xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision 05bcf503)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/mlx4/cq.h>
35 #include <linux/mlx4/qp.h>
36 #include <linux/slab.h>
37 
38 #include "mlx4_ib.h"
39 #include "user.h"
40 
41 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
42 {
43 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
44 	ibcq->comp_handler(ibcq, ibcq->cq_context);
45 }
46 
47 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
48 {
49 	struct ib_event event;
50 	struct ib_cq *ibcq;
51 
52 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
53 		pr_warn("Unexpected event type %d "
54 		       "on CQ %06x\n", type, cq->cqn);
55 		return;
56 	}
57 
58 	ibcq = &to_mibcq(cq)->ibcq;
59 	if (ibcq->event_handler) {
60 		event.device     = ibcq->device;
61 		event.event      = IB_EVENT_CQ_ERR;
62 		event.element.cq = ibcq;
63 		ibcq->event_handler(&event, ibcq->cq_context);
64 	}
65 }
66 
67 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
68 {
69 	return mlx4_buf_offset(&buf->buf, n * sizeof (struct mlx4_cqe));
70 }
71 
72 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
73 {
74 	return get_cqe_from_buf(&cq->buf, n);
75 }
76 
77 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
78 {
79 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
80 
81 	return (!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
82 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
83 }
84 
85 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
86 {
87 	return get_sw_cqe(cq, cq->mcq.cons_index);
88 }
89 
90 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
91 {
92 	struct mlx4_ib_cq *mcq = to_mcq(cq);
93 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
94 
95 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
96 }
97 
98 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
99 {
100 	int err;
101 
102 	err = mlx4_buf_alloc(dev->dev, nent * sizeof(struct mlx4_cqe),
103 			     PAGE_SIZE * 2, &buf->buf);
104 
105 	if (err)
106 		goto out;
107 
108 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
109 				    &buf->mtt);
110 	if (err)
111 		goto err_buf;
112 
113 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
114 	if (err)
115 		goto err_mtt;
116 
117 	return 0;
118 
119 err_mtt:
120 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
121 
122 err_buf:
123 	mlx4_buf_free(dev->dev, nent * sizeof(struct mlx4_cqe),
124 			      &buf->buf);
125 
126 out:
127 	return err;
128 }
129 
130 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
131 {
132 	mlx4_buf_free(dev->dev, (cqe + 1) * sizeof(struct mlx4_cqe), &buf->buf);
133 }
134 
135 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
136 			       struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
137 			       u64 buf_addr, int cqe)
138 {
139 	int err;
140 
141 	*umem = ib_umem_get(context, buf_addr, cqe * sizeof (struct mlx4_cqe),
142 			    IB_ACCESS_LOCAL_WRITE, 1);
143 	if (IS_ERR(*umem))
144 		return PTR_ERR(*umem);
145 
146 	err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
147 			    ilog2((*umem)->page_size), &buf->mtt);
148 	if (err)
149 		goto err_buf;
150 
151 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
152 	if (err)
153 		goto err_mtt;
154 
155 	return 0;
156 
157 err_mtt:
158 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
159 
160 err_buf:
161 	ib_umem_release(*umem);
162 
163 	return err;
164 }
165 
166 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
167 				struct ib_ucontext *context,
168 				struct ib_udata *udata)
169 {
170 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
171 	struct mlx4_ib_cq *cq;
172 	struct mlx4_uar *uar;
173 	int err;
174 
175 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
176 		return ERR_PTR(-EINVAL);
177 
178 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
179 	if (!cq)
180 		return ERR_PTR(-ENOMEM);
181 
182 	entries      = roundup_pow_of_two(entries + 1);
183 	cq->ibcq.cqe = entries - 1;
184 	mutex_init(&cq->resize_mutex);
185 	spin_lock_init(&cq->lock);
186 	cq->resize_buf = NULL;
187 	cq->resize_umem = NULL;
188 
189 	if (context) {
190 		struct mlx4_ib_create_cq ucmd;
191 
192 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
193 			err = -EFAULT;
194 			goto err_cq;
195 		}
196 
197 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
198 					  ucmd.buf_addr, entries);
199 		if (err)
200 			goto err_cq;
201 
202 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
203 					  &cq->db);
204 		if (err)
205 			goto err_mtt;
206 
207 		uar = &to_mucontext(context)->uar;
208 	} else {
209 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
210 		if (err)
211 			goto err_cq;
212 
213 		cq->mcq.set_ci_db  = cq->db.db;
214 		cq->mcq.arm_db     = cq->db.db + 1;
215 		*cq->mcq.set_ci_db = 0;
216 		*cq->mcq.arm_db    = 0;
217 
218 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
219 		if (err)
220 			goto err_db;
221 
222 		uar = &dev->priv_uar;
223 	}
224 
225 	if (dev->eq_table)
226 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
227 
