xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision f8e17c17)
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 
40 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe)
41 {
42 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
43 
44 	ibcq->comp_handler(ibcq, ibcq->cq_context);
45 }
46 
47 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type)
48 {
49 	struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq);
50 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
51 	struct ib_cq *ibcq = &cq->ibcq;
52 	struct ib_event event;
53 
54 	if (type != MLX5_EVENT_TYPE_CQ_ERROR) {
55 		mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n",
56 			     type, mcq->cqn);
57 		return;
58 	}
59 
60 	if (ibcq->event_handler) {
61 		event.device     = &dev->ib_dev;
62 		event.event      = IB_EVENT_CQ_ERR;
63 		event.element.cq = ibcq;
64 		ibcq->event_handler(&event, ibcq->cq_context);
65 	}
66 }
67 
68 static void *get_cqe(struct mlx5_ib_cq *cq, int n)
69 {
70 	return mlx5_frag_buf_get_wqe(&cq->buf.fbc, n);
71 }
72 
73 static u8 sw_ownership_bit(int n, int nent)
74 {
75 	return (n & nent) ? 1 : 0;
76 }
77 
78 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n)
79 {
80 	void *cqe = get_cqe(cq, n & cq->ibcq.cqe);
81 	struct mlx5_cqe64 *cqe64;
82 
83 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
84 
85 	if (likely(get_cqe_opcode(cqe64) != MLX5_CQE_INVALID) &&
86 	    !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) {
87 		return cqe;
88 	} else {
89 		return NULL;
90 	}
91 }
92 
93 static void *next_cqe_sw(struct mlx5_ib_cq *cq)
94 {
95 	return get_sw_cqe(cq, cq->mcq.cons_index);
96 }
97 
98 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx)
99 {
100 	switch (wq->wr_data[idx]) {
101 	case MLX5_IB_WR_UMR:
102 		return 0;
103 
104 	case IB_WR_LOCAL_INV:
105 		return IB_WC_LOCAL_INV;
106 
107 	case IB_WR_REG_MR:
108 		return IB_WC_REG_MR;
109 
110 	default:
111 		pr_warn("unknown completion status\n");
112 		return 0;
113 	}
114 }
115 
116 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
117 			    struct mlx5_ib_wq *wq, int idx)
118 {
119 	wc->wc_flags = 0;
120 	switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) {
121 	case MLX5_OPCODE_RDMA_WRITE_IMM:
122 		wc->wc_flags |= IB_WC_WITH_IMM;
123 		/* fall through */
124 	case MLX5_OPCODE_RDMA_WRITE:
125 		wc->opcode    = IB_WC_RDMA_WRITE;
126 		break;
127 	case MLX5_OPCODE_SEND_IMM:
128 		wc->wc_flags |= IB_WC_WITH_IMM;
129 		/* fall through */
130 	case MLX5_OPCODE_SEND:
131 	case MLX5_OPCODE_SEND_INVAL:
132 		wc->opcode    = IB_WC_SEND;
133 		break;
134 	case MLX5_OPCODE_RDMA_READ:
135 		wc->opcode    = IB_WC_RDMA_READ;
136 		wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
137 		break;
138 	case MLX5_OPCODE_ATOMIC_CS:
139 		wc->opcode    = IB_WC_COMP_SWAP;
140 		wc->byte_len  = 8;
141 		break;
142 	case MLX5_OPCODE_ATOMIC_FA:
143 		wc->opcode    = IB_WC_FETCH_ADD;
144 		wc->byte_len  = 8;
145 		break;
146 	case MLX5_OPCODE_ATOMIC_MASKED_CS:
147 		wc->opcode    = IB_WC_MASKED_COMP_SWAP;
148 		wc->byte_len  = 8;
149 		break;
150 	case MLX5_OPCODE_ATOMIC_MASKED_FA:
151 		wc->opcode    = IB_WC_MASKED_FETCH_ADD;
152 		wc->byte_len  = 8;
153 		break;
154 	case MLX5_OPCODE_UMR:
155 		wc->opcode = get_umr_comp(wq, idx);
156 		break;
157 	}
158 }
159 
160 enum {
161 	MLX5_GRH_IN_BUFFER = 1,
162 	MLX5_GRH_IN_CQE	   = 2,
163 };
164 
165 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe,
166 			     struct mlx5_ib_qp *qp)
167 {
168 	enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1);
169 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
170 	struct mlx5_ib_srq *srq;
171 	struct mlx5_ib_wq *wq;
172 	u16 wqe_ctr;
173 	u8  roce_packet_type;
174 	bool vlan_present;
175 	u8 g;
176 
177 	if (qp->ibqp.srq || qp->ibqp.xrcd) {
178 		struct mlx5_core_srq *msrq = NULL;
179 
180 		if (qp->ibqp.xrcd) {
181 			msrq = mlx5_cmd_get_srq(dev, be32_to_cpu(cqe->srqn));
182 			srq = to_mibsrq(msrq);
183 		} else {
184 			srq = to_msrq(qp->ibqp.srq);
185 		}
186 		if (srq) {
187 			wqe_ctr = be16_to_cpu(cqe->wqe_counter);
188 			wc->wr_id = srq->wrid[wqe_ctr];
189 			mlx5_ib_free_srq_wqe(srq, wqe_ctr);
190 			if (msrq)
