xref: /linux/drivers/infiniband/core/mad.c (revision cfbf1eecd70db9a7a49c42a0613c00f7a2a86dfb)
1 /*
2  * Copyright (c) 2004-2007 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  */
35 #include <linux/dma-mapping.h>
36 #include <rdma/ib_cache.h>
37 
38 #include "mad_priv.h"
39 #include "mad_rmpp.h"
40 #include "smi.h"
41 #include "agent.h"
42 
43 MODULE_LICENSE("Dual BSD/GPL");
44 MODULE_DESCRIPTION("kernel IB MAD API");
45 MODULE_AUTHOR("Hal Rosenstock");
46 MODULE_AUTHOR("Sean Hefty");
47 
48 static struct kmem_cache *ib_mad_cache;
49 
50 static struct list_head ib_mad_port_list;
51 static u32 ib_mad_client_id = 0;
52 
53 /* Port list lock */
54 static spinlock_t ib_mad_port_list_lock;
55 
56 
57 /* Forward declarations */
58 static int method_in_use(struct ib_mad_mgmt_method_table **method,
59 			 struct ib_mad_reg_req *mad_reg_req);
60 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
61 static struct ib_mad_agent_private *find_mad_agent(
62 					struct ib_mad_port_private *port_priv,
63 					struct ib_mad *mad);
64 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
65 				    struct ib_mad_private *mad);
66 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
67 static void timeout_sends(struct work_struct *work);
68 static void local_completions(struct work_struct *work);
69 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
70 			      struct ib_mad_agent_private *agent_priv,
71 			      u8 mgmt_class);
72 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
73 			   struct ib_mad_agent_private *agent_priv);
74 
75 /*
76  * Returns a ib_mad_port_private structure or NULL for a device/port
77  * Assumes ib_mad_port_list_lock is being held
78  */
79 static inline struct ib_mad_port_private *
80 __ib_get_mad_port(struct ib_device *device, int port_num)
81 {
82 	struct ib_mad_port_private *entry;
83 
84 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
85 		if (entry->device == device && entry->port_num == port_num)
86 			return entry;
87 	}
88 	return NULL;
89 }
90 
91 /*
92  * Wrapper function to return a ib_mad_port_private structure or NULL
93  * for a device/port
94  */
95 static inline struct ib_mad_port_private *
96 ib_get_mad_port(struct ib_device *device, int port_num)
97 {
98 	struct ib_mad_port_private *entry;
99 	unsigned long flags;
100 
101 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
102 	entry = __ib_get_mad_port(device, port_num);
103 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
104 
105 	return entry;
106 }
107 
108 static inline u8 convert_mgmt_class(u8 mgmt_class)
109 {
110 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
111 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
112 		0 : mgmt_class;
113 }
114 
115 static int get_spl_qp_index(enum ib_qp_type qp_type)
116 {
117 	switch (qp_type)
118 	{
119 	case IB_QPT_SMI:
120 		return 0;
121 	case IB_QPT_GSI:
122 		return 1;
123 	default:
124 		return -1;
125 	}
126 }
127 
128 static int vendor_class_index(u8 mgmt_class)
129 {
130 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
131 }
132 
133 static int is_vendor_class(u8 mgmt_class)
134 {
135 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
136 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
137 		return 0;
138 	return 1;
139 }
140 
141 static int is_vendor_oui(char *oui)
142 {
143 	if (oui[0] || oui[1] || oui[2])
144 		return 1;
145 	return 0;
146 }
147 
148 static int is_vendor_method_in_use(
149 		struct ib_mad_mgmt_vendor_class *vendor_class,
150 		struct ib_mad_reg_req *mad_reg_req)
151 {
152 	struct ib_mad_mgmt_method_table *method;
153 	int i;
154 
155 	for (i = 0; i < MAX_MGMT_OUI; i++) {
156 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
157 			method = vendor_class->method_table[i];
158 			if (method) {
159 				if (method_in_use(&method, mad_reg_req))
160 					return 1;
161 				else
162 					break;
163 			}
164 		}
165 	}
166 	return 0;
167 }
168 
169 int ib_response_mad(struct ib_mad *mad)
170 {
171 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
172 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
173 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
174 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
175 }
176 EXPORT_SYMBOL(ib_response_mad);
177 
178 /*
179  * ib_register_mad_agent - Register to send/receive MADs
180  */
181 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
182 					   u8 port_num,
183 					   enum ib_qp_type qp_type,
184 					   struct ib_mad_reg_req *mad_reg_req,
185 					   u8 rmpp_version,
186 					   ib_mad_send_handler send_handler,
187 					   ib_mad_recv_handler recv_handler,
188 					   void *context)
189 {
190 	struct ib_mad_port_private *port_priv;
191 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
192 	struct ib_mad_agent_private *mad_agent_priv;
193 	struct ib_mad_reg_req *reg_req = NULL;
194 	struct ib_mad_mgmt_class_table *class;
195 	struct ib_mad_mgmt_vendor_class_table *vendor;
196 	struct ib_mad_mgmt_vendor_class *vendor_class;
197 	struct ib_mad_mgmt_method_table *method;
198 	int ret2, qpn;
199 	unsigned long flags;
200 	u8 mgmt_class, vclass;
201 
202 	/* Validate parameters */
203 	qpn = get_spl_qp_index(qp_type);
204 	if (qpn == -1)
205 		goto error1;
206 
207 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
208 		goto error1;
209 
210 	/* Validate MAD registration request if supplied */
211 	if (mad_reg_req) {
212 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
213 			goto error1;
214 		if (!recv_handler)
215 			goto error1;
216 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
217 			/*
218 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
219 			 * one in this range currently allowed
220 			 */
221 			if (mad_reg_req->mgmt_class !=
222 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
223 				goto error1;
224 		} else if (mad_reg_req->mgmt_class == 0) {
225 			/*
226 			 * Class 0 is reserved in IBA and is used for
227 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
228 			 */
229 			goto error1;
230 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
231 			/*
232 			 * If class is in "new" vendor range,
233 			 * ensure supplied OUI is not zero
234 			 */
235 			if (!is_vendor_oui(mad_reg_req->oui))
236 				goto error1;
237 		}
238 		/* Make sure class supplied is consistent with RMPP */
239 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
240 			if (rmpp_version)
241 				goto error1;
242 		}
243 		/* Make sure class supplied is consistent with QP type */
244 		if (qp_type == IB_QPT_SMI) {
245 			if ((mad_reg_req->mgmt_class !=
246 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
247 			    (mad_reg_req->mgmt_class !=
248 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
249 				goto error1;
250 		} else {
251 			if ((mad_reg_req->mgmt_class ==
252 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
253 			    (mad_reg_req->mgmt_class ==
254 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
255 				goto error1;
256 		}
257 	} else {
258 		/* No registration request supplied */
259 		if (!send_handler)
260 			goto error1;
261 	}
262 
263 	/* Validate device and port */
264 	port_priv = ib_get_mad_port(device, port_num);
265 	if (!port_priv) {
266 		ret = ERR_PTR(-ENODEV);
267 		goto error1;
268 	}
269 
270 	/* Allocate structures */
271 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
272 	if (!mad_agent_priv) {
273 		ret = ERR_PTR(-ENOMEM);
274 		goto error1;
275 	}
276 
277 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
278 						 IB_ACCESS_LOCAL_WRITE);
279 	if (IS_ERR(mad_agent_priv->agent.mr)) {
280 		ret = ERR_PTR(-ENOMEM);
281 		goto error2;
282 	}
283 
284 	if (mad_reg_req) {
285 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
286 		if (!reg_req) {
287 			ret = ERR_PTR(-ENOMEM);
288 			goto error3;
289 		}
290 		/* Make a copy of the MAD registration request */
291 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
292 	}
293 
294 	/* Now, fill in the various structures */
295 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
296 	mad_agent_priv->reg_req = reg_req;
297 	mad_agent_priv->agent.rmpp_version = rmpp_version;
298 	mad_agent_priv->agent.device = device;
299 	mad_agent_priv->agent.recv_handler = recv_handler;
300 	mad_agent_priv->agent.send_handler = send_handler;
301 	mad_agent_priv->agent.context = context;
302 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
303 	mad_agent_priv->agent.port_num = port_num;
304 	spin_lock_init(&mad_agent_priv->lock);
305 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
306 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
307 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
308 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
309 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
310 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
311 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
312 	atomic_set(&mad_agent_priv->refcount, 1);
313 	init_completion(&mad_agent_priv->comp);
314 
315 	spin_lock_irqsave(&port_priv->reg_lock, flags);
316 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
317 
318 	/*
319 	 * Make sure MAD registration (if supplied)
320 	 * is non overlapping with any existing ones
321 	 */
322 	if (mad_reg_req) {
323 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
324 		if (!is_vendor_class(mgmt_class)) {
325 			class = port_priv->version[mad_reg_req->
326 						   mgmt_class_version].class;
327 			if (class) {
328 				method = class->method_table[mgmt_class];
329 				if (method) {
330 					if (method_in_use(&method,
331 							   mad_reg_req))
332 						goto error4;
333 				}
334 			}
335 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
336 						  mgmt_class);
337 		} else {
338 			/* "New" vendor class range */
339 			vendor = port_priv->version[mad_reg_req->
340 						    mgmt_class_version].vendor;
341 			if (vendor) {
342 				vclass = vendor_class_index(mgmt_class);
343 				vendor_class = vendor->vendor_class[vclass];
344 				if (vendor_class) {
345 					if (is_vendor_method_in_use(
346 							vendor_class,
347 							mad_reg_req))
348 						goto error4;
349 				}
350 			}
351 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
352 		}
353 		if (ret2) {
354 			ret = ERR_PTR(ret2);
355 			goto error4;
356 		}
357 	}
358 
359 	/* Add mad agent into port's agent list */
360 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
361 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
362 
363 	return &mad_agent_priv->agent;
364 
365 error4:
366 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
367 	kfree(reg_req);
368 error3:
369 	ib_dereg_mr(mad_agent_priv->agent.mr);
370 error2:
371 	kfree(mad_agent_priv);
372 error1:
373 	return ret;
374 }
375 EXPORT_SYMBOL(ib_register_mad_agent);
376 
377 static inline int is_snooping_sends(int mad_snoop_flags)
378 {
379 	return (mad_snoop_flags &
380 		(/*IB_MAD_SNOOP_POSTED_SENDS |
381 		 IB_MAD_SNOOP_RMPP_SENDS |*/
382 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
383 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
384 }
385 
386 static inline int is_snooping_recvs(int mad_snoop_flags)
387 {
388 	return (mad_snoop_flags &
389 		(IB_MAD_SNOOP_RECVS /*|
390 		 IB_MAD_SNOOP_RMPP_RECVS*/));
391 }
392 
393 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
394 				struct ib_mad_snoop_private *mad_snoop_priv)
395 {
396 	struct ib_mad_snoop_private **new_snoop_table;
397 	unsigned long flags;
398 	int i;
399 
400 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
401 	/* Check for empty slot in array. */
402 	for (i = 0; i < qp_info->snoop_table_size; i++)
