xref: /linux/drivers/infiniband/hw/mlx4/cq.c (revision fbf5df34a4dbcd09d433dd4f0916bf9b2ddb16de)
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 <rdma/mlx4-abi.h>
41 #include <rdma/uverbs_ioctl.h>
42 
43 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
44 {
45 	struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
46 	ibcq->comp_handler(ibcq, ibcq->cq_context);
47 }
48 
49 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
50 {
51 	struct ib_event event;
52 	struct ib_cq *ibcq;
53 
54 	if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
55 		pr_warn("Unexpected event type %d "
56 		       "on CQ %06x\n", type, cq->cqn);
57 		return;
58 	}
59 
60 	ibcq = &to_mibcq(cq)->ibcq;
61 	if (ibcq->event_handler) {
62 		event.device     = ibcq->device;
63 		event.event      = IB_EVENT_CQ_ERR;
64 		event.element.cq = ibcq;
65 		ibcq->event_handler(&event, ibcq->cq_context);
66 	}
67 }
68 
69 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
70 {
71 	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
72 }
73 
74 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
75 {
76 	return get_cqe_from_buf(&cq->buf, n);
77 }
78 
79 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
80 {
81 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
82 	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
83 
84 	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
85 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
86 }
87 
88 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
89 {
90 	return get_sw_cqe(cq, cq->mcq.cons_index);
91 }
92 
93 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
94 {
95 	struct mlx4_ib_cq *mcq = to_mcq(cq);
96 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
97 
98 	return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
99 }
100 
101 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
102 {
103 	int err;
104 
105 	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
106 			     PAGE_SIZE * 2, &buf->buf);
107 
108 	if (err)
109 		goto out;
110 
111 	buf->entry_size = dev->dev->caps.cqe_size;
112 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
113 				    &buf->mtt);
114 	if (err)
115 		goto err_buf;
116 
117 	err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
118 	if (err)
119 		goto err_mtt;
120 
121 	return 0;
122 
123 err_mtt:
124 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
125 
126 err_buf:
127 	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
128 
129 out:
130 	return err;
131 }
132 
133 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
134 {
135 	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
136 }
137 
138 #define CQ_CREATE_FLAGS_SUPPORTED IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION
139 int mlx4_ib_create_user_cq(struct ib_cq *ibcq,
140 			   const struct ib_cq_init_attr *attr,
141 			   struct uverbs_attr_bundle *attrs)
142 {
143 	struct ib_udata *udata = &attrs->driver_udata;
144 	struct ib_device *ibdev = ibcq->device;
145 	int entries = attr->cqe;
146 	int vector = attr->comp_vector;
147 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
148 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
149 	struct mlx4_ib_create_cq ucmd;
150 	int cqe_size = dev->dev->caps.cqe_size;
151 	void *buf_addr;
152 	int shift;
153 	int n;
154 	int err;
155 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
156 		udata, struct mlx4_ib_ucontext, ibucontext);
157 
158 	if (attr->cqe > dev->dev->caps.max_cqes)
159 		return -EINVAL;
160 
161 	if (attr->flags & ~CQ_CREATE_FLAGS_SUPPORTED)
162 		return -EINVAL;