228 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
229 			    cq->db.dma, &cq->mcq, vector, 0);
230 	if (err)
231 		goto err_dbmap;
232 
233 	cq->mcq.comp  = mlx4_ib_cq_comp;
234 	cq->mcq.event = mlx4_ib_cq_event;
235 
236 	if (context)
237 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
238 			err = -EFAULT;
239 			goto err_dbmap;
240 		}
241 
242 	return &cq->ibcq;
243 
244 err_dbmap:
245 	if (context)
246 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
247 
248 err_mtt:
249 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
250 
251 	if (context)
252 		ib_umem_release(cq->umem);
253 	else
254 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
255 
256 err_db:
257 	if (!context)
258 		mlx4_db_free(dev->dev, &cq->db);
259 
260 err_cq:
261 	kfree(cq);
262 
263 	return ERR_PTR(err);
264 }
265 
266 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
267 				  int entries)
268 {
269 	int err;
270 
271 	if (cq->resize_buf)
272 		return -EBUSY;
273 
274 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
275 	if (!cq->resize_buf)
276 		return -ENOMEM;
277 
278 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
279 	if (err) {
280 		kfree(cq->resize_buf);
281 		cq->resize_buf = NULL;
282 		return err;
283 	}
284 
285 	cq->resize_buf->cqe = entries - 1;
286 
287 	return 0;
288 }
289 
290 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
291 				   int entries, struct ib_udata *udata)
292 {
293 	struct mlx4_ib_resize_cq ucmd;
294 	int err;
295 
296 	if (cq->resize_umem)
297 		return -EBUSY;
298 
299 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
300 		return -EFAULT;
301 
302 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
303 	if (!cq->resize_buf)
304 		return -ENOMEM;
305 
306 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
307 				  &cq->resize_umem, ucmd.buf_addr, entries);
308 	if (err) {
309 		kfree(cq->resize_buf);
310 		cq->resize_buf = NULL;
311 		return err;
312 	}
313 
314 	cq->resize_buf->cqe = entries - 1;
315 
316 	return 0;
317 }
318 
319 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
320 {
321 	u32 i;
322 
323 	i = cq->mcq.cons_index;
324 	while (get_sw_cqe(cq, i & cq->ibcq.cqe))
325 		++i;
326 
327 	return i - cq->mcq.cons_index;
328 }
329 
330 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
331 {
332 	struct mlx4_cqe *cqe, *new_cqe;
333 	int i;
334 
335 	i = cq->mcq.cons_index;
336 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
337 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
338 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
339 					   (i + 1) & cq->resize_buf->cqe);
340 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), sizeof(struct mlx4_cqe));
341 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
342 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
343 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
344 	}
345 	++cq->mcq.cons_index;
346 }
347 
348 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
349 {
350 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
351 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
352 	struct mlx4_mtt mtt;
353 	int outst_cqe;
354 	int err;
355 
356 	mutex_lock(&cq->resize_mutex);
357 
358 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
359 		err = -EINVAL;
360 		goto out;
361 	}
362 
363 	entries = roundup_pow_of_two(entries + 1);
364 	if (entries == ibcq->cqe + 1) {
365 		err = 0;
366 		goto out;
367 	}
368 
369 	if (ibcq->uobject) {
370 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
371 		if (err)
372 			goto out;
373 	} else {
374 		/* Can't be smaller than the number of outstanding CQEs */
375 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
376 		if (entries < outst_cqe + 1) {
377 			err = 0;
378 			goto out;
379 		}
380 
381 		err = mlx4_alloc_resize_buf(dev, cq, entries);
382 		if (err)
383 			goto out;
384 	}
385 
386 	mtt = cq->buf.mtt;
387 
388 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
389 	if (err)
390 		goto err_buf;
391 
392 	mlx4_mtt_cleanup(dev->dev, &mtt);
393 	if (ibcq->uobject) {
394 		cq->buf      = cq->resize_buf->buf;
395 		cq->ibcq.cqe = cq->resize_buf->cqe;
396 		ib_umem_release(cq->umem);
397 		cq->umem     = cq->resize_umem;
398 
399 		kfree(cq->resize_buf);
400 		cq->resize_buf = NULL;
401 		cq->resize_umem = NULL;
402 	} else {
403 		struct mlx4_ib_cq_buf tmp_buf;
404 		int tmp_cqe = 0;
405 
406 		spin_lock_irq(&cq->lock);
407 		if (cq->resize_buf) {
408 			mlx4_ib_cq_resize_copy_cqes(cq);
409 			tmp_buf = cq->buf;
410 			tmp_cqe = cq->ibcq.cqe;
411 			cq->buf      = cq->resize_buf->buf;
412 			cq->ibcq.cqe = cq->resize_buf->cqe;
413 
414 			kfree(cq->resize_buf);
415 			cq->resize_buf = NULL;
416 		}
417 		spin_unlock_irq(&cq->lock);
418 
419 		if (tmp_cqe)
420 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
421 	}
422 
423 	goto out;
424 
425 err_buf:
426 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
427 	if (!ibcq->uobject)
428 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
429 				    cq->resize_buf->cqe);
430 
