xref: /openbmc/linux/drivers/infiniband/hw/mlx5/cq.c (revision 752beb5e)
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)
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 	struct mlx5_sig_err_cqe *sig_err_cqe;
427 	struct mlx5_core_mkey *mmkey;
428 	struct mlx5_ib_mr *mr;
429 	uint8_t opcode;
430 	uint32_t qpn;
431 	u16 wqe_ctr;
432 	void *cqe;
433 	int idx;
434 
435 repoll:
436 	cqe = next_cqe_sw(cq);
437 	if (!cqe)
438 		return -EAGAIN;
439 
440 	cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
441 
442 	++cq->mcq.cons_index;
443 
444 	/* Make sure we read CQ entry contents after we've checked the
445 	 * ownership bit.
446 	 */
447 	rmb();
448 
449 	opcode = get_cqe_opcode(cqe64);
450 	if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) {
451 		if (likely(cq->resize_buf)) {
452 			free_cq_buf(dev, &cq->buf);
453 			cq->buf = *cq->resize_buf;
454 			kfree(cq->resize_buf);
455 			cq->resize_buf = NULL;
456 			goto repoll;
457 		} else {
458 			mlx5_ib_warn(dev, "unexpected resize cqe\n");
459 		}
460 	}
461 
462 	qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff;
463 	if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) {
464 		/* We do not have to take the QP table lock here,
465 		 * because CQs will be locked while QPs are removed
466 		 * from the table.
467 		 */
468 		mqp = __mlx5_qp_lookup(dev->mdev, qpn);
469 		*cur_qp = to_mibqp(mqp);
470 	}
471 
472 	wc->qp  = &(*cur_qp)->ibqp;
473 	switch (opcode) {
474 	case MLX5_CQE_REQ:
475 		wq = &(*cur_qp)->sq;
476 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
477 		idx = wqe_ctr & (wq->wqe_cnt - 1);
478 		handle_good_req(wc, cqe64, wq, idx);
479 		wc->wr_id = wq->wrid[idx];
480 		wq->tail = wq->wqe_head[idx] + 1;
481 		wc->status = IB_WC_SUCCESS;
482 		break;
483 	case MLX5_CQE_RESP_WR_IMM:
484 	case MLX5_CQE_RESP_SEND:
485 	case MLX5_CQE_RESP_SEND_IMM:
486 	case MLX5_CQE_RESP_SEND_INV:
487 		handle_responder(wc, cqe64, *cur_qp);
488 		wc->status = IB_WC_SUCCESS;
489 		break;
490 	case MLX5_CQE_RESIZE_CQ:
491 		break;
492 	case MLX5_CQE_REQ_ERR:
493 	case MLX5_CQE_RESP_ERR:
494 		err_cqe = (struct mlx5_err_cqe *)cqe64;
495 		mlx5_handle_error_cqe(dev, err_cqe, wc);
496 		mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n",
497 			    opcode == MLX5_CQE_REQ_ERR ?
498 			    "Requestor" : "Responder", cq->mcq.cqn);
499 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
500 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
501 		if (opcode == MLX5_CQE_REQ_ERR) {
502 			wq = &(*cur_qp)->sq;
503 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
504 			idx = wqe_ctr & (wq->wqe_cnt - 1);
505 			wc->wr_id = wq->wrid[idx];
506 			wq->tail = wq->wqe_head[idx] + 1;
507 		} else {
508 			struct mlx5_ib_srq *srq;
509 
510 			if ((*cur_qp)->ibqp.srq) {
511 				srq = to_msrq((*cur_qp)->ibqp.srq);
512 				wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
513 				wc->wr_id = srq->wrid[wqe_ctr];
514 				mlx5_ib_free_srq_wqe(srq, wqe_ctr);
515 			} else {
516 				wq = &(*cur_qp)->rq;
517 				wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
518 				++wq->tail;
519 			}
520 		}
521 		break;
522 	case MLX5_CQE_SIG_ERR:
523 		sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64;
524 
525 		read_lock(&dev->mdev->priv.mkey_table.lock);
526 		mmkey = __mlx5_mr_lookup(dev->mdev,
527 					 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey)));
528 		mr = to_mibmr(mmkey);
529 		get_sig_err_item(sig_err_cqe, &mr->sig->err_item);
530 		mr->sig->sig_err_exists = true;
531 		mr->sig->sigerr_count++;
532 
533 		mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n",
534 			     cq->mcq.cqn, mr->sig->err_item.key,
535 			     mr->sig->err_item.err_type,
536 			     mr->sig->err_item.sig_err_offset,
537 			     mr->sig->err_item.expected,
538 			     mr->sig->err_item.actual);
539 
540 		read_unlock(&dev->mdev->priv.mkey_table.lock);
541 		goto repoll;
542 	}
543 
544 	return 0;
545 }
546 
547 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries,
548 			struct ib_wc *wc, bool is_fatal_err)
549 {
550 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
551 	struct mlx5_ib_wc *soft_wc, *next;
552 	int npolled = 0;
553 
554 	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
