xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision e553d2a5)
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, struct ib_udata *udata,
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(udata, buf_addr, cqe * cqe_size,
148 			    IB_ACCESS_LOCAL_WRITE, 1);
149 	if (IS_ERR(*umem))
150 		return PTR_ERR(*umem);
151 
152 	n = ib_umem_page_count(*umem);
153 	shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n);
154 	err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt);
155 
156 	if (err)
157 		goto err_buf;
158 
159 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
160 	if (err)
161 		goto err_mtt;
162 
163 	return 0;
164 
165 err_mtt:
166 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
167 
168 err_buf:
169 	ib_umem_release(*umem);
170 
171 	return err;
172 }
173 
174 #define CQ_CREATE_FLAGS_SUPPORTED IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION
175 int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
176 		      struct ib_udata *udata)
177 {
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, udata, &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, udata, &cq->resize_buf->buf,
332 				  &cq->resize_umem, 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 void 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 }
499 
500 static void dump_cqe(void *cqe)
501 {
502 	__be32 *buf = cqe;
503 
504 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
505 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
506 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
507 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
508 }
509 
510 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
511 				     struct ib_wc *wc)
512 {
513 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
514 		pr_debug("local QP operation err "
515 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
516 		       "opcode = %02x)\n",
517 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
518 		       cqe->vendor_err_syndrome,
519 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
520 		dump_cqe(cqe);
521 	}
522 
523 	switch (cqe->syndrome) {
524 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
525 		wc->status = IB_WC_LOC_LEN_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
528 		wc->status = IB_WC_LOC_QP_OP_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
531 		wc->status = IB_WC_LOC_PROT_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
534 		wc->status = IB_WC_WR_FLUSH_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
537 		wc->status = IB_WC_MW_BIND_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
540 		wc->status = IB_WC_BAD_RESP_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
543 		wc->status = IB_WC_LOC_ACCESS_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
546 		wc->status = IB_WC_REM_INV_REQ_ERR;
547 		break;
548 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
549 		wc->status = IB_WC_REM_ACCESS_ERR;
550 		break;
551 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
552 		wc->status = IB_WC_REM_OP_ERR;
553 		break;
554 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
555 		wc->status = IB_WC_RETRY_EXC_ERR;
556 		break;
557 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
558 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
559 		break;
560 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
561 		wc->status = IB_WC_REM_ABORT_ERR;
562 		break;
563 	default:
564 		wc->status = IB_WC_GENERAL_ERR;
565 		break;
566 	}
567 
568 	wc->vendor_err = cqe->vendor_err_syndrome;
569 }
570 
571 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
572 {
573 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
574 				      MLX4_CQE_STATUS_IPV4F     |
575 				      MLX4_CQE_STATUS_IPV4OPT   |
576 				      MLX4_CQE_STATUS_IPV6      |
577 				      MLX4_CQE_STATUS_IPOK)) ==
578 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
579 			    MLX4_CQE_STATUS_IPOK))              &&
580 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
581 				      MLX4_CQE_STATUS_TCP))     &&
582 		checksum == cpu_to_be16(0xffff);
583 }
584 
585 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
586 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
587 {
588 	struct mlx4_ib_proxy_sqp_hdr *hdr;
589 
590 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
591 				   qp->sqp_proxy_rcv[tail].map,
592 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
593 				   DMA_FROM_DEVICE);
594 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
595 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
596 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
597 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
598 	wc->dlid_path_bits = 0;
599 
600 	if (is_eth) {
601 		wc->slid = 0;
602 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
603 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
604 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
605 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
606 	} else {
607 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
608 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
609 	}
610 }
611 
612 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
613 			       struct ib_wc *wc, int *npolled, int is_send)
614 {
615 	struct mlx4_ib_wq *wq;
616 	unsigned cur;
617 	int i;
618 
619 	wq = is_send ? &qp->sq : &qp->rq;
620 	cur = wq->head - wq->tail;
621 
622 	if (cur == 0)
623 		return;
624 
625 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
626 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
627 		wc->status = IB_WC_WR_FLUSH_ERR;
628 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
629 		wq->tail++;
630 		(*npolled)++;
631 		wc->qp = &qp->ibqp;
632 		wc++;
633 	}
634 }
635 
636 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
637 				 struct ib_wc *wc, int *npolled)
638 {
639 	struct mlx4_ib_qp *qp;
640 
641 	*npolled = 0;
642 	/* Find uncompleted WQEs belonging to that cq and return
643 	 * simulated FLUSH_ERR completions
644 	 */
645 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
646 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
647 		if (*npolled >= num_entries)
648 			goto out;
649 	}
650 
651 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
652 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
653 		if (*npolled >= num_entries)
654 			goto out;
655 	}
656 
657 out:
658 	return;
659 }
660 
661 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
662 			    struct mlx4_ib_qp **cur_qp,
663 			    struct ib_wc *wc)
664 {
665 	struct mlx4_cqe *cqe;
666 	struct mlx4_qp *mqp;
667 	struct mlx4_ib_wq *wq;
668 	struct mlx4_ib_srq *srq;
669 	struct mlx4_srq *msrq = NULL;
670 	int is_send;
671 	int is_error;
672 	int is_eth;
673 	u32 g_mlpath_rqpn;
674 	u16 wqe_ctr;
675 	unsigned tail = 0;
676 
677 repoll:
678 	cqe = next_cqe_sw(cq);
679 	if (!cqe)
680 		return -EAGAIN;
681 
682 	if (cq->buf.entry_size == 64)
683 		cqe++;
684 
685 	++cq->mcq.cons_index;
686 
687 	/*
688 	 * Make sure we read CQ entry contents after we've checked the
689 	 * ownership bit.
