xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_cq.c (revision 031beb4e239bfce798af17f5fe8dba8bcaf13d99)
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 <dev/mlx4/driver.h>
38 #include <linux/slab.h>
39 
40 #include "mlx4_ib.h"
41 #include <rdma/mlx4-abi.h>
42 #include <rdma/uverbs_ioctl.h>
43 
44 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
45 {
46 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
47 	ibcq->comp_handler(ibcq, ibcq->cq_context);
48 }
49 
50 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
51 {
52 	struct ib_event event;
53 	struct ib_cq *ibcq;
54 
55 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
56 		pr_warn("Unexpected event type %d "
57 		       "on CQ %06x\n", type, cq->cqn);
58 		return;
59 	}
60 
61 	ibcq = &to_mibcq(cq)->ibcq;
62 	if (ibcq->event_handler) {
63 		event.device     = ibcq->device;
64 		event.event      = IB_EVENT_CQ_ERR;
65 		event.element.cq = ibcq;
66 		ibcq->event_handler(&event, ibcq->cq_context);
67 	}
68 }
69 
70 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
71 {
72 	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
73 }
74 
75 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
76 {
77 	return get_cqe_from_buf(&cq->buf, n);
78 }
79 
80 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
81 {
82 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
83 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
84 
85 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
86 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
87 }
88 
89 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
90 {
91 	return get_sw_cqe(cq, cq->mcq.cons_index);
92 }
93 
94 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
95 {
96 	struct mlx4_ib_cq *mcq = to_mcq(cq);
97 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
98 
99 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
100 }
101 
102 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
103 {
104 	int err;
105 
106 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
107 			     PAGE_SIZE * 2, &buf->buf, GFP_KERNEL);
108 
109 	if (err)
110 		goto out;
111 
112 	buf->entry_size = dev->dev->caps.cqe_size;
113 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
114 				    &buf->mtt);
115 	if (err)
116 		goto err_buf;
117 
118 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf, GFP_KERNEL);
119 	if (err)
120 		goto err_mtt;
121 
122 	return 0;
123 
124 err_mtt:
125 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
126 
127 err_buf:
128 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
129 
130 out:
131 	return err;
132 }
133 
134 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
135 {
136 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
137 }
138 
139 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_udata *udata,
140 			       struct mlx4_ib_cq_buf *buf,
141 			       struct ib_umem **umem, u64 buf_addr, int cqe)
142 {
143 	int err;
144 	int cqe_size = dev->dev->caps.cqe_size;
145 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
146 		udata, struct mlx4_ib_ucontext, ibucontext);
147 
148 	*umem = ib_umem_get(&context->ibucontext, buf_addr, cqe * cqe_size,
149 			    IB_ACCESS_LOCAL_WRITE, 1);
150 	if (IS_ERR(*umem))
151 		return PTR_ERR(*umem);
152 
153 	err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
154 			    ilog2((*umem)->page_size), &buf->mtt);
155 	if (err)
156 		goto err_buf;
157 
158 	err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
159 	if (err)
160 		goto err_mtt;
161 
162 	return 0;
163 
164 err_mtt:
165 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
166 
167 err_buf:
168 	ib_umem_release(*umem);
169 
170 	return err;
171 }
172 
173 #define CQ_CREATE_FLAGS_SUPPORTED IB_CQ_FLAGS_TIMESTAMP_COMPLETION
174 int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
175 		      struct ib_udata *udata)
176 {
177 	struct ib_device *ibdev = ibcq->device;
178 	int entries = attr->cqe;
179 	int vector = attr->comp_vector;
180 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
181 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
182 	struct mlx4_uar *uar;
183 	int err;
184 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
185 		udata, struct mlx4_ib_ucontext, ibucontext);
186 