555 		if (npolled >= num_entries)
556 			break;
557 
558 		mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n",
559 			    cq->mcq.cqn);
560 
561 		if (unlikely(is_fatal_err)) {
562 			soft_wc->wc.status = IB_WC_WR_FLUSH_ERR;
563 			soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
564 		}
565 		wc[npolled++] = soft_wc->wc;
566 		list_del(&soft_wc->list);
567 		kfree(soft_wc);
568 	}
569 
570 	return npolled;
571 }
572 
573 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
574 {
575 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
576 	struct mlx5_ib_qp *cur_qp = NULL;
577 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
578 	struct mlx5_core_dev *mdev = dev->mdev;
579 	unsigned long flags;
580 	int soft_polled = 0;
581 	int npolled;
582 
583 	spin_lock_irqsave(&cq->lock, flags);
584 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
585 		/* make sure no soft wqe's are waiting */
586 		if (unlikely(!list_empty(&cq->wc_list)))
587 			soft_polled = poll_soft_wc(cq, num_entries, wc, true);
588 
589 		mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled,
590 				     wc + soft_polled, &npolled);
591 		goto out;
592 	}
593 
594 	if (unlikely(!list_empty(&cq->wc_list)))
595 		soft_polled = poll_soft_wc(cq, num_entries, wc, false);
596 
597 	for (npolled = 0; npolled < num_entries - soft_polled; npolled++) {
598 		if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled))
599 			break;
600 	}
601 
602 	if (npolled)
603 		mlx5_cq_set_ci(&cq->mcq);
604 out:
605 	spin_unlock_irqrestore(&cq->lock, flags);
606 
607 	return soft_polled + npolled;
608 }
609 
610 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
611 {
612 	struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev;
613 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
614 	void __iomem *uar_page = mdev->priv.uar->map;
615 	unsigned long irq_flags;
616 	int ret = 0;
617 
618 	spin_lock_irqsave(&cq->lock, irq_flags);
619 	if (cq->notify_flags != IB_CQ_NEXT_COMP)
620 		cq->notify_flags = flags & IB_CQ_SOLICITED_MASK;
621 
622 	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list))
623 		ret = 1;
624 	spin_unlock_irqrestore(&cq->lock, irq_flags);
625 
626 	mlx5_cq_arm(&cq->mcq,
627 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
628 		    MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT,
629 		    uar_page, to_mcq(ibcq)->mcq.cons_index);
630 
631 	return ret;
632 }
633 
634 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,
635 			     struct mlx5_ib_cq_buf *buf,
636 			     int nent,
637 			     int cqe_size)
638 {
639 	struct mlx5_frag_buf *frag_buf = &buf->frag_buf;
640 	u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0);
641 	u8 log_wq_sz     = ilog2(cqe_size);
642 	int err;
643 
644 	err = mlx5_frag_buf_alloc_node(dev->mdev,
645 				       nent * cqe_size,
646 				       frag_buf,
647 				       dev->mdev->priv.numa_node);
648 	if (err)
649 		return err;
650 
651 	mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc);
652 
653 	buf->cqe_size = cqe_size;
654 	buf->nent = nent;
655 
656 	return 0;
657 }
658 
659 enum {
660 	MLX5_CQE_RES_FORMAT_HASH = 0,
661 	MLX5_CQE_RES_FORMAT_CSUM = 1,
662 	MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3,
663 };
664 
665 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
666 {
667 	switch (format) {
668 	case MLX5_IB_CQE_RES_FORMAT_HASH:
669 		return MLX5_CQE_RES_FORMAT_HASH;
670 	case MLX5_IB_CQE_RES_FORMAT_CSUM:
671 		return MLX5_CQE_RES_FORMAT_CSUM;
672 	case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX:
673 		if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
674 			return MLX5_CQE_RES_FORMAT_CSUM_STRIDX;
675 		return -EOPNOTSUPP;
676 	default:
677 		return -EINVAL;
678 	}
679 }
680 
681 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
682 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
683 			  int *cqe_size, int *index, int *inlen)
684 {
685 	struct mlx5_ib_create_cq ucmd = {};
686 	size_t ucmdlen;
687 	int page_shift;
688 	__be64 *pas;
689 	int npages;
690 	int ncont;
691 	void *cqc;
692 	int err;
693 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
694 		udata, struct mlx5_ib_ucontext, ibucontext);
695 
696 	ucmdlen = udata->inlen < sizeof(ucmd) ?
