xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision 4f3db074)
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 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
170 				struct ib_ucontext *context,
171 				struct ib_udata *udata)
172 {
173 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
174 	struct mlx4_ib_cq *cq;
175 	struct mlx4_uar *uar;
176 	int err;
177 
178 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
179 		return ERR_PTR(-EINVAL);
180 
181 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
182 	if (!cq)
183 		return ERR_PTR(-ENOMEM);
184 
185 	entries      = roundup_pow_of_two(entries + 1);
186 	cq->ibcq.cqe = entries - 1;
187 	mutex_init(&cq->resize_mutex);
188 	spin_lock_init(&cq->lock);
189 	cq->resize_buf = NULL;
190 	cq->resize_umem = NULL;
191 	INIT_LIST_HEAD(&cq->send_qp_list);
192 	INIT_LIST_HEAD(&cq->recv_qp_list);
193 
194 	if (context) {
195 		struct mlx4_ib_create_cq ucmd;
196 
197 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
198 			err = -EFAULT;
199 			goto err_cq;
200 		}
201 
202 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
203 					  ucmd.buf_addr, entries);
204 		if (err)
205 			goto err_cq;
206 
207 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
208 					  &cq->db);
209 		if (err)
210 			goto err_mtt;
211 
212 		uar = &to_mucontext(context)->uar;
213 	} else {
214 		err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
215 		if (err)
216 			goto err_cq;
217 
218 		cq->mcq.set_ci_db  = cq->db.db;
219 		cq->mcq.arm_db     = cq->db.db + 1;
220 		*cq->mcq.set_ci_db = 0;
221 		*cq->mcq.arm_db    = 0;
222 
223 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
224 		if (err)
225 			goto err_db;
226 
227 		uar = &dev->priv_uar;
228 	}
229 
230 	if (dev->eq_table)
231 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
232 
233 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
234 			    cq->db.dma, &cq->mcq, vector, 0, 0);
235 	if (err)
236 		goto err_dbmap;
237 
238 	if (context)
239 		cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
240 	else
241 		cq->mcq.comp = mlx4_ib_cq_comp;
242 	cq->mcq.event = mlx4_ib_cq_event;
243 
244 	if (context)
245 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
246 			err = -EFAULT;
247 			goto err_dbmap;
248 		}
249 
250 	return &cq->ibcq;
251 
252 err_dbmap:
253 	if (context)
254 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
255 
256 err_mtt:
257 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
258 
259 	if (context)
260 		ib_umem_release(cq->umem);
261 	else
262 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
263 
264 err_db:
265 	if (!context)
266 		mlx4_db_free(dev->dev, &cq->db);
267 
268 err_cq:
269 	kfree(cq);
270 
271 	return ERR_PTR(err);
272 }
273 
274 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
275 				  int entries)
276 {
277 	int err;
278 
279 	if (cq->resize_buf)
280 		return -EBUSY;
281 
282 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
283 	if (!cq->resize_buf)
284 		return -ENOMEM;
285 
286 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
287 	if (err) {
288 		kfree(cq->resize_buf);
289 		cq->resize_buf = NULL;
290 		return err;
291 	}
292 
293 	cq->resize_buf->cqe = entries - 1;
294 
295 	return 0;
296 }
297 
298 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
299 				   int entries, struct ib_udata *udata)
300 {
301 	struct mlx4_ib_resize_cq ucmd;
302 	int err;
303 
304 	if (cq->resize_umem)
305 		return -EBUSY;
306 
307 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
308 		return -EFAULT;
309 
310 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
311 	if (!cq->resize_buf)
312 		return -ENOMEM;
313 
314 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
315 				  &cq->resize_umem, ucmd.buf_addr, entries);
316 	if (err) {
317 		kfree(cq->resize_buf);
318 		cq->resize_buf = NULL;
319 		return err;
320 	}
321 
322 	cq->resize_buf->cqe = entries - 1;
323 
324 	return 0;
325 }
326 
327 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
328 {
329 	u32 i;
330 
331 	i = cq->mcq.cons_index;
332 	while (get_sw_cqe(cq, i))
333 		++i;
334 
335 	return i - cq->mcq.cons_index;
336 }
337 
338 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
339 {
340 	struct mlx4_cqe *cqe, *new_cqe;
341 	int i;
342 	int cqe_size = cq->buf.entry_size;
343 	int cqe_inc = cqe_size == 64 ? 1 : 0;
344 
345 	i = cq->mcq.cons_index;
346 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
347 	cqe += cqe_inc;
348 
349 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
350 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