187 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
188 		return -EINVAL;
189 
190 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
191 		return -EINVAL;
192 
193 	entries      = roundup_pow_of_two(entries + 1);
194 	cq->ibcq.cqe = entries - 1;
195 	mutex_init(&cq->resize_mutex);
196 	spin_lock_init(&cq->lock);
197 	cq->resize_buf = NULL;
198 	cq->resize_umem = NULL;
199 	cq->create_flags = attr->flags;
200 	INIT_LIST_HEAD(&cq->send_qp_list);
201 	INIT_LIST_HEAD(&cq->recv_qp_list);
202 
203 	if (udata) {
204 		struct mlx4_ib_create_cq ucmd;
205 
206 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
207 			err = -EFAULT;
208 			goto err_cq;
209 		}
210 
211 		err = mlx4_ib_get_cq_umem(dev, udata, &cq->buf, &cq->umem,
212 					  ucmd.buf_addr, entries);
213 		if (err)
214 			goto err_cq;
215 
216 		err = mlx4_ib_db_map_user(context, ucmd.db_addr, &cq->db);
217 		if (err)
218 			goto err_mtt;
219 
220 		uar = &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 	cq->mcq.comp = mlx4_ib_cq_comp;
248 	cq->mcq.event = mlx4_ib_cq_event;
249 
250 	if (udata)
251 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
252 			err = -EFAULT;
253 			goto err_cq_free;
254 		}
255 
256 	return 0;
257 
258 err_cq_free:
259 	mlx4_cq_free(dev->dev, &cq->mcq);
260 
261 err_dbmap:
262 	if (udata)
263 		mlx4_ib_db_unmap_user(context, &cq->db);
264 
265 err_mtt:
266 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
267 
268 	ib_umem_release(cq->umem);
269 	if (!udata)
270 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
271 
272 err_db:
273 	if (!udata)
274 		mlx4_db_free(dev->dev, &cq->db);
275 err_cq:
276 	return err;
277 }
278 
279 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
280 				  int entries)
281 {
282 	int err;
283 
284 	if (cq->resize_buf)
285 		return -EBUSY;
286 
287 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
288 	if (!cq->resize_buf)
289 		return -ENOMEM;
290 
291 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
292 	if (err) {
293 		kfree(cq->resize_buf);
294 		cq->resize_buf = NULL;
295 		return err;
296 	}
297 
298 	cq->resize_buf->cqe = entries - 1;
299 
300 	return 0;
301 }
302 
303 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
304 				   int entries, struct ib_udata *udata)
305 {
306 	struct mlx4_ib_resize_cq ucmd;
307 	int err;
308 
309 	if (cq->resize_umem)
310 		return -EBUSY;
311 
312 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
313 		return -EFAULT;
314 
315 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
316 	if (!cq->resize_buf)
317 		return -ENOMEM;
318 
319 	err = mlx4_ib_get_cq_umem(dev, udata, &cq->resize_buf->buf,
320 				  &cq->resize_umem, ucmd.buf_addr, entries);
321 	if (err) {
322 		kfree(cq->resize_buf);
323 		cq->resize_buf = NULL;
324 		return err;
325 	}
326 
327 	cq->resize_buf->cqe = entries - 1;
328 
329 	return 0;
330 }
331 
332 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
333 {
334 	u32 i;
335 
336 	i = cq->mcq.cons_index;
337 	while (get_sw_cqe(cq, i))
338 		++i;
339 
340 	return i - cq->mcq.cons_index;
341 }
342 
343 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
344 {
345 	struct mlx4_cqe *cqe, *new_cqe;
346 	int i;
347 	int cqe_size = cq->buf.entry_size;
348 	int cqe_inc = cqe_size == 64 ? 1 : 0;
349 
350 	i = cq->mcq.cons_index;
351 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
352 	cqe += cqe_inc;
353 
354 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
355 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
356 					   (i + 1) & cq->resize_buf->cqe);
357 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
358 		new_cqe += cqe_inc;
359 
360 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
361 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
362 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
363 		cqe += cqe_inc;
364 	}
365 	++cq->mcq.cons_index;
366 }
367 
368 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
369 {
370 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
371 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
372 	struct mlx4_mtt mtt;
373 	int outst_cqe;
374 	int err;
375 
376 	mutex_lock(&cq->resize_mutex);
377 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
378 		err = -EINVAL;
379 		goto out;
380 	}
381 
382 	entries = roundup_pow_of_two(entries + 1);
