xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision 8730046c)
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 <rdma/mlx4-abi.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, GFP_KERNEL);
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, GFP_KERNEL);
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 #define CQ_CREATE_FLAGS_SUPPORTED IB_CQ_FLAGS_TIMESTAMP_COMPLETION
170 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
171 				const struct ib_cq_init_attr *attr,
172 				struct ib_ucontext *context,
173 				struct ib_udata *udata)
174 {
175 	int entries = attr->cqe;
176 	int vector = attr->comp_vector;
177 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
178 	struct mlx4_ib_cq *cq;
179 	struct mlx4_uar *uar;
180 	int err;
181 
182 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
183 		return ERR_PTR(-EINVAL);
184 
185 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
186 		return ERR_PTR(-EINVAL);
187 
188 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
189 	if (!cq)
190 		return ERR_PTR(-ENOMEM);
191 
192 	entries      = roundup_pow_of_two(entries + 1);
193 	cq->ibcq.cqe = entries - 1;
194 	mutex_init(&cq->resize_mutex);
195 	spin_lock_init(&cq->lock);
196 	cq->resize_buf = NULL;
197 	cq->resize_umem = NULL;
198 	cq->create_flags = attr->flags;
199 	INIT_LIST_HEAD(&cq->send_qp_list);
200 	INIT_LIST_HEAD(&cq->recv_qp_list);
201 
202 	if (context) {
203 		struct mlx4_ib_create_cq ucmd;
204 
205 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
206 			err = -EFAULT;
207 			goto err_cq;
208 		}
209 
210 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
211 					  ucmd.buf_addr, entries);
212 		if (err)
213 			goto err_cq;
214 
215 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
216 					  &cq->db);
217 		if (err)
218 			goto err_mtt;
219 
220 		uar = &to_mucontext(context)->uar;
221 	} else {
222 		err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
223 		if (err)
224 			goto err_cq;
225 
226 		cq->mcq.set_ci_db  = cq->db.db;
227 		cq->mcq.arm_db     = cq->db.db + 1;
228 		*cq->mcq.set_ci_db = 0;
229 		*cq->mcq.arm_db    = 0;
230 
231 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
232 		if (err)
233 			goto err_db;
234 
235 		uar = &dev->priv_uar;
236 	}
237 
238 	if (dev->eq_table)
239 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
240 
241 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
242 			    cq->db.dma, &cq->mcq, vector, 0,
243 			    !!(cq->create_flags & IB_CQ_FLAGS_TIMESTAMP_COMPLETION));
244 	if (err)
245 		goto err_dbmap;
246 
247 	if (context)
248 		cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
249 	else
250 		cq->mcq.comp = mlx4_ib_cq_comp;
251 	cq->mcq.event = mlx4_ib_cq_event;
252 
253 	if (context)
254 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
255 			err = -EFAULT;
256 			goto err_cq_free;
257 		}
258 
259 	return &cq->ibcq;
260 
261 err_cq_free:
262 	mlx4_cq_free(dev->dev, &cq->mcq);
263 
264 err_dbmap:
265 	if (context)
266 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
267 
268 err_mtt:
269 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
270 
271 	if (context)
272 		ib_umem_release(cq->umem);
273 	else
274 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
275 
276 err_db:
277 	if (!context)
278 		mlx4_db_free(dev->dev, &cq->db);
279 
280 err_cq:
281 	kfree(cq);
282 
283 	return ERR_PTR(err);
284 }
285 
286 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
287 				  int entries)
288 {
289 	int err;
290 
291 	if (cq->resize_buf)
292 		return -EBUSY;
293 
294 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
295 	if (!cq->resize_buf)
296 		return -ENOMEM;
297 
298 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
299 	if (err) {
300 		kfree(cq->resize_buf);
301 		cq->resize_buf = NULL;
302 		return err;
303 	}
304 
305 	cq->resize_buf->cqe = entries - 1;
306 
307 	return 0;
308 }
309 
310 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
311 				   int entries, struct ib_udata *udata)
312 {
313 	struct mlx4_ib_resize_cq ucmd;
314 	int err;
315 
316 	if (cq->resize_umem)
317 		return -EBUSY;
318 
319 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
320 		return -EFAULT;
321 
322 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
323 	if (!cq->resize_buf)
324 		return -ENOMEM;
325 
326 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
327 				  &cq->resize_umem, ucmd.buf_addr, entries);
328 	if (err) {
329 		kfree(cq->resize_buf);
330 		cq->resize_buf = NULL;
331 		return err;
332 	}
333 
334 	cq->resize_buf->cqe = entries - 1;
335 
336 	return 0;
337 }
338 
339 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
340 {
341 	u32 i;
342 
343 	i = cq->mcq.cons_index;
344 	while (get_sw_cqe(cq, i))
345 		++i;
346 
347 	return i - cq->mcq.cons_index;