163 
164 	entries      = roundup_pow_of_two(entries + 1);
165 	cq->ibcq.cqe = entries - 1;
166 	mutex_init(&cq->resize_mutex);
167 	spin_lock_init(&cq->lock);
168 	INIT_LIST_HEAD(&cq->send_qp_list);
169 	INIT_LIST_HEAD(&cq->recv_qp_list);
170 
171 	err = ib_copy_validate_udata_in(udata, ucmd, db_addr);
172 	if (err)
173 		goto err_cq;
174 
175 	if (ibcq->umem &&
176 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SW_CQ_INIT))
177 		return -EOPNOTSUPP;
178 
179 	buf_addr = (void *)(unsigned long)ucmd.buf_addr;
180 
181 	if (!ibcq->umem)
182 		ibcq->umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
183 					 entries * cqe_size,
184 					 IB_ACCESS_LOCAL_WRITE);
185 	if (IS_ERR(ibcq->umem)) {
186 		err = PTR_ERR(ibcq->umem);
187 		goto err_cq;
188 	}
189 
190 	shift = mlx4_ib_umem_calc_optimal_mtt_size(cq->ibcq.umem, 0, &n);
191 	if (shift < 0) {
192 		err = shift;
193 		goto err_cq;
194 	}
195 
196 	err = mlx4_mtt_init(dev->dev, n, shift, &cq->buf.mtt);
197 	if (err)
198 		goto err_cq;
199 
200 	err = mlx4_ib_umem_write_mtt(dev, &cq->buf.mtt, cq->ibcq.umem);
201 	if (err)
202 		goto err_mtt;
203 
204 	err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &cq->db);
205 	if (err)
206 		goto err_mtt;
207 
208 	if (dev->eq_table)
209 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
210 
211 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, &context->uar,
212 			    cq->db.dma, &cq->mcq, vector, 0,
213 			    attr->flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION,
214 			    buf_addr, true);
215 	if (err)
216 		goto err_dbmap;
217 
218 	cq->mcq.tasklet_ctx.comp = mlx4_ib_cq_comp;
219 	cq->mcq.event = mlx4_ib_cq_event;
220 	cq->mcq.usage = MLX4_RES_USAGE_USER_VERBS;
221 
222 	if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) {
223 		err = -EFAULT;
224 		goto err_cq_free;
225 	}
226 
227 	return 0;
228 
229 err_cq_free:
230 	mlx4_cq_free(dev->dev, &cq->mcq);
231 
232 err_dbmap:
233 	mlx4_ib_db_unmap_user(context, &cq->db);
234 
235 err_mtt:
236 	mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
237 	/* UMEM is released by ib_core */
238 
239 err_cq:
240 	return err;
241 }
242 
243 int mlx4_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
244 		      struct uverbs_attr_bundle *attrs)
245 {
246 	struct ib_device *ibdev = ibcq->device;
247 	int entries = attr->cqe;
248 	int vector = attr->comp_vector;
249 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
250 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
251 	void *buf_addr;
252 	int err;
253 
254 	if (attr->cqe > dev->dev->caps.max_cqes)
255 		return -EINVAL;
256 
257 	entries      = roundup_pow_of_two(entries + 1);
258 	cq->ibcq.cqe = entries - 1;
259 	mutex_init(&cq->resize_mutex);
260 	spin_lock_init(&cq->lock);
261 	INIT_LIST_HEAD(&cq->send_qp_list);
262 	INIT_LIST_HEAD(&cq->recv_qp_list);
263 
264 	err = mlx4_db_alloc(dev->dev, &cq->db, 1);
265 	if (err)
266 		return err;
267 
268 	cq->mcq.set_ci_db  = cq->db.db;
269 	cq->mcq.arm_db     = cq->db.db + 1;
270 	*cq->mcq.set_ci_db = 0;
271 	*cq->mcq.arm_db    = 0;
272 
273 	err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
274 	if (err)
275 		goto err_db;
276 
277 	buf_addr = &cq->buf.buf;
278 
279 	if (dev->eq_table)
280 		vector = dev->eq_table[vector % ibdev->num_comp_vectors];
281 
282 	err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, &dev->priv_uar,
283 			    cq->db.dma, &cq->mcq, vector, 0, 0,
284 			    buf_addr, false);
285 	if (err)
286 		goto err_buf;
287 
288 	cq->mcq.comp = mlx4_ib_cq_comp;
289 	cq->mcq.event = mlx4_ib_cq_event;