697 		  (sizeof(ucmd) - sizeof(ucmd.flags)) : sizeof(ucmd);
698 
699 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen))
700 		return -EFAULT;
701 
702 	if (ucmdlen == sizeof(ucmd) &&
703 	    (ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD)))
704 		return -EINVAL;
705 
706 	if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128)
707 		return -EINVAL;
708 
709 	*cqe_size = ucmd.cqe_size;
710 
711 	cq->buf.umem =
712 		ib_umem_get(udata, ucmd.buf_addr, entries * ucmd.cqe_size,
713 			    IB_ACCESS_LOCAL_WRITE, 1);
714 	if (IS_ERR(cq->buf.umem)) {
715 		err = PTR_ERR(cq->buf.umem);
716 		return err;
717 	}
718 
719 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &cq->db);
720 	if (err)
721 		goto err_umem;
722 
723 	mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift,
724 			   &ncont, NULL);
725 	mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n",
726 		    ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont);
727 
728 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
729 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont;
730 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
731 	if (!*cqb) {
732 		err = -ENOMEM;
733 		goto err_db;
734 	}
735 
736 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
737 	mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0);
738 
739 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
740 	MLX5_SET(cqc, cqc, log_page_size,
741 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
742 
743 	*index = context->bfregi.sys_pages[0];
744 
745 	if (ucmd.cqe_comp_en == 1) {
746 		int mini_cqe_format;
747 
748 		if (!((*cqe_size == 128 &&
749 		       MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) ||
750 		      (*cqe_size == 64  &&
751 		       MLX5_CAP_GEN(dev->mdev, cqe_compression)))) {
752 			err = -EOPNOTSUPP;
753 			mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n",
754 				     *cqe_size);
755 			goto err_cqb;
756 		}
757 
758 		mini_cqe_format =
759 			mini_cqe_res_format_to_hw(dev,
760 						  ucmd.cqe_comp_res_format);
761 		if (mini_cqe_format < 0) {
762 			err = mini_cqe_format;
763 			mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n",
764 				    ucmd.cqe_comp_res_format, err);
765 			goto err_cqb;
766 		}
767 
768 		MLX5_SET(cqc, cqc, cqe_comp_en, 1);
769 		MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format);
770 	}
771 
772 	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) {
773 		if (*cqe_size != 128 ||
774 		    !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) {
775 			err = -EOPNOTSUPP;
776 			mlx5_ib_warn(dev,
777 				     "CQE padding is not supported for CQE size of %dB!\n",
778 				     *cqe_size);
779 			goto err_cqb;
780 		}
781 
782 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
783 	}
784 
785 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
786 	return 0;
787 
788 err_cqb:
789 	kvfree(*cqb);
790 
791 err_db:
792 	mlx5_ib_db_unmap_user(context, &cq->db);
793 
794 err_umem:
795 	ib_umem_release(cq->buf.umem);
796 	return err;
797 }
798 
799 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata)
800 {
801 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
802 		udata, struct mlx5_ib_ucontext, ibucontext);
803 
804 	mlx5_ib_db_unmap_user(context, &cq->db);
805 	ib_umem_release(cq->buf.umem);
806 }
807 
808 static void init_cq_frag_buf(struct mlx5_ib_cq *cq,
809 			     struct mlx5_ib_cq_buf *buf)
810 {
811 	int i;
812 	void *cqe;
813 	struct mlx5_cqe64 *cqe64;
814 
815 	for (i = 0; i < buf->nent; i++) {
816 		cqe = get_cqe(cq, i);
817 		cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64;
818 		cqe64->op_own = MLX5_CQE_INVALID << 4;
819 	}
820 }
821 
822 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
823 			    int entries, int cqe_size,
824 			    u32 **cqb, int *index, int *inlen)
825 {
826 	__be64 *pas;
827 	void *cqc;
828 	int err;
829 
