xref: /linux/drivers/infiniband/hw/mlx4/cq.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 "user.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_dbmap;
257 		}
258 
259 	return &cq->ibcq;
260 
261 err_dbmap:
262 	if (context)
263 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
264 
265 err_mtt:
266 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
267 
268 	if (context)
269 		ib_umem_release(cq->umem);
270 	else
271 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
272 
273 err_db:
274 	if (!context)
275 		mlx4_db_free(dev->dev, &cq->db);
276 
277 err_cq:
278 	kfree(cq);
279 
280 	return ERR_PTR(err);
281 }
282 
283 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
284 				  int entries)
285 {
286 	int err;
287 
288 	if (cq->resize_buf)
289 		return -EBUSY;
290 
291 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
292 	if (!cq->resize_buf)
293 		return -ENOMEM;
294 
295 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
296 	if (err) {
297 		kfree(cq->resize_buf);
298 		cq->resize_buf = NULL;
299 		return err;
300 	}
301 
302 	cq->resize_buf->cqe = entries - 1;
303 
304 	return 0;
305 }
306 
307 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
308 				   int entries, struct ib_udata *udata)
309 {
310 	struct mlx4_ib_resize_cq ucmd;
311 	int err;
312 
313 	if (cq->resize_umem)
314 		return -EBUSY;
315 
316 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
317 		return -EFAULT;
318 
319 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
320 	if (!cq->resize_buf)
321 		return -ENOMEM;
322 
323 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
324 				  &cq->resize_umem, ucmd.buf_addr, entries);
325 	if (err) {
326 		kfree(cq->resize_buf);
327 		cq->resize_buf = NULL;
328 		return err;
329 	}
330 
331 	cq->resize_buf->cqe = entries - 1;
332 
333 	return 0;
334 }
335 
336 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
337 {
338 	u32 i;
339 
340 	i = cq->mcq.cons_index;
341 	while (get_sw_cqe(cq, i))
342 		++i;
343 
344 	return i - cq->mcq.cons_index;
345 }
346 
347 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
348 {
349 	struct mlx4_cqe *cqe, *new_cqe;
350 	int i;
351 	int cqe_size = cq->buf.entry_size;
352 	int cqe_inc = cqe_size == 64 ? 1 : 0;
353 
354 	i = cq->mcq.cons_index;
355 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
356 	cqe += cqe_inc;
357 
358 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
359 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
360 					   (i + 1) & cq->resize_buf->cqe);
361 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
362 		new_cqe += cqe_inc;
363 
364 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
365 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
366 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
367 		cqe += cqe_inc;
368 	}
369 	++cq->mcq.cons_index;
370 }
371 
372 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
373 {
374 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
375 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
376 	struct mlx4_mtt mtt;
377 	int outst_cqe;
378 	int err;
379 
380 	mutex_lock(&cq->resize_mutex);
381 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
382 		err = -EINVAL;
383 		goto out;
384 	}
385 
386 	entries = roundup_pow_of_two(entries + 1);
387 	if (entries == ibcq->cqe + 1) {
388 		err = 0;
389 		goto out;
390 	}
391 
392 	if (entries > dev->dev->caps.max_cqes + 1) {
393 		err = -EINVAL;
394 		goto out;
395 	}
396 
397 	if (ibcq->uobject) {
398 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
399 		if (err)
400 			goto out;
401 	} else {
402 		/* Can't be smaller than the number of outstanding CQEs */
403 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
404 		if (entries < outst_cqe + 1) {
405 			err = -EINVAL;
406 			goto out;
407 		}
408 
409 		err = mlx4_alloc_resize_buf(dev, cq, entries);
410 		if (err)
411 			goto out;
412 	}
413 
414 	mtt = cq->buf.mtt;
415 
416 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
417 	if (err)
418 		goto err_buf;
419 
420 	mlx4_mtt_cleanup(dev->dev, &mtt);
421 	if (ibcq->uobject) {
422 		cq->buf      = cq->resize_buf->buf;
423 		cq->ibcq.cqe = cq->resize_buf->cqe;
424 		ib_umem_release(cq->umem);