290 	cq->mcq.usage = MLX4_RES_USAGE_DRIVER;
291 
292 	return 0;
293 
294 err_buf:
295 	mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
296 
297 err_db:
298 	mlx4_db_free(dev->dev, &cq->db);
299 	return err;
300 }
301 
302 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
303 				  int entries)
304 {
305 	int err;
306 
307 	if (cq->resize_buf)
308 		return -EBUSY;
309 
310 	cq->resize_buf = kmalloc_obj(*cq->resize_buf);
311 	if (!cq->resize_buf)
312 		return -ENOMEM;
313 
314 	err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
315 	if (err) {
316 		kfree(cq->resize_buf);
317 		cq->resize_buf = NULL;
318 		return err;
319 	}
320 
321 	cq->resize_buf->cqe = entries - 1;
322 
323 	return 0;
324 }
325 
326 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
327 				   int entries, struct ib_udata *udata)
328 {
329 	struct mlx4_ib_resize_cq ucmd;
330 	int cqe_size = dev->dev->caps.cqe_size;
331 	int shift;
332 	int n;
333 	int err;
334 
335 	if (cq->resize_umem)
336 		return -EBUSY;
337 
338 	err = ib_copy_validate_udata_in(udata, ucmd, buf_addr);
339 	if (err)
340 		return err;
341 
342 	cq->resize_buf = kmalloc_obj(*cq->resize_buf);
343 	if (!cq->resize_buf)
344 		return -ENOMEM;
345 
346 	cq->resize_umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
347 				      entries * cqe_size,
348 				      IB_ACCESS_LOCAL_WRITE);
349 	if (IS_ERR(cq->resize_umem)) {
350 		err = PTR_ERR(cq->resize_umem);
351 		goto err_buf;
352 	}
353 
354 	shift = mlx4_ib_umem_calc_optimal_mtt_size(cq->resize_umem, 0, &n);
355 	if (shift < 0) {
356 		err = shift;
357 		goto err_umem;
358 	}
359 
360 	err = mlx4_mtt_init(dev->dev, n, shift, &cq->resize_buf->buf.mtt);
361 	if (err)
362 		goto err_umem;
363 
364 	err = mlx4_ib_umem_write_mtt(dev, &cq->resize_buf->buf.mtt,
365 				     cq->resize_umem);
366 	if (err)
367 		goto err_mtt;
368 
369 	cq->resize_buf->cqe = entries - 1;
370 
371 	return 0;
372 
373 err_mtt:
374 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
375 
376 err_umem:
377 	ib_umem_release(cq->resize_umem);
378 
379 err_buf:
380 	kfree(cq->resize_buf);
381 	cq->resize_buf = NULL;
382 	return err;
383 }
384 
385 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
386 {
387 	u32 i;
388 
389 	i = cq->mcq.cons_index;
390 	while (get_sw_cqe(cq, i))
391 		++i;
392 
393 	return i - cq->mcq.cons_index;
394 }
395 
396 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
397 {
398 	struct mlx4_cqe *cqe, *new_cqe;
399 	int i;
400 	int cqe_size = cq->buf.entry_size;
401 	int cqe_inc = cqe_size == 64 ? 1 : 0;
402 
403 	i = cq->mcq.cons_index;
404 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
405 	cqe += cqe_inc;
406 
407 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
408 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
409 					   (i + 1) & cq->resize_buf->cqe);
410 		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
411 		new_cqe += cqe_inc;
412 
413 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
414 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
415 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
416 		cqe += cqe_inc;
417 	}
418 	++cq->mcq.cons_index;
419 }
420 
421 int mlx4_ib_resize_cq(struct ib_cq *ibcq, unsigned int entries,
422 		      struct ib_udata *udata)
423 {
424 	struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
425 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
426 	struct mlx4_mtt mtt;
427 	int outst_cqe;
428 	int err;
429 
430 	mutex_lock(&cq->resize_mutex);
431 	if (entries > dev->dev->caps.max_cqes) {
432 		err = -EINVAL;