191 				mlx5_core_res_put(&msrq->common);
192 		}
193 	} else {
194 		wq	  = &qp->rq;
195 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
196 		++wq->tail;
197 	}
198 	wc->byte_len = be32_to_cpu(cqe->byte_cnt);
199 
200 	switch (get_cqe_opcode(cqe)) {
201 	case MLX5_CQE_RESP_WR_IMM:
202 		wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
203 		wc->wc_flags	= IB_WC_WITH_IMM;
204 		wc->ex.imm_data = cqe->imm_inval_pkey;
205 		break;
206 	case MLX5_CQE_RESP_SEND:
207 		wc->opcode   = IB_WC_RECV;
208 		wc->wc_flags = IB_WC_IP_CSUM_OK;
209 		if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) &&
210 			       (cqe->hds_ip_ext & CQE_L4_OK))))
211 			wc->wc_flags = 0;
212 		break;
213 	case MLX5_CQE_RESP_SEND_IMM:
214 		wc->opcode	= IB_WC_RECV;
215 		wc->wc_flags	= IB_WC_WITH_IMM;
216 		wc->ex.imm_data = cqe->imm_inval_pkey;
217 		break;
218 	case MLX5_CQE_RESP_SEND_INV:
219 		wc->opcode	= IB_WC_RECV;
220 		wc->wc_flags	= IB_WC_WITH_INVALIDATE;
221 		wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey);
222 		break;
223 	}
224 	wc->src_qp	   = be32_to_cpu(cqe->flags_rqpn) & 0xffffff;
225 	wc->dlid_path_bits = cqe->ml_path;
226 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
227 	wc->wc_flags |= g ? IB_WC_GRH : 0;
228 	if (unlikely(is_qp1(qp->ibqp.qp_type))) {
229 		u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff;
230 
231 		ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey,
232 				    &wc->pkey_index);
233 	} else {
234 		wc->pkey_index = 0;
235 	}
236 
237 	if (ll != IB_LINK_LAYER_ETHERNET) {
238 		wc->slid = be16_to_cpu(cqe->slid);
239 		wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf;
240 		return;
241 	}
242 
243 	wc->slid = 0;
244 	vlan_present = cqe->l4_l3_hdr_type & 0x1;
245 	roce_packet_type   = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3;
246 	if (vlan_present) {
247 		wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff;
248 		wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7;
249 		wc->wc_flags |= IB_WC_WITH_VLAN;
250 	} else {
251 		wc->sl = 0;
252 	}
253 
254 	switch (roce_packet_type) {
255 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH:
256 		wc->network_hdr_type = RDMA_NETWORK_IB;
257 		break;
258 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6:
259 		wc->network_hdr_type = RDMA_NETWORK_IPV6;
260 		break;
261 	case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4:
262 		wc->network_hdr_type = RDMA_NETWORK_IPV4;
263 		break;
264 	}
265 	wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
266 }
267 
268 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe)
269 {
270 	mlx5_ib_warn(dev, "dump error cqe\n");
271 	mlx5_dump_err_cqe(dev->mdev, cqe);
272 }
273 
274 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
275 				  struct mlx5_err_cqe *cqe,
276 				  struct ib_wc *wc)
277 {
278 	int dump = 1;
279 
280 	switch (cqe->syndrome) {
281 	case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR:
282 		wc->status = IB_WC_LOC_LEN_ERR;
283 		break;
284 	case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR:
285 		wc->status = IB_WC_LOC_QP_OP_ERR;
286 		break;
287 	case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR:
288 		wc->status = IB_WC_LOC_PROT_ERR;
289 		break;
290 	case MLX5_CQE_SYNDROME_WR_FLUSH_ERR:
291 		dump = 0;
292 		wc->status = IB_WC_WR_FLUSH_ERR;
293 		break;
294 	case MLX5_CQE_SYNDROME_MW_BIND_ERR:
295 		wc->status = IB_WC_MW_BIND_ERR;
296 		break;
297 	case MLX5_CQE_SYNDROME_BAD_RESP_ERR:
298 		wc->status = IB_WC_BAD_RESP_ERR;
299 		break;
300 	case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR:
301 		wc->status = IB_WC_LOC_ACCESS_ERR;
302 		break;
303 	case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
304 		wc->status = IB_WC_REM_INV_REQ_ERR;
305 		break;
306 	case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR:
307 		wc->status = IB_WC_REM_ACCESS_ERR;
308 		break;
309 	case MLX5_CQE_SYNDROME_REMOTE_OP_ERR:
310 		wc->status = IB_WC_REM_OP_ERR;
311 		break;
312 	case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
313 		wc->status = IB_WC_RETRY_EXC_ERR;
314 		dump = 0;
315 		break;
316 	case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
