xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_cq.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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 	void *buf_addr;
184 	int err;
185 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
186 		udata, struct mlx4_ib_ucontext, ibucontext);
187 
188 	if (entries < 1 || entries > dev->dev->caps.max_cqes)
189 		return -EINVAL;
190 
191 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
192 		return -EINVAL;
193 
194 	entries      = roundup_pow_of_two(entries + 1);
195 	cq->ibcq.cqe = entries - 1;
196 	mutex_init(&cq->resize_mutex);
197 	spin_lock_init(&cq->lock);
198 	cq->resize_buf = NULL;
199 	cq->resize_umem = NULL;
200 	cq->create_flags = attr->flags;
201 	INIT_LIST_HEAD(&cq->send_qp_list);
202 	INIT_LIST_HEAD(&cq->recv_qp_list);
203 
204 	if (udata) {
205 		struct mlx4_ib_create_cq ucmd;
206 
207 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
208 			err = -EFAULT;
209 			goto err_cq;
210 		}
211 
212 		buf_addr = (void *)(unsigned long)ucmd.buf_addr;
213 		err = mlx4_ib_get_cq_umem(dev, udata, &cq->buf, &cq->umem,
214 					  ucmd.buf_addr, entries);
215 		if (err)
216 			goto err_cq;
217 
218 		err = mlx4_ib_db_map_user(context, ucmd.db_addr, &cq->db);
219 		if (err)
220 			goto err_mtt;
221 
222 		uar = &context->uar;
223 	} else {
224 		err = mlx4_db_alloc(dev->dev, &cq->db, 1, GFP_KERNEL);
225 		if (err)
226 			goto err_cq;
227 
228 		cq->mcq.set_ci_db  = cq->db.db;
229 		cq->mcq.arm_db     = cq->db.db + 1;
230 		*cq->mcq.set_ci_db = 0;
231 		*cq->mcq.arm_db    = 0;
232 
233 		err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
234 		if (err)
235 			goto err_db;
236 
237 		buf_addr = &cq->buf.buf;
238 
239 		uar = &dev->priv_uar;
240 	}
241 
242 	if (dev->eq_table)
243 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
244 
245 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
246 			    cq->db.dma, &cq->mcq, vector, 0,
247 			    !!(cq->create_flags & IB_CQ_FLAGS_TIMESTAMP_COMPLETION));
248 	if (err)
249 		goto err_dbmap;
250 
251 	cq->mcq.comp = mlx4_ib_cq_comp;
252 	cq->mcq.event = mlx4_ib_cq_event;
253 
254 	if (udata)
255 		if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
256 			err = -EFAULT;
257 			goto err_cq_free;
258 		}
259 
260 	return 0;
261 
262 err_cq_free:
263 	mlx4_cq_free(dev->dev, &cq->mcq);
264 
265 err_dbmap:
266 	if (udata)
267 		mlx4_ib_db_unmap_user(context, &cq->db);
268 
269 err_mtt:
270 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
271 
272 	ib_umem_release(cq->umem);
273 	if (!udata)
274 		mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
275 
276 err_db:
277 	if (!udata)
278 		mlx4_db_free(dev->dev, &cq->db);
279 err_cq:
280 	return err;
281 }
282 
283 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
284 				  int entries)
285 {
286 	int err;
287 
288 	if (cq->resize_buf)
289 		return -EBUSY;
290 
291 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
292 	if (!cq->resize_buf)
293 		return -ENOMEM;
294 
295 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
296 	if (err) {
297 		kfree(cq->resize_buf);
298 		cq->resize_buf = NULL;
299 		return err;
300 	}
301 
302 	cq->resize_buf->cqe = entries - 1;
303 
304 	return 0;
305 }
306 
307 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
308 				   int entries, struct ib_udata *udata)
309 {
310 	struct mlx4_ib_resize_cq ucmd;
311 	int err;
312 
313 	if (cq->resize_umem)
314 		return -EBUSY;
315 
316 	if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
317 		return -EFAULT;
318 
319 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_KERNEL);
320 	if (!cq->resize_buf)
321 		return -ENOMEM;
322 
323 	err = mlx4_ib_get_cq_umem(dev, udata, &cq->resize_buf->buf,
324 				  &cq->resize_umem, ucmd.buf_addr, entries);
325 	if (err) {
326 		kfree(cq->resize_buf);
327 		cq->resize_buf = NULL;
328 		return err;
329 	}
330 
331 	cq->resize_buf->cqe = entries - 1;
332 
333 	return 0;
334 }
335 
336 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
337 {
338 	u32 i;
339 
340 	i = cq->mcq.cons_index;
341 	while (get_sw_cqe(cq, i))
342 		++i;
343 
344 	return i - cq->mcq.cons_index;
345 }
346 
347 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
348 {
349 	struct mlx4_cqe *cqe, *new_cqe;
350 	int i;
351 	int cqe_size = cq->buf.entry_size;
352 	int cqe_inc = cqe_size == 64 ? 1 : 0;
353 
354 	i = cq->mcq.cons_index;
355 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
356 	cqe += cqe_inc;
357 
358 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
359 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
360 					   (i + 1) & cq->resize_buf->cqe);
