1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  */
37 
38 #include <linux/delay.h>
39 #include <linux/gfp.h>
40 
41 #include "c2.h"
42 #include "c2_vq.h"
43 #include "c2_status.h"
44 
45 #define C2_MAX_ORD_PER_QP 128
46 #define C2_MAX_IRD_PER_QP 128
47 
48 #define C2_HINT_MAKE(q_index, hint_count) (((q_index) << 16) | hint_count)
49 #define C2_HINT_GET_INDEX(hint) (((hint) & 0x7FFF0000) >> 16)
50 #define C2_HINT_GET_COUNT(hint) ((hint) & 0x0000FFFF)
51 
52 #define NO_SUPPORT -1
53 static const u8 c2_opcode[] = {
54 	[IB_WR_SEND] = C2_WR_TYPE_SEND,
55 	[IB_WR_SEND_WITH_IMM] = NO_SUPPORT,
56 	[IB_WR_RDMA_WRITE] = C2_WR_TYPE_RDMA_WRITE,
57 	[IB_WR_RDMA_WRITE_WITH_IMM] = NO_SUPPORT,
58 	[IB_WR_RDMA_READ] = C2_WR_TYPE_RDMA_READ,
59 	[IB_WR_ATOMIC_CMP_AND_SWP] = NO_SUPPORT,
60 	[IB_WR_ATOMIC_FETCH_AND_ADD] = NO_SUPPORT,
61 };
62 
to_c2_state(enum ib_qp_state ib_state)63 static int to_c2_state(enum ib_qp_state ib_state)
64 {
65 	switch (ib_state) {
66 	case IB_QPS_RESET:
67 		return C2_QP_STATE_IDLE;
68 	case IB_QPS_RTS:
69 		return C2_QP_STATE_RTS;
70 	case IB_QPS_SQD:
71 		return C2_QP_STATE_CLOSING;
72 	case IB_QPS_SQE:
73 		return C2_QP_STATE_CLOSING;
74 	case IB_QPS_ERR:
75 		return C2_QP_STATE_ERROR;
76 	default:
77 		return -1;
78 	}
79 }
80 
to_ib_state(enum c2_qp_state c2_state)81 static int to_ib_state(enum c2_qp_state c2_state)
82 {
83 	switch (c2_state) {
84 	case C2_QP_STATE_IDLE:
85 		return IB_QPS_RESET;
86 	case C2_QP_STATE_CONNECTING:
87 		return IB_QPS_RTR;
88 	case C2_QP_STATE_RTS:
89 		return IB_QPS_RTS;
90 	case C2_QP_STATE_CLOSING:
91 		return IB_QPS_SQD;
92 	case C2_QP_STATE_ERROR:
93 		return IB_QPS_ERR;
94 	case C2_QP_STATE_TERMINATE:
95 		return IB_QPS_SQE;
96 	default:
97 		return -1;
98 	}
99 }
100 
to_ib_state_str(int ib_state)101 static const char *to_ib_state_str(int ib_state)
102 {
103 	static const char *state_str[] = {
104 		"IB_QPS_RESET",
105 		"IB_QPS_INIT",
106 		"IB_QPS_RTR",
107 		"IB_QPS_RTS",
108 		"IB_QPS_SQD",
109 		"IB_QPS_SQE",
110 		"IB_QPS_ERR"
111 	};
112 	if (ib_state < IB_QPS_RESET ||
113 	    ib_state > IB_QPS_ERR)
114 		return "<invalid IB QP state>";
115 
116 	ib_state -= IB_QPS_RESET;
117 	return state_str[ib_state];
118 }
119 
c2_set_qp_state(struct c2_qp * qp,int c2_state)120 void c2_set_qp_state(struct c2_qp *qp, int c2_state)
121 {
122 	int new_state = to_ib_state(c2_state);
123 
124 	pr_debug("%s: qp[%p] state modify %s --> %s\n",
125 	       __func__,
126 		qp,
127 		to_ib_state_str(qp->state),
128 		to_ib_state_str(new_state));
129 	qp->state = new_state;
130 }
131 
132 #define C2_QP_NO_ATTR_CHANGE 0xFFFFFFFF
133 
c2_qp_modify(struct c2_dev * c2dev,struct c2_qp * qp,struct ib_qp_attr * attr,int attr_mask)134 int c2_qp_modify(struct c2_dev *c2dev, struct c2_qp *qp,
135 		 struct ib_qp_attr *attr, int attr_mask)
136 {
137 	struct c2wr_qp_modify_req wr;
138 	struct c2wr_qp_modify_rep *reply;
139 	struct c2_vq_req *vq_req;
140 	unsigned long flags;
141 	u8 next_state;
142 	int err;
143 
144 	pr_debug("%s:%d qp=%p, %s --> %s\n",
145 		__func__, __LINE__,
146 		qp,
147 		to_ib_state_str(qp->state),
148 		to_ib_state_str(attr->qp_state));
149 
150 	vq_req = vq_req_alloc(c2dev);
151 	if (!vq_req)
152 		return -ENOMEM;
153 
154 	c2_wr_set_id(&wr, CCWR_QP_MODIFY);
155 	wr.hdr.context = (unsigned long) vq_req;
156 	wr.rnic_handle = c2dev->adapter_handle;
157 	wr.qp_handle = qp->adapter_handle;
158 	wr.ord = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