425 		cq->umem     = cq->resize_umem;
426 
427 		kfree(cq->resize_buf);
428 		cq->resize_buf = NULL;
429 		cq->resize_umem = NULL;
430 	} else {
431 		struct mlx4_ib_cq_buf tmp_buf;
432 		int tmp_cqe = 0;
433 
434 		spin_lock_irq(&cq->lock);
435 		if (cq->resize_buf) {
436 			mlx4_ib_cq_resize_copy_cqes(cq);
437 			tmp_buf = cq->buf;
438 			tmp_cqe = cq->ibcq.cqe;
439 			cq->buf      = cq->resize_buf->buf;
440 			cq->ibcq.cqe = cq->resize_buf->cqe;
441 
442 			kfree(cq->resize_buf);
443 			cq->resize_buf = NULL;
444 		}
445 		spin_unlock_irq(&cq->lock);
446 
447 		if (tmp_cqe)
448 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
449 	}
450 
451 	goto out;
452 
453 err_buf:
454 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
455 	if (!ibcq->uobject)
456 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
457 				    cq->resize_buf->cqe);
458 
459 	kfree(cq->resize_buf);
460 	cq->resize_buf = NULL;
461 
462 	if (cq->resize_umem) {
463 		ib_umem_release(cq->resize_umem);
464 		cq->resize_umem = NULL;
465 	}
466 
467 out:
468 	mutex_unlock(&cq->resize_mutex);
469 
470 	return err;
471 }
472 
473 int mlx4_ib_destroy_cq(struct ib_cq *cq)
474 {
475 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
476 	struct mlx4_ib_cq *mcq = to_mcq(cq);
477 
478 	mlx4_cq_free(dev->dev, &mcq->mcq);
479 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
480 
481 	if (cq->uobject) {
482 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
483 		ib_umem_release(mcq->umem);
484 	} else {
485 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
486 		mlx4_db_free(dev->dev, &mcq->db);
487 	}
488 
489 	kfree(mcq);
490 
491 	return 0;
492 }
493 
494 static void dump_cqe(void *cqe)
495 {
496 	__be32 *buf = cqe;
497 
498 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
499 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
500 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
501 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
502 }
503 
504 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
505 				     struct ib_wc *wc)
506 {
507 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
508 		pr_debug("local QP operation err "
509 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
510 		       "opcode = %02x)\n",
511 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
512 		       cqe->vendor_err_syndrome,
513 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
514 		dump_cqe(cqe);
515 	}
516 
517 	switch (cqe->syndrome) {
518 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
519 		wc->status = IB_WC_LOC_LEN_ERR;
520 		break;
521 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
522 		wc->status = IB_WC_LOC_QP_OP_ERR;
523 		break;
524 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
525 		wc->status = IB_WC_LOC_PROT_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
528 		wc->status = IB_WC_WR_FLUSH_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
531 		wc->status = IB_WC_MW_BIND_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
534 		wc->status = IB_WC_BAD_RESP_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
537 		wc->status = IB_WC_LOC_ACCESS_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
540 		wc->status = IB_WC_REM_INV_REQ_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
543 		wc->status = IB_WC_REM_ACCESS_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
546 		wc->status = IB_WC_REM_OP_ERR;
547 		break;
548 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
549 		wc->status = IB_WC_RETRY_EXC_ERR;
550 		break;
551 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
552 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
553 		break;
554 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
555 		wc->status = IB_WC_REM_ABORT_ERR;
556 		break;
557 	default:
558 		wc->status = IB_WC_GENERAL_ERR;
559 		break;
560 	}
561 
562 	wc->vendor_err = cqe->vendor_err_syndrome;
563 }
564 
565 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
566 {
567 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
568 				      MLX4_CQE_STATUS_IPV4F     |
569 				      MLX4_CQE_STATUS_IPV4OPT   |
570 				      MLX4_CQE_STATUS_IPV6      |
571 				      MLX4_CQE_STATUS_IPOK)) ==