830 	err = mlx5_db_alloc(dev->mdev, &cq->db);
831 	if (err)
832 		return err;
833 
834 	cq->mcq.set_ci_db  = cq->db.db;
835 	cq->mcq.arm_db     = cq->db.db + 1;
836 	cq->mcq.cqe_sz = cqe_size;
837 
838 	err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size);
839 	if (err)
840 		goto err_db;
841 
842 	init_cq_frag_buf(cq, &cq->buf);
843 
844 	*inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
845 		 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) *
846 		 cq->buf.frag_buf.npages;
847 	*cqb = kvzalloc(*inlen, GFP_KERNEL);
848 	if (!*cqb) {
849 		err = -ENOMEM;
850 		goto err_buf;
851 	}
852 
853 	pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas);
854 	mlx5_fill_page_frag_array(&cq->buf.frag_buf, pas);
855 
856 	cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context);
857 	MLX5_SET(cqc, cqc, log_page_size,
858 		 cq->buf.frag_buf.page_shift -
859 		 MLX5_ADAPTER_PAGE_SHIFT);
860 
861 	*index = dev->mdev->priv.uar->index;
862 
863 	return 0;
864 
865 err_buf:
866 	free_cq_buf(dev, &cq->buf);
867 
868 err_db:
869 	mlx5_db_free(dev->mdev, &cq->db);
870 	return err;
871 }
872 
873 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
874 {
875 	free_cq_buf(dev, &cq->buf);
876 	mlx5_db_free(dev->mdev, &cq->db);
877 }
878 
879 static void notify_soft_wc_handler(struct work_struct *work)
880 {
881 	struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq,
882 					     notify_work);
883 
884 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
885 }
886 
887 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
888 				const struct ib_cq_init_attr *attr,
889 				struct ib_udata *udata)
890 {
891 	int entries = attr->cqe;
892 	int vector = attr->comp_vector;
893 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
894 	struct mlx5_ib_cq *cq;
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 ERR_PTR(-EINVAL);
907 
908 	if (check_cq_create_flags(attr->flags))
909 		return ERR_PTR(-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 ERR_PTR(-EINVAL);
914 
915 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
916 	if (!cq)
917 		return ERR_PTR(-ENOMEM);
918 
919 	cq->ibcq.cqe = entries - 1;
920 	mutex_init(&cq->resize_mutex);
921 	spin_lock_init(&cq->lock);
922 	cq->resize_buf = NULL;
923 	cq->resize_umem = NULL;
924 	cq->create_flags = attr->flags;
925 	INIT_LIST_HEAD(&cq->list_send_qp);
926 	INIT_LIST_HEAD(&cq->list_recv_qp);
927 
928 	if (udata) {
929 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
930 				     &index, &inlen);
931 		if (err)
932 			goto err_create;
933 	} else {
934 		cqe_size = cache_line_size() == 128 ? 128 : 64;
935 		err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb,
936 				       &index, &inlen);
937 		if (err)
938 			goto err_create;
939 
940 		INIT_WORK(&cq->notify_work, notify_soft_wc_handler);
941 	}
942 
943 	err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn);
944 	if (err)
945 		goto err_cqb;
946 
947 	cq->cqe_size = cqe_size;
948 
949 	cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context);
950 	MLX5_SET(cqc, cqc, cqe_sz,
951 		 cqe_sz_to_mlx_sz(cqe_size,
952 				  cq->private_flags &
953 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
954 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
955 	MLX5_SET(cqc, cqc, uar_page, index);
956 	MLX5_SET(cqc, cqc, c_eqn, eqn);
957 	MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma);
958 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN)
959 		MLX5_SET(cqc, cqc, oi, 1);
960 
961 	err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen);
962 	if (err)
963 		goto err_cqb;
964 
965 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
966 	cq->mcq.irqn = irqn;
967 	if (udata)
968 		cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp;
969 	else
970 		cq->mcq.comp  = mlx5_ib_cq_comp;
971 	cq->mcq.event = mlx5_ib_cq_event;
972 
973 	INIT_LIST_HEAD(&cq->wc_list);
974 
975 	if (udata)
