xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision b6bec26c)
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 * buf->entry_size);
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 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
81 
82 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
83 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
84 }
85 
86 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
87 {
88 	return get_sw_cqe(cq, cq->mcq.cons_index);
89 }
90 
91 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
92 {
93 	struct mlx4_ib_cq *mcq = to_mcq(cq);
94 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
95 
96 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
97 }
98 
99 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
100 {
101 	int err;
102 
103 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
104 			     PAGE_SIZE * 2, &buf->buf);
105 
106 	if (err)
107 		goto out;
108 
109 	buf->entry_size = dev->dev->caps.cqe_size;
110 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
111 				    &buf->mtt);
112 	if (err)
113 		goto err_buf;
114 
115 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
116 	if (err)
117 		goto err_mtt;
118 
119 	return 0;
120 
121 err_mtt:
122 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
123 
124 err_buf:
125 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
126 
127 out:
128 	return err;
129 }
130 
131 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
132 {
133 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
134 }
135 
136 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
137 			       struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
138 			       u64 buf_addr, int cqe)
139 {
140 	int err;
141 	int cqe_size = dev->dev->caps.cqe_size;
142 
143 	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
144 			    IB_ACCESS_LOCAL_WRITE, 1);
145 	if (IS_ERR(*umem))
146 		return PTR_ERR(*umem);
147 
148 	err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
149 			    ilog2((*umem)->page_size), &buf->mtt);
150 	if (err)
151 		goto err_buf;
152 
153 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
154 	if (err)
155 		goto err_mtt;
156 
157 	return 0;
158 
159 err_mtt:
160 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
161 
162 err_buf:
163 	ib_umem_release(*umem);
164 
165 	return err;
166 }
167 
168 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
169 				struct ib_ucontext *context,
170 				struct ib_udata *udata)
171 {
172 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
173 	struct mlx4_ib_cq *cq;
174 	struct mlx4_uar *uar;
175 	int err;
176 
177 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
178 		return ERR_PTR(-EINVAL);
179 
180 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
181 	if (!cq)
182 		return ERR_PTR(-ENOMEM);
183 
184 	entries      = roundup_pow_of_two(entries + 1);
185 	cq->ibcq.cqe = entries - 1;
186 	mutex_init(&cq->resize_mutex);
187 	spin_lock_init(&cq->lock);
188 	cq->resize_buf = NULL;
189 	cq->resize_umem = NULL;
190 
191 	if (context) {
192 		struct mlx4_ib_create_cq ucmd;
193 
194 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
195 			err = -EFAULT;
196 			goto err_cq;
197 		}
198 
199 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
200 					  ucmd.buf_addr, entries);
201 		if (err)
202 			goto err_cq;
203 
204 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
205 					  &cq->db);
206 		if (err)
207 			goto err_mtt;
208 
209 		uar = &to_mucontext(context)->uar;
210 	} else {
211 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
212 		if (err)
213 			goto err_cq;
214 
215 		cq->mcq.set_ci_db  = cq->db.db;
216 		cq->mcq.arm_db     = cq->db.db + 1;
217 		*cq->mcq.set_ci_db = 0;
218 		*cq->mcq.arm_db    = 0;
219 
220 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
221 		if (err)
222 			goto err_db;
223 
224 		uar = &dev->priv_uar;
225 	}
226 
227 	if (dev->eq_table)
228 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
229 
230 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
231 			    cq->db.dma, &cq->mcq, vector, 0);
232 	if (err)
233 		goto err_dbmap;
234 
235 	cq->mcq.comp  = mlx4_ib_cq_comp;
236 	cq->mcq.event = mlx4_ib_cq_event;
237 
238 	if (context)
239 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
240 			err = -EFAULT;
241 			goto err_dbmap;
242 		}
243 
244 	return &cq->ibcq;
245 
246 err_dbmap:
247 	if (context)
248 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
249 
250 err_mtt:
251 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
252 
253 	if (context)
254 		ib_umem_release(cq->umem);
255 	else
256 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
257 
258 err_db:
259 	if (!context)
260 		mlx4_db_free(dev->dev, &cq->db);
261 
262 err_cq:
263 	kfree(cq);
264 
265 	return ERR_PTR(err);
266 }
267 
268 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
269 				  int entries)
270 {
271 	int err;
272 
273 	if (cq->resize_buf)
274 		return -EBUSY;
275 
276 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
277 	if (!cq->resize_buf)
278 		return -ENOMEM;
279 
280 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
281 	if (err) {
282 		kfree(cq->resize_buf);
283 		cq->resize_buf = NULL;
284 		return err;
285 	}
286 
287 	cq->resize_buf->cqe = entries - 1;
288 
289 	return 0;
290 }
291 
292 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
293 				   int entries, struct ib_udata *udata)
294 {
295 	struct mlx4_ib_resize_cq ucmd;
296 	int err;
297 
298 	if (cq->resize_umem)
299 		return -EBUSY;
300 
301 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
302 		return -EFAULT;
303 
304 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
305 	if (!cq->resize_buf)
306 		return -ENOMEM;
307 
308 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
309 				  &cq->resize_umem, ucmd.buf_addr, entries);
310 	if (err) {
311 		kfree(cq->resize_buf);
312 		cq->resize_buf = NULL;
313 		return err;
314 	}
315 
316 	cq->resize_buf->cqe = entries - 1;
317 
318 	return 0;
319 }
320 
321 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
322 {
323 	u32 i;
324 
325 	i = cq->mcq.cons_index;
326 	while (get_sw_cqe(cq, i & cq->ibcq.cqe))
327 		++i;
328 
329 	return i - cq->mcq.cons_index;
330 }
331 
332 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
333 {
334 	struct mlx4_cqe *cqe, *new_cqe;
335 	int i;
336 	int cqe_size = cq->buf.entry_size;
337 	int cqe_inc = cqe_size == 64 ? 1 : 0;
338 
339 	i = cq->mcq.cons_index;
340 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
341 	cqe += cqe_inc;
342 
343 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
344 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
345 					   (i + 1) & cq->resize_buf->cqe);
346 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
347 		new_cqe += cqe_inc;
348 
349 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
350 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
351 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
352 		cqe += cqe_inc;
353 	}
354 	++cq->mcq.cons_index;
355 }
356 
357 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
358 {
359 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
360 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
361 	struct mlx4_mtt mtt;
362 	int outst_cqe;
363 	int err;
364 
365 	mutex_lock(&cq->resize_mutex);
366 
367 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
368 		err = -EINVAL;
369 		goto out;
370 	}
371 
372 	entries = roundup_pow_of_two(entries + 1);
373 	if (entries == ibcq->cqe + 1) {
374 		err = 0;
375 		goto out;
376 	}
377 
378 	if (ibcq->uobject) {
379 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
380 		if (err)
381 			goto out;
382 	} else {
383 		/* Can't be smaller than the number of outstanding CQEs */
384 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
385 		if (entries < outst_cqe + 1) {
386 			err = 0;
387 			goto out;
388 		}
389 
390 		err = mlx4_alloc_resize_buf(dev, cq, entries);
391 		if (err)
392 			goto out;
393 	}
394 
395 	mtt = cq->buf.mtt;
396 
397 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
398 	if (err)
399 		goto err_buf;
400 
401 	mlx4_mtt_cleanup(dev->dev, &mtt);
402 	if (ibcq->uobject) {
403 		cq->buf      = cq->resize_buf->buf;
404 		cq->ibcq.cqe = cq->resize_buf->cqe;
405 		ib_umem_release(cq->umem);
406 		cq->umem     = cq->resize_umem;
407 
408 		kfree(cq->resize_buf);
409 		cq->resize_buf = NULL;
410 		cq->resize_umem = NULL;
411 	} else {
412 		struct mlx4_ib_cq_buf tmp_buf;
413 		int tmp_cqe = 0;
414 
415 		spin_lock_irq(&cq->lock);
416 		if (cq->resize_buf) {
417 			mlx4_ib_cq_resize_copy_cqes(cq);
418 			tmp_buf = cq->buf;
419 			tmp_cqe = cq->ibcq.cqe;
420 			cq->buf      = cq->resize_buf->buf;
421 			cq->ibcq.cqe = cq->resize_buf->cqe;
422 
423 			kfree(cq->resize_buf);
424 			cq->resize_buf = NULL;
425 		}
426 		spin_unlock_irq(&cq->lock);
427 
428 		if (tmp_cqe)
429 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
430 	}
431 
432 	goto out;
433 
434 err_buf:
435 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
436 	if (!ibcq->uobject)
437 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
438 				    cq->resize_buf->cqe);
439 
440 	kfree(cq->resize_buf);
441 	cq->resize_buf = NULL;
442 
443 	if (cq->resize_umem) {
444 		ib_umem_release(cq->resize_umem);
445 		cq->resize_umem = NULL;
446 	}
447 
448 out:
449 	mutex_unlock(&cq->resize_mutex);
450 
451 	return err;
452 }
453 
454 int mlx4_ib_destroy_cq(struct ib_cq *cq)
455 {
456 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
