xref: /linux/drivers/infiniband/core/mad.c (revision 5e8d780d745c1619aba81fe7166c5a4b5cad2b84)
1 /*
2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
35  */
36 #include <linux/dma-mapping.h>
37 #include <rdma/ib_cache.h>
38 
39 #include "mad_priv.h"
40 #include "mad_rmpp.h"
41 #include "smi.h"
42 #include "agent.h"
43 
44 MODULE_LICENSE("Dual BSD/GPL");
45 MODULE_DESCRIPTION("kernel IB MAD API");
46 MODULE_AUTHOR("Hal Rosenstock");
47 MODULE_AUTHOR("Sean Hefty");
48 
49 static kmem_cache_t *ib_mad_cache;
50 
51 static struct list_head ib_mad_port_list;
52 static u32 ib_mad_client_id = 0;
53 
54 /* Port list lock */
55 static spinlock_t ib_mad_port_list_lock;
56 
57 
58 /* Forward declarations */
59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
60 			 struct ib_mad_reg_req *mad_reg_req);
61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
62 static struct ib_mad_agent_private *find_mad_agent(
63 					struct ib_mad_port_private *port_priv,
64 					struct ib_mad *mad);
65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
66 				    struct ib_mad_private *mad);
67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
68 static void timeout_sends(void *data);
69 static void local_completions(void *data);
70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
71 			      struct ib_mad_agent_private *agent_priv,
72 			      u8 mgmt_class);
73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
74 			   struct ib_mad_agent_private *agent_priv);
75 
76 /*
77  * Returns a ib_mad_port_private structure or NULL for a device/port
78  * Assumes ib_mad_port_list_lock is being held
79  */
80 static inline struct ib_mad_port_private *
81 __ib_get_mad_port(struct ib_device *device, int port_num)
82 {
83 	struct ib_mad_port_private *entry;
84 
85 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
86 		if (entry->device == device && entry->port_num == port_num)
87 			return entry;
88 	}
89 	return NULL;
90 }
91 
92 /*
93  * Wrapper function to return a ib_mad_port_private structure or NULL
94  * for a device/port
95  */
96 static inline struct ib_mad_port_private *
97 ib_get_mad_port(struct ib_device *device, int port_num)
98 {
99 	struct ib_mad_port_private *entry;
100 	unsigned long flags;
101 
102 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
103 	entry = __ib_get_mad_port(device, port_num);
104 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
105 
106 	return entry;
107 }
108 
109 static inline u8 convert_mgmt_class(u8 mgmt_class)
110 {
111 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
112 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
113 		0 : mgmt_class;
114 }
115 
116 static int get_spl_qp_index(enum ib_qp_type qp_type)
117 {
118 	switch (qp_type)
119 	{
120 	case IB_QPT_SMI:
121 		return 0;
122 	case IB_QPT_GSI:
123 		return 1;
124 	default:
125 		return -1;
126 	}
127 }
128 
129 static int vendor_class_index(u8 mgmt_class)
130 {
131 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
132 }
133 
134 static int is_vendor_class(u8 mgmt_class)
135 {
136 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
137 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
138 		return 0;
139 	return 1;
140 }
141 
142 static int is_vendor_oui(char *oui)
143 {
144 	if (oui[0] || oui[1] || oui[2])
145 		return 1;
146 	return 0;
147 }
148 
149 static int is_vendor_method_in_use(
150 		struct ib_mad_mgmt_vendor_class *vendor_class,
151 		struct ib_mad_reg_req *mad_reg_req)
152 {
153 	struct ib_mad_mgmt_method_table *method;
154 	int i;
155 
156 	for (i = 0; i < MAX_MGMT_OUI; i++) {
157 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
158 			method = vendor_class->method_table[i];
159 			if (method) {
160 				if (method_in_use(&method, mad_reg_req))
161 					return 1;
162 				else
163 					break;
164 			}
165 		}
166 	}
167 	return 0;
168 }
169 
170 /*
171  * ib_register_mad_agent - Register to send/receive MADs
172  */
173 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
174 					   u8 port_num,
175 					   enum ib_qp_type qp_type,
176 					   struct ib_mad_reg_req *mad_reg_req,
177 					   u8 rmpp_version,
178 					   ib_mad_send_handler send_handler,
179 					   ib_mad_recv_handler recv_handler,
180 					   void *context)
181 {
182 	struct ib_mad_port_private *port_priv;
183 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
184 	struct ib_mad_agent_private *mad_agent_priv;
185 	struct ib_mad_reg_req *reg_req = NULL;
186 	struct ib_mad_mgmt_class_table *class;
187 	struct ib_mad_mgmt_vendor_class_table *vendor;
188 	struct ib_mad_mgmt_vendor_class *vendor_class;
189 	struct ib_mad_mgmt_method_table *method;
190 	int ret2, qpn;
191 	unsigned long flags;
192 	u8 mgmt_class, vclass;
193 
194 	/* Validate parameters */
195 	qpn = get_spl_qp_index(qp_type);
196 	if (qpn == -1)
197 		goto error1;
198 
199 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
200 		goto error1;
201 
202 	/* Validate MAD registration request if supplied */
203 	if (mad_reg_req) {
204 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
205 			goto error1;
206 		if (!recv_handler)
207 			goto error1;
208 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
209 			/*
210 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
211 			 * one in this range currently allowed
212 			 */
213 			if (mad_reg_req->mgmt_class !=
214 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
215 				goto error1;
216 		} else if (mad_reg_req->mgmt_class == 0) {
217 			/*
218 			 * Class 0 is reserved in IBA and is used for
219 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
220 			 */
221 			goto error1;
222 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
223 			/*
224 			 * If class is in "new" vendor range,
225 			 * ensure supplied OUI is not zero
226 			 */
227 			if (!is_vendor_oui(mad_reg_req->oui))
228 				goto error1;
229 		}
230 		/* Make sure class supplied is consistent with RMPP */
231 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
232 			if (rmpp_version)
233 				goto error1;
234 		}
235 		/* Make sure class supplied is consistent with QP type */
236 		if (qp_type == IB_QPT_SMI) {
237 			if ((mad_reg_req->mgmt_class !=
238 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
239 			    (mad_reg_req->mgmt_class !=
240 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
241 				goto error1;
242 		} else {
243 			if ((mad_reg_req->mgmt_class ==
244 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
245 			    (mad_reg_req->mgmt_class ==
246 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
247 				goto error1;
248 		}
249 	} else {
250 		/* No registration request supplied */
251 		if (!send_handler)
252 			goto error1;
253 	}
254 
255 	/* Validate device and port */
256 	port_priv = ib_get_mad_port(device, port_num);
257 	if (!port_priv) {
258 		ret = ERR_PTR(-ENODEV);
259 		goto error1;
260 	}
261 
262 	/* Allocate structures */
263 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
264 	if (!mad_agent_priv) {
265 		ret = ERR_PTR(-ENOMEM);
266 		goto error1;
267 	}
268 
269 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
270 						 IB_ACCESS_LOCAL_WRITE);
271 	if (IS_ERR(mad_agent_priv->agent.mr)) {
272 		ret = ERR_PTR(-ENOMEM);
273 		goto error2;
274 	}
275 
276 	if (mad_reg_req) {
277 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
278 		if (!reg_req) {
279 			ret = ERR_PTR(-ENOMEM);
280 			goto error3;
281 		}
282 		/* Make a copy of the MAD registration request */
283 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
284 	}
285 
286 	/* Now, fill in the various structures */
287 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
288 	mad_agent_priv->reg_req = reg_req;
289 	mad_agent_priv->agent.rmpp_version = rmpp_version;
290 	mad_agent_priv->agent.device = device;
291 	mad_agent_priv->agent.recv_handler = recv_handler;
292 	mad_agent_priv->agent.send_handler = send_handler;
293 	mad_agent_priv->agent.context = context;
294 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
295 	mad_agent_priv->agent.port_num = port_num;
296 
297 	spin_lock_irqsave(&port_priv->reg_lock, flags);
298 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
299 
300 	/*
301 	 * Make sure MAD registration (if supplied)
302 	 * is non overlapping with any existing ones
303 	 */
304 	if (mad_reg_req) {
305 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
306 		if (!is_vendor_class(mgmt_class)) {
307 			class = port_priv->version[mad_reg_req->
308 						   mgmt_class_version].class;
309 			if (class) {
310 				method = class->method_table[mgmt_class];
311 				if (method) {
312 					if (method_in_use(&method,
313 							   mad_reg_req))
314 						goto error4;
315 				}
316 			}
317 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
318 						  mgmt_class);
319 		} else {
320 			/* "New" vendor class range */
321 			vendor = port_priv->version[mad_reg_req->
322 						    mgmt_class_version].vendor;
323 			if (vendor) {
324 				vclass = vendor_class_index(mgmt_class);
325 				vendor_class = vendor->vendor_class[vclass];
326 				if (vendor_class) {
327 					if (is_vendor_method_in_use(
328 							vendor_class,
329 							mad_reg_req))
330 						goto error4;
331 				}
332 			}
333 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
334 		}
335 		if (ret2) {
336 			ret = ERR_PTR(ret2);
337 			goto error4;
338 		}
339 	}
340 
341 	/* Add mad agent into port's agent list */
342 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
343 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
344 
345 	spin_lock_init(&mad_agent_priv->lock);
346 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
347 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
348 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
349 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
350 	INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
351 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
352 	INIT_WORK(&mad_agent_priv->local_work, local_completions,
353 		   mad_agent_priv);
354 	atomic_set(&mad_agent_priv->refcount, 1);
355 	init_completion(&mad_agent_priv->comp);
356 
357 	return &mad_agent_priv->agent;
358 
359 error4:
360 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
361 	kfree(reg_req);
362 error3:
363 	ib_dereg_mr(mad_agent_priv->agent.mr);
364 error2:
365 	kfree(mad_agent_priv);
366 error1:
367 	return ret;
368 }
369 EXPORT_SYMBOL(ib_register_mad_agent);
370 
371 static inline int is_snooping_sends(int mad_snoop_flags)
372 {
373 	return (mad_snoop_flags &
374 		(/*IB_MAD_SNOOP_POSTED_SENDS |
375 		 IB_MAD_SNOOP_RMPP_SENDS |*/
376 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
377 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
378 }
379 
380 static inline int is_snooping_recvs(int mad_snoop_flags)
381 {
382 	return (mad_snoop_flags &
383 		(IB_MAD_SNOOP_RECVS /*|
384 		 IB_MAD_SNOOP_RMPP_RECVS*/));
385 }
386 
387 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
388 				struct ib_mad_snoop_private *mad_snoop_priv)
389 {
390 	struct ib_mad_snoop_private **new_snoop_table;
391 	unsigned long flags;
392 	int i;
393 
394 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
395 	/* Check for empty slot in array. */
396 	for (i = 0; i < qp_info->snoop_table_size; i++)
397 		if (!qp_info->snoop_table[i])
398 			break;
399 
400 	if (i == qp_info->snoop_table_size) {
401 		/* Grow table. */
402 		new_snoop_table = kmalloc(sizeof mad_snoop_priv *
403 					  qp_info->snoop_table_size + 1,
404 					  GFP_ATOMIC);
405 		if (!new_snoop_table) {
406 			i = -ENOMEM;
407 			goto out;
408 		}
409 		if (qp_info->snoop_table) {
410 			memcpy(new_snoop_table, qp_info->snoop_table,
411 			       sizeof mad_snoop_priv *
412 			       qp_info->snoop_table_size);
413 			kfree(qp_info->snoop_table);
414 		}
415 		qp_info->snoop_table = new_snoop_table;
416 		qp_info->snoop_table_size++;
417 	}
418 	qp_info->snoop_table[i] = mad_snoop_priv;
419 	atomic_inc(&qp_info->snoop_count);
420 out:
421 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
422 	return i;
423 }
424 
425 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
426 					   u8 port_num,
427 					   enum ib_qp_type qp_type,
428 					   int mad_snoop_flags,
429 					   ib_mad_snoop_handler snoop_handler,
430 					   ib_mad_recv_handler recv_handler,
431 					   void *context)
432 {
433 	struct ib_mad_port_private *port_priv;
434 	struct ib_mad_agent *ret;
435 	struct ib_mad_snoop_private *mad_snoop_priv;
436 	int qpn;
437 
438 	/* Validate parameters */
439 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
440 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
441 		ret = ERR_PTR(-EINVAL);
442 		goto error1;
443 	}
444 	qpn = get_spl_qp_index(qp_type);
445 	if (qpn == -1) {
446 		ret = ERR_PTR(-EINVAL);
447 		goto error1;
448 	}
449 	port_priv = ib_get_mad_port(device, port_num);
450 	if (!port_priv) {
451 		ret = ERR_PTR(-ENODEV);
452 		goto error1;
453 	}
454 	/* Allocate structures */
455 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
456 	if (!mad_snoop_priv) {
457 		ret = ERR_PTR(-ENOMEM);
458 		goto error1;
459 	}
460 
461 	/* Now, fill in the various structures */
462 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
463 	mad_snoop_priv->agent.device = device;
464 	mad_snoop_priv->agent.recv_handler = recv_handler;
465 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
466 	mad_snoop_priv->agent.context = context;
467 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
468 	mad_snoop_priv->agent.port_num = port_num;
469 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
470 	init_completion(&mad_snoop_priv->comp);
471 	mad_snoop_priv->snoop_index = register_snoop_agent(
472 						&port_priv->qp_info[qpn],
473 						mad_snoop_priv);
474 	if (mad_snoop_priv->snoop_index < 0) {
475 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
476 		goto error2;
477 	}
478 
479 	atomic_set(&mad_snoop_priv->refcount, 1);
480 	return &mad_snoop_priv->agent;
481 
482 error2:
483 	kfree(mad_snoop_priv);
484 error1:
485 	return ret;
486 }
487 EXPORT_SYMBOL(ib_register_mad_snoop);
488 
489 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
490 {
491 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
492 		complete(&mad_agent_priv->comp);
493 }
494 
495 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
496 {
497 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
498 		complete(&mad_snoop_priv->comp);
499 }
500 
501 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
502 {
503 	struct ib_mad_port_private *port_priv;
504 	unsigned long flags;
505 
506 	/* Note that we could still be handling received MADs */
507 
508 	/*
509 	 * Canceling all sends results in dropping received response
510 	 * MADs, preventing us from queuing additional work
511 	 */
512 	cancel_mads(mad_agent_priv);
513 	port_priv = mad_agent_priv->qp_info->port_priv;
514 	cancel_delayed_work(&mad_agent_priv->timed_work);
515 
516 	spin_lock_irqsave(&port_priv->reg_lock, flags);
517 	remove_mad_reg_req(mad_agent_priv);
518 	list_del(&mad_agent_priv->agent_list);
519 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
520 
521 	flush_workqueue(port_priv->wq);
522 	ib_cancel_rmpp_recvs(mad_agent_priv);
523 
524 	deref_mad_agent(mad_agent_priv);
525 	wait_for_completion(&mad_agent_priv->comp);
526 
527 	kfree(mad_agent_priv->reg_req);
528 	ib_dereg_mr(mad_agent_priv->agent.mr);
529 	kfree(mad_agent_priv);
530 }
531 
532 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
533 {
534 	struct ib_mad_qp_info *qp_info;
535 	unsigned long flags;
536 
537 	qp_info = mad_snoop_priv->qp_info;
538 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
539 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
540 	atomic_dec(&qp_info->snoop_count);
541 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
542 
543 	deref_snoop_agent(mad_snoop_priv);
544 	wait_for_completion(&mad_snoop_priv->comp);
545 
546 	kfree(mad_snoop_priv);
547 }
548 
549 /*
550  * ib_unregister_mad_agent - Unregisters a client from using MAD services
551  */
552 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
553 {
554 	struct ib_mad_agent_private *mad_agent_priv;
555 	struct ib_mad_snoop_private *mad_snoop_priv;
556 
557 	/* If the TID is zero, the agent can only snoop. */
558 	if (mad_agent->hi_tid) {
559 		mad_agent_priv = container_of(mad_agent,
560 					      struct ib_mad_agent_private,
561 					      agent);
562 		unregister_mad_agent(mad_agent_priv);
563 	} else {
564 		mad_snoop_priv = container_of(mad_agent,
565 					      struct ib_mad_snoop_private,
566 					      agent);
567 		unregister_mad_snoop(mad_snoop_priv);
568 	}
569 	return 0;
570 }
571 EXPORT_SYMBOL(ib_unregister_mad_agent);
572 
573 static inline int response_mad(struct ib_mad *mad)
574 {
575 	/* Trap represses are responses although response bit is reset */
576 	return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
577 		(mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
578 }
579 
580 static void dequeue_mad(struct ib_mad_list_head *mad_list)
581 {
582 	struct ib_mad_queue *mad_queue;
583 	unsigned long flags;
584 
585 	BUG_ON(!mad_list->mad_queue);
586 	mad_queue = mad_list->mad_queue;
587 	spin_lock_irqsave(&mad_queue->lock, flags);
588 	list_del(&mad_list->list);
589 	mad_queue->count--;
590 	spin_unlock_irqrestore(&mad_queue->lock, flags);
591 }
592 
593 static void snoop_send(struct ib_mad_qp_info *qp_info,
594 		       struct ib_mad_send_buf *send_buf,
595 		       struct ib_mad_send_wc *mad_send_wc,
596 		       int mad_snoop_flags)
597 {
598 	struct ib_mad_snoop_private *mad_snoop_priv;
599 	unsigned long flags;
600 	int i;
601 
602 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
603 	for (i = 0; i < qp_info->snoop_table_size; i++) {
604 		mad_snoop_priv = qp_info->snoop_table[i];
605 		if (!mad_snoop_priv ||
606 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
607 			continue;
608 
609 		atomic_inc(&mad_snoop_priv->refcount);
610 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
611 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
612 						    send_buf, mad_send_wc);
613 		deref_snoop_agent(mad_snoop_priv);
614 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
615 	}
616 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
617 }
618 
619 static void snoop_recv(struct ib_mad_qp_info *qp_info,
620 		       struct ib_mad_recv_wc *mad_recv_wc,
621 		       int mad_snoop_flags)
622 {
623 	struct ib_mad_snoop_private *mad_snoop_priv;
624 	unsigned long flags;
625 	int i;
626 
627 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
628 	for (i = 0; i < qp_info->snoop_table_size; i++) {
629 		mad_snoop_priv = qp_info->snoop_table[i];
630 		if (!mad_snoop_priv ||
631 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
632 			continue;
633 
634 		atomic_inc(&mad_snoop_priv->refcount);
635 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
636 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
637 						   mad_recv_wc);
638 		deref_snoop_agent(mad_snoop_priv);
639 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
640 	}
641 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
642 }
643 
644 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
645 			 struct ib_wc *wc)
646 {
647 	memset(wc, 0, sizeof *wc);
648 	wc->wr_id = wr_id;
649 	wc->status = IB_WC_SUCCESS;
650 	wc->opcode = IB_WC_RECV;
651 	wc->pkey_index = pkey_index;
652 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
653 	wc->src_qp = IB_QP0;
654 	wc->qp_num = IB_QP0;
655 	wc->slid = slid;
656 	wc->sl = 0;
657 	wc->dlid_path_bits = 0;
658 	wc->port_num = port_num;
659 }
660 
661 /*
662  * Return 0 if SMP is to be sent
663  * Return 1 if SMP was consumed locally (whether or not solicited)
664  * Return < 0 if error
665  */
666 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
667 				  struct ib_mad_send_wr_private *mad_send_wr)
668 {
669 	int ret;
670 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
671 	unsigned long flags;
672 	struct ib_mad_local_private *local;
673 	struct ib_mad_private *mad_priv;
674 	struct ib_mad_port_private *port_priv;
675 	struct ib_mad_agent_private *recv_mad_agent = NULL;
676 	struct ib_device *device = mad_agent_priv->agent.device;
677 	u8 port_num = mad_agent_priv->agent.port_num;
678 	struct ib_wc mad_wc;
679 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
680 
681 	/*
682 	 * Directed route handling starts if the initial LID routed part of
683 	 * a request or the ending LID routed part of a response is empty.
