xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision e639c869)
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);
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);
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 	int shift;
144 	int n;
145 
146 	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
147 			    IB_ACCESS_LOCAL_WRITE, 1);
148 	if (IS_ERR(*umem))
149 		return PTR_ERR(*umem);
150 
151 	n = ib_umem_page_count(*umem);
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_CQ_FLAGS_TIMESTAMP_COMPLETION
174 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
175 				const struct ib_cq_init_attr *attr,
176 				struct ib_ucontext *context,
177 				struct ib_udata *udata)
178 {
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;
183 	struct mlx4_uar *uar;
184 	int err;
185 
186 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
187 		return ERR_PTR(-EINVAL);
188 
189 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
190 		return ERR_PTR(-EINVAL);
191 
192 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
193 	if (!cq)
194 		return ERR_PTR(-ENOMEM);
195 
196 	entries      = roundup_pow_of_two(entries + 1);
197 	cq->ibcq.cqe = entries - 1;
198 	mutex_init(&cq->resize_mutex);
199 	spin_lock_init(&cq->lock);
200 	cq->resize_buf = NULL;
201 	cq->resize_umem = NULL;
202 	cq->create_flags = attr->flags;
203 	INIT_LIST_HEAD(&cq->send_qp_list);
204 	INIT_LIST_HEAD(&cq->recv_qp_list);
205 
206 	if (context) {
207 		struct mlx4_ib_create_cq ucmd;
208 
209 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
210 			err = -EFAULT;
211 			goto err_cq;
212 		}
213 
214 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
215 					  ucmd.buf_addr, entries);
216 		if (err)
217 			goto err_cq;
218 
219 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
220 					  &cq->db);
221 		if (err)
222 			goto err_mtt;
223 
224 		uar = &to_mucontext(context)->uar;
225 		cq->mcq.usage = MLX4_RES_USAGE_USER_VERBS;
226 	} else {
227 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
228 		if (err)
229 			goto err_cq;
230 
231 		cq->mcq.set_ci_db  = cq->db.db;
232 		cq->mcq.arm_db     = cq->db.db + 1;
233 		*cq->mcq.set_ci_db = 0;
234 		*cq->mcq.arm_db    = 0;
235 
236 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
237 		if (err)
238 			goto err_db;
239 
240 		uar = &dev->priv_uar;
241 		cq->mcq.usage = MLX4_RES_USAGE_DRIVER;
242 	}
243 
244 	if (dev->eq_table)
245 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
246 
247 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
248 			    cq->db.dma, &cq->mcq, vector, 0,
249 			    !!(cq->create_flags & IB_CQ_FLAGS_TIMESTAMP_COMPLETION));
250 	if (err)
251 		goto err_dbmap;
252 
253 	if (context)
254 		cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
255 	else
256 		cq->mcq.comp = mlx4_ib_cq_comp;
257 	cq->mcq.event = mlx4_ib_cq_event;
258 
259 	if (context)
260 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
261 			err = -EFAULT;
262 			goto err_cq_free;
263 		}
264 
265 	return &cq->ibcq;
266 
267 err_cq_free:
268 	mlx4_cq_free(dev->dev, &cq->mcq);
269 
270 err_dbmap:
271 	if (context)
272 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
273 
274 err_mtt:
275 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
276 
277 	if (context)
278 		ib_umem_release(cq->umem);
279 	else
280 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
281 
282 err_db:
283 	if (!context)
284 		mlx4_db_free(dev->dev, &cq->db);
285 
286 err_cq:
287 	kfree(cq);
288 
289 	return ERR_PTR(err);
290 }
291 
292 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
293 				  int entries)
294 {
295 	int err;
296 
297 	if (cq->resize_buf)
298 		return -EBUSY;
299 
300 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
301 	if (!cq->resize_buf)
302 		return -ENOMEM;
303 
304 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
305 	if (err) {
306 		kfree(cq->resize_buf);
307 		cq->resize_buf = NULL;
308 		return err;
309 	}
310 
311 	cq->resize_buf->cqe = entries - 1;
312 
313 	return 0;
314 }
315 
316 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
317 				   int entries, struct ib_udata *udata)
318 {
319 	struct mlx4_ib_resize_cq ucmd;
320 	int err;
321 
322 	if (cq->resize_umem)
323 		return -EBUSY;
324 
325 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
326 		return -EFAULT;
327 
328 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
329 	if (!cq->resize_buf)
330 		return -ENOMEM;
331 
332 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
333 				  &cq->resize_umem, ucmd.buf_addr, entries);
334 	if (err) {
335 		kfree(cq->resize_buf);
336 		cq->resize_buf = NULL;
337 		return err;
338 	}
339 
340 	cq->resize_buf->cqe = entries - 1;
341 
342 	return 0;
343 }
344 
345 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
346 {
347 	u32 i;
348 
349 	i = cq->mcq.cons_index;