351 					   (i + 1) & cq->resize_buf->cqe);
352 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
353 		new_cqe += cqe_inc;
354 
355 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
356 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
357 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
358 		cqe += cqe_inc;
359 	}
360 	++cq->mcq.cons_index;
361 }
362 
363 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
364 {
365 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
366 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
367 	struct mlx4_mtt mtt;
368 	int outst_cqe;
369 	int err;
370 
371 	mutex_lock(&cq->resize_mutex);
372 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
373 		err = -EINVAL;
374 		goto out;
375 	}
376 
377 	entries = roundup_pow_of_two(entries + 1);
378 	if (entries == ibcq->cqe + 1) {
379 		err = 0;
380 		goto out;
381 	}
382 
383 	if (entries > dev->dev->caps.max_cqes + 1) {
384 		err = -EINVAL;
385 		goto out;
386 	}
387 
388 	if (ibcq->uobject) {
389 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
390 		if (err)
391 			goto out;
392 	} else {
393 		/* Can't be smaller than the number of outstanding CQEs */
394 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
395 		if (entries < outst_cqe + 1) {
396 			err = -EINVAL;
397 			goto out;
398 		}
399 
400 		err = mlx4_alloc_resize_buf(dev, cq, entries);
401 		if (err)
402 			goto out;
403 	}
404 
405 	mtt = cq->buf.mtt;
406 
407 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
408 	if (err)
409 		goto err_buf;
410 
411 	mlx4_mtt_cleanup(dev->dev, &mtt);
412 	if (ibcq->uobject) {
413 		cq->buf      = cq->resize_buf->buf;
414 		cq->ibcq.cqe = cq->resize_buf->cqe;
415 		ib_umem_release(cq->umem);
416 		cq->umem     = cq->resize_umem;
417 
418 		kfree(cq->resize_buf);
419 		cq->resize_buf = NULL;
420 		cq->resize_umem = NULL;
421 	} else {
422 		struct mlx4_ib_cq_buf tmp_buf;
423 		int tmp_cqe = 0;
424 
425 		spin_lock_irq(&cq->lock);
426 		if (cq->resize_buf) {
427 			mlx4_ib_cq_resize_copy_cqes(cq);
428 			tmp_buf = cq->buf;
429 			tmp_cqe = cq->ibcq.cqe;
430 			cq->buf      = cq->resize_buf->buf;
431 			cq->ibcq.cqe = cq->resize_buf->cqe;
432 
433 			kfree(cq->resize_buf);
434 			cq->resize_buf = NULL;
435 		}
436 		spin_unlock_irq(&cq->lock);
437 
438 		if (tmp_cqe)
439 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
440 	}
441 
442 	goto out;
443 
444 err_buf:
445 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
446 	if (!ibcq->uobject)
447 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
448 				    cq->resize_buf->cqe);
449 
450 	kfree(cq->resize_buf);
451 	cq->resize_buf = NULL;
452 
453 	if (cq->resize_umem) {
454 		ib_umem_release(cq->resize_umem);
455 		cq->resize_umem = NULL;
456 	}
457 
458 out:
459 	mutex_unlock(&cq->resize_mutex);
460 
461 	return err;
462 }
463 
464 int mlx4_ib_destroy_cq(struct ib_cq *cq)
465 {
466 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
467 	struct mlx4_ib_cq *mcq = to_mcq(cq);
468 
469 	mlx4_cq_free(dev->dev, &mcq->mcq);
470 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
471 
472 	if (cq->uobject) {
473 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
474 		ib_umem_release(mcq->umem);
475 	} else {
476 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
477 		mlx4_db_free(dev->dev, &mcq->db);
478 	}
479 
480 	kfree(mcq);
481 
482 	return 0;
483 }
484 
485 static void dump_cqe(void *cqe)
486 {
487 	__be32 *buf = cqe;
488 
489 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
490 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
491 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
492 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
493 }
494 
495 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
496 				     struct ib_wc *wc)
497 {
498 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
499 		pr_debug("local QP operation err "
500 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
501 		       "opcode = %02x)\n",
502 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
503 		       cqe->vendor_err_syndrome,
504 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
505 		dump_cqe(cqe);
506 	}
507 
508 	switch (cqe->syndrome) {
509 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
510 		wc->status = IB_WC_LOC_LEN_ERR;
511 		break;
512 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
513 		wc->status = IB_WC_LOC_QP_OP_ERR;
514 		break;
515 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
516 		wc->status = IB_WC_LOC_PROT_ERR;
517 		break;