684 	 * If we are at the start of the LID routed part, don't update the
685 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
686 	 */
687 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
688 	     IB_LID_PERMISSIVE &&
689 	    !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
690 		ret = -EINVAL;
691 		printk(KERN_ERR PFX "Invalid directed route\n");
692 		goto out;
693 	}
694 	/* Check to post send on QP or process locally */
695 	ret = smi_check_local_smp(smp, device);
696 	if (!ret)
697 		goto out;
698 
699 	local = kmalloc(sizeof *local, GFP_ATOMIC);
700 	if (!local) {
701 		ret = -ENOMEM;
702 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
703 		goto out;
704 	}
705 	local->mad_priv = NULL;
706 	local->recv_mad_agent = NULL;
707 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
708 	if (!mad_priv) {
709 		ret = -ENOMEM;
710 		printk(KERN_ERR PFX "No memory for local response MAD\n");
711 		kfree(local);
712 		goto out;
713 	}
714 
715 	build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
716 		     send_wr->wr.ud.pkey_index,
717 		     send_wr->wr.ud.port_num, &mad_wc);
718 
719 	/* No GRH for DR SMP */
720 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
721 				  (struct ib_mad *)smp,
722 				  (struct ib_mad *)&mad_priv->mad);
723 	switch (ret)
724 	{
725 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
726 		if (response_mad(&mad_priv->mad.mad) &&
727 		    mad_agent_priv->agent.recv_handler) {
728 			local->mad_priv = mad_priv;
729 			local->recv_mad_agent = mad_agent_priv;
730 			/*
731 			 * Reference MAD agent until receive
732 			 * side of local completion handled
733 			 */
734 			atomic_inc(&mad_agent_priv->refcount);
735 		} else
736 			kmem_cache_free(ib_mad_cache, mad_priv);
737 		break;
738 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
739 		kmem_cache_free(ib_mad_cache, mad_priv);
740 		break;
741 	case IB_MAD_RESULT_SUCCESS:
742 		/* Treat like an incoming receive MAD */
743 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
744 					    mad_agent_priv->agent.port_num);
745 		if (port_priv) {
746 			mad_priv->mad.mad.mad_hdr.tid =
747 				((struct ib_mad *)smp)->mad_hdr.tid;
748 			recv_mad_agent = find_mad_agent(port_priv,
749 						        &mad_priv->mad.mad);
750 		}
751 		if (!port_priv || !recv_mad_agent) {
752 			kmem_cache_free(ib_mad_cache, mad_priv);
753 			kfree(local);
754 			ret = 0;
755 			goto out;
756 		}
757 		local->mad_priv = mad_priv;
758 		local->recv_mad_agent = recv_mad_agent;
759 		break;
760 	default:
761 		kmem_cache_free(ib_mad_cache, mad_priv);
762 		kfree(local);
763 		ret = -EINVAL;
764 		goto out;
765 	}
766 
767 	local->mad_send_wr = mad_send_wr;
768 	/* Reference MAD agent until send side of local completion handled */
769 	atomic_inc(&mad_agent_priv->refcount);
770 	/* Queue local completion to local list */
771 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
772 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
773 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
774 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
775 		   &mad_agent_priv->local_work);
776 	ret = 1;
777 out:
778 	return ret;
779 }
780 
781 static int get_pad_size(int hdr_len, int data_len)
782 {
783 	int seg_size, pad;
784 
785 	seg_size = sizeof(struct ib_mad) - hdr_len;
786 	if (data_len && seg_size) {
787 		pad = seg_size - data_len % seg_size;
788 		return pad == seg_size ? 0 : pad;
789 	} else
790 		return seg_size;
791 }
792 
793 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
794 {
795 	struct ib_rmpp_segment *s, *t;
796 
797 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
798 		list_del(&s->list);
799 		kfree(s);
800 	}
801 }
802 
803 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
804 				gfp_t gfp_mask)
805 {
806 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
807 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
808 	struct ib_rmpp_segment *seg = NULL;
809 	int left, seg_size, pad;
810 
811 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
812 	seg_size = send_buf->seg_size;
813 	pad = send_wr->pad;
814 
815 	/* Allocate data segments. */
816 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
817 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
818 		if (!seg) {
819 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
820 			       "alloc failed for len %zd, gfp %#x\n",
821 			       sizeof (*seg) + seg_size, gfp_mask);
822 			free_send_rmpp_list(send_wr);
823 			return -ENOMEM;
824 		}
825 		seg->num = ++send_buf->seg_count;
826 		list_add_tail(&seg->list, &send_wr->rmpp_list);
827 	}
828 
829 	/* Zero any padding */
830 	if (pad)
831 		memset(seg->data + seg_size - pad, 0, pad);
832 
833 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
834 					  agent.rmpp_version;
835 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
836 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
837 
838 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
839 					struct ib_rmpp_segment, list);
840 	send_wr->last_ack_seg = send_wr->cur_seg;
841 	return 0;
842 }
843 
844 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
845 					    u32 remote_qpn, u16 pkey_index,
846 					    int rmpp_active,
847 					    int hdr_len, int data_len,
848 					    gfp_t gfp_mask)
849 {
850 	struct ib_mad_agent_private *mad_agent_priv;
851 	struct ib_mad_send_wr_private *mad_send_wr;
852 	int pad, message_size, ret, size;
853 	void *buf;
854 
855 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
856 				      agent);
857 	pad = get_pad_size(hdr_len, data_len);
858 	message_size = hdr_len + data_len + pad;
859 
860 	if ((!mad_agent->rmpp_version &&
861 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
862 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
863 		return ERR_PTR(-EINVAL);
864 
865 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
866 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
867 	if (!buf)
868 		return ERR_PTR(-ENOMEM);
869 
870 	mad_send_wr = buf + size;
871 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
872 	mad_send_wr->send_buf.mad = buf;
873 	mad_send_wr->send_buf.hdr_len = hdr_len;
874 	mad_send_wr->send_buf.data_len = data_len;
875 	mad_send_wr->pad = pad;
876 
877 	mad_send_wr->mad_agent_priv = mad_agent_priv;
878 	mad_send_wr->sg_list[0].length = hdr_len;
879 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
880 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
881 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
882 
883 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
884 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
885 	mad_send_wr->send_wr.num_sge = 2;
886 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
887 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
888 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
889 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
890 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
891 
892 	if (rmpp_active) {
893 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
894 		if (ret) {
895 			kfree(buf);
896 			return ERR_PTR(ret);
897 		}
898 	}
899 
900 	mad_send_wr->send_buf.mad_agent = mad_agent;
901 	atomic_inc(&mad_agent_priv->refcount);
902 	return &mad_send_wr->send_buf;
903 }
904 EXPORT_SYMBOL(ib_create_send_mad);
905 
906 int ib_get_mad_data_offset(u8 mgmt_class)
907 {
908 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
909 		return IB_MGMT_SA_HDR;
910 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
911 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
912 		 (mgmt_class == IB_MGMT_CLASS_BIS))
913 		return IB_MGMT_DEVICE_HDR;
914 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
915 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
916 		return IB_MGMT_VENDOR_HDR;
917 	else
918 		return IB_MGMT_MAD_HDR;
919 }
920 EXPORT_SYMBOL(ib_get_mad_data_offset);
921 
922 int ib_is_mad_class_rmpp(u8 mgmt_class)
923 {
924 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
925 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
926 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
927 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
928 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
929 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
930 		return 1;
931 	return 0;
932 }
933 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
934 
935 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
936 {
937 	struct ib_mad_send_wr_private *mad_send_wr;
938 	struct list_head *list;
939 
940 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
941 				   send_buf);
942 	list = &mad_send_wr->cur_seg->list;
943 
944 	if (mad_send_wr->cur_seg->num < seg_num) {
945 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
946 			if (mad_send_wr->cur_seg->num == seg_num)
947 				break;
948 	} else if (mad_send_wr->cur_seg->num > seg_num) {
949 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
950 			if (mad_send_wr->cur_seg->num == seg_num)
951 				break;
952 	}
953 	return mad_send_wr->cur_seg->data;
954 }
955 EXPORT_SYMBOL(ib_get_rmpp_segment);
956 
957 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
958 {
959 	if (mad_send_wr->send_buf.seg_count)
960 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
961 					   mad_send_wr->seg_num);
962 	else
963 		return mad_send_wr->send_buf.mad +
964 		       mad_send_wr->send_buf.hdr_len;
965 }
966 
967 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
968 {
969 	struct ib_mad_agent_private *mad_agent_priv;
970 	struct ib_mad_send_wr_private *mad_send_wr;
971 
972 	mad_agent_priv = container_of(send_buf->mad_agent,
973 				      struct ib_mad_agent_private, agent);
974 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
975 				   send_buf);
976 
977 	free_send_rmpp_list(mad_send_wr);
978 	kfree(send_buf->mad);
979 	deref_mad_agent(mad_agent_priv);
980 }
981 EXPORT_SYMBOL(ib_free_send_mad);
982 
983 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
984 {
985 	struct ib_mad_qp_info *qp_info;
986 	struct list_head *list;
987 	struct ib_send_wr *bad_send_wr;
988 	struct ib_mad_agent *mad_agent;
989 	struct ib_sge *sge;
990 	unsigned long flags;
991 	int ret;
992 
993 	/* Set WR ID to find mad_send_wr upon completion */
994 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
995 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
996 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
997 
998 	mad_agent = mad_send_wr->send_buf.mad_agent;
999 	sge = mad_send_wr->sg_list;
1000 	sge[0].addr = dma_map_single(mad_agent->device->dma_device,
1001 				     mad_send_wr->send_buf.mad,
1002 				     sge[0].length,
1003 				     DMA_TO_DEVICE);
1004 	pci_unmap_addr_set(mad_send_wr, header_mapping, sge[0].addr);
1005 
1006 	sge[1].addr = dma_map_single(mad_agent->device->dma_device,
1007 				     ib_get_payload(mad_send_wr),
1008 				     sge[1].length,
1009 				     DMA_TO_DEVICE);
1010 	pci_unmap_addr_set(mad_send_wr, payload_mapping, sge[1].addr);
1011 
1012 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1013 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1014 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1015 				   &bad_send_wr);
1016 		list = &qp_info->send_queue.list;
1017 	} else {
1018 		ret = 0;
1019 		list = &qp_info->overflow_list;
1020 	}
1021 
1022 	if (!ret) {
1023 		qp_info->send_queue.count++;
1024 		list_add_tail(&mad_send_wr->mad_list.list, list);
1025 	}
1026 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1027 	if (ret) {
1028 		dma_unmap_single(mad_agent->device->dma_device,
1029 				 pci_unmap_addr(mad_send_wr, header_mapping),
1030 				 sge[0].length, DMA_TO_DEVICE);
1031 		dma_unmap_single(mad_agent->device->dma_device,
1032 				 pci_unmap_addr(mad_send_wr, payload_mapping),
1033 				 sge[1].length, DMA_TO_DEVICE);
1034 	}
1035 	return ret;
1036 }
1037 
1038 /*
1039  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1040  *  with the registered client
1041  */
1042 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1043 		     struct ib_mad_send_buf **bad_send_buf)
1044 {
1045 	struct ib_mad_agent_private *mad_agent_priv;
1046 	struct ib_mad_send_buf *next_send_buf;
1047 	struct ib_mad_send_wr_private *mad_send_wr;
1048 	unsigned long flags;
1049 	int ret = -EINVAL;
1050 
1051 	/* Walk list of send WRs and post each on send list */
1052 	for (; send_buf; send_buf = next_send_buf) {
1053 
1054 		mad_send_wr = container_of(send_buf,
1055 					   struct ib_mad_send_wr_private,
1056 					   send_buf);
1057 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1058 
1059 		if (!send_buf->mad_agent->send_handler ||
1060 		    (send_buf->timeout_ms &&
1061 		     !send_buf->mad_agent->recv_handler)) {
1062 			ret = -EINVAL;
1063 			goto error;
1064 		}
1065 
1066 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1067 			if (mad_agent_priv->agent.rmpp_version) {
1068 				ret = -EINVAL;
1069 				goto error;
1070 			}
1071 		}
1072 
1073 		/*
1074 		 * Save pointer to next work request to post in case the
1075 		 * current one completes, and the user modifies the work
1076 		 * request associated with the completion
1077 		 */
1078 		next_send_buf = send_buf->next;
1079 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1080 
1081 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1082 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1083 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1084 						     mad_send_wr);
1085 			if (ret < 0)		/* error */
1086 				goto error;
1087 			else if (ret == 1)	/* locally consumed */
1088 				continue;
1089 		}
1090 
1091 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1092 		/* Timeout will be updated after send completes */