403 		if (!qp_info->snoop_table[i])
404 			break;
405 
406 	if (i == qp_info->snoop_table_size) {
407 		/* Grow table. */
408 		new_snoop_table = krealloc(qp_info->snoop_table,
409 					   sizeof mad_snoop_priv *
410 					   (qp_info->snoop_table_size + 1),
411 					   GFP_ATOMIC);
412 		if (!new_snoop_table) {
413 			i = -ENOMEM;
414 			goto out;
415 		}
416 
417 		qp_info->snoop_table = new_snoop_table;
418 		qp_info->snoop_table_size++;
419 	}
420 	qp_info->snoop_table[i] = mad_snoop_priv;
421 	atomic_inc(&qp_info->snoop_count);
422 out:
423 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
424 	return i;
425 }
426 
427 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
428 					   u8 port_num,
429 					   enum ib_qp_type qp_type,
430 					   int mad_snoop_flags,
431 					   ib_mad_snoop_handler snoop_handler,
432 					   ib_mad_recv_handler recv_handler,
433 					   void *context)
434 {
435 	struct ib_mad_port_private *port_priv;
436 	struct ib_mad_agent *ret;
437 	struct ib_mad_snoop_private *mad_snoop_priv;
438 	int qpn;
439 
440 	/* Validate parameters */
441 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
442 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
443 		ret = ERR_PTR(-EINVAL);
444 		goto error1;
445 	}
446 	qpn = get_spl_qp_index(qp_type);
447 	if (qpn == -1) {
448 		ret = ERR_PTR(-EINVAL);
449 		goto error1;
450 	}
451 	port_priv = ib_get_mad_port(device, port_num);
452 	if (!port_priv) {
453 		ret = ERR_PTR(-ENODEV);
454 		goto error1;
455 	}
456 	/* Allocate structures */
457 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
458 	if (!mad_snoop_priv) {
459 		ret = ERR_PTR(-ENOMEM);
460 		goto error1;
461 	}
462 
463 	/* Now, fill in the various structures */
464 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
465 	mad_snoop_priv->agent.device = device;
466 	mad_snoop_priv->agent.recv_handler = recv_handler;
467 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
468 	mad_snoop_priv->agent.context = context;
469 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
470 	mad_snoop_priv->agent.port_num = port_num;
471 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
472 	init_completion(&mad_snoop_priv->comp);
473 	mad_snoop_priv->snoop_index = register_snoop_agent(
474 						&port_priv->qp_info[qpn],
475 						mad_snoop_priv);
476 	if (mad_snoop_priv->snoop_index < 0) {
477 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
478 		goto error2;
479 	}
480 
481 	atomic_set(&mad_snoop_priv->refcount, 1);
482 	return &mad_snoop_priv->agent;
483 
484 error2:
485 	kfree(mad_snoop_priv);
486 error1:
487 	return ret;
488 }
489 EXPORT_SYMBOL(ib_register_mad_snoop);
490 
491 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
492 {
493 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
494 		complete(&mad_agent_priv->comp);
495 }
496 
497 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
498 {
499 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
500 		complete(&mad_snoop_priv->comp);
501 }
502 
503 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
504 {
505 	struct ib_mad_port_private *port_priv;
506 	unsigned long flags;
507 
508 	/* Note that we could still be handling received MADs */
509 
510 	/*
511 	 * Canceling all sends results in dropping received response
512 	 * MADs, preventing us from queuing additional work
513 	 */
514 	cancel_mads(mad_agent_priv);
515 	port_priv = mad_agent_priv->qp_info->port_priv;
516 	cancel_delayed_work(&mad_agent_priv->timed_work);
517 
518 	spin_lock_irqsave(&port_priv->reg_lock, flags);
519 	remove_mad_reg_req(mad_agent_priv);
520 	list_del(&mad_agent_priv->agent_list);
521 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
522 
523 	flush_workqueue(port_priv->wq);
524 	ib_cancel_rmpp_recvs(mad_agent_priv);
525 
526 	deref_mad_agent(mad_agent_priv);
527 	wait_for_completion(&mad_agent_priv->comp);
528 
529 	kfree(mad_agent_priv->reg_req);
530 	ib_dereg_mr(mad_agent_priv->agent.mr);
531 	kfree(mad_agent_priv);
532 }
533 
534 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
535 {
536 	struct ib_mad_qp_info *qp_info;
537 	unsigned long flags;
538 
539 	qp_info = mad_snoop_priv->qp_info;
540 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
541 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
542 	atomic_dec(&qp_info->snoop_count);
543 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
544 
545 	deref_snoop_agent(mad_snoop_priv);
546 	wait_for_completion(&mad_snoop_priv->comp);
547 
548 	kfree(mad_snoop_priv);
549 }
550 
551 /*
552  * ib_unregister_mad_agent - Unregisters a client from using MAD services
553  */
554 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
555 {
556 	struct ib_mad_agent_private *mad_agent_priv;
557 	struct ib_mad_snoop_private *mad_snoop_priv;
558 
559 	/* If the TID is zero, the agent can only snoop. */
560 	if (mad_agent->hi_tid) {
561 		mad_agent_priv = container_of(mad_agent,
562 					      struct ib_mad_agent_private,
563 					      agent);
564 		unregister_mad_agent(mad_agent_priv);
565 	} else {
566 		mad_snoop_priv = container_of(mad_agent,
567 					      struct ib_mad_snoop_private,
568 					      agent);
569 		unregister_mad_snoop(mad_snoop_priv);
570 	}
571 	return 0;
572 }
573 EXPORT_SYMBOL(ib_unregister_mad_agent);
574 
575 static void dequeue_mad(struct ib_mad_list_head *mad_list)
576 {
577 	struct ib_mad_queue *mad_queue;
578 	unsigned long flags;
579 
580 	BUG_ON(!mad_list->mad_queue);
581 	mad_queue = mad_list->mad_queue;
582 	spin_lock_irqsave(&mad_queue->lock, flags);
583 	list_del(&mad_list->list);
584 	mad_queue->count--;
585 	spin_unlock_irqrestore(&mad_queue->lock, flags);
586 }
587 
588 static void snoop_send(struct ib_mad_qp_info *qp_info,
589 		       struct ib_mad_send_buf *send_buf,
590 		       struct ib_mad_send_wc *mad_send_wc,
591 		       int mad_snoop_flags)
592 {
593 	struct ib_mad_snoop_private *mad_snoop_priv;
594 	unsigned long flags;
595 	int i;
596 
597 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
598 	for (i = 0; i < qp_info->snoop_table_size; i++) {
599 		mad_snoop_priv = qp_info->snoop_table[i];
600 		if (!mad_snoop_priv ||
601 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
602 			continue;
603 
604 		atomic_inc(&mad_snoop_priv->refcount);
605 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
606 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
607 						    send_buf, mad_send_wc);
608 		deref_snoop_agent(mad_snoop_priv);
609 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
610 	}
611 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
612 }
613 
614 static void snoop_recv(struct ib_mad_qp_info *qp_info,
615 		       struct ib_mad_recv_wc *mad_recv_wc,
616 		       int mad_snoop_flags)
617 {
618 	struct ib_mad_snoop_private *mad_snoop_priv;
619 	unsigned long flags;
620 	int i;
621 
622 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
623 	for (i = 0; i < qp_info->snoop_table_size; i++) {
624 		mad_snoop_priv = qp_info->snoop_table[i];
625 		if (!mad_snoop_priv ||
626 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
627 			continue;
628 
629 		atomic_inc(&mad_snoop_priv->refcount);
630 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
631 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
632 						   mad_recv_wc);
633 		deref_snoop_agent(mad_snoop_priv);
634 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
635 	}
636 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
637 }
638 
639 static void build_smp_wc(struct ib_qp *qp,
640 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
641 			 struct ib_wc *wc)
642 {
643 	memset(wc, 0, sizeof *wc);
644 	wc->wr_id = wr_id;
645 	wc->status = IB_WC_SUCCESS;
646 	wc->opcode = IB_WC_RECV;
647 	wc->pkey_index = pkey_index;
648 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
649 	wc->src_qp = IB_QP0;
650 	wc->qp = qp;
651 	wc->slid = slid;
652 	wc->sl = 0;
653 	wc->dlid_path_bits = 0;
654 	wc->port_num = port_num;
655 }
656 
657 /*
658  * Return 0 if SMP is to be sent
659  * Return 1 if SMP was consumed locally (whether or not solicited)
660  * Return < 0 if error
661  */
662 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
663 				  struct ib_mad_send_wr_private *mad_send_wr)
664 {
665 	int ret = 0;
666 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
667 	unsigned long flags;
668 	struct ib_mad_local_private *local;
669 	struct ib_mad_private *mad_priv;
670 	struct ib_mad_port_private *port_priv;
671 	struct ib_mad_agent_private *recv_mad_agent = NULL;
672 	struct ib_device *device = mad_agent_priv->agent.device;
673 	u8 port_num;
674 	struct ib_wc mad_wc;
675 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
676 
677 	if (device->node_type == RDMA_NODE_IB_SWITCH &&
678 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
679 		port_num = send_wr->wr.ud.port_num;
680 	else
681 		port_num = mad_agent_priv->agent.port_num;
682 
683 	/*
684 	 * Directed route handling starts if the initial LID routed part of
685 	 * a request or the ending LID routed part of a response is empty.
686 	 * If we are at the start of the LID routed part, don't update the
687 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
688 	 */
689 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
690 	     IB_LID_PERMISSIVE &&
691 	     smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
692 	     IB_SMI_DISCARD) {
693 		ret = -EINVAL;
694 		printk(KERN_ERR PFX "Invalid directed route\n");
695 		goto out;
696 	}
697 
698 	/* Check to post send on QP or process locally */
699 	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
700 	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
701 		goto out;
702 
703 	local = kmalloc(sizeof *local, GFP_ATOMIC);
704 	if (!local) {
705 		ret = -ENOMEM;
706 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
707 		goto out;
708 	}
709 	local->mad_priv = NULL;
710 	local->recv_mad_agent = NULL;
711 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
712 	if (!mad_priv) {
713 		ret = -ENOMEM;
714 		printk(KERN_ERR PFX "No memory for local response MAD\n");
715 		kfree(local);
716 		goto out;
717 	}
718 
719 	build_smp_wc(mad_agent_priv->agent.qp,
720 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
721 		     send_wr->wr.ud.pkey_index,
722 		     send_wr->wr.ud.port_num, &mad_wc);
723 
724 	/* No GRH for DR SMP */
725 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
726 				  (struct ib_mad *)smp,
727 				  (struct ib_mad *)&mad_priv->mad);
728 	switch (ret)
729 	{
730 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
731 		if (ib_response_mad(&mad_priv->mad.mad) &&
732 		    mad_agent_priv->agent.recv_handler) {
733 			local->mad_priv = mad_priv;
734 			local->recv_mad_agent = mad_agent_priv;
735 			/*
736 			 * Reference MAD agent until receive
737 			 * side of local completion handled
738 			 */
739 			atomic_inc(&mad_agent_priv->refcount);
740 		} else
741 			kmem_cache_free(ib_mad_cache, mad_priv);
742 		break;
743 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
744 		kmem_cache_free(ib_mad_cache, mad_priv);
745 		break;
746 	case IB_MAD_RESULT_SUCCESS:
747 		/* Treat like an incoming receive MAD */
748 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
749 					    mad_agent_priv->agent.port_num);
750 		if (port_priv) {
751 			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
752 			recv_mad_agent = find_mad_agent(port_priv,
753 						        &mad_priv->mad.mad);
754 		}
755 		if (!port_priv || !recv_mad_agent) {
756 			/*
757 			 * No receiving agent so drop packet and
758 			 * generate send completion.