159 	wr.ird = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
160 	wr.sq_depth = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
161 	wr.rq_depth = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
162 
163 	if (attr_mask & IB_QP_STATE) {
164 		/* Ensure the state is valid */
165 		if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR) {
166 			err = -EINVAL;
167 			goto bail0;
168 		}
169 
170 		wr.next_qp_state = cpu_to_be32(to_c2_state(attr->qp_state));
171 
172 		if (attr->qp_state == IB_QPS_ERR) {
173 			spin_lock_irqsave(&qp->lock, flags);
174 			if (qp->cm_id && qp->state == IB_QPS_RTS) {
175 				pr_debug("Generating CLOSE event for QP-->ERR, "
176 					"qp=%p, cm_id=%p\n",qp,qp->cm_id);
177 				/* Generate an CLOSE event */
178 				vq_req->cm_id = qp->cm_id;
179 				vq_req->event = IW_CM_EVENT_CLOSE;
180 			}
181 			spin_unlock_irqrestore(&qp->lock, flags);
182 		}
183 		next_state =  attr->qp_state;
184 
185 	} else if (attr_mask & IB_QP_CUR_STATE) {
186 
187 		if (attr->cur_qp_state != IB_QPS_RTR &&
188 		    attr->cur_qp_state != IB_QPS_RTS &&
189 		    attr->cur_qp_state != IB_QPS_SQD &&
190 		    attr->cur_qp_state != IB_QPS_SQE) {
191 			err = -EINVAL;
192 			goto bail0;
193 		} else
194 			wr.next_qp_state =
195 			    cpu_to_be32(to_c2_state(attr->cur_qp_state));
196 
197 		next_state = attr->cur_qp_state;
198 
199 	} else {
200 		err = 0;
201 		goto bail0;
202 	}
203 
204 	/* reference the request struct */
205 	vq_req_get(c2dev, vq_req);
206 
207 	err = vq_send_wr(c2dev, (union c2wr *) & wr);
208 	if (err) {
209 		vq_req_put(c2dev, vq_req);
210 		goto bail0;
211 	}
212 
213 	err = vq_wait_for_reply(c2dev, vq_req);
214 	if (err)
215 		goto bail0;
216 
217 	reply = (struct c2wr_qp_modify_rep *) (unsigned long) vq_req->reply_msg;
218 	if (!reply) {
219 		err = -ENOMEM;
220 		goto bail0;
221 	}
222 
223 	err = c2_errno(reply);
224 	if (!err)
225 		qp->state = next_state;
226 #ifdef DEBUG
227 	else
228 		pr_debug("%s: c2_errno=%d\n", __func__, err);
229 #endif
230 	/*
231 	 * If we're going to error and generating the event here, then
232 	 * we need to remove the reference because there will be no
233 	 * close event generated by the adapter
234 	*/
235 	spin_lock_irqsave(&qp->lock, flags);
236 	if (vq_req->event==IW_CM_EVENT_CLOSE && qp->cm_id) {
237 		qp->cm_id->rem_ref(qp->cm_id);
238 		qp->cm_id = NULL;
239 	}
240 	spin_unlock_irqrestore(&qp->lock, flags);
241 
242 	vq_repbuf_free(c2dev, reply);
243       bail0:
244 	vq_req_free(c2dev, vq_req);
245 
246 	pr_debug("%s:%d qp=%p, cur_state=%s\n",
247 		__func__, __LINE__,
248 		qp,
249 		to_ib_state_str(qp->state));
250 	return err;
251 }
252 
c2_qp_set_read_limits(struct c2_dev * c2dev,struct c2_qp * qp,int ord,int ird)253 int c2_qp_set_read_limits(struct c2_dev *c2dev, struct c2_qp *qp,
254 			  int ord, int ird)
255 {
256 	struct c2wr_qp_modify_req wr;
257 	struct c2wr_qp_modify_rep *reply;
258 	struct c2_vq_req *vq_req;
259 	int err;
260 
261 	vq_req = vq_req_alloc(c2dev);
262 	if (!vq_req)
263 		return -ENOMEM;
264 
265 	c2_wr_set_id(&wr, CCWR_QP_MODIFY);
266 	wr.hdr.context = (unsigned long) vq_req;
267 	wr.rnic_handle = c2dev->adapter_handle;
268 	wr.qp_handle = qp->adapter_handle;
269 	wr.ord = cpu_to_be32(ord);
270 	wr.ird = cpu_to_be32(ird);
271 	wr.sq_depth = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
272 	wr.rq_depth = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
273 	wr.next_qp_state = cpu_to_be32(C2_QP_NO_ATTR_CHANGE);