1093 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1094 		mad_send_wr->retries = send_buf->retries;
1095 		/* Reference for work request to QP + response */
1096 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1097 		mad_send_wr->status = IB_WC_SUCCESS;
1098 
1099 		/* Reference MAD agent until send completes */
1100 		atomic_inc(&mad_agent_priv->refcount);
1101 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1102 		list_add_tail(&mad_send_wr->agent_list,
1103 			      &mad_agent_priv->send_list);
1104 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1105 
1106 		if (mad_agent_priv->agent.rmpp_version) {
1107 			ret = ib_send_rmpp_mad(mad_send_wr);
1108 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1109 				ret = ib_send_mad(mad_send_wr);
1110 		} else
1111 			ret = ib_send_mad(mad_send_wr);
1112 		if (ret < 0) {
1113 			/* Fail send request */
1114 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1115 			list_del(&mad_send_wr->agent_list);
1116 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1117 			atomic_dec(&mad_agent_priv->refcount);
1118 			goto error;
1119 		}
1120 	}
1121 	return 0;
1122 error:
1123 	if (bad_send_buf)
1124 		*bad_send_buf = send_buf;
1125 	return ret;
1126 }
1127 EXPORT_SYMBOL(ib_post_send_mad);
1128 
1129 /*
1130  * ib_free_recv_mad - Returns data buffers used to receive
1131  *  a MAD to the access layer
1132  */
1133 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1134 {
1135 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1136 	struct ib_mad_private_header *mad_priv_hdr;
1137 	struct ib_mad_private *priv;
1138 	struct list_head free_list;
1139 
1140 	INIT_LIST_HEAD(&free_list);
1141 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1142 
1143 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1144 					&free_list, list) {
1145 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1146 					   recv_buf);
1147 		mad_priv_hdr = container_of(mad_recv_wc,
1148 					    struct ib_mad_private_header,
1149 					    recv_wc);
1150 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1151 				    header);
1152 		kmem_cache_free(ib_mad_cache, priv);
1153 	}
1154 }
1155 EXPORT_SYMBOL(ib_free_recv_mad);
1156 
1157 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1158 					u8 rmpp_version,
1159 					ib_mad_send_handler send_handler,
1160 					ib_mad_recv_handler recv_handler,
1161 					void *context)
1162 {
1163 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1164 }
1165 EXPORT_SYMBOL(ib_redirect_mad_qp);
1166 
1167 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1168 		      struct ib_wc *wc)
1169 {
1170 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1171 	return 0;
1172 }
1173 EXPORT_SYMBOL(ib_process_mad_wc);
1174 
1175 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1176 			 struct ib_mad_reg_req *mad_reg_req)
1177 {
1178 	int i;
1179 
1180 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1181 	     i < IB_MGMT_MAX_METHODS;
1182 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1183 			       1+i)) {
1184 		if ((*method)->agent[i]) {
1185 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1186 			return -EINVAL;
1187 		}
1188 	}
1189 	return 0;
1190 }
1191 
1192 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1193 {
1194 	/* Allocate management method table */
1195 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1196 	if (!*method) {
1197 		printk(KERN_ERR PFX "No memory for "
1198 		       "ib_mad_mgmt_method_table\n");
1199 		return -ENOMEM;
1200 	}
1201 
1202 	return 0;
1203 }
1204 
1205 /*
1206  * Check to see if there are any methods still in use
1207  */
1208 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1209 {
1210 	int i;
1211 
1212 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1213 		if (method->agent[i])
1214 			return 1;
1215 	return 0;
1216 }
1217 
1218 /*
1219  * Check to see if there are any method tables for this class still in use
1220  */
1221 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1222 {
1223 	int i;
1224 
1225 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1226 		if (class->method_table[i])
1227 			return 1;
1228 	return 0;
1229 }
1230 
1231 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1232 {
1233 	int i;
1234 
1235 	for (i = 0; i < MAX_MGMT_OUI; i++)
1236 		if (vendor_class->method_table[i])
1237 			return 1;
1238 	return 0;
1239 }
1240 
1241 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1242 			   char *oui)
1243 {
1244 	int i;
1245 
1246 	for (i = 0; i < MAX_MGMT_OUI; i++)
1247                 /* Is there matching OUI for this vendor class ? */
1248                 if (!memcmp(vendor_class->oui[i], oui, 3))
1249 			return i;
1250 
1251 	return -1;
1252 }
1253 
1254 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1255 {
1256 	int i;
1257 
1258 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1259 		if (vendor->vendor_class[i])
1260 			return 1;
1261 
1262 	return 0;
1263 }
1264 
1265 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1266 				     struct ib_mad_agent_private *agent)
1267 {
1268 	int i;
1269 
1270 	/* Remove any methods for this mad agent */
1271 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1272 		if (method->agent[i] == agent) {
1273 			method->agent[i] = NULL;
1274 		}
1275 	}
1276 }
1277 
1278 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1279 			      struct ib_mad_agent_private *agent_priv,
1280 			      u8 mgmt_class)
1281 {
1282 	struct ib_mad_port_private *port_priv;
1283 	struct ib_mad_mgmt_class_table **class;
1284 	struct ib_mad_mgmt_method_table **method;
1285 	int i, ret;
1286 
1287 	port_priv = agent_priv->qp_info->port_priv;
1288 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1289 	if (!*class) {
1290 		/* Allocate management class table for "new" class version */
1291 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1292 		if (!*class) {
1293 			printk(KERN_ERR PFX "No memory for "
1294 			       "ib_mad_mgmt_class_table\n");
1295 			ret = -ENOMEM;
1296 			goto error1;
1297 		}
1298 
1299 		/* Allocate method table for this management class */
1300 		method = &(*class)->method_table[mgmt_class];
1301 		if ((ret = allocate_method_table(method)))
1302 			goto error2;
1303 	} else {
1304 		method = &(*class)->method_table[mgmt_class];
1305 		if (!*method) {
1306 			/* Allocate method table for this management class */
1307 			if ((ret = allocate_method_table(method)))
1308 				goto error1;
1309 		}
1310 	}
1311 
1312 	/* Now, make sure methods are not already in use */
1313 	if (method_in_use(method, mad_reg_req))
1314 		goto error3;
1315 
1316 	/* Finally, add in methods being registered */
1317 	for (i = find_first_bit(mad_reg_req->method_mask,
1318 				IB_MGMT_MAX_METHODS);
1319 	     i < IB_MGMT_MAX_METHODS;
1320 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1321 			       1+i)) {
1322 		(*method)->agent[i] = agent_priv;
1323 	}
1324 	return 0;
1325 
1326 error3:
1327 	/* Remove any methods for this mad agent */
1328 	remove_methods_mad_agent(*method, agent_priv);
1329 	/* Now, check to see if there are any methods in use */
1330 	if (!check_method_table(*method)) {
1331 		/* If not, release management method table */
1332 		kfree(*method);
1333 		*method = NULL;
1334 	}
1335 	ret = -EINVAL;
1336 	goto error1;
1337 error2:
1338 	kfree(*class);
1339 	*class = NULL;
1340 error1:
1341 	return ret;
1342 }
1343 
1344 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1345 			   struct ib_mad_agent_private *agent_priv)
1346 {
1347 	struct ib_mad_port_private *port_priv;
1348 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1349 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1350 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1351 	struct ib_mad_mgmt_method_table **method;
1352 	int i, ret = -ENOMEM;
1353 	u8 vclass;
1354 
1355 	/* "New" vendor (with OUI) class */
1356 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1357 	port_priv = agent_priv->qp_info->port_priv;
1358 	vendor_table = &port_priv->version[
1359 				mad_reg_req->mgmt_class_version].vendor;
1360 	if (!*vendor_table) {
1361 		/* Allocate mgmt vendor class table for "new" class version */
1362 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1363 		if (!vendor) {
1364 			printk(KERN_ERR PFX "No memory for "
1365 			       "ib_mad_mgmt_vendor_class_table\n");
1366 			goto error1;
1367 		}
1368 
1369 		*vendor_table = vendor;
1370 	}
1371 	if (!(*vendor_table)->vendor_class[vclass]) {
1372 		/* Allocate table for this management vendor class */
1373 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1374 		if (!vendor_class) {
1375 			printk(KERN_ERR PFX "No memory for "
1376 			       "ib_mad_mgmt_vendor_class\n");
1377 			goto error2;
1378 		}
1379 
1380 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1381 	}
1382 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1383 		/* Is there matching OUI for this vendor class ? */
1384 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1385 			    mad_reg_req->oui, 3)) {
1386 			method = &(*vendor_table)->vendor_class[
1387 						vclass]->method_table[i];
1388 			BUG_ON(!*method);
1389 			goto check_in_use;
1390 		}
1391 	}
1392 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1393 		/* OUI slot available ? */
1394 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1395 				vclass]->oui[i])) {
1396 			method = &(*vendor_table)->vendor_class[
1397 				vclass]->method_table[i];
1398 			BUG_ON(*method);
1399 			/* Allocate method table for this OUI */
1400 			if ((ret = allocate_method_table(method)))
1401 				goto error3;
1402 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1403 			       mad_reg_req->oui, 3);
1404 			goto check_in_use;
1405 		}
1406 	}
1407 	printk(KERN_ERR PFX "All OUI slots in use\n");
1408 	goto error3;
1409 
1410 check_in_use:
1411 	/* Now, make sure methods are not already in use */
1412 	if (method_in_use(method, mad_reg_req))
1413 		goto error4;
1414 
1415 	/* Finally, add in methods being registered */
1416 	for (i = find_first_bit(mad_reg_req->method_mask,
1417 				IB_MGMT_MAX_METHODS);
1418 	     i < IB_MGMT_MAX_METHODS;
1419 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1420 			       1+i)) {
1421 		(*method)->agent[i] = agent_priv;
1422 	}
1423 	return 0;
1424 
1425 error4:
1426 	/* Remove any methods for this mad agent */
1427 	remove_methods_mad_agent(*method, agent_priv);
1428 	/* Now, check to see if there are any methods in use */
1429 	if (!check_method_table(*method)) {
1430 		/* If not, release management method table */
1431 		kfree(*method);
1432 		*method = NULL;
1433 	}
1434 	ret = -EINVAL;
1435 error3:
1436 	if (vendor_class) {
1437 		(*vendor_table)->vendor_class[vclass] = NULL;
1438 		kfree(vendor_class);
1439 	}
1440 error2:
1441 	if (vendor) {
1442 		*vendor_table = NULL;
1443 		kfree(vendor);
1444 	}
1445 error1:
1446 	return ret;
1447 }
1448 
1449 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1450 {
1451 	struct ib_mad_port_private *port_priv;
1452 	struct ib_mad_mgmt_class_table *class;
1453 	struct ib_mad_mgmt_method_table *method;
1454 	struct ib_mad_mgmt_vendor_class_table *vendor;
1455 	struct ib_mad_mgmt_vendor_class *vendor_class;
1456 	int index;
1457 	u8 mgmt_class;
1458 
1459 	/*
1460 	 * Was MAD registration request supplied
1461 	 * with original registration ?
1462 	 */
1463 	if (!agent_priv->reg_req) {
1464 		goto out;
1465 	}
1466 
1467 	port_priv = agent_priv->qp_info->port_priv;
1468 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1469 	class = port_priv->version[
1470 			agent_priv->reg_req->mgmt_class_version].class;
1471 	if (!class)
1472 		goto vendor_check;
1473 
1474 	method = class->method_table[mgmt_class];
1475 	if (method) {
1476 		/* Remove any methods for this mad agent */
1477 		remove_methods_mad_agent(method, agent_priv);
1478 		/* Now, check to see if there are any methods still in use */
1479 		if (!check_method_table(method)) {
1480 			/* If not, release management method table */
1481 			 kfree(method);
1482 			 class->method_table[mgmt_class] = NULL;
1483 			 /* Any management classes left ? */
1484 			if (!check_class_table(class)) {
1485 				/* If not, release management class table */
1486 				kfree(class);
1487 				port_priv->version[
1488 					agent_priv->reg_req->
1489 					mgmt_class_version].class = NULL;
1490 			}
1491 		}
1492 	}
1493 
1494 vendor_check:
1495 	if (!is_vendor_class(mgmt_class))
1496 		goto out;
1497 
1498 	/* normalize mgmt_class to vendor range 2 */
1499 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1500 	vendor = port_priv->version[
1501 			agent_priv->reg_req->mgmt_class_version].vendor;
1502 
1503 	if (!vendor)
1504 		goto out;
1505 
1506 	vendor_class = vendor->vendor_class[mgmt_class];
1507 	if (vendor_class) {
1508 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1509 		if (index < 0)
1510 			goto out;
1511 		method = vendor_class->method_table[index];
1512 		if (method) {
1513 			/* Remove any methods for this mad agent */
1514 			remove_methods_mad_agent(method, agent_priv);
1515 			/*
1516 			 * Now, check to see if there are
1517 			 * any methods still in use
1518 			 */
1519 			if (!check_method_table(method)) {
1520 				/* If not, release management method table */
1521 				kfree(method);
1522 				vendor_class->method_table[index] = NULL;
1523 				memset(vendor_class->oui[index], 0, 3);
1524 				/* Any OUIs left ? */
1525 				if (!check_vendor_class(vendor_class)) {
1526 					/* If not, release vendor class table */
1527 					kfree(vendor_class);
1528 					vendor->vendor_class[mgmt_class] = NULL;
1529 					/* Any other vendor classes left ? */
1530 					if (!check_vendor_table(vendor)) {
1531 						kfree(vendor);
1532 						port_priv->version[
1533 							agent_priv->reg_req->
1534 							mgmt_class_version].
