xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cq.c (revision 930beb5a)
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);
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 
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 
192 	if (context) {
193 		struct mlx4_ib_create_cq ucmd;
194 
195 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
196 			err = -EFAULT;
197 			goto err_cq;
198 		}
199 
200 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
201 					  ucmd.buf_addr, entries);
202 		if (err)
203 			goto err_cq;
204 
205 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
206 					  &cq->db);
207 		if (err)
208 			goto err_mtt;
209 
210 		uar = &to_mucontext(context)->uar;
211 	} else {
212 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
213 		if (err)
214 			goto err_cq;
215 
216 		cq->mcq.set_ci_db  = cq->db.db;
217 		cq->mcq.arm_db     = cq->db.db + 1;
218 		*cq->mcq.set_ci_db = 0;
219 		*cq->mcq.arm_db    = 0;
220 
221 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
222 		if (err)
223 			goto err_db;
224 
225 		uar = &dev->priv_uar;
226 	}
227 
228 	if (dev->eq_table)
229 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
230 
231 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
232 			    cq->db.dma, &cq->mcq, vector, 0, 0);
233 	if (err)
234 		goto err_dbmap;
235 
236 	cq->mcq.comp  = mlx4_ib_cq_comp;
237 	cq->mcq.event = mlx4_ib_cq_event;
238 
239 	if (context)
240 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
241 			err = -EFAULT;
242 			goto err_dbmap;
243 		}
244 
245 	return &cq->ibcq;
246 
247 err_dbmap:
248 	if (context)
249 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
250 
251 err_mtt:
252 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
253 
254 	if (context)
255 		ib_umem_release(cq->umem);
256 	else
257 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
258 
259 err_db:
260 	if (!context)
261 		mlx4_db_free(dev->dev, &cq->db);
262 
263 err_cq:
264 	kfree(cq);
265 
266 	return ERR_PTR(err);
267 }
268 
269 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
270 				  int entries)
271 {
272 	int err;
273 
274 	if (cq->resize_buf)
275 		return -EBUSY;
276 
277 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
278 	if (!cq->resize_buf)
279 		return -ENOMEM;
280 
281 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
282 	if (err) {
283 		kfree(cq->resize_buf);
284 		cq->resize_buf = NULL;
285 		return err;
286 	}
287 
288 	cq->resize_buf->cqe = entries - 1;
289 
290 	return 0;
291 }
292 
293 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
294 				   int entries, struct ib_udata *udata)
295 {
296 	struct mlx4_ib_resize_cq ucmd;
297 	int err;
298 
299 	if (cq->resize_umem)
300 		return -EBUSY;
301 
302 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
303 		return -EFAULT;
304 
305 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
306 	if (!cq->resize_buf)
307 		return -ENOMEM;
308 
309 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
310 				  &cq->resize_umem, ucmd.buf_addr, entries);
311 	if (err) {
312 		kfree(cq->resize_buf);
313 		cq->resize_buf = NULL;
314 		return err;
315 	}
316 
317 	cq->resize_buf->cqe = entries - 1;
318 
319 	return 0;
320 }
321 
322 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
323 {
324 	u32 i;
325 
326 	i = cq->mcq.cons_index;
327 	while (get_sw_cqe(cq, i))
328 		++i;
329 
330 	return i - cq->mcq.cons_index;
331 }
332 
333 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
334 {
335 	struct mlx4_cqe *cqe, *new_cqe;
336 	int i;
337 	int cqe_size = cq->buf.entry_size;
338 	int cqe_inc = cqe_size == 64 ? 1 : 0;
339 
340 	i = cq->mcq.cons_index;
341 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
342 	cqe += cqe_inc;
343 
344 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
345 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
346 					   (i + 1) & cq->resize_buf->cqe);
347 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
348 		new_cqe += cqe_inc;
349 
350 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
351 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
352 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
353 		cqe += cqe_inc;
354 	}