431 	kfree(cq->resize_buf);
432 	cq->resize_buf = NULL;
433 
434 	if (cq->resize_umem) {
435 		ib_umem_release(cq->resize_umem);
436 		cq->resize_umem = NULL;
437 	}
438 
439 out:
440 	mutex_unlock(&cq->resize_mutex);
441 	return err;
442 }
443 
444 int mlx4_ib_destroy_cq(struct ib_cq *cq)
445 {
446 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
447 	struct mlx4_ib_cq *mcq = to_mcq(cq);
448 
449 	mlx4_cq_free(dev->dev, &mcq->mcq);
450 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
451 
452 	if (cq->uobject) {
453 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
454 		ib_umem_release(mcq->umem);
455 	} else {
456 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
457 		mlx4_db_free(dev->dev, &mcq->db);
458 	}
459 
460 	kfree(mcq);
461 
462 	return 0;
463 }
464 
465 static void dump_cqe(void *cqe)
466 {
467 	__be32 *buf = cqe;
468 
469 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
470 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
471 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
472 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
473 }
474 
475 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
476 				     struct ib_wc *wc)
477 {
478 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
479 		pr_debug("local QP operation err "
480 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
481 		       "opcode = %02x)\n",
482 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
483 		       cqe->vendor_err_syndrome,
484 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
485 		dump_cqe(cqe);
486 	}
487 
488 	switch (cqe->syndrome) {
489 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
490 		wc->status = IB_WC_LOC_LEN_ERR;
491 		break;
492 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
493 		wc->status = IB_WC_LOC_QP_OP_ERR;
494 		break;
495 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
496 		wc->status = IB_WC_LOC_PROT_ERR;
497 		break;
498 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
499 		wc->status = IB_WC_WR_FLUSH_ERR;
500 		break;
501 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
502 		wc->status = IB_WC_MW_BIND_ERR;
503 		break;
504 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
505 		wc->status = IB_WC_BAD_RESP_ERR;
506 		break;
507 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
508 		wc->status = IB_WC_LOC_ACCESS_ERR;
509 		break;
510 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
511 		wc->status = IB_WC_REM_INV_REQ_ERR;
512 		break;
513 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
514 		wc->status = IB_WC_REM_ACCESS_ERR;
515 		break;
516 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
517 		wc->status = IB_WC_REM_OP_ERR;
518 		break;
519 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
520 		wc->status = IB_WC_RETRY_EXC_ERR;
521 		break;
522 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
523 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
524 		break;
525 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
526 		wc->status = IB_WC_REM_ABORT_ERR;
527 		break;
528 	default:
529 		wc->status = IB_WC_GENERAL_ERR;
530 		break;
531 	}
532 
533 	wc->vendor_err = cqe->vendor_err_syndrome;
534 }
535 
536 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
537 {
538 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
539 				      MLX4_CQE_STATUS_IPV4F     |
540 				      MLX4_CQE_STATUS_IPV4OPT   |
541 				      MLX4_CQE_STATUS_IPV6      |
542 				      MLX4_CQE_STATUS_IPOK)) ==
543 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
544 			    MLX4_CQE_STATUS_IPOK))              &&
545 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
546 				      MLX4_CQE_STATUS_TCP))     &&
547 		checksum == cpu_to_be16(0xffff);
548 }
549 
550 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
551 			   unsigned tail, struct mlx4_cqe *cqe)
552 {
553 	struct mlx4_ib_proxy_sqp_hdr *hdr;
554 
555 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
556 				   qp->sqp_proxy_rcv[tail].map,
557 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
558 				   DMA_FROM_DEVICE);
559 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
560 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
561 	wc->slid	= be16_to_cpu(hdr->tun.slid_mac_47_32);
562 	wc->sl		= (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
563 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
564 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
565 	wc->dlid_path_bits = 0;
566 
567 	return 0;
568 }
569 
570 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
571 			    struct mlx4_ib_qp **cur_qp,
572 			    struct ib_wc *wc)
573 {
574 	struct mlx4_cqe *cqe;
575 	struct mlx4_qp *mqp;
576 	struct mlx4_ib_wq *wq;
577 	struct mlx4_ib_srq *srq;
578 	int is_send;
579 	int is_error;
580 	u32 g_mlpath_rqpn;
581 	u16 wqe_ctr;
582 	unsigned tail = 0;
583 
584 repoll:
585 	cqe = next_cqe_sw(cq);
586 	if (!cqe)
587 		return -EAGAIN;
588 
589 	++cq->mcq.cons_index;
590 
591 	/*
592 	 * Make sure we read CQ entry contents after we've checked the
593 	 * ownership bit.
