xref: /linux/drivers/infiniband/core/mad.c (revision 12871a0bd67dd4db4418e1daafcd46e9d329ef10)
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  * Copyright (c) 2009 HNR Consulting. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  *
35  */
36 #include <linux/dma-mapping.h>
37 #include <linux/slab.h>
38 #include <rdma/ib_cache.h>
39 
40 #include "mad_priv.h"
41 #include "mad_rmpp.h"
42 #include "smi.h"
43 #include "agent.h"
44 
45 MODULE_LICENSE("Dual BSD/GPL");
46 MODULE_DESCRIPTION("kernel IB MAD API");
47 MODULE_AUTHOR("Hal Rosenstock");
48 MODULE_AUTHOR("Sean Hefty");
49 
50 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
51 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
52 
53 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
54 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
55 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
56 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
57 
58 static struct kmem_cache *ib_mad_cache;
59 
60 static struct list_head ib_mad_port_list;
61 static u32 ib_mad_client_id = 0;
62 
63 /* Port list lock */
64 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
65 
66 /* Forward declarations */
67 static int method_in_use(struct ib_mad_mgmt_method_table **method,
68 			 struct ib_mad_reg_req *mad_reg_req);
69 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
70 static struct ib_mad_agent_private *find_mad_agent(
71 					struct ib_mad_port_private *port_priv,
72 					struct ib_mad *mad);
73 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
74 				    struct ib_mad_private *mad);
75 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
76 static void timeout_sends(struct work_struct *work);
77 static void local_completions(struct work_struct *work);
78 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
79 			      struct ib_mad_agent_private *agent_priv,
80 			      u8 mgmt_class);
81 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
82 			   struct ib_mad_agent_private *agent_priv);
83 
84 /*
85  * Returns a ib_mad_port_private structure or NULL for a device/port
86  * Assumes ib_mad_port_list_lock is being held
87  */
88 static inline struct ib_mad_port_private *
89 __ib_get_mad_port(struct ib_device *device, int port_num)
90 {
91 	struct ib_mad_port_private *entry;
92 
93 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
94 		if (entry->device == device && entry->port_num == port_num)
95 			return entry;
96 	}
97 	return NULL;
98 }
99 
100 /*
101  * Wrapper function to return a ib_mad_port_private structure or NULL
102  * for a device/port
103  */
104 static inline struct ib_mad_port_private *
105 ib_get_mad_port(struct ib_device *device, int port_num)
106 {
107 	struct ib_mad_port_private *entry;
108 	unsigned long flags;
109 
110 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
111 	entry = __ib_get_mad_port(device, port_num);
112 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
113 
114 	return entry;
115 }
116 
117 static inline u8 convert_mgmt_class(u8 mgmt_class)
118 {
119 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
120 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
121 		0 : mgmt_class;
122 }
123 
124 static int get_spl_qp_index(enum ib_qp_type qp_type)
125 {
126 	switch (qp_type)
127 	{
128 	case IB_QPT_SMI:
129 		return 0;
130 	case IB_QPT_GSI:
131 		return 1;
132 	default:
133 		return -1;
134 	}
135 }
136 
137 static int vendor_class_index(u8 mgmt_class)
138 {
139 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
140 }
141 
142 static int is_vendor_class(u8 mgmt_class)
143 {
144 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
145 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
146 		return 0;
147 	return 1;
148 }
149 
150 static int is_vendor_oui(char *oui)
151 {
152 	if (oui[0] || oui[1] || oui[2])
153 		return 1;
154 	return 0;
155 }
156 
157 static int is_vendor_method_in_use(
158 		struct ib_mad_mgmt_vendor_class *vendor_class,
159 		struct ib_mad_reg_req *mad_reg_req)
160 {
161 	struct ib_mad_mgmt_method_table *method;
162 	int i;
163 
164 	for (i = 0; i < MAX_MGMT_OUI; i++) {
165 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
166 			method = vendor_class->method_table[i];
167 			if (method) {
168 				if (method_in_use(&method, mad_reg_req))
169 					return 1;
170 				else
171 					break;
172 			}
173 		}
174 	}
175 	return 0;
176 }
177 
178 int ib_response_mad(struct ib_mad *mad)
179 {
180 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
181 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
182 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
183 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
184 }
185 EXPORT_SYMBOL(ib_response_mad);
186 
187 /*
188  * ib_register_mad_agent - Register to send/receive MADs
189  */
190 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
191 					   u8 port_num,
192 					   enum ib_qp_type qp_type,
193 					   struct ib_mad_reg_req *mad_reg_req,
194 					   u8 rmpp_version,
195 					   ib_mad_send_handler send_handler,
196 					   ib_mad_recv_handler recv_handler,
197 					   void *context)
198 {
199 	struct ib_mad_port_private *port_priv;
200 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
201 	struct ib_mad_agent_private *mad_agent_priv;
202 	struct ib_mad_reg_req *reg_req = NULL;
203 	struct ib_mad_mgmt_class_table *class;
204 	struct ib_mad_mgmt_vendor_class_table *vendor;
205 	struct ib_mad_mgmt_vendor_class *vendor_class;
206 	struct ib_mad_mgmt_method_table *method;
207 	int ret2, qpn;
208 	unsigned long flags;
209 	u8 mgmt_class, vclass;
210 
211 	/* Validate parameters */
212 	qpn = get_spl_qp_index(qp_type);
213 	if (qpn == -1)
214 		goto error1;
215 
216 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
217 		goto error1;
218 
219 	/* Validate MAD registration request if supplied */
220 	if (mad_reg_req) {
221 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
222 			goto error1;
223 		if (!recv_handler)
224 			goto error1;
225 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
226 			/*
227 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
228 			 * one in this range currently allowed
229 			 */
230 			if (mad_reg_req->mgmt_class !=
231 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
232 				goto error1;
233 		} else if (mad_reg_req->mgmt_class == 0) {
234 			/*
235 			 * Class 0 is reserved in IBA and is used for
236 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
237 			 */
238 			goto error1;
239 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
240 			/*
241 			 * If class is in "new" vendor range,
242 			 * ensure supplied OUI is not zero
243 			 */
244 			if (!is_vendor_oui(mad_reg_req->oui))
245 				goto error1;
246 		}
247 		/* Make sure class supplied is consistent with RMPP */
248 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
249 			if (rmpp_version)
250 				goto error1;
251 		}
252 		/* Make sure class supplied is consistent with QP type */
253 		if (qp_type == IB_QPT_SMI) {
254 			if ((mad_reg_req->mgmt_class !=
255 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
256 			    (mad_reg_req->mgmt_class !=
257 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
258 				goto error1;
259 		} else {
260 			if ((mad_reg_req->mgmt_class ==
261 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
262 			    (mad_reg_req->mgmt_class ==
263 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
264 				goto error1;
265 		}
266 	} else {
267 		/* No registration request supplied */
268 		if (!send_handler)
269 			goto error1;
270 	}
271 
272 	/* Validate device and port */
273 	port_priv = ib_get_mad_port(device, port_num);
274 	if (!port_priv) {
275 		ret = ERR_PTR(-ENODEV);
276 		goto error1;
277 	}
278 
279 	/* Verify the QP requested is supported.  For example, Ethernet devices
280 	 * will not have QP0 */
281 	if (!port_priv->qp_info[qpn].qp) {
282 		ret = ERR_PTR(-EPROTONOSUPPORT);
283 		goto error1;
284 	}
285 
286 	/* Allocate structures */
287 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
288 	if (!mad_agent_priv) {
289 		ret = ERR_PTR(-ENOMEM);
290 		goto error1;
291 	}
292 
293 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
294 						 IB_ACCESS_LOCAL_WRITE);
295 	if (IS_ERR(mad_agent_priv->agent.mr)) {
296 		ret = ERR_PTR(-ENOMEM);
297 		goto error2;
298 	}
299 
300 	if (mad_reg_req) {
301 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
302 		if (!reg_req) {
303 			ret = ERR_PTR(-ENOMEM);
304 			goto error3;
305 		}
306 	}
307 
308 	/* Now, fill in the various structures */
309 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
310 	mad_agent_priv->reg_req = reg_req;
311 	mad_agent_priv->agent.rmpp_version = rmpp_version;
312 	mad_agent_priv->agent.device = device;
313 	mad_agent_priv->agent.recv_handler = recv_handler;
314 	mad_agent_priv->agent.send_handler = send_handler;
315 	mad_agent_priv->agent.context = context;
316 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
317 	mad_agent_priv->agent.port_num = port_num;
318 	spin_lock_init(&mad_agent_priv->lock);
319 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
320 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
321 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
322 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
323 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
324 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
325 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
326 	atomic_set(&mad_agent_priv->refcount, 1);
327 	init_completion(&mad_agent_priv->comp);
328 
329 	spin_lock_irqsave(&port_priv->reg_lock, flags);
330 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
331 
332 	/*
333 	 * Make sure MAD registration (if supplied)
334 	 * is non overlapping with any existing ones
335 	 */
336 	if (mad_reg_req) {
337 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
338 		if (!is_vendor_class(mgmt_class)) {
339 			class = port_priv->version[mad_reg_req->
340 						   mgmt_class_version].class;
341 			if (class) {
342 				method = class->method_table[mgmt_class];
343 				if (method) {
344 					if (method_in_use(&method,
345 							   mad_reg_req))
346 						goto error4;
347 				}
348 			}
349 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
350 						  mgmt_class);
351 		} else {
352 			/* "New" vendor class range */
353 			vendor = port_priv->version[mad_reg_req->
354 						    mgmt_class_version].vendor;
355 			if (vendor) {
356 				vclass = vendor_class_index(mgmt_class);
357 				vendor_class = vendor->vendor_class[vclass];
358 				if (vendor_class) {
359 					if (is_vendor_method_in_use(
360 							vendor_class,
361 							mad_reg_req))
362 						goto error4;
363 				}
364 			}
365 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
366 		}
367 		if (ret2) {
368 			ret = ERR_PTR(ret2);
369 			goto error4;
370 		}
371 	}
372 
373 	/* Add mad agent into port's agent list */
374 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
375 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
376 
377 	return &mad_agent_priv->agent;
378 
379 error4:
380 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
381 	kfree(reg_req);
382 error3:
383 	ib_dereg_mr(mad_agent_priv->agent.mr);
384 error2:
385 	kfree(mad_agent_priv);
386 error1:
387 	return ret;
388 }
389 EXPORT_SYMBOL(ib_register_mad_agent);
390 
391 static inline int is_snooping_sends(int mad_snoop_flags)
392 {
393 	return (mad_snoop_flags &
394 		(/*IB_MAD_SNOOP_POSTED_SENDS |
395 		 IB_MAD_SNOOP_RMPP_SENDS |*/
396 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
397 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
398 }
399 
400 static inline int is_snooping_recvs(int mad_snoop_flags)
401 {
402 	return (mad_snoop_flags &
403 		(IB_MAD_SNOOP_RECVS /*|
404 		 IB_MAD_SNOOP_RMPP_RECVS*/));
405 }
406 
407 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
