xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_cq.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
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 <dev/mlx4/cq.h>
35 #include <dev/mlx4/qp.h>
36 #include <dev/mlx4/srq.h>
37 #include <linux/slab.h>
38 
39 #include "mlx4_ib.h"
40 #include "user.h"
41 
42 /* Which firmware version adds support for Resize CQ */
43 #define MLX4_FW_VER_RESIZE_CQ  mlx4_fw_ver(2, 5, 0)
44 #define MLX4_FW_VER_IGNORE_OVERRUN_CQ mlx4_fw_ver(2, 7, 8200)
45 
46 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
47 {
48 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
49 	ibcq->comp_handler(ibcq, ibcq->cq_context);
50 }
51 
52 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
53 {
54 	struct ib_event event;
55 	struct ib_cq *ibcq;
56 
57 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
58 		pr_warn("Unexpected event type %d "
59 		       "on CQ %06x\n", type, cq->cqn);
60 		return;
61 	}
62 
63 	ibcq = &to_mibcq(cq)->ibcq;
64 	if (ibcq->event_handler) {
65 		event.device     = ibcq->device;
66 		event.event      = IB_EVENT_CQ_ERR;
67 		event.element.cq = ibcq;
68 		ibcq->event_handler(&event, ibcq->cq_context);
69 	}
70 }
71 
72 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
73 {
74 	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
75 }
76 
77 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
78 {
79 	return get_cqe_from_buf(&cq->buf, n);
80 }
81 
82 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
83 {
84 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
85 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
86 
87 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
88 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
89 }
90 
91 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
92 {
93 	return get_sw_cqe(cq, cq->mcq.cons_index);
94 }
95 
96 int mlx4_ib_modify_cq(struct ib_cq *cq,
97 		      struct ib_cq_attr *cq_attr,
98 		      int cq_attr_mask)
99 {
100 	int err = 0;
101 	struct mlx4_ib_cq *mcq = to_mcq(cq);
102 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
103 
104 	if (cq_attr_mask & IB_CQ_CAP_FLAGS) {
105 		if (cq_attr->cq_cap_flags & IB_CQ_TIMESTAMP)
106 			return -ENOTSUPP;
107 
108 		if (cq_attr->cq_cap_flags & IB_CQ_IGNORE_OVERRUN) {
109 			if (dev->dev->caps.cq_flags & MLX4_DEV_CAP_CQ_FLAG_IO)
110 				err = mlx4_cq_ignore_overrun(dev->dev, &mcq->mcq);
111 			else
112 				err = -ENOSYS;
113 		}
114 	}
115 
116 	if (!err)
117 		if (cq_attr_mask & IB_CQ_MODERATION)
118 			err = mlx4_cq_modify(dev->dev, &mcq->mcq,
119 					cq_attr->moderation.cq_count,
120 					cq_attr->moderation.cq_period);
121 
122 	return err;
123 }
124 
125 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
126 {
127 	int err;
128 
129 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
130 			     PAGE_SIZE * 2, &buf->buf);
131 
132 	if (err)
133 		goto out;
134 
135 	buf->entry_size = dev->dev->caps.cqe_size;
136 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
137 				    &buf->mtt);
138 	if (err)
139 		goto err_buf;
140 
141 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
142 	if (err)
143 		goto err_mtt;
144 
145 	return 0;
146 
147 err_mtt:
148 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
149 
150 err_buf:
151 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
152 
153 out:
154 	return err;
155 }
156 
157 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
158 {
159 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
160 }
161 
162 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
163 			       struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
164 			       u64 buf_addr, int cqe)
165 {
166 	int err;
167 	int cqe_size = dev->dev->caps.cqe_size;
168 	int shift;
169 	int n;
170 
171 	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
172 			    IB_ACCESS_LOCAL_WRITE, 1);
173 	if (IS_ERR(*umem))
174 		return PTR_ERR(*umem);
175 
176 	n = ib_umem_page_count(*umem);
177 	shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n);
178 	err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt);
179 
180 	if (err)
181 		goto err_buf;
182 
183 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
184 	if (err)
185 		goto err_mtt;
186 
187 	return 0;
188 
189 err_mtt:
190 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
191 
192 err_buf:
193 	ib_umem_release(*umem);
194 
195 	return err;
196 }
197 
198 /* we don't support system timestamping */