317 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
318 		dump = 0;
319 		break;
320 	case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR:
321 		wc->status = IB_WC_REM_ABORT_ERR;
322 		break;
323 	default:
324 		wc->status = IB_WC_GENERAL_ERR;
325 		break;
326 	}
327 
328 	wc->vendor_err = cqe->vendor_err_synd;
329 	if (dump)
330 		dump_cqe(dev, cqe);
331 }
332 
333 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
334 {
335 	mlx5_frag_buf_free(dev->mdev, &buf->frag_buf);
336 }
337 
338 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
339 			     struct ib_sig_err *item)
340 {
341 	u16 syndrome = be16_to_cpu(cqe->syndrome);
342 
343 #define GUARD_ERR   (1 << 13)
344 #define APPTAG_ERR  (1 << 12)
345 #define REFTAG_ERR  (1 << 11)
346 
347 	if (syndrome & GUARD_ERR) {
348 		item->err_type = IB_SIG_BAD_GUARD;
349 		item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16;
350 		item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16;
351 	} else
352 	if (syndrome & REFTAG_ERR) {
353 		item->err_type = IB_SIG_BAD_REFTAG;
354 		item->expected = be32_to_cpu(cqe->expected_reftag);
355 		item->actual = be32_to_cpu(cqe->actual_reftag);
356 	} else
357 	if (syndrome & APPTAG_ERR) {
358 		item->err_type = IB_SIG_BAD_APPTAG;
359 		item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff;
360 		item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff;
361 	} else {
362 		pr_err("Got signature completion error with bad syndrome %04x\n",
363 		       syndrome);
364 	}
365 
366 	item->sig_err_offset = be64_to_cpu(cqe->err_offset);
367 	item->key = be32_to_cpu(cqe->mkey);
368 }
369 
370 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
371 		    int *npolled, int is_send)
372 {
373 	struct mlx5_ib_wq *wq;
374 	unsigned int cur;
375 	int np;
376 	int i;
377 
378 	wq = (is_send) ? &qp->sq : &qp->rq;
379 	cur = wq->head - wq->tail;
380 	np = *npolled;
381 
382 	if (cur == 0)
383 		return;
384 
385 	for (i = 0;  i < cur && np < num_entries; i++) {
386 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
387 		wc->status = IB_WC_WR_FLUSH_ERR;
388 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
389 		wq->tail++;
390 		np++;
391 		wc->qp = &qp->ibqp;
392 		wc++;
393 	}
394 	*npolled = np;
395 }
396 
397 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
398 				 struct ib_wc *wc, int *npolled)
399 {
400 	struct mlx5_ib_qp *qp;
401 
402 	*npolled = 0;
403 	/* Find uncompleted WQEs belonging to that cq and return mmics ones */
404 	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
405 		sw_comp(qp, num_entries, wc + *npolled, npolled, true);
406 		if (*npolled >= num_entries)
407 			return;
408 	}
409 
410 	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
411 		sw_comp(qp, num_entries, wc + *npolled, npolled, false);
412 		if (*npolled >= num_entries)
413 			return;
414 	}
415 }
416 
417 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
418 			 struct mlx5_ib_qp **cur_qp,
419 			 struct ib_wc *wc)
420 {
421 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
422 	struct mlx5_err_cqe *err_cqe;
423 	struct mlx5_cqe64 *cqe64;
424 	struct mlx5_core_qp *mqp;
425 	struct mlx5_ib_wq *wq;
426 	uint8_t opcode;
427 	uint32_t qpn;
428 	u16 wqe_ctr;
429 	void *cqe;
430 	int idx;
431 
432 repoll:
433 	cqe = next_cqe_sw(cq);
434 	if (!cqe)
435 		return -EAGAIN;
436 
437 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
438 
439 	++cq->mcq.cons_index;
440 
441 	/* Make sure we read CQ entry contents after we've checked the
442 	 * ownership bit.
443 	 */
444 	rmb();
445 
446 	opcode = get_cqe_opcode(cqe64);
447 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
448 		if (likely(cq->resize_buf)) {
449 			free_cq_buf(dev, &cq->buf);
450 			cq->buf = *cq->resize_buf;
451 			kfree(cq->resize_buf);
452 			cq->resize_buf = NULL;
453 			goto repoll;
454 		} else {
455 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
456 		}
457 	}
458 
459 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
460 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
461 		/* We do not have to take the QP table lock here,
462 		 * because CQs will be locked while QPs are removed
463 		 * from the table.