361 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
362 		new_cqe += cqe_inc;
363 
364 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
365 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
366 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
367 		cqe += cqe_inc;
368 	}
369 	++cq->mcq.cons_index;
370 }
371 
372 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
373 {
374 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
375 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
376 	struct mlx4_mtt mtt;
377 	int outst_cqe;
378 	int err;
379 
380 	mutex_lock(&cq->resize_mutex);
381 	if (entries < 1 || entries > dev->dev->caps.max_cqes) {
382 		err = -EINVAL;
383 		goto out;
384 	}
385 
386 	entries = roundup_pow_of_two(entries + 1);
387 	if (entries == ibcq->cqe + 1) {
388 		err = 0;
389 		goto out;
390 	}
391 
392 	if (entries > dev->dev->caps.max_cqes + 1) {
393 		err = -EINVAL;
394 		goto out;
395 	}
396 
397 	if (ibcq->uobject) {
398 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
399 		if (err)
400 			goto out;
401 	} else {
402 		/* Can't be smaller than the number of outstanding CQEs */
403 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
404 		if (entries < outst_cqe + 1) {
405 			err = -EINVAL;
406 			goto out;
407 		}
408 
409 		err = mlx4_alloc_resize_buf(dev, cq, entries);
410 		if (err)
411 			goto out;
412 	}
413 
414 	mtt = cq->buf.mtt;
415 
416 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
417 	if (err)
418 		goto err_buf;
419 
420 	mlx4_mtt_cleanup(dev->dev, &mtt);
421 	if (ibcq->uobject) {
422 		cq->buf      = cq->resize_buf->buf;
423 		cq->ibcq.cqe = cq->resize_buf->cqe;
424 		ib_umem_release(cq->umem);
425 		cq->umem     = cq->resize_umem;
426 
427 		kfree(cq->resize_buf);
428 		cq->resize_buf = NULL;
429 		cq->resize_umem = NULL;
430 	} else {
431 		struct mlx4_ib_cq_buf tmp_buf;
432 		int tmp_cqe = 0;
433 
434 		spin_lock_irq(&cq->lock);
435 		if (cq->resize_buf) {
436 			mlx4_ib_cq_resize_copy_cqes(cq);
437 			tmp_buf = cq->buf;
438 			tmp_cqe = cq->ibcq.cqe;
439 			cq->buf      = cq->resize_buf->buf;
440 			cq->ibcq.cqe = cq->resize_buf->cqe;
441 
442 			kfree(cq->resize_buf);
443 			cq->resize_buf = NULL;
444 		}
445 		spin_unlock_irq(&cq->lock);
446 
447 		if (tmp_cqe)
448 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
449 	}
450 
451 	goto out;
452 
453 err_buf:
454 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
455 	if (!ibcq->uobject)
456 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
457 				    cq->resize_buf->cqe);
458 
459 	kfree(cq->resize_buf);
460 	cq->resize_buf = NULL;
461 
462 	ib_umem_release(cq->resize_umem);
463 	cq->resize_umem = NULL;
464 out:
465 	mutex_unlock(&cq->resize_mutex);
466 
467 	return err;
468 }
469 
470 void mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
471 {
472 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
473 	struct mlx4_ib_cq *mcq = to_mcq(cq);
474 
475 	mlx4_cq_free(dev->dev, &mcq->mcq);
476 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
477 
478 	if (udata) {
479 		mlx4_ib_db_unmap_user(
480 			rdma_udata_to_drv_context(
481 				udata,
482 				struct mlx4_ib_ucontext,
483 				ibucontext),
484 			&mcq->db);
485 	} else {
486 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
487 		mlx4_db_free(dev->dev, &mcq->db);
488 	}
489 	ib_umem_release(mcq->umem);
490 }
491 
492 static void dump_cqe(void *cqe)
493 {
494 	__be32 *buf = cqe;
495 
496 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
497 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
498 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
499 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
500 }
501 
502 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
503 				     struct ib_wc *wc)
504 {
505 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
506 		pr_debug("local QP operation err "
507 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
508 		       "opcode = %02x)\n",
509 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
510 		       cqe->vendor_err_syndrome,
511 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
512 		dump_cqe(cqe);
513 	}
514 
515 	switch (cqe->syndrome) {
516 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
517 		wc->status = IB_WC_LOC_LEN_ERR;
518 		break;
519 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
520 		wc->status = IB_WC_LOC_QP_OP_ERR;
521 		break;
522 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
523 		wc->status = IB_WC_LOC_PROT_ERR;
524 		break;
525 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
526 		wc->status = IB_WC_WR_FLUSH_ERR;
527 		break;