199 #define CQ_CREATE_FLAGS_SUPPORTED IB_CQ_TIMESTAMP
200 
201 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev,
202 				struct ib_cq_init_attr *attr,
203 				struct ib_ucontext *context,
204 				struct ib_udata *udata)
205 {
206 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
207 	struct mlx4_ib_cq *cq;
208 	struct mlx4_uar *uar;
209 	int err;
210 	int entries = attr->cqe;
211 	int vector = attr->comp_vector;
212 
213 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
214 		return ERR_PTR(-EINVAL);
215 
216 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
217 		return ERR_PTR(-EINVAL);
218 
219 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
220 	if (!cq)
221 		return ERR_PTR(-ENOMEM);
222 
223 	entries      = roundup_pow_of_two(entries + 1);
224 	cq->ibcq.cqe = entries - 1;
225 	mutex_init(&cq->resize_mutex);
226 	spin_lock_init(&cq->lock);
227 	cq->resize_buf = NULL;
228 	cq->resize_umem = NULL;
229 	cq->create_flags = attr->flags;
230 
231 	if (context) {
232 		struct mlx4_ib_create_cq ucmd;
233 
234 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
235 			err = -EFAULT;
236 			goto err_cq;
237 		}
238 
239 		err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
240 					  ucmd.buf_addr, entries);
241 		if (err)
242 			goto err_cq;
243 
244 		err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
245 					  &cq->db);
246 		if (err)
247 			goto err_mtt;
248 
249 		uar = &to_mucontext(context)->uar;
250 	} else {
251 		err = mlx4_db_alloc(dev->dev, &cq->db, 1);
252 		if (err)
253 			goto err_cq;
254 
255 		cq->mcq.set_ci_db  = cq->db.db;
256 		cq->mcq.arm_db     = cq->db.db + 1;
257 		*cq->mcq.set_ci_db = 0;
258 		*cq->mcq.arm_db    = 0;
259 
260 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
261 		if (err)
262 			goto err_db;
263 
264 		uar = &dev->priv_uar;
265 	}
266 
267 	if (dev->eq_table)
268 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
269 
270 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
271 			    cq->db.dma, &cq->mcq, vector, 0,
272 			    !!(cq->create_flags & IB_CQ_TIMESTAMP));
273 	if (err)
274 		goto err_dbmap;
275 
276 	cq->mcq.comp  = mlx4_ib_cq_comp;
277 	cq->mcq.event = mlx4_ib_cq_event;
278 
279 	if (context)
280 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
281 			err = -EFAULT;
282 			goto err_dbmap;
283 		}
284 
285 	return &cq->ibcq;
286 
287 err_dbmap:
288 	if (context)
289 		mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
290 
291 err_mtt:
292 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
293 
294 	if (context)
295 		ib_umem_release(cq->umem);
296 	else
297 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
298 
299 err_db:
300 	if (!context)
301 		mlx4_db_free(dev->dev, &cq->db);
302 
303 err_cq:
304 	kfree(cq);
305 
306 	return ERR_PTR(err);
307 }
308 
309 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
310 				  int entries)
311 {
312 	int err;
313 
314 	if (cq->resize_buf)
315 		return -EBUSY;
316 
317 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
318 	if (!cq->resize_buf)
319 		return -ENOMEM;
320 
321 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
322 	if (err) {
323 		kfree(cq->resize_buf);
324 		cq->resize_buf = NULL;
325 		return err;
326 	}
327 
328 	cq->resize_buf->cqe = entries - 1;
329 
330 	return 0;
331 }
332 
333 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
334 				   int entries, struct ib_udata *udata)
335 {
336 	struct mlx4_ib_resize_cq ucmd;
337 	int err;
338 
339 	if (cq->resize_umem)
340 		return -EBUSY;
341 
342 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
343 		return -EFAULT;
344 
345 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
346 	if (!cq->resize_buf)
347 		return -ENOMEM;
348 
349 	err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
350 				  &cq->resize_umem, ucmd.buf_addr, entries);
351 	if (err) {
352 		kfree(cq->resize_buf);
353 		cq->resize_buf = NULL;
354 		return err;
355 	}
356 
357 	cq->resize_buf->cqe = entries - 1;
358 
359 	return 0;
360 }
361 
362 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
363 {
364 	u32 i;
365 
366 	i = cq->mcq.cons_index;
367 	while (get_sw_cqe(cq, i))
368 		++i;
369 
370 	return i - cq->mcq.cons_index;
371 }
372 
373 static int mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
374 {
375 	struct mlx4_cqe *cqe, *new_cqe;
376 	int i;
377 	int cqe_size = cq->buf.entry_size;
378 	int cqe_inc = cqe_size == 64 ? 1 : 0;
379 	struct mlx4_cqe *start_cqe;