759 			 */
760 			kmem_cache_free(ib_mad_cache, mad_priv);
761 			break;
762 		}
763 		local->mad_priv = mad_priv;
764 		local->recv_mad_agent = recv_mad_agent;
765 		break;
766 	default:
767 		kmem_cache_free(ib_mad_cache, mad_priv);
768 		kfree(local);
769 		ret = -EINVAL;
770 		goto out;
771 	}
772 
773 	local->mad_send_wr = mad_send_wr;
774 	/* Reference MAD agent until send side of local completion handled */
775 	atomic_inc(&mad_agent_priv->refcount);
776 	/* Queue local completion to local list */
777 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
778 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
779 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
780 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
781 		   &mad_agent_priv->local_work);
782 	ret = 1;
783 out:
784 	return ret;
785 }
786 
787 static int get_pad_size(int hdr_len, int data_len)
788 {
789 	int seg_size, pad;
790 
791 	seg_size = sizeof(struct ib_mad) - hdr_len;
792 	if (data_len && seg_size) {
793 		pad = seg_size - data_len % seg_size;
794 		return pad == seg_size ? 0 : pad;
795 	} else
796 		return seg_size;
797 }
798 
799 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
800 {
801 	struct ib_rmpp_segment *s, *t;
802 
803 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
804 		list_del(&s->list);
805 		kfree(s);
806 	}
807 }
808 
809 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
810 				gfp_t gfp_mask)
811 {
812 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
813 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
814 	struct ib_rmpp_segment *seg = NULL;
815 	int left, seg_size, pad;
816 
817 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
818 	seg_size = send_buf->seg_size;
819 	pad = send_wr->pad;
820 
821 	/* Allocate data segments. */
822 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
823 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
824 		if (!seg) {
825 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
826 			       "alloc failed for len %zd, gfp %#x\n",
827 			       sizeof (*seg) + seg_size, gfp_mask);
828 			free_send_rmpp_list(send_wr);
829 			return -ENOMEM;
830 		}
831 		seg->num = ++send_buf->seg_count;
832 		list_add_tail(&seg->list, &send_wr->rmpp_list);
833 	}
834 
835 	/* Zero any padding */
836 	if (pad)
837 		memset(seg->data + seg_size - pad, 0, pad);
838 
839 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
840 					  agent.rmpp_version;
841 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
842 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
843 
844 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
845 					struct ib_rmpp_segment, list);
846 	send_wr->last_ack_seg = send_wr->cur_seg;
847 	return 0;
848 }
849 
850 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
851 					    u32 remote_qpn, u16 pkey_index,
852 					    int rmpp_active,
853 					    int hdr_len, int data_len,
854 					    gfp_t gfp_mask)
855 {
856 	struct ib_mad_agent_private *mad_agent_priv;
857 	struct ib_mad_send_wr_private *mad_send_wr;
858 	int pad, message_size, ret, size;
859 	void *buf;
860 
861 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
862 				      agent);
863 	pad = get_pad_size(hdr_len, data_len);
864 	message_size = hdr_len + data_len + pad;
865 
866 	if ((!mad_agent->rmpp_version &&
867 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
868 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
869 		return ERR_PTR(-EINVAL);
870 
871 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
872 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
873 	if (!buf)
874 		return ERR_PTR(-ENOMEM);
875 
876 	mad_send_wr = buf + size;
877 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
878 	mad_send_wr->send_buf.mad = buf;
879 	mad_send_wr->send_buf.hdr_len = hdr_len;
880 	mad_send_wr->send_buf.data_len = data_len;
881 	mad_send_wr->pad = pad;
882 
883 	mad_send_wr->mad_agent_priv = mad_agent_priv;
884 	mad_send_wr->sg_list[0].length = hdr_len;
885 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
886 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
887 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
888 
889 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
890 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
891 	mad_send_wr->send_wr.num_sge = 2;
892 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
893 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
894 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
895 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
896 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
897 
898 	if (rmpp_active) {
899 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
900 		if (ret) {
901 			kfree(buf);
902 			return ERR_PTR(ret);
903 		}
904 	}
905 
906 	mad_send_wr->send_buf.mad_agent = mad_agent;
907 	atomic_inc(&mad_agent_priv->refcount);
908 	return &mad_send_wr->send_buf;
909 }
910 EXPORT_SYMBOL(ib_create_send_mad);
911 
912 int ib_get_mad_data_offset(u8 mgmt_class)
913 {
914 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
915 		return IB_MGMT_SA_HDR;
916 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
917 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
918 		 (mgmt_class == IB_MGMT_CLASS_BIS))
919 		return IB_MGMT_DEVICE_HDR;
920 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
921 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
922 		return IB_MGMT_VENDOR_HDR;
923 	else
924 		return IB_MGMT_MAD_HDR;
925 }
926 EXPORT_SYMBOL(ib_get_mad_data_offset);
927 
928 int ib_is_mad_class_rmpp(u8 mgmt_class)
929 {
930 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
931 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
932 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
933 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
934 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
935 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
936 		return 1;
937 	return 0;
938 }
939 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
940 
941 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
942 {
943 	struct ib_mad_send_wr_private *mad_send_wr;
944 	struct list_head *list;
945 
946 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
947 				   send_buf);
948 	list = &mad_send_wr->cur_seg->list;
949 
950 	if (mad_send_wr->cur_seg->num < seg_num) {
951 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
952 			if (mad_send_wr->cur_seg->num == seg_num)
953 				break;
954 	} else if (mad_send_wr->cur_seg->num > seg_num) {
955 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
956 			if (mad_send_wr->cur_seg->num == seg_num)
957 				break;
958 	}
959 	return mad_send_wr->cur_seg->data;
960 }
961 EXPORT_SYMBOL(ib_get_rmpp_segment);
962 
963 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
964 {
965 	if (mad_send_wr->send_buf.seg_count)
966 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
967 					   mad_send_wr->seg_num);
968 	else
969 		return mad_send_wr->send_buf.mad +
970 		       mad_send_wr->send_buf.hdr_len;
971 }
972 
973 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
974 {
975 	struct ib_mad_agent_private *mad_agent_priv;
976 	struct ib_mad_send_wr_private *mad_send_wr;
977 
978 	mad_agent_priv = container_of(send_buf->mad_agent,
979 				      struct ib_mad_agent_private, agent);
980 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
981 				   send_buf);
982 
983 	free_send_rmpp_list(mad_send_wr);
984 	kfree(send_buf->mad);
985 	deref_mad_agent(mad_agent_priv);
986 }
987 EXPORT_SYMBOL(ib_free_send_mad);
988 
989 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
990 {
991 	struct ib_mad_qp_info *qp_info;
992 	struct list_head *list;
993 	struct ib_send_wr *bad_send_wr;
994 	struct ib_mad_agent *mad_agent;
995 	struct ib_sge *sge;
996 	unsigned long flags;
997 	int ret;
998 
999 	/* Set WR ID to find mad_send_wr upon completion */
1000 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1001 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1002 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1003 
1004 	mad_agent = mad_send_wr->send_buf.mad_agent;
1005 	sge = mad_send_wr->sg_list;
1006 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1007 					mad_send_wr->send_buf.mad,
1008 					sge[0].length,
1009 					DMA_TO_DEVICE);
1010 	mad_send_wr->header_mapping = sge[0].addr;
1011 
1012 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1013 					ib_get_payload(mad_send_wr),
1014 					sge[1].length,
1015 					DMA_TO_DEVICE);
1016 	mad_send_wr->payload_mapping = sge[1].addr;
1017 
1018 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1019 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1020 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1021 				   &bad_send_wr);
1022 		list = &qp_info->send_queue.list;
1023 	} else {
1024 		ret = 0;
1025 		list = &qp_info->overflow_list;
1026 	}
1027 
1028 	if (!ret) {
1029 		qp_info->send_queue.count++;
1030 		list_add_tail(&mad_send_wr->mad_list.list, list);
1031 	}
1032 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1033 	if (ret) {
1034 		ib_dma_unmap_single(mad_agent->device,
1035 				    mad_send_wr->header_mapping,
1036 				    sge[0].length, DMA_TO_DEVICE);
1037 		ib_dma_unmap_single(mad_agent->device,
1038 				    mad_send_wr->payload_mapping,
1039 				    sge[1].length, DMA_TO_DEVICE);
1040 	}
1041 	return ret;
1042 }
1043 
1044 /*
1045  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1046  *  with the registered client
1047  */
1048 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1049 		     struct ib_mad_send_buf **bad_send_buf)
1050 {
1051 	struct ib_mad_agent_private *mad_agent_priv;
1052 	struct ib_mad_send_buf *next_send_buf;
1053 	struct ib_mad_send_wr_private *mad_send_wr;
1054 	unsigned long flags;
1055 	int ret = -EINVAL;
1056 
1057 	/* Walk list of send WRs and post each on send list */
1058 	for (; send_buf; send_buf = next_send_buf) {
1059 
1060 		mad_send_wr = container_of(send_buf,
1061 					   struct ib_mad_send_wr_private,
1062 					   send_buf);
1063 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1064 
1065 		if (!send_buf->mad_agent->send_handler ||
1066 		    (send_buf->timeout_ms &&
1067 		     !send_buf->mad_agent->recv_handler)) {
1068 			ret = -EINVAL;
1069 			goto error;
1070 		}
1071 
1072 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1073 			if (mad_agent_priv->agent.rmpp_version) {
1074 				ret = -EINVAL;
1075 				goto error;
1076 			}
1077 		}
1078 
1079 		/*
1080 		 * Save pointer to next work request to post in case the
1081 		 * current one completes, and the user modifies the work
1082 		 * request associated with the completion
1083 		 */
1084 		next_send_buf = send_buf->next;
1085 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1086 
1087 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1088 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1089 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1090 						     mad_send_wr);
1091 			if (ret < 0)		/* error */
1092 				goto error;
1093 			else if (ret == 1)	/* locally consumed */
1094 				continue;
1095 		}
1096 
1097 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1098 		/* Timeout will be updated after send completes */
1099 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1100 		mad_send_wr->max_retries = send_buf->retries;
1101 		mad_send_wr->retries_left = send_buf->retries;
1102 		send_buf->retries = 0;
1103 		/* Reference for work request to QP + response */
1104 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1105 		mad_send_wr->status = IB_WC_SUCCESS;
1106 
1107 		/* Reference MAD agent until send completes */
1108 		atomic_inc(&mad_agent_priv->refcount);
1109 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1110 		list_add_tail(&mad_send_wr->agent_list,
1111 			      &mad_agent_priv->send_list);
1112 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1113 
1114 		if (mad_agent_priv->agent.rmpp_version) {
1115 			ret = ib_send_rmpp_mad(mad_send_wr);
1116 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1117 				ret = ib_send_mad(mad_send_wr);
1118 		} else
1119 			ret = ib_send_mad(mad_send_wr);
1120 		if (ret < 0) {
1121 			/* Fail send request */
1122 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1123 			list_del(&mad_send_wr->agent_list);
1124 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1125 			atomic_dec(&mad_agent_priv->refcount);
1126 			goto error;
1127 		}
1128 	}
1129 	return 0;
1130 error:
1131 	if (bad_send_buf)
1132 		*bad_send_buf = send_buf;
1133 	return ret;
