xref: /linux/drivers/infiniband/hw/mlx4/cq.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 #include <rdma/uverbs_ioctl.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);
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);
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,
139 			       struct mlx4_ib_cq_buf *buf,
140 			       struct ib_umem **umem, u64 buf_addr, int cqe)
141 {
142 	int err;
143 	int cqe_size = dev->dev->caps.cqe_size;
144 	int shift;
145 	int n;
146 
147 	*umem = ib_umem_get(&dev->ib_dev, buf_addr, cqe * cqe_size,
148 			    IB_ACCESS_LOCAL_WRITE);
149 	if (IS_ERR(*umem))
150 		return PTR_ERR(*umem);
151 
152 	shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n);
153 	err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt);
154 
155 	if (err)
156 		goto err_buf;
157 
158 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
159 	if (err)
160 		goto err_mtt;
161 
162 	return 0;
163 
164 err_mtt:
165 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
166 
167 err_buf:
168 	ib_umem_release(*umem);
169 
170 	return err;
171 }
172 
173 #define CQ_CREATE_FLAGS_SUPPORTED IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION
174 int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
175 		      struct uverbs_attr_bundle *attrs)
176 {
177 	struct ib_udata *udata = &attrs->driver_udata;
178 	struct ib_device *ibdev = ibcq->device;
179 	int entries = attr->cqe;
180 	int vector = attr->comp_vector;
181 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
182 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
183 	struct mlx4_uar *uar;
184 	void *buf_addr;
185 	int err;
186 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
187 		udata, struct mlx4_ib_ucontext, ibucontext);
188 
189 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
190 		return -EINVAL;
191 
192 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
193 		return -EINVAL;
194 
195 	entries      = roundup_pow_of_two(entries + 1);
196 	cq->ibcq.cqe = entries - 1;
197 	mutex_init(&cq->resize_mutex);
198 	spin_lock_init(&cq->lock);
199 	cq->resize_buf = NULL;
200 	cq->resize_umem = NULL;
201 	cq->create_flags = attr->flags;
202 	INIT_LIST_HEAD(&cq->send_qp_list);
203 	INIT_LIST_HEAD(&cq->recv_qp_list);
204 
205 	if (udata) {
206 		struct mlx4_ib_create_cq ucmd;
207 
208 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
209 			err = -EFAULT;
210 			goto err_cq;
211 		}
212 
213 		buf_addr = (void *)(unsigned long)ucmd.buf_addr;
214 		err = mlx4_ib_get_cq_umem(dev, &cq->buf, &cq->umem,
215 					  ucmd.buf_addr, entries);
216 		if (err)
217 			goto err_cq;
218 
219 		err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &cq->db);
220 		if (err)
221 			goto err_mtt;
222 
223 		uar = &context->uar;
224 		cq->mcq.usage = MLX4_RES_USAGE_USER_VERBS;
225 	} else {
226 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
227 		if (err)
228 			goto err_cq;
229 
230 		cq->mcq.set_ci_db  = cq->db.db;
231 		cq->mcq.arm_db     = cq->db.db + 1;
232 		*cq->mcq.set_ci_db = 0;
233 		*cq->mcq.arm_db    = 0;
234 
235 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
236 		if (err)
237 			goto err_db;
238 
239 		buf_addr = &cq->buf.buf;
240 
241 		uar = &dev->priv_uar;
242 		cq->mcq.usage = MLX4_RES_USAGE_DRIVER;
243 	}
244 
245 	if (dev->eq_table)
246 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
247 
248 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar, cq->db.dma,
249 			    &cq->mcq, vector, 0,
250 			    !!(cq->create_flags &
251 			       IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION),
252 			    buf_addr, !!udata);
253 	if (err)
254 		goto err_dbmap;
255 
256 	if (udata)
257 		cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
258 	else
259 		cq->mcq.comp = mlx4_ib_cq_comp;
260 	cq->mcq.event = mlx4_ib_cq_event;
261 
262 	if (udata)
263 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
264 			err = -EFAULT;
265 			goto err_cq_free;
266 		}
267 
268 	return 0;
269 
270 err_cq_free:
271 	mlx4_cq_free(dev->dev, &cq->mcq);
272 
273 err_dbmap:
274 	if (udata)
275 		mlx4_ib_db_unmap_user(context, &cq->db);
276 
277 err_mtt:
278 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
279 
280 	ib_umem_release(cq->umem);
281 	if (!udata)
282 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
283 
284 err_db:
285 	if (!udata)
286 		mlx4_db_free(dev->dev, &cq->db);
287 err_cq:
288 	return err;
289 }
290 
291 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
292 				  int entries)
293 {
294 	int err;
295 
296 	if (cq->resize_buf)
297 		return -EBUSY;
298 
299 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
300 	if (!cq->resize_buf)
301 		return -ENOMEM;
302 
303 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
304 	if (err) {
305 		kfree(cq->resize_buf);
306 		cq->resize_buf = NULL;
307 		return err;
308 	}
309 
310 	cq->resize_buf->cqe = entries - 1;
311 
312 	return 0;
313 }
314 
315 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