350 	while (get_sw_cqe(cq, i))
351 		++i;
352 
353 	return i - cq->mcq.cons_index;
354 }
355 
356 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
357 {
358 	struct mlx4_cqe *cqe, *new_cqe;
359 	int i;
360 	int cqe_size = cq->buf.entry_size;
361 	int cqe_inc = cqe_size == 64 ? 1 : 0;
362 
363 	i = cq->mcq.cons_index;
364 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
365 	cqe += cqe_inc;
366 
367 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
368 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
369 					   (i + 1) & cq->resize_buf->cqe);
370 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
371 		new_cqe += cqe_inc;
372 
373 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
374 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
375 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
376 		cqe += cqe_inc;
377 	}
378 	++cq->mcq.cons_index;
379 }
380 
381 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
382 {
383 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
384 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
385 	struct mlx4_mtt mtt;
386 	int outst_cqe;
387 	int err;
388 
389 	mutex_lock(&cq->resize_mutex);
390 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
391 		err = -EINVAL;
392 		goto out;
393 	}
394 
395 	entries = roundup_pow_of_two(entries + 1);
396 	if (entries == ibcq->cqe + 1) {
397 		err = 0;
398 		goto out;
399 	}
400 
401 	if (entries > dev->dev->caps.max_cqes + 1) {
402 		err = -EINVAL;
403 		goto out;
404 	}
405 
406 	if (ibcq->uobject) {
407 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
408 		if (err)
409 			goto out;
410 	} else {
411 		/* Can't be smaller than the number of outstanding CQEs */
412 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
413 		if (entries < outst_cqe + 1) {
414 			err = -EINVAL;
415 			goto out;
416 		}
417 
418 		err = mlx4_alloc_resize_buf(dev, cq, entries);
419 		if (err)
420 			goto out;
421 	}
422 
423 	mtt = cq->buf.mtt;
424 
425 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
426 	if (err)
427 		goto err_buf;
428 
429 	mlx4_mtt_cleanup(dev->dev, &mtt);
430 	if (ibcq->uobject) {
431 		cq->buf      = cq->resize_buf->buf;
432 		cq->ibcq.cqe = cq->resize_buf->cqe;
433 		ib_umem_release(cq->umem);
434 		cq->umem     = cq->resize_umem;
435 
436 		kfree(cq->resize_buf);
437 		cq->resize_buf = NULL;
438 		cq->resize_umem = NULL;
439 	} else {
440 		struct mlx4_ib_cq_buf tmp_buf;
441 		int tmp_cqe = 0;
442 
443 		spin_lock_irq(&cq->lock);
444 		if (cq->resize_buf) {
445 			mlx4_ib_cq_resize_copy_cqes(cq);
446 			tmp_buf = cq->buf;
447 			tmp_cqe = cq->ibcq.cqe;
448 			cq->buf      = cq->resize_buf->buf;
449 			cq->ibcq.cqe = cq->resize_buf->cqe;
450 
451 			kfree(cq->resize_buf);
452 			cq->resize_buf = NULL;
453 		}
454 		spin_unlock_irq(&cq->lock);
455 
456 		if (tmp_cqe)
457 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
458 	}
459 
460 	goto out;
461 
462 err_buf:
463 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
464 	if (!ibcq->uobject)
465 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
466 				    cq->resize_buf->cqe);
467 
468 	kfree(cq->resize_buf);
469 	cq->resize_buf = NULL;
470 
471 	if (cq->resize_umem) {
472 		ib_umem_release(cq->resize_umem);
473 		cq->resize_umem = NULL;
474 	}
475 
476 out:
477 	mutex_unlock(&cq->resize_mutex);
478 
479 	return err;
480 }
481 
482 int mlx4_ib_destroy_cq(struct ib_cq *cq)
483 {
484 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
485 	struct mlx4_ib_cq *mcq = to_mcq(cq);
486 
487 	mlx4_cq_free(dev->dev, &mcq->mcq);
488 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
489 
490 	if (cq->uobject) {
491 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
492 		ib_umem_release(mcq->umem);
493 	} else {
494 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
495 		mlx4_db_free(dev->dev, &mcq->db);
496 	}
497 
498 	kfree(mcq);
499 
500 	return 0;
501 }
502 
503 static void dump_cqe(void *cqe)
504 {
505 	__be32 *buf = cqe;
506 
507 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
508 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
509 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
510 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
511 }
512 
513 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
514 				     struct ib_wc *wc)
515 {
516 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
517 		pr_debug("local QP operation err "
518 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
519 		       "opcode = %02x)\n",
520 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
521 		       cqe->vendor_err_syndrome,
522 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
523 		dump_cqe(cqe);
524 	}
525 
526 	switch (cqe->syndrome) {
527 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
528 		wc->status = IB_WC_LOC_LEN_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