380 
381 	i = cq->mcq.cons_index;
382 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
383 	start_cqe = cqe;
384 	cqe += cqe_inc;
385 
386 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
387 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
388 					   (i + 1) & cq->resize_buf->cqe);
389 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
390 		new_cqe += cqe_inc;
391 
392 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
393 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
394 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
395 		if (cqe == start_cqe) {
396 			pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n", cq->mcq.cqn);
397 			return -ENOMEM;
398 		}
399 		cqe += cqe_inc;
400 
401 	}
402 	++cq->mcq.cons_index;
403 	return 0;
404 }
405 
406 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
407 {
408 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
409 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
410 	struct mlx4_mtt mtt;
411 	int outst_cqe;
412 	int err;
413 
414 	if (dev->dev->caps.fw_ver < MLX4_FW_VER_RESIZE_CQ)
415 		return -ENOSYS;
416 
417 	mutex_lock(&cq->resize_mutex);
418 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
419 		err = -EINVAL;
420 		goto out;
421 	}
422 
423 	entries = roundup_pow_of_two(entries + 1);
424 	if (entries == ibcq->cqe + 1) {
425 		err = 0;
426 		goto out;
427 	}
428 
429 	if (entries > dev->dev->caps.max_cqes + 1) {
430 		err = -EINVAL;
431 		goto out;
432 	}
433 
434 	if (ibcq->uobject) {
435 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
436 		if (err)
437 			goto out;
438 	} else {
439 		/* Can't be smaller than the number of outstanding CQEs */
440 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
441 		if (entries < outst_cqe + 1) {
442 			err = 0;
443 			goto out;
444 		}
445 
446 		err = mlx4_alloc_resize_buf(dev, cq, entries);
447 		if (err)
448 			goto out;
449 	}
450 
451 	mtt = cq->buf.mtt;
452 
453 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
454 	if (err)
455 		goto err_buf;
456 
457 	mlx4_mtt_cleanup(dev->dev, &mtt);
458 	if (ibcq->uobject) {
459 		cq->buf      = cq->resize_buf->buf;
460 		cq->ibcq.cqe = cq->resize_buf->cqe;
461 		ib_umem_release(cq->umem);
462 		cq->umem     = cq->resize_umem;
463 
464 		kfree(cq->resize_buf);
465 		cq->resize_buf = NULL;
466 		cq->resize_umem = NULL;
467 	} else {
468 		struct mlx4_ib_cq_buf tmp_buf;
469 		int tmp_cqe = 0;
470 
471 		spin_lock_irq(&cq->lock);
472 		if (cq->resize_buf) {
473 			err = mlx4_ib_cq_resize_copy_cqes(cq);
474 			tmp_buf = cq->buf;
475 			tmp_cqe = cq->ibcq.cqe;
476 			cq->buf      = cq->resize_buf->buf;
477 			cq->ibcq.cqe = cq->resize_buf->cqe;
478 
479 			kfree(cq->resize_buf);
480 			cq->resize_buf = NULL;
481 		}
482 		spin_unlock_irq(&cq->lock);
483 
484 		if (tmp_cqe)
485 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
486 	}
487 
488 	goto out;
489 
490 err_buf:
491 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
492 	if (!ibcq->uobject)
493 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
494 				    cq->resize_buf->cqe);
495 
496 	kfree(cq->resize_buf);
497 	cq->resize_buf = NULL;
498 
499 	if (cq->resize_umem) {
500 		ib_umem_release(cq->resize_umem);
501 		cq->resize_umem = NULL;
502 	}
503 
504 out:
505 	mutex_unlock(&cq->resize_mutex);
506 
507 	return err;
508 }
509 
510 int mlx4_ib_ignore_overrun_cq(struct ib_cq *ibcq)
511 {
512 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
513 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
514 
515 	if (dev->dev->caps.fw_ver < MLX4_FW_VER_IGNORE_OVERRUN_CQ)
516 		return -ENOSYS;
517 
518 	return mlx4_cq_ignore_overrun(dev->dev, &cq->mcq);
519 }
520 
521 int mlx4_ib_destroy_cq(struct ib_cq *cq)
522 {
523 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
524 	struct mlx4_ib_cq *mcq = to_mcq(cq);
525 
526 	mlx4_cq_free(dev->dev, &mcq->mcq);
527 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
528 
529 	if (cq->uobject) {
530 		mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
531 		ib_umem_release(mcq->umem);
532 	} else {
533 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
534 		mlx4_db_free(dev->dev, &mcq->db);
535 	}
536 
537 	kfree(mcq);
538 
539 	return 0;
540 }
541 
542 static void dump_cqe(void *cqe)
543 {
544 	__be32 *buf = cqe;
545 
546 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
547 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