518 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
519 		wc->status = IB_WC_WR_FLUSH_ERR;
520 		break;
521 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
522 		wc->status = IB_WC_MW_BIND_ERR;
523 		break;
524 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
525 		wc->status = IB_WC_BAD_RESP_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
528 		wc->status = IB_WC_LOC_ACCESS_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
531 		wc->status = IB_WC_REM_INV_REQ_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
534 		wc->status = IB_WC_REM_ACCESS_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
537 		wc->status = IB_WC_REM_OP_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
540 		wc->status = IB_WC_RETRY_EXC_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
543 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
546 		wc->status = IB_WC_REM_ABORT_ERR;
547 		break;
548 	default:
549 		wc->status = IB_WC_GENERAL_ERR;
550 		break;
551 	}
552 
553 	wc->vendor_err = cqe->vendor_err_syndrome;
554 }
555 
556 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
557 {
558 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
559 				      MLX4_CQE_STATUS_IPV4F     |
560 				      MLX4_CQE_STATUS_IPV4OPT   |
561 				      MLX4_CQE_STATUS_IPV6      |
562 				      MLX4_CQE_STATUS_IPOK)) ==
563 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
564 			    MLX4_CQE_STATUS_IPOK))              &&
565 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
566 				      MLX4_CQE_STATUS_TCP))     &&
567 		checksum == cpu_to_be16(0xffff);
568 }
569 
570 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
571 			   unsigned tail, struct mlx4_cqe *cqe, int is_eth)
572 {
573 	struct mlx4_ib_proxy_sqp_hdr *hdr;
574 
575 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
576 				   qp->sqp_proxy_rcv[tail].map,
577 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
578 				   DMA_FROM_DEVICE);
579 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
580 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
581 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
582 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
583 	wc->dlid_path_bits = 0;
584 
585 	if (is_eth) {
586 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
587 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
588 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
589 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
590 	} else {
591 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
592 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
593 	}
594 
595 	return 0;
596 }
597 
598 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
599 			       struct ib_wc *wc, int *npolled, int is_send)
600 {
601 	struct mlx4_ib_wq *wq;
602 	unsigned cur;
603 	int i;
604 
605 	wq = is_send ? &qp->sq : &qp->rq;
606 	cur = wq->head - wq->tail;
607 
608 	if (cur == 0)
609 		return;
610 
611 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
612 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
613 		wc->status = IB_WC_WR_FLUSH_ERR;
614 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
615 		wq->tail++;
616 		(*npolled)++;
617 		wc->qp = &qp->ibqp;
618 		wc++;
619 	}
620 }
621 
622 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
623 				 struct ib_wc *wc, int *npolled)
624 {
625 	struct mlx4_ib_qp *qp;
626 
627 	*npolled = 0;
628 	/* Find uncompleted WQEs belonging to that cq and retrun
629 	 * simulated FLUSH_ERR completions
630 	 */
631 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
632 		mlx4_ib_qp_sw_comp(qp, num_entries, wc, npolled, 1);
633 		if (*npolled >= num_entries)
634 			goto out;
635 	}
636 
637 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
638 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
639 		if (*npolled >= num_entries)
640 			goto out;
641 	}
642 
643 out:
644 	return;
645 }
646 
647 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
648 			    struct mlx4_ib_qp **cur_qp,
649 			    struct ib_wc *wc)
650 {
651 	struct mlx4_cqe *cqe;
652 	struct mlx4_qp *mqp;
653 	struct mlx4_ib_wq *wq;
654 	struct mlx4_ib_srq *srq;
655 	struct mlx4_srq *msrq = NULL;
656 	int is_send;
657 	int is_error;
658 	int is_eth;
659 	u32 g_mlpath_rqpn;
660 	u16 wqe_ctr;
661 	unsigned tail = 0;
662 
663 repoll:
664 	cqe = next_cqe_sw(cq);
665 	if (!cqe)
666 		return -EAGAIN;
667 
668 	if (cq->buf.entry_size == 64)
669 		cqe++;
670 
671 	++cq->mcq.cons_index;
672 
673 	/*
674 	 * Make sure we read CQ entry contents after we've checked the
675 	 * ownership bit.