457 	struct mlx4_ib_cq *mcq = to_mcq(cq);
458 
459 	mlx4_cq_free(dev->dev, &mcq->mcq);
460 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
461 
462 	if (cq->uobject) {
463 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
464 		ib_umem_release(mcq->umem);
465 	} else {
466 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
467 		mlx4_db_free(dev->dev, &mcq->db);
468 	}
469 
470 	kfree(mcq);
471 
472 	return 0;
473 }
474 
475 static void dump_cqe(void *cqe)
476 {
477 	__be32 *buf = cqe;
478 
479 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
480 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
481 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
482 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
483 }
484 
485 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
486 				     struct ib_wc *wc)
487 {
488 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
489 		pr_debug("local QP operation err "
490 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
491 		       "opcode = %02x)\n",
492 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
493 		       cqe->vendor_err_syndrome,
494 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
495 		dump_cqe(cqe);
496 	}
497 
498 	switch (cqe->syndrome) {
499 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
500 		wc->status = IB_WC_LOC_LEN_ERR;
501 		break;
502 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
503 		wc->status = IB_WC_LOC_QP_OP_ERR;
504 		break;
505 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
506 		wc->status = IB_WC_LOC_PROT_ERR;
507 		break;
508 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
509 		wc->status = IB_WC_WR_FLUSH_ERR;
510 		break;
511 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
512 		wc->status = IB_WC_MW_BIND_ERR;
513 		break;
514 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
515 		wc->status = IB_WC_BAD_RESP_ERR;
516 		break;
517 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
518 		wc->status = IB_WC_LOC_ACCESS_ERR;
519 		break;
520 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
521 		wc->status = IB_WC_REM_INV_REQ_ERR;
522 		break;
523 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
524 		wc->status = IB_WC_REM_ACCESS_ERR;
525 		break;
526 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
527 		wc->status = IB_WC_REM_OP_ERR;
528 		break;
529 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
530 		wc->status = IB_WC_RETRY_EXC_ERR;
531 		break;
532 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
533 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
534 		break;
535 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
536 		wc->status = IB_WC_REM_ABORT_ERR;
537 		break;
538 	default:
539 		wc->status = IB_WC_GENERAL_ERR;
540 		break;
541 	}
542 
543 	wc->vendor_err = cqe->vendor_err_syndrome;
544 }
545 
546 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
547 {
548 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
549 				      MLX4_CQE_STATUS_IPV4F     |
550 				      MLX4_CQE_STATUS_IPV4OPT   |
551 				      MLX4_CQE_STATUS_IPV6      |
552 				      MLX4_CQE_STATUS_IPOK)) ==
553 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
554 			    MLX4_CQE_STATUS_IPOK))              &&
555 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
556 				      MLX4_CQE_STATUS_TCP))     &&
557 		checksum == cpu_to_be16(0xffff);
558 }
559 
560 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
561 			   unsigned tail, struct mlx4_cqe *cqe)
562 {
563 	struct mlx4_ib_proxy_sqp_hdr *hdr;
564 
565 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
566 				   qp->sqp_proxy_rcv[tail].map,
567 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
568 				   DMA_FROM_DEVICE);
569 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
570 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
571 	wc->slid	= be16_to_cpu(hdr->tun.slid_mac_47_32);
572 	wc->sl		= (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
573 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
574 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
575 	wc->dlid_path_bits = 0;
576 
577 	return 0;
578 }
579 
580 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
581 			    struct mlx4_ib_qp **cur_qp,
582 			    struct ib_wc *wc)
583 {
584 	struct mlx4_cqe *cqe;
585 	struct mlx4_qp *mqp;
586 	struct mlx4_ib_wq *wq;
587 	struct mlx4_ib_srq *srq;
588 	int is_send;
589 	int is_error;
590 	u32 g_mlpath_rqpn;
591 	u16 wqe_ctr;
592 	unsigned tail = 0;
593 
594 repoll:
595 	cqe = next_cqe_sw(cq);
596 	if (!cqe)
597 		return -EAGAIN;
598 
599 	if (cq->buf.entry_size == 64)
600 		cqe++;
601 
602 	++cq->mcq.cons_index;
603 
604 	/*
605 	 * Make sure we read CQ entry contents after we've checked the
606 	 * ownership bit.