531 		wc->status = IB_WC_LOC_QP_OP_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
534 		wc->status = IB_WC_LOC_PROT_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
537 		wc->status = IB_WC_WR_FLUSH_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
540 		wc->status = IB_WC_MW_BIND_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
543 		wc->status = IB_WC_BAD_RESP_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
546 		wc->status = IB_WC_LOC_ACCESS_ERR;
547 		break;
548 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
549 		wc->status = IB_WC_REM_INV_REQ_ERR;
550 		break;
551 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
552 		wc->status = IB_WC_REM_ACCESS_ERR;
553 		break;
554 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
555 		wc->status = IB_WC_REM_OP_ERR;
556 		break;
557 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
558 		wc->status = IB_WC_RETRY_EXC_ERR;
559 		break;
560 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
561 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
562 		break;
563 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
564 		wc->status = IB_WC_REM_ABORT_ERR;
565 		break;
566 	default:
567 		wc->status = IB_WC_GENERAL_ERR;
568 		break;
569 	}
570 
571 	wc->vendor_err = cqe->vendor_err_syndrome;
572 }
573 
574 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
575 {
576 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
577 				      MLX4_CQE_STATUS_IPV4F     |
578 				      MLX4_CQE_STATUS_IPV4OPT   |
579 				      MLX4_CQE_STATUS_IPV6      |
580 				      MLX4_CQE_STATUS_IPOK)) ==
581 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
582 			    MLX4_CQE_STATUS_IPOK))              &&
583 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
584 				      MLX4_CQE_STATUS_TCP))     &&
585 		checksum == cpu_to_be16(0xffff);
586 }
587 
588 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
589 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
590 {
591 	struct mlx4_ib_proxy_sqp_hdr *hdr;
592 
593 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
594 				   qp->sqp_proxy_rcv[tail].map,
595 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
596 				   DMA_FROM_DEVICE);
597 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
598 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
599 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
600 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
601 	wc->dlid_path_bits = 0;
602 
603 	if (is_eth) {
604 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
605 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
606 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
607 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
608 	} else {
609 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
610 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
611 	}
612 }
613 
614 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
615 			       struct ib_wc *wc, int *npolled, int is_send)
616 {
617 	struct mlx4_ib_wq *wq;
618 	unsigned cur;
619 	int i;
620 
621 	wq = is_send ? &qp->sq : &qp->rq;
622 	cur = wq->head - wq->tail;
623 
624 	if (cur == 0)
625 		return;
626 
627 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
628 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
629 		wc->status = IB_WC_WR_FLUSH_ERR;
630 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
631 		wq->tail++;
632 		(*npolled)++;
633 		wc->qp = &qp->ibqp;
634 		wc++;
635 	}
636 }
637 
638 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
639 				 struct ib_wc *wc, int *npolled)
640 {
641 	struct mlx4_ib_qp *qp;
642 
643 	*npolled = 0;
644 	/* Find uncompleted WQEs belonging to that cq and return
645 	 * simulated FLUSH_ERR completions
646 	 */
647 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
648 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
649 		if (*npolled >= num_entries)
650 			goto out;
651 	}
652 
653 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
654 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
655 		if (*npolled >= num_entries)
656 			goto out;
657 	}
658 
659 out:
660 	return;
661 }
662 
663 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
664 			    struct mlx4_ib_qp **cur_qp,
665 			    struct ib_wc *wc)
666 {
667 	struct mlx4_cqe *cqe;
668 	struct mlx4_qp *mqp;
669 	struct mlx4_ib_wq *wq;
670 	struct mlx4_ib_srq *srq;
671 	struct mlx4_srq *msrq = NULL;
672 	int is_send;
673 	int is_error;
674 	int is_eth;
675 	u32 g_mlpath_rqpn;
676 	u16 wqe_ctr;
677 	unsigned tail = 0;
678 
679 repoll:
680 	cqe = next_cqe_sw(cq);
681 	if (!cqe)
682 		return -EAGAIN;
683 
684 	if (cq->buf.entry_size == 64)
685 		cqe++;
686 
687 	++cq->mcq.cons_index;
688 
689 	/*
690 	 * Make sure we read CQ entry contents after we've checked the
691 	 * ownership bit.