355 	++cq->mcq.cons_index;
356 }
357 
358 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
359 {
360 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
361 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
362 	struct mlx4_mtt mtt;
363 	int outst_cqe;
364 	int err;
365 
366 	mutex_lock(&cq->resize_mutex);
367 
368 	if (entries < 1) {
369 		err = -EINVAL;
370 		goto out;
371 	}
372 
373 	entries = roundup_pow_of_two(entries + 1);
374 	if (entries == ibcq->cqe + 1) {
375 		err = 0;
376 		goto out;
377 	}
378 
379 	if (entries > dev->dev->caps.max_cqes) {
380 		err = -EINVAL;
381 		goto out;
382 	}
383 
384 	if (ibcq->uobject) {
385 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
386 		if (err)
387 			goto out;
388 	} else {
389 		/* Can't be smaller than the number of outstanding CQEs */
390 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
391 		if (entries < outst_cqe + 1) {
392 			err = 0;
393 			goto out;
394 		}
395 
396 		err = mlx4_alloc_resize_buf(dev, cq, entries);
397 		if (err)
398 			goto out;
399 	}
400 
401 	mtt = cq->buf.mtt;
402 
403 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
404 	if (err)
405 		goto err_buf;
406 
407 	mlx4_mtt_cleanup(dev->dev, &mtt);
408 	if (ibcq->uobject) {
409 		cq->buf      = cq->resize_buf->buf;
410 		cq->ibcq.cqe = cq->resize_buf->cqe;
411 		ib_umem_release(cq->umem);
412 		cq->umem     = cq->resize_umem;
413 
414 		kfree(cq->resize_buf);
415 		cq->resize_buf = NULL;
416 		cq->resize_umem = NULL;
417 	} else {
418 		struct mlx4_ib_cq_buf tmp_buf;
419 		int tmp_cqe = 0;
420 
421 		spin_lock_irq(&cq->lock);
422 		if (cq->resize_buf) {
423 			mlx4_ib_cq_resize_copy_cqes(cq);
424 			tmp_buf = cq->buf;
425 			tmp_cqe = cq->ibcq.cqe;
426 			cq->buf      = cq->resize_buf->buf;
427 			cq->ibcq.cqe = cq->resize_buf->cqe;
428 
429 			kfree(cq->resize_buf);
430 			cq->resize_buf = NULL;
431 		}
432 		spin_unlock_irq(&cq->lock);
433 
434 		if (tmp_cqe)
435 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
436 	}
437 
438 	goto out;
439 
440 err_buf:
441 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
442 	if (!ibcq->uobject)
443 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
444 				    cq->resize_buf->cqe);
445 
446 	kfree(cq->resize_buf);
447 	cq->resize_buf = NULL;
448 
449 	if (cq->resize_umem) {
450 		ib_umem_release(cq->resize_umem);
451 		cq->resize_umem = NULL;
452 	}
453 
454 out:
455 	mutex_unlock(&cq->resize_mutex);
456 
457 	return err;
458 }
459 
460 int mlx4_ib_destroy_cq(struct ib_cq *cq)
461 {
462 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
463 	struct mlx4_ib_cq *mcq = to_mcq(cq);
464 
465 	mlx4_cq_free(dev->dev, &mcq->mcq);
466 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
467 
468 	if (cq->uobject) {
469 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
470 		ib_umem_release(mcq->umem);
471 	} else {
472 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
473 		mlx4_db_free(dev->dev, &mcq->db);
474 	}
475 
476 	kfree(mcq);
477 
478 	return 0;
479 }
480 
481 static void dump_cqe(void *cqe)
482 {
483 	__be32 *buf = cqe;
484 
485 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
486 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
487 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
488 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
489 }
490 
491 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
492 				     struct ib_wc *wc)
493 {
494 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
495 		pr_debug("local QP operation err "
496 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
497 		       "opcode = %02x)\n",
498 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
499 		       cqe->vendor_err_syndrome,
500 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
501 		dump_cqe(cqe);
502 	}
503 
504 	switch (cqe->syndrome) {
505 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
506 		wc->status = IB_WC_LOC_LEN_ERR;
507 		break;
508 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
509 		wc->status = IB_WC_LOC_QP_OP_ERR;
510 		break;
511 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
512 		wc->status = IB_WC_LOC_PROT_ERR;
513 		break;
