xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision ee89bd6b)
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/mlx4/srq.h>
37 #include <linux/slab.h>
38 
39 #include "mlx4_ib.h"
40 #include "user.h"
41 
42 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
43 {
44 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
45 	ibcq->comp_handler(ibcq, ibcq->cq_context);
46 }
47 
48 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
49 {
50 	struct ib_event event;
51 	struct ib_cq *ibcq;
52 
53 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
54 		pr_warn("Unexpected event type %d "
55 		       "on CQ %06x\n", type, cq->cqn);
56 		return;
57 	}
58 
59 	ibcq = &to_mibcq(cq)->ibcq;
60 	if (ibcq->event_handler) {
61 		event.device     = ibcq->device;
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_from_buf(struct mlx4_ib_cq_buf *buf, int n)
69 {
70 	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
71 }
72 
73 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
74 {
75 	return get_cqe_from_buf(&cq->buf, n);
76 }
77 
78 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
79 {
80 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
81 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
82 
83 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
84 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
85 }
86 
87 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
88 {
89 	return get_sw_cqe(cq, cq->mcq.cons_index);
90 }
91 
92 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
93 {
94 	struct mlx4_ib_cq *mcq = to_mcq(cq);
95 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
96 
97 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
98 }
99 
100 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
101 {
102 	int err;
103 
104 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
105 			     PAGE_SIZE * 2, &buf->buf);
106 
107 	if (err)
108 		goto out;
109 
110 	buf->entry_size = dev->dev->caps.cqe_size;
111 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
112 				    &buf->mtt);
113 	if (err)
114 		goto err_buf;
115 
116 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
117 	if (err)
118 		goto err_mtt;
119 
120 	return 0;
121 
122 err_mtt:
123 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
124 
125 err_buf:
126 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
127 
128 out:
129 	return err;
130 }
131 
132 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
133 {
134 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
135 }
136 
137 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
138 			       struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
139 			       u64 buf_addr, int cqe)
140 {
141 	int err;
142 	int cqe_size = dev->dev->caps.cqe_size;
143 
144 	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
145 			    IB_ACCESS_LOCAL_WRITE, 1);
146 	if (IS_ERR(*umem))
147 		return PTR_ERR(*umem);
148 
149 	err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
150 			    ilog2((*umem)->page_size), &buf->mtt);
151 	if (err)
152 		goto err_buf;
153 
154 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
155 	if (err)
156 		goto err_mtt;
157 
158 	return 0;
159 
160 err_mtt:
161 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
162 
163 err_buf:
164 	ib_umem_release(*umem);
165 
166 	return err;
167 }
168 
169 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
170 				struct ib_ucontext *context,
171 				struct ib_udata *udata)
172 {
173 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
174 	struct mlx4_ib_cq *cq;
175 	struct mlx4_uar *uar;
176 	int err;
177 
178 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
179 		return ERR_PTR(-EINVAL);
180 
181 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
182 	if (!cq)
183 		return ERR_PTR(-ENOMEM);
184 
185 	entries      = roundup_pow_of_two(entries + 1);
186 	cq->ibcq.cqe = entries - 1;
187 	mutex_init(&cq->resize_mutex);
188 	spin_lock_init(&cq->lock);
189 	cq->resize_buf = NULL;
190 	cq->resize_umem = NULL;
191 
192 	if (context) {
193 		struct mlx4_ib_create_cq ucmd;
194 
195 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
196 			err = -EFAULT;
197 			goto err_cq;
198 		}
199 
200 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
201 					  ucmd.buf_addr, entries);
202 		if (err)
203 			goto err_cq;
204 
205 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
206 					  &cq->db);
207 		if (err)
208 			goto err_mtt;
209 
210 		uar = &to_mucontext(context)->uar;
211 	} else {
212 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
213 		if (err)
214 			goto err_cq;
215 
216 		cq->mcq.set_ci_db  = cq->db.db;
217 		cq->mcq.arm_db     = cq->db.db + 1;
218 		*cq->mcq.set_ci_db = 0;
219 		*cq->mcq.arm_db    = 0;
220 
221 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
222 		if (err)
223 			goto err_db;
224 
225 		uar = &dev->priv_uar;
226 	}
227 
228 	if (dev->eq_table)
229 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
230 
231 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