692 	 */
693 	rmb();
694 
695 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
696 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
697 		MLX4_CQE_OPCODE_ERROR;
698 
699 	/* Resize CQ in progress */
700 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
701 		if (cq->resize_buf) {
702 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
703 
704 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
705 			cq->buf      = cq->resize_buf->buf;
706 			cq->ibcq.cqe = cq->resize_buf->cqe;
707 
708 			kfree(cq->resize_buf);
709 			cq->resize_buf = NULL;
710 		}
711 
712 		goto repoll;
713 	}
714 
715 	if (!*cur_qp ||
716 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
717 		/*
718 		 * We do not have to take the QP table lock here,
719 		 * because CQs will be locked while QPs are removed
720 		 * from the table.
721 		 */
722 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
723 				       be32_to_cpu(cqe->vlan_my_qpn));
724 		*cur_qp = to_mibqp(mqp);
725 	}
726 
727 	wc->qp = &(*cur_qp)->ibqp;
728 
729 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
730 		u32 srq_num;
731 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
732 		srq_num       = g_mlpath_rqpn & 0xffffff;
733 		/* SRQ is also in the radix tree */
734 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
735 				       srq_num);
736 	}
737 
738 	if (is_send) {
739 		wq = &(*cur_qp)->sq;
740 		if (!(*cur_qp)->sq_signal_bits) {
741 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
742 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
743 		}
744 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
745 		++wq->tail;
746 	} else if ((*cur_qp)->ibqp.srq) {
747 		srq = to_msrq((*cur_qp)->ibqp.srq);
748 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
749 		wc->wr_id = srq->wrid[wqe_ctr];
750 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
751 	} else if (msrq) {
752 		srq = to_mibsrq(msrq);
753 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
754 		wc->wr_id = srq->wrid[wqe_ctr];
755 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
756 	} else {
757 		wq	  = &(*cur_qp)->rq;
758 		tail	  = wq->tail & (wq->wqe_cnt - 1);
759 		wc->wr_id = wq->wrid[tail];
760 		++wq->tail;
761 	}
762 
763 	if (unlikely(is_error)) {
764 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
765 		return 0;
766 	}
767 
768 	wc->status = IB_WC_SUCCESS;
769 
770 	if (is_send) {
771 		wc->wc_flags = 0;
772 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
773 		case MLX4_OPCODE_RDMA_WRITE_IMM:
774 			wc->wc_flags |= IB_WC_WITH_IMM;
775 			/* fall through */
776 		case MLX4_OPCODE_RDMA_WRITE:
777 			wc->opcode    = IB_WC_RDMA_WRITE;
778 			break;
779 		case MLX4_OPCODE_SEND_IMM:
780 			wc->wc_flags |= IB_WC_WITH_IMM;
781 			/* fall through */
782 		case MLX4_OPCODE_SEND:
783 		case MLX4_OPCODE_SEND_INVAL:
784 			wc->opcode    = IB_WC_SEND;
785 			break;
786 		case MLX4_OPCODE_RDMA_READ:
787 			wc->opcode    = IB_WC_RDMA_READ;
788 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
789 			break;
790 		case MLX4_OPCODE_ATOMIC_CS:
791 			wc->opcode    = IB_WC_COMP_SWAP;
792 			wc->byte_len  = 8;
793 			break;
794 		case MLX4_OPCODE_ATOMIC_FA:
795 			wc->opcode    = IB_WC_FETCH_ADD;
796 			wc->byte_len  = 8;
797 			break;
798 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
799 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
800 			wc->byte_len  = 8;
801 			break;
802 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
803 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
804 			wc->byte_len  = 8;
805 			break;
806 		case MLX4_OPCODE_LSO:
807 			wc->opcode    = IB_WC_LSO;
808 			break;
809 		case MLX4_OPCODE_FMR:
810 			wc->opcode    = IB_WC_REG_MR;
811 			break;
812 		case MLX4_OPCODE_LOCAL_INVAL:
813 			wc->opcode    = IB_WC_LOCAL_INV;
814 			break;
815 		}
816 	} else {
817 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
818 
819 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
820 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
821 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
822 			wc->wc_flags	= IB_WC_WITH_IMM;
823 			wc->ex.imm_data = cqe->immed_rss_invalid;
824 			break;
825 		case MLX4_RECV_OPCODE_SEND_INVAL:
826 			wc->opcode	= IB_WC_RECV;
827 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
828 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
829 			break;
830 		case MLX4_RECV_OPCODE_SEND:
831 			wc->opcode   = IB_WC_RECV;
832 			wc->wc_flags = 0;
833 			break;
834 		case MLX4_RECV_OPCODE_SEND_IMM:
835 			wc->opcode	= IB_WC_RECV;