514 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
515 		wc->status = IB_WC_WR_FLUSH_ERR;
516 		break;
517 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
518 		wc->status = IB_WC_MW_BIND_ERR;
519 		break;
520 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
521 		wc->status = IB_WC_BAD_RESP_ERR;
522 		break;
523 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
524 		wc->status = IB_WC_LOC_ACCESS_ERR;
525 		break;
526 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
527 		wc->status = IB_WC_REM_INV_REQ_ERR;
528 		break;
529 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
530 		wc->status = IB_WC_REM_ACCESS_ERR;
531 		break;
532 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
533 		wc->status = IB_WC_REM_OP_ERR;
534 		break;
535 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
536 		wc->status = IB_WC_RETRY_EXC_ERR;
537 		break;
538 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
539 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
540 		break;
541 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
542 		wc->status = IB_WC_REM_ABORT_ERR;
543 		break;
544 	default:
545 		wc->status = IB_WC_GENERAL_ERR;
546 		break;
547 	}
548 
549 	wc->vendor_err = cqe->vendor_err_syndrome;
550 }
551 
552 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
553 {
554 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
555 				      MLX4_CQE_STATUS_IPV4F     |
556 				      MLX4_CQE_STATUS_IPV4OPT   |
557 				      MLX4_CQE_STATUS_IPV6      |
558 				      MLX4_CQE_STATUS_IPOK)) ==
559 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
560 			    MLX4_CQE_STATUS_IPOK))              &&
561 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
562 				      MLX4_CQE_STATUS_TCP))     &&
563 		checksum == cpu_to_be16(0xffff);
564 }
565 
566 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
567 			   unsigned tail, struct mlx4_cqe *cqe)
568 {
569 	struct mlx4_ib_proxy_sqp_hdr *hdr;
570 
571 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
572 				   qp->sqp_proxy_rcv[tail].map,
573 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
574 				   DMA_FROM_DEVICE);
575 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
576 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
577 	wc->slid	= be16_to_cpu(hdr->tun.slid_mac_47_32);
578 	wc->sl		= (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
579 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
580 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
581 	wc->dlid_path_bits = 0;
582 
583 	return 0;
584 }
585 
586 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
587 			    struct mlx4_ib_qp **cur_qp,
588 			    struct ib_wc *wc)
589 {
590 	struct mlx4_cqe *cqe;
591 	struct mlx4_qp *mqp;
592 	struct mlx4_ib_wq *wq;
593 	struct mlx4_ib_srq *srq;
594 	struct mlx4_srq *msrq = NULL;
595 	int is_send;
596 	int is_error;
597 	u32 g_mlpath_rqpn;
598 	u16 wqe_ctr;
599 	unsigned tail = 0;
600 
601 repoll:
602 	cqe = next_cqe_sw(cq);
603 	if (!cqe)
604 		return -EAGAIN;
605 
606 	if (cq->buf.entry_size == 64)
607 		cqe++;
608 
609 	++cq->mcq.cons_index;
610 
611 	/*
612 	 * Make sure we read CQ entry contents after we've checked the
613 	 * ownership bit.
614 	 */
615 	rmb();
616 
617 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
618 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
619 		MLX4_CQE_OPCODE_ERROR;
620 
621 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
622 		     is_send)) {
623 		pr_warn("Completion for NOP opcode detected!\n");
624 		return -EINVAL;
625 	}
626 
627 	/* Resize CQ in progress */
628 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
629 		if (cq->resize_buf) {
630 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
631 
632 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
633 			cq->buf      = cq->resize_buf->buf;
634 			cq->ibcq.cqe = cq->resize_buf->cqe;
635 
636 			kfree(cq->resize_buf);
637 			cq->resize_buf = NULL;
638 		}
639 
640 		goto repoll;
641 	}
642 
643 	if (!*cur_qp ||
644 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
645 		/*
646 		 * We do not have to take the QP table lock here,
647 		 * because CQs will be locked while QPs are removed
648 		 * from the table.