433 		goto out;
434 	}
435 
436 	entries = roundup_pow_of_two(entries + 1);
437 	if (entries == ibcq->cqe + 1) {
438 		err = 0;
439 		goto out;
440 	}
441 
442 	if (entries > dev->dev->caps.max_cqes + 1) {
443 		err = -EINVAL;
444 		goto out;
445 	}
446 
447 	if (ibcq->uobject) {
448 		err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
449 		if (err)
450 			goto out;
451 	} else {
452 		/* Can't be smaller than the number of outstanding CQEs */
453 		outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
454 		if (entries < outst_cqe + 1) {
455 			err = -EINVAL;
456 			goto out;
457 		}
458 
459 		err = mlx4_alloc_resize_buf(dev, cq, entries);
460 		if (err)
461 			goto out;
462 	}
463 
464 	mtt = cq->buf.mtt;
465 
466 	err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
467 	if (err)
468 		goto err_buf;
469 
470 	mlx4_mtt_cleanup(dev->dev, &mtt);
471 	if (ibcq->uobject) {
472 		cq->buf      = cq->resize_buf->buf;
473 		cq->ibcq.cqe = cq->resize_buf->cqe;
474 		ib_umem_release(cq->ibcq.umem);
475 		cq->ibcq.umem     = cq->resize_umem;
476 
477 		kfree(cq->resize_buf);
478 		cq->resize_buf = NULL;
479 		cq->resize_umem = NULL;
480 	} else {
481 		struct mlx4_ib_cq_buf tmp_buf;
482 		int tmp_cqe = 0;
483 
484 		spin_lock_irq(&cq->lock);
485 		if (cq->resize_buf) {
486 			mlx4_ib_cq_resize_copy_cqes(cq);
487 			tmp_buf = cq->buf;
488 			tmp_cqe = cq->ibcq.cqe;
489 			cq->buf      = cq->resize_buf->buf;
490 			cq->ibcq.cqe = cq->resize_buf->cqe;
491 
492 			kfree(cq->resize_buf);
493 			cq->resize_buf = NULL;
494 		}
495 		spin_unlock_irq(&cq->lock);
496 
497 		if (tmp_cqe)
498 			mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
499 	}
500 
501 	goto out;
502 
503 err_buf:
504 	mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
505 	if (!ibcq->uobject)
506 		mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
507 				    cq->resize_buf->cqe);
508 
509 	kfree(cq->resize_buf);
510 	cq->resize_buf = NULL;
511 
512 	ib_umem_release(cq->resize_umem);
513 	cq->resize_umem = NULL;
514 out:
515 	mutex_unlock(&cq->resize_mutex);
516 
517 	return err;
518 }
519 
520 int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
521 {
522 	struct mlx4_ib_dev *dev = to_mdev(cq->device);
523 	struct mlx4_ib_cq *mcq = to_mcq(cq);
524 
525 	mlx4_cq_free(dev->dev, &mcq->mcq);
526 	mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
527 
528 	if (udata) {
529 		mlx4_ib_db_unmap_user(
530 			rdma_udata_to_drv_context(
531 				udata,
532 				struct mlx4_ib_ucontext,
533 				ibucontext),
534 			&mcq->db);
535 		/* UMEM is released by ib_core */
536 	} else {
537 		mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
538 		mlx4_db_free(dev->dev, &mcq->db);
539 	}
540 	return 0;
541 }
542 
543 static void dump_cqe(void *cqe)
544 {
545 	__be32 *buf = cqe;
546 
547 	pr_debug("CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
548 	       be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
549 	       be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
550 	       be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
551 }
552 
553 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
554 				     struct ib_wc *wc)
555 {
556 	if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
557 		pr_debug("local QP operation err "
558 		       "(QPN %06x, WQE index %x, vendor syndrome %02x, "
559 		       "opcode = %02x)\n",
560 		       be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
561 		       cqe->vendor_err_syndrome,
562 		       cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
563 		dump_cqe(cqe);