408 				struct ib_mad_snoop_private *mad_snoop_priv)
409 {
410 	struct ib_mad_snoop_private **new_snoop_table;
411 	unsigned long flags;
412 	int i;
413 
414 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
415 	/* Check for empty slot in array. */
416 	for (i = 0; i < qp_info->snoop_table_size; i++)
417 		if (!qp_info->snoop_table[i])
418 			break;
419 
420 	if (i == qp_info->snoop_table_size) {
421 		/* Grow table. */
422 		new_snoop_table = krealloc(qp_info->snoop_table,
423 					   sizeof mad_snoop_priv *
424 					   (qp_info->snoop_table_size + 1),
425 					   GFP_ATOMIC);
426 		if (!new_snoop_table) {
427 			i = -ENOMEM;
428 			goto out;
429 		}
430 
431 		qp_info->snoop_table = new_snoop_table;
432 		qp_info->snoop_table_size++;
433 	}
434 	qp_info->snoop_table[i] = mad_snoop_priv;
435 	atomic_inc(&qp_info->snoop_count);
436 out:
437 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
438 	return i;
439 }
440 
441 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
442 					   u8 port_num,
443 					   enum ib_qp_type qp_type,
444 					   int mad_snoop_flags,
445 					   ib_mad_snoop_handler snoop_handler,
446 					   ib_mad_recv_handler recv_handler,
447 					   void *context)
448 {
449 	struct ib_mad_port_private *port_priv;
450 	struct ib_mad_agent *ret;
451 	struct ib_mad_snoop_private *mad_snoop_priv;
452 	int qpn;
453 
454 	/* Validate parameters */
455 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
456 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
457 		ret = ERR_PTR(-EINVAL);
458 		goto error1;
459 	}
460 	qpn = get_spl_qp_index(qp_type);
461 	if (qpn == -1) {
462 		ret = ERR_PTR(-EINVAL);
463 		goto error1;
464 	}
465 	port_priv = ib_get_mad_port(device, port_num);
466 	if (!port_priv) {
467 		ret = ERR_PTR(-ENODEV);
468 		goto error1;
469 	}
470 	/* Allocate structures */
471 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
472 	if (!mad_snoop_priv) {
473 		ret = ERR_PTR(-ENOMEM);
474 		goto error1;
475 	}
476 
477 	/* Now, fill in the various structures */
478 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
479 	mad_snoop_priv->agent.device = device;
480 	mad_snoop_priv->agent.recv_handler = recv_handler;
481 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
482 	mad_snoop_priv->agent.context = context;
483 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
484 	mad_snoop_priv->agent.port_num = port_num;
485 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
486 	init_completion(&mad_snoop_priv->comp);
487 	mad_snoop_priv->snoop_index = register_snoop_agent(
488 						&port_priv->qp_info[qpn],
489 						mad_snoop_priv);
490 	if (mad_snoop_priv->snoop_index < 0) {
491 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
492 		goto error2;
493 	}
494 
495 	atomic_set(&mad_snoop_priv->refcount, 1);
496 	return &mad_snoop_priv->agent;
497 
498 error2:
499 	kfree(mad_snoop_priv);
500 error1:
501 	return ret;
502 }
503 EXPORT_SYMBOL(ib_register_mad_snoop);
504 
505 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
506 {
507 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
508 		complete(&mad_agent_priv->comp);
509 }
510 
511 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
512 {
513 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
514 		complete(&mad_snoop_priv->comp);
515 }
516 
517 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
518 {
519 	struct ib_mad_port_private *port_priv;
520 	unsigned long flags;
521 
522 	/* Note that we could still be handling received MADs */
523 
524 	/*
525 	 * Canceling all sends results in dropping received response
526 	 * MADs, preventing us from queuing additional work
527 	 */
528 	cancel_mads(mad_agent_priv);
529 	port_priv = mad_agent_priv->qp_info->port_priv;
530 	cancel_delayed_work(&mad_agent_priv->timed_work);
531 
532 	spin_lock_irqsave(&port_priv->reg_lock, flags);
533 	remove_mad_reg_req(mad_agent_priv);
534 	list_del(&mad_agent_priv->agent_list);
535 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
536 
537 	flush_workqueue(port_priv->wq);
538 	ib_cancel_rmpp_recvs(mad_agent_priv);
539 
540 	deref_mad_agent(mad_agent_priv);
541 	wait_for_completion(&mad_agent_priv->comp);
542 
543 	kfree(mad_agent_priv->reg_req);
544 	ib_dereg_mr(mad_agent_priv->agent.mr);
545 	kfree(mad_agent_priv);
546 }
547 
548 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
549 {
550 	struct ib_mad_qp_info *qp_info;
551 	unsigned long flags;
552 
553 	qp_info = mad_snoop_priv->qp_info;
554 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
555 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
556 	atomic_dec(&qp_info->snoop_count);
557 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
558 
559 	deref_snoop_agent(mad_snoop_priv);
560 	wait_for_completion(&mad_snoop_priv->comp);
561 
562 	kfree(mad_snoop_priv);
563 }
564 
565 /*
566  * ib_unregister_mad_agent - Unregisters a client from using MAD services
567  */
568 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
569 {
570 	struct ib_mad_agent_private *mad_agent_priv;
571 	struct ib_mad_snoop_private *mad_snoop_priv;
572 
573 	/* If the TID is zero, the agent can only snoop. */
574 	if (mad_agent->hi_tid) {
575 		mad_agent_priv = container_of(mad_agent,
576 					      struct ib_mad_agent_private,
577 					      agent);
578 		unregister_mad_agent(mad_agent_priv);
579 	} else {
580 		mad_snoop_priv = container_of(mad_agent,
581 					      struct ib_mad_snoop_private,
582 					      agent);
583 		unregister_mad_snoop(mad_snoop_priv);
584 	}
585 	return 0;
586 }
587 EXPORT_SYMBOL(ib_unregister_mad_agent);
588 
589 static void dequeue_mad(struct ib_mad_list_head *mad_list)
590 {
591 	struct ib_mad_queue *mad_queue;
592 	unsigned long flags;
593 
594 	BUG_ON(!mad_list->mad_queue);
595 	mad_queue = mad_list->mad_queue;
596 	spin_lock_irqsave(&mad_queue->lock, flags);
597 	list_del(&mad_list->list);
598 	mad_queue->count--;
599 	spin_unlock_irqrestore(&mad_queue->lock, flags);
600 }
601 
602 static void snoop_send(struct ib_mad_qp_info *qp_info,
603 		       struct ib_mad_send_buf *send_buf,
604 		       struct ib_mad_send_wc *mad_send_wc,
605 		       int mad_snoop_flags)
606 {
607 	struct ib_mad_snoop_private *mad_snoop_priv;
608 	unsigned long flags;
609 	int i;
610 
611 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
612 	for (i = 0; i < qp_info->snoop_table_size; i++) {
613 		mad_snoop_priv = qp_info->snoop_table[i];
614 		if (!mad_snoop_priv ||
615 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
616 			continue;
617 
618 		atomic_inc(&mad_snoop_priv->refcount);
619 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
620 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
621 						    send_buf, mad_send_wc);
622 		deref_snoop_agent(mad_snoop_priv);
623 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
624 	}
625 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
626 }
627 
628 static void snoop_recv(struct ib_mad_qp_info *qp_info,
629 		       struct ib_mad_recv_wc *mad_recv_wc,
630 		       int mad_snoop_flags)
631 {
632 	struct ib_mad_snoop_private *mad_snoop_priv;
633 	unsigned long flags;
634 	int i;
635 
636 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
637 	for (i = 0; i < qp_info->snoop_table_size; i++) {
638 		mad_snoop_priv = qp_info->snoop_table[i];
639 		if (!mad_snoop_priv ||
640 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
641 			continue;
642 
643 		atomic_inc(&mad_snoop_priv->refcount);
644 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
645 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
646 						   mad_recv_wc);
647 		deref_snoop_agent(mad_snoop_priv);
648 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
649 	}
650 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
651 }
652 
653 static void build_smp_wc(struct ib_qp *qp,
654 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
655 			 struct ib_wc *wc)
656 {
657 	memset(wc, 0, sizeof *wc);
658 	wc->wr_id = wr_id;
659 	wc->status = IB_WC_SUCCESS;
660 	wc->opcode = IB_WC_RECV;
661 	wc->pkey_index = pkey_index;
662 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
663 	wc->src_qp = IB_QP0;
664 	wc->qp = qp;
665 	wc->slid = slid;
666 	wc->sl = 0;
667 	wc->dlid_path_bits = 0;
668 	wc->port_num = port_num;
669 }
670 
671 /*
672  * Return 0 if SMP is to be sent
673  * Return 1 if SMP was consumed locally (whether or not solicited)
674  * Return < 0 if error
675  */
676 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
677 				  struct ib_mad_send_wr_private *mad_send_wr)
678 {
679 	int ret = 0;
680 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
681 	unsigned long flags;
682 	struct ib_mad_local_private *local;
683 	struct ib_mad_private *mad_priv;
684 	struct ib_mad_port_private *port_priv;
685 	struct ib_mad_agent_private *recv_mad_agent = NULL;
686 	struct ib_device *device = mad_agent_priv->agent.device;
687 	u8 port_num;
688 	struct ib_wc mad_wc;
689 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
690 
691 	if (device->node_type == RDMA_NODE_IB_SWITCH &&
692 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
693 		port_num = send_wr->wr.ud.port_num;
694 	else
695 		port_num = mad_agent_priv->agent.port_num;
696 
697 	/*
698 	 * Directed route handling starts if the initial LID routed part of
699 	 * a request or the ending LID routed part of a response is empty.
700 	 * If we are at the start of the LID routed part, don't update the
701 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
702 	 */
703 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
704 	     IB_LID_PERMISSIVE &&
705 	     smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
706 	     IB_SMI_DISCARD) {
707 		ret = -EINVAL;
708 		printk(KERN_ERR PFX "Invalid directed route\n");
709 		goto out;
710 	}
711 
712 	/* Check to post send on QP or process locally */
713 	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
714 	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
715 		goto out;
716 
717 	local = kmalloc(sizeof *local, GFP_ATOMIC);
718 	if (!local) {
719 		ret = -ENOMEM;
720 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
721 		goto out;
722 	}
723 	local->mad_priv = NULL;
724 	local->recv_mad_agent = NULL;
725 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
726 	if (!mad_priv) {
727 		ret = -ENOMEM;
728 		printk(KERN_ERR PFX "No memory for local response MAD\n");
729 		kfree(local);
730 		goto out;
731 	}
732 
733 	build_smp_wc(mad_agent_priv->agent.qp,
734 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
735 		     send_wr->wr.ud.pkey_index,
736 		     send_wr->wr.ud.port_num, &mad_wc);
737 
738 	/* No GRH for DR SMP */
739 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
740 				  (struct ib_mad *)smp,
741 				  (struct ib_mad *)&mad_priv->mad);
742 	switch (ret)
743 	{
744 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
745 		if (ib_response_mad(&mad_priv->mad.mad) &&
746 		    mad_agent_priv->agent.recv_handler) {
747 			local->mad_priv = mad_priv;
748 			local->recv_mad_agent = mad_agent_priv;
749 			/*
750 			 * Reference MAD agent until receive
751 			 * side of local completion handled
752 			 */
753 			atomic_inc(&mad_agent_priv->refcount);
754 		} else
755 			kmem_cache_free(ib_mad_cache, mad_priv);
756 		break;
757 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
758 		kmem_cache_free(ib_mad_cache, mad_priv);
759 		break;
760 	case IB_MAD_RESULT_SUCCESS:
761 		/* Treat like an incoming receive MAD */
762 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
763 					    mad_agent_priv->agent.port_num);
764 		if (port_priv) {
765 			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
766 			recv_mad_agent = find_mad_agent(port_priv,
767 						        &mad_priv->mad.mad);
768 		}
769 		if (!port_priv || !recv_mad_agent) {
770 			/*
771 			 * No receiving agent so drop packet and
772 			 * generate send completion.