464 		 */
465 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
466 		*cur_qp = to_mibqp(mqp);
467 	}
468 
469 	wc->qp  = &(*cur_qp)->ibqp;
470 	switch (opcode) {
471 	case MLX5_CQE_REQ:
472 		wq = &(*cur_qp)->sq;
473 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
474 		idx = wqe_ctr & (wq->wqe_cnt - 1);
475 		handle_good_req(wc, cqe64, wq, idx);
476 		wc->wr_id = wq->wrid[idx];
477 		wq->tail = wq->wqe_head[idx] + 1;
478 		wc->status = IB_WC_SUCCESS;
479 		break;
480 	case MLX5_CQE_RESP_WR_IMM:
481 	case MLX5_CQE_RESP_SEND:
482 	case MLX5_CQE_RESP_SEND_IMM:
483 	case MLX5_CQE_RESP_SEND_INV:
484 		handle_responder(wc, cqe64, *cur_qp);
485 		wc->status = IB_WC_SUCCESS;
486 		break;
487 	case MLX5_CQE_RESIZE_CQ:
488 		break;
489 	case MLX5_CQE_REQ_ERR:
490 	case MLX5_CQE_RESP_ERR:
491 		err_cqe = (struct mlx5_err_cqe *)cqe64;
492 		mlx5_handle_error_cqe(dev, err_cqe, wc);
493 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
494 			    opcode == MLX5_CQE_REQ_ERR ?
495 			    "Requestor" : "Responder", cq->mcq.cqn);
496 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
497 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
498 		if (opcode == MLX5_CQE_REQ_ERR) {
499 			wq = &(*cur_qp)->sq;
500 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
501 			idx = wqe_ctr & (wq->wqe_cnt - 1);
502 			wc->wr_id = wq->wrid[idx];
503 			wq->tail = wq->wqe_head[idx] + 1;
504 		} else {
505 			struct mlx5_ib_srq *srq;
506 
507 			if ((*cur_qp)->ibqp.srq) {
508 				srq = to_msrq((*cur_qp)->ibqp.srq);
509 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
510 				wc->wr_id = srq->wrid[wqe_ctr];
511 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
512 			} else {
513 				wq = &(*cur_qp)->rq;
514 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
515 				++wq->tail;
516 			}
517 		}
518 		break;
519 	case MLX5_CQE_SIG_ERR: {
520 		struct mlx5_sig_err_cqe *sig_err_cqe =
521 			(struct mlx5_sig_err_cqe *)cqe64;
522 		struct mlx5_core_sig_ctx *sig;
523 
524 		xa_lock(&dev->sig_mrs);
525 		sig = xa_load(&dev->sig_mrs,
526 				mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
527 		get_sig_err_item(sig_err_cqe, &sig->err_item);
528 		sig->sig_err_exists = true;
529 		sig->sigerr_count++;
530 
531 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
532 			     cq->mcq.cqn, sig->err_item.key,
533 			     sig->err_item.err_type,
534 			     sig->err_item.sig_err_offset,
535 			     sig->err_item.expected,
536 			     sig->err_item.actual);
537 
538 		xa_unlock(&dev->sig_mrs);
539 		goto repoll;
540 	}
541 	}
542 
543 	return 0;
544 }
545 
546 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
547 			struct ib_wc *wc, bool is_fatal_err)
548 {
549 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
550 	struct mlx5_ib_wc *soft_wc, *next;
551 	int npolled = 0;
552 
553 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
554 		if (npolled >= num_entries)
555 			break;
556 
557 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
558 			    cq->mcq.cqn);
559 
560 		if (unlikely(is_fatal_err)) {
561 			soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
562 			soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
563 		}
564 		wc[npolled++] = soft_wc->wc;
565 		list_del(&soft_wc->list);
566 		kfree(soft_wc);
567 	}
568 
569 	return npolled;
570 }
571 
572 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
573 {
574 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
575 	struct mlx5_ib_qp *cur_qp = NULL;
576 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
577 	struct mlx5_core_dev *mdev = dev->mdev;
578 	unsigned long flags;
579 	int soft_polled = 0;
580 	int npolled;
581 
582 	spin_lock_irqsave(&cq->lock, flags);
583 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
584 		/* make sure no soft wqe's are waiting */
585 		if (unlikely(!list_empty(&cq->wc_list)))
586 			soft_polled = poll_soft_wc(cq, num_entries, wc, true);
587 
588 		mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
589 				     wc + soft_polled, &npolled);
590 		goto out;
591 	}
592 
593 	if (unlikely(!list_empty(&cq->wc_list)))
594 		soft_polled = poll_soft_wc(cq, num_entries, wc, false);
595 
596 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
597 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
598 			break;
599 	}
600 
601 	if (npolled)
602 		mlx5_cq_set_ci(&cq->mcq);
603 out:
604 	spin_unlock_irqrestore(&cq->lock, flags);
605 
606 	return soft_polled + npolled;
607 }
608 
609 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
610 {
611 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
612 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
613 	void __iomem *uar_page = mdev->priv.uar->map;
614 	unsigned long irq_flags;
615 	int ret = 0;
616 
617 	spin_lock_irqsave(&cq->lock, irq_flags);
618 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
619 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
620 
621 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
622 		ret = 1;
623 	spin_unlock_irqrestore(&cq->lock, irq_flags);
624 
625 	mlx5_cq_arm(&cq->mcq,
626 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
627 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
628 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
629 
630 	return ret;
631 }
632 
633 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,
634 			     struct mlx5_ib_cq_buf *buf,
635 			     int nent,
636 			     int cqe_size)
637 {
638 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