348 }
349 
350 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
351 {
352 	struct mlx4_cqe *cqe, *new_cqe;
353 	int i;
354 	int cqe_size = cq->buf.entry_size;
355 	int cqe_inc = cqe_size == 64 ? 1 : 0;
356 
357 	i = cq->mcq.cons_index;
358 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
359 	cqe += cqe_inc;
360 
361 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
362 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
363 					   (i + 1) & cq->resize_buf->cqe);
364 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
365 		new_cqe += cqe_inc;
366 
367 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
368 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
369 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
370 		cqe += cqe_inc;
371 	}
372 	++cq->mcq.cons_index;
373 }
374 
375 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
376 {
377 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
378 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
379 	struct mlx4_mtt mtt;
380 	int outst_cqe;
381 	int err;
382 
383 	mutex_lock(&cq->resize_mutex);
384 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
385 		err = -EINVAL;
386 		goto out;
387 	}
388 
389 	entries = roundup_pow_of_two(entries + 1);
390 	if (entries == ibcq->cqe + 1) {
391 		err = 0;
392 		goto out;
393 	}
394 
395 	if (entries > dev->dev->caps.max_cqes + 1) {
396 		err = -EINVAL;
397 		goto out;
398 	}
399 
400 	if (ibcq->uobject) {
401 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
402 		if (err)
403 			goto out;
404 	} else {
405 		/* Can't be smaller than the number of outstanding CQEs */
406 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
407 		if (entries < outst_cqe + 1) {
408 			err = -EINVAL;
409 			goto out;
410 		}
411 
412 		err = mlx4_alloc_resize_buf(dev, cq, entries);
413 		if (err)
414 			goto out;
415 	}
416 
417 	mtt = cq->buf.mtt;
418 
419 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
420 	if (err)
421 		goto err_buf;
422 
423 	mlx4_mtt_cleanup(dev->dev, &mtt);
424 	if (ibcq->uobject) {
425 		cq->buf      = cq->resize_buf->buf;
426 		cq->ibcq.cqe = cq->resize_buf->cqe;
427 		ib_umem_release(cq->umem);
428 		cq->umem     = cq->resize_umem;
429 
430 		kfree(cq->resize_buf);
431 		cq->resize_buf = NULL;
432 		cq->resize_umem = NULL;
433 	} else {
434 		struct mlx4_ib_cq_buf tmp_buf;
435 		int tmp_cqe = 0;
436 
437 		spin_lock_irq(&cq->lock);
438 		if (cq->resize_buf) {
439 			mlx4_ib_cq_resize_copy_cqes(cq);
440 			tmp_buf = cq->buf;
441 			tmp_cqe = cq->ibcq.cqe;
442 			cq->buf      = cq->resize_buf->buf;
443 			cq->ibcq.cqe = cq->resize_buf->cqe;
444 
445 			kfree(cq->resize_buf);
446 			cq->resize_buf = NULL;
447 		}
448 		spin_unlock_irq(&cq->lock);
449 
450 		if (tmp_cqe)
451 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
452 	}
453 
454 	goto out;
455 
456 err_buf:
457 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
458 	if (!ibcq->uobject)
459 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
460 				    cq->resize_buf->cqe);
461 
462 	kfree(cq->resize_buf);
463 	cq->resize_buf = NULL;
464 
465 	if (cq->resize_umem) {
466 		ib_umem_release(cq->resize_umem);
467 		cq->resize_umem = NULL;
468 	}
469 
470 out:
471 	mutex_unlock(&cq->resize_mutex);
472 
473 	return err;
474 }
475 
476 int mlx4_ib_destroy_cq(struct ib_cq *cq)
477 {
478 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
479 	struct mlx4_ib_cq *mcq = to_mcq(cq);
480 
481 	mlx4_cq_free(dev->dev, &mcq->mcq);
482 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
483 
484 	if (cq->uobject) {
485 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
486 		ib_umem_release(mcq->umem);
487 	} else {
488 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
489 		mlx4_db_free(dev->dev, &mcq->db);
490 	}
491 
492 	kfree(mcq);
493 
494 	return 0;
495 }
496 
497 static void dump_cqe(void *cqe)
498 {
499 	__be32 *buf = cqe;
500 
501 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
502 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
503 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
504 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
505 }
506 
507 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
508 				     struct ib_wc *wc)
509 {
510 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
511 		pr_debug("local QP operation err "
512 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
513 		       "opcode = %02x)\n",
514 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
515 		       cqe->vendor_err_syndrome,
516 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
517 		dump_cqe(cqe);
518 	}
519 
520 	switch (cqe->syndrome) {
521 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
522 		wc->status = IB_WC_LOC_LEN_ERR;
523 		break;
524 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