649 		 */
650 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
651 				       be32_to_cpu(cqe->vlan_my_qpn));
652 		if (unlikely(!mqp)) {
653 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
654 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
655 			return -EINVAL;
656 		}
657 
658 		*cur_qp = to_mibqp(mqp);
659 	}
660 
661 	wc->qp = &(*cur_qp)->ibqp;
662 
663 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
664 		u32 srq_num;
665 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
666 		srq_num       = g_mlpath_rqpn & 0xffffff;
667 		/* SRQ is also in the radix tree */
668 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
669 				       srq_num);
670 		if (unlikely(!msrq)) {
671 			pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
672 				cq->mcq.cqn, srq_num);
673 			return -EINVAL;
674 		}
675 	}
676 
677 	if (is_send) {
678 		wq = &(*cur_qp)->sq;
679 		if (!(*cur_qp)->sq_signal_bits) {
680 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
681 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
682 		}
683 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
684 		++wq->tail;
685 	} else if ((*cur_qp)->ibqp.srq) {
686 		srq = to_msrq((*cur_qp)->ibqp.srq);
687 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
688 		wc->wr_id = srq->wrid[wqe_ctr];
689 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
690 	} else if (msrq) {
691 		srq = to_mibsrq(msrq);
692 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
693 		wc->wr_id = srq->wrid[wqe_ctr];
694 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
695 	} else {
696 		wq	  = &(*cur_qp)->rq;
697 		tail	  = wq->tail & (wq->wqe_cnt - 1);
698 		wc->wr_id = wq->wrid[tail];
699 		++wq->tail;
700 	}
701 
702 	if (unlikely(is_error)) {
703 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
704 		return 0;
705 	}
706 
707 	wc->status = IB_WC_SUCCESS;
708 
709 	if (is_send) {
710 		wc->wc_flags = 0;
711 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
712 		case MLX4_OPCODE_RDMA_WRITE_IMM:
713 			wc->wc_flags |= IB_WC_WITH_IMM;
714 		case MLX4_OPCODE_RDMA_WRITE:
715 			wc->opcode    = IB_WC_RDMA_WRITE;
716 			break;
717 		case MLX4_OPCODE_SEND_IMM:
718 			wc->wc_flags |= IB_WC_WITH_IMM;
719 		case MLX4_OPCODE_SEND:
720 		case MLX4_OPCODE_SEND_INVAL:
721 			wc->opcode    = IB_WC_SEND;
722 			break;
723 		case MLX4_OPCODE_RDMA_READ:
724 			wc->opcode    = IB_WC_RDMA_READ;
725 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
726 			break;
727 		case MLX4_OPCODE_ATOMIC_CS:
728 			wc->opcode    = IB_WC_COMP_SWAP;
729 			wc->byte_len  = 8;
730 			break;
731 		case MLX4_OPCODE_ATOMIC_FA:
732 			wc->opcode    = IB_WC_FETCH_ADD;
733 			wc->byte_len  = 8;
734 			break;
735 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
736 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
737 			wc->byte_len  = 8;
738 			break;
739 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
740 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
741 			wc->byte_len  = 8;
742 			break;
743 		case MLX4_OPCODE_BIND_MW:
744 			wc->opcode    = IB_WC_BIND_MW;
745 			break;
746 		case MLX4_OPCODE_LSO:
747 			wc->opcode    = IB_WC_LSO;
748 			break;
749 		case MLX4_OPCODE_FMR:
750 			wc->opcode    = IB_WC_FAST_REG_MR;
751 			break;
752 		case MLX4_OPCODE_LOCAL_INVAL:
753 			wc->opcode    = IB_WC_LOCAL_INV;
754 			break;
755 		}
756 	} else {
757 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
758 
759 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
760 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
761 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
762 			wc->wc_flags	= IB_WC_WITH_IMM;
763 			wc->ex.imm_data = cqe->immed_rss_invalid;
764 			break;
765 		case MLX4_RECV_OPCODE_SEND_INVAL:
766 			wc->opcode	= IB_WC_RECV;
767 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
768 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
769 			break;
770 		case MLX4_RECV_OPCODE_SEND:
771 			wc->opcode   = IB_WC_RECV;
772 			wc->wc_flags = 0;
773 			break;
774 		case MLX4_RECV_OPCODE_SEND_IMM:
775 			wc->opcode	= IB_WC_RECV;
776 			wc->wc_flags	= IB_WC_WITH_IMM;
777 			wc->ex.imm_data = cqe->immed_rss_invalid;
778 			break;
779 		}
780 
781 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
782 			if ((*cur_qp)->mlx4_ib_qp_type &
783 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
784 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
785 				return use_tunnel_data(*cur_qp, cq, wc, tail, cqe);
786 		}
787 
788 		wc->slid	   = be16_to_cpu(cqe->rlid);
789 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
790 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
791 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
792 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
793 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
794 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
795 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
796 		if (rdma_port_get_link_layer(wc->qp->device,
797 				(*cur_qp)->port) == IB_LINK_LAYER_ETHERNET)
798 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
799 		else
800 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
801 	}
802 
803 	return 0;
804 }
805 
806 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
807 {
808 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
809 	struct mlx4_ib_qp *cur_qp = NULL;
810 	unsigned long flags;
811 	int npolled;
812 	int err = 0;
813 
814 	spin_lock_irqsave(&cq->lock, flags);
815 
816 	for (npolled = 0; npolled < num_entries; ++npolled) {
817 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
818 		if (err)
819 			break;
820 	}
821 
822 	mlx4_cq_set_ci(&cq->mcq);
823 
824 	spin_unlock_irqrestore(&cq->lock, flags);
825 
826 	if (err == 0 || err == -EAGAIN)
827 		return npolled;
828 	else
829 		return err;
830 }
831 
832 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
833 {
834 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
835 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
836 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
837 		    to_mdev(ibcq->device)->uar_map,
838 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
839 
840 	return 0;
841 }
842 
843 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
844 {
845 	u32 prod_index;
846 	int nfreed = 0;
847 	struct mlx4_cqe *cqe, *dest;
848 	u8 owner_bit;
849 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
850 
851 	/*
852 	 * First we need to find the current producer index, so we
853 	 * know where to start cleaning from.  It doesn't matter if HW
854 	 * adds new entries after this loop -- the QP we're worried
855 	 * about is already in RESET, so the new entries won't come
856 	 * from our QP and therefore don't need to be checked.
857 	 */
858 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
859 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
860 			break;
861 
862 	/*
863 	 * Now sweep backwards through the CQ, removing CQ entries
864 	 * that match our QP by copying older entries on top of them.
865 	 */
866 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
867 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
868 		cqe += cqe_inc;
869 
870 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
871 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
872 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
873 			++nfreed;
874 		} else if (nfreed) {
875 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
876 			dest += cqe_inc;
877 
878 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
879 			memcpy(dest, cqe, sizeof *cqe);
880 			dest->owner_sr_opcode = owner_bit |
881 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
882 		}
883 	}
884 
885 	if (nfreed) {
886 		cq->mcq.cons_index += nfreed;
887 		/*
888 		 * Make sure update of buffer contents is done before
889 		 * updating consumer index.
890 		 */
891 		wmb();
892 		mlx4_cq_set_ci(&cq->mcq);
893 	}
894 }
895 
896 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
897 {
898 	spin_lock_irq(&cq->lock);
899 	__mlx4_ib_cq_clean(cq, qpn, srq);
900 	spin_unlock_irq(&cq->lock);
901 }
902