xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision baa7eb025ab14f3cba2e35c0a8648f9c9f01d24f)
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 		printk(KERN_WARNING "mlx4_ib: 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 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
226 			    cq->db.dma, &cq->mcq, vector, 0);
227 	if (err)
228 		goto err_dbmap;
229 
230 	cq->mcq.comp  = mlx4_ib_cq_comp;
231 	cq->mcq.event = mlx4_ib_cq_event;
232 
233 	if (context)
234 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
235 			err = -EFAULT;
236 			goto err_dbmap;
237 		}
238 
239 	return &cq->ibcq;
240 
241 err_dbmap:
242 	if (context)
243 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
244 
245 err_mtt:
246 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
247 
248 	if (context)
249 		ib_umem_release(cq->umem);
250 	else
251 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
252 
253 err_db:
254 	if (!context)
255 		mlx4_db_free(dev->dev, &cq->db);
256 
257 err_cq:
258 	kfree(cq);
259 
260 	return ERR_PTR(err);
261 }
262 
263 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
264 				  int entries)
265 {
266 	int err;
267 
268 	if (cq->resize_buf)
269 		return -EBUSY;
270 
271 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
272 	if (!cq->resize_buf)
273 		return -ENOMEM;
274 
275 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
276 	if (err) {
277 		kfree(cq->resize_buf);
278 		cq->resize_buf = NULL;
279 		return err;
280 	}
281 
282 	cq->resize_buf->cqe = entries - 1;
283 
284 	return 0;
285 }
286 
287 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
288 				   int entries, struct ib_udata *udata)
289 {
290 	struct mlx4_ib_resize_cq ucmd;
291 	int err;
292 
293 	if (cq->resize_umem)
294 		return -EBUSY;
295 
296 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
297 		return -EFAULT;
298 
299 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
300 	if (!cq->resize_buf)
301 		return -ENOMEM;
302 
303 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
304 				  &cq->resize_umem, ucmd.buf_addr, entries);
305 	if (err) {
306 		kfree(cq->resize_buf);
307 		cq->resize_buf = NULL;
308 		return err;
309 	}
310 
311 	cq->resize_buf->cqe = entries - 1;
312 
313 	return 0;
314 }
315 
316 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
317 {
318 	u32 i;
319 
320 	i = cq->mcq.cons_index;
321 	while (get_sw_cqe(cq, i & cq->ibcq.cqe))
322 		++i;
323 
324 	return i - cq->mcq.cons_index;
325 }
326 
327 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
328 {
329 	struct mlx4_cqe *cqe, *new_cqe;
330 	int i;
331 
332 	i = cq->mcq.cons_index;
333 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
334 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
335 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
336 					   (i + 1) & cq->resize_buf->cqe);
337 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), sizeof(struct mlx4_cqe));
338 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
339 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
340 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
341 	}
342 	++cq->mcq.cons_index;
343 }
344 
345 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
346 {
347 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
348 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
349 	struct mlx4_mtt mtt;
350 	int outst_cqe;
351 	int err;
352 
353 	mutex_lock(&cq->resize_mutex);
354 
355 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
356 		err = -EINVAL;
357 		goto out;
358 	}
359 
360 	entries = roundup_pow_of_two(entries + 1);
361 	if (entries == ibcq->cqe + 1) {
362 		err = 0;
363 		goto out;
364 	}
365 
366 	if (ibcq->uobject) {
367 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
368 		if (err)
369 			goto out;
370 	} else {
371 		/* Can't be smaller than the number of outstanding CQEs */
372 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
373 		if (entries < outst_cqe + 1) {
374 			err = 0;
375 			goto out;
376 		}
377 
378 		err = mlx4_alloc_resize_buf(dev, cq, entries);
379 		if (err)
380 			goto out;
381 	}
382 
383 	mtt = cq->buf.mtt;
384 
385 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
386 	if (err)
387 		goto err_buf;
388 
389 	mlx4_mtt_cleanup(dev->dev, &mtt);
390 	if (ibcq->uobject) {
391 		cq->buf      = cq->resize_buf->buf;
392 		cq->ibcq.cqe = cq->resize_buf->cqe;
393 		ib_umem_release(cq->umem);
394 		cq->umem     = cq->resize_umem;
395 
396 		kfree(cq->resize_buf);
397 		cq->resize_buf = NULL;
398 		cq->resize_umem = NULL;
399 	} else {
400 		spin_lock_irq(&cq->lock);
401 		if (cq->resize_buf) {
402 			mlx4_ib_cq_resize_copy_cqes(cq);
403 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