528 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
529 		wc->status = IB_WC_MW_BIND_ERR;
530 		break;
531 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
532 		wc->status = IB_WC_BAD_RESP_ERR;
533 		break;
534 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
535 		wc->status = IB_WC_LOC_ACCESS_ERR;
536 		break;
537 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
538 		wc->status = IB_WC_REM_INV_REQ_ERR;
539 		break;
540 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
541 		wc->status = IB_WC_REM_ACCESS_ERR;
542 		break;
543 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
544 		wc->status = IB_WC_REM_OP_ERR;
545 		break;
546 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
547 		wc->status = IB_WC_RETRY_EXC_ERR;
548 		break;
549 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
550 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
551 		break;
552 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
553 		wc->status = IB_WC_REM_ABORT_ERR;
554 		break;
555 	default:
556 		wc->status = IB_WC_GENERAL_ERR;
557 		break;
558 	}
559 
560 	wc->vendor_err = cqe->vendor_err_syndrome;
561 }
562 
563 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
564 {
565 	return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4      |
566 				      MLX4_CQE_STATUS_IPV4F     |
567 				      MLX4_CQE_STATUS_IPV4OPT   |
568 				      MLX4_CQE_STATUS_IPV6      |
569 				      MLX4_CQE_STATUS_IPOK)) ==
570 		cpu_to_be16(MLX4_CQE_STATUS_IPV4        |
571 			    MLX4_CQE_STATUS_IPOK))              &&
572 		(status & cpu_to_be16(MLX4_CQE_STATUS_UDP       |
573 				      MLX4_CQE_STATUS_TCP))     &&
574 		checksum == cpu_to_be16(0xffff);
575 }
576 
577 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
578 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
579 {
580 	struct mlx4_ib_proxy_sqp_hdr *hdr;
581 
582 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
583 				   qp->sqp_proxy_rcv[tail].map,
584 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
585 				   DMA_FROM_DEVICE);
586 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
587 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
588 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
589 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
590 	wc->dlid_path_bits = 0;
591 
592 	if (is_eth) {
593 		wc->slid = 0;
594 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
595 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
596 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
597 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
598 	} else {
599 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
600 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
601 	}
602 }
603 
604 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
605 			       struct ib_wc *wc, int *npolled, int is_send)
606 {
607 	struct mlx4_ib_wq *wq;
608 	unsigned cur;
609 	int i;
610 
611 	wq = is_send ? &qp->sq : &qp->rq;
612 	cur = wq->head - wq->tail;
613 
614 	if (cur == 0)
615 		return;
616 
617 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
618 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
619 		wc->status = IB_WC_WR_FLUSH_ERR;
620 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
621 		wq->tail++;
622 		(*npolled)++;
623 		wc->qp = &qp->ibqp;
624 		wc++;
625 	}
626 }
627 
628 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
629 				 struct ib_wc *wc, int *npolled)
630 {
631 	struct mlx4_ib_qp *qp;
632 
633 	*npolled = 0;
634 	/* Find uncompleted WQEs belonging to that cq and retrun
635 	 * simulated FLUSH_ERR completions
636 	 */
637 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
638 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
639 		if (*npolled >= num_entries)
640 			goto out;
641 	}
642 
643 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
644 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
645 		if (*npolled >= num_entries)
646 			goto out;
647 	}
648 
649 out:
650 	return;
651 }
652 
653 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
654 			    struct mlx4_ib_qp **cur_qp,
655 			    struct ib_wc *wc)
656 {
657 	struct mlx4_cqe *cqe;
658 	struct mlx4_qp *mqp;
659 	struct mlx4_ib_wq *wq;
660 	struct mlx4_ib_srq *srq;
661 	struct mlx4_srq *msrq = NULL;
662 	int is_send;
663 	int is_error;
664 	int is_eth;
665 	u32 g_mlpath_rqpn;
666 	u16 wqe_ctr;
667 	unsigned tail = 0;
668 
669 repoll:
670 	cqe = next_cqe_sw(cq);
671 	if (!cqe)
672 		return -EAGAIN;
673 
674 	if (cq->buf.entry_size == 64)
675 		cqe++;
676 
677 	++cq->mcq.cons_index;
678 
679 	/*
680 	 * Make sure we read CQ entry contents after we've checked the
681 	 * ownership bit.
