xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_cq.c (revision d6eb98610fa65663bf0df4574b7cb2c5c4ffda71)
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 <dev/mlx4/cq.h>
35 #include <dev/mlx4/qp.h>
36 #include <dev/mlx4/srq.h>
37 #include <dev/mlx4/driver.h>
38 #include <linux/slab.h>
39 
40 #include "mlx4_ib.h"
41 #include <rdma/mlx4-abi.h>
42 
43 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
44 {
45 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
46 	ibcq->comp_handler(ibcq, ibcq->cq_context);
47 }
48 
49 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
50 {
51 	struct ib_event event;
52 	struct ib_cq *ibcq;
53 
54 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
55 		pr_warn("Unexpected event type %d "
56 		       "on CQ %06x\n", type, cq->cqn);
57 		return;
58 	}
59 
60 	ibcq = &to_mibcq(cq)->ibcq;
61 	if (ibcq->event_handler) {
62 		event.device     = ibcq->device;
63 		event.event      = IB_EVENT_CQ_ERR;
64 		event.element.cq = ibcq;
65 		ibcq->event_handler(&event, ibcq->cq_context);
66 	}
67 }
68 
69 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
70 {
71 	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
72 }
73 
74 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
75 {
76 	return get_cqe_from_buf(&cq->buf, n);
77 }
78 
79 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
80 {
81 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
82 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
83 
84 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
85 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
86 }
87 
88 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
89 {
90 	return get_sw_cqe(cq, cq->mcq.cons_index);
91 }
92 
93 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
94 {
95 	struct mlx4_ib_cq *mcq = to_mcq(cq);
96 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
97 
98 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
99 }
100 
101 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
102 {
103 	int err;
104 
105 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
106 			     PAGE_SIZE * 2, &buf->buf, GFP_KERNEL);
107 
108 	if (err)
109 		goto out;
110 
111 	buf->entry_size = dev->dev->caps.cqe_size;
112 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
113 				    &buf->mtt);
114 	if (err)
115 		goto err_buf;
116 
117 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf, GFP_KERNEL);
118 	if (err)
119 		goto err_mtt;
120 
121 	return 0;
122 
123 err_mtt:
124 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
125 
126 err_buf:
127 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
128 
129 out:
130 	return err;
131 }
132 
133 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
134 {
135 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
136 }
137 
138 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
139 			       struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
140 			       u64 buf_addr, int cqe)
141 {
142 	int err;
143 	int cqe_size = dev->dev->caps.cqe_size;
144 
145 	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
146 			    IB_ACCESS_LOCAL_WRITE, 1);
147 	if (IS_ERR(*umem))
148 		return PTR_ERR(*umem);
149 
150 	err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
151 			    ilog2((*umem)->page_size), &buf->mtt);
152 	if (err)
153 		goto err_buf;
154 
155 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
156 	if (err)
157 		goto err_mtt;
158 
159 	return 0;
160 
161 err_mtt:
162 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
163 
164 err_buf:
165 	ib_umem_release(*umem);
166 
167 	return err;
168 }
169 
170 #define CQ_CREATE_FLAGS_SUPPORTED IB_CQ_FLAGS_TIMESTAMP_COMPLETION
171 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
172 				const struct ib_cq_init_attr *attr,
173 				struct ib_ucontext *context,
174 				struct ib_udata *udata)
175 {
176 	int entries = attr->cqe;
177 	int vector = attr->comp_vector;
178 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
179 	struct mlx4_ib_cq *cq;
180 	struct mlx4_uar *uar;
181 	int err;
182 
183 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
184 		return ERR_PTR(-EINVAL);
185 
186 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
187 		return ERR_PTR(-EINVAL);
188 
189 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
190 	if (!cq)
191 		return ERR_PTR(-ENOMEM);
192 
193 	entries      = roundup_pow_of_two(entries + 1);
194 	cq->ibcq.cqe = entries - 1;