232 			    cq->db.dma, &cq->mcq, vector, 0, 0);
233 	if (err)
234 		goto err_dbmap;
235 
236 	cq->mcq.comp  = mlx4_ib_cq_comp;
237 	cq->mcq.event = mlx4_ib_cq_event;
238 
239 	if (context)
240 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
241 			err = -EFAULT;
242 			goto err_dbmap;
243 		}
244 
245 	return &cq->ibcq;
246 
247 err_dbmap:
248 	if (context)
249 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
250 
251 err_mtt:
252 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
253 
254 	if (context)
255 		ib_umem_release(cq->umem);
256 	else
257 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
258 
259 err_db:
260 	if (!context)
261 		mlx4_db_free(dev->dev, &cq->db);
262 
263 err_cq:
264 	kfree(cq);
265 
266 	return ERR_PTR(err);
267 }
268 
269 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
270 				  int entries)
271 {
272 	int err;
273 
274 	if (cq->resize_buf)
275 		return -EBUSY;
276 
277 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
278 	if (!cq->resize_buf)
279 		return -ENOMEM;
280 
281 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
282 	if (err) {
283 		kfree(cq->resize_buf);
284 		cq->resize_buf = NULL;
285 		return err;
286 	}
287 
288 	cq->resize_buf->cqe = entries - 1;
289 
290 	return 0;
291 }
292 
293 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
294 				   int entries, struct ib_udata *udata)
295 {
296 	struct mlx4_ib_resize_cq ucmd;
297 	int err;
298 
299 	if (cq->resize_umem)
300 		return -EBUSY;
301 
302 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
303 		return -EFAULT;
304 
305 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
306 	if (!cq->resize_buf)
307 		return -ENOMEM;
308 
309 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
310 				  &cq->resize_umem, ucmd.buf_addr, entries);
311 	if (err) {
312 		kfree(cq->resize_buf);
313 		cq->resize_buf = NULL;
314 		return err;
315 	}
316 
317 	cq->resize_buf->cqe = entries - 1;
318 
319 	return 0;
320 }
321 
322 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
323 {
324 	u32 i;
325 
326 	i = cq->mcq.cons_index;
327 	while (get_sw_cqe(cq, i & cq->ibcq.cqe))
328 		++i;
329 
330 	return i - cq->mcq.cons_index;
331 }
332 
333 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
334 {
335 	struct mlx4_cqe *cqe, *new_cqe;
336 	int i;
337 	int cqe_size = cq->buf.entry_size;
338 	int cqe_inc = cqe_size == 64 ? 1 : 0;
339 
340 	i = cq->mcq.cons_index;
341 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
342 	cqe += cqe_inc;
343 
344 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
345 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
346 					   (i + 1) & cq->resize_buf->cqe);
347 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
348 		new_cqe += cqe_inc;
349 
350 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
351 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
352 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
353 		cqe += cqe_inc;
354 	}
355 	++cq->mcq.cons_index;
356 }
357 
358 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
359 {
360 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
361 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
362 	struct mlx4_mtt mtt;
363 	int outst_cqe;
364 	int err;
365 
366 	mutex_lock(&cq->resize_mutex);
367 
368 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
369 		err = -EINVAL;
370 		goto out;
371 	}
372 
373 	entries = roundup_pow_of_two(entries + 1);
374 	if (entries == ibcq->cqe + 1) {
375 		err = 0;
376 		goto out;
377 	}
378 
379 	if (ibcq->uobject) {
380 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
381 		if (err)
382 			goto out;
383 	} else {
384 		/* Can't be smaller than the number of outstanding CQEs */
385 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
386 		if (entries < outst_cqe + 1) {
387 			err = 0;
388 			goto out;
389 		}
390 
391 		err = mlx4_alloc_resize_buf(dev, cq, entries);
392 		if (err)
393 			goto out;
394 	}
395 
396 	mtt = cq->buf.mtt;
397 
398 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
399 	if (err)
400 		goto err_buf;
401 
402 	mlx4_mtt_cleanup(dev->dev, &mtt);
403 	if (ibcq->uobject) {
404 		cq->buf      = cq->resize_buf->buf;
405 		cq->ibcq.cqe = cq->resize_buf->cqe;
406 		ib_umem_release(cq->umem);
407 		cq->umem     = cq->resize_umem;
408 
409 		kfree(cq->resize_buf);
410 		cq->resize_buf = NULL;
411 		cq->resize_umem = NULL;
412 	} else {
413 		struct mlx4_ib_cq_buf tmp_buf;
414 		int tmp_cqe = 0;
415 
416 		spin_lock_irq(&cq->lock);
417 		if (cq->resize_buf) {
418 			mlx4_ib_cq_resize_copy_cqes(cq);
419 			tmp_buf = cq->buf;
420 			tmp_cqe = cq->ibcq.cqe;
421 			cq->buf      = cq->resize_buf->buf;
422 			cq->ibcq.cqe = cq->resize_buf->cqe;
423 
424 			kfree(cq->resize_buf);
425 			cq->resize_buf = NULL;
426 		}
427 		spin_unlock_irq(&cq->lock);
428 
429 		if (tmp_cqe)
430 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
431 	}
432 
433 	goto out;
434 
435 err_buf:
436 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
437 	if (!ibcq->uobject)
438 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