564 	}
565 
566 	switch (cqe->syndrome) {
567 	case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
568 		wc->status = IB_WC_LOC_LEN_ERR;
569 		break;
570 	case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
571 		wc->status = IB_WC_LOC_QP_OP_ERR;
572 		break;
573 	case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
574 		wc->status = IB_WC_LOC_PROT_ERR;
575 		break;
576 	case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
577 		wc->status = IB_WC_WR_FLUSH_ERR;
578 		break;
579 	case MLX4_CQE_SYNDROME_MW_BIND_ERR:
580 		wc->status = IB_WC_MW_BIND_ERR;
581 		break;
582 	case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
583 		wc->status = IB_WC_BAD_RESP_ERR;
584 		break;
585 	case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
586 		wc->status = IB_WC_LOC_ACCESS_ERR;
587 		break;
588 	case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
589 		wc->status = IB_WC_REM_INV_REQ_ERR;
590 		break;
591 	case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
592 		wc->status = IB_WC_REM_ACCESS_ERR;
593 		break;
594 	case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
595 		wc->status = IB_WC_REM_OP_ERR;
596 		break;
597 	case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
598 		wc->status = IB_WC_RETRY_EXC_ERR;
599 		break;
600 	case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
601 		wc->status = IB_WC_RNR_RETRY_EXC_ERR;
602 		break;
603 	case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
604 		wc->status = IB_WC_REM_ABORT_ERR;
605 		break;
606 	default:
607 		wc->status = IB_WC_GENERAL_ERR;
608 		break;
609 	}
610 
611 	wc->vendor_err = cqe->vendor_err_syndrome;
612 }
613 
614 static int mlx4_ib_ipoib_csum_ok(__be16 status, u8 badfcs_enc, __be16 checksum)
615 {
616 	return ((badfcs_enc & MLX4_CQE_STATUS_L4_CSUM) ||
617 		((status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
618 		 (status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
619 				       MLX4_CQE_STATUS_UDP)) &&
620 		 (checksum == cpu_to_be16(0xffff))));
621 }
622 
623 static void use_tunnel_data(struct mlx4_ib_qp *qp, struct mlx4_ib_cq *cq, struct ib_wc *wc,
624 			    unsigned tail, struct mlx4_cqe *cqe, int is_eth)
625 {
626 	struct mlx4_ib_proxy_sqp_hdr *hdr;
627 
628 	ib_dma_sync_single_for_cpu(qp->ibqp.device,
629 				   qp->sqp_proxy_rcv[tail].map,
630 				   sizeof (struct mlx4_ib_proxy_sqp_hdr),
631 				   DMA_FROM_DEVICE);
632 	hdr = (struct mlx4_ib_proxy_sqp_hdr *) (qp->sqp_proxy_rcv[tail].addr);
633 	wc->pkey_index	= be16_to_cpu(hdr->tun.pkey_index);
634 	wc->src_qp	= be32_to_cpu(hdr->tun.flags_src_qp) & 0xFFFFFF;
635 	wc->wc_flags   |= (hdr->tun.g_ml_path & 0x80) ? (IB_WC_GRH) : 0;
636 	wc->dlid_path_bits = 0;
637 
638 	if (is_eth) {
639 		wc->slid = 0;
640 		wc->vlan_id = be16_to_cpu(hdr->tun.sl_vid);
641 		memcpy(&(wc->smac[0]), (char *)&hdr->tun.mac_31_0, 4);
642 		memcpy(&(wc->smac[4]), (char *)&hdr->tun.slid_mac_47_32, 2);
643 		wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
644 	} else {
645 		wc->slid        = be16_to_cpu(hdr->tun.slid_mac_47_32);
646 		wc->sl          = (u8) (be16_to_cpu(hdr->tun.sl_vid) >> 12);
647 	}
648 }
649 
650 static void mlx4_ib_qp_sw_comp(struct mlx4_ib_qp *qp, int num_entries,
651 			       struct ib_wc *wc, int *npolled, int is_send)
652 {
653 	struct mlx4_ib_wq *wq;
654 	unsigned cur;
655 	int i;
656 
657 	wq = is_send ? &qp->sq : &qp->rq;
658 	cur = wq->head - wq->tail;
659 
660 	if (cur == 0)
661 		return;
662 
663 	for (i = 0;  i < cur && *npolled < num_entries; i++) {
664 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