383 	if (entries == ibcq->cqe + 1) {
384 		err = 0;
385 		goto out;
386 	}
387 
388 	if (entries > dev->dev->caps.max_cqes + 1) {
389 		err = -EINVAL;
390 		goto out;
391 	}
392 
393 	if (ibcq->uobject) {
394 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
395 		if (err)
396 			goto out;
397 	} else {
398 		/* Can't be smaller than the number of outstanding CQEs */
399 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
400 		if (entries < outst_cqe + 1) {
401 			err = -EINVAL;
402 			goto out;
403 		}
404 
405 		err = mlx4_alloc_resize_buf(dev, cq, entries);
406 		if (err)
407 			goto out;
408 	}
409 
410 	mtt = cq->buf.mtt;
411 
412 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
413 	if (err)
414 		goto err_buf;
415 
416 	mlx4_mtt_cleanup(dev->dev, &mtt);
417 	if (ibcq->uobject) {
418 		cq->buf      = cq->resize_buf->buf;
419 		cq->ibcq.cqe = cq->resize_buf->cqe;
420 		ib_umem_release(cq->umem);
421 		cq->umem     = cq->resize_umem;
422 
423 		kfree(cq->resize_buf);
424 		cq->resize_buf = NULL;
425 		cq->resize_umem = NULL;
426 	} else {
427 		struct mlx4_ib_cq_buf tmp_buf;
428 		int tmp_cqe = 0;
429 
430 		spin_lock_irq(&cq->lock);
431 		if (cq->resize_buf) {
432 			mlx4_ib_cq_resize_copy_cqes(cq);
433 			tmp_buf = cq->buf;
434 			tmp_cqe = cq->ibcq.cqe;
435 			cq->buf      = cq->resize_buf->buf;
436 			cq->ibcq.cqe = cq->resize_buf->cqe;
437 
438 			kfree(cq->resize_buf);
439 			cq->resize_buf = NULL;
440 		}
441 		spin_unlock_irq(&cq->lock);
442 
443 		if (tmp_cqe)
444 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
445 	}
446 
447 	goto out;
448 
449 err_buf:
450 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
451 	if (!ibcq->uobject)
452 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
453 				    cq->resize_buf->cqe);
454 
455 	kfree(cq->resize_buf);
456 	cq->resize_buf = NULL;
457 
458 	ib_umem_release(cq->resize_umem);
459 	cq->resize_umem = NULL;
460 out:
461 	mutex_unlock(&cq->resize_mutex);
462 
463 	return err;
464 }
465 
466 void mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
467 {
468 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
469 	struct mlx4_ib_cq *mcq = to_mcq(cq);
470 
471 	mlx4_cq_free(dev->dev, &mcq->mcq);
472 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
473 
474 	if (udata) {
475 		mlx4_ib_db_unmap_user(
476 			rdma_udata_to_drv_context(
477 				udata,
478 				struct mlx4_ib_ucontext,
479 				ibucontext),
480 			&mcq->db);
481 	} else {
482 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
483 		mlx4_db_free(dev->dev, &mcq->db);
484 	}
485 	ib_umem_release(mcq->umem);
486 }
487 
488 static void dump_cqe(void *cqe)
489 {
490 	__be32 *buf = cqe;
491 
492 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
493 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
494 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
495 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
496 }
497 
498 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
499 				     struct ib_wc *wc)
500 {
501 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
502 		pr_debug("local QP operation err "
503 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
504 		       "opcode = %02x)\n",
505 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
506 		       cqe->vendor_err_syndrome,
507 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
508 		dump_cqe(cqe);
509 	}
510 
511 	switch (cqe->syndrome) {
512 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
513 		wc->status = IB_WC_LOC_LEN_ERR;
514 		break;
515 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
516 		wc->status = IB_WC_LOC_QP_OP_ERR;
517 		break;
518 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
519 		wc->status = IB_WC_LOC_PROT_ERR;
520 		break;
521 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
522 		wc->status = IB_WC_WR_FLUSH_ERR;
523 		break;
524 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
525 		wc->status = IB_WC_MW_BIND_ERR;
526 		break;
527 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
528 		wc->status = IB_WC_BAD_RESP_ERR;
529 		break;
530 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
531 		wc->status = IB_WC_LOC_ACCESS_ERR;
532 		break;
533 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