639 	u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0);
640 	u8 log_wq_sz     = ilog2(cqe_size);
641 	int err;
642 
643 	err = mlx5_frag_buf_alloc_node(dev->mdev,
644 				       nent * cqe_size,
645 				       frag_buf,
646 				       dev->mdev->priv.numa_node);
647 	if (err)
648 		return err;
649 
650 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
651 
652 	buf->cqe_size = cqe_size;
653 	buf->nent = nent;
654 
655 	return 0;
656 }
657 
658 enum {
659 	MLX5_CQE_RES_FORMAT_HASH = 0,
660 	MLX5_CQE_RES_FORMAT_CSUM = 1,
661 	MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3,
662 };
663 
664 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
665 {
666 	switch (format) {
667 	case MLX5_IB_CQE_RES_FORMAT_HASH:
668 		return MLX5_CQE_RES_FORMAT_HASH;
669 	case MLX5_IB_CQE_RES_FORMAT_CSUM:
670 		return MLX5_CQE_RES_FORMAT_CSUM;
671 	case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX:
672 		if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
673 			return MLX5_CQE_RES_FORMAT_CSUM_STRIDX;
674 		return -EOPNOTSUPP;
675 	default:
676 		return -EINVAL;
677 	}
678 }
679 
680 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
681 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
682 			  int *cqe_size, int *index, int *inlen)
683 {
684 	struct mlx5_ib_create_cq ucmd = {};
685 	size_t ucmdlen;
686 	int page_shift;
687 	__be64 *pas;
688 	int npages;
689 	int ncont;
690 	void *cqc;
691 	int err;
692 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
693 		udata, struct mlx5_ib_ucontext, ibucontext);
694 
695 	ucmdlen = udata->inlen < sizeof(ucmd) ?
696 		  (sizeof(ucmd) - sizeof(ucmd.flags)) : sizeof(ucmd);
697 
698 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
699 		return -EFAULT;
700 
701 	if (ucmdlen == sizeof(ucmd) &&
702 	    (ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD)))
703 		return -EINVAL;
704 
705 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
706 		return -EINVAL;
707 
708 	*cqe_size = ucmd.cqe_size;
709 
710 	cq->buf.umem =
711 		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
712 			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
713 	if (IS_ERR(cq->buf.umem)) {
714 		err = PTR_ERR(cq->buf.umem);
715 		return err;
716 	}
717 
718 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &cq->db);
719 	if (err)
720 		goto err_umem;
721 
722 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
723 			   &ncont, NULL);
724 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
725 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
726 
727 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
728 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
729 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
730 	if (!*cqb) {
731 		err = -ENOMEM;
732 		goto err_db;
733 	}
734 
735 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
736 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
737 
738 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
739 	MLX5_SET(cqc, cqc, log_page_size,
740 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
741 
742 	*index = context->bfregi.sys_pages[0];
743 
744 	if (ucmd.cqe_comp_en == 1) {
745 		int mini_cqe_format;
746 
747 		if (!((*cqe_size == 128 &&
748 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
749 		      (*cqe_size == 64  &&
750 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
751 			err = -EOPNOTSUPP;
752 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
753 				     *cqe_size);
754 			goto err_cqb;
755 		}
756 
757 		mini_cqe_format =
758 			mini_cqe_res_format_to_hw(dev,
759 						  ucmd.cqe_comp_res_format);
760 		if (mini_cqe_format < 0) {
761 			err = mini_cqe_format;
762 			mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n",
763 				    ucmd.cqe_comp_res_format, err);
764 			goto err_cqb;
765 		}
766 
767 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
768 		MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format);
769 	}
770 
771 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
772 		if (*cqe_size != 128 ||
773 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
774 			err = -EOPNOTSUPP;
775 			mlx5_ib_warn(dev,
776 				     "CQE padding is not supported for CQE size of %dB!\n",
777 				     *cqe_size);
778 			goto err_cqb;
779 		}
780 
781 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
782 	}
783 
784 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
785 	return 0;
786 
787 err_cqb:
788 	kvfree(*cqb);
789 
790 err_db:
791 	mlx5_ib_db_unmap_user(context, &cq->db);
792 
793 err_umem:
794 	ib_umem_release(cq->buf.umem);
795 	return err;
796 }
797 
798 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
799 {
800 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
801 		udata, struct mlx5_ib_ucontext, ibucontext);
802 
803 	mlx5_ib_db_unmap_user(context, &cq->db);
804 	ib_umem_release(cq->buf.umem);
805 }
806 
807 static void init_cq_frag_buf(struct mlx5_ib_cq *cq,
808 			     struct mlx5_ib_cq_buf *buf)
809 {
810 	int i;
811 	void *cqe;
812 	struct mlx5_cqe64 *cqe64;
813 
814 	for (i = 0; i < buf->nent; i++) {
815 		cqe = get_cqe(cq, i);
816 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
817 		cqe64->op_own = MLX5_CQE_INVALID << 4;
818 	}
819 }
820 
821 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
822 			    int entries, int cqe_size,
823 			    u32 **cqb, int *index, int *inlen)
824 {
825 	__be64 *pas;
826 	void *cqc;
827 	int err;
828 
829 	err = mlx5_db_alloc(dev->mdev, &cq->db);