525 		wc->status = IB_WC_LOC_QP_OP_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
528 		wc->status = IB_WC_LOC_PROT_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
531 		wc->status = IB_WC_WR_FLUSH_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
534 		wc->status = IB_WC_MW_BIND_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
537 		wc->status = IB_WC_BAD_RESP_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
540 		wc->status = IB_WC_LOC_ACCESS_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
543 		wc->status = IB_WC_REM_INV_REQ_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
546 		wc->status = IB_WC_REM_ACCESS_ERR;
547 		break;
548 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
549 		wc->status = IB_WC_REM_OP_ERR;
550 		break;
551 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
552 		wc->status = IB_WC_RETRY_EXC_ERR;
553 		break;
554 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
555 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
556 		break;
557 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
558 		wc->status = IB_WC_REM_ABORT_ERR;
559 		break;
560 	default:
561 		wc->status = IB_WC_GENERAL_ERR;
562 		break;
563 	}
564 
565 	wc->vendor_err = cqe->vendor_err_syndrome;
566 }
567 
568 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
569 {
570 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
571 				      MLX4_CQE_STATUS_IPV4F     |
572 				      MLX4_CQE_STATUS_IPV4OPT   |
573 				      MLX4_CQE_STATUS_IPV6      |
574 				      MLX4_CQE_STATUS_IPOK)) ==
575 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
576 			    MLX4_CQE_STATUS_IPOK))              &&
577 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
578 				      MLX4_CQE_STATUS_TCP))     &&
579 		checksum == cpu_to_be16(0xffff);
580 }
581 
582 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
583 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
584 {
585 	struct mlx4_ib_proxy_sqp_hdr *hdr;
586 
587 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
588 				   qp->sqp_proxy_rcv[tail].map,
589 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
590 				   DMA_FROM_DEVICE);
591 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
592 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
593 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
594 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
595 	wc->dlid_path_bits = 0;
596 
597 	if (is_eth) {
598 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
599 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
600 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
601 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
602 	} else {
603 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
604 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
605 	}
606 }
607 
608 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
609 			       struct ib_wc *wc, int *npolled, int is_send)
610 {
611 	struct mlx4_ib_wq *wq;
612 	unsigned cur;
613 	int i;
614 
615 	wq = is_send ? &qp->sq : &qp->rq;
616 	cur = wq->head - wq->tail;
617 
618 	if (cur == 0)
619 		return;
620 
621 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
622 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
623 		wc->status = IB_WC_WR_FLUSH_ERR;
624 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
625 		wq->tail++;
626 		(*npolled)++;
627 		wc->qp = &qp->ibqp;
628 		wc++;
629 	}
630 }
631 
632 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
633 				 struct ib_wc *wc, int *npolled)
634 {
635 	struct mlx4_ib_qp *qp;
636 
637 	*npolled = 0;
638 	/* Find uncompleted WQEs belonging to that cq and retrun
639 	 * simulated FLUSH_ERR completions
640 	 */
641 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
642 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
643 		if (*npolled >= num_entries)
644 			goto out;
645 	}
646 
647 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
648 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
649 		if (*npolled >= num_entries)
650 			goto out;
651 	}
652 
653 out:
654 	return;
655 }
656 
657 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
658 			    struct mlx4_ib_qp **cur_qp,
659 			    struct ib_wc *wc)
660 {
661 	struct mlx4_cqe *cqe;
662 	struct mlx4_qp *mqp;
663 	struct mlx4_ib_wq *wq;
664 	struct mlx4_ib_srq *srq;
665 	struct mlx4_srq *msrq = NULL;
666 	int is_send;
667 	int is_error;
668 	int is_eth;
669 	u32 g_mlpath_rqpn;
670 	u16 wqe_ctr;
671 	unsigned tail = 0;
672 
673 repoll:
674 	cqe = next_cqe_sw(cq);
675 	if (!cqe)
676 		return -EAGAIN;
677 
678 	if (cq->buf.entry_size == 64)
679 		cqe++;
680 
681 	++cq->mcq.cons_index;
682 
683 	/*
684 	 * Make sure we read CQ entry contents after we've checked the
685 	 * ownership bit.