439 				    cq->resize_buf->cqe);
440 
441 	kfree(cq->resize_buf);
442 	cq->resize_buf = NULL;
443 
444 	if (cq->resize_umem) {
445 		ib_umem_release(cq->resize_umem);
446 		cq->resize_umem = NULL;
447 	}
448 
449 out:
450 	mutex_unlock(&cq->resize_mutex);
451 
452 	return err;
453 }
454 
455 int mlx4_ib_destroy_cq(struct ib_cq *cq)
456 {
457 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
458 	struct mlx4_ib_cq *mcq = to_mcq(cq);
459 
460 	mlx4_cq_free(dev->dev, &mcq->mcq);
461 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
462 
463 	if (cq->uobject) {
464 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
465 		ib_umem_release(mcq->umem);
466 	} else {
467 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
468 		mlx4_db_free(dev->dev, &mcq->db);
469 	}
470 
471 	kfree(mcq);
472 
473 	return 0;
474 }
475 
476 static void dump_cqe(void *cqe)
477 {
478 	__be32 *buf = cqe;
479 
480 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
481 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
482 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
483 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
484 }
485 
486 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
487 				     struct ib_wc *wc)
488 {
489 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
490 		pr_debug("local QP operation err "
491 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
492 		       "opcode = %02x)\n",
493 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
494 		       cqe->vendor_err_syndrome,
495 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
496 		dump_cqe(cqe);
497 	}
498 
499 	switch (cqe->syndrome) {
500 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
501 		wc->status = IB_WC_LOC_LEN_ERR;
502 		break;
503 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
504 		wc->status = IB_WC_LOC_QP_OP_ERR;
505 		break;
506 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
507 		wc->status = IB_WC_LOC_PROT_ERR;
508 		break;
509 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
510 		wc->status = IB_WC_WR_FLUSH_ERR;
511 		break;
512 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
513 		wc->status = IB_WC_MW_BIND_ERR;
514 		break;
515 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
516 		wc->status = IB_WC_BAD_RESP_ERR;
517 		break;
518 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
519 		wc->status = IB_WC_LOC_ACCESS_ERR;
520 		break;
521 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
522 		wc->status = IB_WC_REM_INV_REQ_ERR;
523 		break;
524 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
525 		wc->status = IB_WC_REM_ACCESS_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
528 		wc->status = IB_WC_REM_OP_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
531 		wc->status = IB_WC_RETRY_EXC_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
534 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
537 		wc->status = IB_WC_REM_ABORT_ERR;
538 		break;
539 	default:
540 		wc->status = IB_WC_GENERAL_ERR;
541 		break;
542 	}
543 
544 	wc->vendor_err = cqe->vendor_err_syndrome;
545 }
546 
547 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
548 {
549 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
550 				      MLX4_CQE_STATUS_IPV4F     |
551 				      MLX4_CQE_STATUS_IPV4OPT   |
552 				      MLX4_CQE_STATUS_IPV6      |
553 				      MLX4_CQE_STATUS_IPOK)) ==
554 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
555 			    MLX4_CQE_STATUS_IPOK))              &&
556 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
557 				      MLX4_CQE_STATUS_TCP))     &&
558 		checksum == cpu_to_be16(0xffff);
559 }
560 
561 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
562 			   unsigned tail, struct mlx4_cqe *cqe)
563 {
564 	struct mlx4_ib_proxy_sqp_hdr *hdr;
565 
566 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
567 				   qp->sqp_proxy_rcv[tail].map,
568 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
569 				   DMA_FROM_DEVICE);
570 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
571 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
572 	wc->slid	= be16_to_cpu(hdr->tun.slid_mac_47_32);
573 	wc->sl		= (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
574 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
575 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
576 	wc->dlid_path_bits = 0;
577 
578 	return 0;
579 }
580 
581 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
582 			    struct mlx4_ib_qp **cur_qp,
583 			    struct ib_wc *wc)
584 {
585 	struct mlx4_cqe *cqe;
586 	struct mlx4_qp *mqp;
587 	struct mlx4_ib_wq *wq;
588 	struct mlx4_ib_srq *srq;
589 	struct mlx4_srq *msrq = NULL;
590 	int is_send;
591 	int is_error;
592 	u32 g_mlpath_rqpn;
593 	u16 wqe_ctr;
594 	unsigned tail = 0;
595 
596 repoll:
597 	cqe = next_cqe_sw(cq);
598 	if (!cqe)
599 		return -EAGAIN;
600 
601 	if (cq->buf.entry_size == 64)
602 		cqe++;
603 
604 	++cq->mcq.cons_index;
605 
606 	/*
607 	 * Make sure we read CQ entry contents after we've checked the
608 	 * ownership bit.