316 				   int entries, struct ib_udata *udata)
317 {
318 	struct mlx4_ib_resize_cq ucmd;
319 	int err;
320 
321 	if (cq->resize_umem)
322 		return -EBUSY;
323 
324 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
325 		return -EFAULT;
326 
327 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
328 	if (!cq->resize_buf)
329 		return -ENOMEM;
330 
331 	err = mlx4_ib_get_cq_umem(dev, &cq->resize_buf->buf, &cq->resize_umem,
332 				  ucmd.buf_addr, entries);
333 	if (err) {
334 		kfree(cq->resize_buf);
335 		cq->resize_buf = NULL;
336 		return err;
337 	}
338 
339 	cq->resize_buf->cqe = entries - 1;
340 
341 	return 0;
342 }
343 
344 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
345 {
346 	u32 i;
347 
348 	i = cq->mcq.cons_index;
349 	while (get_sw_cqe(cq, i))
350 		++i;
351 
352 	return i - cq->mcq.cons_index;
353 }
354 
355 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
356 {
357 	struct mlx4_cqe *cqe, *new_cqe;
358 	int i;
359 	int cqe_size = cq->buf.entry_size;
360 	int cqe_inc = cqe_size == 64 ? 1 : 0;
361 
362 	i = cq->mcq.cons_index;
363 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
364 	cqe += cqe_inc;
365 
366 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
367 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
368 					   (i + 1) & cq->resize_buf->cqe);
369 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
370 		new_cqe += cqe_inc;
371 
372 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
373 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
374 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
375 		cqe += cqe_inc;
376 	}
377 	++cq->mcq.cons_index;
378 }
379 
380 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
381 {
382 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
383 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
384 	struct mlx4_mtt mtt;
385 	int outst_cqe;
386 	int err;
387 
388 	mutex_lock(&cq->resize_mutex);
389 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
390 		err = -EINVAL;
391 		goto out;
392 	}
393 
394 	entries = roundup_pow_of_two(entries + 1);
395 	if (entries == ibcq->cqe + 1) {
396 		err = 0;
397 		goto out;
398 	}
399 
400 	if (entries > dev->dev->caps.max_cqes + 1) {
401 		err = -EINVAL;
402 		goto out;
403 	}
404 
405 	if (ibcq->uobject) {
406 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
407 		if (err)
408 			goto out;
409 	} else {
410 		/* Can't be smaller than the number of outstanding CQEs */
411 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
412 		if (entries < outst_cqe + 1) {
413 			err = -EINVAL;
414 			goto out;
415 		}
416 
417 		err = mlx4_alloc_resize_buf(dev, cq, entries);
418 		if (err)
419 			goto out;
420 	}
421 
422 	mtt = cq->buf.mtt;
423 
424 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
425 	if (err)
426 		goto err_buf;
427 
428 	mlx4_mtt_cleanup(dev->dev, &mtt);
429 	if (ibcq->uobject) {
430 		cq->buf      = cq->resize_buf->buf;
431 		cq->ibcq.cqe = cq->resize_buf->cqe;
432 		ib_umem_release(cq->umem);
433 		cq->umem     = cq->resize_umem;
434 
435 		kfree(cq->resize_buf);
436 		cq->resize_buf = NULL;
437 		cq->resize_umem = NULL;
438 	} else {
439 		struct mlx4_ib_cq_buf tmp_buf;
440 		int tmp_cqe = 0;
441 
442 		spin_lock_irq(&cq->lock);
443 		if (cq->resize_buf) {
444 			mlx4_ib_cq_resize_copy_cqes(cq);
445 			tmp_buf = cq->buf;
446 			tmp_cqe = cq->ibcq.cqe;
447 			cq->buf      = cq->resize_buf->buf;
448 			cq->ibcq.cqe = cq->resize_buf->cqe;
449 
450 			kfree(cq->resize_buf);
451 			cq->resize_buf = NULL;
452 		}
453 		spin_unlock_irq(&cq->lock);
454 
455 		if (tmp_cqe)
456 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
457 	}
458 
459 	goto out;
460 
461 err_buf:
462 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
463 	if (!ibcq->uobject)
464 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
465 				    cq->resize_buf->cqe);
466 
467 	kfree(cq->resize_buf);
468 	cq->resize_buf = NULL;
469 
470 	ib_umem_release(cq->resize_umem);
471 	cq->resize_umem = NULL;
472 out:
473 	mutex_unlock(&cq->resize_mutex);
474 
475 	return err;
476 }
477 
478 int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
479 {
480 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
481 	struct mlx4_ib_cq *mcq = to_mcq(cq);
482 
483 	mlx4_cq_free(dev->dev, &mcq->mcq);
484 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
485 
486 	if (udata) {
487 		mlx4_ib_db_unmap_user(
488 			rdma_udata_to_drv_context(
489 				udata,
490 				struct mlx4_ib_ucontext,
491 				ibucontext),
492 			&mcq->db);
493 	} else {
494 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
495 		mlx4_db_free(dev->dev, &mcq->db);
496 	}
497 	ib_umem_release(mcq->umem);
498 	return 0;
499 }
500 
501 static void dump_cqe(void *cqe)
502 {
503 	__be32 *buf = cqe;
504 
505 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
506 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
507 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