572 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
573 			    MLX4_CQE_STATUS_IPOK))              &&
574 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
575 				      MLX4_CQE_STATUS_TCP))     &&
576 		checksum == cpu_to_be16(0xffff);
577 }
578 
579 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
580 			   unsigned tail, struct mlx4_cqe *cqe, int is_eth)
581 {
582 	struct mlx4_ib_proxy_sqp_hdr *hdr;
583 
584 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
585 				   qp->sqp_proxy_rcv[tail].map,
586 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
587 				   DMA_FROM_DEVICE);
588 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
589 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
590 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
591 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
592 	wc->dlid_path_bits = 0;
593 
594 	if (is_eth) {
595 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
596 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
597 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
598 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
599 	} else {
600 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
601 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
602 	}
603 
604 	return 0;
605 }
606 
607 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
608 			       struct ib_wc *wc, int *npolled, int is_send)
609 {
610 	struct mlx4_ib_wq *wq;
611 	unsigned cur;
612 	int i;
613 
614 	wq = is_send ? &qp->sq : &qp->rq;
615 	cur = wq->head - wq->tail;
616 
617 	if (cur == 0)
618 		return;
619 
620 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
621 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
622 		wc->status = IB_WC_WR_FLUSH_ERR;
623 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
624 		wq->tail++;
625 		(*npolled)++;
626 		wc->qp = &qp->ibqp;
627 		wc++;
628 	}
629 }
630 
631 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
632 				 struct ib_wc *wc, int *npolled)
633 {
634 	struct mlx4_ib_qp *qp;
635 
636 	*npolled = 0;
637 	/* Find uncompleted WQEs belonging to that cq and retrun
638 	 * simulated FLUSH_ERR completions
639 	 */
640 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
641 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
642 		if (*npolled >= num_entries)
643 			goto out;
644 	}
645 
646 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
647 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
648 		if (*npolled >= num_entries)
649 			goto out;
650 	}
651 
652 out:
653 	return;
654 }
655 
656 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
657 			    struct mlx4_ib_qp **cur_qp,
658 			    struct ib_wc *wc)
659 {
660 	struct mlx4_cqe *cqe;
661 	struct mlx4_qp *mqp;
662 	struct mlx4_ib_wq *wq;
663 	struct mlx4_ib_srq *srq;
664 	struct mlx4_srq *msrq = NULL;
665 	int is_send;
666 	int is_error;
667 	int is_eth;
668 	u32 g_mlpath_rqpn;
669 	u16 wqe_ctr;
670 	unsigned tail = 0;
671 
672 repoll:
673 	cqe = next_cqe_sw(cq);
674 	if (!cqe)
675 		return -EAGAIN;
676 
677 	if (cq->buf.entry_size == 64)
678 		cqe++;
679 
680 	++cq->mcq.cons_index;
681 
682 	/*
683 	 * Make sure we read CQ entry contents after we've checked the
684 	 * ownership bit.
685 	 */
686 	rmb();
687 
688 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
689 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
690 		MLX4_CQE_OPCODE_ERROR;
691 
692 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
693 		     is_send)) {
694 		pr_warn("Completion for NOP opcode detected!\n");
695 		return -EINVAL;
696 	}
697 
698 	/* Resize CQ in progress */
699 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
700 		if (cq->resize_buf) {
701 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
702 
703 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
704 			cq->buf      = cq->resize_buf->buf;
705 			cq->ibcq.cqe = cq->resize_buf->cqe;
706 
707 			kfree(cq->resize_buf);
708 			cq->resize_buf = NULL;
709 		}
710 
711 		goto repoll;
712 	}
713 
714 	if (!*cur_qp ||
715 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
716 		/*
717 		 * We do not have to take the QP table lock here,
718 		 * because CQs will be locked while QPs are removed
719 		 * from the table.