609 	 */
610 	rmb();
611 
612 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
613 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
614 		MLX4_CQE_OPCODE_ERROR;
615 
616 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
617 		     is_send)) {
618 		pr_warn("Completion for NOP opcode detected!\n");
619 		return -EINVAL;
620 	}
621 
622 	/* Resize CQ in progress */
623 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
624 		if (cq->resize_buf) {
625 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
626 
627 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
628 			cq->buf      = cq->resize_buf->buf;
629 			cq->ibcq.cqe = cq->resize_buf->cqe;
630 
631 			kfree(cq->resize_buf);
632 			cq->resize_buf = NULL;
633 		}
634 
635 		goto repoll;
636 	}
637 
638 	if (!*cur_qp ||
639 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
640 		/*
641 		 * We do not have to take the QP table lock here,
642 		 * because CQs will be locked while QPs are removed
643 		 * from the table.
644 		 */
645 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
646 				       be32_to_cpu(cqe->vlan_my_qpn));
647 		if (unlikely(!mqp)) {
648 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
649 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
650 			return -EINVAL;
651 		}
652 
653 		*cur_qp = to_mibqp(mqp);
654 	}
655 
656 	wc->qp = &(*cur_qp)->ibqp;
657 
658 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
659 		u32 srq_num;
660 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
661 		srq_num       = g_mlpath_rqpn & 0xffffff;
662 		/* SRQ is also in the radix tree */
663 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
664 				       srq_num);
665 		if (unlikely(!msrq)) {
666 			pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
667 				cq->mcq.cqn, srq_num);
668 			return -EINVAL;
669 		}
670 	}
671 
672 	if (is_send) {
673 		wq = &(*cur_qp)->sq;
674 		if (!(*cur_qp)->sq_signal_bits) {
675 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
676 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
677 		}
678 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
679 		++wq->tail;
680 	} else if ((*cur_qp)->ibqp.srq) {
681 		srq = to_msrq((*cur_qp)->ibqp.srq);
682 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
683 		wc->wr_id = srq->wrid[wqe_ctr];
684 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
685 	} else if (msrq) {
686 		srq = to_mibsrq(msrq);
687 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
688 		wc->wr_id = srq->wrid[wqe_ctr];
689 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
690 	} else {
691 		wq	  = &(*cur_qp)->rq;
692 		tail	  = wq->tail & (wq->wqe_cnt - 1);
693 		wc->wr_id = wq->wrid[tail];
694 		++wq->tail;
695 	}
696 
697 	if (unlikely(is_error)) {
698 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
699 		return 0;
700 	}
701 
702 	wc->status = IB_WC_SUCCESS;
703 
704 	if (is_send) {
705 		wc->wc_flags = 0;
706 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
707 		case MLX4_OPCODE_RDMA_WRITE_IMM:
708 			wc->wc_flags |= IB_WC_WITH_IMM;
709 		case MLX4_OPCODE_RDMA_WRITE:
710 			wc->opcode    = IB_WC_RDMA_WRITE;
711 			break;
712 		case MLX4_OPCODE_SEND_IMM:
713 			wc->wc_flags |= IB_WC_WITH_IMM;
714 		case MLX4_OPCODE_SEND:
715 		case MLX4_OPCODE_SEND_INVAL:
716 			wc->opcode    = IB_WC_SEND;
717 			break;
718 		case MLX4_OPCODE_RDMA_READ:
719 			wc->opcode    = IB_WC_RDMA_READ;
720 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
721 			break;
722 		case MLX4_OPCODE_ATOMIC_CS:
723 			wc->opcode    = IB_WC_COMP_SWAP;
724 			wc->byte_len  = 8;
725 			break;
726 		case MLX4_OPCODE_ATOMIC_FA:
727 			wc->opcode    = IB_WC_FETCH_ADD;
728 			wc->byte_len  = 8;
729 			break;
730 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
731 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
732 			wc->byte_len  = 8;
733 			break;
734 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
735 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
736 			wc->byte_len  = 8;
737 			break;
738 		case MLX4_OPCODE_BIND_MW:
739 			wc->opcode    = IB_WC_BIND_MW;
740 			break;
741 		case MLX4_OPCODE_LSO:
742 			wc->opcode    = IB_WC_LSO;
743 			break;
744 		case MLX4_OPCODE_FMR:
745 			wc->opcode    = IB_WC_FAST_REG_MR;
746 			break;
747 		case MLX4_OPCODE_LOCAL_INVAL:
748 			wc->opcode    = IB_WC_LOCAL_INV;
749 			break;
750 		}
751 	} else {
752 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
753 
754 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
755 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