686 	 */
687 	rmb();
688 
689 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
690 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
691 		MLX4_CQE_OPCODE_ERROR;
692 
693 	/* Resize CQ in progress */
694 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
695 		if (cq->resize_buf) {
696 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
697 
698 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
699 			cq->buf      = cq->resize_buf->buf;
700 			cq->ibcq.cqe = cq->resize_buf->cqe;
701 
702 			kfree(cq->resize_buf);
703 			cq->resize_buf = NULL;
704 		}
705 
706 		goto repoll;
707 	}
708 
709 	if (!*cur_qp ||
710 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
711 		/*
712 		 * We do not have to take the QP table lock here,
713 		 * because CQs will be locked while QPs are removed
714 		 * from the table.
715 		 */
716 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
717 				       be32_to_cpu(cqe->vlan_my_qpn));
718 		*cur_qp = to_mibqp(mqp);
719 	}
720 
721 	wc->qp = &(*cur_qp)->ibqp;
722 
723 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
724 		u32 srq_num;
725 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
726 		srq_num       = g_mlpath_rqpn & 0xffffff;
727 		/* SRQ is also in the radix tree */
728 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
729 				       srq_num);
730 	}
731 
732 	if (is_send) {
733 		wq = &(*cur_qp)->sq;
734 		if (!(*cur_qp)->sq_signal_bits) {
735 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
736 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
737 		}
738 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
739 		++wq->tail;
740 	} else if ((*cur_qp)->ibqp.srq) {
741 		srq = to_msrq((*cur_qp)->ibqp.srq);
742 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
743 		wc->wr_id = srq->wrid[wqe_ctr];
744 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
745 	} else if (msrq) {
746 		srq = to_mibsrq(msrq);
747 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
748 		wc->wr_id = srq->wrid[wqe_ctr];
749 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
750 	} else {
751 		wq	  = &(*cur_qp)->rq;
752 		tail	  = wq->tail & (wq->wqe_cnt - 1);
753 		wc->wr_id = wq->wrid[tail];
754 		++wq->tail;
755 	}
756 
757 	if (unlikely(is_error)) {
758 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
759 		return 0;
760 	}
761 
762 	wc->status = IB_WC_SUCCESS;
763 
764 	if (is_send) {
765 		wc->wc_flags = 0;
766 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
767 		case MLX4_OPCODE_RDMA_WRITE_IMM:
768 			wc->wc_flags |= IB_WC_WITH_IMM;
769 		case MLX4_OPCODE_RDMA_WRITE:
770 			wc->opcode    = IB_WC_RDMA_WRITE;
771 			break;
772 		case MLX4_OPCODE_SEND_IMM:
773 			wc->wc_flags |= IB_WC_WITH_IMM;
774 		case MLX4_OPCODE_SEND:
775 		case MLX4_OPCODE_SEND_INVAL:
776 			wc->opcode    = IB_WC_SEND;
777 			break;
778 		case MLX4_OPCODE_RDMA_READ:
779 			wc->opcode    = IB_WC_RDMA_READ;
780 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
781 			break;
782 		case MLX4_OPCODE_ATOMIC_CS:
783 			wc->opcode    = IB_WC_COMP_SWAP;
784 			wc->byte_len  = 8;
785 			break;
786 		case MLX4_OPCODE_ATOMIC_FA:
787 			wc->opcode    = IB_WC_FETCH_ADD;
788 			wc->byte_len  = 8;
789 			break;
790 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
791 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
792 			wc->byte_len  = 8;
793 			break;
794 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
795 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
796 			wc->byte_len  = 8;
797 			break;
798 		case MLX4_OPCODE_LSO:
799 			wc->opcode    = IB_WC_LSO;
800 			break;
801 		case MLX4_OPCODE_FMR:
802 			wc->opcode    = IB_WC_REG_MR;
803 			break;
804 		case MLX4_OPCODE_LOCAL_INVAL:
805 			wc->opcode    = IB_WC_LOCAL_INV;
806 			break;
807 		}
808 	} else {
809 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
810 
811 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
812 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
813 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
814 			wc->wc_flags	= IB_WC_WITH_IMM;
815 			wc->ex.imm_data = cqe->immed_rss_invalid;
816 			break;
817 		case MLX4_RECV_OPCODE_SEND_INVAL:
818 			wc->opcode	= IB_WC_RECV;
819 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
820 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
821 			break;
822 		case MLX4_RECV_OPCODE_SEND:
823 			wc->opcode   = IB_WC_RECV;
824 			wc->wc_flags = 0;
825 			break;
826 		case MLX4_RECV_OPCODE_SEND_IMM:
827 			wc->opcode	= IB_WC_RECV;
828 			wc->wc_flags	= IB_WC_WITH_IMM;