548 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
549 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
550 }
551 
552 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
553 				     struct ib_wc *wc)
554 {
555 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
556 		pr_debug("local QP operation err "
557 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
558 		       "opcode = %02x)\n",
559 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
560 		       cqe->vendor_err_syndrome,
561 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
562 		dump_cqe(cqe);
563 	}
564 
565 	switch (cqe->syndrome) {
566 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
567 		wc->status = IB_WC_LOC_LEN_ERR;
568 		break;
569 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
570 		wc->status = IB_WC_LOC_QP_OP_ERR;
571 		break;
572 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
573 		wc->status = IB_WC_LOC_PROT_ERR;
574 		break;
575 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
576 		wc->status = IB_WC_WR_FLUSH_ERR;
577 		break;
578 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
579 		wc->status = IB_WC_MW_BIND_ERR;
580 		break;
581 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
582 		wc->status = IB_WC_BAD_RESP_ERR;
583 		break;
584 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
585 		wc->status = IB_WC_LOC_ACCESS_ERR;
586 		break;
587 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
588 		wc->status = IB_WC_REM_INV_REQ_ERR;
589 		break;
590 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
591 		wc->status = IB_WC_REM_ACCESS_ERR;
592 		break;
593 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
594 		wc->status = IB_WC_REM_OP_ERR;
595 		break;
596 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
597 		wc->status = IB_WC_RETRY_EXC_ERR;
598 		break;
599 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
600 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
601 		break;
602 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
603 		wc->status = IB_WC_REM_ABORT_ERR;
604 		break;
605 	default:
606 		wc->status = IB_WC_GENERAL_ERR;
607 		break;
608 	}
609 
610 	wc->vendor_err = cqe->vendor_err_syndrome;
611 }
612 
613 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
614 {
615 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
616 				      MLX4_CQE_STATUS_IPV4F     |
617 				      MLX4_CQE_STATUS_IPV4OPT   |
618 				      MLX4_CQE_STATUS_IPV6      |
619 				      MLX4_CQE_STATUS_IPOK)) ==
620 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
621 			    MLX4_CQE_STATUS_IPOK))              &&
622 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
623 				      MLX4_CQE_STATUS_TCP))     &&
624 		checksum == cpu_to_be16(0xffff);
625 }
626 
627 static int use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
628 			   unsigned tail, struct mlx4_cqe *cqe, int is_eth)
629 {
630 	struct mlx4_ib_proxy_sqp_hdr *hdr;
631 
632 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
633 				   qp->sqp_proxy_rcv[tail].map,
634 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
635 				   DMA_FROM_DEVICE);
636 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
637 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
638 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
639 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
640 	wc->dlid_path_bits = 0;
641 
642 	if (is_eth) {
643 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
644 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
645 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
646 	} else {
647 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
648 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
649 	}
650 
651 	return 0;
652 }
653 
654 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
655 			    struct mlx4_ib_qp **cur_qp,
656 			    struct ib_wc *wc)
657 {
658 	struct mlx4_cqe *cqe;
659 	struct mlx4_qp *mqp;
660 	struct mlx4_ib_wq *wq;
661 	struct mlx4_ib_srq *srq;
662 	struct mlx4_srq *msrq = NULL;
663 	int is_send;
664 	int is_error;
665 	u32 g_mlpath_rqpn;
666 	u16 wqe_ctr;
667 	unsigned tail = 0;
668 	int timestamp_en = !!(cq->create_flags & IB_CQ_TIMESTAMP);
669 
670 
671 repoll:
672 	cqe = next_cqe_sw(cq);
673 	if (!cqe)
674 		return -EAGAIN;
675 
676 	if (cq->buf.entry_size == 64)
677 		cqe++;
678 
679 	++cq->mcq.cons_index;
680 
681 	/*
682 	 * Make sure we read CQ entry contents after we've checked the
683 	 * ownership bit.