607 	 */
608 	rmb();
609 
610 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
611 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
612 		MLX4_CQE_OPCODE_ERROR;
613 
614 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
615 		     is_send)) {
616 		pr_warn("Completion for NOP opcode detected!\n");
617 		return -EINVAL;
618 	}
619 
620 	/* Resize CQ in progress */
621 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
622 		if (cq->resize_buf) {
623 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
624 
625 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
626 			cq->buf      = cq->resize_buf->buf;
627 			cq->ibcq.cqe = cq->resize_buf->cqe;
628 
629 			kfree(cq->resize_buf);
630 			cq->resize_buf = NULL;
631 		}
632 
633 		goto repoll;
634 	}
635 
636 	if (!*cur_qp ||
637 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
638 		/*
639 		 * We do not have to take the QP table lock here,
640 		 * because CQs will be locked while QPs are removed
641 		 * from the table.
642 		 */
643 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
644 				       be32_to_cpu(cqe->vlan_my_qpn));
645 		if (unlikely(!mqp)) {
646 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
647 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
648 			return -EINVAL;
649 		}
650 
651 		*cur_qp = to_mibqp(mqp);
652 	}
653 
654 	wc->qp = &(*cur_qp)->ibqp;
655 
656 	if (is_send) {
657 		wq = &(*cur_qp)->sq;
658 		if (!(*cur_qp)->sq_signal_bits) {
659 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
660 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
661 		}
662 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
663 		++wq->tail;
664 	} else if ((*cur_qp)->ibqp.srq) {
665 		srq = to_msrq((*cur_qp)->ibqp.srq);
666 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
667 		wc->wr_id = srq->wrid[wqe_ctr];
668 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
669 	} else {
670 		wq	  = &(*cur_qp)->rq;
671 		tail	  = wq->tail & (wq->wqe_cnt - 1);
672 		wc->wr_id = wq->wrid[tail];
673 		++wq->tail;
674 	}
675 
676 	if (unlikely(is_error)) {
677 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
678 		return 0;
679 	}
680 
681 	wc->status = IB_WC_SUCCESS;
682 
683 	if (is_send) {
684 		wc->wc_flags = 0;
685 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
686 		case MLX4_OPCODE_RDMA_WRITE_IMM:
687 			wc->wc_flags |= IB_WC_WITH_IMM;
688 		case MLX4_OPCODE_RDMA_WRITE:
689 			wc->opcode    = IB_WC_RDMA_WRITE;
690 			break;
691 		case MLX4_OPCODE_SEND_IMM:
692 			wc->wc_flags |= IB_WC_WITH_IMM;
693 		case MLX4_OPCODE_SEND:
694 		case MLX4_OPCODE_SEND_INVAL:
695 			wc->opcode    = IB_WC_SEND;
696 			break;
697 		case MLX4_OPCODE_RDMA_READ:
698 			wc->opcode    = IB_WC_RDMA_READ;
699 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
700 			break;
701 		case MLX4_OPCODE_ATOMIC_CS:
702 			wc->opcode    = IB_WC_COMP_SWAP;
703 			wc->byte_len  = 8;
704 			break;
705 		case MLX4_OPCODE_ATOMIC_FA:
706 			wc->opcode    = IB_WC_FETCH_ADD;
707 			wc->byte_len  = 8;
708 			break;
709 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
710 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
711 			wc->byte_len  = 8;
712 			break;
713 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
714 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
715 			wc->byte_len  = 8;
716 			break;
717 		case MLX4_OPCODE_BIND_MW:
718 			wc->opcode    = IB_WC_BIND_MW;
719 			break;
720 		case MLX4_OPCODE_LSO:
721 			wc->opcode    = IB_WC_LSO;
722 			break;
723 		case MLX4_OPCODE_FMR:
724 			wc->opcode    = IB_WC_FAST_REG_MR;
725 			break;
726 		case MLX4_OPCODE_LOCAL_INVAL:
727 			wc->opcode    = IB_WC_LOCAL_INV;
728 			break;
729 		}
730 	} else {
731 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
732 
733 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
734 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
735 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
736 			wc->wc_flags	= IB_WC_WITH_IMM;
737 			wc->ex.imm_data = cqe->immed_rss_invalid;
738 			break;
739 		case MLX4_RECV_OPCODE_SEND_INVAL:
740 			wc->opcode	= IB_WC_RECV;
741 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
742 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