534 		wc->status = IB_WC_REM_INV_REQ_ERR;
535 		break;
536 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
537 		wc->status = IB_WC_REM_ACCESS_ERR;
538 		break;
539 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
540 		wc->status = IB_WC_REM_OP_ERR;
541 		break;
542 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
543 		wc->status = IB_WC_RETRY_EXC_ERR;
544 		break;
545 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
546 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
547 		break;
548 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
549 		wc->status = IB_WC_REM_ABORT_ERR;
550 		break;
551 	default:
552 		wc->status = IB_WC_GENERAL_ERR;
553 		break;
554 	}
555 
556 	wc->vendor_err = cqe->vendor_err_syndrome;
557 }
558 
559 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
560 {
561 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
562 				      MLX4_CQE_STATUS_IPV4F     |
563 				      MLX4_CQE_STATUS_IPV4OPT   |
564 				      MLX4_CQE_STATUS_IPV6      |
565 				      MLX4_CQE_STATUS_IPOK)) ==
566 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
567 			    MLX4_CQE_STATUS_IPOK))              &&
568 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
569 				      MLX4_CQE_STATUS_TCP))     &&
570 		checksum == cpu_to_be16(0xffff);
571 }
572 
573 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
574 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
575 {
576 	struct mlx4_ib_proxy_sqp_hdr *hdr;
577 
578 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
579 				   qp->sqp_proxy_rcv[tail].map,
580 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
581 				   DMA_FROM_DEVICE);
582 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
583 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
584 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
585 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
586 	wc->dlid_path_bits = 0;
587 
588 	if (is_eth) {
589 		wc->slid = 0;
590 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
591 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
592 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
593 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
594 	} else {
595 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
596 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
597 	}
598 }
599 
600 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
601 			       struct ib_wc *wc, int *npolled, int is_send)
602 {
603 	struct mlx4_ib_wq *wq;
604 	unsigned cur;
605 	int i;
606 
607 	wq = is_send ? &qp->sq : &qp->rq;
608 	cur = wq->head - wq->tail;
609 
610 	if (cur == 0)
611 		return;
612 
613 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
614 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
615 		wc->status = IB_WC_WR_FLUSH_ERR;
616 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
617 		wq->tail++;
618 		(*npolled)++;
619 		wc->qp = &qp->ibqp;
620 		wc++;
621 	}
622 }
623 
624 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
625 				 struct ib_wc *wc, int *npolled)
626 {
627 	struct mlx4_ib_qp *qp;
628 
629 	*npolled = 0;
630 	/* Find uncompleted WQEs belonging to that cq and retrun
631 	 * simulated FLUSH_ERR completions
632 	 */
633 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
634 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
635 		if (*npolled >= num_entries)
636 			goto out;
637 	}
638 
639 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
640 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
641 		if (*npolled >= num_entries)
642 			goto out;
643 	}
644 
645 out:
646 	return;
647 }
648 
649 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
650 			    struct mlx4_ib_qp **cur_qp,
651 			    struct ib_wc *wc)
652 {
653 	struct mlx4_cqe *cqe;
654 	struct mlx4_qp *mqp;
655 	struct mlx4_ib_wq *wq;
656 	struct mlx4_ib_srq *srq;
657 	struct mlx4_srq *msrq = NULL;
658 	int is_send;
659 	int is_error;
660 	int is_eth;
661 	u32 g_mlpath_rqpn;
662 	u16 wqe_ctr;
663 	unsigned tail = 0;
664 
665 repoll:
666 	cqe = next_cqe_sw(cq);
667 	if (!cqe)
668 		return -EAGAIN;
669 
670 	if (cq->buf.entry_size == 64)
671 		cqe++;
672 
673 	++cq->mcq.cons_index;
674 
675 	/*
676 	 * Make sure we read CQ entry contents after we've checked the
677 	 * ownership bit.