684 	 */
685 	rmb();
686 
687 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
688 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
689 		MLX4_CQE_OPCODE_ERROR;
690 
691 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
692 		     is_send)) {
693 		pr_warn("Completion for NOP opcode detected!\n");
694 		return -EINVAL;
695 	}
696 
697 	/* Resize CQ in progress */
698 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
699 		if (cq->resize_buf) {
700 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
701 
702 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
703 			cq->buf      = cq->resize_buf->buf;
704 			cq->ibcq.cqe = cq->resize_buf->cqe;
705 
706 			kfree(cq->resize_buf);
707 			cq->resize_buf = NULL;
708 		}
709 
710 		goto repoll;
711 	}
712 
713 	if (!*cur_qp ||
714 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
715 		/*
716 		 * We do not have to take the QP table lock here,
717 		 * because CQs will be locked while QPs are removed
718 		 * from the table.
719 		 */
720 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
721 				       be32_to_cpu(cqe->vlan_my_qpn));
722 		if (unlikely(!mqp)) {
723 			pr_warn("CQ %06x with entry for unknown QPN %06x\n",
724 			       cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
725 			return -EINVAL;
726 		}
727 
728 		*cur_qp = to_mibqp(mqp);
729 	}
730 
731 	wc->qp = &(*cur_qp)->ibqp;
732 
733 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
734 		u32 srq_num;
735 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
736 		srq_num       = g_mlpath_rqpn & 0xffffff;
737 		/* SRQ is also in the radix tree */
738 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
739 				       srq_num);
740 		if (unlikely(!msrq)) {
741 			pr_warn("CQ %06x with entry for unknown SRQN %06x\n",
742 				cq->mcq.cqn, srq_num);
743 			return -EINVAL;
744 		}
745 	}
746 
747 	if (is_send) {
748 		wq = &(*cur_qp)->sq;
749 		if (!(*cur_qp)->sq_signal_bits) {
750 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
751 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
752 		}
753 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
754 		++wq->tail;
755 	} else if ((*cur_qp)->ibqp.srq) {
756 		srq = to_msrq((*cur_qp)->ibqp.srq);
757 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
758 		wc->wr_id = srq->wrid[wqe_ctr];
759 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
760 	} else if (msrq) {
761 		srq = to_mibsrq(msrq);
762 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
763 		wc->wr_id = srq->wrid[wqe_ctr];
764 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
765 	} else {
766 		wq	  = &(*cur_qp)->rq;
767 		tail	  = wq->tail & (wq->wqe_cnt - 1);
768 		wc->wr_id = wq->wrid[tail];
769 		++wq->tail;
770 	}
771 
772 	if (unlikely(is_error)) {
773 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
774 		return 0;
775 	}
776 
777 	wc->status = IB_WC_SUCCESS;
778 
779 	if (is_send) {
780 		wc->wc_flags = 0;
781 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
782 		case MLX4_OPCODE_RDMA_WRITE_IMM:
783 			wc->wc_flags |= IB_WC_WITH_IMM;
784 			/* fall through */
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 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
853 			if ((*cur_qp)->mlx4_ib_qp_type &
854 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
855 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
856 				return use_tunnel_data
857 					(*cur_qp, cq, wc, tail, cqe,
858 					 rdma_port_get_link_layer
859 						(wc->qp->device,
860 						 (*cur_qp)->port) ==
861 						IB_LINK_LAYER_ETHERNET);
862 		}
863 
864 		if (timestamp_en) {
865 			/* currently, only CQ_CREATE_WITH_TIMESTAMPING_RAW is
866 			 * supported. CQ_CREATE_WITH_TIMESTAMPING_SYS isn't
867 			 * supported */
868 			if (cq->create_flags & IB_CQ_TIMESTAMP_TO_SYS_TIME) {
869 				wc->ts.timestamp = 0;
870 			} else {
871 				wc->ts.timestamp =
872 					((u64)(be32_to_cpu(cqe->timestamp_16_47)
873 					       + !cqe->timestamp_0_15) << 16)
874 					| be16_to_cpu(cqe->timestamp_0_15);
875 				wc->wc_flags |= IB_WC_WITH_TIMESTAMP;
876 			}
877 		} else {
878 			wc->wc_flags |= IB_WC_WITH_SLID;