773 			 */
774 			kmem_cache_free(ib_mad_cache, mad_priv);
775 			break;
776 		}
777 		local->mad_priv = mad_priv;
778 		local->recv_mad_agent = recv_mad_agent;
779 		break;
780 	default:
781 		kmem_cache_free(ib_mad_cache, mad_priv);
782 		kfree(local);
783 		ret = -EINVAL;
784 		goto out;
785 	}
786 
787 	local->mad_send_wr = mad_send_wr;
788 	/* Reference MAD agent until send side of local completion handled */
789 	atomic_inc(&mad_agent_priv->refcount);
790 	/* Queue local completion to local list */
791 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
792 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
793 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
794 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
795 		   &mad_agent_priv->local_work);
796 	ret = 1;
797 out:
798 	return ret;
799 }
800 
801 static int get_pad_size(int hdr_len, int data_len)
802 {
803 	int seg_size, pad;
804 
805 	seg_size = sizeof(struct ib_mad) - hdr_len;
806 	if (data_len && seg_size) {
807 		pad = seg_size - data_len % seg_size;
808 		return pad == seg_size ? 0 : pad;
809 	} else
810 		return seg_size;
811 }
812 
813 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
814 {
815 	struct ib_rmpp_segment *s, *t;
816 
817 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
818 		list_del(&s->list);
819 		kfree(s);
820 	}
821 }
822 
823 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
824 				gfp_t gfp_mask)
825 {
826 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
827 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
828 	struct ib_rmpp_segment *seg = NULL;
829 	int left, seg_size, pad;
830 
831 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
832 	seg_size = send_buf->seg_size;
833 	pad = send_wr->pad;
834 
835 	/* Allocate data segments. */
836 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
837 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
838 		if (!seg) {
839 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
840 			       "alloc failed for len %zd, gfp %#x\n",
841 			       sizeof (*seg) + seg_size, gfp_mask);
842 			free_send_rmpp_list(send_wr);
843 			return -ENOMEM;
844 		}
845 		seg->num = ++send_buf->seg_count;
846 		list_add_tail(&seg->list, &send_wr->rmpp_list);
847 	}
848 
849 	/* Zero any padding */
850 	if (pad)
851 		memset(seg->data + seg_size - pad, 0, pad);
852 
853 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
854 					  agent.rmpp_version;
855 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
856 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
857 
858 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
859 					struct ib_rmpp_segment, list);
860 	send_wr->last_ack_seg = send_wr->cur_seg;
861 	return 0;
862 }
863 
864 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
865 					    u32 remote_qpn, u16 pkey_index,
866 					    int rmpp_active,
867 					    int hdr_len, int data_len,
868 					    gfp_t gfp_mask)
869 {
870 	struct ib_mad_agent_private *mad_agent_priv;
871 	struct ib_mad_send_wr_private *mad_send_wr;
872 	int pad, message_size, ret, size;
873 	void *buf;
874 
875 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
876 				      agent);
877 	pad = get_pad_size(hdr_len, data_len);
878 	message_size = hdr_len + data_len + pad;
879 
880 	if ((!mad_agent->rmpp_version &&
881 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
882 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
883 		return ERR_PTR(-EINVAL);
884 
885 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
886 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
887 	if (!buf)
888 		return ERR_PTR(-ENOMEM);
889 
890 	mad_send_wr = buf + size;
891 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
892 	mad_send_wr->send_buf.mad = buf;
893 	mad_send_wr->send_buf.hdr_len = hdr_len;
894 	mad_send_wr->send_buf.data_len = data_len;
895 	mad_send_wr->pad = pad;
896 
897 	mad_send_wr->mad_agent_priv = mad_agent_priv;
898 	mad_send_wr->sg_list[0].length = hdr_len;
899 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
900 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
901 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
902 
903 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
904 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
905 	mad_send_wr->send_wr.num_sge = 2;
906 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
907 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
908 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
909 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
910 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
911 
912 	if (rmpp_active) {
913 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
914 		if (ret) {
915 			kfree(buf);
916 			return ERR_PTR(ret);
917 		}
918 	}
919 
920 	mad_send_wr->send_buf.mad_agent = mad_agent;
921 	atomic_inc(&mad_agent_priv->refcount);
922 	return &mad_send_wr->send_buf;
923 }
924 EXPORT_SYMBOL(ib_create_send_mad);
925 
926 int ib_get_mad_data_offset(u8 mgmt_class)
927 {
928 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
929 		return IB_MGMT_SA_HDR;
930 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
931 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
932 		 (mgmt_class == IB_MGMT_CLASS_BIS))
933 		return IB_MGMT_DEVICE_HDR;
934 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
935 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
936 		return IB_MGMT_VENDOR_HDR;
937 	else
938 		return IB_MGMT_MAD_HDR;
939 }
940 EXPORT_SYMBOL(ib_get_mad_data_offset);
941 
942 int ib_is_mad_class_rmpp(u8 mgmt_class)
943 {
944 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
945 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
946 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
947 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
948 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
949 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
950 		return 1;
951 	return 0;
952 }
953 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
954 
955 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
956 {
957 	struct ib_mad_send_wr_private *mad_send_wr;
958 	struct list_head *list;
959 
960 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
961 				   send_buf);
962 	list = &mad_send_wr->cur_seg->list;
963 
964 	if (mad_send_wr->cur_seg->num < seg_num) {
965 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
966 			if (mad_send_wr->cur_seg->num == seg_num)
967 				break;
968 	} else if (mad_send_wr->cur_seg->num > seg_num) {
969 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
970 			if (mad_send_wr->cur_seg->num == seg_num)
971 				break;
972 	}
973 	return mad_send_wr->cur_seg->data;
974 }
975 EXPORT_SYMBOL(ib_get_rmpp_segment);
976 
977 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
978 {
979 	if (mad_send_wr->send_buf.seg_count)
980 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
981 					   mad_send_wr->seg_num);
982 	else
983 		return mad_send_wr->send_buf.mad +
984 		       mad_send_wr->send_buf.hdr_len;
985 }
986 
987 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
988 {
989 	struct ib_mad_agent_private *mad_agent_priv;
990 	struct ib_mad_send_wr_private *mad_send_wr;
991 
992 	mad_agent_priv = container_of(send_buf->mad_agent,
993 				      struct ib_mad_agent_private, agent);
994 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
995 				   send_buf);
996 
997 	free_send_rmpp_list(mad_send_wr);
998 	kfree(send_buf->mad);
999 	deref_mad_agent(mad_agent_priv);
1000 }
1001 EXPORT_SYMBOL(ib_free_send_mad);
1002 
1003 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1004 {
1005 	struct ib_mad_qp_info *qp_info;
1006 	struct list_head *list;
1007 	struct ib_send_wr *bad_send_wr;
1008 	struct ib_mad_agent *mad_agent;
1009 	struct ib_sge *sge;
1010 	unsigned long flags;
1011 	int ret;
1012 
1013 	/* Set WR ID to find mad_send_wr upon completion */
1014 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1015 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1016 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1017 
1018 	mad_agent = mad_send_wr->send_buf.mad_agent;
1019 	sge = mad_send_wr->sg_list;
1020 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1021 					mad_send_wr->send_buf.mad,
1022 					sge[0].length,
1023 					DMA_TO_DEVICE);
1024 	mad_send_wr->header_mapping = sge[0].addr;
1025 
1026 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1027 					ib_get_payload(mad_send_wr),
1028 					sge[1].length,
1029 					DMA_TO_DEVICE);
1030 	mad_send_wr->payload_mapping = sge[1].addr;
1031 
1032 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1033 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1034 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1035 				   &bad_send_wr);
1036 		list = &qp_info->send_queue.list;
1037 	} else {
1038 		ret = 0;
1039 		list = &qp_info->overflow_list;
1040 	}
1041 
1042 	if (!ret) {
1043 		qp_info->send_queue.count++;
1044 		list_add_tail(&mad_send_wr->mad_list.list, list);
1045 	}
1046 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1047 	if (ret) {
1048 		ib_dma_unmap_single(mad_agent->device,
1049 				    mad_send_wr->header_mapping,
1050 				    sge[0].length, DMA_TO_DEVICE);
1051 		ib_dma_unmap_single(mad_agent->device,
1052 				    mad_send_wr->payload_mapping,
1053 				    sge[1].length, DMA_TO_DEVICE);
1054 	}
1055 	return ret;
1056 }
1057 
1058 /*
1059  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1060  *  with the registered client
1061  */
1062 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1063 		     struct ib_mad_send_buf **bad_send_buf)
1064 {
1065 	struct ib_mad_agent_private *mad_agent_priv;
1066 	struct ib_mad_send_buf *next_send_buf;
1067 	struct ib_mad_send_wr_private *mad_send_wr;
1068 	unsigned long flags;
1069 	int ret = -EINVAL;
1070 
1071 	/* Walk list of send WRs and post each on send list */
1072 	for (; send_buf; send_buf = next_send_buf) {
1073 
1074 		mad_send_wr = container_of(send_buf,
1075 					   struct ib_mad_send_wr_private,
1076 					   send_buf);
1077 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1078 
1079 		if (!send_buf->mad_agent->send_handler ||
1080 		    (send_buf->timeout_ms &&
1081 		     !send_buf->mad_agent->recv_handler)) {
1082 			ret = -EINVAL;
1083 			goto error;
1084 		}
1085 
1086 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1087 			if (mad_agent_priv->agent.rmpp_version) {
1088 				ret = -EINVAL;
1089 				goto error;
1090 			}
1091 		}
1092 
1093 		/*
1094 		 * Save pointer to next work request to post in case the
1095 		 * current one completes, and the user modifies the work
1096 		 * request associated with the completion
1097 		 */
1098 		next_send_buf = send_buf->next;
1099 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1100 
1101 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1102 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1103 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1104 						     mad_send_wr);
1105 			if (ret < 0)		/* error */
1106 				goto error;
1107 			else if (ret == 1)	/* locally consumed */
1108 				continue;
1109 		}
1110 
1111 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1112 		/* Timeout will be updated after send completes */
1113 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1114 		mad_send_wr->max_retries = send_buf->retries;
1115 		mad_send_wr->retries_left = send_buf->retries;
1116 		send_buf->retries = 0;
1117 		/* Reference for work request to QP + response */
1118 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1119 		mad_send_wr->status = IB_WC_SUCCESS;
1120 
1121 		/* Reference MAD agent until send completes */
1122 		atomic_inc(&mad_agent_priv->refcount);
1123 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1124 		list_add_tail(&mad_send_wr->agent_list,
1125 			      &mad_agent_priv->send_list);
1126 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1127 
1128 		if (mad_agent_priv->agent.rmpp_version) {
1129 			ret = ib_send_rmpp_mad(mad_send_wr);
1130 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1131 				ret = ib_send_mad(mad_send_wr);
1132 		} else
1133 			ret = ib_send_mad(mad_send_wr);
1134 		if (ret < 0) {
1135 			/* Fail send request */
1136 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1137 			list_del(&mad_send_wr->agent_list);
