1031 		}
1032 
1033 
1034 	kvfree(cqb);
1035 	return 0;
1036 
1037 err_cmd:
1038 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
1039 
1040 err_cqb:
1041 	kvfree(cqb);
1042 	if (udata)
1043 		destroy_cq_user(cq, udata);
1044 	else
1045 		destroy_cq_kernel(dev, cq);
1046 	return err;
1047 }
1048 
1049 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1050 {
1051 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1052 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1053 	int ret;
1054 
1055 	ret = mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1056 	if (ret)
1057 		return ret;
1058 
1059 	if (udata)
1060 		destroy_cq_user(mcq, udata);
1061 	else
1062 		destroy_cq_kernel(dev, mcq);
1063 	return 0;
1064 }
1065 
1066 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1067 {
1068 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1069 }
1070 
1071 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1072 {
1073 	struct mlx5_cqe64 *cqe64, *dest64;
1074 	void *cqe, *dest;
1075 	u32 prod_index;
1076 	int nfreed = 0;
1077 	u8 owner_bit;
1078 
1079 	if (!cq)
1080 		return;
1081 
1082 	/* First we need to find the current producer index, so we
1083 	 * know where to start cleaning from.  It doesn't matter if HW
1084 	 * adds new entries after this loop -- the QP we're worried
1085 	 * about is already in RESET, so the new entries won't come
1086 	 * from our QP and therefore don't need to be checked.
1087 	 */
1088 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1089 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1090 			break;
1091 
1092 	/* Now sweep backwards through the CQ, removing CQ entries
1093 	 * that match our QP by copying older entries on top of them.
1094 	 */
1095 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1096 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1097 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1098 		if (is_equal_rsn(cqe64, rsn)) {
1099 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1100 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1101 			++nfreed;
1102 		} else if (nfreed) {
1103 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1104 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1105 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1106 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1107 			dest64->op_own = owner_bit |
1108 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1109 		}
1110 	}
1111 
1112 	if (nfreed) {
1113 		cq->mcq.cons_index += nfreed;
1114 		/* Make sure update of buffer contents is done before
1115 		 * updating consumer index.
1116 		 */
1117 		wmb();
1118 		mlx5_cq_set_ci(&cq->mcq);
1119 	}
1120 }
1121 
1122 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1123 {
1124 	if (!cq)
1125 		return;
1126 
1127 	spin_lock_irq(&cq->lock);
1128 	__mlx5_ib_cq_clean(cq, qpn, srq);
1129 	spin_unlock_irq(&cq->lock);
1130 }
1131 
1132 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1133 {
1134 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1135 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1136 	int err;
1137 
1138 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1139 		return -EOPNOTSUPP;
1140 
1141 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1142 		return -EINVAL;
1143 
1144 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1145 					     cq_period, cq_count);
1146 	if (err)
1147 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1148 
1149 	return err;
1150 }
1151 
1152 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1153 		       int entries, struct ib_udata *udata,
1154 		       int *cqe_size)
1155 {
1156 	struct mlx5_ib_resize_cq ucmd;
1157 	struct ib_umem *umem;
1158 	int err;
1159 
1160 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1161 	if (err)
1162 		return err;
1163 
1164 	if (ucmd.reserved0 || ucmd.reserved1)
1165 		return -EINVAL;
1166 
1167 	/* check multiplication overflow */
1168 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1169 		return -EINVAL;
1170 
1171 	umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
1172 			   (size_t)ucmd.cqe_size * entries,
1173 			   IB_ACCESS_LOCAL_WRITE);
1174 	if (IS_ERR(umem)) {
1175 		err = PTR_ERR(umem);
1176 		return err;
1177 	}
1178 
1179 	cq->resize_umem = umem;
1180 	*cqe_size = ucmd.cqe_size;
1181 
1182 	return 0;
1183 }
1184 
1185 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1186 			 int entries, int cqe_size)
1187 {
1188 	int err;
1189 
1190 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1191 	if (!cq->resize_buf)
1192 		return -ENOMEM;
1193 
1194 	err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size);
1195 	if (err)
1196 		goto ex;
1197 
1198 	init_cq_frag_buf(cq->resize_buf);
1199 
1200 	return 0;
1201 
1202 ex:
1203 	kfree(cq->resize_buf);
1204 	return err;
1205 }
1206 
1207 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1208 {
1209 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1210 	struct mlx5_cqe64 *scqe64;
1211 	struct mlx5_cqe64 *dcqe64;
1212 	void *start_cqe;
1213 	void *scqe;
1214 	void *dcqe;
1215 	int ssize;
1216 	int dsize;
1217 	int i;
1218 	u8 sw_own;
1219 
1220 	ssize = cq->buf.cqe_size;
1221 	dsize = cq->resize_buf->cqe_size;
1222 	if (ssize != dsize) {
1223 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1224 		return -EINVAL;
1225 	}
1226 
1227 	i = cq->mcq.cons_index;
1228 	scqe = get_sw_cqe(cq, i);
1229 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1230 	start_cqe = scqe;
1231 	if (!scqe) {
1232 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1233 		return -EINVAL;
1234 	}
1235 
1236 	while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) {
1237 		dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc,
1238 					     (i + 1) & cq->resize_buf->nent);
1239 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1240 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1241 		memcpy(dcqe, scqe, dsize);