682 	 */
683 	rmb();
684 
685 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
686 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
687 		MLX4_CQE_OPCODE_ERROR;
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 		*cur_qp = to_mibqp(mqp);
715 	}
716 
717 	wc->qp = &(*cur_qp)->ibqp;
718 
719 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
720 		u32 srq_num;
721 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
722 		srq_num       = g_mlpath_rqpn & 0xffffff;
723 		/* SRQ is also in the radix tree */
724 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
725 				       srq_num);
726 	}
727 
728 	if (is_send) {
729 		wq = &(*cur_qp)->sq;
730 		if (!(*cur_qp)->sq_signal_bits) {
731 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
732 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
733 		}
734 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
735 		++wq->tail;
736 	} else if ((*cur_qp)->ibqp.srq) {
737 		srq = to_msrq((*cur_qp)->ibqp.srq);
738 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
739 		wc->wr_id = srq->wrid[wqe_ctr];
740 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
741 	} else if (msrq) {
742 		srq = to_mibsrq(msrq);
743 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
744 		wc->wr_id = srq->wrid[wqe_ctr];
745 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
746 	} else {
747 		wq	  = &(*cur_qp)->rq;
748 		tail	  = wq->tail & (wq->wqe_cnt - 1);
749 		wc->wr_id = wq->wrid[tail];
750 		++wq->tail;
751 	}
752 
753 	if (unlikely(is_error)) {
754 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
755 		return 0;
756 	}
757 
758 	wc->status = IB_WC_SUCCESS;
759 
760 	if (is_send) {
761 		wc->wc_flags = 0;
762 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
763 		case MLX4_OPCODE_RDMA_WRITE_IMM:
764 			wc->wc_flags |= IB_WC_WITH_IMM;
765 		case MLX4_OPCODE_RDMA_WRITE:
766 			wc->opcode    = IB_WC_RDMA_WRITE;
767 			break;
768 		case MLX4_OPCODE_SEND_IMM:
769 			wc->wc_flags |= IB_WC_WITH_IMM;
770 		case MLX4_OPCODE_SEND:
771 		case MLX4_OPCODE_SEND_INVAL:
772 			wc->opcode    = IB_WC_SEND;
773 			break;
774 		case MLX4_OPCODE_RDMA_READ:
775 			wc->opcode    = IB_WC_RDMA_READ;
776 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
777 			break;
778 		case MLX4_OPCODE_ATOMIC_CS:
779 			wc->opcode    = IB_WC_COMP_SWAP;
780 			wc->byte_len  = 8;
781 			break;
782 		case MLX4_OPCODE_ATOMIC_FA:
783 			wc->opcode    = IB_WC_FETCH_ADD;
784 			wc->byte_len  = 8;
785 			break;
786 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
787 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
788 			wc->byte_len  = 8;
789 			break;
790 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
791 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
792 			wc->byte_len  = 8;
793 			break;
794 		case MLX4_OPCODE_LSO:
795 			wc->opcode    = IB_WC_LSO;
796 			break;
797 		case MLX4_OPCODE_FMR:
798 			wc->opcode    = IB_WC_REG_MR;
799 			break;
800 		case MLX4_OPCODE_LOCAL_INVAL:
801 			wc->opcode    = IB_WC_LOCAL_INV;
802 			break;
803 		}
804 	} else {
805 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
806 
807 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
808 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
809 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
810 			wc->wc_flags	= IB_WC_WITH_IMM;
811 			wc->ex.imm_data = cqe->immed_rss_invalid;
812 			break;
813 		case MLX4_RECV_OPCODE_SEND_INVAL:
814 			wc->opcode	= IB_WC_RECV;
815 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
816 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
817 			break;
818 		case MLX4_RECV_OPCODE_SEND:
819 			wc->opcode   = IB_WC_RECV;
820 			wc->wc_flags = 0;
821 			break;
822 		case MLX4_RECV_OPCODE_SEND_IMM:
823 			wc->opcode	= IB_WC_RECV;
824 			wc->wc_flags	= IB_WC_WITH_IMM;
825 			wc->ex.imm_data = cqe->immed_rss_invalid;
826 			break;
827 		}
828 
829 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
830 						  (*cur_qp)->port) ==
831 			  IB_LINK_LAYER_ETHERNET);
832 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
833 			if ((*cur_qp)->mlx4_ib_qp_type &
834 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
835 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
836 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
837 						is_eth);