404 			cq->buf      = cq->resize_buf->buf;
405 			cq->ibcq.cqe = cq->resize_buf->cqe;
406 
407 			kfree(cq->resize_buf);
408 			cq->resize_buf = NULL;
409 		}
410 		spin_unlock_irq(&cq->lock);
411 	}
412 
413 	goto out;
414 
415 err_buf:
416 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
417 	if (!ibcq->uobject)
418 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
419 				    cq->resize_buf->cqe);
420 
421 	kfree(cq->resize_buf);
422 	cq->resize_buf = NULL;
423 
424 	if (cq->resize_umem) {
425 		ib_umem_release(cq->resize_umem);
426 		cq->resize_umem = NULL;
427 	}
428 
429 out:
430 	mutex_unlock(&cq->resize_mutex);
431 	return err;
432 }
433 
434 int mlx4_ib_destroy_cq(struct ib_cq *cq)
435 {
436 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
437 	struct mlx4_ib_cq *mcq = to_mcq(cq);
438 
439 	mlx4_cq_free(dev->dev, &mcq->mcq);
440 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
441 
442 	if (cq->uobject) {
443 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
444 		ib_umem_release(mcq->umem);
445 	} else {
446 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
447 		mlx4_db_free(dev->dev, &mcq->db);
448 	}
449 
450 	kfree(mcq);
451 
452 	return 0;
453 }
454 
455 static void dump_cqe(void *cqe)
456 {
457 	__be32 *buf = cqe;
458 
459 	printk(KERN_DEBUG "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
460 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
461 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
462 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
463 }
464 
465 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
466 				     struct ib_wc *wc)
467 {
468 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
469 		printk(KERN_DEBUG "local QP operation err "
470 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
471 		       "opcode = %02x)\n",
472 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
473 		       cqe->vendor_err_syndrome,
474 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
475 		dump_cqe(cqe);
476 	}
477 
478 	switch (cqe->syndrome) {
479 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
480 		wc->status = IB_WC_LOC_LEN_ERR;
481 		break;
482 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
483 		wc->status = IB_WC_LOC_QP_OP_ERR;
484 		break;
485 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
486 		wc->status = IB_WC_LOC_PROT_ERR;
487 		break;
488 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
489 		wc->status = IB_WC_WR_FLUSH_ERR;
490 		break;
491 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
492 		wc->status = IB_WC_MW_BIND_ERR;
493 		break;
494 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
495 		wc->status = IB_WC_BAD_RESP_ERR;
496 		break;
497 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
498 		wc->status = IB_WC_LOC_ACCESS_ERR;
499 		break;
500 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
501 		wc->status = IB_WC_REM_INV_REQ_ERR;
502 		break;
503 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
504 		wc->status = IB_WC_REM_ACCESS_ERR;
505 		break;
506 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
507 		wc->status = IB_WC_REM_OP_ERR;
508 		break;
509 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
510 		wc->status = IB_WC_RETRY_EXC_ERR;
511 		break;
512 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
513 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
514 		break;
515 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
516 		wc->status = IB_WC_REM_ABORT_ERR;
517 		break;
518 	default:
519 		wc->status = IB_WC_GENERAL_ERR;
520 		break;
521 	}
522 
523 	wc->vendor_err = cqe->vendor_err_syndrome;
524 }
525 
526 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
527 {
528 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
529 				      MLX4_CQE_STATUS_IPV4F     |
530 				      MLX4_CQE_STATUS_IPV4OPT   |
531 				      MLX4_CQE_STATUS_IPV6      |
532 				      MLX4_CQE_STATUS_IPOK)) ==
533 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
534 			    MLX4_CQE_STATUS_IPOK))              &&
535 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
536 				      MLX4_CQE_STATUS_TCP))     &&
537 		checksum == cpu_to_be16(0xffff);
538 }
539 
540 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
541 			    struct mlx4_ib_qp **cur_qp,
542 			    struct ib_wc *wc)
543 {
544 	struct mlx4_cqe *cqe;
545 	struct mlx4_qp *mqp;
546 	struct mlx4_ib_wq *wq;
547 	struct mlx4_ib_srq *srq;
548 	int is_send;
549 	int is_error;
550 	u32 g_mlpath_rqpn;
551 	u16 wqe_ctr;
552 
553 repoll:
554 	cqe = next_cqe_sw(cq);
555 	if (!cqe)
556 		return -EAGAIN;
557 
558 	++cq->mcq.cons_index;
559 
560 	/*
561 	 * Make sure we read CQ entry contents after we've checked the
562 	 * ownership bit.