678 	 */
679 	rmb();
680 
681 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
682 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
683 		MLX4_CQE_OPCODE_ERROR;
684 
685 	/* Resize CQ in progress */
686 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
687 		if (cq->resize_buf) {
688 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
689 
690 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
691 			cq->buf      = cq->resize_buf->buf;
692 			cq->ibcq.cqe = cq->resize_buf->cqe;
693 
694 			kfree(cq->resize_buf);
695 			cq->resize_buf = NULL;
696 		}
697 
698 		goto repoll;
699 	}
700 
701 	if (!*cur_qp ||
702 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
703 		/*
704 		 * We do not have to take the QP table lock here,
705 		 * because CQs will be locked while QPs are removed
706 		 * from the table.
707 		 */
708 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
709 				       be32_to_cpu(cqe->vlan_my_qpn));
710 		*cur_qp = to_mibqp(mqp);
711 	}
712 
713 	wc->qp = &(*cur_qp)->ibqp;
714 
715 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
716 		u32 srq_num;
717 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
718 		srq_num       = g_mlpath_rqpn & 0xffffff;
719 		/* SRQ is also in the radix tree */
720 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
721 				       srq_num);
722 	}
723 
724 	if (is_send) {
725 		wq = &(*cur_qp)->sq;
726 		if (!(*cur_qp)->sq_signal_bits) {
727 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
728 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
729 		}
730 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
731 		++wq->tail;
732 	} else if ((*cur_qp)->ibqp.srq) {
733 		srq = to_msrq((*cur_qp)->ibqp.srq);
734 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
735 		wc->wr_id = srq->wrid[wqe_ctr];
736 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
737 	} else if (msrq) {
738 		srq = to_mibsrq(msrq);
739 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
740 		wc->wr_id = srq->wrid[wqe_ctr];
741 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
742 	} else {
743 		wq	  = &(*cur_qp)->rq;
744 		tail	  = wq->tail & (wq->wqe_cnt - 1);
745 		wc->wr_id = wq->wrid[tail];
746 		++wq->tail;
747 	}
748 
749 	if (unlikely(is_error)) {
750 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
751 		return 0;
752 	}
753 
754 	wc->status = IB_WC_SUCCESS;
755 
756 	if (is_send) {
757 		wc->wc_flags = 0;
758 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
759 		case MLX4_OPCODE_RDMA_WRITE_IMM:
760 			wc->wc_flags |= IB_WC_WITH_IMM;
761 		case MLX4_OPCODE_RDMA_WRITE:
762 			wc->opcode    = IB_WC_RDMA_WRITE;
763 			break;
764 		case MLX4_OPCODE_SEND_IMM:
765 			wc->wc_flags |= IB_WC_WITH_IMM;
766 		case MLX4_OPCODE_SEND:
767 		case MLX4_OPCODE_SEND_INVAL:
768 			wc->opcode    = IB_WC_SEND;
769 			break;
770 		case MLX4_OPCODE_RDMA_READ:
771 			wc->opcode    = IB_WC_RDMA_READ;
772 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
773 			break;
774 		case MLX4_OPCODE_ATOMIC_CS:
775 			wc->opcode    = IB_WC_COMP_SWAP;
776 			wc->byte_len  = 8;
777 			break;
778 		case MLX4_OPCODE_ATOMIC_FA:
779 			wc->opcode    = IB_WC_FETCH_ADD;
780 			wc->byte_len  = 8;
781 			break;
782 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
783 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
784 			wc->byte_len  = 8;
785 			break;
786 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
787 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
788 			wc->byte_len  = 8;
789 			break;
790 		case MLX4_OPCODE_LSO:
791 			wc->opcode    = IB_WC_LSO;
792 			break;
793 		case MLX4_OPCODE_FMR:
794 			wc->opcode    = IB_WC_REG_MR;
795 			break;
796 		case MLX4_OPCODE_LOCAL_INVAL:
797 			wc->opcode    = IB_WC_LOCAL_INV;
798 			break;
799 		}
800 	} else {
801 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
802 
803 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
804 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
805 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
806 			wc->wc_flags	= IB_WC_WITH_IMM;
807 			wc->ex.imm_data = cqe->immed_rss_invalid;
808 			break;
809 		case MLX4_RECV_OPCODE_SEND_INVAL:
810 			wc->opcode	= IB_WC_RECV;
811 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
812 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
813 			break;
814 		case MLX4_RECV_OPCODE_SEND:
815 			wc->opcode   = IB_WC_RECV;
816 			wc->wc_flags = 0;
817 			break;
818 		case MLX4_RECV_OPCODE_SEND_IMM:
819 			wc->opcode	= IB_WC_RECV;
820 			wc->wc_flags	= IB_WC_WITH_IMM;
821 			wc->ex.imm_data = cqe->immed_rss_invalid;
822 			break;
823 		}
824 
825 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