274 
275 	/* reference the request struct */
276 	vq_req_get(c2dev, vq_req);
277 
278 	err = vq_send_wr(c2dev, (union c2wr *) & wr);
279 	if (err) {
280 		vq_req_put(c2dev, vq_req);
281 		goto bail0;
282 	}
283 
284 	err = vq_wait_for_reply(c2dev, vq_req);
285 	if (err)
286 		goto bail0;
287 
288 	reply = (struct c2wr_qp_modify_rep *) (unsigned long)
289 		vq_req->reply_msg;
290 	if (!reply) {
291 		err = -ENOMEM;
292 		goto bail0;
293 	}
294 
295 	err = c2_errno(reply);
296 	vq_repbuf_free(c2dev, reply);
297       bail0:
298 	vq_req_free(c2dev, vq_req);
299 	return err;
300 }
301 
destroy_qp(struct c2_dev * c2dev,struct c2_qp * qp)302 static int destroy_qp(struct c2_dev *c2dev, struct c2_qp *qp)
303 {
304 	struct c2_vq_req *vq_req;
305 	struct c2wr_qp_destroy_req wr;
306 	struct c2wr_qp_destroy_rep *reply;
307 	unsigned long flags;
308 	int err;
309 
310 	/*
311 	 * Allocate a verb request message
312 	 */
313 	vq_req = vq_req_alloc(c2dev);
314 	if (!vq_req) {
315 		return -ENOMEM;
316 	}
317 
318 	/*
319 	 * Initialize the WR
320 	 */
321 	c2_wr_set_id(&wr, CCWR_QP_DESTROY);
322 	wr.hdr.context = (unsigned long) vq_req;
323 	wr.rnic_handle = c2dev->adapter_handle;
324 	wr.qp_handle = qp->adapter_handle;
325 
326 	/*
327 	 * reference the request struct.  dereferenced in the int handler.
328 	 */
329 	vq_req_get(c2dev, vq_req);
330 
331 	spin_lock_irqsave(&qp->lock, flags);
332 	if (qp->cm_id && qp->state == IB_QPS_RTS) {
333 		pr_debug("destroy_qp: generating CLOSE event for QP-->ERR, "
334 			"qp=%p, cm_id=%p\n",qp,qp->cm_id);
335 		/* Generate an CLOSE event */
336 		vq_req->qp = qp;
337 		vq_req->cm_id = qp->cm_id;
338 		vq_req->event = IW_CM_EVENT_CLOSE;
339 	}
340 	spin_unlock_irqrestore(&qp->lock, flags);
341 
342 	/*
343 	 * Send WR to adapter
344 	 */
345 	err = vq_send_wr(c2dev, (union c2wr *) & wr);
346 	if (err) {
347 		vq_req_put(c2dev, vq_req);
348 		goto bail0;
349 	}
350 
351 	/*
352 	 * Wait for reply from adapter
353 	 */
354 	err = vq_wait_for_reply(c2dev, vq_req);
355 	if (err) {
356 		goto bail0;
357 	}
358 
359 	/*
360 	 * Process reply
361 	 */
362 	reply = (struct c2wr_qp_destroy_rep *) (unsigned long) (vq_req->reply_msg);
363 	if (!reply) {
364 		err = -ENOMEM;
365 		goto bail0;
366 	}
367 
368 	spin_lock_irqsave(&qp->lock, flags);
369 	if (qp->cm_id) {
370 		qp->cm_id->rem_ref(qp->cm_id);
371 		qp->cm_id = NULL;
372 	}
373 	spin_unlock_irqrestore(&qp->lock, flags);
374 
375 	vq_repbuf_free(c2dev, reply);
376       bail0:
377 	vq_req_free(c2dev, vq_req);
378 	return err;
379 }
380 
c2_alloc_qpn(struct c2_dev * c2dev,struct c2_qp * qp)381 static int c2_alloc_qpn(struct c2_dev *c2dev, struct c2_qp *qp)
382 {
383 	int ret;
384 
385         do {
386 		spin_lock_irq(&c2dev->qp_table.lock);
387 		ret = idr_get_new_above(&c2dev->qp_table.idr, qp,
388 					c2dev->qp_table.last++, &qp->qpn);
389 		spin_unlock_irq(&c2dev->qp_table.lock);
390         } while ((ret == -EAGAIN) &&
391 	 	 idr_pre_get(&c2dev->qp_table.idr, GFP_KERNEL));
392 	return ret;
393 }
394 
c2_free_qpn(struct c2_dev * c2dev,int qpn)395 static void c2_free_qpn(struct c2_dev *c2dev, int qpn)
396 {
397 	spin_lock_irq(&c2dev->qp_table.lock);
398 	idr_remove(&c2dev->qp_table.idr, qpn);
399 	spin_unlock_irq(&c2dev->qp_table.lock);
400 }
401 
c2_find_qpn(struct c2_dev * c2dev,int qpn)402 struct c2_qp *c2_find_qpn(struct c2_dev *c2dev, int qpn)
403 {
404 	unsigned long flags;
405 	struct c2_qp *qp;
406 
407 	spin_lock_irqsave(&c2dev->qp_table.lock, flags);