1535 							vendor = NULL;
1536 					}
1537 				}
1538 			}
1539 		}
1540 	}
1541 
1542 out:
1543 	return;
1544 }
1545 
1546 static struct ib_mad_agent_private *
1547 find_mad_agent(struct ib_mad_port_private *port_priv,
1548 	       struct ib_mad *mad)
1549 {
1550 	struct ib_mad_agent_private *mad_agent = NULL;
1551 	unsigned long flags;
1552 
1553 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1554 	if (response_mad(mad)) {
1555 		u32 hi_tid;
1556 		struct ib_mad_agent_private *entry;
1557 
1558 		/*
1559 		 * Routing is based on high 32 bits of transaction ID
1560 		 * of MAD.
1561 		 */
1562 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1563 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1564 			if (entry->agent.hi_tid == hi_tid) {
1565 				mad_agent = entry;
1566 				break;
1567 			}
1568 		}
1569 	} else {
1570 		struct ib_mad_mgmt_class_table *class;
1571 		struct ib_mad_mgmt_method_table *method;
1572 		struct ib_mad_mgmt_vendor_class_table *vendor;
1573 		struct ib_mad_mgmt_vendor_class *vendor_class;
1574 		struct ib_vendor_mad *vendor_mad;
1575 		int index;
1576 
1577 		/*
1578 		 * Routing is based on version, class, and method
1579 		 * For "newer" vendor MADs, also based on OUI
1580 		 */
1581 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1582 			goto out;
1583 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1584 			class = port_priv->version[
1585 					mad->mad_hdr.class_version].class;
1586 			if (!class)
1587 				goto out;
1588 			method = class->method_table[convert_mgmt_class(
1589 							mad->mad_hdr.mgmt_class)];
1590 			if (method)
1591 				mad_agent = method->agent[mad->mad_hdr.method &
1592 							  ~IB_MGMT_METHOD_RESP];
1593 		} else {
1594 			vendor = port_priv->version[
1595 					mad->mad_hdr.class_version].vendor;
1596 			if (!vendor)
1597 				goto out;
1598 			vendor_class = vendor->vendor_class[vendor_class_index(
1599 						mad->mad_hdr.mgmt_class)];
1600 			if (!vendor_class)
1601 				goto out;
1602 			/* Find matching OUI */
1603 			vendor_mad = (struct ib_vendor_mad *)mad;
1604 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1605 			if (index == -1)
1606 				goto out;
1607 			method = vendor_class->method_table[index];
1608 			if (method) {
1609 				mad_agent = method->agent[mad->mad_hdr.method &
1610 							  ~IB_MGMT_METHOD_RESP];
1611 			}
1612 		}
1613 	}
1614 
1615 	if (mad_agent) {
1616 		if (mad_agent->agent.recv_handler)
1617 			atomic_inc(&mad_agent->refcount);
1618 		else {
1619 			printk(KERN_NOTICE PFX "No receive handler for client "
1620 			       "%p on port %d\n",
1621 			       &mad_agent->agent, port_priv->port_num);
1622 			mad_agent = NULL;
1623 		}
1624 	}
1625 out:
1626 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1627 
1628 	return mad_agent;
1629 }
1630 
1631 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1632 {
1633 	int valid = 0;
1634 
1635 	/* Make sure MAD base version is understood */
1636 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1637 		printk(KERN_ERR PFX "MAD received with unsupported base "
1638 		       "version %d\n", mad->mad_hdr.base_version);
1639 		goto out;
1640 	}
1641 
1642 	/* Filter SMI packets sent to other than QP0 */
1643 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1644 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1645 		if (qp_num == 0)
1646 			valid = 1;
1647 	} else {
1648 		/* Filter GSI packets sent to QP0 */
1649 		if (qp_num != 0)
1650 			valid = 1;
1651 	}
1652 
1653 out:
1654 	return valid;
1655 }
1656 
1657 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1658 		       struct ib_mad_hdr *mad_hdr)
1659 {
1660 	struct ib_rmpp_mad *rmpp_mad;
1661 
1662 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1663 	return !mad_agent_priv->agent.rmpp_version ||
1664 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1665 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1666 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1667 }
1668 
1669 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1670 				     struct ib_mad_recv_wc *rwc)
1671 {
1672 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1673 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1674 }
1675 
1676 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1677 				   struct ib_mad_send_wr_private *wr,
1678 				   struct ib_mad_recv_wc *rwc )
1679 {
1680 	struct ib_ah_attr attr;
1681 	u8 send_resp, rcv_resp;
1682 	union ib_gid sgid;
1683 	struct ib_device *device = mad_agent_priv->agent.device;
1684 	u8 port_num = mad_agent_priv->agent.port_num;
1685 	u8 lmc;
1686 
1687 	send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1688 		     mad_hdr.method & IB_MGMT_METHOD_RESP;
1689 	rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1690 
1691 	if (send_resp == rcv_resp)
1692 		/* both requests, or both responses. GIDs different */
1693 		return 0;
1694 
1695 	if (ib_query_ah(wr->send_buf.ah, &attr))
1696 		/* Assume not equal, to avoid false positives. */
1697 		return 0;
1698 
1699 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1700 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1701 		/* one has GID, other does not.  Assume different */
1702 		return 0;
1703 
1704 	if (!send_resp && rcv_resp) {
1705 		/* is request/response. */
1706 		if (!(attr.ah_flags & IB_AH_GRH)) {
1707 			if (ib_get_cached_lmc(device, port_num, &lmc))
1708 				return 0;
1709 			return (!lmc || !((attr.src_path_bits ^
1710 					   rwc->wc->dlid_path_bits) &
1711 					  ((1 << lmc) - 1)));
1712 		} else {
1713 			if (ib_get_cached_gid(device, port_num,
1714 					      attr.grh.sgid_index, &sgid))
1715 				return 0;
1716 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1717 				       16);
1718 		}
1719 	}
1720 
1721 	if (!(attr.ah_flags & IB_AH_GRH))
1722 		return attr.dlid == rwc->wc->slid;
1723 	else
1724 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1725 			       16);
1726 }
1727 
1728 static inline int is_direct(u8 class)
1729 {
1730 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1731 }
1732 
1733 struct ib_mad_send_wr_private*
1734 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1735 		 struct ib_mad_recv_wc *wc)
1736 {
1737 	struct ib_mad_send_wr_private *wr;
1738 	struct ib_mad *mad;
1739 
1740 	mad = (struct ib_mad *)wc->recv_buf.mad;
1741 
1742 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1743 		if ((wr->tid == mad->mad_hdr.tid) &&
1744 		    rcv_has_same_class(wr, wc) &&
1745 		    /*
1746 		     * Don't check GID for direct routed MADs.
1747 		     * These might have permissive LIDs.
1748 		     */
1749 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1750 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1751 			return wr;
1752 	}
1753 
1754 	/*
1755 	 * It's possible to receive the response before we've
1756 	 * been notified that the send has completed
1757 	 */
1758 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1759 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1760 		    wr->tid == mad->mad_hdr.tid &&
1761 		    wr->timeout &&
1762 		    rcv_has_same_class(wr, wc) &&
1763 		    /*
1764 		     * Don't check GID for direct routed MADs.
1765 		     * These might have permissive LIDs.
1766 		     */
1767 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1768 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1769 			/* Verify request has not been canceled */
1770 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1771 	}
1772 	return NULL;
1773 }
1774 
1775 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1776 {
1777 	mad_send_wr->timeout = 0;
1778 	if (mad_send_wr->refcount == 1)
1779 		list_move_tail(&mad_send_wr->agent_list,
1780 			      &mad_send_wr->mad_agent_priv->done_list);
1781 }
1782 
1783 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1784 				 struct ib_mad_recv_wc *mad_recv_wc)
1785 {
1786 	struct ib_mad_send_wr_private *mad_send_wr;
1787 	struct ib_mad_send_wc mad_send_wc;
1788 	unsigned long flags;
1789 
1790 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1791 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1792 	if (mad_agent_priv->agent.rmpp_version) {
1793 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1794 						      mad_recv_wc);
1795 		if (!mad_recv_wc) {
1796 			deref_mad_agent(mad_agent_priv);
1797 			return;
1798 		}
1799 	}
1800 
1801 	/* Complete corresponding request */
1802 	if (response_mad(mad_recv_wc->recv_buf.mad)) {
1803 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1804 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1805 		if (!mad_send_wr) {
1806 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1807 			ib_free_recv_mad(mad_recv_wc);
1808 			deref_mad_agent(mad_agent_priv);
1809 			return;
1810 		}
1811 		ib_mark_mad_done(mad_send_wr);
1812 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1813 
1814 		/* Defined behavior is to complete response before request */
1815 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1816 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1817 						   mad_recv_wc);
1818 		atomic_dec(&mad_agent_priv->refcount);
1819 
1820 		mad_send_wc.status = IB_WC_SUCCESS;
1821 		mad_send_wc.vendor_err = 0;
1822 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1823 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1824 	} else {
1825 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1826 						   mad_recv_wc);
1827 		deref_mad_agent(mad_agent_priv);
1828 	}
1829 }
1830 
1831 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1832 				     struct ib_wc *wc)
1833 {
1834 	struct ib_mad_qp_info *qp_info;
1835 	struct ib_mad_private_header *mad_priv_hdr;
1836 	struct ib_mad_private *recv, *response;
1837 	struct ib_mad_list_head *mad_list;
1838 	struct ib_mad_agent_private *mad_agent;
1839 
1840 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1841 	if (!response)
1842 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1843 		       "for response buffer\n");
1844 
1845 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1846 	qp_info = mad_list->mad_queue->qp_info;
1847 	dequeue_mad(mad_list);
1848 
1849 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1850 				    mad_list);
1851 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1852 	dma_unmap_single(port_priv->device->dma_device,
1853 			 pci_unmap_addr(&recv->header, mapping),
1854 			 sizeof(struct ib_mad_private) -
1855 			 sizeof(struct ib_mad_private_header),
1856 			 DMA_FROM_DEVICE);
1857 
1858 	/* Setup MAD receive work completion from "normal" work completion */
1859 	recv->header.wc = *wc;
1860 	recv->header.recv_wc.wc = &recv->header.wc;
1861 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1862 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1863 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1864 
1865 	if (atomic_read(&qp_info->snoop_count))
1866 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1867 
1868 	/* Validate MAD */
1869 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1870 		goto out;
1871 
1872 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1873 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1874 		if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1875 					    port_priv->device->node_type,
1876 					    port_priv->port_num,
1877 					    port_priv->device->phys_port_cnt))
1878 			goto out;
1879 		if (!smi_check_forward_dr_smp(&recv->mad.smp))
1880 			goto local;
1881 		if (!smi_handle_dr_smp_send(&recv->mad.smp,
1882 					    port_priv->device->node_type,
1883 					    port_priv->port_num))
1884 			goto out;
1885 		if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1886 			goto out;
1887 	}
1888 
1889 local:
1890 	/* Give driver "right of first refusal" on incoming MAD */
1891 	if (port_priv->device->process_mad) {
1892 		int ret;
1893 
1894 		if (!response) {
1895 			printk(KERN_ERR PFX "No memory for response MAD\n");
1896 			/*
1897 			 * Is it better to assume that
1898 			 * it wouldn't be processed ?