836 			wc->wc_flags	= IB_WC_WITH_IMM;
837 			wc->ex.imm_data = cqe->immed_rss_invalid;
838 			break;
839 		}
840 
841 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
842 						  (*cur_qp)->port) ==
843 			  IB_LINK_LAYER_ETHERNET);
844 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
845 			if ((*cur_qp)->mlx4_ib_qp_type &
846 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
847 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
848 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
849 						is_eth);
850 				return 0;
851 			}
852 		}
853 
854 		wc->slid	   = be16_to_cpu(cqe->rlid);
855 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
856 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
857 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
858 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
859 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
860 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
861 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
862 		if (is_eth) {
863 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
864 			if (be32_to_cpu(cqe->vlan_my_qpn) &
865 					MLX4_CQE_CVLAN_PRESENT_MASK) {
866 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
867 					MLX4_CQE_VID_MASK;
868 			} else {
869 				wc->vlan_id = 0xffff;
870 			}
871 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
872 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
873 		} else {
874 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
875 			wc->vlan_id = 0xffff;
876 		}
877 	}
878 
879 	return 0;
880 }
881 
882 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
883 {
884 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
885 	struct mlx4_ib_qp *cur_qp = NULL;
886 	unsigned long flags;
887 	int npolled;
888 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
889 
890 	spin_lock_irqsave(&cq->lock, flags);
891 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
892 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
893 		goto out;
894 	}
895 
896 	for (npolled = 0; npolled < num_entries; ++npolled) {
897 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
898 			break;
899 	}
900 
901 	mlx4_cq_set_ci(&cq->mcq);
902 
903 out:
904 	spin_unlock_irqrestore(&cq->lock, flags);
905 
906 	return npolled;
907 }
908 
909 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
910 {
911 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
912 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
913 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
914 		    to_mdev(ibcq->device)->uar_map,
915 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
916 
917 	return 0;
918 }
919 
920 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
921 {
922 	u32 prod_index;
923 	int nfreed = 0;
924 	struct mlx4_cqe *cqe, *dest;
925 	u8 owner_bit;
926 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
927 
928 	/*
929 	 * First we need to find the current producer index, so we
930 	 * know where to start cleaning from.  It doesn't matter if HW
931 	 * adds new entries after this loop -- the QP we're worried
932 	 * about is already in RESET, so the new entries won't come
933 	 * from our QP and therefore don't need to be checked.
934 	 */
935 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
936 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
937 			break;
938 
939 	/*
940 	 * Now sweep backwards through the CQ, removing CQ entries
941 	 * that match our QP by copying older entries on top of them.
942 	 */
943 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
944 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
945 		cqe += cqe_inc;
946 
947 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
948 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
949 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
950 			++nfreed;
951 		} else if (nfreed) {
952 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
953 			dest += cqe_inc;
954 
955 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
956 			memcpy(dest, cqe, sizeof *cqe);
957 			dest->owner_sr_opcode = owner_bit |
958 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
959 		}
960 	}
961 
962 	if (nfreed) {
963 		cq->mcq.cons_index += nfreed;
964 		/*
965 		 * Make sure update of buffer contents is done before
966 		 * updating consumer index.
967 		 */
968 		wmb();
969 		mlx4_cq_set_ci(&cq->mcq);
970 	}
971 }
972 
973 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
974 {
975 	spin_lock_irq(&cq->lock);
976 	__mlx4_ib_cq_clean(cq, qpn, srq);
977 	spin_unlock_irq(&cq->lock);
978 }
979