838 				return 0;
839 			}
840 		}
841 
842 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
843 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
844 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
845 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
846 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
847 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
848 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
849 		if (is_eth) {
850 			wc->slid = 0;
851 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
852 			if (be32_to_cpu(cqe->vlan_my_qpn) &
853 					MLX4_CQE_CVLAN_PRESENT_MASK) {
854 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
855 					MLX4_CQE_VID_MASK;
856 			} else {
857 				wc->vlan_id = 0xffff;
858 			}
859 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
860 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
861 		} else {
862 			wc->slid = be16_to_cpu(cqe->rlid);
863 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
864 			wc->vlan_id = 0xffff;
865 		}
866 	}
867 
868 	return 0;
869 }
870 
871 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
872 {
873 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
874 	struct mlx4_ib_qp *cur_qp = NULL;
875 	unsigned long flags;
876 	int npolled;
877 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
878 
879 	spin_lock_irqsave(&cq->lock, flags);
880 	if (unlikely(mdev->dev->persist->state &
881 		     MLX4_DEVICE_STATE_INTERNAL_ERROR)) {
882 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
883 		goto out;
884 	}
885 
886 	for (npolled = 0; npolled < num_entries; ++npolled) {
887 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
888 			break;
889 	}
890 
891 	mlx4_cq_set_ci(&cq->mcq);
892 
893 out:
894 	spin_unlock_irqrestore(&cq->lock, flags);
895 
896 	return npolled;
897 }
898 
899 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
900 {
901 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
902 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
903 
904 	if (unlikely(mdev->dev->persist->state &
905 		     MLX4_DEVICE_STATE_INTERNAL_ERROR))
906 		return -1;
907 
908 	mlx4_cq_arm(&cq->mcq,
909 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
910 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
911 		    mdev->uar_map,
912 		    MLX4_GET_DOORBELL_LOCK(&mdev->uar_lock));
913 
914 	return 0;
915 }
916 
917 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
918 {
919 	u32 prod_index;
920 	int nfreed = 0;
921 	struct mlx4_cqe *cqe, *dest;
922 	u8 owner_bit;
923 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
924 
925 	/*
926 	 * First we need to find the current producer index, so we
927 	 * know where to start cleaning from.  It doesn't matter if HW
928 	 * adds new entries after this loop -- the QP we're worried
929 	 * about is already in RESET, so the new entries won't come
930 	 * from our QP and therefore don't need to be checked.
931 	 */
932 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
933 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
934 			break;
935 
936 	/*
937 	 * Now sweep backwards through the CQ, removing CQ entries
938 	 * that match our QP by copying older entries on top of them.
939 	 */
940 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
941 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
942 		cqe += cqe_inc;
943 
944 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
945 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
946 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
947 			++nfreed;
948 		} else if (nfreed) {
949 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
950 			dest += cqe_inc;
951 
952 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
953 			memcpy(dest, cqe, sizeof *cqe);
954 			dest->owner_sr_opcode = owner_bit |
955 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
956 		}
957 	}
958 
959 	if (nfreed) {
960 		cq->mcq.cons_index += nfreed;
961 		/*
962 		 * Make sure update of buffer contents is done before
963 		 * updating consumer index.
964 		 */
965 		wmb();
966 		mlx4_cq_set_ci(&cq->mcq);
967 	}
968 }
969 
970 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
971 {
972 	spin_lock_irq(&cq->lock);
973 	__mlx4_ib_cq_clean(cq, qpn, srq);
974 	spin_unlock_irq(&cq->lock);
975 }
976