976 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
977 			err = -EFAULT;
978 			goto err_cmd;
979 		}
980 
981 
982 	kvfree(cqb);
983 	return &cq->ibcq;
984 
985 err_cmd:
986 	mlx5_core_destroy_cq(dev->mdev, &cq->mcq);
987 
988 err_cqb:
989 	kvfree(cqb);
990 	if (udata)
991 		destroy_cq_user(cq, udata);
992 	else
993 		destroy_cq_kernel(dev, cq);
994 
995 err_create:
996 	kfree(cq);
997 
998 	return ERR_PTR(err);
999 }
1000 
1001 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
1002 {
1003 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1004 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1005 
1006 	mlx5_core_destroy_cq(dev->mdev, &mcq->mcq);
1007 	if (udata)
1008 		destroy_cq_user(mcq, udata);
1009 	else
1010 		destroy_cq_kernel(dev, mcq);
1011 
1012 	kfree(mcq);
1013 
1014 	return 0;
1015 }
1016 
1017 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn)
1018 {
1019 	return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff);
1020 }
1021 
1022 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq)
1023 {
1024 	struct mlx5_cqe64 *cqe64, *dest64;
1025 	void *cqe, *dest;
1026 	u32 prod_index;
1027 	int nfreed = 0;
1028 	u8 owner_bit;
1029 
1030 	if (!cq)
1031 		return;
1032 
1033 	/* First we need to find the current producer index, so we
1034 	 * know where to start cleaning from.  It doesn't matter if HW
1035 	 * adds new entries after this loop -- the QP we're worried
1036 	 * about is already in RESET, so the new entries won't come
1037 	 * from our QP and therefore don't need to be checked.
1038 	 */
1039 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++)
1040 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
1041 			break;
1042 
1043 	/* Now sweep backwards through the CQ, removing CQ entries
1044 	 * that match our QP by copying older entries on top of them.
1045 	 */
1046 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
1047 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
1048 		cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64;
1049 		if (is_equal_rsn(cqe64, rsn)) {
1050 			if (srq && (ntohl(cqe64->srqn) & 0xffffff))
1051 				mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter));
1052 			++nfreed;
1053 		} else if (nfreed) {
1054 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
1055 			dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64;
1056 			owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK;
1057 			memcpy(dest, cqe, cq->mcq.cqe_sz);
1058 			dest64->op_own = owner_bit |
1059 				(dest64->op_own & ~MLX5_CQE_OWNER_MASK);
1060 		}
1061 	}
1062 
1063 	if (nfreed) {
1064 		cq->mcq.cons_index += nfreed;
1065 		/* Make sure update of buffer contents is done before
1066 		 * updating consumer index.
1067 		 */
1068 		wmb();
1069 		mlx5_cq_set_ci(&cq->mcq);
1070 	}
1071 }
1072 
1073 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq)
1074 {
1075 	if (!cq)
1076 		return;
1077 
1078 	spin_lock_irq(&cq->lock);
1079 	__mlx5_ib_cq_clean(cq, qpn, srq);
1080 	spin_unlock_irq(&cq->lock);
1081 }
1082 
1083 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1084 {
1085 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
1086 	struct mlx5_ib_cq *mcq = to_mcq(cq);
1087 	int err;
1088 
1089 	if (!MLX5_CAP_GEN(dev->mdev, cq_moderation))
1090 		return -EOPNOTSUPP;
1091 
1092 	if (cq_period > MLX5_MAX_CQ_PERIOD)
1093 		return -EINVAL;
1094 
1095 	err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq,
1096 					     cq_period, cq_count);
1097 	if (err)
1098 		mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn);
1099 
1100 	return err;
1101 }
1102 
1103 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1104 		       int entries, struct ib_udata *udata, int *npas,
1105 		       int *page_shift, int *cqe_size)
1106 {
1107 	struct mlx5_ib_resize_cq ucmd;
1108 	struct ib_umem *umem;
1109 	int err;
1110 	int npages;
1111 
1112 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