195 	mutex_init(&cq->resize_mutex);
196 	spin_lock_init(&cq->lock);
197 	cq->resize_buf = NULL;
198 	cq->resize_umem = NULL;
199 	cq->create_flags = attr->flags;
200 	INIT_LIST_HEAD(&cq->send_qp_list);
201 	INIT_LIST_HEAD(&cq->recv_qp_list);
202 
203 	if (context) {
204 		struct mlx4_ib_create_cq ucmd;
205 
206 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
207 			err = -EFAULT;
208 			goto err_cq;
209 		}
210 
211 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
212 					  ucmd.buf_addr, entries);
213 		if (err)
214 			goto err_cq;
215 
216 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
217 					  &cq->db);
218 		if (err)
219 			goto err_mtt;
220 
221 		uar = &to_mucontext(context)->uar;
222 	} else {
223 		err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
224 		if (err)
225 			goto err_cq;
226 
227 		cq->mcq.set_ci_db  = cq->db.db;
228 		cq->mcq.arm_db     = cq->db.db + 1;
229 		*cq->mcq.set_ci_db = 0;
230 		*cq->mcq.arm_db    = 0;
231 
232 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
233 		if (err)
234 			goto err_db;
235 
236 		uar = &dev->priv_uar;
237 	}
238 
239 	if (dev->eq_table)
240 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
241 
242 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
243 			    cq->db.dma, &cq->mcq, vector, 0,
244 			    !!(cq->create_flags & IB_CQ_FLAGS_TIMESTAMP_COMPLETION));
245 	if (err)
246 		goto err_dbmap;
247 
248 	cq->mcq.comp = mlx4_ib_cq_comp;
249 	cq->mcq.event = mlx4_ib_cq_event;
250 
251 	if (context)
252 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
253 			err = -EFAULT;
254 			goto err_cq_free;
255 		}
256 
257 	return &cq->ibcq;
258 
259 err_cq_free:
260 	mlx4_cq_free(dev->dev, &cq->mcq);
261 
262 err_dbmap:
263 	if (context)
264 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
265 
266 err_mtt:
267 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
268 
269 	if (context)
270 		ib_umem_release(cq->umem);
271 	else
272 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
273 
274 err_db:
275 	if (!context)
276 		mlx4_db_free(dev->dev, &cq->db);
277 
278 err_cq:
279 	kfree(cq);
280 
281 	return ERR_PTR(err);
282 }
283 
284 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
285 				  int entries)
286 {
287 	int err;
288 
289 	if (cq->resize_buf)
290 		return -EBUSY;
291 
292 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
293 	if (!cq->resize_buf)
294 		return -ENOMEM;
295 
296 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
297 	if (err) {
298 		kfree(cq->resize_buf);
299 		cq->resize_buf = NULL;
300 		return err;
301 	}
302 
303 	cq->resize_buf->cqe = entries - 1;
304 
305 	return 0;
306 }
307 
308 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
309 				   int entries, struct ib_udata *udata)
310 {
311 	struct mlx4_ib_resize_cq ucmd;
312 	int err;
313 
314 	if (cq->resize_umem)
315 		return -EBUSY;
316 
317 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
318 		return -EFAULT;
319 
320 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
321 	if (!cq->resize_buf)
322 		return -ENOMEM;
323 
324 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
325 				  &cq->resize_umem, ucmd.buf_addr, entries);
326 	if (err) {
327 		kfree(cq->resize_buf);
328 		cq->resize_buf = NULL;
329 		return err;
330 	}
331 
332 	cq->resize_buf->cqe = entries - 1;
333 
334 	return 0;
335 }
336 
337 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
338 {
339 	u32 i;
340 
341 	i = cq->mcq.cons_index;
342 	while (get_sw_cqe(cq, i))
343 		++i;
344 
345 	return i - cq->mcq.cons_index;
346 }
347 
348 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
349 {
350 	struct mlx4_cqe *cqe, *new_cqe;
351 	int i;
352 	int cqe_size = cq->buf.entry_size;
353 	int cqe_inc = cqe_size == 64 ? 1 : 0;
354 
355 	i = cq->mcq.cons_index;
356 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
357 	cqe += cqe_inc;
358 
359 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
360 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
361 					   (i + 1) & cq->resize_buf->cqe);
362 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
363 		new_cqe += cqe_inc;
364 
365 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
366 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
367 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
368 		cqe += cqe_inc;
369 	}
370 	++cq->mcq.cons_index;
371 }
372 
373 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