1138 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1139 			atomic_dec(&mad_agent_priv->refcount);
1140 			goto error;
1141 		}
1142 	}
1143 	return 0;
1144 error:
1145 	if (bad_send_buf)
1146 		*bad_send_buf = send_buf;
1147 	return ret;
1148 }
1149 EXPORT_SYMBOL(ib_post_send_mad);
1150 
1151 /*
1152  * ib_free_recv_mad - Returns data buffers used to receive
1153  *  a MAD to the access layer
1154  */
1155 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1156 {
1157 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1158 	struct ib_mad_private_header *mad_priv_hdr;
1159 	struct ib_mad_private *priv;
1160 	struct list_head free_list;
1161 
1162 	INIT_LIST_HEAD(&free_list);
1163 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1164 
1165 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1166 					&free_list, list) {
1167 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1168 					   recv_buf);
1169 		mad_priv_hdr = container_of(mad_recv_wc,
1170 					    struct ib_mad_private_header,
1171 					    recv_wc);
1172 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1173 				    header);
1174 		kmem_cache_free(ib_mad_cache, priv);
1175 	}
1176 }
1177 EXPORT_SYMBOL(ib_free_recv_mad);
1178 
1179 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1180 					u8 rmpp_version,
1181 					ib_mad_send_handler send_handler,
1182 					ib_mad_recv_handler recv_handler,
1183 					void *context)
1184 {
1185 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1186 }
1187 EXPORT_SYMBOL(ib_redirect_mad_qp);
1188 
1189 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1190 		      struct ib_wc *wc)
1191 {
1192 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1193 	return 0;
1194 }
1195 EXPORT_SYMBOL(ib_process_mad_wc);
1196 
1197 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1198 			 struct ib_mad_reg_req *mad_reg_req)
1199 {
1200 	int i;
1201 
1202 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1203 		if ((*method)->agent[i]) {
1204 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1205 			return -EINVAL;
1206 		}
1207 	}
1208 	return 0;
1209 }
1210 
1211 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1212 {
1213 	/* Allocate management method table */
1214 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1215 	if (!*method) {
1216 		printk(KERN_ERR PFX "No memory for "
1217 		       "ib_mad_mgmt_method_table\n");
1218 		return -ENOMEM;
1219 	}
1220 
1221 	return 0;
1222 }
1223 
1224 /*
1225  * Check to see if there are any methods still in use
1226  */
1227 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1228 {
1229 	int i;
1230 
1231 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1232 		if (method->agent[i])
1233 			return 1;
1234 	return 0;
1235 }
1236 
1237 /*
1238  * Check to see if there are any method tables for this class still in use
1239  */
1240 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1241 {
1242 	int i;
1243 
1244 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1245 		if (class->method_table[i])
1246 			return 1;
1247 	return 0;
1248 }
1249 
1250 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1251 {
1252 	int i;
1253 
1254 	for (i = 0; i < MAX_MGMT_OUI; i++)
1255 		if (vendor_class->method_table[i])
1256 			return 1;
1257 	return 0;
1258 }
1259 
1260 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1261 			   char *oui)
1262 {
1263 	int i;
1264 
1265 	for (i = 0; i < MAX_MGMT_OUI; i++)
1266 		/* Is there matching OUI for this vendor class ? */
1267 		if (!memcmp(vendor_class->oui[i], oui, 3))
1268 			return i;
1269 
1270 	return -1;
1271 }
1272 
1273 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1274 {
1275 	int i;
1276 
1277 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1278 		if (vendor->vendor_class[i])
1279 			return 1;
1280 
1281 	return 0;
1282 }
1283 
1284 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1285 				     struct ib_mad_agent_private *agent)
1286 {
1287 	int i;
1288 
1289 	/* Remove any methods for this mad agent */
1290 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1291 		if (method->agent[i] == agent) {
1292 			method->agent[i] = NULL;
1293 		}
1294 	}
1295 }
1296 
1297 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1298 			      struct ib_mad_agent_private *agent_priv,
1299 			      u8 mgmt_class)
1300 {
1301 	struct ib_mad_port_private *port_priv;
1302 	struct ib_mad_mgmt_class_table **class;
1303 	struct ib_mad_mgmt_method_table **method;
1304 	int i, ret;
1305 
1306 	port_priv = agent_priv->qp_info->port_priv;
1307 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1308 	if (!*class) {
1309 		/* Allocate management class table for "new" class version */
1310 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1311 		if (!*class) {
1312 			printk(KERN_ERR PFX "No memory for "
1313 			       "ib_mad_mgmt_class_table\n");
1314 			ret = -ENOMEM;
1315 			goto error1;
1316 		}
1317 
1318 		/* Allocate method table for this management class */
1319 		method = &(*class)->method_table[mgmt_class];
1320 		if ((ret = allocate_method_table(method)))
1321 			goto error2;
1322 	} else {
1323 		method = &(*class)->method_table[mgmt_class];
1324 		if (!*method) {
1325 			/* Allocate method table for this management class */
1326 			if ((ret = allocate_method_table(method)))
1327 				goto error1;
1328 		}
1329 	}
1330 
1331 	/* Now, make sure methods are not already in use */
1332 	if (method_in_use(method, mad_reg_req))
1333 		goto error3;
1334 
1335 	/* Finally, add in methods being registered */
1336 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1337 		(*method)->agent[i] = agent_priv;
1338 
1339 	return 0;
1340 
1341 error3:
1342 	/* Remove any methods for this mad agent */
1343 	remove_methods_mad_agent(*method, agent_priv);
1344 	/* Now, check to see if there are any methods in use */
1345 	if (!check_method_table(*method)) {
1346 		/* If not, release management method table */
1347 		kfree(*method);
1348 		*method = NULL;
1349 	}
1350 	ret = -EINVAL;
1351 	goto error1;
1352 error2:
1353 	kfree(*class);
1354 	*class = NULL;
1355 error1:
1356 	return ret;
1357 }
1358 
1359 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1360 			   struct ib_mad_agent_private *agent_priv)
1361 {
1362 	struct ib_mad_port_private *port_priv;
1363 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1364 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1365 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1366 	struct ib_mad_mgmt_method_table **method;
1367 	int i, ret = -ENOMEM;
1368 	u8 vclass;
1369 
1370 	/* "New" vendor (with OUI) class */
1371 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1372 	port_priv = agent_priv->qp_info->port_priv;
1373 	vendor_table = &port_priv->version[
1374 				mad_reg_req->mgmt_class_version].vendor;
1375 	if (!*vendor_table) {
1376 		/* Allocate mgmt vendor class table for "new" class version */
1377 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1378 		if (!vendor) {
1379 			printk(KERN_ERR PFX "No memory for "
1380 			       "ib_mad_mgmt_vendor_class_table\n");
1381 			goto error1;
1382 		}
1383 
1384 		*vendor_table = vendor;
1385 	}
1386 	if (!(*vendor_table)->vendor_class[vclass]) {
1387 		/* Allocate table for this management vendor class */
1388 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1389 		if (!vendor_class) {
1390 			printk(KERN_ERR PFX "No memory for "
1391 			       "ib_mad_mgmt_vendor_class\n");
1392 			goto error2;
1393 		}
1394 
1395 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1396 	}
1397 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1398 		/* Is there matching OUI for this vendor class ? */
1399 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1400 			    mad_reg_req->oui, 3)) {
1401 			method = &(*vendor_table)->vendor_class[
1402 						vclass]->method_table[i];
1403 			BUG_ON(!*method);
1404 			goto check_in_use;
1405 		}
1406 	}
1407 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1408 		/* OUI slot available ? */
1409 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1410 				vclass]->oui[i])) {
1411 			method = &(*vendor_table)->vendor_class[
1412 				vclass]->method_table[i];
1413 			BUG_ON(*method);
1414 			/* Allocate method table for this OUI */
1415 			if ((ret = allocate_method_table(method)))
1416 				goto error3;
1417 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1418 			       mad_reg_req->oui, 3);
1419 			goto check_in_use;
1420 		}
1421 	}
1422 	printk(KERN_ERR PFX "All OUI slots in use\n");
1423 	goto error3;
1424 
1425 check_in_use:
1426 	/* Now, make sure methods are not already in use */
1427 	if (method_in_use(method, mad_reg_req))
1428 		goto error4;
1429 
1430 	/* Finally, add in methods being registered */
1431 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1432 		(*method)->agent[i] = agent_priv;
1433 
1434 	return 0;
1435 
1436 error4:
1437 	/* Remove any methods for this mad agent */
1438 	remove_methods_mad_agent(*method, agent_priv);
1439 	/* Now, check to see if there are any methods in use */
1440 	if (!check_method_table(*method)) {
1441 		/* If not, release management method table */
1442 		kfree(*method);
1443 		*method = NULL;
1444 	}
1445 	ret = -EINVAL;
1446 error3:
1447 	if (vendor_class) {
1448 		(*vendor_table)->vendor_class[vclass] = NULL;
1449 		kfree(vendor_class);
1450 	}
1451 error2:
1452 	if (vendor) {
1453 		*vendor_table = NULL;
1454 		kfree(vendor);
1455 	}
1456 error1:
1457 	return ret;
1458 }
1459 
1460 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1461 {
1462 	struct ib_mad_port_private *port_priv;
1463 	struct ib_mad_mgmt_class_table *class;
1464 	struct ib_mad_mgmt_method_table *method;
1465 	struct ib_mad_mgmt_vendor_class_table *vendor;
1466 	struct ib_mad_mgmt_vendor_class *vendor_class;
1467 	int index;
1468 	u8 mgmt_class;
1469 
1470 	/*
1471 	 * Was MAD registration request supplied
1472 	 * with original registration ?
1473 	 */
1474 	if (!agent_priv->reg_req) {
1475 		goto out;
1476 	}
1477 
1478 	port_priv = agent_priv->qp_info->port_priv;
1479 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1480 	class = port_priv->version[
1481 			agent_priv->reg_req->mgmt_class_version].class;
1482 	if (!class)
1483 		goto vendor_check;
1484 
1485 	method = class->method_table[mgmt_class];
1486 	if (method) {
1487 		/* Remove any methods for this mad agent */
1488 		remove_methods_mad_agent(method, agent_priv);
1489 		/* Now, check to see if there are any methods still in use */
1490 		if (!check_method_table(method)) {
1491 			/* If not, release management method table */
1492 			 kfree(method);
1493 			 class->method_table[mgmt_class] = NULL;
1494 			 /* Any management classes left ? */
1495 			if (!check_class_table(class)) {
1496 				/* If not, release management class table */
1497 				kfree(class);
1498 				port_priv->version[
1499 					agent_priv->reg_req->
1500 					mgmt_class_version].class = NULL;
1501 			}
1502 		}
1503 	}
1504 
1505 vendor_check:
1506 	if (!is_vendor_class(mgmt_class))
1507 		goto out;
1508 
1509 	/* normalize mgmt_class to vendor range 2 */
1510 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1511 	vendor = port_priv->version[
1512 			agent_priv->reg_req->mgmt_class_version].vendor;
1513 
1514 	if (!vendor)
1515 		goto out;
1516 
1517 	vendor_class = vendor->vendor_class[mgmt_class];
1518 	if (vendor_class) {
1519 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1520 		if (index < 0)
1521 			goto out;
1522 		method = vendor_class->method_table[index];
1523 		if (method) {
1524 			/* Remove any methods for this mad agent */
1525 			remove_methods_mad_agent(method, agent_priv);
1526 			/*
1527 			 * Now, check to see if there are
1528 			 * any methods still in use
1529 			 */
1530 			if (!check_method_table(method)) {
1531 				/* If not, release management method table */
1532 				kfree(method);
1533 				vendor_class->method_table[index] = NULL;
1534 				memset(vendor_class->oui[index], 0, 3);
1535 				/* Any OUIs left ? */
1536 				if (!check_vendor_class(vendor_class)) {
1537 					/* If not, release vendor class table */
1538 					kfree(vendor_class);
1539 					vendor->vendor_class[mgmt_class] = NULL;
1540 					/* Any other vendor classes left ? */
1541 					if (!check_vendor_table(vendor)) {
1542 						kfree(vendor);
1543 						port_priv->version[
1544 							agent_priv->reg_req->
1545 							mgmt_class_version].