720 		 */
721 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
722 				       be32_to_cpu(cqe->vlan_my_qpn));
723 		if (unlikely(!mqp)) {
724 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
725 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
726 			return -EINVAL;
727 		}
728 
729 		*cur_qp = to_mibqp(mqp);
730 	}
731 
732 	wc->qp = &(*cur_qp)->ibqp;
733 
734 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
735 		u32 srq_num;
736 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
737 		srq_num       = g_mlpath_rqpn & 0xffffff;
738 		/* SRQ is also in the radix tree */
739 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
740 				       srq_num);
741 		if (unlikely(!msrq)) {
742 			pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
743 				cq->mcq.cqn, srq_num);
744 			return -EINVAL;
745 		}
746 	}
747 
748 	if (is_send) {
749 		wq = &(*cur_qp)->sq;
750 		if (!(*cur_qp)->sq_signal_bits) {
751 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
752 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
753 		}
754 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
755 		++wq->tail;
756 	} else if ((*cur_qp)->ibqp.srq) {
757 		srq = to_msrq((*cur_qp)->ibqp.srq);
758 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
759 		wc->wr_id = srq->wrid[wqe_ctr];
760 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
761 	} else if (msrq) {
762 		srq = to_mibsrq(msrq);
763 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
764 		wc->wr_id = srq->wrid[wqe_ctr];
765 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
766 	} else {
767 		wq	  = &(*cur_qp)->rq;
768 		tail	  = wq->tail & (wq->wqe_cnt - 1);
769 		wc->wr_id = wq->wrid[tail];
770 		++wq->tail;
771 	}
772 
773 	if (unlikely(is_error)) {
774 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
775 		return 0;
776 	}
777 
778 	wc->status = IB_WC_SUCCESS;
779 
780 	if (is_send) {
781 		wc->wc_flags = 0;
782 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
783 		case MLX4_OPCODE_RDMA_WRITE_IMM:
784 			wc->wc_flags |= IB_WC_WITH_IMM;
785 		case MLX4_OPCODE_RDMA_WRITE:
786 			wc->opcode    = IB_WC_RDMA_WRITE;
787 			break;
788 		case MLX4_OPCODE_SEND_IMM:
789 			wc->wc_flags |= IB_WC_WITH_IMM;
790 		case MLX4_OPCODE_SEND:
791 		case MLX4_OPCODE_SEND_INVAL:
792 			wc->opcode    = IB_WC_SEND;
793 			break;
794 		case MLX4_OPCODE_RDMA_READ:
795 			wc->opcode    = IB_WC_RDMA_READ;
796 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
797 			break;
798 		case MLX4_OPCODE_ATOMIC_CS:
799 			wc->opcode    = IB_WC_COMP_SWAP;
800 			wc->byte_len  = 8;
801 			break;
802 		case MLX4_OPCODE_ATOMIC_FA:
803 			wc->opcode    = IB_WC_FETCH_ADD;
804 			wc->byte_len  = 8;
805 			break;
806 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
807 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
808 			wc->byte_len  = 8;
809 			break;
810 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
811 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
812 			wc->byte_len  = 8;
813 			break;
814 		case MLX4_OPCODE_BIND_MW:
815 			wc->opcode    = IB_WC_BIND_MW;
816 			break;
817 		case MLX4_OPCODE_LSO:
818 			wc->opcode    = IB_WC_LSO;
819 			break;
820 		case MLX4_OPCODE_FMR:
821 			wc->opcode    = IB_WC_FAST_REG_MR;
822 			break;
823 		case MLX4_OPCODE_LOCAL_INVAL:
824 			wc->opcode    = IB_WC_LOCAL_INV;
825 			break;
826 		}
827 	} else {
828 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
829 
830 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
831 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
832 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
833 			wc->wc_flags	= IB_WC_WITH_IMM;
834 			wc->ex.imm_data = cqe->immed_rss_invalid;
835 			break;
836 		case MLX4_RECV_OPCODE_SEND_INVAL:
837 			wc->opcode	= IB_WC_RECV;
838 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
839 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
840 			break;
841 		case MLX4_RECV_OPCODE_SEND:
842 			wc->opcode   = IB_WC_RECV;
843 			wc->wc_flags = 0;
844 			break;
845 		case MLX4_RECV_OPCODE_SEND_IMM:
846 			wc->opcode	= IB_WC_RECV;
847 			wc->wc_flags	= IB_WC_WITH_IMM;
848 			wc->ex.imm_data = cqe->immed_rss_invalid;
849 			break;
850 		}
851 
852 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
853 						  (*cur_qp)->port) ==
854 			  IB_LINK_LAYER_ETHERNET);
855 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