374 {
375 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
376 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
377 	struct mlx4_mtt mtt;
378 	int outst_cqe;
379 	int err;
380 
381 	mutex_lock(&cq->resize_mutex);
382 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
383 		err = -EINVAL;
384 		goto out;
385 	}
386 
387 	entries = roundup_pow_of_two(entries + 1);
388 	if (entries == ibcq->cqe + 1) {
389 		err = 0;
390 		goto out;
391 	}
392 
393 	if (entries > dev->dev->caps.max_cqes + 1) {
394 		err = -EINVAL;
395 		goto out;
396 	}
397 
398 	if (ibcq->uobject) {
399 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
400 		if (err)
401 			goto out;
402 	} else {
403 		/* Can't be smaller than the number of outstanding CQEs */
404 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
405 		if (entries < outst_cqe + 1) {
406 			err = -EINVAL;
407 			goto out;
408 		}
409 
410 		err = mlx4_alloc_resize_buf(dev, cq, entries);
411 		if (err)
412 			goto out;
413 	}
414 
415 	mtt = cq->buf.mtt;
416 
417 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
418 	if (err)
419 		goto err_buf;
420 
421 	mlx4_mtt_cleanup(dev->dev, &mtt);
422 	if (ibcq->uobject) {
423 		cq->buf      = cq->resize_buf->buf;
424 		cq->ibcq.cqe = cq->resize_buf->cqe;
425 		ib_umem_release(cq->umem);
426 		cq->umem     = cq->resize_umem;
427 
428 		kfree(cq->resize_buf);
429 		cq->resize_buf = NULL;
430 		cq->resize_umem = NULL;
431 	} else {
432 		struct mlx4_ib_cq_buf tmp_buf;
433 		int tmp_cqe = 0;
434 
435 		spin_lock_irq(&cq->lock);
436 		if (cq->resize_buf) {
437 			mlx4_ib_cq_resize_copy_cqes(cq);
438 			tmp_buf = cq->buf;
439 			tmp_cqe = cq->ibcq.cqe;
440 			cq->buf      = cq->resize_buf->buf;
441 			cq->ibcq.cqe = cq->resize_buf->cqe;
442 
443 			kfree(cq->resize_buf);
444 			cq->resize_buf = NULL;
445 		}
446 		spin_unlock_irq(&cq->lock);
447 
448 		if (tmp_cqe)
449 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
450 	}
451 
452 	goto out;
453 
454 err_buf:
455 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
456 	if (!ibcq->uobject)
457 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
458 				    cq->resize_buf->cqe);
459 
460 	kfree(cq->resize_buf);
461 	cq->resize_buf = NULL;
462 
463 	if (cq->resize_umem) {
464 		ib_umem_release(cq->resize_umem);
465 		cq->resize_umem = NULL;
466 	}
467 
468 out:
469 	mutex_unlock(&cq->resize_mutex);
470 
471 	return err;
472 }
473 
474 int mlx4_ib_destroy_cq(struct ib_cq *cq)
475 {
476 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
477 	struct mlx4_ib_cq *mcq = to_mcq(cq);
478 
479 	mlx4_cq_free(dev->dev, &mcq->mcq);
480 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
481 
482 	if (cq->uobject) {
483 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
484 		ib_umem_release(mcq->umem);
485 	} else {
486 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
487 		mlx4_db_free(dev->dev, &mcq->db);
488 	}
489 
490 	kfree(mcq);
491 
492 	return 0;
493 }
494 
495 static void dump_cqe(void *cqe)
496 {
497 	__be32 *buf = cqe;
498 
499 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
500 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
501 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
502 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
503 }
504 
505 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
506 				     struct ib_wc *wc)
507 {
508 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
509 		pr_debug("local QP operation err "
510 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
511 		       "opcode = %02x)\n",
512 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
513 		       cqe->vendor_err_syndrome,
514 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
515 		dump_cqe(cqe);
516 	}
517 
518 	switch (cqe->syndrome) {
519 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
520 		wc->status = IB_WC_LOC_LEN_ERR;
521 		break;
522 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
523 		wc->status = IB_WC_LOC_QP_OP_ERR;
524 		break;
525 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
526 		wc->status = IB_WC_LOC_PROT_ERR;
527 		break;
528 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
529 		wc->status = IB_WC_WR_FLUSH_ERR;
530 		break;
531 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
532 		wc->status = IB_WC_MW_BIND_ERR;
533 		break;
534 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
535 		wc->status = IB_WC_BAD_RESP_ERR;
536 		break;