830 	if (err)
831 		return err;
832 
833 	cq->mcq.set_ci_db  = cq->db.db;
834 	cq->mcq.arm_db     = cq->db.db + 1;
835 	cq->mcq.cqe_sz = cqe_size;
836 
837 	err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size);
838 	if (err)
839 		goto err_db;
840 
841 	init_cq_frag_buf(cq, &cq->buf);
842 
843 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
844 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) *
845 		 cq->buf.frag_buf.npages;
846 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
847 	if (!*cqb) {
848 		err = -ENOMEM;
849 		goto err_buf;
850 	}
851 
852 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
853 	mlx5_fill_page_frag_array(&cq->buf.frag_buf, pas);
854 
855 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
856 	MLX5_SET(cqc, cqc, log_page_size,
857 		 cq->buf.frag_buf.page_shift -
858 		 MLX5_ADAPTER_PAGE_SHIFT);
859 
860 	*index = dev->mdev->priv.uar->index;
861 
862 	return 0;
863 
864 err_buf:
865 	free_cq_buf(dev, &cq->buf);
866 
867 err_db:
868 	mlx5_db_free(dev->mdev, &cq->db);
869 	return err;
870 }
871 
872 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
873 {
874 	free_cq_buf(dev, &cq->buf);
875 	mlx5_db_free(dev->mdev, &cq->db);
876 }
877 
878 static void notify_soft_wc_handler(struct work_struct *work)
879 {
880 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
881 					     notify_work);
882 
883 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
884 }
885 
886 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
887 		      struct ib_udata *udata)
888 {
889 	struct ib_device *ibdev = ibcq->device;
890 	int entries = attr->cqe;
891 	int vector = attr->comp_vector;
892 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
893 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
894 	u32 out[MLX5_ST_SZ_DW(create_cq_out)];
895 	int uninitialized_var(index);
896 	int uninitialized_var(inlen);
897 	u32 *cqb = NULL;
898 	void *cqc;
899 	int cqe_size;
900 	unsigned int irqn;
901 	int eqn;
902 	int err;
903 
904 	if (entries < 0 ||
905 	    (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))))
906 		return -EINVAL;
907 
908 	if (check_cq_create_flags(attr->flags))
909 		return -EOPNOTSUPP;
910 
911 	entries = roundup_pow_of_two(entries + 1);
912 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))
913 		return -EINVAL;
914 
915 	cq->ibcq.cqe = entries - 1;
916 	mutex_init(&cq->resize_mutex);
917 	spin_lock_init(&cq->lock);
918 	cq->resize_buf = NULL;
919 	cq->resize_umem = NULL;
920 	cq->create_flags = attr->flags;
921 	INIT_LIST_HEAD(&cq->list_send_qp);
922 	INIT_LIST_HEAD(&cq->list_recv_qp);
923 
924 	if (udata) {
925 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
926 				     &index, &inlen);
927 		if (err)
928 			return err;
929 	} else {
930 		cqe_size = cache_line_size() == 128 ? 128 : 64;
931 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
932 				       &index, &inlen);
933 		if (err)
934 			return err;
935 
936 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
937 	}
938 
939 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
940 	if (err)
941 		goto err_cqb;
942 
943 	cq->cqe_size = cqe_size;
944 
945 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
946 	MLX5_SET(cqc, cqc, cqe_sz,
947 		 cqe_sz_to_mlx_sz(cqe_size,
948 				  cq->private_flags &
949 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
950 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
951 	MLX5_SET(cqc, cqc, uar_page, index);
952 	MLX5_SET(cqc, cqc, c_eqn, eqn);
953 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
954 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
955 		MLX5_SET(cqc, cqc, oi, 1);
956 
957 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out));
958 	if (err)
959 		goto err_cqb;
960 
961 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
962 	cq->mcq.irqn = irqn;
963 	if (udata)
964 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
965 	else
966 		cq->mcq.comp  = mlx5_ib_cq_comp;
967 	cq->mcq.event = mlx5_ib_cq_event;
968 
969 	INIT_LIST_HEAD(&cq->wc_list);
970 
971 	if (udata)
972 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
973 			err = -EFAULT;
974 			goto err_cmd;
975 		}
976 
977 
978 	kvfree(cqb);
979 	return 0;
980 
981 err_cmd:
982 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
983 
984 err_cqb:
985 	kvfree(cqb);
986 	if (udata)
987 		destroy_cq_user(cq, udata);
988 	else
989 		destroy_cq_kernel(dev, cq);
990 	return err;
991 }
992 
993 void mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
994 {
995 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
996 	struct mlx5_ib_cq *mcq = to_mcq(cq);
997 
998 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
999 	if (udata)
1000 		destroy_cq_user(mcq, udata);
1001 	else
1002 		destroy_cq_kernel(dev, mcq);
1003 }
1004 
1005 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1006 {
1007 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1008 }
1009 
1010 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1011 {
1012 	struct mlx5_cqe64 *cqe64, *dest64;
1013 	void *cqe, *dest;
1014 	u32 prod_index;
1015 	int nfreed = 0;
1016 	u8 owner_bit;
1017 
1018 	if (!cq)
1019 		return;
1020 
1021 	/* First we need to find the current producer index, so we
1022 	 * know where to start cleaning from.  It doesn't matter if HW
1023 	 * adds new entries after this loop -- the QP we're worried
1024 	 * about is already in RESET, so the new entries won't come
1025 	 * from our QP and therefore don't need to be checked.