676 	 */
677 	rmb();
678 
679 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
680 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
681 		MLX4_CQE_OPCODE_ERROR;
682 
683 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
684 		     is_send)) {
685 		pr_warn("Completion for NOP opcode detected!\n");
686 		return -EINVAL;
687 	}
688 
689 	/* Resize CQ in progress */
690 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
691 		if (cq->resize_buf) {
692 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
693 
694 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
695 			cq->buf      = cq->resize_buf->buf;
696 			cq->ibcq.cqe = cq->resize_buf->cqe;
697 
698 			kfree(cq->resize_buf);
699 			cq->resize_buf = NULL;
700 		}
701 
702 		goto repoll;
703 	}
704 
705 	if (!*cur_qp ||
706 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
707 		/*
708 		 * We do not have to take the QP table lock here,
709 		 * because CQs will be locked while QPs are removed
710 		 * from the table.
711 		 */
712 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
713 				       be32_to_cpu(cqe->vlan_my_qpn));
714 		if (unlikely(!mqp)) {
715 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
716 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
717 			return -EINVAL;
718 		}
719 
720 		*cur_qp = to_mibqp(mqp);
721 	}
722 
723 	wc->qp = &(*cur_qp)->ibqp;
724 
725 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
726 		u32 srq_num;
727 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
728 		srq_num       = g_mlpath_rqpn & 0xffffff;
729 		/* SRQ is also in the radix tree */
730 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
731 				       srq_num);
732 		if (unlikely(!msrq)) {
733 			pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
734 				cq->mcq.cqn, srq_num);
735 			return -EINVAL;
736 		}
737 	}
738 
739 	if (is_send) {
740 		wq = &(*cur_qp)->sq;
741 		if (!(*cur_qp)->sq_signal_bits) {
742 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
743 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
744 		}
745 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
746 		++wq->tail;
747 	} else if ((*cur_qp)->ibqp.srq) {
748 		srq = to_msrq((*cur_qp)->ibqp.srq);
749 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
750 		wc->wr_id = srq->wrid[wqe_ctr];
751 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
752 	} else if (msrq) {
753 		srq = to_mibsrq(msrq);
754 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
755 		wc->wr_id = srq->wrid[wqe_ctr];
756 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
757 	} else {
758 		wq	  = &(*cur_qp)->rq;
759 		tail	  = wq->tail & (wq->wqe_cnt - 1);
760 		wc->wr_id = wq->wrid[tail];
761 		++wq->tail;
762 	}
763 
764 	if (unlikely(is_error)) {
765 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
766 		return 0;
767 	}
768 
769 	wc->status = IB_WC_SUCCESS;
770 
771 	if (is_send) {
772 		wc->wc_flags = 0;
773 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
774 		case MLX4_OPCODE_RDMA_WRITE_IMM:
775 			wc->wc_flags |= IB_WC_WITH_IMM;
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 		case MLX4_OPCODE_SEND:
782 		case MLX4_OPCODE_SEND_INVAL:
783 			wc->opcode    = IB_WC_SEND;
784 			break;
785 		case MLX4_OPCODE_RDMA_READ:
786 			wc->opcode    = IB_WC_RDMA_READ;
787 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
788 			break;
789 		case MLX4_OPCODE_ATOMIC_CS:
790 			wc->opcode    = IB_WC_COMP_SWAP;
791 			wc->byte_len  = 8;
792 			break;
793 		case MLX4_OPCODE_ATOMIC_FA:
794 			wc->opcode    = IB_WC_FETCH_ADD;
795 			wc->byte_len  = 8;
796 			break;
797 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
798 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
799 			wc->byte_len  = 8;
800 			break;
801 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
802 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
803 			wc->byte_len  = 8;
804 			break;
805 		case MLX4_OPCODE_BIND_MW:
806 			wc->opcode    = IB_WC_BIND_MW;
807 			break;
808 		case MLX4_OPCODE_LSO:
809 			wc->opcode    = IB_WC_LSO;
810 			break;
811 		case MLX4_OPCODE_FMR:
812 			wc->opcode    = IB_WC_FAST_REG_MR;
813 			break;
814 		case MLX4_OPCODE_LOCAL_INVAL:
815 			wc->opcode    = IB_WC_LOCAL_INV;
816 			break;
817 		}
818 	} else {