537 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
538 		wc->status = IB_WC_LOC_ACCESS_ERR;
539 		break;
540 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
541 		wc->status = IB_WC_REM_INV_REQ_ERR;
542 		break;
543 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
544 		wc->status = IB_WC_REM_ACCESS_ERR;
545 		break;
546 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
547 		wc->status = IB_WC_REM_OP_ERR;
548 		break;
549 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
550 		wc->status = IB_WC_RETRY_EXC_ERR;
551 		break;
552 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
553 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
554 		break;
555 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
556 		wc->status = IB_WC_REM_ABORT_ERR;
557 		break;
558 	default:
559 		wc->status = IB_WC_GENERAL_ERR;
560 		break;
561 	}
562 
563 	wc->vendor_err = cqe->vendor_err_syndrome;
564 }
565 
566 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
567 {
568 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
569 				      MLX4_CQE_STATUS_IPV4F     |
570 				      MLX4_CQE_STATUS_IPV4OPT   |
571 				      MLX4_CQE_STATUS_IPV6      |
572 				      MLX4_CQE_STATUS_IPOK)) ==
573 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
574 			    MLX4_CQE_STATUS_IPOK))              &&
575 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
576 				      MLX4_CQE_STATUS_TCP))     &&
577 		checksum == cpu_to_be16(0xffff);
578 }
579 
580 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
581 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
582 {
583 	struct mlx4_ib_proxy_sqp_hdr *hdr;
584 
585 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
586 				   qp->sqp_proxy_rcv[tail].map,
587 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
588 				   DMA_FROM_DEVICE);
589 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
590 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
591 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
592 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
593 	wc->dlid_path_bits = 0;
594 
595 	if (is_eth) {
596 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
597 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
598 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
599 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
600 	} else {
601 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
602 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
603 	}
604 }
605 
606 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
607 			       struct ib_wc *wc, int *npolled, int is_send)
608 {
609 	struct mlx4_ib_wq *wq;
610 	unsigned cur;
611 	int i;
612 
613 	wq = is_send ? &qp->sq : &qp->rq;
614 	cur = wq->head - wq->tail;
615 
616 	if (cur == 0)
617 		return;
618 
619 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
620 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
621 		wc->status = IB_WC_WR_FLUSH_ERR;
622 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
623 		wq->tail++;
624 		(*npolled)++;
625 		wc->qp = &qp->ibqp;
626 		wc++;
627 	}
628 }
629 
630 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
631 				 struct ib_wc *wc, int *npolled)
632 {
633 	struct mlx4_ib_qp *qp;
634 
635 	*npolled = 0;
636 	/* Find uncompleted WQEs belonging to that cq and retrun
637 	 * simulated FLUSH_ERR completions
638 	 */
639 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
640 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
641 		if (*npolled >= num_entries)
642 			goto out;
643 	}
644 
645 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
646 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
647 		if (*npolled >= num_entries)
648 			goto out;
649 	}
650 
651 out:
652 	return;
653 }
654 
655 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
656 			    struct mlx4_ib_qp **cur_qp,
657 			    struct ib_wc *wc)
658 {
659 	struct mlx4_cqe *cqe;
660 	struct mlx4_qp *mqp;
661 	struct mlx4_ib_wq *wq;
662 	struct mlx4_ib_srq *srq;
663 	struct mlx4_srq *msrq = NULL;
664 	int is_send;
665 	int is_error;
666 	int is_eth;
667 	u32 g_mlpath_rqpn;
668 	u16 wqe_ctr;
669 	unsigned tail = 0;
670 
671 repoll:
672 	cqe = next_cqe_sw(cq);
673 	if (!cqe)
674 		return -EAGAIN;
675 
676 	if (cq->buf.entry_size == 64)
677 		cqe++;
678 
679 	++cq->mcq.cons_index;
680 
681 	/*
682 	 * Make sure we read CQ entry contents after we've checked the
683 	 * ownership bit.