879 			wc->slid	   = be16_to_cpu(cqe->rlid);
880 		}
881 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
882 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
883 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
884 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
885 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
886 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
887 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
888 		if (!timestamp_en) {
889 			if (rdma_port_get_link_layer(wc->qp->device,
890 						     (*cur_qp)->port) ==
891 						      IB_LINK_LAYER_ETHERNET)
892 				wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
893 			else
894 				wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
895 			wc->wc_flags	  |= IB_WC_WITH_SL;
896 		}
897 		if ((be32_to_cpu(cqe->vlan_my_qpn) &
898 		    MLX4_CQE_VLAN_PRESENT_MASK) && !timestamp_en) {
899 			wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
900 				MLX4_CQE_VID_MASK;
901 			wc->wc_flags	  |= IB_WC_WITH_VLAN;
902 		} else {
903 			wc->vlan_id = 0xffff;
904 		}
905 		if (!timestamp_en) {
906 			memcpy(wc->smac, cqe->smac, 6);
907 			wc->wc_flags |= IB_WC_WITH_SMAC;
908 		}
909 	}
910 
911 	return 0;
912 }
913 
914 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
915 {
916 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
917 	struct mlx4_ib_qp *cur_qp = NULL;
918 	unsigned long flags;
919 	int npolled;
920 	int err = 0;
921 
922 	spin_lock_irqsave(&cq->lock, flags);
923 
924 	for (npolled = 0; npolled < num_entries; ++npolled) {
925 		err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
926 		if (err)
927 			break;
928 	}
929 
930 	mlx4_cq_set_ci(&cq->mcq);
931 
932 	spin_unlock_irqrestore(&cq->lock, flags);
933 
934 	if (err == 0 || err == -EAGAIN)
935 		return npolled;
936 	else
937 		return err;
938 }
939 
940 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
941 {
942 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
943 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
944 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
945 		    to_mdev(ibcq->device)->priv_uar.map,
946 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
947 
948 	return 0;
949 }
950 
951 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
952 {
953 	u32 prod_index;
954 	int nfreed = 0;
955 	struct mlx4_cqe *cqe, *dest;
956 	u8 owner_bit;
957 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
958 
959 	/*
960 	 * First we need to find the current producer index, so we
961 	 * know where to start cleaning from.  It doesn't matter if HW
962 	 * adds new entries after this loop -- the QP we're worried
963 	 * about is already in RESET, so the new entries won't come
964 	 * from our QP and therefore don't need to be checked.
965 	 */
966 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
967 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
968 			break;
969 
970 	/*
971 	 * Now sweep backwards through the CQ, removing CQ entries
972 	 * that match our QP by copying older entries on top of them.
973 	 */
974 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
975 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
976 		cqe += cqe_inc;
977 
978 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
979 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
980 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
981 			++nfreed;
982 		} else if (nfreed) {
983 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
984 			dest += cqe_inc;
985 
986 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
987 			memcpy(dest, cqe, sizeof *cqe);
988 			dest->owner_sr_opcode = owner_bit |
989 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
990 		}
991 	}
992 
993 	if (nfreed) {
994 		cq->mcq.cons_index += nfreed;
995 		/*
996 		 * Make sure update of buffer contents is done before
997 		 * updating consumer index.
998 		 */
999 		wmb();
1000 		mlx4_cq_set_ci(&cq->mcq);
1001 	}
1002 }
1003 
1004 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
1005 {
1006 	spin_lock_irq(&cq->lock);
1007 	__mlx4_ib_cq_clean(cq, qpn, srq);
1008 	spin_unlock_irq(&cq->lock);
1009 }
1010