408 	qp = idr_find(&c2dev->qp_table.idr, qpn);
409 	spin_unlock_irqrestore(&c2dev->qp_table.lock, flags);
410 	return qp;
411 }
412 
c2_alloc_qp(struct c2_dev * c2dev,struct c2_pd * pd,struct ib_qp_init_attr * qp_attrs,struct c2_qp * qp)413 int c2_alloc_qp(struct c2_dev *c2dev,
414 		struct c2_pd *pd,
415 		struct ib_qp_init_attr *qp_attrs, struct c2_qp *qp)
416 {
417 	struct c2wr_qp_create_req wr;
418 	struct c2wr_qp_create_rep *reply;
419 	struct c2_vq_req *vq_req;
420 	struct c2_cq *send_cq = to_c2cq(qp_attrs->send_cq);
421 	struct c2_cq *recv_cq = to_c2cq(qp_attrs->recv_cq);
422 	unsigned long peer_pa;
423 	u32 q_size, msg_size, mmap_size;
424 	void __iomem *mmap;
425 	int err;
426 
427 	err = c2_alloc_qpn(c2dev, qp);
428 	if (err)
429 		return err;
430 	qp->ibqp.qp_num = qp->qpn;
431 	qp->ibqp.qp_type = IB_QPT_RC;
432 
433 	/* Allocate the SQ and RQ shared pointers */
434 	qp->sq_mq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
435 					 &qp->sq_mq.shared_dma, GFP_KERNEL);
436 	if (!qp->sq_mq.shared) {
437 		err = -ENOMEM;
438 		goto bail0;
439 	}
440 
441 	qp->rq_mq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
442 					 &qp->rq_mq.shared_dma, GFP_KERNEL);
443 	if (!qp->rq_mq.shared) {
444 		err = -ENOMEM;
445 		goto bail1;
446 	}
447 
448 	/* Allocate the verbs request */
449 	vq_req = vq_req_alloc(c2dev);
450 	if (vq_req == NULL) {
451 		err = -ENOMEM;
452 		goto bail2;
453 	}
454 
455 	/* Initialize the work request */
456 	memset(&wr, 0, sizeof(wr));
457 	c2_wr_set_id(&wr, CCWR_QP_CREATE);
458 	wr.hdr.context = (unsigned long) vq_req;
459 	wr.rnic_handle = c2dev->adapter_handle;
460 	wr.sq_cq_handle = send_cq->adapter_handle;
461 	wr.rq_cq_handle = recv_cq->adapter_handle;
462 	wr.sq_depth = cpu_to_be32(qp_attrs->cap.max_send_wr + 1);
463 	wr.rq_depth = cpu_to_be32(qp_attrs->cap.max_recv_wr + 1);
464 	wr.srq_handle = 0;
465 	wr.flags = cpu_to_be32(QP_RDMA_READ | QP_RDMA_WRITE | QP_MW_BIND |
466 			       QP_ZERO_STAG | QP_RDMA_READ_RESPONSE);
467 	wr.send_sgl_depth = cpu_to_be32(qp_attrs->cap.max_send_sge);
468 	wr.recv_sgl_depth = cpu_to_be32(qp_attrs->cap.max_recv_sge);
469 	wr.rdma_write_sgl_depth = cpu_to_be32(qp_attrs->cap.max_send_sge);
470 	wr.shared_sq_ht = cpu_to_be64(qp->sq_mq.shared_dma);
471 	wr.shared_rq_ht = cpu_to_be64(qp->rq_mq.shared_dma);
472 	wr.ord = cpu_to_be32(C2_MAX_ORD_PER_QP);
473 	wr.ird = cpu_to_be32(C2_MAX_IRD_PER_QP);
474 	wr.pd_id = pd->pd_id;
475 	wr.user_context = (unsigned long) qp;
476 
477 	vq_req_get(c2dev, vq_req);
478 
479 	/* Send the WR to the adapter */
480 	err = vq_send_wr(c2dev, (union c2wr *) & wr);
481 	if (err) {
482 		vq_req_put(c2dev, vq_req);
483 		goto bail3;
484 	}
485 
486 	/* Wait for the verb reply  */
487 	err = vq_wait_for_reply(c2dev, vq_req);
488 	if (err) {
489 		goto bail3;
490 	}
491 
492 	/* Process the reply */
493 	reply = (struct c2wr_qp_create_rep *) (unsigned long) (vq_req->reply_msg);
494 	if (!reply) {
495 		err = -ENOMEM;
496 		goto bail3;
497 	}
498 
499 	if ((err = c2_wr_get_result(reply)) != 0) {
500 		goto bail4;
501 	}
502 
503 	/* Fill in the kernel QP struct */
504 	atomic_set(&qp->refcount, 1);
505 	qp->adapter_handle = reply->qp_handle;
506 	qp->state = IB_QPS_RESET;
507 	qp->send_sgl_depth = qp_attrs->cap.max_send_sge;
508 	qp->rdma_write_sgl_depth = qp_attrs->cap.max_send_sge;
509 	qp->recv_sgl_depth = qp_attrs->cap.max_recv_sge;
510 	init_waitqueue_head(&qp->wait);
511 
512 	/* Initialize the SQ MQ */