684 	 */
685 	rmb();
686 
687 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
688 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
689 		MLX4_CQE_OPCODE_ERROR;
690 
691 	/* Resize CQ in progress */
692 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
693 		if (cq->resize_buf) {
694 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
695 
696 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
697 			cq->buf      = cq->resize_buf->buf;
698 			cq->ibcq.cqe = cq->resize_buf->cqe;
699 
700 			kfree(cq->resize_buf);
701 			cq->resize_buf = NULL;
702 		}
703 
704 		goto repoll;
705 	}
706 
707 	if (!*cur_qp ||
708 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
709 		/*
710 		 * We do not have to take the QP table lock here,
711 		 * because CQs will be locked while QPs are removed
712 		 * from the table.
713 		 */
714 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
715 				       be32_to_cpu(cqe->vlan_my_qpn));
716 		*cur_qp = to_mibqp(mqp);
717 	}
718 
719 	wc->qp = &(*cur_qp)->ibqp;
720 
721 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
722 		u32 srq_num;
723 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
724 		srq_num       = g_mlpath_rqpn & 0xffffff;
725 		/* SRQ is also in the radix tree */
726 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
727 				       srq_num);
728 	}
729 
730 	if (is_send) {
731 		wq = &(*cur_qp)->sq;
732 		if (!(*cur_qp)->sq_signal_bits) {
733 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
734 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
735 		}
736 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
737 		++wq->tail;
738 	} else if ((*cur_qp)->ibqp.srq) {
739 		srq = to_msrq((*cur_qp)->ibqp.srq);
740 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
741 		wc->wr_id = srq->wrid[wqe_ctr];
742 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
743 	} else if (msrq) {
744 		srq = to_mibsrq(msrq);
745 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
746 		wc->wr_id = srq->wrid[wqe_ctr];
747 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
748 	} else {
749 		wq	  = &(*cur_qp)->rq;
750 		tail	  = wq->tail & (wq->wqe_cnt - 1);
751 		wc->wr_id = wq->wrid[tail];
752 		++wq->tail;
753 	}
754 
755 	if (unlikely(is_error)) {
756 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
757 		return 0;
758 	}
759 
760 	wc->status = IB_WC_SUCCESS;
761 
762 	if (is_send) {
763 		wc->wc_flags = 0;
764 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
765 		case MLX4_OPCODE_RDMA_WRITE_IMM:
766 			wc->wc_flags |= IB_WC_WITH_IMM;
767 		case MLX4_OPCODE_RDMA_WRITE:
768 			wc->opcode    = IB_WC_RDMA_WRITE;
769 			break;
770 		case MLX4_OPCODE_SEND_IMM:
771 			wc->wc_flags |= IB_WC_WITH_IMM;
772 		case MLX4_OPCODE_SEND:
773 		case MLX4_OPCODE_SEND_INVAL:
774 			wc->opcode    = IB_WC_SEND;
775 			break;
776 		case MLX4_OPCODE_RDMA_READ:
777 			wc->opcode    = IB_WC_RDMA_READ;
778 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
779 			break;
780 		case MLX4_OPCODE_ATOMIC_CS:
781 			wc->opcode    = IB_WC_COMP_SWAP;
782 			wc->byte_len  = 8;
783 			break;
784 		case MLX4_OPCODE_ATOMIC_FA:
785 			wc->opcode    = IB_WC_FETCH_ADD;
786 			wc->byte_len  = 8;
787 			break;
788 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
789 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
790 			wc->byte_len  = 8;
791 			break;
792 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
793 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
794 			wc->byte_len  = 8;
795 			break;
796 		case MLX4_OPCODE_LSO:
797 			wc->opcode    = IB_WC_LSO;
798 			break;
799 		case MLX4_OPCODE_FMR:
800 			wc->opcode    = IB_WC_REG_MR;
801 			break;
802 		case MLX4_OPCODE_LOCAL_INVAL:
803 			wc->opcode    = IB_WC_LOCAL_INV;
804 			break;
805 		}
806 	} else {
807 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
808 
809 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
810 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
811 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
812 			wc->wc_flags	= IB_WC_WITH_IMM;
813 			wc->ex.imm_data = cqe->immed_rss_invalid;
814 			break;
815 		case MLX4_RECV_OPCODE_SEND_INVAL:
816 			wc->opcode	= IB_WC_RECV;
817 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
818 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
819 			break;
820 		case MLX4_RECV_OPCODE_SEND:
821 			wc->opcode   = IB_WC_RECV;
822 			wc->wc_flags = 0;
823 			break;
824 		case MLX4_RECV_OPCODE_SEND_IMM:
825 			wc->opcode	= IB_WC_RECV;
826 			wc->wc_flags	= IB_WC_WITH_IMM;
827 			wc->ex.imm_data = cqe->immed_rss_invalid;
828 			break;
829 		}
830 
831 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