1546 							vendor = NULL;
1547 					}
1548 				}
1549 			}
1550 		}
1551 	}
1552 
1553 out:
1554 	return;
1555 }
1556 
1557 static struct ib_mad_agent_private *
1558 find_mad_agent(struct ib_mad_port_private *port_priv,
1559 	       struct ib_mad *mad)
1560 {
1561 	struct ib_mad_agent_private *mad_agent = NULL;
1562 	unsigned long flags;
1563 
1564 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1565 	if (ib_response_mad(mad)) {
1566 		u32 hi_tid;
1567 		struct ib_mad_agent_private *entry;
1568 
1569 		/*
1570 		 * Routing is based on high 32 bits of transaction ID
1571 		 * of MAD.
1572 		 */
1573 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1574 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1575 			if (entry->agent.hi_tid == hi_tid) {
1576 				mad_agent = entry;
1577 				break;
1578 			}
1579 		}
1580 	} else {
1581 		struct ib_mad_mgmt_class_table *class;
1582 		struct ib_mad_mgmt_method_table *method;
1583 		struct ib_mad_mgmt_vendor_class_table *vendor;
1584 		struct ib_mad_mgmt_vendor_class *vendor_class;
1585 		struct ib_vendor_mad *vendor_mad;
1586 		int index;
1587 
1588 		/*
1589 		 * Routing is based on version, class, and method
1590 		 * For "newer" vendor MADs, also based on OUI
1591 		 */
1592 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1593 			goto out;
1594 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1595 			class = port_priv->version[
1596 					mad->mad_hdr.class_version].class;
1597 			if (!class)
1598 				goto out;
1599 			method = class->method_table[convert_mgmt_class(
1600 							mad->mad_hdr.mgmt_class)];
1601 			if (method)
1602 				mad_agent = method->agent[mad->mad_hdr.method &
1603 							  ~IB_MGMT_METHOD_RESP];
1604 		} else {
1605 			vendor = port_priv->version[
1606 					mad->mad_hdr.class_version].vendor;
1607 			if (!vendor)
1608 				goto out;
1609 			vendor_class = vendor->vendor_class[vendor_class_index(
1610 						mad->mad_hdr.mgmt_class)];
1611 			if (!vendor_class)
1612 				goto out;
1613 			/* Find matching OUI */
1614 			vendor_mad = (struct ib_vendor_mad *)mad;
1615 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1616 			if (index == -1)
1617 				goto out;
1618 			method = vendor_class->method_table[index];
1619 			if (method) {
1620 				mad_agent = method->agent[mad->mad_hdr.method &
1621 							  ~IB_MGMT_METHOD_RESP];
1622 			}
1623 		}
1624 	}
1625 
1626 	if (mad_agent) {
1627 		if (mad_agent->agent.recv_handler)
1628 			atomic_inc(&mad_agent->refcount);
1629 		else {
1630 			printk(KERN_NOTICE PFX "No receive handler for client "
1631 			       "%p on port %d\n",
1632 			       &mad_agent->agent, port_priv->port_num);
1633 			mad_agent = NULL;
1634 		}
1635 	}
1636 out:
1637 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1638 
1639 	return mad_agent;
1640 }
1641 
1642 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1643 {
1644 	int valid = 0;
1645 
1646 	/* Make sure MAD base version is understood */
1647 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1648 		printk(KERN_ERR PFX "MAD received with unsupported base "
1649 		       "version %d\n", mad->mad_hdr.base_version);
1650 		goto out;
1651 	}
1652 
1653 	/* Filter SMI packets sent to other than QP0 */
1654 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1655 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1656 		if (qp_num == 0)
1657 			valid = 1;
1658 	} else {
1659 		/* Filter GSI packets sent to QP0 */
1660 		if (qp_num != 0)
1661 			valid = 1;
1662 	}
1663 
1664 out:
1665 	return valid;
1666 }
1667 
1668 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1669 		       struct ib_mad_hdr *mad_hdr)
1670 {
1671 	struct ib_rmpp_mad *rmpp_mad;
1672 
1673 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1674 	return !mad_agent_priv->agent.rmpp_version ||
1675 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1676 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1677 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1678 }
1679 
1680 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1681 				     struct ib_mad_recv_wc *rwc)
1682 {
1683 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1684 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1685 }
1686 
1687 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1688 				   struct ib_mad_send_wr_private *wr,
1689 				   struct ib_mad_recv_wc *rwc )
1690 {
1691 	struct ib_ah_attr attr;
1692 	u8 send_resp, rcv_resp;
1693 	union ib_gid sgid;
1694 	struct ib_device *device = mad_agent_priv->agent.device;
1695 	u8 port_num = mad_agent_priv->agent.port_num;
1696 	u8 lmc;
1697 
1698 	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1699 	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1700 
1701 	if (send_resp == rcv_resp)
1702 		/* both requests, or both responses. GIDs different */
1703 		return 0;
1704 
1705 	if (ib_query_ah(wr->send_buf.ah, &attr))
1706 		/* Assume not equal, to avoid false positives. */
1707 		return 0;
1708 
1709 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1710 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1711 		/* one has GID, other does not.  Assume different */
1712 		return 0;
1713 
1714 	if (!send_resp && rcv_resp) {
1715 		/* is request/response. */
1716 		if (!(attr.ah_flags & IB_AH_GRH)) {
1717 			if (ib_get_cached_lmc(device, port_num, &lmc))
1718 				return 0;
1719 			return (!lmc || !((attr.src_path_bits ^
1720 					   rwc->wc->dlid_path_bits) &
1721 					  ((1 << lmc) - 1)));
1722 		} else {
1723 			if (ib_get_cached_gid(device, port_num,
1724 					      attr.grh.sgid_index, &sgid))
1725 				return 0;
1726 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1727 				       16);
1728 		}
1729 	}
1730 
1731 	if (!(attr.ah_flags & IB_AH_GRH))
1732 		return attr.dlid == rwc->wc->slid;
1733 	else
1734 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1735 			       16);
1736 }
1737 
1738 static inline int is_direct(u8 class)
1739 {
1740 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1741 }
1742 
1743 struct ib_mad_send_wr_private*
1744 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1745 		 struct ib_mad_recv_wc *wc)
1746 {
1747 	struct ib_mad_send_wr_private *wr;
1748 	struct ib_mad *mad;
1749 
1750 	mad = (struct ib_mad *)wc->recv_buf.mad;
1751 
1752 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1753 		if ((wr->tid == mad->mad_hdr.tid) &&
1754 		    rcv_has_same_class(wr, wc) &&
1755 		    /*
1756 		     * Don't check GID for direct routed MADs.
1757 		     * These might have permissive LIDs.
1758 		     */
1759 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1760 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1761 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1762 	}
1763 
1764 	/*
1765 	 * It's possible to receive the response before we've
1766 	 * been notified that the send has completed
1767 	 */
1768 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1769 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1770 		    wr->tid == mad->mad_hdr.tid &&
1771 		    wr->timeout &&
1772 		    rcv_has_same_class(wr, wc) &&
1773 		    /*
1774 		     * Don't check GID for direct routed MADs.
1775 		     * These might have permissive LIDs.
1776 		     */
1777 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1778 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1779 			/* Verify request has not been canceled */
1780 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1781 	}
1782 	return NULL;
1783 }
1784 
1785 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1786 {
1787 	mad_send_wr->timeout = 0;
1788 	if (mad_send_wr->refcount == 1)
1789 		list_move_tail(&mad_send_wr->agent_list,
1790 			      &mad_send_wr->mad_agent_priv->done_list);
1791 }
1792 
1793 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1794 				 struct ib_mad_recv_wc *mad_recv_wc)
1795 {
1796 	struct ib_mad_send_wr_private *mad_send_wr;
1797 	struct ib_mad_send_wc mad_send_wc;
1798 	unsigned long flags;
1799 
1800 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1801 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1802 	if (mad_agent_priv->agent.rmpp_version) {
1803 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1804 						      mad_recv_wc);
1805 		if (!mad_recv_wc) {
1806 			deref_mad_agent(mad_agent_priv);
1807 			return;
1808 		}
1809 	}
1810 
1811 	/* Complete corresponding request */
1812 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1813 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1814 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1815 		if (!mad_send_wr) {
1816 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1817 			ib_free_recv_mad(mad_recv_wc);
1818 			deref_mad_agent(mad_agent_priv);
1819 			return;
1820 		}
1821 		ib_mark_mad_done(mad_send_wr);
1822 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1823 
1824 		/* Defined behavior is to complete response before request */
1825 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1826 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1827 						   mad_recv_wc);
1828 		atomic_dec(&mad_agent_priv->refcount);
1829 
1830 		mad_send_wc.status = IB_WC_SUCCESS;
1831 		mad_send_wc.vendor_err = 0;
1832 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1833 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1834 	} else {
1835 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1836 						   mad_recv_wc);
1837 		deref_mad_agent(mad_agent_priv);
1838 	}
1839 }
1840 
1841 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1842 				     struct ib_wc *wc)
1843 {
1844 	struct ib_mad_qp_info *qp_info;
1845 	struct ib_mad_private_header *mad_priv_hdr;
1846 	struct ib_mad_private *recv, *response = NULL;
1847 	struct ib_mad_list_head *mad_list;
1848 	struct ib_mad_agent_private *mad_agent;
1849 	int port_num;
1850 
1851 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1852 	qp_info = mad_list->mad_queue->qp_info;
1853 	dequeue_mad(mad_list);
1854 
1855 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1856 				    mad_list);
1857 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1858 	ib_dma_unmap_single(port_priv->device,
1859 			    recv->header.mapping,
1860 			    sizeof(struct ib_mad_private) -
1861 			      sizeof(struct ib_mad_private_header),
1862 			    DMA_FROM_DEVICE);
1863 
1864 	/* Setup MAD receive work completion from "normal" work completion */
1865 	recv->header.wc = *wc;
1866 	recv->header.recv_wc.wc = &recv->header.wc;
1867 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1868 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1869 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1870 
1871 	if (atomic_read(&qp_info->snoop_count))
1872 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1873 
1874 	/* Validate MAD */
1875 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1876 		goto out;
1877 
1878 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1879 	if (!response) {
1880 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1881 		       "for response buffer\n");
1882 		goto out;
1883 	}
1884 
1885 	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1886 		port_num = wc->port_num;
1887 	else
1888 		port_num = port_priv->port_num;
1889 
1890 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1891 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1892 		enum smi_forward_action retsmi;
1893 
1894 		if (smi_handle_dr_smp_recv(&recv->mad.smp,
1895 					   port_priv->device->node_type,
1896 					   port_num,
1897 					   port_priv->device->phys_port_cnt) ==
1898 					   IB_SMI_DISCARD)
1899 			goto out;
1900 
1901 		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1902 		if (retsmi == IB_SMI_LOCAL)
1903 			goto local;
1904 
1905 		if (retsmi == IB_SMI_SEND) { /* don't forward */
1906 			if (smi_handle_dr_smp_send(&recv->mad.smp,