690 	 */
691 	rmb();
692 
693 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
694 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
695 		MLX4_CQE_OPCODE_ERROR;
696 
697 	/* Resize CQ in progress */
698 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
699 		if (cq->resize_buf) {
700 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
701 
702 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
703 			cq->buf      = cq->resize_buf->buf;
704 			cq->ibcq.cqe = cq->resize_buf->cqe;
705 
706 			kfree(cq->resize_buf);
707 			cq->resize_buf = NULL;
708 		}
709 
710 		goto repoll;
711 	}
712 
713 	if (!*cur_qp ||
714 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
715 		/*
716 		 * We do not have to take the QP table lock here,
717 		 * because CQs will be locked while QPs are removed
718 		 * from the table.
719 		 */
720 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
721 				       be32_to_cpu(cqe->vlan_my_qpn));
722 		*cur_qp = to_mibqp(mqp);
723 	}
724 
725 	wc->qp = &(*cur_qp)->ibqp;
726 
727 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
728 		u32 srq_num;
729 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
730 		srq_num       = g_mlpath_rqpn & 0xffffff;
731 		/* SRQ is also in the radix tree */
732 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
733 				       srq_num);
734 	}
735 
736 	if (is_send) {
737 		wq = &(*cur_qp)->sq;
738 		if (!(*cur_qp)->sq_signal_bits) {
739 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
740 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
741 		}
742 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
743 		++wq->tail;
744 	} else if ((*cur_qp)->ibqp.srq) {
745 		srq = to_msrq((*cur_qp)->ibqp.srq);
746 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
747 		wc->wr_id = srq->wrid[wqe_ctr];
748 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
749 	} else if (msrq) {
750 		srq = to_mibsrq(msrq);
751 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
752 		wc->wr_id = srq->wrid[wqe_ctr];
753 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
754 	} else {
755 		wq	  = &(*cur_qp)->rq;
756 		tail	  = wq->tail & (wq->wqe_cnt - 1);
757 		wc->wr_id = wq->wrid[tail];
758 		++wq->tail;
759 	}
760 
761 	if (unlikely(is_error)) {
762 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
763 		return 0;
764 	}
765 
766 	wc->status = IB_WC_SUCCESS;
767 
768 	if (is_send) {
769 		wc->wc_flags = 0;
770 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
771 		case MLX4_OPCODE_RDMA_WRITE_IMM:
772 			wc->wc_flags |= IB_WC_WITH_IMM;
773 			/* fall through */
774 		case MLX4_OPCODE_RDMA_WRITE:
775 			wc->opcode    = IB_WC_RDMA_WRITE;
776 			break;
777 		case MLX4_OPCODE_SEND_IMM:
778 			wc->wc_flags |= IB_WC_WITH_IMM;
779 			/* fall through */
780 		case MLX4_OPCODE_SEND:
781 		case MLX4_OPCODE_SEND_INVAL:
782 			wc->opcode    = IB_WC_SEND;
783 			break;
784 		case MLX4_OPCODE_RDMA_READ:
785 			wc->opcode    = IB_WC_RDMA_READ;
786 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
787 			break;
788 		case MLX4_OPCODE_ATOMIC_CS:
789 			wc->opcode    = IB_WC_COMP_SWAP;
790 			wc->byte_len  = 8;
791 			break;
792 		case MLX4_OPCODE_ATOMIC_FA:
793 			wc->opcode    = IB_WC_FETCH_ADD;
794 			wc->byte_len  = 8;
795 			break;
796 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
797 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
798 			wc->byte_len  = 8;
799 			break;
800 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
801 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
802 			wc->byte_len  = 8;
803 			break;
804 		case MLX4_OPCODE_LSO:
805 			wc->opcode    = IB_WC_LSO;
806 			break;
807 		case MLX4_OPCODE_FMR:
808 			wc->opcode    = IB_WC_REG_MR;
809 			break;
810 		case MLX4_OPCODE_LOCAL_INVAL:
811 			wc->opcode    = IB_WC_LOCAL_INV;
812 			break;
813 		}
814 	} else {
815 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
816 
817 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
818 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
819 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
820 			wc->wc_flags	= IB_WC_WITH_IMM;
821 			wc->ex.imm_data = cqe->immed_rss_invalid;
822 			break;
823 		case MLX4_RECV_OPCODE_SEND_INVAL:
824 			wc->opcode	= IB_WC_RECV;
825 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
826 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
827 			break;
828 		case MLX4_RECV_OPCODE_SEND:
829 			wc->opcode   = IB_WC_RECV;
830 			wc->wc_flags = 0;
831 			break;
832 		case MLX4_RECV_OPCODE_SEND_IMM:
833 			wc->opcode	= IB_WC_RECV;
834 			wc->wc_flags	= IB_WC_WITH_IMM;