756 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
757 			wc->wc_flags	= IB_WC_WITH_IMM;
758 			wc->ex.imm_data = cqe->immed_rss_invalid;
759 			break;
760 		case MLX4_RECV_OPCODE_SEND_INVAL:
761 			wc->opcode	= IB_WC_RECV;
762 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
763 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
764 			break;
765 		case MLX4_RECV_OPCODE_SEND:
766 			wc->opcode   = IB_WC_RECV;
767 			wc->wc_flags = 0;
768 			break;
769 		case MLX4_RECV_OPCODE_SEND_IMM:
770 			wc->opcode	= IB_WC_RECV;
771 			wc->wc_flags	= IB_WC_WITH_IMM;
772 			wc->ex.imm_data = cqe->immed_rss_invalid;
773 			break;
774 		}
775 
776 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
777 			if ((*cur_qp)->mlx4_ib_qp_type &
778 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
779 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
780 				return use_tunnel_data(*cur_qp, cq, wc, tail, cqe);
781 		}
782 
783 		wc->slid	   = be16_to_cpu(cqe->rlid);
784 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
785 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
786 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
787 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
788 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
789 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
790 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
791 		if (rdma_port_get_link_layer(wc->qp->device,
792 				(*cur_qp)->port) == IB_LINK_LAYER_ETHERNET)
793 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
794 		else
795 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
796 	}
797 
798 	return 0;
799 }
800 
801 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
802 {
803 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
804 	struct mlx4_ib_qp *cur_qp = NULL;
805 	unsigned long flags;
806 	int npolled;
807 	int err = 0;
808 
809 	spin_lock_irqsave(&cq->lock, flags);
810 
811 	for (npolled = 0; npolled < num_entries; ++npolled) {
812 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
813 		if (err)
814 			break;
815 	}
816 
817 	mlx4_cq_set_ci(&cq->mcq);
818 
819 	spin_unlock_irqrestore(&cq->lock, flags);
820 
821 	if (err == 0 || err == -EAGAIN)
822 		return npolled;
823 	else
824 		return err;
825 }
826 
827 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
828 {
829 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
830 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
831 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
832 		    to_mdev(ibcq->device)->uar_map,
833 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
834 
835 	return 0;
836 }
837 
838 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
839 {
840 	u32 prod_index;
841 	int nfreed = 0;
842 	struct mlx4_cqe *cqe, *dest;
843 	u8 owner_bit;
844 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
845 
846 	/*
847 	 * First we need to find the current producer index, so we
848 	 * know where to start cleaning from.  It doesn't matter if HW
849 	 * adds new entries after this loop -- the QP we're worried
850 	 * about is already in RESET, so the new entries won't come
851 	 * from our QP and therefore don't need to be checked.
852 	 */
853 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
854 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
855 			break;
856 
857 	/*
858 	 * Now sweep backwards through the CQ, removing CQ entries
859 	 * that match our QP by copying older entries on top of them.
860 	 */
861 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
862 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
863 		cqe += cqe_inc;
864 
865 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
866 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
867 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
868 			++nfreed;
869 		} else if (nfreed) {
870 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
871 			dest += cqe_inc;
872 
873 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
874 			memcpy(dest, cqe, sizeof *cqe);
875 			dest->owner_sr_opcode = owner_bit |
876 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
877 		}
878 	}
879 
880 	if (nfreed) {
881 		cq->mcq.cons_index += nfreed;
882 		/*
883 		 * Make sure update of buffer contents is done before
884 		 * updating consumer index.
885 		 */
886 		wmb();
887 		mlx4_cq_set_ci(&cq->mcq);
888 	}
889 }
890 
891 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
892 {
893 	spin_lock_irq(&cq->lock);
894 	__mlx4_ib_cq_clean(cq, qpn, srq);
895 	spin_unlock_irq(&cq->lock);
896 }
897