1899 			 */
1900 			goto out;
1901 		}
1902 
1903 		ret = port_priv->device->process_mad(port_priv->device, 0,
1904 						     port_priv->port_num,
1905 						     wc, &recv->grh,
1906 						     &recv->mad.mad,
1907 						     &response->mad.mad);
1908 		if (ret & IB_MAD_RESULT_SUCCESS) {
1909 			if (ret & IB_MAD_RESULT_CONSUMED)
1910 				goto out;
1911 			if (ret & IB_MAD_RESULT_REPLY) {
1912 				agent_send_response(&response->mad.mad,
1913 						    &recv->grh, wc,
1914 						    port_priv->device,
1915 						    port_priv->port_num,
1916 						    qp_info->qp->qp_num);
1917 				goto out;
1918 			}
1919 		}
1920 	}
1921 
1922 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1923 	if (mad_agent) {
1924 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1925 		/*
1926 		 * recv is freed up in error cases in ib_mad_complete_recv
1927 		 * or via recv_handler in ib_mad_complete_recv()
1928 		 */
1929 		recv = NULL;
1930 	}
1931 
1932 out:
1933 	/* Post another receive request for this QP */
1934 	if (response) {
1935 		ib_mad_post_receive_mads(qp_info, response);
1936 		if (recv)
1937 			kmem_cache_free(ib_mad_cache, recv);
1938 	} else
1939 		ib_mad_post_receive_mads(qp_info, recv);
1940 }
1941 
1942 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1943 {
1944 	struct ib_mad_send_wr_private *mad_send_wr;
1945 	unsigned long delay;
1946 
1947 	if (list_empty(&mad_agent_priv->wait_list)) {
1948 		cancel_delayed_work(&mad_agent_priv->timed_work);
1949 	} else {
1950 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1951 					 struct ib_mad_send_wr_private,
1952 					 agent_list);
1953 
1954 		if (time_after(mad_agent_priv->timeout,
1955 			       mad_send_wr->timeout)) {
1956 			mad_agent_priv->timeout = mad_send_wr->timeout;
1957 			cancel_delayed_work(&mad_agent_priv->timed_work);
1958 			delay = mad_send_wr->timeout - jiffies;
1959 			if ((long)delay <= 0)
1960 				delay = 1;
1961 			queue_delayed_work(mad_agent_priv->qp_info->
1962 					   port_priv->wq,
1963 					   &mad_agent_priv->timed_work, delay);
1964 		}
1965 	}
1966 }
1967 
1968 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1969 {
1970 	struct ib_mad_agent_private *mad_agent_priv;
1971 	struct ib_mad_send_wr_private *temp_mad_send_wr;
1972 	struct list_head *list_item;
1973 	unsigned long delay;
1974 
1975 	mad_agent_priv = mad_send_wr->mad_agent_priv;
1976 	list_del(&mad_send_wr->agent_list);
1977 
1978 	delay = mad_send_wr->timeout;
1979 	mad_send_wr->timeout += jiffies;
1980 
1981 	if (delay) {
1982 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1983 			temp_mad_send_wr = list_entry(list_item,
1984 						struct ib_mad_send_wr_private,
1985 						agent_list);
1986 			if (time_after(mad_send_wr->timeout,
1987 				       temp_mad_send_wr->timeout))
1988 				break;
1989 		}
1990 	}
1991 	else
1992 		list_item = &mad_agent_priv->wait_list;
1993 	list_add(&mad_send_wr->agent_list, list_item);
1994 
1995 	/* Reschedule a work item if we have a shorter timeout */
1996 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1997 		cancel_delayed_work(&mad_agent_priv->timed_work);
1998 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1999 				   &mad_agent_priv->timed_work, delay);
2000 	}
2001 }
2002 
2003 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2004 			  int timeout_ms)
2005 {
2006 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2007 	wait_for_response(mad_send_wr);
2008 }
2009 
2010 /*
2011  * Process a send work completion
2012  */
2013 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2014 			     struct ib_mad_send_wc *mad_send_wc)
2015 {
2016 	struct ib_mad_agent_private	*mad_agent_priv;
2017 	unsigned long			flags;
2018 	int				ret;
2019 
2020 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2021 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2022 	if (mad_agent_priv->agent.rmpp_version) {
2023 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2024 		if (ret == IB_RMPP_RESULT_CONSUMED)
2025 			goto done;
2026 	} else
2027 		ret = IB_RMPP_RESULT_UNHANDLED;
2028 
2029 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2030 	    mad_send_wr->status == IB_WC_SUCCESS) {
2031 		mad_send_wr->status = mad_send_wc->status;
2032 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2033 	}
2034 
2035 	if (--mad_send_wr->refcount > 0) {
2036 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2037 		    mad_send_wr->status == IB_WC_SUCCESS) {
2038 			wait_for_response(mad_send_wr);
2039 		}
2040 		goto done;
2041 	}
2042 
2043 	/* Remove send from MAD agent and notify client of completion */
2044 	list_del(&mad_send_wr->agent_list);
2045 	adjust_timeout(mad_agent_priv);
2046 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2047 
2048 	if (mad_send_wr->status != IB_WC_SUCCESS )
2049 		mad_send_wc->status = mad_send_wr->status;
2050 	if (ret == IB_RMPP_RESULT_INTERNAL)
2051 		ib_rmpp_send_handler(mad_send_wc);
2052 	else
2053 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2054 						   mad_send_wc);
2055 
2056 	/* Release reference on agent taken when sending */
2057 	deref_mad_agent(mad_agent_priv);
2058 	return;
2059 done:
2060 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2061 }
2062 
2063 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2064 				     struct ib_wc *wc)
2065 {
2066 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2067 	struct ib_mad_list_head		*mad_list;
2068 	struct ib_mad_qp_info		*qp_info;
2069 	struct ib_mad_queue		*send_queue;
2070 	struct ib_send_wr		*bad_send_wr;
2071 	struct ib_mad_send_wc		mad_send_wc;
2072 	unsigned long flags;
2073 	int ret;
2074 
2075 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2076 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2077 				   mad_list);
2078 	send_queue = mad_list->mad_queue;
2079 	qp_info = send_queue->qp_info;
2080 
2081 retry:
2082 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2083 			 pci_unmap_addr(mad_send_wr, header_mapping),
2084 			 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2085 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
2086 			 pci_unmap_addr(mad_send_wr, payload_mapping),
2087 			 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2088 	queued_send_wr = NULL;
2089 	spin_lock_irqsave(&send_queue->lock, flags);
2090 	list_del(&mad_list->list);
2091 
2092 	/* Move queued send to the send queue */
2093 	if (send_queue->count-- > send_queue->max_active) {
2094 		mad_list = container_of(qp_info->overflow_list.next,
2095 					struct ib_mad_list_head, list);
2096 		queued_send_wr = container_of(mad_list,
2097 					struct ib_mad_send_wr_private,
2098 					mad_list);
2099 		list_move_tail(&mad_list->list, &send_queue->list);
2100 	}
2101 	spin_unlock_irqrestore(&send_queue->lock, flags);
2102 
2103 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2104 	mad_send_wc.status = wc->status;
2105 	mad_send_wc.vendor_err = wc->vendor_err;
2106 	if (atomic_read(&qp_info->snoop_count))
2107 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2108 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2109 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2110 
2111 	if (queued_send_wr) {
2112 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2113 				   &bad_send_wr);
2114 		if (ret) {
2115 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2116 			mad_send_wr = queued_send_wr;
2117 			wc->status = IB_WC_LOC_QP_OP_ERR;
2118 			goto retry;
2119 		}
2120 	}
2121 }
2122 
2123 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2124 {
2125 	struct ib_mad_send_wr_private *mad_send_wr;
2126 	struct ib_mad_list_head *mad_list;
2127 	unsigned long flags;
2128 
2129 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2130 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2131 		mad_send_wr = container_of(mad_list,
2132 					   struct ib_mad_send_wr_private,
2133 					   mad_list);
2134 		mad_send_wr->retry = 1;
2135 	}
2136 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2137 }
2138 
2139 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2140 			      struct ib_wc *wc)
2141 {
2142 	struct ib_mad_list_head *mad_list;
2143 	struct ib_mad_qp_info *qp_info;
2144 	struct ib_mad_send_wr_private *mad_send_wr;
2145 	int ret;
2146 
2147 	/* Determine if failure was a send or receive */
2148 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2149 	qp_info = mad_list->mad_queue->qp_info;
2150 	if (mad_list->mad_queue == &qp_info->recv_queue)
2151 		/*
2152 		 * Receive errors indicate that the QP has entered the error
2153 		 * state - error handling/shutdown code will cleanup
2154 		 */
2155 		return;
2156 
2157 	/*
2158 	 * Send errors will transition the QP to SQE - move
2159 	 * QP to RTS and repost flushed work requests
2160 	 */
2161 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2162 				   mad_list);
2163 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2164 		if (mad_send_wr->retry) {
2165 			/* Repost send */
2166 			struct ib_send_wr *bad_send_wr;
2167 
2168 			mad_send_wr->retry = 0;
2169 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2170 					&bad_send_wr);
2171 			if (ret)
2172 				ib_mad_send_done_handler(port_priv, wc);
2173 		} else
2174 			ib_mad_send_done_handler(port_priv, wc);
2175 	} else {
2176 		struct ib_qp_attr *attr;
2177 
2178 		/* Transition QP to RTS and fail offending send */
2179 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2180 		if (attr) {
2181 			attr->qp_state = IB_QPS_RTS;
2182 			attr->cur_qp_state = IB_QPS_SQE;
2183 			ret = ib_modify_qp(qp_info->qp, attr,
2184 					   IB_QP_STATE | IB_QP_CUR_STATE);
2185 			kfree(attr);
2186 			if (ret)
2187 				printk(KERN_ERR PFX "mad_error_handler - "
2188 				       "ib_modify_qp to RTS : %d\n", ret);
2189 			else
2190 				mark_sends_for_retry(qp_info);
2191 		}
2192 		ib_mad_send_done_handler(port_priv, wc);
2193 	}
2194 }
2195 
2196 /*
2197  * IB MAD completion callback
2198  */
2199 static void ib_mad_completion_handler(void *data)
2200 {
2201 	struct ib_mad_port_private *port_priv;
2202 	struct ib_wc wc;
2203 
2204 	port_priv = (struct ib_mad_port_private *)data;
2205 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2206 
2207 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2208 		if (wc.status == IB_WC_SUCCESS) {
2209 			switch (wc.opcode) {
2210 			case IB_WC_SEND:
2211 				ib_mad_send_done_handler(port_priv, &wc);
2212 				break;
2213 			case IB_WC_RECV:
2214 				ib_mad_recv_done_handler(port_priv, &wc);
2215 				break;
2216 			default:
2217 				BUG_ON(1);
2218 				break;
2219 			}
2220 		} else
2221 			mad_error_handler(port_priv, &wc);
2222 	}
2223 }
2224 
2225 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2226 {
2227 	unsigned long flags;
2228 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2229 	struct ib_mad_send_wc mad_send_wc;
2230 	struct list_head cancel_list;
2231 
2232 	INIT_LIST_HEAD(&cancel_list);
2233 
2234 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2235 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2236 				 &mad_agent_priv->send_list, agent_list) {
2237 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2238  			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2239 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2240 		}
2241 	}
2242 
2243 	/* Empty wait list to prevent receives from finding a request */
2244 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2245 	/* Empty local completion list as well */
2246 	list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2247 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2248 
2249 	/* Report all cancelled requests */
2250 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2251 	mad_send_wc.vendor_err = 0;
2252 