1026 	 */
1027 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1028 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1029 			break;
1030 
1031 	/* Now sweep backwards through the CQ, removing CQ entries
1032 	 * that match our QP by copying older entries on top of them.
1033 	 */
1034 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1035 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1036 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1037 		if (is_equal_rsn(cqe64, rsn)) {
1038 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1039 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1040 			++nfreed;
1041 		} else if (nfreed) {
1042 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1043 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1044 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1045 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1046 			dest64->op_own = owner_bit |
1047 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1048 		}
1049 	}
1050 
1051 	if (nfreed) {
1052 		cq->mcq.cons_index += nfreed;
1053 		/* Make sure update of buffer contents is done before
1054 		 * updating consumer index.
1055 		 */
1056 		wmb();
1057 		mlx5_cq_set_ci(&cq->mcq);
1058 	}
1059 }
1060 
1061 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1062 {
1063 	if (!cq)
1064 		return;
1065 
1066 	spin_lock_irq(&cq->lock);
1067 	__mlx5_ib_cq_clean(cq, qpn, srq);
1068 	spin_unlock_irq(&cq->lock);
1069 }
1070 
1071 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1072 {
1073 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1074 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1075 	int err;
1076 
1077 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1078 		return -EOPNOTSUPP;
1079 
1080 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1081 		return -EINVAL;
1082 
1083 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1084 					     cq_period, cq_count);
1085 	if (err)
1086 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1087 
1088 	return err;
1089 }
1090 
1091 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1092 		       int entries, struct ib_udata *udata, int *npas,
1093 		       int *page_shift, int *cqe_size)
1094 {
1095 	struct mlx5_ib_resize_cq ucmd;
1096 	struct ib_umem *umem;
1097 	int err;
1098 	int npages;
1099 
1100 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1101 	if (err)
1102 		return err;
1103 
1104 	if (ucmd.reserved0 || ucmd.reserved1)
1105 		return -EINVAL;
1106 
1107 	/* check multiplication overflow */
1108 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1109 		return -EINVAL;
1110 
1111 	umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
1112 			   (size_t)ucmd.cqe_size * entries,
1113 			   IB_ACCESS_LOCAL_WRITE);
1114 	if (IS_ERR(umem)) {
1115 		err = PTR_ERR(umem);
1116 		return err;
1117 	}
1118 
1119 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1120 			   npas, NULL);
1121 
1122 	cq->resize_umem = umem;
1123 	*cqe_size = ucmd.cqe_size;
1124 
1125 	return 0;
1126 }
1127 
1128 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1129 			 int entries, int cqe_size)
1130 {
1131 	int err;
1132 
1133 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1134 	if (!cq->resize_buf)
1135 		return -ENOMEM;
1136 
1137 	err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size);
1138 	if (err)
1139 		goto ex;
1140 
1141 	init_cq_frag_buf(cq, cq->resize_buf);
1142 
1143 	return 0;
1144 
1145 ex:
1146 	kfree(cq->resize_buf);
1147 	return err;
1148 }
1149 
1150 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1151 {
1152 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1153 	struct mlx5_cqe64 *scqe64;
1154 	struct mlx5_cqe64 *dcqe64;
1155 	void *start_cqe;
1156 	void *scqe;
1157 	void *dcqe;
1158 	int ssize;
1159 	int dsize;
1160 	int i;
1161 	u8 sw_own;
1162 
1163 	ssize = cq->buf.cqe_size;
1164 	dsize = cq->resize_buf->cqe_size;
1165 	if (ssize != dsize) {
1166 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1167 		return -EINVAL;
1168 	}
1169 
1170 	i = cq->mcq.cons_index;
1171 	scqe = get_sw_cqe(cq, i);
1172 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1173 	start_cqe = scqe;
1174 	if (!scqe) {
1175 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1176 		return -EINVAL;
1177 	}
1178 
1179 	while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) {
1180 		dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc,
1181 					     (i + 1) & cq->resize_buf->nent);
1182 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1183 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1184 		memcpy(dcqe, scqe, dsize);
1185 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1186 
1187 		++i;
1188 		scqe = get_sw_cqe(cq, i);
1189 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1190 		if (!scqe) {
1191 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1192 			return -EINVAL;
1193 		}
1194 
1195 		if (scqe == start_cqe) {