1907 						   port_priv->device->node_type,
1908 						   port_num) == IB_SMI_DISCARD)
1909 				goto out;
1910 
1911 			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1912 				goto out;
1913 		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1914 			/* forward case for switches */
1915 			memcpy(response, recv, sizeof(*response));
1916 			response->header.recv_wc.wc = &response->header.wc;
1917 			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1918 			response->header.recv_wc.recv_buf.grh = &response->grh;
1919 
1920 			agent_send_response(&response->mad.mad,
1921 					    &response->grh, wc,
1922 					    port_priv->device,
1923 					    smi_get_fwd_port(&recv->mad.smp),
1924 					    qp_info->qp->qp_num);
1925 
1926 			goto out;
1927 		}
1928 	}
1929 
1930 local:
1931 	/* Give driver "right of first refusal" on incoming MAD */
1932 	if (port_priv->device->process_mad) {
1933 		int ret;
1934 
1935 		ret = port_priv->device->process_mad(port_priv->device, 0,
1936 						     port_priv->port_num,
1937 						     wc, &recv->grh,
1938 						     &recv->mad.mad,
1939 						     &response->mad.mad);
1940 		if (ret & IB_MAD_RESULT_SUCCESS) {
1941 			if (ret & IB_MAD_RESULT_CONSUMED)
1942 				goto out;
1943 			if (ret & IB_MAD_RESULT_REPLY) {
1944 				agent_send_response(&response->mad.mad,
1945 						    &recv->grh, wc,
1946 						    port_priv->device,
1947 						    port_num,
1948 						    qp_info->qp->qp_num);
1949 				goto out;
1950 			}
1951 		}
1952 	}
1953 
1954 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1955 	if (mad_agent) {
1956 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1957 		/*
1958 		 * recv is freed up in error cases in ib_mad_complete_recv
1959 		 * or via recv_handler in ib_mad_complete_recv()
1960 		 */
1961 		recv = NULL;
1962 	}
1963 
1964 out:
1965 	/* Post another receive request for this QP */
1966 	if (response) {
1967 		ib_mad_post_receive_mads(qp_info, response);
1968 		if (recv)
1969 			kmem_cache_free(ib_mad_cache, recv);
1970 	} else
1971 		ib_mad_post_receive_mads(qp_info, recv);
1972 }
1973 
1974 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1975 {
1976 	struct ib_mad_send_wr_private *mad_send_wr;
1977 	unsigned long delay;
1978 
1979 	if (list_empty(&mad_agent_priv->wait_list)) {
1980 		__cancel_delayed_work(&mad_agent_priv->timed_work);
1981 	} else {
1982 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1983 					 struct ib_mad_send_wr_private,
1984 					 agent_list);
1985 
1986 		if (time_after(mad_agent_priv->timeout,
1987 			       mad_send_wr->timeout)) {
1988 			mad_agent_priv->timeout = mad_send_wr->timeout;
1989 			__cancel_delayed_work(&mad_agent_priv->timed_work);
1990 			delay = mad_send_wr->timeout - jiffies;
1991 			if ((long)delay <= 0)
1992 				delay = 1;
1993 			queue_delayed_work(mad_agent_priv->qp_info->
1994 					   port_priv->wq,
1995 					   &mad_agent_priv->timed_work, delay);
1996 		}
1997 	}
1998 }
1999 
2000 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2001 {
2002 	struct ib_mad_agent_private *mad_agent_priv;
2003 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2004 	struct list_head *list_item;
2005 	unsigned long delay;
2006 
2007 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2008 	list_del(&mad_send_wr->agent_list);
2009 
2010 	delay = mad_send_wr->timeout;
2011 	mad_send_wr->timeout += jiffies;
2012 
2013 	if (delay) {
2014 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2015 			temp_mad_send_wr = list_entry(list_item,
2016 						struct ib_mad_send_wr_private,
2017 						agent_list);
2018 			if (time_after(mad_send_wr->timeout,
2019 				       temp_mad_send_wr->timeout))
2020 				break;
2021 		}
2022 	}
2023 	else
2024 		list_item = &mad_agent_priv->wait_list;
2025 	list_add(&mad_send_wr->agent_list, list_item);
2026 
2027 	/* Reschedule a work item if we have a shorter timeout */
2028 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2029 		__cancel_delayed_work(&mad_agent_priv->timed_work);
2030 		queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2031 				   &mad_agent_priv->timed_work, delay);
2032 	}
2033 }
2034 
2035 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2036 			  int timeout_ms)
2037 {
2038 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2039 	wait_for_response(mad_send_wr);
2040 }
2041 
2042 /*
2043  * Process a send work completion
2044  */
2045 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2046 			     struct ib_mad_send_wc *mad_send_wc)
2047 {
2048 	struct ib_mad_agent_private	*mad_agent_priv;
2049 	unsigned long			flags;
2050 	int				ret;
2051 
2052 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2053 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2054 	if (mad_agent_priv->agent.rmpp_version) {
2055 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2056 		if (ret == IB_RMPP_RESULT_CONSUMED)
2057 			goto done;
2058 	} else
2059 		ret = IB_RMPP_RESULT_UNHANDLED;
2060 
2061 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2062 	    mad_send_wr->status == IB_WC_SUCCESS) {
2063 		mad_send_wr->status = mad_send_wc->status;
2064 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2065 	}
2066 
2067 	if (--mad_send_wr->refcount > 0) {
2068 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2069 		    mad_send_wr->status == IB_WC_SUCCESS) {
2070 			wait_for_response(mad_send_wr);
2071 		}
2072 		goto done;
2073 	}
2074 
2075 	/* Remove send from MAD agent and notify client of completion */
2076 	list_del(&mad_send_wr->agent_list);
2077 	adjust_timeout(mad_agent_priv);
2078 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2079 
2080 	if (mad_send_wr->status != IB_WC_SUCCESS )
2081 		mad_send_wc->status = mad_send_wr->status;
2082 	if (ret == IB_RMPP_RESULT_INTERNAL)
2083 		ib_rmpp_send_handler(mad_send_wc);
2084 	else
2085 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2086 						   mad_send_wc);
2087 
2088 	/* Release reference on agent taken when sending */
2089 	deref_mad_agent(mad_agent_priv);
2090 	return;
2091 done:
2092 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2093 }
2094 
2095 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2096 				     struct ib_wc *wc)
2097 {
2098 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2099 	struct ib_mad_list_head		*mad_list;
2100 	struct ib_mad_qp_info		*qp_info;
2101 	struct ib_mad_queue		*send_queue;
2102 	struct ib_send_wr		*bad_send_wr;
2103 	struct ib_mad_send_wc		mad_send_wc;
2104 	unsigned long flags;
2105 	int ret;
2106 
2107 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2108 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2109 				   mad_list);
2110 	send_queue = mad_list->mad_queue;
2111 	qp_info = send_queue->qp_info;
2112 
2113 retry:
2114 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2115 			    mad_send_wr->header_mapping,
2116 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2117 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2118 			    mad_send_wr->payload_mapping,
2119 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2120 	queued_send_wr = NULL;
2121 	spin_lock_irqsave(&send_queue->lock, flags);
2122 	list_del(&mad_list->list);
2123 
2124 	/* Move queued send to the send queue */
2125 	if (send_queue->count-- > send_queue->max_active) {
2126 		mad_list = container_of(qp_info->overflow_list.next,
2127 					struct ib_mad_list_head, list);
2128 		queued_send_wr = container_of(mad_list,
2129 					struct ib_mad_send_wr_private,
2130 					mad_list);
2131 		list_move_tail(&mad_list->list, &send_queue->list);
2132 	}
2133 	spin_unlock_irqrestore(&send_queue->lock, flags);
2134 
2135 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2136 	mad_send_wc.status = wc->status;
2137 	mad_send_wc.vendor_err = wc->vendor_err;
2138 	if (atomic_read(&qp_info->snoop_count))
2139 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2140 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2141 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2142 
2143 	if (queued_send_wr) {
2144 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2145 				   &bad_send_wr);
2146 		if (ret) {
2147 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2148 			mad_send_wr = queued_send_wr;
2149 			wc->status = IB_WC_LOC_QP_OP_ERR;
2150 			goto retry;
2151 		}
2152 	}
2153 }
2154 
2155 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2156 {
2157 	struct ib_mad_send_wr_private *mad_send_wr;
2158 	struct ib_mad_list_head *mad_list;
2159 	unsigned long flags;
2160 
2161 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2162 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2163 		mad_send_wr = container_of(mad_list,
2164 					   struct ib_mad_send_wr_private,
2165 					   mad_list);
2166 		mad_send_wr->retry = 1;
2167 	}
2168 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2169 }
2170 
2171 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2172 			      struct ib_wc *wc)
2173 {
2174 	struct ib_mad_list_head *mad_list;
2175 	struct ib_mad_qp_info *qp_info;
2176 	struct ib_mad_send_wr_private *mad_send_wr;
2177 	int ret;
2178 
2179 	/* Determine if failure was a send or receive */
2180 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2181 	qp_info = mad_list->mad_queue->qp_info;
2182 	if (mad_list->mad_queue == &qp_info->recv_queue)
2183 		/*
2184 		 * Receive errors indicate that the QP has entered the error
2185 		 * state - error handling/shutdown code will cleanup
2186 		 */
2187 		return;
2188 
2189 	/*
2190 	 * Send errors will transition the QP to SQE - move
2191 	 * QP to RTS and repost flushed work requests
2192 	 */
2193 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2194 				   mad_list);
2195 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2196 		if (mad_send_wr->retry) {
2197 			/* Repost send */
2198 			struct ib_send_wr *bad_send_wr;
2199 
2200 			mad_send_wr->retry = 0;
2201 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2202 					&bad_send_wr);
2203 			if (ret)
2204 				ib_mad_send_done_handler(port_priv, wc);
2205 		} else
2206 			ib_mad_send_done_handler(port_priv, wc);
2207 	} else {
2208 		struct ib_qp_attr *attr;
2209 
2210 		/* Transition QP to RTS and fail offending send */
2211 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2212 		if (attr) {
2213 			attr->qp_state = IB_QPS_RTS;
2214 			attr->cur_qp_state = IB_QPS_SQE;
2215 			ret = ib_modify_qp(qp_info->qp, attr,
2216 					   IB_QP_STATE | IB_QP_CUR_STATE);
2217 			kfree(attr);
2218 			if (ret)
2219 				printk(KERN_ERR PFX "mad_error_handler - "
2220 				       "ib_modify_qp to RTS : %d\n", ret);
2221 			else
2222 				mark_sends_for_retry(qp_info);
2223 		}
2224 		ib_mad_send_done_handler(port_priv, wc);
2225 	}
2226 }
2227 
2228 /*
2229  * IB MAD completion callback
2230  */
2231 static void ib_mad_completion_handler(struct work_struct *work)
2232 {
2233 	struct ib_mad_port_private *port_priv;
2234 	struct ib_wc wc;
2235 
2236 	port_priv = container_of(work, struct ib_mad_port_private, work);
2237 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2238 
2239 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2240 		if (wc.status == IB_WC_SUCCESS) {
2241 			switch (wc.opcode) {
2242 			case IB_WC_SEND:
2243 				ib_mad_send_done_handler(port_priv, &wc);
2244 				break;
2245 			case IB_WC_RECV:
2246 				ib_mad_recv_done_handler(port_priv, &wc);
2247 				break;
2248 			default:
2249 				BUG_ON(1);
2250 				break;
2251 			}
2252 		} else
2253 			mad_error_handler(port_priv, &wc);
2254 	}
2255 }
2256 
2257 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2258 {
2259 	unsigned long flags;
2260 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2261 	struct ib_mad_send_wc mad_send_wc;
2262 	struct list_head cancel_list;
2263 
2264 	INIT_LIST_HEAD(&cancel_list);
2265 