2253 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2254 				 &cancel_list, agent_list) {
2255 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2256 		list_del(&mad_send_wr->agent_list);
2257 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2258 						   &mad_send_wc);
2259 		atomic_dec(&mad_agent_priv->refcount);
2260 	}
2261 }
2262 
2263 static struct ib_mad_send_wr_private*
2264 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2265 	     struct ib_mad_send_buf *send_buf)
2266 {
2267 	struct ib_mad_send_wr_private *mad_send_wr;
2268 
2269 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2270 			    agent_list) {
2271 		if (&mad_send_wr->send_buf == send_buf)
2272 			return mad_send_wr;
2273 	}
2274 
2275 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2276 			    agent_list) {
2277 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2278 		    &mad_send_wr->send_buf == send_buf)
2279 			return mad_send_wr;
2280 	}
2281 	return NULL;
2282 }
2283 
2284 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2285 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2286 {
2287 	struct ib_mad_agent_private *mad_agent_priv;
2288 	struct ib_mad_send_wr_private *mad_send_wr;
2289 	unsigned long flags;
2290 	int active;
2291 
2292 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2293 				      agent);
2294 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2295 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2296 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2297 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2298 		return -EINVAL;
2299 	}
2300 
2301 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2302 	if (!timeout_ms) {
2303 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2304 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2305 	}
2306 
2307 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2308 	if (active)
2309 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2310 	else
2311 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2312 
2313 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2314 	return 0;
2315 }
2316 EXPORT_SYMBOL(ib_modify_mad);
2317 
2318 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2319 		   struct ib_mad_send_buf *send_buf)
2320 {
2321 	ib_modify_mad(mad_agent, send_buf, 0);
2322 }
2323 EXPORT_SYMBOL(ib_cancel_mad);
2324 
2325 static void local_completions(void *data)
2326 {
2327 	struct ib_mad_agent_private *mad_agent_priv;
2328 	struct ib_mad_local_private *local;
2329 	struct ib_mad_agent_private *recv_mad_agent;
2330 	unsigned long flags;
2331 	int recv = 0;
2332 	struct ib_wc wc;
2333 	struct ib_mad_send_wc mad_send_wc;
2334 
2335 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2336 
2337 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2338 	while (!list_empty(&mad_agent_priv->local_list)) {
2339 		local = list_entry(mad_agent_priv->local_list.next,
2340 				   struct ib_mad_local_private,
2341 				   completion_list);
2342 		list_del(&local->completion_list);
2343 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2344 		if (local->mad_priv) {
2345 			recv_mad_agent = local->recv_mad_agent;
2346 			if (!recv_mad_agent) {
2347 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2348 				goto local_send_completion;
2349 			}
2350 
2351 			recv = 1;
2352 			/*
2353 			 * Defined behavior is to complete response
2354 			 * before request
2355 			 */
2356 			build_smp_wc((unsigned long) local->mad_send_wr,
2357 				     be16_to_cpu(IB_LID_PERMISSIVE),
2358 				     0, recv_mad_agent->agent.port_num, &wc);
2359 
2360 			local->mad_priv->header.recv_wc.wc = &wc;
2361 			local->mad_priv->header.recv_wc.mad_len =
2362 						sizeof(struct ib_mad);
2363 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2364 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2365 				 &local->mad_priv->header.recv_wc.rmpp_list);
2366 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2367 			local->mad_priv->header.recv_wc.recv_buf.mad =
2368 						&local->mad_priv->mad.mad;
2369 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2370 				snoop_recv(recv_mad_agent->qp_info,
2371 					  &local->mad_priv->header.recv_wc,
2372 					   IB_MAD_SNOOP_RECVS);
2373 			recv_mad_agent->agent.recv_handler(
2374 						&recv_mad_agent->agent,
2375 						&local->mad_priv->header.recv_wc);
2376 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2377 			atomic_dec(&recv_mad_agent->refcount);
2378 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2379 		}
2380 
2381 local_send_completion:
2382 		/* Complete send */
2383 		mad_send_wc.status = IB_WC_SUCCESS;
2384 		mad_send_wc.vendor_err = 0;
2385 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2386 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2387 			snoop_send(mad_agent_priv->qp_info,
2388 				   &local->mad_send_wr->send_buf,
2389 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2390 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2391 						   &mad_send_wc);
2392 
2393 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2394 		atomic_dec(&mad_agent_priv->refcount);
2395 		if (!recv)
2396 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2397 		kfree(local);
2398 	}
2399 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2400 }
2401 
2402 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2403 {
2404 	int ret;
2405 
2406 	if (!mad_send_wr->retries--)
2407 		return -ETIMEDOUT;
2408 
2409 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2410 
2411 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2412 		ret = ib_retry_rmpp(mad_send_wr);
2413 		switch (ret) {
2414 		case IB_RMPP_RESULT_UNHANDLED:
2415 			ret = ib_send_mad(mad_send_wr);
2416 			break;
2417 		case IB_RMPP_RESULT_CONSUMED:
2418 			ret = 0;
2419 			break;
2420 		default:
2421 			ret = -ECOMM;
2422 			break;
2423 		}
2424 	} else
2425 		ret = ib_send_mad(mad_send_wr);
2426 
2427 	if (!ret) {
2428 		mad_send_wr->refcount++;
2429 		list_add_tail(&mad_send_wr->agent_list,
2430 			      &mad_send_wr->mad_agent_priv->send_list);
2431 	}
2432 	return ret;
2433 }
2434 
2435 static void timeout_sends(void *data)
2436 {
2437 	struct ib_mad_agent_private *mad_agent_priv;
2438 	struct ib_mad_send_wr_private *mad_send_wr;
2439 	struct ib_mad_send_wc mad_send_wc;
2440 	unsigned long flags, delay;
2441 
2442 	mad_agent_priv = (struct ib_mad_agent_private *)data;
2443 	mad_send_wc.vendor_err = 0;
2444 
2445 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2446 	while (!list_empty(&mad_agent_priv->wait_list)) {
2447 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2448 					 struct ib_mad_send_wr_private,
2449 					 agent_list);
2450 
2451 		if (time_after(mad_send_wr->timeout, jiffies)) {
2452 			delay = mad_send_wr->timeout - jiffies;
2453 			if ((long)delay <= 0)
2454 				delay = 1;
2455 			queue_delayed_work(mad_agent_priv->qp_info->
2456 					   port_priv->wq,
2457 					   &mad_agent_priv->timed_work, delay);
2458 			break;
2459 		}
2460 
2461 		list_del(&mad_send_wr->agent_list);
2462 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2463 		    !retry_send(mad_send_wr))
2464 			continue;
2465 
2466 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2467 
2468 		if (mad_send_wr->status == IB_WC_SUCCESS)
2469 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2470 		else
2471 			mad_send_wc.status = mad_send_wr->status;
2472 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2473 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2474 						   &mad_send_wc);
2475 
2476 		atomic_dec(&mad_agent_priv->refcount);
2477 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2478 	}
2479 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2480 }
2481 
2482 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2483 {
2484 	struct ib_mad_port_private *port_priv = cq->cq_context;
2485 	unsigned long flags;
2486 
2487 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2488 	if (!list_empty(&port_priv->port_list))
2489 		queue_work(port_priv->wq, &port_priv->work);
2490 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2491 }
2492 
2493 /*
2494  * Allocate receive MADs and post receive WRs for them
2495  */
2496 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2497 				    struct ib_mad_private *mad)
2498 {
2499 	unsigned long flags;
2500 	int post, ret;
2501 	struct ib_mad_private *mad_priv;
2502 	struct ib_sge sg_list;
2503 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2504 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2505 
2506 	/* Initialize common scatter list fields */
2507 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2508 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2509 
2510 	/* Initialize common receive WR fields */
2511 	recv_wr.next = NULL;
2512 	recv_wr.sg_list = &sg_list;
2513 	recv_wr.num_sge = 1;
2514 
2515 	do {
2516 		/* Allocate and map receive buffer */
2517 		if (mad) {
2518 			mad_priv = mad;
2519 			mad = NULL;
2520 		} else {
2521 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2522 			if (!mad_priv) {
2523 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2524 				ret = -ENOMEM;
2525 				break;
2526 			}
2527 		}
2528 		sg_list.addr = dma_map_single(qp_info->port_priv->
2529 					      	device->dma_device,
2530 					      &mad_priv->grh,
2531 					      sizeof *mad_priv -
2532 					      	sizeof mad_priv->header,
2533 					      DMA_FROM_DEVICE);
2534 		pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2535 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2536 		mad_priv->header.mad_list.mad_queue = recv_queue;
2537 
2538 		/* Post receive WR */
2539 		spin_lock_irqsave(&recv_queue->lock, flags);
2540 		post = (++recv_queue->count < recv_queue->max_active);
2541 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2542 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2543 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2544 		if (ret) {
2545 			spin_lock_irqsave(&recv_queue->lock, flags);
2546 			list_del(&mad_priv->header.mad_list.list);
2547 			recv_queue->count--;
2548 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2549 			dma_unmap_single(qp_info->port_priv->device->dma_device,
2550 					 pci_unmap_addr(&mad_priv->header,
2551 							mapping),
2552 					 sizeof *mad_priv -
2553 					   sizeof mad_priv->header,
2554 					 DMA_FROM_DEVICE);
2555 			kmem_cache_free(ib_mad_cache, mad_priv);
2556 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2557 			break;
2558 		}
2559 	} while (post);
2560 
2561 	return ret;
2562 }
2563 
2564 /*
2565  * Return all the posted receive MADs
2566  */
2567 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2568 {
2569 	struct ib_mad_private_header *mad_priv_hdr;
2570 	struct ib_mad_private *recv;
2571 	struct ib_mad_list_head *mad_list;
2572 
2573 	while (!list_empty(&qp_info->recv_queue.list)) {
2574 
2575 		mad_list = list_entry(qp_info->recv_queue.list.next,
2576 				      struct ib_mad_list_head, list);
2577 		mad_priv_hdr = container_of(mad_list,
2578 					    struct ib_mad_private_header,
2579 					    mad_list);
2580 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2581 				    header);
2582 
2583 		/* Remove from posted receive MAD list */
2584 		list_del(&mad_list->list);
2585 
2586 		dma_unmap_single(qp_info->port_priv->device->dma_device,
2587 				 pci_unmap_addr(&recv->header, mapping),
2588 				 sizeof(struct ib_mad_private) -
2589 				 sizeof(struct ib_mad_private_header),
2590 				 DMA_FROM_DEVICE);
2591 		kmem_cache_free(ib_mad_cache, recv);
2592 	}
2593 
2594 	qp_info->recv_queue.count = 0;
2595 }
2596 
2597 /*
2598  * Start the port
2599  */
2600 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2601 {
2602 	int ret, i;
2603 	struct ib_qp_attr *attr;
2604 	struct ib_qp *qp;
2605 