1134 }
1135 EXPORT_SYMBOL(ib_post_send_mad);
1136 
1137 /*
1138  * ib_free_recv_mad - Returns data buffers used to receive
1139  *  a MAD to the access layer
1140  */
1141 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1142 {
1143 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1144 	struct ib_mad_private_header *mad_priv_hdr;
1145 	struct ib_mad_private *priv;
1146 	struct list_head free_list;
1147 
1148 	INIT_LIST_HEAD(&free_list);
1149 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1150 
1151 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1152 					&free_list, list) {
1153 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1154 					   recv_buf);
1155 		mad_priv_hdr = container_of(mad_recv_wc,
1156 					    struct ib_mad_private_header,
1157 					    recv_wc);
1158 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1159 				    header);
1160 		kmem_cache_free(ib_mad_cache, priv);
1161 	}
1162 }
1163 EXPORT_SYMBOL(ib_free_recv_mad);
1164 
1165 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1166 					u8 rmpp_version,
1167 					ib_mad_send_handler send_handler,
1168 					ib_mad_recv_handler recv_handler,
1169 					void *context)
1170 {
1171 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1172 }
1173 EXPORT_SYMBOL(ib_redirect_mad_qp);
1174 
1175 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1176 		      struct ib_wc *wc)
1177 {
1178 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1179 	return 0;
1180 }
1181 EXPORT_SYMBOL(ib_process_mad_wc);
1182 
1183 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1184 			 struct ib_mad_reg_req *mad_reg_req)
1185 {
1186 	int i;
1187 
1188 	for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1189 	     i < IB_MGMT_MAX_METHODS;
1190 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1191 			       1+i)) {
1192 		if ((*method)->agent[i]) {
1193 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1194 			return -EINVAL;
1195 		}
1196 	}
1197 	return 0;
1198 }
1199 
1200 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1201 {
1202 	/* Allocate management method table */
1203 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1204 	if (!*method) {
1205 		printk(KERN_ERR PFX "No memory for "
1206 		       "ib_mad_mgmt_method_table\n");
1207 		return -ENOMEM;
1208 	}
1209 
1210 	return 0;
1211 }
1212 
1213 /*
1214  * Check to see if there are any methods still in use
1215  */
1216 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1217 {
1218 	int i;
1219 
1220 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1221 		if (method->agent[i])
1222 			return 1;
1223 	return 0;
1224 }
1225 
1226 /*
1227  * Check to see if there are any method tables for this class still in use
1228  */
1229 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1230 {
1231 	int i;
1232 
1233 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1234 		if (class->method_table[i])
1235 			return 1;
1236 	return 0;
1237 }
1238 
1239 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1240 {
1241 	int i;
1242 
1243 	for (i = 0; i < MAX_MGMT_OUI; i++)
1244 		if (vendor_class->method_table[i])
1245 			return 1;
1246 	return 0;
1247 }
1248 
1249 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1250 			   char *oui)
1251 {
1252 	int i;
1253 
1254 	for (i = 0; i < MAX_MGMT_OUI; i++)
1255 		/* Is there matching OUI for this vendor class ? */
1256 		if (!memcmp(vendor_class->oui[i], oui, 3))
1257 			return i;
1258 
1259 	return -1;
1260 }
1261 
1262 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1263 {
1264 	int i;
1265 
1266 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1267 		if (vendor->vendor_class[i])
1268 			return 1;
1269 
1270 	return 0;
1271 }
1272 
1273 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1274 				     struct ib_mad_agent_private *agent)
1275 {
1276 	int i;
1277 
1278 	/* Remove any methods for this mad agent */
1279 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1280 		if (method->agent[i] == agent) {
1281 			method->agent[i] = NULL;
1282 		}
1283 	}
1284 }
1285 
1286 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1287 			      struct ib_mad_agent_private *agent_priv,
1288 			      u8 mgmt_class)
1289 {
1290 	struct ib_mad_port_private *port_priv;
1291 	struct ib_mad_mgmt_class_table **class;
1292 	struct ib_mad_mgmt_method_table **method;
1293 	int i, ret;
1294 
1295 	port_priv = agent_priv->qp_info->port_priv;
1296 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1297 	if (!*class) {
1298 		/* Allocate management class table for "new" class version */
1299 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1300 		if (!*class) {
1301 			printk(KERN_ERR PFX "No memory for "
1302 			       "ib_mad_mgmt_class_table\n");
1303 			ret = -ENOMEM;
1304 			goto error1;
1305 		}
1306 
1307 		/* Allocate method table for this management class */
1308 		method = &(*class)->method_table[mgmt_class];
1309 		if ((ret = allocate_method_table(method)))
1310 			goto error2;
1311 	} else {
1312 		method = &(*class)->method_table[mgmt_class];
1313 		if (!*method) {
1314 			/* Allocate method table for this management class */
1315 			if ((ret = allocate_method_table(method)))
1316 				goto error1;
1317 		}
1318 	}
1319 
1320 	/* Now, make sure methods are not already in use */
1321 	if (method_in_use(method, mad_reg_req))
1322 		goto error3;
1323 
1324 	/* Finally, add in methods being registered */
1325 	for (i = find_first_bit(mad_reg_req->method_mask,
1326 				IB_MGMT_MAX_METHODS);
1327 	     i < IB_MGMT_MAX_METHODS;
1328 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1329 			       1+i)) {
1330 		(*method)->agent[i] = agent_priv;
1331 	}
1332 	return 0;
1333 
1334 error3:
1335 	/* Remove any methods for this mad agent */
1336 	remove_methods_mad_agent(*method, agent_priv);
1337 	/* Now, check to see if there are any methods in use */
1338 	if (!check_method_table(*method)) {
1339 		/* If not, release management method table */
1340 		kfree(*method);
1341 		*method = NULL;
1342 	}
1343 	ret = -EINVAL;
1344 	goto error1;
1345 error2:
1346 	kfree(*class);
1347 	*class = NULL;
1348 error1:
1349 	return ret;
1350 }
1351 
1352 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1353 			   struct ib_mad_agent_private *agent_priv)
1354 {
1355 	struct ib_mad_port_private *port_priv;
1356 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1357 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1358 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1359 	struct ib_mad_mgmt_method_table **method;
1360 	int i, ret = -ENOMEM;
1361 	u8 vclass;
1362 
1363 	/* "New" vendor (with OUI) class */
1364 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1365 	port_priv = agent_priv->qp_info->port_priv;
1366 	vendor_table = &port_priv->version[
1367 				mad_reg_req->mgmt_class_version].vendor;
1368 	if (!*vendor_table) {
1369 		/* Allocate mgmt vendor class table for "new" class version */
1370 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1371 		if (!vendor) {
1372 			printk(KERN_ERR PFX "No memory for "
1373 			       "ib_mad_mgmt_vendor_class_table\n");
1374 			goto error1;
1375 		}
1376 
1377 		*vendor_table = vendor;
1378 	}
1379 	if (!(*vendor_table)->vendor_class[vclass]) {
1380 		/* Allocate table for this management vendor class */
1381 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1382 		if (!vendor_class) {
1383 			printk(KERN_ERR PFX "No memory for "
1384 			       "ib_mad_mgmt_vendor_class\n");
1385 			goto error2;
1386 		}
1387 
1388 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1389 	}
1390 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1391 		/* Is there matching OUI for this vendor class ? */
1392 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1393 			    mad_reg_req->oui, 3)) {
1394 			method = &(*vendor_table)->vendor_class[
1395 						vclass]->method_table[i];
1396 			BUG_ON(!*method);
1397 			goto check_in_use;
1398 		}
1399 	}
1400 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1401 		/* OUI slot available ? */
1402 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1403 				vclass]->oui[i])) {
1404 			method = &(*vendor_table)->vendor_class[
1405 				vclass]->method_table[i];
1406 			BUG_ON(*method);
1407 			/* Allocate method table for this OUI */
1408 			if ((ret = allocate_method_table(method)))
1409 				goto error3;
1410 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1411 			       mad_reg_req->oui, 3);
1412 			goto check_in_use;
1413 		}
1414 	}
1415 	printk(KERN_ERR PFX "All OUI slots in use\n");
1416 	goto error3;
1417 
1418 check_in_use:
1419 	/* Now, make sure methods are not already in use */
1420 	if (method_in_use(method, mad_reg_req))
1421 		goto error4;
1422 
1423 	/* Finally, add in methods being registered */
1424 	for (i = find_first_bit(mad_reg_req->method_mask,
1425 				IB_MGMT_MAX_METHODS);
1426 	     i < IB_MGMT_MAX_METHODS;
1427 	     i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1428 			       1+i)) {
1429 		(*method)->agent[i] = agent_priv;
1430 	}
1431 	return 0;
1432 
1433 error4:
1434 	/* Remove any methods for this mad agent */
1435 	remove_methods_mad_agent(*method, agent_priv);
1436 	/* Now, check to see if there are any methods in use */
1437 	if (!check_method_table(*method)) {
1438 		/* If not, release management method table */
1439 		kfree(*method);
1440 		*method = NULL;
1441 	}
1442 	ret = -EINVAL;
1443 error3:
1444 	if (vendor_class) {
1445 		(*vendor_table)->vendor_class[vclass] = NULL;
1446 		kfree(vendor_class);
1447 	}
1448 error2:
1449 	if (vendor) {
1450 		*vendor_table = NULL;
1451 		kfree(vendor);
1452 	}
1453 error1:
1454 	return ret;
1455 }
1456 
1457 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1458 {
1459 	struct ib_mad_port_private *port_priv;
1460 	struct ib_mad_mgmt_class_table *class;
1461 	struct ib_mad_mgmt_method_table *method;
1462 	struct ib_mad_mgmt_vendor_class_table *vendor;
1463 	struct ib_mad_mgmt_vendor_class *vendor_class;
1464 	int index;
1465 	u8 mgmt_class;
1466 
1467 	/*
1468 	 * Was MAD registration request supplied
1469 	 * with original registration ?
1470 	 */
1471 	if (!agent_priv->reg_req) {
1472 		goto out;
1473 	}
1474 
1475 	port_priv = agent_priv->qp_info->port_priv;
1476 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1477 	class = port_priv->version[
1478 			agent_priv->reg_req->mgmt_class_version].class;
1479 	if (!class)
1480 		goto vendor_check;
1481 
1482 	method = class->method_table[mgmt_class];
1483 	if (method) {
1484 		/* Remove any methods for this mad agent */
1485 		remove_methods_mad_agent(method, agent_priv);
1486 		/* Now, check to see if there are any methods still in use */
1487 		if (!check_method_table(method)) {
1488 			/* If not, release management method table */
1489 			 kfree(method);
1490 			 class->method_table[mgmt_class] = NULL;
1491 			 /* Any management classes left ? */
1492 			if (!check_class_table(class)) {
1493 				/* If not, release management class table */
1494 				kfree(class);
1495 				port_priv->version[
1496 					agent_priv->reg_req->
1497 					mgmt_class_version].class = NULL;
1498 			}
1499 		}
1500 	}
1501 
1502 vendor_check:
1503 	if (!is_vendor_class(mgmt_class))
1504 		goto out;
1505 
1506 	/* normalize mgmt_class to vendor range 2 */
1507 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1508 	vendor = port_priv->version[
1509 			agent_priv->reg_req->mgmt_class_version].vendor;
1510 
1511 	if (!vendor)
1512 		goto out;
1513 
1514 	vendor_class = vendor->vendor_class[mgmt_class];
1515 	if (vendor_class) {
1516 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1517 		if (index < 0)
1518 			goto out;
1519 		method = vendor_class->method_table[index];
1520 		if (method) {
1521 			/* Remove any methods for this mad agent */
1522 			remove_methods_mad_agent(method, agent_priv);
1523 			/*
1524 			 * Now, check to see if there are
1525 			 * any methods still in use
1526 			 */
1527 			if (!check_method_table(method)) {
1528 				/* If not, release management method table */
1529 				kfree(method);
1530 				vendor_class->method_table[index] = NULL;
1531 				memset(vendor_class->oui[index], 0, 3);
1532 				/* Any OUIs left ? */
1533 				if (!check_vendor_class(vendor_class)) {
1534 					/* If not, release vendor class table */
1535 					kfree(vendor_class);
1536 					vendor->vendor_class[mgmt_class] = NULL;
1537 					/* Any other vendor classes left ? */
1538 					if (!check_vendor_table(vendor)) {
1539 						kfree(vendor);
1540 						port_priv->version[
1541 							agent_priv->reg_req->
1542 							mgmt_class_version].