743 			break;
744 		case MLX4_RECV_OPCODE_SEND:
745 			wc->opcode   = IB_WC_RECV;
746 			wc->wc_flags = 0;
747 			break;
748 		case MLX4_RECV_OPCODE_SEND_IMM:
749 			wc->opcode	= IB_WC_RECV;
750 			wc->wc_flags	= IB_WC_WITH_IMM;
751 			wc->ex.imm_data = cqe->immed_rss_invalid;
752 			break;
753 		}
754 
755 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
756 			if ((*cur_qp)->mlx4_ib_qp_type &
757 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
758 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
759 				return use_tunnel_data(*cur_qp, cq, wc, tail, cqe);
760 		}
761 
762 		wc->slid	   = be16_to_cpu(cqe->rlid);
763 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
764 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
765 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
766 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
767 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
768 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
769 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
770 		if (rdma_port_get_link_layer(wc->qp->device,
771 				(*cur_qp)->port) == IB_LINK_LAYER_ETHERNET)
772 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
773 		else
774 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
775 	}
776 
777 	return 0;
778 }
779 
780 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
781 {
782 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
783 	struct mlx4_ib_qp *cur_qp = NULL;
784 	unsigned long flags;
785 	int npolled;
786 	int err = 0;
787 
788 	spin_lock_irqsave(&cq->lock, flags);
789 
790 	for (npolled = 0; npolled < num_entries; ++npolled) {
791 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
792 		if (err)
793 			break;
794 	}
795 
796 	mlx4_cq_set_ci(&cq->mcq);
797 
798 	spin_unlock_irqrestore(&cq->lock, flags);
799 
800 	if (err == 0 || err == -EAGAIN)
801 		return npolled;
802 	else
803 		return err;
804 }
805 
806 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
807 {
808 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
809 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
810 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
811 		    to_mdev(ibcq->device)->uar_map,
812 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
813 
814 	return 0;
815 }
816 
817 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
818 {
819 	u32 prod_index;
820 	int nfreed = 0;
821 	struct mlx4_cqe *cqe, *dest;
822 	u8 owner_bit;
823 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
824 
825 	/*
826 	 * First we need to find the current producer index, so we
827 	 * know where to start cleaning from.  It doesn't matter if HW
828 	 * adds new entries after this loop -- the QP we're worried
829 	 * about is already in RESET, so the new entries won't come
830 	 * from our QP and therefore don't need to be checked.
831 	 */
832 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
833 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
834 			break;
835 
836 	/*
837 	 * Now sweep backwards through the CQ, removing CQ entries
838 	 * that match our QP by copying older entries on top of them.
839 	 */
840 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
841 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
842 		cqe += cqe_inc;
843 
844 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
845 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
846 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
847 			++nfreed;
848 		} else if (nfreed) {
849 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
850 			dest += cqe_inc;
851 
852 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
853 			memcpy(dest, cqe, sizeof *cqe);
854 			dest->owner_sr_opcode = owner_bit |
855 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
856 		}
857 	}
858 
859 	if (nfreed) {
860 		cq->mcq.cons_index += nfreed;
861 		/*
862 		 * Make sure update of buffer contents is done before
863 		 * updating consumer index.
864 		 */
865 		wmb();
866 		mlx4_cq_set_ci(&cq->mcq);
867 	}
868 }
869 
870 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
871 {
872 	spin_lock_irq(&cq->lock);
873 	__mlx4_ib_cq_clean(cq, qpn, srq);
874 	spin_unlock_irq(&cq->lock);
875 }
876