594 	 */
595 	rmb();
596 
597 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
598 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
599 		MLX4_CQE_OPCODE_ERROR;
600 
601 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
602 		     is_send)) {
603 		pr_warn("Completion for NOP opcode detected!\n");
604 		return -EINVAL;
605 	}
606 
607 	/* Resize CQ in progress */
608 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
609 		if (cq->resize_buf) {
610 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
611 
612 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
613 			cq->buf      = cq->resize_buf->buf;
614 			cq->ibcq.cqe = cq->resize_buf->cqe;
615 
616 			kfree(cq->resize_buf);
617 			cq->resize_buf = NULL;
618 		}
619 
620 		goto repoll;
621 	}
622 
623 	if (!*cur_qp ||
624 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
625 		/*
626 		 * We do not have to take the QP table lock here,
627 		 * because CQs will be locked while QPs are removed
628 		 * from the table.
629 		 */
630 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
631 				       be32_to_cpu(cqe->vlan_my_qpn));
632 		if (unlikely(!mqp)) {
633 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
634 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
635 			return -EINVAL;
636 		}
637 
638 		*cur_qp = to_mibqp(mqp);
639 	}
640 
641 	wc->qp = &(*cur_qp)->ibqp;
642 
643 	if (is_send) {
644 		wq = &(*cur_qp)->sq;
645 		if (!(*cur_qp)->sq_signal_bits) {
646 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
647 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
648 		}
649 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
650 		++wq->tail;
651 	} else if ((*cur_qp)->ibqp.srq) {
652 		srq = to_msrq((*cur_qp)->ibqp.srq);
653 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
654 		wc->wr_id = srq->wrid[wqe_ctr];
655 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
656 	} else {
657 		wq	  = &(*cur_qp)->rq;
658 		tail	  = wq->tail & (wq->wqe_cnt - 1);
659 		wc->wr_id = wq->wrid[tail];
660 		++wq->tail;
661 	}
662 
663 	if (unlikely(is_error)) {
664 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
665 		return 0;
666 	}
667 
668 	wc->status = IB_WC_SUCCESS;
669 
670 	if (is_send) {
671 		wc->wc_flags = 0;
672 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
673 		case MLX4_OPCODE_RDMA_WRITE_IMM:
674 			wc->wc_flags |= IB_WC_WITH_IMM;
675 		case MLX4_OPCODE_RDMA_WRITE:
676 			wc->opcode    = IB_WC_RDMA_WRITE;
677 			break;
678 		case MLX4_OPCODE_SEND_IMM:
679 			wc->wc_flags |= IB_WC_WITH_IMM;
680 		case MLX4_OPCODE_SEND:
681 		case MLX4_OPCODE_SEND_INVAL:
682 			wc->opcode    = IB_WC_SEND;
683 			break;
684 		case MLX4_OPCODE_RDMA_READ:
685 			wc->opcode    = IB_WC_RDMA_READ;
686 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
687 			break;
688 		case MLX4_OPCODE_ATOMIC_CS:
689 			wc->opcode    = IB_WC_COMP_SWAP;
690 			wc->byte_len  = 8;
691 			break;
692 		case MLX4_OPCODE_ATOMIC_FA:
693 			wc->opcode    = IB_WC_FETCH_ADD;
694 			wc->byte_len  = 8;
695 			break;
696 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
697 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
698 			wc->byte_len  = 8;
699 			break;
700 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
701 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
702 			wc->byte_len  = 8;
703 			break;
704 		case MLX4_OPCODE_BIND_MW:
705 			wc->opcode    = IB_WC_BIND_MW;
706 			break;
707 		case MLX4_OPCODE_LSO:
708 			wc->opcode    = IB_WC_LSO;
709 			break;
710 		case MLX4_OPCODE_FMR:
711 			wc->opcode    = IB_WC_FAST_REG_MR;
712 			break;
713 		case MLX4_OPCODE_LOCAL_INVAL:
714 			wc->opcode    = IB_WC_LOCAL_INV;
715 			break;
716 		}
717 	} else {
718 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
719 
720 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
721 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
722 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
723 			wc->wc_flags	= IB_WC_WITH_IMM;
724 			wc->ex.imm_data = cqe->immed_rss_invalid;
725 			break;
726 		case MLX4_RECV_OPCODE_SEND_INVAL:
727 			wc->opcode	= IB_WC_RECV;
728 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