829 			wc->ex.imm_data = cqe->immed_rss_invalid;
830 			break;
831 		}
832 
833 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
834 						  (*cur_qp)->port) ==
835 			  IB_LINK_LAYER_ETHERNET);
836 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
837 			if ((*cur_qp)->mlx4_ib_qp_type &
838 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
839 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
840 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
841 						is_eth);
842 				return 0;
843 			}
844 		}
845 
846 		wc->slid	   = be16_to_cpu(cqe->rlid);
847 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
848 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
849 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
850 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
851 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
852 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
853 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
854 		if (is_eth) {
855 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
856 			if (be32_to_cpu(cqe->vlan_my_qpn) &
857 					MLX4_CQE_CVLAN_PRESENT_MASK) {
858 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
859 					MLX4_CQE_VID_MASK;
860 			} else {
861 				wc->vlan_id = 0xffff;
862 			}
863 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
864 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
865 		} else {
866 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
867 			wc->vlan_id = 0xffff;
868 		}
869 	}
870 
871 	return 0;
872 }
873 
874 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
875 {
876 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
877 	struct mlx4_ib_qp *cur_qp = NULL;
878 	unsigned long flags;
879 	int npolled;
880 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
881 
882 	spin_lock_irqsave(&cq->lock, flags);
883 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
884 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
885 		goto out;
886 	}
887 
888 	for (npolled = 0; npolled < num_entries; ++npolled) {
889 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
890 			break;
891 	}
892 
893 	mlx4_cq_set_ci(&cq->mcq);
894 
895 out:
896 	spin_unlock_irqrestore(&cq->lock, flags);
897 
898 	return npolled;
899 }
900 
901 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
902 {
903 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
904 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
905 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
906 		    to_mdev(ibcq->device)->uar_map,
907 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
908 
909 	return 0;
910 }
911 
912 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
913 {
914 	u32 prod_index;
915 	int nfreed = 0;
916 	struct mlx4_cqe *cqe, *dest;
917 	u8 owner_bit;
918 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
919 
920 	/*
921 	 * First we need to find the current producer index, so we
922 	 * know where to start cleaning from.  It doesn't matter if HW
923 	 * adds new entries after this loop -- the QP we're worried
924 	 * about is already in RESET, so the new entries won't come
925 	 * from our QP and therefore don't need to be checked.
926 	 */
927 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
928 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
929 			break;
930 
931 	/*
932 	 * Now sweep backwards through the CQ, removing CQ entries
933 	 * that match our QP by copying older entries on top of them.
934 	 */
935 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
936 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
937 		cqe += cqe_inc;
938 
939 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
940 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
941 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
942 			++nfreed;
943 		} else if (nfreed) {
944 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
945 			dest += cqe_inc;
946 
947 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
948 			memcpy(dest, cqe, sizeof *cqe);
949 			dest->owner_sr_opcode = owner_bit |
950 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
951 		}
952 	}
953 
954 	if (nfreed) {
955 		cq->mcq.cons_index += nfreed;
956 		/*
957 		 * Make sure update of buffer contents is done before
958 		 * updating consumer index.
959 		 */
960 		wmb();
961 		mlx4_cq_set_ci(&cq->mcq);
962 	}
963 }
964 
965 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
966 {
967 	spin_lock_irq(&cq->lock);
968 	__mlx4_ib_cq_clean(cq, qpn, srq);
969 	spin_unlock_irq(&cq->lock);
970 }
971