2266 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2267 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2268 				 &mad_agent_priv->send_list, agent_list) {
2269 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2270 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2271 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2272 		}
2273 	}
2274 
2275 	/* Empty wait list to prevent receives from finding a request */
2276 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2277 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2278 
2279 	/* Report all cancelled requests */
2280 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2281 	mad_send_wc.vendor_err = 0;
2282 
2283 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2284 				 &cancel_list, agent_list) {
2285 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2286 		list_del(&mad_send_wr->agent_list);
2287 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2288 						   &mad_send_wc);
2289 		atomic_dec(&mad_agent_priv->refcount);
2290 	}
2291 }
2292 
2293 static struct ib_mad_send_wr_private*
2294 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2295 	     struct ib_mad_send_buf *send_buf)
2296 {
2297 	struct ib_mad_send_wr_private *mad_send_wr;
2298 
2299 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2300 			    agent_list) {
2301 		if (&mad_send_wr->send_buf == send_buf)
2302 			return mad_send_wr;
2303 	}
2304 
2305 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2306 			    agent_list) {
2307 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2308 		    &mad_send_wr->send_buf == send_buf)
2309 			return mad_send_wr;
2310 	}
2311 	return NULL;
2312 }
2313 
2314 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2315 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2316 {
2317 	struct ib_mad_agent_private *mad_agent_priv;
2318 	struct ib_mad_send_wr_private *mad_send_wr;
2319 	unsigned long flags;
2320 	int active;
2321 
2322 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2323 				      agent);
2324 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2325 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2326 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2327 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2328 		return -EINVAL;
2329 	}
2330 
2331 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2332 	if (!timeout_ms) {
2333 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2334 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2335 	}
2336 
2337 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2338 	if (active)
2339 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2340 	else
2341 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2342 
2343 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2344 	return 0;
2345 }
2346 EXPORT_SYMBOL(ib_modify_mad);
2347 
2348 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2349 		   struct ib_mad_send_buf *send_buf)
2350 {
2351 	ib_modify_mad(mad_agent, send_buf, 0);
2352 }
2353 EXPORT_SYMBOL(ib_cancel_mad);
2354 
2355 static void local_completions(struct work_struct *work)
2356 {
2357 	struct ib_mad_agent_private *mad_agent_priv;
2358 	struct ib_mad_local_private *local;
2359 	struct ib_mad_agent_private *recv_mad_agent;
2360 	unsigned long flags;
2361 	int free_mad;
2362 	struct ib_wc wc;
2363 	struct ib_mad_send_wc mad_send_wc;
2364 
2365 	mad_agent_priv =
2366 		container_of(work, struct ib_mad_agent_private, local_work);
2367 
2368 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2369 	while (!list_empty(&mad_agent_priv->local_list)) {
2370 		local = list_entry(mad_agent_priv->local_list.next,
2371 				   struct ib_mad_local_private,
2372 				   completion_list);
2373 		list_del(&local->completion_list);
2374 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2375 		free_mad = 0;
2376 		if (local->mad_priv) {
2377 			recv_mad_agent = local->recv_mad_agent;
2378 			if (!recv_mad_agent) {
2379 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2380 				free_mad = 1;
2381 				goto local_send_completion;
2382 			}
2383 
2384 			/*
2385 			 * Defined behavior is to complete response
2386 			 * before request
2387 			 */
2388 			build_smp_wc(recv_mad_agent->agent.qp,
2389 				     (unsigned long) local->mad_send_wr,
2390 				     be16_to_cpu(IB_LID_PERMISSIVE),
2391 				     0, recv_mad_agent->agent.port_num, &wc);
2392 
2393 			local->mad_priv->header.recv_wc.wc = &wc;
2394 			local->mad_priv->header.recv_wc.mad_len =
2395 						sizeof(struct ib_mad);
2396 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2397 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2398 				 &local->mad_priv->header.recv_wc.rmpp_list);
2399 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2400 			local->mad_priv->header.recv_wc.recv_buf.mad =
2401 						&local->mad_priv->mad.mad;
2402 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2403 				snoop_recv(recv_mad_agent->qp_info,
2404 					  &local->mad_priv->header.recv_wc,
2405 					   IB_MAD_SNOOP_RECVS);
2406 			recv_mad_agent->agent.recv_handler(
2407 						&recv_mad_agent->agent,
2408 						&local->mad_priv->header.recv_wc);
2409 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2410 			atomic_dec(&recv_mad_agent->refcount);
2411 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2412 		}
2413 
2414 local_send_completion:
2415 		/* Complete send */
2416 		mad_send_wc.status = IB_WC_SUCCESS;
2417 		mad_send_wc.vendor_err = 0;
2418 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2419 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2420 			snoop_send(mad_agent_priv->qp_info,
2421 				   &local->mad_send_wr->send_buf,
2422 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2423 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2424 						   &mad_send_wc);
2425 
2426 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2427 		atomic_dec(&mad_agent_priv->refcount);
2428 		if (free_mad)
2429 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2430 		kfree(local);
2431 	}
2432 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2433 }
2434 
2435 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2436 {
2437 	int ret;
2438 
2439 	if (!mad_send_wr->retries_left)
2440 		return -ETIMEDOUT;
2441 
2442 	mad_send_wr->retries_left--;
2443 	mad_send_wr->send_buf.retries++;
2444 
2445 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2446 
2447 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2448 		ret = ib_retry_rmpp(mad_send_wr);
2449 		switch (ret) {
2450 		case IB_RMPP_RESULT_UNHANDLED:
2451 			ret = ib_send_mad(mad_send_wr);
2452 			break;
2453 		case IB_RMPP_RESULT_CONSUMED:
2454 			ret = 0;
2455 			break;
2456 		default:
2457 			ret = -ECOMM;
2458 			break;
2459 		}
2460 	} else
2461 		ret = ib_send_mad(mad_send_wr);
2462 
2463 	if (!ret) {
2464 		mad_send_wr->refcount++;
2465 		list_add_tail(&mad_send_wr->agent_list,
2466 			      &mad_send_wr->mad_agent_priv->send_list);
2467 	}
2468 	return ret;
2469 }
2470 
2471 static void timeout_sends(struct work_struct *work)
2472 {
2473 	struct ib_mad_agent_private *mad_agent_priv;
2474 	struct ib_mad_send_wr_private *mad_send_wr;
2475 	struct ib_mad_send_wc mad_send_wc;
2476 	unsigned long flags, delay;
2477 
2478 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2479 				      timed_work.work);
2480 	mad_send_wc.vendor_err = 0;
2481 
2482 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2483 	while (!list_empty(&mad_agent_priv->wait_list)) {
2484 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2485 					 struct ib_mad_send_wr_private,
2486 					 agent_list);
2487 
2488 		if (time_after(mad_send_wr->timeout, jiffies)) {
2489 			delay = mad_send_wr->timeout - jiffies;
2490 			if ((long)delay <= 0)
2491 				delay = 1;
2492 			queue_delayed_work(mad_agent_priv->qp_info->
2493 					   port_priv->wq,
2494 					   &mad_agent_priv->timed_work, delay);
2495 			break;
2496 		}
2497 
2498 		list_del(&mad_send_wr->agent_list);
2499 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2500 		    !retry_send(mad_send_wr))
2501 			continue;
2502 
2503 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2504 
2505 		if (mad_send_wr->status == IB_WC_SUCCESS)
2506 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2507 		else
2508 			mad_send_wc.status = mad_send_wr->status;
2509 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2510 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2511 						   &mad_send_wc);
2512 
2513 		atomic_dec(&mad_agent_priv->refcount);
2514 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2515 	}
2516 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2517 }
2518 
2519 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2520 {
2521 	struct ib_mad_port_private *port_priv = cq->cq_context;
2522 	unsigned long flags;
2523 
2524 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2525 	if (!list_empty(&port_priv->port_list))
2526 		queue_work(port_priv->wq, &port_priv->work);
2527 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2528 }
2529 
2530 /*
2531  * Allocate receive MADs and post receive WRs for them
2532  */
2533 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2534 				    struct ib_mad_private *mad)
2535 {
2536 	unsigned long flags;
2537 	int post, ret;
2538 	struct ib_mad_private *mad_priv;
2539 	struct ib_sge sg_list;
2540 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2541 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2542 
2543 	/* Initialize common scatter list fields */
2544 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2545 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2546 
2547 	/* Initialize common receive WR fields */
2548 	recv_wr.next = NULL;
2549 	recv_wr.sg_list = &sg_list;
2550 	recv_wr.num_sge = 1;
2551 
2552 	do {
2553 		/* Allocate and map receive buffer */
2554 		if (mad) {
2555 			mad_priv = mad;
2556 			mad = NULL;
2557 		} else {
2558 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2559 			if (!mad_priv) {
2560 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2561 				ret = -ENOMEM;
2562 				break;
2563 			}
2564 		}
2565 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2566 						 &mad_priv->grh,
2567 						 sizeof *mad_priv -
2568 						   sizeof mad_priv->header,
2569 						 DMA_FROM_DEVICE);
2570 		mad_priv->header.mapping = sg_list.addr;
2571 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2572 		mad_priv->header.mad_list.mad_queue = recv_queue;
2573 
2574 		/* Post receive WR */
2575 		spin_lock_irqsave(&recv_queue->lock, flags);
2576 		post = (++recv_queue->count < recv_queue->max_active);
2577 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2578 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2579 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2580 		if (ret) {
2581 			spin_lock_irqsave(&recv_queue->lock, flags);
2582 			list_del(&mad_priv->header.mad_list.list);
2583 			recv_queue->count--;
2584 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2585 			ib_dma_unmap_single(qp_info->port_priv->device,
2586 					    mad_priv->header.mapping,
2587 					    sizeof *mad_priv -
2588 					      sizeof mad_priv->header,
2589 					    DMA_FROM_DEVICE);
2590 			kmem_cache_free(ib_mad_cache, mad_priv);
2591 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2592 			break;
2593 		}
2594 	} while (post);
2595 
2596 	return ret;
2597 }
2598 
2599 /*
2600  * Return all the posted receive MADs
2601  */
2602 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2603 {
2604 	struct ib_mad_private_header *mad_priv_hdr;
2605 	struct ib_mad_private *recv;
2606 	struct ib_mad_list_head *mad_list;
2607 
2608 	if (!qp_info->qp)
2609 		return;
2610 
2611 	while (!list_empty(&qp_info->recv_queue.list)) {
2612 
2613 		mad_list = list_entry(qp_info->recv_queue.list.next,