563 	 */
564 	rmb();
565 
566 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
567 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
568 		MLX4_CQE_OPCODE_ERROR;
569 
570 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
571 		     is_send)) {
572 		printk(KERN_WARNING "Completion for NOP opcode detected!\n");
573 		return -EINVAL;
574 	}
575 
576 	/* Resize CQ in progress */
577 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
578 		if (cq->resize_buf) {
579 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
580 
581 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
582 			cq->buf      = cq->resize_buf->buf;
583 			cq->ibcq.cqe = cq->resize_buf->cqe;
584 
585 			kfree(cq->resize_buf);
586 			cq->resize_buf = NULL;
587 		}
588 
589 		goto repoll;
590 	}
591 
592 	if (!*cur_qp ||
593 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
594 		/*
595 		 * We do not have to take the QP table lock here,
596 		 * because CQs will be locked while QPs are removed
597 		 * from the table.
598 		 */
599 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
600 				       be32_to_cpu(cqe->vlan_my_qpn));
601 		if (unlikely(!mqp)) {
602 			printk(KERN_WARNING "CQ %06x with entry for unknown QPN %06x\n",
603 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
604 			return -EINVAL;
605 		}
606 
607 		*cur_qp = to_mibqp(mqp);
608 	}
609 
610 	wc->qp = &(*cur_qp)->ibqp;
611 
612 	if (is_send) {
613 		wq = &(*cur_qp)->sq;
614 		if (!(*cur_qp)->sq_signal_bits) {
615 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
616 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
617 		}
618 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
619 		++wq->tail;
620 	} else if ((*cur_qp)->ibqp.srq) {
621 		srq = to_msrq((*cur_qp)->ibqp.srq);
622 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
623 		wc->wr_id = srq->wrid[wqe_ctr];
624 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
625 	} else {
626 		wq	  = &(*cur_qp)->rq;
627 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
628 		++wq->tail;
629 	}
630 
631 	if (unlikely(is_error)) {
632 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
633 		return 0;
634 	}
635 
636 	wc->status = IB_WC_SUCCESS;
637 
638 	if (is_send) {
639 		wc->wc_flags = 0;
640 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
641 		case MLX4_OPCODE_RDMA_WRITE_IMM:
642 			wc->wc_flags |= IB_WC_WITH_IMM;
643 		case MLX4_OPCODE_RDMA_WRITE:
644 			wc->opcode    = IB_WC_RDMA_WRITE;
645 			break;
646 		case MLX4_OPCODE_SEND_IMM:
647 			wc->wc_flags |= IB_WC_WITH_IMM;
648 		case MLX4_OPCODE_SEND:
649 		case MLX4_OPCODE_SEND_INVAL:
650 			wc->opcode    = IB_WC_SEND;
651 			break;
652 		case MLX4_OPCODE_RDMA_READ:
653 			wc->opcode    = IB_WC_RDMA_READ;
654 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
655 			break;
656 		case MLX4_OPCODE_ATOMIC_CS:
657 			wc->opcode    = IB_WC_COMP_SWAP;
658 			wc->byte_len  = 8;
659 			break;
660 		case MLX4_OPCODE_ATOMIC_FA:
661 			wc->opcode    = IB_WC_FETCH_ADD;
662 			wc->byte_len  = 8;
663 			break;
664 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
665 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
666 			wc->byte_len  = 8;
667 			break;
668 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
669 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
670 			wc->byte_len  = 8;
671 			break;
672 		case MLX4_OPCODE_BIND_MW:
673 			wc->opcode    = IB_WC_BIND_MW;
674 			break;
675 		case MLX4_OPCODE_LSO:
676 			wc->opcode    = IB_WC_LSO;
677 			break;
678 		case MLX4_OPCODE_FMR:
679 			wc->opcode    = IB_WC_FAST_REG_MR;
680 			break;
681 		case MLX4_OPCODE_LOCAL_INVAL:
682 			wc->opcode    = IB_WC_LOCAL_INV;
683 			break;
684 		}
685 	} else {
686 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
687 
688 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
689 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
690 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