665 		wc->status = IB_WC_WR_FLUSH_ERR;
666 		wc->vendor_err = MLX4_CQE_SYNDROME_WR_FLUSH_ERR;
667 		wq->tail++;
668 		(*npolled)++;
669 		wc->qp = &qp->ibqp;
670 		wc++;
671 	}
672 }
673 
674 static void mlx4_ib_poll_sw_comp(struct mlx4_ib_cq *cq, int num_entries,
675 				 struct ib_wc *wc, int *npolled)
676 {
677 	struct mlx4_ib_qp *qp;
678 
679 	*npolled = 0;
680 	/* Find uncompleted WQEs belonging to that cq and return
681 	 * simulated FLUSH_ERR completions
682 	 */
683 	list_for_each_entry(qp, &cq->send_qp_list, cq_send_list) {
684 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 1);
685 		if (*npolled >= num_entries)
686 			goto out;
687 	}
688 
689 	list_for_each_entry(qp, &cq->recv_qp_list, cq_recv_list) {
690 		mlx4_ib_qp_sw_comp(qp, num_entries, wc + *npolled, npolled, 0);
691 		if (*npolled >= num_entries)
692 			goto out;
693 	}
694 
695 out:
696 	return;
697 }
698 
699 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
700 			    struct mlx4_ib_qp **cur_qp,
701 			    struct ib_wc *wc)
702 {
703 	struct mlx4_cqe *cqe;
704 	struct mlx4_qp *mqp;
705 	struct mlx4_ib_wq *wq;
706 	struct mlx4_ib_srq *srq;
707 	struct mlx4_srq *msrq = NULL;
708 	int is_send;
709 	int is_error;
710 	int is_eth;
711 	u32 g_mlpath_rqpn;
712 	u16 wqe_ctr;
713 	unsigned tail = 0;
714 
715 repoll:
716 	cqe = next_cqe_sw(cq);
717 	if (!cqe)
718 		return -EAGAIN;
719 
720 	if (cq->buf.entry_size == 64)
721 		cqe++;
722 
723 	++cq->mcq.cons_index;
724 
725 	/*
726 	 * Make sure we read CQ entry contents after we've checked the
727 	 * ownership bit.
728 	 */
729 	rmb();
730 
731 	is_send  = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
732 	is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
733 		MLX4_CQE_OPCODE_ERROR;
734 
735 	/* Resize CQ in progress */
736 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
737 		if (cq->resize_buf) {
738 			struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
739 
740 			mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
741 			cq->buf      = cq->resize_buf->buf;
742 			cq->ibcq.cqe = cq->resize_buf->cqe;
743 
744 			kfree(cq->resize_buf);
745 			cq->resize_buf = NULL;
746 		}
747 
748 		goto repoll;
749 	}
750 
751 	if (!*cur_qp ||
752 	    (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
753 		/*
754 		 * We do not have to take the QP table lock here,
755 		 * because CQs will be locked while QPs are removed
756 		 * from the table.
757 		 */
758 		mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
759 				       be32_to_cpu(cqe->vlan_my_qpn));
760 		*cur_qp = to_mibqp(mqp);
761 	}
762 
763 	wc->qp = &(*cur_qp)->ibqp;
764 
765 	if (wc->qp->qp_type == IB_QPT_XRC_TGT) {
766 		u32 srq_num;
767 		g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
768 		srq_num       = g_mlpath_rqpn & 0xffffff;
769 		/* SRQ is also in the radix tree */
770 		msrq = mlx4_srq_lookup(to_mdev(cq->ibcq.device)->dev,
771 				       srq_num);
772 	}
773 
774 	if (is_send) {
775 		wq = &(*cur_qp)->sq;
776 		if (!(*cur_qp)->sq_signal_bits) {
777 			wqe_ctr = be16_to_cpu(cqe->wqe_index);
778 			wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
779 		}
780 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
781 		++wq->tail;
782 	} else if ((*cur_qp)->ibqp.srq) {
783 		srq = to_msrq((*cur_qp)->ibqp.srq);
784 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
785 		wc->wr_id = srq->wrid[wqe_ctr];
786 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
787 	} else if (msrq) {
788 		srq = to_mibsrq(msrq);
789 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