1242 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1243 
1244 		++i;
1245 		scqe = get_sw_cqe(cq, i);
1246 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1247 		if (!scqe) {
1248 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1249 			return -EINVAL;
1250 		}
1251 
1252 		if (scqe == start_cqe) {
1253 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1254 				cq->mcq.cqn);
1255 			return -ENOMEM;
1256 		}
1257 	}
1258 	++cq->mcq.cons_index;
1259 	return 0;
1260 }
1261 
1262 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1263 {
1264 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1265 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1266 	void *cqc;
1267 	u32 *in;
1268 	int err;
1269 	int npas;
1270 	__be64 *pas;
1271 	unsigned int page_offset_quantized = 0;
1272 	unsigned int page_shift;
1273 	int inlen;
1274 	int cqe_size;
1275 	unsigned long flags;
1276 
1277 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1278 		pr_info("Firmware does not support resize CQ\n");
1279 		return -ENOSYS;
1280 	}
1281 
1282 	if (entries < 1 ||
1283 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1284 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1285 			     entries,
1286 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1287 		return -EINVAL;
1288 	}
1289 
1290 	entries = roundup_pow_of_two(entries + 1);
1291 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1292 		return -EINVAL;
1293 
1294 	if (entries == ibcq->cqe + 1)
1295 		return 0;
1296 
1297 	mutex_lock(&cq->resize_mutex);
1298 	if (udata) {
1299 		unsigned long page_size;
1300 
1301 		err = resize_user(dev, cq, entries, udata, &cqe_size);
1302 		if (err)
1303 			goto ex;
1304 
1305 		page_size = mlx5_umem_find_best_cq_quantized_pgoff(
1306 			cq->resize_umem, cqc, log_page_size,
1307 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
1308 			&page_offset_quantized);
1309 		if (!page_size) {
1310 			err = -EINVAL;
1311 			goto ex_resize;
1312 		}
1313 		npas = ib_umem_num_dma_blocks(cq->resize_umem, page_size);
1314 		page_shift = order_base_2(page_size);
1315 	} else {
1316 		struct mlx5_frag_buf *frag_buf;
1317 
1318 		cqe_size = 64;
1319 		err = resize_kernel(dev, cq, entries, cqe_size);
1320 		if (err)
1321 			goto ex;
1322 		frag_buf = &cq->resize_buf->frag_buf;
1323 		npas = frag_buf->npages;
1324 		page_shift = frag_buf->page_shift;
1325 	}
1326 
1327 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1328 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1329 
1330 	in = kvzalloc(inlen, GFP_KERNEL);
1331 	if (!in) {
1332 		err = -ENOMEM;
1333 		goto ex_resize;
1334 	}
1335 
1336 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1337 	if (udata)
1338 		mlx5_ib_populate_pas(cq->resize_umem, 1UL << page_shift, pas,
1339 				     0);
1340 	else
1341 		mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas);
1342 
1343 	MLX5_SET(modify_cq_in, in,
1344 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1345 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1346 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1347 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1348 
1349 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1350 
1351 	MLX5_SET(cqc, cqc, log_page_size,
1352 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1353 	MLX5_SET(cqc, cqc, page_offset, page_offset_quantized);
1354 	MLX5_SET(cqc, cqc, cqe_sz,
1355 		 cqe_sz_to_mlx_sz(cqe_size,
1356 				  cq->private_flags &
1357 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1358 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1359 
1360 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1361 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1362 
1363 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1364 	if (err)
1365 		goto ex_alloc;
1366 
1367 	if (udata) {
1368 		cq->ibcq.cqe = entries - 1;
1369 		ib_umem_release(cq->buf.umem);
1370 		cq->buf.umem = cq->resize_umem;
1371 		cq->resize_umem = NULL;
1372 	} else {
1373 		struct mlx5_ib_cq_buf tbuf;
1374 		int resized = 0;
1375 
1376 		spin_lock_irqsave(&cq->lock, flags);
1377 		if (cq->resize_buf) {
1378 			err = copy_resize_cqes(cq);
1379 			if (!err) {
1380 				tbuf = cq->buf;
1381 				cq->buf = *cq->resize_buf;
1382 				kfree(cq->resize_buf);
1383 				cq->resize_buf = NULL;
1384 				resized = 1;
1385 			}
1386 		}
1387 		cq->ibcq.cqe = entries - 1;
1388 		spin_unlock_irqrestore(&cq->lock, flags);
1389 		if (resized)
1390 			free_cq_buf(dev, &tbuf);
1391 	}
1392 	mutex_unlock(&cq->resize_mutex);
1393 
1394 	kvfree(in);
1395 	return 0;
1396 
1397 ex_alloc:
1398 	kvfree(in);
1399 
1400 ex_resize:
1401 	ib_umem_release(cq->resize_umem);
1402 	if (!udata) {
1403 		free_cq_buf(dev, cq->resize_buf);
1404 		cq->resize_buf = NULL;
1405 	}
1406 ex:
1407 	mutex_unlock(&cq->resize_mutex);
1408 	return err;
1409 }
1410 
1411 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq)
1412 {
1413 	struct mlx5_ib_cq *cq;
1414 
1415 	if (!ibcq)
1416 		return 128;
1417 
1418 	cq = to_mcq(ibcq);
1419 	return cq->cqe_size;
1420 }
1421 
1422 /* Called from atomic context */
1423 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1424 {
1425 	struct mlx5_ib_wc *soft_wc;
1426 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1427 	unsigned long flags;
1428 
1429 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1430 	if (!soft_wc)
1431 		return -ENOMEM;
1432 
1433 	soft_wc->wc = *wc;
1434 	spin_lock_irqsave(&cq->lock, flags);
1435 	list_add_tail(&soft_wc->list, &cq->wc_list);
1436 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1437 	    wc->status != IB_WC_SUCCESS) {
1438 		cq->notify_flags = 0;
1439 		schedule_work(&cq->notify_work);
1440 	}
1441 	spin_unlock_irqrestore(&cq->lock, flags);
1442 
1443 	return 0;
1444 }
1445