819 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
820 
821 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
822 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
823 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
824 			wc->wc_flags	= IB_WC_WITH_IMM;
825 			wc->ex.imm_data = cqe->immed_rss_invalid;
826 			break;
827 		case MLX4_RECV_OPCODE_SEND_INVAL:
828 			wc->opcode	= IB_WC_RECV;
829 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
830 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
831 			break;
832 		case MLX4_RECV_OPCODE_SEND:
833 			wc->opcode   = IB_WC_RECV;
834 			wc->wc_flags = 0;
835 			break;
836 		case MLX4_RECV_OPCODE_SEND_IMM:
837 			wc->opcode	= IB_WC_RECV;
838 			wc->wc_flags	= IB_WC_WITH_IMM;
839 			wc->ex.imm_data = cqe->immed_rss_invalid;
840 			break;
841 		}
842 
843 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
844 						  (*cur_qp)->port) ==
845 			  IB_LINK_LAYER_ETHERNET);
846 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
847 			if ((*cur_qp)->mlx4_ib_qp_type &
848 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
849 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
850 				return use_tunnel_data(*cur_qp, cq, wc, tail,
851 						       cqe, is_eth);
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_VLAN_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 	int err = 0;
889 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
890 
891 	spin_lock_irqsave(&cq->lock, flags);
892 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
893 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
894 		goto out;
895 	}
896 
897 	for (npolled = 0; npolled < num_entries; ++npolled) {
898 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
899 		if (err)
900 			break;
901 	}
902 
903 	mlx4_cq_set_ci(&cq->mcq);
904 
905 out:
906 	spin_unlock_irqrestore(&cq->lock, flags);
907 
908 	if (err == 0 || err == -EAGAIN)
909 		return npolled;
910 	else
911 		return err;
912 }
913 
914 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
915 {
916 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
917 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
918 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
919 		    to_mdev(ibcq->device)->uar_map,
920 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
921 
922 	return 0;
923 }
924 
925 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
926 {
927 	u32 prod_index;
928 	int nfreed = 0;
929 	struct mlx4_cqe *cqe, *dest;
930 	u8 owner_bit;
931 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
932 
933 	/*
934 	 * First we need to find the current producer index, so we
935 	 * know where to start cleaning from.  It doesn't matter if HW
936 	 * adds new entries after this loop -- the QP we're worried
937 	 * about is already in RESET, so the new entries won't come
938 	 * from our QP and therefore don't need to be checked.
939 	 */
940 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
941 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
942 			break;
943 
944 	/*
945 	 * Now sweep backwards through the CQ, removing CQ entries
946 	 * that match our QP by copying older entries on top of them.
947 	 */
948 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
949 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
950 		cqe += cqe_inc;
951 
952 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
953 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
954 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
955 			++nfreed;
956 		} else if (nfreed) {
957 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
958 			dest += cqe_inc;
959 
960 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
961 			memcpy(dest, cqe, sizeof *cqe);
962 			dest->owner_sr_opcode = owner_bit |
963 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
964 		}
965 	}
966 
967 	if (nfreed) {
968 		cq->mcq.cons_index += nfreed;
969 		/*
970 		 * Make sure update of buffer contents is done before
971 		 * updating consumer index.
972 		 */
973 		wmb();
974 		mlx4_cq_set_ci(&cq->mcq);
975 	}
976 }
977 
978 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
979 {
980 	spin_lock_irq(&cq->lock);
981 	__mlx4_ib_cq_clean(cq, qpn, srq);
982 	spin_unlock_irq(&cq->lock);
983 }
984