790 		wc->wr_id = srq->wrid[wqe_ctr];
791 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
792 	} else {
793 		wq	  = &(*cur_qp)->rq;
794 		tail	  = wq->tail & (wq->wqe_cnt - 1);
795 		wc->wr_id = wq->wrid[tail];
796 		++wq->tail;
797 	}
798 
799 	if (unlikely(is_error)) {
800 		mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
801 		return 0;
802 	}
803 
804 	wc->status = IB_WC_SUCCESS;
805 
806 	if (is_send) {
807 		wc->wc_flags = 0;
808 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
809 		case MLX4_OPCODE_RDMA_WRITE_IMM:
810 			wc->wc_flags |= IB_WC_WITH_IMM;
811 			fallthrough;
812 		case MLX4_OPCODE_RDMA_WRITE:
813 			wc->opcode    = IB_WC_RDMA_WRITE;
814 			break;
815 		case MLX4_OPCODE_SEND_IMM:
816 			wc->wc_flags |= IB_WC_WITH_IMM;
817 			fallthrough;
818 		case MLX4_OPCODE_SEND:
819 		case MLX4_OPCODE_SEND_INVAL:
820 			wc->opcode    = IB_WC_SEND;
821 			break;
822 		case MLX4_OPCODE_RDMA_READ:
823 			wc->opcode    = IB_WC_RDMA_READ;
824 			wc->byte_len  = be32_to_cpu(cqe->byte_cnt);
825 			break;
826 		case MLX4_OPCODE_ATOMIC_CS:
827 			wc->opcode    = IB_WC_COMP_SWAP;
828 			wc->byte_len  = 8;
829 			break;
830 		case MLX4_OPCODE_ATOMIC_FA:
831 			wc->opcode    = IB_WC_FETCH_ADD;
832 			wc->byte_len  = 8;
833 			break;
834 		case MLX4_OPCODE_MASKED_ATOMIC_CS:
835 			wc->opcode    = IB_WC_MASKED_COMP_SWAP;
836 			wc->byte_len  = 8;
837 			break;
838 		case MLX4_OPCODE_MASKED_ATOMIC_FA:
839 			wc->opcode    = IB_WC_MASKED_FETCH_ADD;
840 			wc->byte_len  = 8;
841 			break;
842 		case MLX4_OPCODE_LSO:
843 			wc->opcode    = IB_WC_LSO;
844 			break;
845 		case MLX4_OPCODE_FMR:
846 			wc->opcode    = IB_WC_REG_MR;
847 			break;
848 		case MLX4_OPCODE_LOCAL_INVAL:
849 			wc->opcode    = IB_WC_LOCAL_INV;
850 			break;
851 		}
852 	} else {
853 		wc->byte_len = be32_to_cpu(cqe->byte_cnt);
854 
855 		switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
856 		case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
857 			wc->opcode	= IB_WC_RECV_RDMA_WITH_IMM;
858 			wc->wc_flags	= IB_WC_WITH_IMM;
859 			wc->ex.imm_data = cqe->immed_rss_invalid;
860 			break;
861 		case MLX4_RECV_OPCODE_SEND_INVAL:
862 			wc->opcode	= IB_WC_RECV;
863 			wc->wc_flags	= IB_WC_WITH_INVALIDATE;
864 			wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
865 			break;
866 		case MLX4_RECV_OPCODE_SEND:
867 			wc->opcode   = IB_WC_RECV;
868 			wc->wc_flags = 0;
869 			break;
870 		case MLX4_RECV_OPCODE_SEND_IMM:
871 			wc->opcode	= IB_WC_RECV;
872 			wc->wc_flags	= IB_WC_WITH_IMM;
873 			wc->ex.imm_data = cqe->immed_rss_invalid;
874 			break;
875 		}
876 
877 		is_eth = (rdma_port_get_link_layer(wc->qp->device,
878 						  (*cur_qp)->port) ==
879 			  IB_LINK_LAYER_ETHERNET);
880 		if (mlx4_is_mfunc(to_mdev(cq->ibcq.device)->dev)) {
881 			if ((*cur_qp)->mlx4_ib_qp_type &
882 			    (MLX4_IB_QPT_PROXY_SMI_OWNER |
883 			     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
884 				use_tunnel_data(*cur_qp, cq, wc, tail, cqe,
885 						is_eth);
886 				return 0;
887 			}
888 		}
889 
890 		g_mlpath_rqpn	   = be32_to_cpu(cqe->g_mlpath_rqpn);
891 		wc->src_qp	   = g_mlpath_rqpn & 0xffffff;
892 		wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
893 		wc->wc_flags	  |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
894 		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
895 		wc->wc_flags	  |= mlx4_ib_ipoib_csum_ok(cqe->status,
896 					cqe->badfcs_enc,
897 					cqe->checksum) ? IB_WC_IP_CSUM_OK : 0;
898 		if (is_eth) {
899 			wc->slid = 0;
900 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 13;