1543 							vendor = NULL;
1544 					}
1545 				}
1546 			}
1547 		}
1548 	}
1549 
1550 out:
1551 	return;
1552 }
1553 
1554 static struct ib_mad_agent_private *
1555 find_mad_agent(struct ib_mad_port_private *port_priv,
1556 	       struct ib_mad *mad)
1557 {
1558 	struct ib_mad_agent_private *mad_agent = NULL;
1559 	unsigned long flags;
1560 
1561 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1562 	if (ib_response_mad(mad)) {
1563 		u32 hi_tid;
1564 		struct ib_mad_agent_private *entry;
1565 
1566 		/*
1567 		 * Routing is based on high 32 bits of transaction ID
1568 		 * of MAD.
1569 		 */
1570 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1571 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1572 			if (entry->agent.hi_tid == hi_tid) {
1573 				mad_agent = entry;
1574 				break;
1575 			}
1576 		}
1577 	} else {
1578 		struct ib_mad_mgmt_class_table *class;
1579 		struct ib_mad_mgmt_method_table *method;
1580 		struct ib_mad_mgmt_vendor_class_table *vendor;
1581 		struct ib_mad_mgmt_vendor_class *vendor_class;
1582 		struct ib_vendor_mad *vendor_mad;
1583 		int index;
1584 
1585 		/*
1586 		 * Routing is based on version, class, and method
1587 		 * For "newer" vendor MADs, also based on OUI
1588 		 */
1589 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1590 			goto out;
1591 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1592 			class = port_priv->version[
1593 					mad->mad_hdr.class_version].class;
1594 			if (!class)
1595 				goto out;
1596 			method = class->method_table[convert_mgmt_class(
1597 							mad->mad_hdr.mgmt_class)];
1598 			if (method)
1599 				mad_agent = method->agent[mad->mad_hdr.method &
1600 							  ~IB_MGMT_METHOD_RESP];
1601 		} else {
1602 			vendor = port_priv->version[
1603 					mad->mad_hdr.class_version].vendor;
1604 			if (!vendor)
1605 				goto out;
1606 			vendor_class = vendor->vendor_class[vendor_class_index(
1607 						mad->mad_hdr.mgmt_class)];
1608 			if (!vendor_class)
1609 				goto out;
1610 			/* Find matching OUI */
1611 			vendor_mad = (struct ib_vendor_mad *)mad;
1612 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1613 			if (index == -1)
1614 				goto out;
1615 			method = vendor_class->method_table[index];
1616 			if (method) {
1617 				mad_agent = method->agent[mad->mad_hdr.method &
1618 							  ~IB_MGMT_METHOD_RESP];
1619 			}
1620 		}
1621 	}
1622 
1623 	if (mad_agent) {
1624 		if (mad_agent->agent.recv_handler)
1625 			atomic_inc(&mad_agent->refcount);
1626 		else {
1627 			printk(KERN_NOTICE PFX "No receive handler for client "
1628 			       "%p on port %d\n",
1629 			       &mad_agent->agent, port_priv->port_num);
1630 			mad_agent = NULL;
1631 		}
1632 	}
1633 out:
1634 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1635 
1636 	return mad_agent;
1637 }
1638 
1639 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1640 {
1641 	int valid = 0;
1642 
1643 	/* Make sure MAD base version is understood */
1644 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1645 		printk(KERN_ERR PFX "MAD received with unsupported base "
1646 		       "version %d\n", mad->mad_hdr.base_version);
1647 		goto out;
1648 	}
1649 
1650 	/* Filter SMI packets sent to other than QP0 */
1651 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1652 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1653 		if (qp_num == 0)
1654 			valid = 1;
1655 	} else {
1656 		/* Filter GSI packets sent to QP0 */
1657 		if (qp_num != 0)
1658 			valid = 1;
1659 	}
1660 
1661 out:
1662 	return valid;
1663 }
1664 
1665 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1666 		       struct ib_mad_hdr *mad_hdr)
1667 {
1668 	struct ib_rmpp_mad *rmpp_mad;
1669 
1670 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1671 	return !mad_agent_priv->agent.rmpp_version ||
1672 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1673 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1674 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1675 }
1676 
1677 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1678 				     struct ib_mad_recv_wc *rwc)
1679 {
1680 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1681 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1682 }
1683 
1684 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1685 				   struct ib_mad_send_wr_private *wr,
1686 				   struct ib_mad_recv_wc *rwc )
1687 {
1688 	struct ib_ah_attr attr;
1689 	u8 send_resp, rcv_resp;
1690 	union ib_gid sgid;
1691 	struct ib_device *device = mad_agent_priv->agent.device;
1692 	u8 port_num = mad_agent_priv->agent.port_num;
1693 	u8 lmc;
1694 
1695 	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1696 	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1697 
1698 	if (send_resp == rcv_resp)
1699 		/* both requests, or both responses. GIDs different */
1700 		return 0;
1701 
1702 	if (ib_query_ah(wr->send_buf.ah, &attr))
1703 		/* Assume not equal, to avoid false positives. */
1704 		return 0;
1705 
1706 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1707 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1708 		/* one has GID, other does not.  Assume different */
1709 		return 0;
1710 
1711 	if (!send_resp && rcv_resp) {
1712 		/* is request/response. */
1713 		if (!(attr.ah_flags & IB_AH_GRH)) {
1714 			if (ib_get_cached_lmc(device, port_num, &lmc))
1715 				return 0;
1716 			return (!lmc || !((attr.src_path_bits ^
1717 					   rwc->wc->dlid_path_bits) &
1718 					  ((1 << lmc) - 1)));
1719 		} else {
1720 			if (ib_get_cached_gid(device, port_num,
1721 					      attr.grh.sgid_index, &sgid))
1722 				return 0;
1723 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1724 				       16);
1725 		}
1726 	}
1727 
1728 	if (!(attr.ah_flags & IB_AH_GRH))
1729 		return attr.dlid == rwc->wc->slid;
1730 	else
1731 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1732 			       16);
1733 }
1734 
1735 static inline int is_direct(u8 class)
1736 {
1737 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1738 }
1739 
1740 struct ib_mad_send_wr_private*
1741 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1742 		 struct ib_mad_recv_wc *wc)
1743 {
1744 	struct ib_mad_send_wr_private *wr;
1745 	struct ib_mad *mad;
1746 
1747 	mad = (struct ib_mad *)wc->recv_buf.mad;
1748 
1749 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1750 		if ((wr->tid == mad->mad_hdr.tid) &&
1751 		    rcv_has_same_class(wr, wc) &&
1752 		    /*
1753 		     * Don't check GID for direct routed MADs.
1754 		     * These might have permissive LIDs.
1755 		     */
1756 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1757 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1758 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1759 	}
1760 
1761 	/*
1762 	 * It's possible to receive the response before we've
1763 	 * been notified that the send has completed
1764 	 */
1765 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1766 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1767 		    wr->tid == mad->mad_hdr.tid &&
1768 		    wr->timeout &&
1769 		    rcv_has_same_class(wr, wc) &&
1770 		    /*
1771 		     * Don't check GID for direct routed MADs.
1772 		     * These might have permissive LIDs.
1773 		     */
1774 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1775 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1776 			/* Verify request has not been canceled */
1777 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1778 	}
1779 	return NULL;
1780 }
1781 
1782 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1783 {
1784 	mad_send_wr->timeout = 0;
1785 	if (mad_send_wr->refcount == 1)
1786 		list_move_tail(&mad_send_wr->agent_list,
1787 			      &mad_send_wr->mad_agent_priv->done_list);
1788 }
1789 
1790 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1791 				 struct ib_mad_recv_wc *mad_recv_wc)
1792 {
1793 	struct ib_mad_send_wr_private *mad_send_wr;
1794 	struct ib_mad_send_wc mad_send_wc;
1795 	unsigned long flags;
1796 
1797 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1798 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1799 	if (mad_agent_priv->agent.rmpp_version) {
1800 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1801 						      mad_recv_wc);
1802 		if (!mad_recv_wc) {
1803 			deref_mad_agent(mad_agent_priv);
1804 			return;
1805 		}
1806 	}
1807 
1808 	/* Complete corresponding request */
1809 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1810 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1811 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1812 		if (!mad_send_wr) {
1813 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1814 			ib_free_recv_mad(mad_recv_wc);
1815 			deref_mad_agent(mad_agent_priv);
1816 			return;
1817 		}
1818 		ib_mark_mad_done(mad_send_wr);
1819 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1820 
1821 		/* Defined behavior is to complete response before request */
1822 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1823 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1824 						   mad_recv_wc);
1825 		atomic_dec(&mad_agent_priv->refcount);
1826 
1827 		mad_send_wc.status = IB_WC_SUCCESS;
1828 		mad_send_wc.vendor_err = 0;
1829 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1830 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1831 	} else {
1832 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1833 						   mad_recv_wc);
1834 		deref_mad_agent(mad_agent_priv);
1835 	}
1836 }
1837 
1838 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1839 				     struct ib_wc *wc)
1840 {
1841 	struct ib_mad_qp_info *qp_info;
1842 	struct ib_mad_private_header *mad_priv_hdr;
1843 	struct ib_mad_private *recv, *response = NULL;
1844 	struct ib_mad_list_head *mad_list;
1845 	struct ib_mad_agent_private *mad_agent;
1846 	int port_num;
1847 
1848 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1849 	qp_info = mad_list->mad_queue->qp_info;
1850 	dequeue_mad(mad_list);
1851 
1852 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1853 				    mad_list);
1854 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1855 	ib_dma_unmap_single(port_priv->device,
1856 			    recv->header.mapping,
1857 			    sizeof(struct ib_mad_private) -
1858 			      sizeof(struct ib_mad_private_header),
1859 			    DMA_FROM_DEVICE);
1860 
1861 	/* Setup MAD receive work completion from "normal" work completion */
1862 	recv->header.wc = *wc;
1863 	recv->header.recv_wc.wc = &recv->header.wc;
1864 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1865 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1866 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1867 
1868 	if (atomic_read(&qp_info->snoop_count))
1869 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1870 
1871 	/* Validate MAD */
1872 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1873 		goto out;
1874 
1875 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1876 	if (!response) {
1877 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1878 		       "for response buffer\n");
1879 		goto out;
1880 	}
1881 
1882 	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1883 		port_num = wc->port_num;
1884 	else
1885 		port_num = port_priv->port_num;
1886 
1887 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1888 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1889 		enum smi_forward_action retsmi;
1890 
1891 		if (smi_handle_dr_smp_recv(&recv->mad.smp,
1892 					   port_priv->device->node_type,
1893 					   port_num,
1894 					   port_priv->device->phys_port_cnt) ==
1895 					   IB_SMI_DISCARD)
1896 			goto out;
1897 
1898 		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1899 		if (retsmi == IB_SMI_LOCAL)
1900 			goto local;
1901 
1902 		if (retsmi == IB_SMI_SEND) { /* don't forward */
1903 			if (smi_handle_dr_smp_send(&recv->mad.smp,
1904 						   port_priv->device->node_type,
1905 						   port_num) == IB_SMI_DISCARD)
1906 				goto out;
1907 
1908 			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1909 				goto out;