508 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
509 }
510 
511 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
512 				     struct ib_wc *wc)
513 {
514 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
515 		pr_debug("local QP operation err "
516 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
517 		       "opcode = %02x)\n",
518 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
519 		       cqe->vendor_err_syndrome,
520 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
521 		dump_cqe(cqe);
522 	}
523 
524 	switch (cqe->syndrome) {
525 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
526 		wc->status = IB_WC_LOC_LEN_ERR;
527 		break;
528 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
529 		wc->status = IB_WC_LOC_QP_OP_ERR;
530 		break;
531 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
532 		wc->status = IB_WC_LOC_PROT_ERR;
533 		break;
534 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
535 		wc->status = IB_WC_WR_FLUSH_ERR;
536 		break;
537 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
538 		wc->status = IB_WC_MW_BIND_ERR;
539 		break;
540 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
541 		wc->status = IB_WC_BAD_RESP_ERR;
542 		break;
543 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
544 		wc->status = IB_WC_LOC_ACCESS_ERR;
545 		break;
546 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
547 		wc->status = IB_WC_REM_INV_REQ_ERR;
548 		break;
549 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
550 		wc->status = IB_WC_REM_ACCESS_ERR;
551 		break;
552 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
553 		wc->status = IB_WC_REM_OP_ERR;
554 		break;
555 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
556 		wc->status = IB_WC_RETRY_EXC_ERR;
557 		break;
558 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
559 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
560 		break;
561 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
562 		wc->status = IB_WC_REM_ABORT_ERR;
563 		break;
564 	default:
565 		wc->status = IB_WC_GENERAL_ERR;
566 		break;
567 	}
568 
569 	wc->vendor_err = cqe->vendor_err_syndrome;
570 }
571 
572 static int mlx4_ib_ipoib_csum_ok(__be16 status, u8 badfcs_enc, __be16 checksum)
573 {
574 	return ((badfcs_enc & MLX4_CQE_STATUS_L4_CSUM) ||
575 		((status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
576 		 (status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
577 				       MLX4_CQE_STATUS_UDP)) &&
578 		 (checksum == cpu_to_be16(0xffff))));
579 }
580 
581 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
582 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
583 {
584 	struct mlx4_ib_proxy_sqp_hdr *hdr;
585 
586 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
587 				   qp->sqp_proxy_rcv[tail].map,
588 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
589 				   DMA_FROM_DEVICE);
590 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
591 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
592 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
593 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
594 	wc->dlid_path_bits = 0;
595 
596 	if (is_eth) {
597 		wc->slid = 0;
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 return
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 			fallthrough;
770 		case MLX4_OPCODE_RDMA_WRITE:
771 			wc->opcode    = IB_WC_RDMA_WRITE;
772 			break;
773 		case MLX4_OPCODE_SEND_IMM:
774 			wc->wc_flags |= IB_WC_WITH_IMM;
775 			fallthrough;
776 		case MLX4_OPCODE_SEND:
777 		case MLX4_OPCODE_SEND_INVAL:
778 			wc->opcode    = IB_WC_SEND;
779 			break;
780 		case MLX4_OPCODE_RDMA_READ:
781 			wc->opcode    = IB_WC_RDMA_READ;
782 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
783 			break;
784 		case MLX4_OPCODE_ATOMIC_CS:
785 			wc->opcode    = IB_WC_COMP_SWAP;
786 			wc->byte_len  = 8;
787 			break;
788 		case MLX4_OPCODE_ATOMIC_FA:
789 			wc->opcode    = IB_WC_FETCH_ADD;
790 			wc->byte_len  = 8;
791 			break;
792 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
793 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
794 			wc->byte_len  = 8;
795 			break;
796 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
797 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
798 			wc->byte_len  = 8;
799 			break;
800 		case MLX4_OPCODE_LSO:
801 			wc->opcode    = IB_WC_LSO;
802 			break;
803 		case MLX4_OPCODE_FMR:
804 			wc->opcode    = IB_WC_REG_MR;
805 			break;
806 		case MLX4_OPCODE_LOCAL_INVAL:
807 			wc->opcode    = IB_WC_LOCAL_INV;
808 			break;
809 		}
810 	} else {
811 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
812 
813 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
814 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
815 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
816 			wc->wc_flags	= IB_WC_WITH_IMM;
817 			wc->ex.imm_data = cqe->immed_rss_invalid;
818 			break;
819 		case MLX4_RECV_OPCODE_SEND_INVAL:
820 			wc->opcode	= IB_WC_RECV;
821 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
822 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
823 			break;
824 		case MLX4_RECV_OPCODE_SEND:
825 			wc->opcode   = IB_WC_RECV;
826 			wc->wc_flags = 0;
827 			break;
828 		case MLX4_RECV_OPCODE_SEND_IMM:
829 			wc->opcode	= IB_WC_RECV;
830 			wc->wc_flags	= IB_WC_WITH_IMM;
831 			wc->ex.imm_data = cqe->immed_rss_invalid;
832 			break;
833 		}
834 
835 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
836 						  (*cur_qp)->port) ==
837 			  IB_LINK_LAYER_ETHERNET);
838 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
839 			if ((*cur_qp)->mlx4_ib_qp_type &
840 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
841 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
842 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
843 						is_eth);
844 				return 0;
845 			}
846 		}
847 
848 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
849 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
850 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
851 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
852 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
853 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
854 					cqe->badfcs_enc,
855 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
856 		if (is_eth) {
857 			wc->slid = 0;
858 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
859 			if (be32_to_cpu(cqe->vlan_my_qpn) &
860 					MLX4_CQE_CVLAN_PRESENT_MASK) {
861 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
862 					MLX4_CQE_VID_MASK;
863 			} else {
864 				wc->vlan_id = 0xffff;
865 			}
866 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
867 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
868 		} else {
869 			wc->slid = be16_to_cpu(cqe->rlid);
870 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
871 			wc->vlan_id = 0xffff;
872 		}
873 	}
874 
875 	return 0;
876 }
877 
878 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
879 {
880 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
881 	struct mlx4_ib_qp *cur_qp = NULL;
882 	unsigned long flags;
883 	int npolled;
884 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
885 
886 	spin_lock_irqsave(&cq->lock, flags);
887 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
888 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
889 		goto out;
890 	}
891 
892 	for (npolled = 0; npolled < num_entries; ++npolled) {
893 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
894 			break;
895 	}
896 
897 	mlx4_cq_set_ci(&cq->mcq);
898 
899 out:
900 	spin_unlock_irqrestore(&cq->lock, flags);
901 
902 	return npolled;
903 }
904 
905 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
906 {
907 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
908 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
909 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
910 		    to_mdev(ibcq->device)->uar_map,
911 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
912 
913 	return 0;
914 }
915 
916 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
917 {
918 	u32 prod_index;
919 	int nfreed = 0;
920 	struct mlx4_cqe *cqe, *dest;
921 	u8 owner_bit;
922 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
923 
924 	/*
925 	 * First we need to find the current producer index, so we
926 	 * know where to start cleaning from.  It doesn't matter if HW
927 	 * adds new entries after this loop -- the QP we're worried
928 	 * about is already in RESET, so the new entries won't come
929 	 * from our QP and therefore don't need to be checked.
930 	 */
931 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
932 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
933 			break;
934 
935 	/*
936 	 * Now sweep backwards through the CQ, removing CQ entries
937 	 * that match our QP by copying older entries on top of them.
938 	 */
939 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
940 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
941 		cqe += cqe_inc;
942 
943 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
944 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
945 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
946 			++nfreed;
947 		} else if (nfreed) {
948 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
949 			dest += cqe_inc;
950 
951 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
952 			memcpy(dest, cqe, sizeof *cqe);
953 			dest->owner_sr_opcode = owner_bit |
954 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
955 		}
956 	}
957 
958 	if (nfreed) {
959 		cq->mcq.cons_index += nfreed;
960 		/*
961 		 * Make sure update of buffer contents is done before
962 		 * updating consumer index.
963 		 */
964 		wmb();
965 		mlx4_cq_set_ci(&cq->mcq);
966 	}
967 }
968 
969 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
970 {
971 	spin_lock_irq(&cq->lock);
972 	__mlx4_ib_cq_clean(cq, qpn, srq);
973 	spin_unlock_irq(&cq->lock);
974 }
975