1113 	if (err)
1114 		return err;
1115 
1116 	if (ucmd.reserved0 || ucmd.reserved1)
1117 		return -EINVAL;
1118 
1119 	/* check multiplication overflow */
1120 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
1121 		return -EINVAL;
1122 
1123 	umem = ib_umem_get(udata, ucmd.buf_addr,
1124 			   (size_t)ucmd.cqe_size * entries,
1125 			   IB_ACCESS_LOCAL_WRITE, 1);
1126 	if (IS_ERR(umem)) {
1127 		err = PTR_ERR(umem);
1128 		return err;
1129 	}
1130 
1131 	mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift,
1132 			   npas, NULL);
1133 
1134 	cq->resize_umem = umem;
1135 	*cqe_size = ucmd.cqe_size;
1136 
1137 	return 0;
1138 }
1139 
1140 static void un_resize_user(struct mlx5_ib_cq *cq)
1141 {
1142 	ib_umem_release(cq->resize_umem);
1143 }
1144 
1145 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1146 			 int entries, int cqe_size)
1147 {
1148 	int err;
1149 
1150 	cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL);
1151 	if (!cq->resize_buf)
1152 		return -ENOMEM;
1153 
1154 	err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size);
1155 	if (err)
1156 		goto ex;
1157 
1158 	init_cq_frag_buf(cq, cq->resize_buf);
1159 
1160 	return 0;
1161 
1162 ex:
1163 	kfree(cq->resize_buf);
1164 	return err;
1165 }
1166 
1167 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
1168 {
1169 	free_cq_buf(dev, cq->resize_buf);
1170 	cq->resize_buf = NULL;
1171 }
1172 
1173 static int copy_resize_cqes(struct mlx5_ib_cq *cq)
1174 {
1175 	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
1176 	struct mlx5_cqe64 *scqe64;
1177 	struct mlx5_cqe64 *dcqe64;
1178 	void *start_cqe;
1179 	void *scqe;
1180 	void *dcqe;
1181 	int ssize;
1182 	int dsize;
1183 	int i;
1184 	u8 sw_own;
1185 
1186 	ssize = cq->buf.cqe_size;
1187 	dsize = cq->resize_buf->cqe_size;
1188 	if (ssize != dsize) {
1189 		mlx5_ib_warn(dev, "resize from different cqe size is not supported\n");
1190 		return -EINVAL;
1191 	}
1192 
1193 	i = cq->mcq.cons_index;
1194 	scqe = get_sw_cqe(cq, i);
1195 	scqe64 = ssize == 64 ? scqe : scqe + 64;
1196 	start_cqe = scqe;
1197 	if (!scqe) {
1198 		mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1199 		return -EINVAL;
1200 	}
1201 
1202 	while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) {
1203 		dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc,
1204 					     (i + 1) & cq->resize_buf->nent);
1205 		dcqe64 = dsize == 64 ? dcqe : dcqe + 64;
1206 		sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent);
1207 		memcpy(dcqe, scqe, dsize);
1208 		dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own;
1209 
1210 		++i;
1211 		scqe = get_sw_cqe(cq, i);
1212 		scqe64 = ssize == 64 ? scqe : scqe + 64;
1213 		if (!scqe) {
1214 			mlx5_ib_warn(dev, "expected cqe in sw ownership\n");
1215 			return -EINVAL;
1216 		}
1217 
1218 		if (scqe == start_cqe) {
1219 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n",
1220 				cq->mcq.cqn);
1221 			return -ENOMEM;
1222 		}
1223 	}
1224 	++cq->mcq.cons_index;
1225 	return 0;
1226 }
1227 
1228 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
1229 {
1230 	struct mlx5_ib_dev *dev = to_mdev(ibcq->device);
1231 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1232 	void *cqc;
1233 	u32 *in;
1234 	int err;
1235 	int npas;
1236 	__be64 *pas;
1237 	int page_shift;
1238 	int inlen;
1239 	int uninitialized_var(cqe_size);
1240 	unsigned long flags;
1241 
1242 	if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) {
1243 		pr_info("Firmware does not support resize CQ\n");
1244 		return -ENOSYS;
1245 	}
1246 
1247 	if (entries < 1 ||
1248 	    entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) {
1249 		mlx5_ib_warn(dev, "wrong entries number %d, max %d\n",
1250 			     entries,
1251 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz));
1252 		return -EINVAL;
1253 	}
1254 
1255 	entries = roundup_pow_of_two(entries + 1);