1910 		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1911 			/* forward case for switches */
1912 			memcpy(response, recv, sizeof(*response));
1913 			response->header.recv_wc.wc = &response->header.wc;
1914 			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1915 			response->header.recv_wc.recv_buf.grh = &response->grh;
1916 
1917 			agent_send_response(&response->mad.mad,
1918 					    &response->grh, wc,
1919 					    port_priv->device,
1920 					    smi_get_fwd_port(&recv->mad.smp),
1921 					    qp_info->qp->qp_num);
1922 
1923 			goto out;
1924 		}
1925 	}
1926 
1927 local:
1928 	/* Give driver "right of first refusal" on incoming MAD */
1929 	if (port_priv->device->process_mad) {
1930 		int ret;
1931 
1932 		ret = port_priv->device->process_mad(port_priv->device, 0,
1933 						     port_priv->port_num,
1934 						     wc, &recv->grh,
1935 						     &recv->mad.mad,
1936 						     &response->mad.mad);
1937 		if (ret & IB_MAD_RESULT_SUCCESS) {
1938 			if (ret & IB_MAD_RESULT_CONSUMED)
1939 				goto out;
1940 			if (ret & IB_MAD_RESULT_REPLY) {
1941 				agent_send_response(&response->mad.mad,
1942 						    &recv->grh, wc,
1943 						    port_priv->device,
1944 						    port_num,
1945 						    qp_info->qp->qp_num);
1946 				goto out;
1947 			}
1948 		}
1949 	}
1950 
1951 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1952 	if (mad_agent) {
1953 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1954 		/*
1955 		 * recv is freed up in error cases in ib_mad_complete_recv
1956 		 * or via recv_handler in ib_mad_complete_recv()
1957 		 */
1958 		recv = NULL;
1959 	}
1960 
1961 out:
1962 	/* Post another receive request for this QP */
1963 	if (response) {
1964 		ib_mad_post_receive_mads(qp_info, response);
1965 		if (recv)
1966 			kmem_cache_free(ib_mad_cache, recv);
1967 	} else
1968 		ib_mad_post_receive_mads(qp_info, recv);
1969 }
1970 
1971 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1972 {
1973 	struct ib_mad_send_wr_private *mad_send_wr;
1974 	unsigned long delay;
1975 
1976 	if (list_empty(&mad_agent_priv->wait_list)) {
1977 		cancel_delayed_work(&mad_agent_priv->timed_work);
1978 	} else {
1979 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1980 					 struct ib_mad_send_wr_private,
1981 					 agent_list);
1982 
1983 		if (time_after(mad_agent_priv->timeout,
1984 			       mad_send_wr->timeout)) {
1985 			mad_agent_priv->timeout = mad_send_wr->timeout;
1986 			cancel_delayed_work(&mad_agent_priv->timed_work);
1987 			delay = mad_send_wr->timeout - jiffies;
1988 			if ((long)delay <= 0)
1989 				delay = 1;
1990 			queue_delayed_work(mad_agent_priv->qp_info->
1991 					   port_priv->wq,
1992 					   &mad_agent_priv->timed_work, delay);
1993 		}
1994 	}
1995 }
1996 
1997 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1998 {
1999 	struct ib_mad_agent_private *mad_agent_priv;
2000 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2001 	struct list_head *list_item;
2002 	unsigned long delay;
2003 
2004 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2005 	list_del(&mad_send_wr->agent_list);
2006 
2007 	delay = mad_send_wr->timeout;
2008 	mad_send_wr->timeout += jiffies;
2009 
2010 	if (delay) {
2011 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2012 			temp_mad_send_wr = list_entry(list_item,
2013 						struct ib_mad_send_wr_private,
2014 						agent_list);
2015 			if (time_after(mad_send_wr->timeout,
2016 				       temp_mad_send_wr->timeout))
2017 				break;
2018 		}
2019 	}
2020 	else
2021 		list_item = &mad_agent_priv->wait_list;
2022 	list_add(&mad_send_wr->agent_list, list_item);
2023 
2024 	/* Reschedule a work item if we have a shorter timeout */
2025 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2026 		cancel_delayed_work(&mad_agent_priv->timed_work);
2027 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2028 				   &mad_agent_priv->timed_work, delay);
2029 	}
2030 }
2031 
2032 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2033 			  int timeout_ms)
2034 {
2035 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2036 	wait_for_response(mad_send_wr);
2037 }
2038 
2039 /*
2040  * Process a send work completion
2041  */
2042 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2043 			     struct ib_mad_send_wc *mad_send_wc)
2044 {
2045 	struct ib_mad_agent_private	*mad_agent_priv;
2046 	unsigned long			flags;
2047 	int				ret;
2048 
2049 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2050 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2051 	if (mad_agent_priv->agent.rmpp_version) {
2052 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2053 		if (ret == IB_RMPP_RESULT_CONSUMED)
2054 			goto done;
2055 	} else
2056 		ret = IB_RMPP_RESULT_UNHANDLED;
2057 
2058 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2059 	    mad_send_wr->status == IB_WC_SUCCESS) {
2060 		mad_send_wr->status = mad_send_wc->status;
2061 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2062 	}
2063 
2064 	if (--mad_send_wr->refcount > 0) {
2065 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2066 		    mad_send_wr->status == IB_WC_SUCCESS) {
2067 			wait_for_response(mad_send_wr);
2068 		}
2069 		goto done;
2070 	}
2071 
2072 	/* Remove send from MAD agent and notify client of completion */
2073 	list_del(&mad_send_wr->agent_list);
2074 	adjust_timeout(mad_agent_priv);
2075 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2076 
2077 	if (mad_send_wr->status != IB_WC_SUCCESS )
2078 		mad_send_wc->status = mad_send_wr->status;
2079 	if (ret == IB_RMPP_RESULT_INTERNAL)
2080 		ib_rmpp_send_handler(mad_send_wc);
2081 	else
2082 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2083 						   mad_send_wc);
2084 
2085 	/* Release reference on agent taken when sending */
2086 	deref_mad_agent(mad_agent_priv);
2087 	return;
2088 done:
2089 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2090 }
2091 
2092 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2093 				     struct ib_wc *wc)
2094 {
2095 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2096 	struct ib_mad_list_head		*mad_list;
2097 	struct ib_mad_qp_info		*qp_info;
2098 	struct ib_mad_queue		*send_queue;
2099 	struct ib_send_wr		*bad_send_wr;
2100 	struct ib_mad_send_wc		mad_send_wc;
2101 	unsigned long flags;
2102 	int ret;
2103 
2104 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2105 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2106 				   mad_list);
2107 	send_queue = mad_list->mad_queue;
2108 	qp_info = send_queue->qp_info;
2109 
2110 retry:
2111 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2112 			    mad_send_wr->header_mapping,
2113 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2114 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2115 			    mad_send_wr->payload_mapping,
2116 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2117 	queued_send_wr = NULL;
2118 	spin_lock_irqsave(&send_queue->lock, flags);
2119 	list_del(&mad_list->list);
2120 
2121 	/* Move queued send to the send queue */
2122 	if (send_queue->count-- > send_queue->max_active) {
2123 		mad_list = container_of(qp_info->overflow_list.next,
2124 					struct ib_mad_list_head, list);
2125 		queued_send_wr = container_of(mad_list,
2126 					struct ib_mad_send_wr_private,
2127 					mad_list);
2128 		list_move_tail(&mad_list->list, &send_queue->list);
2129 	}
2130 	spin_unlock_irqrestore(&send_queue->lock, flags);
2131 
2132 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2133 	mad_send_wc.status = wc->status;
2134 	mad_send_wc.vendor_err = wc->vendor_err;
2135 	if (atomic_read(&qp_info->snoop_count))
2136 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2137 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2138 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2139 
2140 	if (queued_send_wr) {
2141 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2142 				   &bad_send_wr);
2143 		if (ret) {
2144 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2145 			mad_send_wr = queued_send_wr;
2146 			wc->status = IB_WC_LOC_QP_OP_ERR;
2147 			goto retry;
2148 		}
2149 	}
2150 }
2151 
2152 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2153 {
2154 	struct ib_mad_send_wr_private *mad_send_wr;
2155 	struct ib_mad_list_head *mad_list;
2156 	unsigned long flags;
2157 
2158 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2159 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2160 		mad_send_wr = container_of(mad_list,
2161 					   struct ib_mad_send_wr_private,
2162 					   mad_list);
2163 		mad_send_wr->retry = 1;
2164 	}
2165 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2166 }
2167 
2168 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2169 			      struct ib_wc *wc)
2170 {
2171 	struct ib_mad_list_head *mad_list;
2172 	struct ib_mad_qp_info *qp_info;
2173 	struct ib_mad_send_wr_private *mad_send_wr;
2174 	int ret;
2175 
2176 	/* Determine if failure was a send or receive */
2177 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2178 	qp_info = mad_list->mad_queue->qp_info;
2179 	if (mad_list->mad_queue == &qp_info->recv_queue)
2180 		/*
2181 		 * Receive errors indicate that the QP has entered the error
2182 		 * state - error handling/shutdown code will cleanup
2183 		 */
2184 		return;
2185 
2186 	/*
2187 	 * Send errors will transition the QP to SQE - move
2188 	 * QP to RTS and repost flushed work requests
2189 	 */
2190 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2191 				   mad_list);
2192 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2193 		if (mad_send_wr->retry) {
2194 			/* Repost send */
2195 			struct ib_send_wr *bad_send_wr;
2196 
2197 			mad_send_wr->retry = 0;
2198 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2199 					&bad_send_wr);
2200 			if (ret)
2201 				ib_mad_send_done_handler(port_priv, wc);
2202 		} else
2203 			ib_mad_send_done_handler(port_priv, wc);
2204 	} else {
2205 		struct ib_qp_attr *attr;
2206 
2207 		/* Transition QP to RTS and fail offending send */
2208 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2209 		if (attr) {
2210 			attr->qp_state = IB_QPS_RTS;
2211 			attr->cur_qp_state = IB_QPS_SQE;
2212 			ret = ib_modify_qp(qp_info->qp, attr,
2213 					   IB_QP_STATE | IB_QP_CUR_STATE);
2214 			kfree(attr);
2215 			if (ret)
2216 				printk(KERN_ERR PFX "mad_error_handler - "
2217 				       "ib_modify_qp to RTS : %d\n", ret);
2218 			else
2219 				mark_sends_for_retry(qp_info);
2220 		}
2221 		ib_mad_send_done_handler(port_priv, wc);
2222 	}
2223 }
2224 
2225 /*
2226  * IB MAD completion callback
2227  */
2228 static void ib_mad_completion_handler(struct work_struct *work)
2229 {
2230 	struct ib_mad_port_private *port_priv;
2231 	struct ib_wc wc;
2232 
2233 	port_priv = container_of(work, struct ib_mad_port_private, work);
2234 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2235 
2236 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2237 		if (wc.status == IB_WC_SUCCESS) {
2238 			switch (wc.opcode) {
2239 			case IB_WC_SEND:
2240 				ib_mad_send_done_handler(port_priv, &wc);
2241 				break;
2242 			case IB_WC_RECV:
2243 				ib_mad_recv_done_handler(port_priv, &wc);
2244 				break;
2245 			default:
2246 				BUG_ON(1);
2247 				break;
2248 			}
2249 		} else
2250 			mad_error_handler(port_priv, &wc);
2251 	}
2252 }
2253 
2254 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2255 {
2256 	unsigned long flags;
2257 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2258 	struct ib_mad_send_wc mad_send_wc;
2259 	struct list_head cancel_list;
2260 
2261 	INIT_LIST_HEAD(&cancel_list);
2262 
2263 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2264 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2265 				 &mad_agent_priv->send_list, agent_list) {
2266 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2267 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2268 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2269 		}
2270 	}
2271 
2272 	/* Empty wait list to prevent receives from finding a request */