513 	q_size = be32_to_cpu(reply->sq_depth);
514 	msg_size = be32_to_cpu(reply->sq_msg_size);
515 	peer_pa = c2dev->pa + be32_to_cpu(reply->sq_mq_start);
516 	mmap_size = PAGE_ALIGN(sizeof(struct c2_mq_shared) + msg_size * q_size);
517 	mmap = ioremap_nocache(peer_pa, mmap_size);
518 	if (!mmap) {
519 		err = -ENOMEM;
520 		goto bail5;
521 	}
522 
523 	c2_mq_req_init(&qp->sq_mq,
524 		       be32_to_cpu(reply->sq_mq_index),
525 		       q_size,
526 		       msg_size,
527 		       mmap + sizeof(struct c2_mq_shared),	/* pool start */
528 		       mmap,				/* peer */
529 		       C2_MQ_ADAPTER_TARGET);
530 
531 	/* Initialize the RQ mq */
532 	q_size = be32_to_cpu(reply->rq_depth);
533 	msg_size = be32_to_cpu(reply->rq_msg_size);
534 	peer_pa = c2dev->pa + be32_to_cpu(reply->rq_mq_start);
535 	mmap_size = PAGE_ALIGN(sizeof(struct c2_mq_shared) + msg_size * q_size);
536 	mmap = ioremap_nocache(peer_pa, mmap_size);
537 	if (!mmap) {
538 		err = -ENOMEM;
539 		goto bail6;
540 	}
541 
542 	c2_mq_req_init(&qp->rq_mq,
543 		       be32_to_cpu(reply->rq_mq_index),
544 		       q_size,
545 		       msg_size,
546 		       mmap + sizeof(struct c2_mq_shared),	/* pool start */
547 		       mmap,				/* peer */
548 		       C2_MQ_ADAPTER_TARGET);
549 
550 	vq_repbuf_free(c2dev, reply);
551 	vq_req_free(c2dev, vq_req);
552 
553 	return 0;
554 
555       bail6:
556 	iounmap(qp->sq_mq.peer);
557       bail5:
558 	destroy_qp(c2dev, qp);
559       bail4:
560 	vq_repbuf_free(c2dev, reply);
561       bail3:
562 	vq_req_free(c2dev, vq_req);
563       bail2:
564 	c2_free_mqsp(qp->rq_mq.shared);
565       bail1:
566 	c2_free_mqsp(qp->sq_mq.shared);
567       bail0:
568 	c2_free_qpn(c2dev, qp->qpn);
569 	return err;
570 }
571 
c2_lock_cqs(struct c2_cq * send_cq,struct c2_cq * recv_cq)572 static inline void c2_lock_cqs(struct c2_cq *send_cq, struct c2_cq *recv_cq)
573 {
574 	if (send_cq == recv_cq)
575 		spin_lock_irq(&send_cq->lock);
576 	else if (send_cq > recv_cq) {
577 		spin_lock_irq(&send_cq->lock);
578 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
579 	} else {
580 		spin_lock_irq(&recv_cq->lock);
581 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
582 	}
583 }
584 
c2_unlock_cqs(struct c2_cq * send_cq,struct c2_cq * recv_cq)585 static inline void c2_unlock_cqs(struct c2_cq *send_cq, struct c2_cq *recv_cq)
586 {
587 	if (send_cq == recv_cq)
588 		spin_unlock_irq(&send_cq->lock);
589 	else if (send_cq > recv_cq) {
590 		spin_unlock(&recv_cq->lock);
591 		spin_unlock_irq(&send_cq->lock);
592 	} else {
593 		spin_unlock(&send_cq->lock);
594 		spin_unlock_irq(&recv_cq->lock);
595 	}
596 }
597 
c2_free_qp(struct c2_dev * c2dev,struct c2_qp * qp)598 void c2_free_qp(struct c2_dev *c2dev, struct c2_qp *qp)
599 {
600 	struct c2_cq *send_cq;
601 	struct c2_cq *recv_cq;
602 
603 	send_cq = to_c2cq(qp->ibqp.send_cq);
604 	recv_cq = to_c2cq(qp->ibqp.recv_cq);
605 
606 	/*
607 	 * Lock CQs here, so that CQ polling code can do QP lookup
608 	 * without taking a lock.
609 	 */
610 	c2_lock_cqs(send_cq, recv_cq);
611 	c2_free_qpn(c2dev, qp->qpn);
612 	c2_unlock_cqs(send_cq, recv_cq);
613 
614 	/*
615 	 * Destroy qp in the rnic...
616 	 */
617 	destroy_qp(c2dev, qp);
618 
619 	/*
620 	 * Mark any unreaped CQEs as null and void.
621 	 */
622 	c2_cq_clean(c2dev, qp, send_cq->cqn);
623 	if (send_cq != recv_cq)
624 		c2_cq_clean(c2dev, qp, recv_cq->cqn);
625 	/*
626 	 * Unmap the MQs and return the shared pointers
627 	 * to the message pool.