835 			wc->ex.imm_data = cqe->immed_rss_invalid;
836 			break;
837 		}
838 
839 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
840 						  (*cur_qp)->port) ==
841 			  IB_LINK_LAYER_ETHERNET);
842 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
843 			if ((*cur_qp)->mlx4_ib_qp_type &
844 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
845 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
846 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
847 						is_eth);
848 				return 0;
849 			}
850 		}
851 
852 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
853 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
854 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
855 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
856 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
857 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
858 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
859 		if (is_eth) {
860 			wc->slid = 0;
861 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
862 			if (be32_to_cpu(cqe->vlan_my_qpn) &
863 					MLX4_CQE_CVLAN_PRESENT_MASK) {
864 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
865 					MLX4_CQE_VID_MASK;
866 			} else {
867 				wc->vlan_id = 0xffff;
868 			}
869 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
870 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
871 		} else {
872 			wc->slid = be16_to_cpu(cqe->rlid);
873 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
874 			wc->vlan_id = 0xffff;
875 		}
876 	}
877 
878 	return 0;
879 }
880 
881 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
882 {
883 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
884 	struct mlx4_ib_qp *cur_qp = NULL;
885 	unsigned long flags;
886 	int npolled;
887 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
888 
889 	spin_lock_irqsave(&cq->lock, flags);
890 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
891 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
892 		goto out;
893 	}
894 
895 	for (npolled = 0; npolled < num_entries; ++npolled) {
896 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
897 			break;
898 	}
899 
900 	mlx4_cq_set_ci(&cq->mcq);
901 
902 out:
903 	spin_unlock_irqrestore(&cq->lock, flags);
904 
905 	return npolled;
906 }
907 
908 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
909 {
910 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
911 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
912 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
913 		    to_mdev(ibcq->device)->uar_map,
914 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
915 
916 	return 0;
917 }
918 
919 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
920 {
921 	u32 prod_index;
922 	int nfreed = 0;
923 	struct mlx4_cqe *cqe, *dest;
924 	u8 owner_bit;
925 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
926 
927 	/*
928 	 * First we need to find the current producer index, so we
929 	 * know where to start cleaning from.  It doesn't matter if HW
930 	 * adds new entries after this loop -- the QP we're worried
931 	 * about is already in RESET, so the new entries won't come
932 	 * from our QP and therefore don't need to be checked.
933 	 */
934 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
935 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
936 			break;
937 
938 	/*
939 	 * Now sweep backwards through the CQ, removing CQ entries
940 	 * that match our QP by copying older entries on top of them.
941 	 */
942 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
943 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
944 		cqe += cqe_inc;
945 
946 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
947 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
948 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
949 			++nfreed;
950 		} else if (nfreed) {
951 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
952 			dest += cqe_inc;
953 
954 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
955 			memcpy(dest, cqe, sizeof *cqe);
956 			dest->owner_sr_opcode = owner_bit |
957 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
958 		}
959 	}
960 
961 	if (nfreed) {
962 		cq->mcq.cons_index += nfreed;
963 		/*
964 		 * Make sure update of buffer contents is done before
965 		 * updating consumer index.
966 		 */
967 		wmb();
968 		mlx4_cq_set_ci(&cq->mcq);
969 	}
970 }
971 
972 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
973 {
974 	spin_lock_irq(&cq->lock);
975 	__mlx4_ib_cq_clean(cq, qpn, srq);
976 	spin_unlock_irq(&cq->lock);
977 }
978