856 			if ((*cur_qp)->mlx4_ib_qp_type &
857 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
858 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
859 				return use_tunnel_data(*cur_qp, cq, wc, tail,
860 						       cqe, is_eth);
861 		}
862 
863 		wc->slid	   = be16_to_cpu(cqe->rlid);
864 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
865 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
866 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
867 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
868 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
869 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
870 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
871 		if (is_eth) {
872 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
873 			if (be32_to_cpu(cqe->vlan_my_qpn) &
874 					MLX4_CQE_CVLAN_PRESENT_MASK) {
875 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
876 					MLX4_CQE_VID_MASK;
877 			} else {
878 				wc->vlan_id = 0xffff;
879 			}
880 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
881 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
882 		} else {
883 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
884 			wc->vlan_id = 0xffff;
885 		}
886 	}
887 
888 	return 0;
889 }
890 
891 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
892 {
893 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
894 	struct mlx4_ib_qp *cur_qp = NULL;
895 	unsigned long flags;
896 	int npolled;
897 	int err = 0;
898 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
899 
900 	spin_lock_irqsave(&cq->lock, flags);
901 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
902 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
903 		goto out;
904 	}
905 
906 	for (npolled = 0; npolled < num_entries; ++npolled) {
907 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
908 		if (err)
909 			break;
910 	}
911 
912 	mlx4_cq_set_ci(&cq->mcq);
913 
914 out:
915 	spin_unlock_irqrestore(&cq->lock, flags);
916 
917 	if (err == 0 || err == -EAGAIN)
918 		return npolled;
919 	else
920 		return err;
921 }
922 
923 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
924 {
925 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
926 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
927 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
928 		    to_mdev(ibcq->device)->uar_map,
929 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
930 
931 	return 0;
932 }
933 
934 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
935 {
936 	u32 prod_index;
937 	int nfreed = 0;
938 	struct mlx4_cqe *cqe, *dest;
939 	u8 owner_bit;
940 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
941 
942 	/*
943 	 * First we need to find the current producer index, so we
944 	 * know where to start cleaning from.  It doesn't matter if HW
945 	 * adds new entries after this loop -- the QP we're worried
946 	 * about is already in RESET, so the new entries won't come
947 	 * from our QP and therefore don't need to be checked.
948 	 */
949 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
950 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
951 			break;
952 
953 	/*
954 	 * Now sweep backwards through the CQ, removing CQ entries
955 	 * that match our QP by copying older entries on top of them.
956 	 */
957 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
958 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
959 		cqe += cqe_inc;
960 
961 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
962 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
963 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
964 			++nfreed;
965 		} else if (nfreed) {
966 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
967 			dest += cqe_inc;
968 
969 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
970 			memcpy(dest, cqe, sizeof *cqe);
971 			dest->owner_sr_opcode = owner_bit |
972 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
973 		}
974 	}
975 
976 	if (nfreed) {
977 		cq->mcq.cons_index += nfreed;
978 		/*
979 		 * Make sure update of buffer contents is done before
980 		 * updating consumer index.
981 		 */
982 		wmb();
983 		mlx4_cq_set_ci(&cq->mcq);
984 	}
985 }
986 
987 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
988 {
989 	spin_lock_irq(&cq->lock);
990 	__mlx4_ib_cq_clean(cq, qpn, srq);
991 	spin_unlock_irq(&cq->lock);
992 }
993