628 	 */
629 	iounmap(qp->sq_mq.peer);
630 	iounmap(qp->rq_mq.peer);
631 	c2_free_mqsp(qp->sq_mq.shared);
632 	c2_free_mqsp(qp->rq_mq.shared);
633 
634 	atomic_dec(&qp->refcount);
635 	wait_event(qp->wait, !atomic_read(&qp->refcount));
636 }
637 
638 /*
639  * Function: move_sgl
640  *
641  * Description:
642  * Move an SGL from the user's work request struct into a CCIL Work Request
643  * message, swapping to WR byte order and ensure the total length doesn't
644  * overflow.
645  *
646  * IN:
647  * dst		- ptr to CCIL Work Request message SGL memory.
648  * src		- ptr to the consumers SGL memory.
649  *
650  * OUT: none
651  *
652  * Return:
653  * CCIL status codes.
654  */
655 static int
move_sgl(struct c2_data_addr * dst,struct ib_sge * src,int count,u32 * p_len,u8 * actual_count)656 move_sgl(struct c2_data_addr * dst, struct ib_sge *src, int count, u32 * p_len,
657 	 u8 * actual_count)
658 {
659 	u32 tot = 0;		/* running total */
660 	u8 acount = 0;		/* running total non-0 len sge's */
661 
662 	while (count > 0) {
663 		/*
664 		 * If the addition of this SGE causes the
665 		 * total SGL length to exceed 2^32-1, then
666 		 * fail-n-bail.
667 		 *
668 		 * If the current total plus the next element length
669 		 * wraps, then it will go negative and be less than the
670 		 * current total...
671 		 */
672 		if ((tot + src->length) < tot) {
673 			return -EINVAL;
674 		}
675 		/*
676 		 * Bug: 1456 (as well as 1498 & 1643)
677 		 * Skip over any sge's supplied with len=0
678 		 */
679 		if (src->length) {
680 			tot += src->length;
681 			dst->stag = cpu_to_be32(src->lkey);
682 			dst->to = cpu_to_be64(src->addr);
683 			dst->length = cpu_to_be32(src->length);
684 			dst++;
685 			acount++;
686 		}
687 		src++;
688 		count--;
689 	}
690 
691 	if (acount == 0) {
692 		/*
693 		 * Bug: 1476 (as well as 1498, 1456 and 1643)
694 		 * Setup the SGL in the WR to make it easier for the RNIC.
695 		 * This way, the FW doesn't have to deal with special cases.
696 		 * Setting length=0 should be sufficient.
697 		 */
698 		dst->stag = 0;
699 		dst->to = 0;
700 		dst->length = 0;
701 	}
702 
703 	*p_len = tot;
704 	*actual_count = acount;
705 	return 0;
706 }
707 
708 /*
709  * Function: c2_activity (private function)
710  *
711  * Description:
712  * Post an mq index to the host->adapter activity fifo.
713  *
714  * IN:
715  * c2dev	- ptr to c2dev structure
716  * mq_index	- mq index to post
717  * shared	- value most recently written to shared
718  *
719  * OUT:
720  *
721  * Return:
722  * none
723  */
c2_activity(struct c2_dev * c2dev,u32 mq_index,u16 shared)724 static inline void c2_activity(struct c2_dev *c2dev, u32 mq_index, u16 shared)
725 {
726 	/*
727 	 * First read the register to see if the FIFO is full, and if so,
728 	 * spin until it's not.  This isn't perfect -- there is no
729 	 * synchronization among the clients of the register, but in
730 	 * practice it prevents multiple CPU from hammering the bus
731 	 * with PCI RETRY. Note that when this does happen, the card
732 	 * cannot get on the bus and the card and system hang in a
733 	 * deadlock -- thus the need for this code. [TOT]
734 	 */
735 	while (readl(c2dev->regs + PCI_BAR0_ADAPTER_HINT) & 0x80000000)
736 		udelay(10);
737 
738 	__raw_writel(C2_HINT_MAKE(mq_index, shared),
739 		     c2dev->regs + PCI_BAR0_ADAPTER_HINT);
740 }
741 
742 /*
743  * Function: qp_wr_post
744  *
745  * Description:
746  * This in-line function allocates a MQ msg, then moves the host-copy of
747  * the completed WR into msg.  Then it posts the message.
748  *
749  * IN:
750  * q		- ptr to user MQ.
751  * wr		- ptr to host-copy of the WR.
752  * qp		- ptr to user qp
753  * size		- Number of bytes to post.  Assumed to be divisible by 4.
754  *
755  * OUT: none
756  *
757  * Return:
758  * CCIL status codes.
759  */
qp_wr_post(struct c2_mq * q,union c2wr * wr,struct c2_qp * qp,u32 size)760 static int qp_wr_post(struct c2_mq *q, union c2wr * wr, struct c2_qp *qp, u32 size)
761 {
762 	union c2wr *msg;
763 
764 	msg = c2_mq_alloc(q);
765 	if (msg == NULL) {
766 		return -EINVAL;
767 	}
768 #ifdef CCMSGMAGIC
769 	((c2wr_hdr_t *) wr)->magic = cpu_to_be32(CCWR_MAGIC);
770 #endif
771 
772 	/*
773 	 * Since all header fields in the WR are the same as the
774 	 * CQE, set the following so the adapter need not.