2273 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2274 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2275 
2276 	/* Report all cancelled requests */
2277 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2278 	mad_send_wc.vendor_err = 0;
2279 
2280 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2281 				 &cancel_list, agent_list) {
2282 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2283 		list_del(&mad_send_wr->agent_list);
2284 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2285 						   &mad_send_wc);
2286 		atomic_dec(&mad_agent_priv->refcount);
2287 	}
2288 }
2289 
2290 static struct ib_mad_send_wr_private*
2291 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2292 	     struct ib_mad_send_buf *send_buf)
2293 {
2294 	struct ib_mad_send_wr_private *mad_send_wr;
2295 
2296 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2297 			    agent_list) {
2298 		if (&mad_send_wr->send_buf == send_buf)
2299 			return mad_send_wr;
2300 	}
2301 
2302 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2303 			    agent_list) {
2304 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2305 		    &mad_send_wr->send_buf == send_buf)
2306 			return mad_send_wr;
2307 	}
2308 	return NULL;
2309 }
2310 
2311 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2312 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2313 {
2314 	struct ib_mad_agent_private *mad_agent_priv;
2315 	struct ib_mad_send_wr_private *mad_send_wr;
2316 	unsigned long flags;
2317 	int active;
2318 
2319 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2320 				      agent);
2321 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2322 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2323 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2324 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2325 		return -EINVAL;
2326 	}
2327 
2328 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2329 	if (!timeout_ms) {
2330 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2331 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2332 	}
2333 
2334 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2335 	if (active)
2336 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2337 	else
2338 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2339 
2340 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2341 	return 0;
2342 }
2343 EXPORT_SYMBOL(ib_modify_mad);
2344 
2345 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2346 		   struct ib_mad_send_buf *send_buf)
2347 {
2348 	ib_modify_mad(mad_agent, send_buf, 0);
2349 }
2350 EXPORT_SYMBOL(ib_cancel_mad);
2351 
2352 static void local_completions(struct work_struct *work)
2353 {
2354 	struct ib_mad_agent_private *mad_agent_priv;
2355 	struct ib_mad_local_private *local;
2356 	struct ib_mad_agent_private *recv_mad_agent;
2357 	unsigned long flags;
2358 	int free_mad;
2359 	struct ib_wc wc;
2360 	struct ib_mad_send_wc mad_send_wc;
2361 
2362 	mad_agent_priv =
2363 		container_of(work, struct ib_mad_agent_private, local_work);
2364 
2365 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2366 	while (!list_empty(&mad_agent_priv->local_list)) {
2367 		local = list_entry(mad_agent_priv->local_list.next,
2368 				   struct ib_mad_local_private,
2369 				   completion_list);
2370 		list_del(&local->completion_list);
2371 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2372 		free_mad = 0;
2373 		if (local->mad_priv) {
2374 			recv_mad_agent = local->recv_mad_agent;
2375 			if (!recv_mad_agent) {
2376 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2377 				free_mad = 1;
2378 				goto local_send_completion;
2379 			}
2380 
2381 			/*
2382 			 * Defined behavior is to complete response
2383 			 * before request
2384 			 */
2385 			build_smp_wc(recv_mad_agent->agent.qp,
2386 				     (unsigned long) local->mad_send_wr,
2387 				     be16_to_cpu(IB_LID_PERMISSIVE),
2388 				     0, recv_mad_agent->agent.port_num, &wc);
2389 
2390 			local->mad_priv->header.recv_wc.wc = &wc;
2391 			local->mad_priv->header.recv_wc.mad_len =
2392 						sizeof(struct ib_mad);
2393 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2394 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2395 				 &local->mad_priv->header.recv_wc.rmpp_list);
2396 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2397 			local->mad_priv->header.recv_wc.recv_buf.mad =
2398 						&local->mad_priv->mad.mad;
2399 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2400 				snoop_recv(recv_mad_agent->qp_info,
2401 					  &local->mad_priv->header.recv_wc,
2402 					   IB_MAD_SNOOP_RECVS);
2403 			recv_mad_agent->agent.recv_handler(
2404 						&recv_mad_agent->agent,
2405 						&local->mad_priv->header.recv_wc);
2406 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2407 			atomic_dec(&recv_mad_agent->refcount);
2408 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2409 		}
2410 
2411 local_send_completion:
2412 		/* Complete send */
2413 		mad_send_wc.status = IB_WC_SUCCESS;
2414 		mad_send_wc.vendor_err = 0;
2415 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2416 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2417 			snoop_send(mad_agent_priv->qp_info,
2418 				   &local->mad_send_wr->send_buf,
2419 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2420 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2421 						   &mad_send_wc);
2422 
2423 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2424 		atomic_dec(&mad_agent_priv->refcount);
2425 		if (free_mad)
2426 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2427 		kfree(local);
2428 	}
2429 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2430 }
2431 
2432 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2433 {
2434 	int ret;
2435 
2436 	if (!mad_send_wr->retries_left)
2437 		return -ETIMEDOUT;
2438 
2439 	mad_send_wr->retries_left--;
2440 	mad_send_wr->send_buf.retries++;
2441 
2442 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2443 
2444 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2445 		ret = ib_retry_rmpp(mad_send_wr);
2446 		switch (ret) {
2447 		case IB_RMPP_RESULT_UNHANDLED:
2448 			ret = ib_send_mad(mad_send_wr);
2449 			break;
2450 		case IB_RMPP_RESULT_CONSUMED:
2451 			ret = 0;
2452 			break;
2453 		default:
2454 			ret = -ECOMM;
2455 			break;
2456 		}
2457 	} else
2458 		ret = ib_send_mad(mad_send_wr);
2459 
2460 	if (!ret) {
2461 		mad_send_wr->refcount++;
2462 		list_add_tail(&mad_send_wr->agent_list,
2463 			      &mad_send_wr->mad_agent_priv->send_list);
2464 	}
2465 	return ret;
2466 }
2467 
2468 static void timeout_sends(struct work_struct *work)
2469 {
2470 	struct ib_mad_agent_private *mad_agent_priv;
2471 	struct ib_mad_send_wr_private *mad_send_wr;
2472 	struct ib_mad_send_wc mad_send_wc;
2473 	unsigned long flags, delay;
2474 
2475 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2476 				      timed_work.work);
2477 	mad_send_wc.vendor_err = 0;
2478 
2479 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2480 	while (!list_empty(&mad_agent_priv->wait_list)) {
2481 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2482 					 struct ib_mad_send_wr_private,
2483 					 agent_list);
2484 
2485 		if (time_after(mad_send_wr->timeout, jiffies)) {
2486 			delay = mad_send_wr->timeout - jiffies;
2487 			if ((long)delay <= 0)
2488 				delay = 1;
2489 			queue_delayed_work(mad_agent_priv->qp_info->
2490 					   port_priv->wq,
2491 					   &mad_agent_priv->timed_work, delay);
2492 			break;
2493 		}
2494 
2495 		list_del(&mad_send_wr->agent_list);
2496 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2497 		    !retry_send(mad_send_wr))
2498 			continue;
2499 
2500 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2501 
2502 		if (mad_send_wr->status == IB_WC_SUCCESS)
2503 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2504 		else
2505 			mad_send_wc.status = mad_send_wr->status;
2506 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2507 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2508 						   &mad_send_wc);
2509 
2510 		atomic_dec(&mad_agent_priv->refcount);
2511 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2512 	}
2513 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2514 }
2515 
2516 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2517 {
2518 	struct ib_mad_port_private *port_priv = cq->cq_context;
2519 	unsigned long flags;
2520 
2521 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2522 	if (!list_empty(&port_priv->port_list))
2523 		queue_work(port_priv->wq, &port_priv->work);
2524 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2525 }
2526 
2527 /*
2528  * Allocate receive MADs and post receive WRs for them
2529  */
2530 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2531 				    struct ib_mad_private *mad)
2532 {
2533 	unsigned long flags;
2534 	int post, ret;
2535 	struct ib_mad_private *mad_priv;
2536 	struct ib_sge sg_list;
2537 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2538 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2539 
2540 	/* Initialize common scatter list fields */
2541 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2542 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2543 
2544 	/* Initialize common receive WR fields */
2545 	recv_wr.next = NULL;
2546 	recv_wr.sg_list = &sg_list;
2547 	recv_wr.num_sge = 1;
2548 
2549 	do {
2550 		/* Allocate and map receive buffer */
2551 		if (mad) {
2552 			mad_priv = mad;
2553 			mad = NULL;
2554 		} else {
2555 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2556 			if (!mad_priv) {
2557 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2558 				ret = -ENOMEM;
2559 				break;
2560 			}
2561 		}
2562 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2563 						 &mad_priv->grh,
2564 						 sizeof *mad_priv -
2565 						   sizeof mad_priv->header,
2566 						 DMA_FROM_DEVICE);
2567 		mad_priv->header.mapping = sg_list.addr;
2568 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2569 		mad_priv->header.mad_list.mad_queue = recv_queue;
2570 
2571 		/* Post receive WR */
2572 		spin_lock_irqsave(&recv_queue->lock, flags);
2573 		post = (++recv_queue->count < recv_queue->max_active);
2574 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2575 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2576 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2577 		if (ret) {
2578 			spin_lock_irqsave(&recv_queue->lock, flags);
2579 			list_del(&mad_priv->header.mad_list.list);
2580 			recv_queue->count--;
2581 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2582 			ib_dma_unmap_single(qp_info->port_priv->device,
2583 					    mad_priv->header.mapping,
2584 					    sizeof *mad_priv -
2585 					      sizeof mad_priv->header,
2586 					    DMA_FROM_DEVICE);
2587 			kmem_cache_free(ib_mad_cache, mad_priv);
2588 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2589 			break;
2590 		}
2591 	} while (post);
2592 
2593 	return ret;
2594 }
2595 
2596 /*
2597  * Return all the posted receive MADs
2598  */
2599 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2600 {
2601 	struct ib_mad_private_header *mad_priv_hdr;
2602 	struct ib_mad_private *recv;
2603 	struct ib_mad_list_head *mad_list;
2604 
2605 	while (!list_empty(&qp_info->recv_queue.list)) {
2606 
2607 		mad_list = list_entry(qp_info->recv_queue.list.next,
2608 				      struct ib_mad_list_head, list);
2609 		mad_priv_hdr = container_of(mad_list,
2610 					    struct ib_mad_private_header,
2611 					    mad_list);
2612 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2613 				    header);
2614 
2615 		/* Remove from posted receive MAD list */
2616 		list_del(&mad_list->list);
2617 
2618 		ib_dma_unmap_single(qp_info->port_priv->device,
2619 				    recv->header.mapping,
2620 				    sizeof(struct ib_mad_private) -
2621 				      sizeof(struct ib_mad_private_header),
2622 				    DMA_FROM_DEVICE);
2623 		kmem_cache_free(ib_mad_cache, recv);
2624 	}
2625 