832 						  (*cur_qp)->port) ==
833 			  IB_LINK_LAYER_ETHERNET);
834 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
835 			if ((*cur_qp)->mlx4_ib_qp_type &
836 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
837 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
838 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
839 						is_eth);
840 				return 0;
841 			}
842 		}
843 
844 		wc->slid	   = be16_to_cpu(cqe->rlid);
845 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
846 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
847 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
848 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
849 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
850 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
851 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
852 		if (is_eth) {
853 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
854 			if (be32_to_cpu(cqe->vlan_my_qpn) &
855 					MLX4_CQE_CVLAN_PRESENT_MASK) {
856 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
857 					MLX4_CQE_VID_MASK;
858 			} else {
859 				wc->vlan_id = 0xffff;
860 			}
861 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
862 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
863 		} else {
864 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
865 			wc->vlan_id = 0xffff;
866 		}
867 	}
868 
869 	return 0;
870 }
871 
872 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
873 {
874 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
875 	struct mlx4_ib_qp *cur_qp = NULL;
876 	unsigned long flags;
877 	int npolled;
878 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
879 
880 	spin_lock_irqsave(&cq->lock, flags);
881 	if (unlikely(mdev->dev->persist->state &
882 		     MLX4_DEVICE_STATE_INTERNAL_ERROR)) {
883 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
884 		goto out;
885 	}
886 
887 	for (npolled = 0; npolled < num_entries; ++npolled) {
888 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
889 			break;
890 	}
891 
892 	mlx4_cq_set_ci(&cq->mcq);
893 
894 out:
895 	spin_unlock_irqrestore(&cq->lock, flags);
896 
897 	return npolled;
898 }
899 
900 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
901 {
902 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
903 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
904 
905 	if (unlikely(mdev->dev->persist->state &
906 		     MLX4_DEVICE_STATE_INTERNAL_ERROR))
907 		return -1;
908 
909 	mlx4_cq_arm(&cq->mcq,
910 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
911 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
912 		    mdev->uar_map,
913 		    MLX4_GET_DOORBELL_LOCK(&mdev->uar_lock));
914 
915 	return 0;
916 }
917 
918 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
919 {
920 	u32 prod_index;
921 	int nfreed = 0;
922 	struct mlx4_cqe *cqe, *dest;
923 	u8 owner_bit;
924 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
925 
926 	/*
927 	 * First we need to find the current producer index, so we
928 	 * know where to start cleaning from.  It doesn't matter if HW
929 	 * adds new entries after this loop -- the QP we're worried
930 	 * about is already in RESET, so the new entries won't come
931 	 * from our QP and therefore don't need to be checked.
932 	 */
933 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
934 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
935 			break;
936 
937 	/*
938 	 * Now sweep backwards through the CQ, removing CQ entries
939 	 * that match our QP by copying older entries on top of them.
940 	 */
941 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
942 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
943 		cqe += cqe_inc;
944 
945 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
946 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
947 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
948 			++nfreed;
949 		} else if (nfreed) {
950 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
951 			dest += cqe_inc;
952 
953 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
954 			memcpy(dest, cqe, sizeof *cqe);
955 			dest->owner_sr_opcode = owner_bit |
956 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
957 		}
958 	}
959 
960 	if (nfreed) {
961 		cq->mcq.cons_index += nfreed;
962 		/*
963 		 * Make sure update of buffer contents is done before
964 		 * updating consumer index.
965 		 */
966 		wmb();
967 		mlx4_cq_set_ci(&cq->mcq);
968 	}
969 }
970 
971 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
972 {
973 	spin_lock_irq(&cq->lock);
974 	__mlx4_ib_cq_clean(cq, qpn, srq);
975 	spin_unlock_irq(&cq->lock);
976 }
977