775 	 */
776 	c2_wr_set_result(wr, CCERR_PENDING);
777 
778 	/*
779 	 * Copy the wr down to the adapter
780 	 */
781 	memcpy((void *) msg, (void *) wr, size);
782 
783 	c2_mq_produce(q);
784 	return 0;
785 }
786 
787 
c2_post_send(struct ib_qp * ibqp,struct ib_send_wr * ib_wr,struct ib_send_wr ** bad_wr)788 int c2_post_send(struct ib_qp *ibqp, struct ib_send_wr *ib_wr,
789 		 struct ib_send_wr **bad_wr)
790 {
791 	struct c2_dev *c2dev = to_c2dev(ibqp->device);
792 	struct c2_qp *qp = to_c2qp(ibqp);
793 	union c2wr wr;
794 	unsigned long lock_flags;
795 	int err = 0;
796 
797 	u32 flags;
798 	u32 tot_len;
799 	u8 actual_sge_count;
800 	u32 msg_size;
801 
802 	if (qp->state > IB_QPS_RTS) {
803 		err = -EINVAL;
804 		goto out;
805 	}
806 
807 	while (ib_wr) {
808 
809 		flags = 0;
810 		wr.sqwr.sq_hdr.user_hdr.hdr.context = ib_wr->wr_id;
811 		if (ib_wr->send_flags & IB_SEND_SIGNALED) {
812 			flags |= SQ_SIGNALED;
813 		}
814 
815 		switch (ib_wr->opcode) {
816 		case IB_WR_SEND:
817 		case IB_WR_SEND_WITH_INV:
818 			if (ib_wr->opcode == IB_WR_SEND) {
819 				if (ib_wr->send_flags & IB_SEND_SOLICITED)
820 					c2_wr_set_id(&wr, C2_WR_TYPE_SEND_SE);
821 				else
822 					c2_wr_set_id(&wr, C2_WR_TYPE_SEND);
823 				wr.sqwr.send.remote_stag = 0;
824 			} else {
825 				if (ib_wr->send_flags & IB_SEND_SOLICITED)
826 					c2_wr_set_id(&wr, C2_WR_TYPE_SEND_SE_INV);
827 				else
828 					c2_wr_set_id(&wr, C2_WR_TYPE_SEND_INV);
829 				wr.sqwr.send.remote_stag =
830 					cpu_to_be32(ib_wr->ex.invalidate_rkey);
831 			}
832 
833 			msg_size = sizeof(struct c2wr_send_req) +
834 				sizeof(struct c2_data_addr) * ib_wr->num_sge;
835 			if (ib_wr->num_sge > qp->send_sgl_depth) {
836 				err = -EINVAL;
837 				break;
838 			}
839 			if (ib_wr->send_flags & IB_SEND_FENCE) {
840 				flags |= SQ_READ_FENCE;
841 			}
842 			err = move_sgl((struct c2_data_addr *) & (wr.sqwr.send.data),
843 				       ib_wr->sg_list,
844 				       ib_wr->num_sge,
845 				       &tot_len, &actual_sge_count);
846 			wr.sqwr.send.sge_len = cpu_to_be32(tot_len);
847 			c2_wr_set_sge_count(&wr, actual_sge_count);
848 			break;
849 		case IB_WR_RDMA_WRITE:
850 			c2_wr_set_id(&wr, C2_WR_TYPE_RDMA_WRITE);
851 			msg_size = sizeof(struct c2wr_rdma_write_req) +
852 			    (sizeof(struct c2_data_addr) * ib_wr->num_sge);
853 			if (ib_wr->num_sge > qp->rdma_write_sgl_depth) {
854 				err = -EINVAL;
855 				break;
856 			}
857 			if (ib_wr->send_flags & IB_SEND_FENCE) {
858 				flags |= SQ_READ_FENCE;
859 			}
860 			wr.sqwr.rdma_write.remote_stag =
861 			    cpu_to_be32(ib_wr->wr.rdma.rkey);
862 			wr.sqwr.rdma_write.remote_to =
863 			    cpu_to_be64(ib_wr->wr.rdma.remote_addr);
864 			err = move_sgl((struct c2_data_addr *)
865 				       & (wr.sqwr.rdma_write.data),
866 				       ib_wr->sg_list,
867 				       ib_wr->num_sge,
868 				       &tot_len, &actual_sge_count);
869 			wr.sqwr.rdma_write.sge_len = cpu_to_be32(tot_len);
870 			c2_wr_set_sge_count(&wr, actual_sge_count);
871 			break;
872 		case IB_WR_RDMA_READ:
873 			c2_wr_set_id(&wr, C2_WR_TYPE_RDMA_READ);
874 			msg_size = sizeof(struct c2wr_rdma_read_req);
875 
876 			/* IWarp only suppots 1 sge for RDMA reads */
877 			if (ib_wr->num_sge > 1) {
878 				err = -EINVAL;
879 				break;
880 			}
881 
882 			/*
883 			 * Move the local and remote stag/to/len into the WR.