729 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
730 			break;
731 		case MLX4_RECV_OPCODE_SEND:
732 			wc->opcode   = IB_WC_RECV;
733 			wc->wc_flags = 0;
734 			break;
735 		case MLX4_RECV_OPCODE_SEND_IMM:
736 			wc->opcode	= IB_WC_RECV;
737 			wc->wc_flags	= IB_WC_WITH_IMM;
738 			wc->ex.imm_data = cqe->immed_rss_invalid;
739 			break;
740 		}
741 
742 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
743 			if ((*cur_qp)->mlx4_ib_qp_type &
744 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
745 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
746 				return use_tunnel_data(*cur_qp, cq, wc, tail, cqe);
747 		}
748 
749 		wc->slid	   = be16_to_cpu(cqe->rlid);
750 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
751 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
752 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
753 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
754 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
755 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
756 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
757 		if (rdma_port_get_link_layer(wc->qp->device,
758 				(*cur_qp)->port) == IB_LINK_LAYER_ETHERNET)
759 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
760 		else
761 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
762 	}
763 
764 	return 0;
765 }
766 
767 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
768 {
769 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
770 	struct mlx4_ib_qp *cur_qp = NULL;
771 	unsigned long flags;
772 	int npolled;
773 	int err = 0;
774 
775 	spin_lock_irqsave(&cq->lock, flags);
776 
777 	for (npolled = 0; npolled < num_entries; ++npolled) {
778 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
779 		if (err)
780 			break;
781 	}
782 
783 	mlx4_cq_set_ci(&cq->mcq);
784 
785 	spin_unlock_irqrestore(&cq->lock, flags);
786 
787 	if (err == 0 || err == -EAGAIN)
788 		return npolled;
789 	else
790 		return err;
791 }
792 
793 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
794 {
795 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
796 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
797 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
798 		    to_mdev(ibcq->device)->uar_map,
799 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
800 
801 	return 0;
802 }
803 
804 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
805 {
806 	u32 prod_index;
807 	int nfreed = 0;
808 	struct mlx4_cqe *cqe, *dest;
809 	u8 owner_bit;
810 
811 	/*
812 	 * First we need to find the current producer index, so we
813 	 * know where to start cleaning from.  It doesn't matter if HW
814 	 * adds new entries after this loop -- the QP we're worried
815 	 * about is already in RESET, so the new entries won't come
816 	 * from our QP and therefore don't need to be checked.
817 	 */
818 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
819 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
820 			break;
821 
822 	/*
823 	 * Now sweep backwards through the CQ, removing CQ entries
824 	 * that match our QP by copying older entries on top of them.
825 	 */
826 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
827 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
828 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
829 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
830 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
831 			++nfreed;
832 		} else if (nfreed) {
833 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
834 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
835 			memcpy(dest, cqe, sizeof *cqe);
836 			dest->owner_sr_opcode = owner_bit |
837 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
838 		}
839 	}
840 
841 	if (nfreed) {
842 		cq->mcq.cons_index += nfreed;
843 		/*
844 		 * Make sure update of buffer contents is done before
845 		 * updating consumer index.
846 		 */
847 		wmb();
848 		mlx4_cq_set_ci(&cq->mcq);
849 	}
850 }
851 
852 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
853 {
854 	spin_lock_irq(&cq->lock);
855 	__mlx4_ib_cq_clean(cq, qpn, srq);
856 	spin_unlock_irq(&cq->lock);
857 }
858