2626 	qp_info->recv_queue.count = 0;
2627 }
2628 
2629 /*
2630  * Start the port
2631  */
2632 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2633 {
2634 	int ret, i;
2635 	struct ib_qp_attr *attr;
2636 	struct ib_qp *qp;
2637 
2638 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2639 	if (!attr) {
2640 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2641 		return -ENOMEM;
2642 	}
2643 
2644 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2645 		qp = port_priv->qp_info[i].qp;
2646 		/*
2647 		 * PKey index for QP1 is irrelevant but
2648 		 * one is needed for the Reset to Init transition
2649 		 */
2650 		attr->qp_state = IB_QPS_INIT;
2651 		attr->pkey_index = 0;
2652 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2653 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2654 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2655 		if (ret) {
2656 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2657 			       "INIT: %d\n", i, ret);
2658 			goto out;
2659 		}
2660 
2661 		attr->qp_state = IB_QPS_RTR;
2662 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2663 		if (ret) {
2664 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2665 			       "RTR: %d\n", i, ret);
2666 			goto out;
2667 		}
2668 
2669 		attr->qp_state = IB_QPS_RTS;
2670 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2671 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2672 		if (ret) {
2673 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2674 			       "RTS: %d\n", i, ret);
2675 			goto out;
2676 		}
2677 	}
2678 
2679 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2680 	if (ret) {
2681 		printk(KERN_ERR PFX "Failed to request completion "
2682 		       "notification: %d\n", ret);
2683 		goto out;
2684 	}
2685 
2686 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2687 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2688 		if (ret) {
2689 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2690 			goto out;
2691 		}
2692 	}
2693 out:
2694 	kfree(attr);
2695 	return ret;
2696 }
2697 
2698 static void qp_event_handler(struct ib_event *event, void *qp_context)
2699 {
2700 	struct ib_mad_qp_info	*qp_info = qp_context;
2701 
2702 	/* It's worse than that! He's dead, Jim! */
2703 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2704 		event->event, qp_info->qp->qp_num);
2705 }
2706 
2707 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2708 			   struct ib_mad_queue *mad_queue)
2709 {
2710 	mad_queue->qp_info = qp_info;
2711 	mad_queue->count = 0;
2712 	spin_lock_init(&mad_queue->lock);
2713 	INIT_LIST_HEAD(&mad_queue->list);
2714 }
2715 
2716 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2717 			struct ib_mad_qp_info *qp_info)
2718 {
2719 	qp_info->port_priv = port_priv;
2720 	init_mad_queue(qp_info, &qp_info->send_queue);
2721 	init_mad_queue(qp_info, &qp_info->recv_queue);
2722 	INIT_LIST_HEAD(&qp_info->overflow_list);
2723 	spin_lock_init(&qp_info->snoop_lock);
2724 	qp_info->snoop_table = NULL;
2725 	qp_info->snoop_table_size = 0;
2726 	atomic_set(&qp_info->snoop_count, 0);
2727 }
2728 
2729 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2730 			 enum ib_qp_type qp_type)
2731 {
2732 	struct ib_qp_init_attr	qp_init_attr;
2733 	int ret;
2734 
2735 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2736 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2737 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2738 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2739 	qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2740 	qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2741 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2742 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2743 	qp_init_attr.qp_type = qp_type;
2744 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2745 	qp_init_attr.qp_context = qp_info;
2746 	qp_init_attr.event_handler = qp_event_handler;
2747 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2748 	if (IS_ERR(qp_info->qp)) {
2749 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2750 		       get_spl_qp_index(qp_type));
2751 		ret = PTR_ERR(qp_info->qp);
2752 		goto error;
2753 	}
2754 	/* Use minimum queue sizes unless the CQ is resized */
2755 	qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2756 	qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2757 	return 0;
2758 
2759 error:
2760 	return ret;
2761 }
2762 
2763 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2764 {
2765 	ib_destroy_qp(qp_info->qp);
2766 	kfree(qp_info->snoop_table);
2767 }
2768 
2769 /*
2770  * Open the port
2771  * Create the QP, PD, MR, and CQ if needed
2772  */
2773 static int ib_mad_port_open(struct ib_device *device,
2774 			    int port_num)
2775 {
2776 	int ret, cq_size;
2777 	struct ib_mad_port_private *port_priv;
2778 	unsigned long flags;
2779 	char name[sizeof "ib_mad123"];
2780 
2781 	/* Create new device info */
2782 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2783 	if (!port_priv) {
2784 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2785 		return -ENOMEM;
2786 	}
2787 
2788 	port_priv->device = device;
2789 	port_priv->port_num = port_num;
2790 	spin_lock_init(&port_priv->reg_lock);
2791 	INIT_LIST_HEAD(&port_priv->agent_list);
2792 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2793 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2794 
2795 	cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2796 	port_priv->cq = ib_create_cq(port_priv->device,
2797 				     ib_mad_thread_completion_handler,
2798 				     NULL, port_priv, cq_size, 0);
2799 	if (IS_ERR(port_priv->cq)) {
2800 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2801 		ret = PTR_ERR(port_priv->cq);
2802 		goto error3;
2803 	}
2804 
2805 	port_priv->pd = ib_alloc_pd(device);
2806 	if (IS_ERR(port_priv->pd)) {
2807 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2808 		ret = PTR_ERR(port_priv->pd);
2809 		goto error4;
2810 	}
2811 
2812 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2813 	if (IS_ERR(port_priv->mr)) {
2814 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2815 		ret = PTR_ERR(port_priv->mr);
2816 		goto error5;
2817 	}
2818 
2819 	ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2820 	if (ret)
2821 		goto error6;
2822 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2823 	if (ret)
2824 		goto error7;
2825 
2826 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2827 	port_priv->wq = create_singlethread_workqueue(name);
2828 	if (!port_priv->wq) {
2829 		ret = -ENOMEM;
2830 		goto error8;
2831 	}
2832 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2833 
2834 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2835 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2836 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2837 
2838 	ret = ib_mad_port_start(port_priv);
2839 	if (ret) {
2840 		printk(KERN_ERR PFX "Couldn't start port\n");
2841 		goto error9;
2842 	}
2843 
2844 	return 0;
2845 
2846 error9:
2847 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2848 	list_del_init(&port_priv->port_list);
2849 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2850 
2851 	destroy_workqueue(port_priv->wq);
2852 error8:
2853 	destroy_mad_qp(&port_priv->qp_info[1]);
2854 error7:
2855 	destroy_mad_qp(&port_priv->qp_info[0]);
2856 error6:
2857 	ib_dereg_mr(port_priv->mr);
2858 error5:
2859 	ib_dealloc_pd(port_priv->pd);
2860 error4:
2861 	ib_destroy_cq(port_priv->cq);
2862 	cleanup_recv_queue(&port_priv->qp_info[1]);
2863 	cleanup_recv_queue(&port_priv->qp_info[0]);
2864 error3:
2865 	kfree(port_priv);
2866 
2867 	return ret;
2868 }
2869 
2870 /*
2871  * Close the port
2872  * If there are no classes using the port, free the port
2873  * resources (CQ, MR, PD, QP) and remove the port's info structure
2874  */
2875 static int ib_mad_port_close(struct ib_device *device, int port_num)
2876 {
2877 	struct ib_mad_port_private *port_priv;
2878 	unsigned long flags;
2879 
2880 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2881 	port_priv = __ib_get_mad_port(device, port_num);
2882 	if (port_priv == NULL) {
2883 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2884 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2885 		return -ENODEV;
2886 	}
2887 	list_del_init(&port_priv->port_list);
2888 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2889 
2890 	destroy_workqueue(port_priv->wq);
2891 	destroy_mad_qp(&port_priv->qp_info[1]);
2892 	destroy_mad_qp(&port_priv->qp_info[0]);
2893 	ib_dereg_mr(port_priv->mr);
2894 	ib_dealloc_pd(port_priv->pd);
2895 	ib_destroy_cq(port_priv->cq);
2896 	cleanup_recv_queue(&port_priv->qp_info[1]);
2897 	cleanup_recv_queue(&port_priv->qp_info[0]);
2898 	/* XXX: Handle deallocation of MAD registration tables */
2899 
2900 	kfree(port_priv);
2901 
2902 	return 0;
2903 }
2904 
2905 static void ib_mad_init_device(struct ib_device *device)
2906 {
2907 	int start, end, i;
2908 
2909 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2910 		return;
2911 
2912 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2913 		start = 0;
2914 		end   = 0;
2915 	} else {
2916 		start = 1;
2917 		end   = device->phys_port_cnt;
2918 	}
2919 
2920 	for (i = start; i <= end; i++) {
2921 		if (ib_mad_port_open(device, i)) {
2922 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2923 			       device->name, i);
2924 			goto error;
2925 		}
2926 		if (ib_agent_port_open(device, i)) {
2927 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2928 			       "for agents\n",
2929 			       device->name, i);
2930 			goto error_agent;
2931 		}
2932 	}
2933 	return;
2934 
2935 error_agent:
2936 	if (ib_mad_port_close(device, i))
2937 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2938 		       device->name, i);
2939 
2940 error:
2941 	i--;
2942 
2943 	while (i >= start) {
2944 		if (ib_agent_port_close(device, i))
2945 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2946 			       "for agents\n",
2947 			       device->name, i);
2948 		if (ib_mad_port_close(device, i))
2949 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2950 			       device->name, i);
2951 		i--;
2952 	}
2953 }
2954 
2955 static void ib_mad_remove_device(struct ib_device *device)
2956 {
2957 	int i, num_ports, cur_port;
2958 
2959 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2960 		num_ports = 1;
2961 		cur_port = 0;
2962 	} else {
2963 		num_ports = device->phys_port_cnt;
2964 		cur_port = 1;
2965 	}
2966 	for (i = 0; i < num_ports; i++, cur_port++) {
2967 		if (ib_agent_port_close(device, cur_port))
2968 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2969 			       "for agents\n",
2970 			       device->name, cur_port);
2971 		if (ib_mad_port_close(device, cur_port))
2972 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2973 			       device->name, cur_port);
2974 	}
2975 }
2976 
2977 static struct ib_client mad_client = {
2978 	.name   = "mad",
2979 	.add = ib_mad_init_device,
2980 	.remove = ib_mad_remove_device
2981 };
2982 
2983 static int __init ib_mad_init_module(void)
2984 {
2985 	int ret;
2986 
2987 	spin_lock_init(&ib_mad_port_list_lock);
2988 
2989 	ib_mad_cache = kmem_cache_create("ib_mad",
2990 					 sizeof(struct ib_mad_private),
2991 					 0,
2992 					 SLAB_HWCACHE_ALIGN,
2993 					 NULL);
2994 	if (!ib_mad_cache) {
2995 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2996 		ret = -ENOMEM;
2997 		goto error1;
2998 	}
2999 
3000 	INIT_LIST_HEAD(&ib_mad_port_list);
3001 
3002 	if (ib_register_client(&mad_client)) {
3003 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3004 		ret = -EINVAL;
3005 		goto error2;
3006 	}
3007 
3008 	return 0;
3009 
3010 error2:
3011 	kmem_cache_destroy(ib_mad_cache);
3012 error1:
3013 	return ret;
3014 }
3015 
3016 static void __exit ib_mad_cleanup_module(void)
3017 {
3018 	ib_unregister_client(&mad_client);
3019 	kmem_cache_destroy(ib_mad_cache);
3020 }
3021 
3022 module_init(ib_mad_init_module);
3023 module_exit(ib_mad_cleanup_module);
3024 
3025