826 						  (*cur_qp)->port) ==
827 			  IB_LINK_LAYER_ETHERNET);
828 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
829 			if ((*cur_qp)->mlx4_ib_qp_type &
830 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
831 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
832 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
833 						is_eth);
834 				return 0;
835 			}
836 		}
837 
838 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
839 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
840 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
841 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
842 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
843 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
844 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
845 		if (is_eth) {
846 			wc->slid = 0;
847 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
848 			if (be32_to_cpu(cqe->vlan_my_qpn) &
849 					MLX4_CQE_CVLAN_PRESENT_MASK) {
850 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
851 					MLX4_CQE_VID_MASK;
852 			} else {
853 				wc->vlan_id = 0xffff;
854 			}
855 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
856 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
857 		} else {
858 			wc->slid = be16_to_cpu(cqe->rlid);
859 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
860 			wc->vlan_id = 0xffff;
861 		}
862 	}
863 
864 	return 0;
865 }
866 
867 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
868 {
869 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
870 	struct mlx4_ib_qp *cur_qp = NULL;
871 	unsigned long flags;
872 	int npolled;
873 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
874 
875 	spin_lock_irqsave(&cq->lock, flags);
876 	if (unlikely(mdev->dev->persist->state &
877 		     MLX4_DEVICE_STATE_INTERNAL_ERROR)) {
878 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
879 		goto out;
880 	}
881 
882 	for (npolled = 0; npolled < num_entries; ++npolled) {
883 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
884 			break;
885 	}
886 
887 	mlx4_cq_set_ci(&cq->mcq);
888 
889 out:
890 	spin_unlock_irqrestore(&cq->lock, flags);
891 
892 	return npolled;
893 }
894 
895 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
896 {
897 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
898 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
899 
900 	if (unlikely(mdev->dev->persist->state &
901 		     MLX4_DEVICE_STATE_INTERNAL_ERROR))
902 		return -1;
903 
904 	mlx4_cq_arm(&cq->mcq,
905 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
906 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
907 		    mdev->uar_map,
908 		    MLX4_GET_DOORBELL_LOCK(&mdev->uar_lock));
909 
910 	return 0;
911 }
912 
913 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
914 {
915 	u32 prod_index;
916 	int nfreed = 0;
917 	struct mlx4_cqe *cqe, *dest;
918 	u8 owner_bit;
919 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
920 
921 	/*
922 	 * First we need to find the current producer index, so we
923 	 * know where to start cleaning from.  It doesn't matter if HW
924 	 * adds new entries after this loop -- the QP we're worried
925 	 * about is already in RESET, so the new entries won't come
926 	 * from our QP and therefore don't need to be checked.
927 	 */
928 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
929 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
930 			break;
931 
932 	/*
933 	 * Now sweep backwards through the CQ, removing CQ entries
934 	 * that match our QP by copying older entries on top of them.
935 	 */
936 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
937 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
938 		cqe += cqe_inc;
939 
940 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
941 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
942 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
943 			++nfreed;
944 		} else if (nfreed) {
945 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
946 			dest += cqe_inc;
947 
948 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
949 			memcpy(dest, cqe, sizeof *cqe);
950 			dest->owner_sr_opcode = owner_bit |
951 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
952 		}
953 	}
954 
955 	if (nfreed) {
956 		cq->mcq.cons_index += nfreed;
957 		/*
958 		 * Make sure update of buffer contents is done before
959 		 * updating consumer index.
960 		 */
961 		wmb();
962 		mlx4_cq_set_ci(&cq->mcq);
963 	}
964 }
965 
966 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
967 {
968 	spin_lock_irq(&cq->lock);
969 	__mlx4_ib_cq_clean(cq, qpn, srq);
970 	spin_unlock_irq(&cq->lock);
971 }
972