1256 	if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1)
1257 		return -EINVAL;
1258 
1259 	if (entries == ibcq->cqe + 1)
1260 		return 0;
1261 
1262 	mutex_lock(&cq->resize_mutex);
1263 	if (udata) {
1264 		err = resize_user(dev, cq, entries, udata, &npas, &page_shift,
1265 				  &cqe_size);
1266 	} else {
1267 		cqe_size = 64;
1268 		err = resize_kernel(dev, cq, entries, cqe_size);
1269 		if (!err) {
1270 			struct mlx5_frag_buf *frag_buf = &cq->resize_buf->frag_buf;
1271 
1272 			npas = frag_buf->npages;
1273 			page_shift = frag_buf->page_shift;
1274 		}
1275 	}
1276 
1277 	if (err)
1278 		goto ex;
1279 
1280 	inlen = MLX5_ST_SZ_BYTES(modify_cq_in) +
1281 		MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas;
1282 
1283 	in = kvzalloc(inlen, GFP_KERNEL);
1284 	if (!in) {
1285 		err = -ENOMEM;
1286 		goto ex_resize;
1287 	}
1288 
1289 	pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas);
1290 	if (udata)
1291 		mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift,
1292 				     pas, 0);
1293 	else
1294 		mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas);
1295 
1296 	MLX5_SET(modify_cq_in, in,
1297 		 modify_field_select_resize_field_select.resize_field_select.resize_field_select,
1298 		 MLX5_MODIFY_CQ_MASK_LOG_SIZE  |
1299 		 MLX5_MODIFY_CQ_MASK_PG_OFFSET |
1300 		 MLX5_MODIFY_CQ_MASK_PG_SIZE);
1301 
1302 	cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context);
1303 
1304 	MLX5_SET(cqc, cqc, log_page_size,
1305 		 page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1306 	MLX5_SET(cqc, cqc, cqe_sz,
1307 		 cqe_sz_to_mlx_sz(cqe_size,
1308 				  cq->private_flags &
1309 				  MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD));
1310 	MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries));
1311 
1312 	MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE);
1313 	MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn);
1314 
1315 	err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen);
1316 	if (err)
1317 		goto ex_alloc;
1318 
1319 	if (udata) {
1320 		cq->ibcq.cqe = entries - 1;
1321 		ib_umem_release(cq->buf.umem);
1322 		cq->buf.umem = cq->resize_umem;
1323 		cq->resize_umem = NULL;
1324 	} else {
1325 		struct mlx5_ib_cq_buf tbuf;
1326 		int resized = 0;
1327 
1328 		spin_lock_irqsave(&cq->lock, flags);
1329 		if (cq->resize_buf) {
1330 			err = copy_resize_cqes(cq);
1331 			if (!err) {
1332 				tbuf = cq->buf;
1333 				cq->buf = *cq->resize_buf;
1334 				kfree(cq->resize_buf);
1335 				cq->resize_buf = NULL;
1336 				resized = 1;
1337 			}
1338 		}
1339 		cq->ibcq.cqe = entries - 1;
1340 		spin_unlock_irqrestore(&cq->lock, flags);
1341 		if (resized)
1342 			free_cq_buf(dev, &tbuf);
1343 	}
1344 	mutex_unlock(&cq->resize_mutex);
1345 
1346 	kvfree(in);
1347 	return 0;
1348 
1349 ex_alloc:
1350 	kvfree(in);
1351 
1352 ex_resize:
1353 	if (udata)
1354 		un_resize_user(cq);
1355 	else
1356 		un_resize_kernel(dev, cq);
1357 ex:
1358 	mutex_unlock(&cq->resize_mutex);
1359 	return err;
1360 }
1361 
1362 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq)
1363 {
1364 	struct mlx5_ib_cq *cq;
1365 
1366 	if (!ibcq)
1367 		return 128;
1368 
1369 	cq = to_mcq(ibcq);
1370 	return cq->cqe_size;
1371 }
1372 
1373 /* Called from atomic context */
1374 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc)
1375 {
1376 	struct mlx5_ib_wc *soft_wc;
1377 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
1378 	unsigned long flags;
1379 
1380 	soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC);
1381 	if (!soft_wc)
1382 		return -ENOMEM;
1383 
1384 	soft_wc->wc = *wc;
1385 	spin_lock_irqsave(&cq->lock, flags);
1386 	list_add_tail(&soft_wc->list, &cq->wc_list);
1387 	if (cq->notify_flags == IB_CQ_NEXT_COMP ||
1388 	    wc->status != IB_WC_SUCCESS) {
1389 		cq->notify_flags = 0;
1390 		schedule_work(&cq->notify_work);
1391 	}
1392 	spin_unlock_irqrestore(&cq->lock, flags);
1393 
1394 	return 0;
1395 }
1396