1196 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1197 				cq->mcq.cqn);
1198 			return -ENOMEM;
1199 		}
1200 	}
1201 	++cq->mcq.cons_index;
1202 	return 0;
1203 }
1204 
1205 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1206 {
1207 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1208 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1209 	void *cqc;
1210 	u32 *in;
1211 	int err;
1212 	int npas;
1213 	__be64 *pas;
1214 	int page_shift;
1215 	int inlen;
1216 	int uninitialized_var(cqe_size);
1217 	unsigned long flags;
1218 
1219 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1220 		pr_info("Firmware does not support resize CQ\n");
1221 		return -ENOSYS;
1222 	}
1223 
1224 	if (entries < 1 ||
1225 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1226 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1227 			     entries,
1228 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1229 		return -EINVAL;
1230 	}
1231 
1232 	entries = roundup_pow_of_two(entries + 1);
1233 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1234 		return -EINVAL;
1235 
1236 	if (entries == ibcq->cqe + 1)
1237 		return 0;
1238 
1239 	mutex_lock(&cq->resize_mutex);
1240 	if (udata) {
1241 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1242 				  &cqe_size);
1243 	} else {
1244 		cqe_size = 64;
1245 		err = resize_kernel(dev, cq, entries, cqe_size);
1246 		if (!err) {
1247 			struct mlx5_frag_buf *frag_buf = &cq->resize_buf->frag_buf;
1248 
1249 			npas = frag_buf->npages;
1250 			page_shift = frag_buf->page_shift;
1251 		}
1252 	}
1253 
1254 	if (err)
1255 		goto ex;
1256 
1257 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1258 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1259 
1260 	in = kvzalloc(inlen, GFP_KERNEL);
1261 	if (!in) {
1262 		err = -ENOMEM;
1263 		goto ex_resize;
1264 	}
1265 
1266 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1267 	if (udata)
1268 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1269 				     pas, 0);
1270 	else
1271 		mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas);
1272 
1273 	MLX5_SET(modify_cq_in, in,
1274 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1275 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1276 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1277 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1278 
1279 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1280 
1281 	MLX5_SET(cqc, cqc, log_page_size,
1282 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1283 	MLX5_SET(cqc, cqc, cqe_sz,
1284 		 cqe_sz_to_mlx_sz(cqe_size,
1285 				  cq->private_flags &
1286 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1287 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1288 
1289 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1290 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1291 
1292 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1293 	if (err)
1294 		goto ex_alloc;
1295 
1296 	if (udata) {
1297 		cq->ibcq.cqe = entries - 1;
1298 		ib_umem_release(cq->buf.umem);
1299 		cq->buf.umem = cq->resize_umem;
1300 		cq->resize_umem = NULL;
1301 	} else {
1302 		struct mlx5_ib_cq_buf tbuf;
1303 		int resized = 0;
1304 
1305 		spin_lock_irqsave(&cq->lock, flags);
1306 		if (cq->resize_buf) {
1307 			err = copy_resize_cqes(cq);
1308 			if (!err) {
1309 				tbuf = cq->buf;
1310 				cq->buf = *cq->resize_buf;
1311 				kfree(cq->resize_buf);
1312 				cq->resize_buf = NULL;
1313 				resized = 1;
1314 			}
1315 		}
1316 		cq->ibcq.cqe = entries - 1;
1317 		spin_unlock_irqrestore(&cq->lock, flags);
1318 		if (resized)
1319 			free_cq_buf(dev, &tbuf);
1320 	}
1321 	mutex_unlock(&cq->resize_mutex);
1322 
1323 	kvfree(in);
1324 	return 0;
1325 
1326 ex_alloc:
1327 	kvfree(in);
1328 
1329 ex_resize:
1330 	ib_umem_release(cq->resize_umem);
1331 	if (!udata) {
1332 		free_cq_buf(dev, cq->resize_buf);
1333 		cq->resize_buf = NULL;
1334 	}
1335 ex:
1336 	mutex_unlock(&cq->resize_mutex);
1337 	return err;
1338 }
1339 
1340 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq)
1341 {
1342 	struct mlx5_ib_cq *cq;
1343 
1344 	if (!ibcq)
1345 		return 128;
1346 
1347 	cq = to_mcq(ibcq);
1348 	return cq->cqe_size;
1349 }
1350 
1351 /* Called from atomic context */
1352 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1353 {
1354 	struct mlx5_ib_wc *soft_wc;
1355 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1356 	unsigned long flags;
1357 
1358 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1359 	if (!soft_wc)
1360 		return -ENOMEM;
1361 
1362 	soft_wc->wc = *wc;
1363 	spin_lock_irqsave(&cq->lock, flags);
1364 	list_add_tail(&soft_wc->list, &cq->wc_list);
1365 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1366 	    wc->status != IB_WC_SUCCESS) {
1367 		cq->notify_flags = 0;
1368 		schedule_work(&cq->notify_work);
1369 	}
1370 	spin_unlock_irqrestore(&cq->lock, flags);
1371 
1372 	return 0;
1373 }
1374