884 			 */
885 			wr.sqwr.rdma_read.local_stag =
886 			    cpu_to_be32(ib_wr->sg_list->lkey);
887 			wr.sqwr.rdma_read.local_to =
888 			    cpu_to_be64(ib_wr->sg_list->addr);
889 			wr.sqwr.rdma_read.remote_stag =
890 			    cpu_to_be32(ib_wr->wr.rdma.rkey);
891 			wr.sqwr.rdma_read.remote_to =
892 			    cpu_to_be64(ib_wr->wr.rdma.remote_addr);
893 			wr.sqwr.rdma_read.length =
894 			    cpu_to_be32(ib_wr->sg_list->length);
895 			break;
896 		default:
897 			/* error */
898 			msg_size = 0;
899 			err = -EINVAL;
900 			break;
901 		}
902 
903 		/*
904 		 * If we had an error on the last wr build, then
905 		 * break out.  Possible errors include bogus WR
906 		 * type, and a bogus SGL length...
907 		 */
908 		if (err) {
909 			break;
910 		}
911 
912 		/*
913 		 * Store flags
914 		 */
915 		c2_wr_set_flags(&wr, flags);
916 
917 		/*
918 		 * Post the puppy!
919 		 */
920 		spin_lock_irqsave(&qp->lock, lock_flags);
921 		err = qp_wr_post(&qp->sq_mq, &wr, qp, msg_size);
922 		if (err) {
923 			spin_unlock_irqrestore(&qp->lock, lock_flags);
924 			break;
925 		}
926 
927 		/*
928 		 * Enqueue mq index to activity FIFO.
929 		 */
930 		c2_activity(c2dev, qp->sq_mq.index, qp->sq_mq.hint_count);
931 		spin_unlock_irqrestore(&qp->lock, lock_flags);
932 
933 		ib_wr = ib_wr->next;
934 	}
935 
936 out:
937 	if (err)
938 		*bad_wr = ib_wr;
939 	return err;
940 }
941 
c2_post_receive(struct ib_qp * ibqp,struct ib_recv_wr * ib_wr,struct ib_recv_wr ** bad_wr)942 int c2_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *ib_wr,
943 		    struct ib_recv_wr **bad_wr)
944 {
945 	struct c2_dev *c2dev = to_c2dev(ibqp->device);
946 	struct c2_qp *qp = to_c2qp(ibqp);
947 	union c2wr wr;
948 	unsigned long lock_flags;
949 	int err = 0;
950 
951 	if (qp->state > IB_QPS_RTS) {
952 		err = -EINVAL;
953 		goto out;
954 	}
955 
956 	/*
957 	 * Try and post each work request
958 	 */
959 	while (ib_wr) {
960 		u32 tot_len;
961 		u8 actual_sge_count;
962 
963 		if (ib_wr->num_sge > qp->recv_sgl_depth) {
964 			err = -EINVAL;
965 			break;
966 		}
967 
968 		/*
969 		 * Create local host-copy of the WR
970 		 */
971 		wr.rqwr.rq_hdr.user_hdr.hdr.context = ib_wr->wr_id;
972 		c2_wr_set_id(&wr, CCWR_RECV);
973 		c2_wr_set_flags(&wr, 0);
974 
975 		/* sge_count is limited to eight bits. */
976 		BUG_ON(ib_wr->num_sge >= 256);
977 		err = move_sgl((struct c2_data_addr *) & (wr.rqwr.data),
978 			       ib_wr->sg_list,
979 			       ib_wr->num_sge, &tot_len, &actual_sge_count);
980 		c2_wr_set_sge_count(&wr, actual_sge_count);
981 
982 		/*
983 		 * If we had an error on the last wr build, then
984 		 * break out.  Possible errors include bogus WR
985 		 * type, and a bogus SGL length...
986 		 */
987 		if (err) {
988 			break;
989 		}
990 
991 		spin_lock_irqsave(&qp->lock, lock_flags);
992 		err = qp_wr_post(&qp->rq_mq, &wr, qp, qp->rq_mq.msg_size);
993 		if (err) {
994 			spin_unlock_irqrestore(&qp->lock, lock_flags);
995 			break;
996 		}
997 
998 		/*
999 		 * Enqueue mq index to activity FIFO
1000 		 */
1001 		c2_activity(c2dev, qp->rq_mq.index, qp->rq_mq.hint_count);
1002 		spin_unlock_irqrestore(&qp->lock, lock_flags);
1003 
1004 		ib_wr = ib_wr->next;
1005 	}
1006 
1007 out:
1008 	if (err)
1009 		*bad_wr = ib_wr;
1010 	return err;
1011 }
1012 
c2_init_qp_table(struct c2_dev * c2dev)1013 void __devinit c2_init_qp_table(struct c2_dev *c2dev)
1014 {
1015 	spin_lock_init(&c2dev->qp_table.lock);
1016 	idr_init(&c2dev->qp_table.idr);
1017 }
1018 
c2_cleanup_qp_table(struct c2_dev * c2dev)1019 void __devexit c2_cleanup_qp_table(struct c2_dev *c2dev)
1020 {
1021 	idr_destroy(&c2dev->qp_table.idr);
1022 }
1023