691 			wc->wc_flags	= IB_WC_WITH_IMM;
692 			wc->ex.imm_data = cqe->immed_rss_invalid;
693 			break;
694 		case MLX4_RECV_OPCODE_SEND_INVAL:
695 			wc->opcode	= IB_WC_RECV;
696 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
697 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
698 			break;
699 		case MLX4_RECV_OPCODE_SEND:
700 			wc->opcode   = IB_WC_RECV;
701 			wc->wc_flags = 0;
702 			break;
703 		case MLX4_RECV_OPCODE_SEND_IMM:
704 			wc->opcode	= IB_WC_RECV;
705 			wc->wc_flags	= IB_WC_WITH_IMM;
706 			wc->ex.imm_data = cqe->immed_rss_invalid;
707 			break;
708 		}
709 
710 		wc->slid	   = be16_to_cpu(cqe->rlid);
711 		wc->sl		   = be16_to_cpu(cqe->sl_vid) >> 12;
712 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
713 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
714 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
715 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
716 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
717 		wc->csum_ok	   = mlx4_ib_ipoib_csum_ok(cqe->status, cqe->checksum);
718 	}
719 
720 	return 0;
721 }
722 
723 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
724 {
725 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
726 	struct mlx4_ib_qp *cur_qp = NULL;
727 	unsigned long flags;
728 	int npolled;
729 	int err = 0;
730 
731 	spin_lock_irqsave(&cq->lock, flags);
732 
733 	for (npolled = 0; npolled < num_entries; ++npolled) {
734 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
735 		if (err)
736 			break;
737 	}
738 
739 	if (npolled)
740 		mlx4_cq_set_ci(&cq->mcq);
741 
742 	spin_unlock_irqrestore(&cq->lock, flags);
743 
744 	if (err == 0 || err == -EAGAIN)
745 		return npolled;
746 	else
747 		return err;
748 }
749 
750 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
751 {
752 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
753 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
754 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
755 		    to_mdev(ibcq->device)->uar_map,
756 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
757 
758 	return 0;
759 }
760 
761 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
762 {
763 	u32 prod_index;
764 	int nfreed = 0;
765 	struct mlx4_cqe *cqe, *dest;
766 	u8 owner_bit;
767 
768 	/*
769 	 * First we need to find the current producer index, so we
770 	 * know where to start cleaning from.  It doesn't matter if HW
771 	 * adds new entries after this loop -- the QP we're worried
772 	 * about is already in RESET, so the new entries won't come
773 	 * from our QP and therefore don't need to be checked.
774 	 */
775 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
776 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
777 			break;
778 
779 	/*
780 	 * Now sweep backwards through the CQ, removing CQ entries
781 	 * that match our QP by copying older entries on top of them.
782 	 */
783 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
784 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
785 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
786 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
787 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
788 			++nfreed;
789 		} else if (nfreed) {
790 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
791 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
792 			memcpy(dest, cqe, sizeof *cqe);
793 			dest->owner_sr_opcode = owner_bit |
794 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
795 		}
796 	}
797 
798 	if (nfreed) {
799 		cq->mcq.cons_index += nfreed;
800 		/*
801 		 * Make sure update of buffer contents is done before
802 		 * updating consumer index.
803 		 */
804 		wmb();
805 		mlx4_cq_set_ci(&cq->mcq);
806 	}
807 }
808 
809 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
810 {
811 	spin_lock_irq(&cq->lock);
812 	__mlx4_ib_cq_clean(cq, qpn, srq);
813 	spin_unlock_irq(&cq->lock);
814 }
815