2606 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2607  	if (!attr) {
2608 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2609 		return -ENOMEM;
2610 	}
2611 
2612 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2613 		qp = port_priv->qp_info[i].qp;
2614 		/*
2615 		 * PKey index for QP1 is irrelevant but
2616 		 * one is needed for the Reset to Init transition
2617 		 */
2618 		attr->qp_state = IB_QPS_INIT;
2619 		attr->pkey_index = 0;
2620 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2621 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2622 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2623 		if (ret) {
2624 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2625 			       "INIT: %d\n", i, ret);
2626 			goto out;
2627 		}
2628 
2629 		attr->qp_state = IB_QPS_RTR;
2630 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2631 		if (ret) {
2632 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2633 			       "RTR: %d\n", i, ret);
2634 			goto out;
2635 		}
2636 
2637 		attr->qp_state = IB_QPS_RTS;
2638 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2639 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2640 		if (ret) {
2641 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2642 			       "RTS: %d\n", i, ret);
2643 			goto out;
2644 		}
2645 	}
2646 
2647 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2648 	if (ret) {
2649 		printk(KERN_ERR PFX "Failed to request completion "
2650 		       "notification: %d\n", ret);
2651 		goto out;
2652 	}
2653 
2654 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2655 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2656 		if (ret) {
2657 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2658 			goto out;
2659 		}
2660 	}
2661 out:
2662 	kfree(attr);
2663 	return ret;
2664 }
2665 
2666 static void qp_event_handler(struct ib_event *event, void *qp_context)
2667 {
2668 	struct ib_mad_qp_info	*qp_info = qp_context;
2669 
2670 	/* It's worse than that! He's dead, Jim! */
2671 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2672 		event->event, qp_info->qp->qp_num);
2673 }
2674 
2675 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2676 			   struct ib_mad_queue *mad_queue)
2677 {
2678 	mad_queue->qp_info = qp_info;
2679 	mad_queue->count = 0;
2680 	spin_lock_init(&mad_queue->lock);
2681 	INIT_LIST_HEAD(&mad_queue->list);
2682 }
2683 
2684 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2685 			struct ib_mad_qp_info *qp_info)
2686 {
2687 	qp_info->port_priv = port_priv;
2688 	init_mad_queue(qp_info, &qp_info->send_queue);
2689 	init_mad_queue(qp_info, &qp_info->recv_queue);
2690 	INIT_LIST_HEAD(&qp_info->overflow_list);
2691 	spin_lock_init(&qp_info->snoop_lock);
2692 	qp_info->snoop_table = NULL;
2693 	qp_info->snoop_table_size = 0;
2694 	atomic_set(&qp_info->snoop_count, 0);
2695 }
2696 
2697 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2698 			 enum ib_qp_type qp_type)
2699 {
2700 	struct ib_qp_init_attr	qp_init_attr;
2701 	int ret;
2702 
2703 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2704 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2705 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2706 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2707 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2708 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2709 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2710 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2711 	qp_init_attr.qp_type = qp_type;
2712 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2713 	qp_init_attr.qp_context = qp_info;
2714 	qp_init_attr.event_handler = qp_event_handler;
2715 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2716 	if (IS_ERR(qp_info->qp)) {
2717 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2718 		       get_spl_qp_index(qp_type));
2719 		ret = PTR_ERR(qp_info->qp);
2720 		goto error;
2721 	}
2722 	/* Use minimum queue sizes unless the CQ is resized */
2723 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2724 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2725 	return 0;
2726 
2727 error:
2728 	return ret;
2729 }
2730 
2731 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2732 {
2733 	ib_destroy_qp(qp_info->qp);
2734 	kfree(qp_info->snoop_table);
2735 }
2736 
2737 /*
2738  * Open the port
2739  * Create the QP, PD, MR, and CQ if needed
2740  */
2741 static int ib_mad_port_open(struct ib_device *device,
2742 			    int port_num)
2743 {
2744 	int ret, cq_size;
2745 	struct ib_mad_port_private *port_priv;
2746 	unsigned long flags;
2747 	char name[sizeof "ib_mad123"];
2748 
2749 	/* Create new device info */
2750 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2751 	if (!port_priv) {
2752 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2753 		return -ENOMEM;
2754 	}
2755 
2756 	port_priv->device = device;
2757 	port_priv->port_num = port_num;
2758 	spin_lock_init(&port_priv->reg_lock);
2759 	INIT_LIST_HEAD(&port_priv->agent_list);
2760 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2761 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2762 
2763 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2764 	port_priv->cq = ib_create_cq(port_priv->device,
2765 				     ib_mad_thread_completion_handler,
2766 				     NULL, port_priv, cq_size);
2767 	if (IS_ERR(port_priv->cq)) {
2768 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2769 		ret = PTR_ERR(port_priv->cq);
2770 		goto error3;
2771 	}
2772 
2773 	port_priv->pd = ib_alloc_pd(device);
2774 	if (IS_ERR(port_priv->pd)) {
2775 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2776 		ret = PTR_ERR(port_priv->pd);
2777 		goto error4;
2778 	}
2779 
2780 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2781 	if (IS_ERR(port_priv->mr)) {
2782 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2783 		ret = PTR_ERR(port_priv->mr);
2784 		goto error5;
2785 	}
2786 
2787 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2788 	if (ret)
2789 		goto error6;
2790 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2791 	if (ret)
2792 		goto error7;
2793 
2794 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2795 	port_priv->wq = create_singlethread_workqueue(name);
2796 	if (!port_priv->wq) {
2797 		ret = -ENOMEM;
2798 		goto error8;
2799 	}
2800 	INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2801 
2802 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2803 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2804 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2805 
2806 	ret = ib_mad_port_start(port_priv);
2807 	if (ret) {
2808 		printk(KERN_ERR PFX "Couldn't start port\n");
2809 		goto error9;
2810 	}
2811 
2812 	return 0;
2813 
2814 error9:
2815 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2816 	list_del_init(&port_priv->port_list);
2817 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2818 
2819 	destroy_workqueue(port_priv->wq);
2820 error8:
2821 	destroy_mad_qp(&port_priv->qp_info[1]);
2822 error7:
2823 	destroy_mad_qp(&port_priv->qp_info[0]);
2824 error6:
2825 	ib_dereg_mr(port_priv->mr);
2826 error5:
2827 	ib_dealloc_pd(port_priv->pd);
2828 error4:
2829 	ib_destroy_cq(port_priv->cq);
2830 	cleanup_recv_queue(&port_priv->qp_info[1]);
2831 	cleanup_recv_queue(&port_priv->qp_info[0]);
2832 error3:
2833 	kfree(port_priv);
2834 
2835 	return ret;
2836 }
2837 
2838 /*
2839  * Close the port
2840  * If there are no classes using the port, free the port
2841  * resources (CQ, MR, PD, QP) and remove the port's info structure
2842  */
2843 static int ib_mad_port_close(struct ib_device *device, int port_num)
2844 {
2845 	struct ib_mad_port_private *port_priv;
2846 	unsigned long flags;
2847 
2848 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2849 	port_priv = __ib_get_mad_port(device, port_num);
2850 	if (port_priv == NULL) {
2851 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2852 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2853 		return -ENODEV;
2854 	}
2855 	list_del_init(&port_priv->port_list);
2856 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2857 
2858 	destroy_workqueue(port_priv->wq);
2859 	destroy_mad_qp(&port_priv->qp_info[1]);
2860 	destroy_mad_qp(&port_priv->qp_info[0]);
2861 	ib_dereg_mr(port_priv->mr);
2862 	ib_dealloc_pd(port_priv->pd);
2863 	ib_destroy_cq(port_priv->cq);
2864 	cleanup_recv_queue(&port_priv->qp_info[1]);
2865 	cleanup_recv_queue(&port_priv->qp_info[0]);
2866 	/* XXX: Handle deallocation of MAD registration tables */
2867 
2868 	kfree(port_priv);
2869 
2870 	return 0;
2871 }
2872 
2873 static void ib_mad_init_device(struct ib_device *device)
2874 {
2875 	int start, end, i;
2876 
2877 	if (device->node_type == IB_NODE_SWITCH) {
2878 		start = 0;
2879 		end   = 0;
2880 	} else {
2881 		start = 1;
2882 		end   = device->phys_port_cnt;
2883 	}
2884 
2885 	for (i = start; i <= end; i++) {
2886 		if (ib_mad_port_open(device, i)) {
2887 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2888 			       device->name, i);
2889 			goto error;
2890 		}
2891 		if (ib_agent_port_open(device, i)) {
2892 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2893 			       "for agents\n",
2894 			       device->name, i);
2895 			goto error_agent;
2896 		}
2897 	}
2898 	return;
2899 
2900 error_agent:
2901 	if (ib_mad_port_close(device, i))
2902 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2903 		       device->name, i);
2904 
2905 error:
2906 	i--;
2907 
2908 	while (i >= start) {
2909 		if (ib_agent_port_close(device, i))
2910 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2911 			       "for agents\n",
2912 			       device->name, i);
2913 		if (ib_mad_port_close(device, i))
2914 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2915 			       device->name, i);
2916 		i--;
2917 	}
2918 }
2919 
2920 static void ib_mad_remove_device(struct ib_device *device)
2921 {
2922 	int i, num_ports, cur_port;
2923 
2924 	if (device->node_type == IB_NODE_SWITCH) {
2925 		num_ports = 1;
2926 		cur_port = 0;
2927 	} else {
2928 		num_ports = device->phys_port_cnt;
2929 		cur_port = 1;
2930 	}
2931 	for (i = 0; i < num_ports; i++, cur_port++) {
2932 		if (ib_agent_port_close(device, cur_port))
2933 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2934 			       "for agents\n",
2935 			       device->name, cur_port);
2936 		if (ib_mad_port_close(device, cur_port))
2937 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2938 			       device->name, cur_port);
2939 	}
2940 }
2941 
2942 static struct ib_client mad_client = {
2943 	.name   = "mad",
2944 	.add = ib_mad_init_device,
2945 	.remove = ib_mad_remove_device
2946 };
2947 
2948 static int __init ib_mad_init_module(void)
2949 {
2950 	int ret;
2951 
2952 	spin_lock_init(&ib_mad_port_list_lock);
2953 
2954 	ib_mad_cache = kmem_cache_create("ib_mad",
2955 					 sizeof(struct ib_mad_private),
2956 					 0,
2957 					 SLAB_HWCACHE_ALIGN,
2958 					 NULL,
2959 					 NULL);
2960 	if (!ib_mad_cache) {
2961 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2962 		ret = -ENOMEM;
2963 		goto error1;
2964 	}
2965 
2966 	INIT_LIST_HEAD(&ib_mad_port_list);
2967 
2968 	if (ib_register_client(&mad_client)) {
2969 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2970 		ret = -EINVAL;
2971 		goto error2;
2972 	}
2973 
2974 	return 0;
2975 
2976 error2:
2977 	kmem_cache_destroy(ib_mad_cache);
2978 error1:
2979 	return ret;
2980 }
2981 
2982 static void __exit ib_mad_cleanup_module(void)
2983 {
2984 	ib_unregister_client(&mad_client);
2985 
2986 	if (kmem_cache_destroy(ib_mad_cache)) {
2987 		printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2988 	}
2989 }
2990 
2991 module_init(ib_mad_init_module);
2992 module_exit(ib_mad_cleanup_module);
2993 
2994