901 			if (be32_to_cpu(cqe->vlan_my_qpn) &
902 					MLX4_CQE_CVLAN_PRESENT_MASK) {
903 				wc->vlan_id = be16_to_cpu(cqe->sl_vid) &
904 					MLX4_CQE_VID_MASK;
905 			} else {
906 				wc->vlan_id = 0xffff;
907 			}
908 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
909 			wc->wc_flags |= (IB_WC_WITH_VLAN | IB_WC_WITH_SMAC);
910 		} else {
911 			wc->slid = be16_to_cpu(cqe->rlid);
912 			wc->sl  = be16_to_cpu(cqe->sl_vid) >> 12;
913 			wc->vlan_id = 0xffff;
914 		}
915 	}
916 
917 	return 0;
918 }
919 
920 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
921 {
922 	struct mlx4_ib_cq *cq = to_mcq(ibcq);
923 	struct mlx4_ib_qp *cur_qp = NULL;
924 	unsigned long flags;
925 	int npolled;
926 	struct mlx4_ib_dev *mdev = to_mdev(cq->ibcq.device);
927 
928 	spin_lock_irqsave(&cq->lock, flags);
929 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
930 		mlx4_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
931 		goto out;
932 	}
933 
934 	for (npolled = 0; npolled < num_entries; ++npolled) {
935 		if (mlx4_ib_poll_one(cq, &cur_qp, wc + npolled))
936 			break;
937 	}
938 
939 	mlx4_cq_set_ci(&cq->mcq);
940 
941 out:
942 	spin_unlock_irqrestore(&cq->lock, flags);
943 
944 	return npolled;
945 }
946 
947 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
948 {
949 	mlx4_cq_arm(&to_mcq(ibcq)->mcq,
950 		    (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
951 		    MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
952 		    to_mdev(ibcq->device)->uar_map,
953 		    MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
954 
955 	return 0;
956 }
957 
958 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
959 {
960 	u32 prod_index;
961 	int nfreed = 0;
962 	struct mlx4_cqe *cqe, *dest;
963 	u8 owner_bit;
964 	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
965 
966 	/*
967 	 * First we need to find the current producer index, so we
968 	 * know where to start cleaning from.  It doesn't matter if HW
969 	 * adds new entries after this loop -- the QP we're worried
970 	 * about is already in RESET, so the new entries won't come
971 	 * from our QP and therefore don't need to be checked.
972 	 */
973 	for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
974 		if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
975 			break;
976 
977 	/*
978 	 * Now sweep backwards through the CQ, removing CQ entries
979 	 * that match our QP by copying older entries on top of them.
980 	 */
981 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
982 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
983 		cqe += cqe_inc;
984 
985 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
986 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
987 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
988 			++nfreed;
989 		} else if (nfreed) {
990 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
991 			dest += cqe_inc;
992 
993 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
994 			memcpy(dest, cqe, sizeof *cqe);
995 			dest->owner_sr_opcode = owner_bit |
996 				(dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
997 		}
998 	}
999 
1000 	if (nfreed) {
1001 		cq->mcq.cons_index += nfreed;
1002 		/*
1003 		 * Make sure update of buffer contents is done before
1004 		 * updating consumer index.
1005 		 */
1006 		wmb();
1007 		mlx4_cq_set_ci(&cq->mcq);
1008 	}
1009 }
1010 
1011 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
1012 {
1013 	spin_lock_irq(&cq->lock);
1014 	__mlx4_ib_cq_clean(cq, qpn, srq);
1015 	spin_unlock_irq(&cq->lock);
1016 }
1017