2614 				      struct ib_mad_list_head, list);
2615 		mad_priv_hdr = container_of(mad_list,
2616 					    struct ib_mad_private_header,
2617 					    mad_list);
2618 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2619 				    header);
2620 
2621 		/* Remove from posted receive MAD list */
2622 		list_del(&mad_list->list);
2623 
2624 		ib_dma_unmap_single(qp_info->port_priv->device,
2625 				    recv->header.mapping,
2626 				    sizeof(struct ib_mad_private) -
2627 				      sizeof(struct ib_mad_private_header),
2628 				    DMA_FROM_DEVICE);
2629 		kmem_cache_free(ib_mad_cache, recv);
2630 	}
2631 
2632 	qp_info->recv_queue.count = 0;
2633 }
2634 
2635 /*
2636  * Start the port
2637  */
2638 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2639 {
2640 	int ret, i;
2641 	struct ib_qp_attr *attr;
2642 	struct ib_qp *qp;
2643 
2644 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2645 	if (!attr) {
2646 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2647 		return -ENOMEM;
2648 	}
2649 
2650 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2651 		qp = port_priv->qp_info[i].qp;
2652 		if (!qp)
2653 			continue;
2654 
2655 		/*
2656 		 * PKey index for QP1 is irrelevant but
2657 		 * one is needed for the Reset to Init transition
2658 		 */
2659 		attr->qp_state = IB_QPS_INIT;
2660 		attr->pkey_index = 0;
2661 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2662 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2663 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2664 		if (ret) {
2665 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2666 			       "INIT: %d\n", i, ret);
2667 			goto out;
2668 		}
2669 
2670 		attr->qp_state = IB_QPS_RTR;
2671 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2672 		if (ret) {
2673 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2674 			       "RTR: %d\n", i, ret);
2675 			goto out;
2676 		}
2677 
2678 		attr->qp_state = IB_QPS_RTS;
2679 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2680 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2681 		if (ret) {
2682 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2683 			       "RTS: %d\n", i, ret);
2684 			goto out;
2685 		}
2686 	}
2687 
2688 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2689 	if (ret) {
2690 		printk(KERN_ERR PFX "Failed to request completion "
2691 		       "notification: %d\n", ret);
2692 		goto out;
2693 	}
2694 
2695 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2696 		if (!port_priv->qp_info[i].qp)
2697 			continue;
2698 
2699 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2700 		if (ret) {
2701 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2702 			goto out;
2703 		}
2704 	}
2705 out:
2706 	kfree(attr);
2707 	return ret;
2708 }
2709 
2710 static void qp_event_handler(struct ib_event *event, void *qp_context)
2711 {
2712 	struct ib_mad_qp_info	*qp_info = qp_context;
2713 
2714 	/* It's worse than that! He's dead, Jim! */
2715 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2716 		event->event, qp_info->qp->qp_num);
2717 }
2718 
2719 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2720 			   struct ib_mad_queue *mad_queue)
2721 {
2722 	mad_queue->qp_info = qp_info;
2723 	mad_queue->count = 0;
2724 	spin_lock_init(&mad_queue->lock);
2725 	INIT_LIST_HEAD(&mad_queue->list);
2726 }
2727 
2728 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2729 			struct ib_mad_qp_info *qp_info)
2730 {
2731 	qp_info->port_priv = port_priv;
2732 	init_mad_queue(qp_info, &qp_info->send_queue);
2733 	init_mad_queue(qp_info, &qp_info->recv_queue);
2734 	INIT_LIST_HEAD(&qp_info->overflow_list);
2735 	spin_lock_init(&qp_info->snoop_lock);
2736 	qp_info->snoop_table = NULL;
2737 	qp_info->snoop_table_size = 0;
2738 	atomic_set(&qp_info->snoop_count, 0);
2739 }
2740 
2741 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2742 			 enum ib_qp_type qp_type)
2743 {
2744 	struct ib_qp_init_attr	qp_init_attr;
2745 	int ret;
2746 
2747 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2748 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2749 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2750 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2751 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
2752 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2753 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2754 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2755 	qp_init_attr.qp_type = qp_type;
2756 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2757 	qp_init_attr.qp_context = qp_info;
2758 	qp_init_attr.event_handler = qp_event_handler;
2759 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2760 	if (IS_ERR(qp_info->qp)) {
2761 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2762 		       get_spl_qp_index(qp_type));
2763 		ret = PTR_ERR(qp_info->qp);
2764 		goto error;
2765 	}
2766 	/* Use minimum queue sizes unless the CQ is resized */
2767 	qp_info->send_queue.max_active = mad_sendq_size;
2768 	qp_info->recv_queue.max_active = mad_recvq_size;
2769 	return 0;
2770 
2771 error:
2772 	return ret;
2773 }
2774 
2775 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2776 {
2777 	if (!qp_info->qp)
2778 		return;
2779 
2780 	ib_destroy_qp(qp_info->qp);
2781 	kfree(qp_info->snoop_table);
2782 }
2783 
2784 /*
2785  * Open the port
2786  * Create the QP, PD, MR, and CQ if needed
2787  */
2788 static int ib_mad_port_open(struct ib_device *device,
2789 			    int port_num)
2790 {
2791 	int ret, cq_size;
2792 	struct ib_mad_port_private *port_priv;
2793 	unsigned long flags;
2794 	char name[sizeof "ib_mad123"];
2795 	int has_smi;
2796 
2797 	/* Create new device info */
2798 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2799 	if (!port_priv) {
2800 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2801 		return -ENOMEM;
2802 	}
2803 
2804 	port_priv->device = device;
2805 	port_priv->port_num = port_num;
2806 	spin_lock_init(&port_priv->reg_lock);
2807 	INIT_LIST_HEAD(&port_priv->agent_list);
2808 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2809 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2810 
2811 	cq_size = mad_sendq_size + mad_recvq_size;
2812 	has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2813 	if (has_smi)
2814 		cq_size *= 2;
2815 
2816 	port_priv->cq = ib_create_cq(port_priv->device,
2817 				     ib_mad_thread_completion_handler,
2818 				     NULL, port_priv, cq_size, 0);
2819 	if (IS_ERR(port_priv->cq)) {
2820 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2821 		ret = PTR_ERR(port_priv->cq);
2822 		goto error3;
2823 	}
2824 
2825 	port_priv->pd = ib_alloc_pd(device);
2826 	if (IS_ERR(port_priv->pd)) {
2827 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2828 		ret = PTR_ERR(port_priv->pd);
2829 		goto error4;
2830 	}
2831 
2832 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2833 	if (IS_ERR(port_priv->mr)) {
2834 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2835 		ret = PTR_ERR(port_priv->mr);
2836 		goto error5;
2837 	}
2838 
2839 	if (has_smi) {
2840 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2841 		if (ret)
2842 			goto error6;
2843 	}
2844 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2845 	if (ret)
2846 		goto error7;
2847 
2848 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2849 	port_priv->wq = create_singlethread_workqueue(name);
2850 	if (!port_priv->wq) {
2851 		ret = -ENOMEM;
2852 		goto error8;
2853 	}
2854 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2855 
2856 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2857 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2858 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2859 
2860 	ret = ib_mad_port_start(port_priv);
2861 	if (ret) {
2862 		printk(KERN_ERR PFX "Couldn't start port\n");
2863 		goto error9;
2864 	}
2865 
2866 	return 0;
2867 
2868 error9:
2869 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2870 	list_del_init(&port_priv->port_list);
2871 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2872 
2873 	destroy_workqueue(port_priv->wq);
2874 error8:
2875 	destroy_mad_qp(&port_priv->qp_info[1]);
2876 error7:
2877 	destroy_mad_qp(&port_priv->qp_info[0]);
2878 error6:
2879 	ib_dereg_mr(port_priv->mr);
2880 error5:
2881 	ib_dealloc_pd(port_priv->pd);
2882 error4:
2883 	ib_destroy_cq(port_priv->cq);
2884 	cleanup_recv_queue(&port_priv->qp_info[1]);
2885 	cleanup_recv_queue(&port_priv->qp_info[0]);
2886 error3:
2887 	kfree(port_priv);
2888 
2889 	return ret;
2890 }
2891 
2892 /*
2893  * Close the port
2894  * If there are no classes using the port, free the port
2895  * resources (CQ, MR, PD, QP) and remove the port's info structure
2896  */
2897 static int ib_mad_port_close(struct ib_device *device, int port_num)
2898 {
2899 	struct ib_mad_port_private *port_priv;
2900 	unsigned long flags;
2901 
2902 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2903 	port_priv = __ib_get_mad_port(device, port_num);
2904 	if (port_priv == NULL) {
2905 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2906 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2907 		return -ENODEV;
2908 	}
2909 	list_del_init(&port_priv->port_list);
2910 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2911 
2912 	destroy_workqueue(port_priv->wq);
2913 	destroy_mad_qp(&port_priv->qp_info[1]);
2914 	destroy_mad_qp(&port_priv->qp_info[0]);
2915 	ib_dereg_mr(port_priv->mr);
2916 	ib_dealloc_pd(port_priv->pd);
2917 	ib_destroy_cq(port_priv->cq);
2918 	cleanup_recv_queue(&port_priv->qp_info[1]);
2919 	cleanup_recv_queue(&port_priv->qp_info[0]);
2920 	/* XXX: Handle deallocation of MAD registration tables */
2921 
2922 	kfree(port_priv);
2923 
2924 	return 0;
2925 }
2926 
2927 static void ib_mad_init_device(struct ib_device *device)
2928 {
2929 	int start, end, i;
2930 
2931 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2932 		return;
2933 
2934 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2935 		start = 0;
2936 		end   = 0;
2937 	} else {
2938 		start = 1;
2939 		end   = device->phys_port_cnt;
2940 	}
2941 
2942 	for (i = start; i <= end; i++) {
2943 		if (ib_mad_port_open(device, i)) {
2944 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2945 			       device->name, i);
2946 			goto error;
2947 		}
2948 		if (ib_agent_port_open(device, i)) {
2949 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2950 			       "for agents\n",
2951 			       device->name, i);
2952 			goto error_agent;
2953 		}
2954 	}
2955 	return;
2956 
2957 error_agent:
2958 	if (ib_mad_port_close(device, i))
2959 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2960 		       device->name, i);
2961 
2962 error:
2963 	i--;
2964 
2965 	while (i >= start) {
2966 		if (ib_agent_port_close(device, i))
2967 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2968 			       "for agents\n",
2969 			       device->name, i);
2970 		if (ib_mad_port_close(device, i))
2971 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2972 			       device->name, i);
2973 		i--;
2974 	}
2975 }
2976 
2977 static void ib_mad_remove_device(struct ib_device *device)
2978 {
2979 	int i, num_ports, cur_port;
2980 
2981 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2982 		return;
2983 
2984 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2985 		num_ports = 1;
2986 		cur_port = 0;
2987 	} else {
2988 		num_ports = device->phys_port_cnt;
2989 		cur_port = 1;
2990 	}
2991 	for (i = 0; i < num_ports; i++, cur_port++) {
2992 		if (ib_agent_port_close(device, cur_port))
2993 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2994 			       "for agents\n",
2995 			       device->name, cur_port);
2996 		if (ib_mad_port_close(device, cur_port))
2997 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2998 			       device->name, cur_port);
2999 	}
3000 }
3001 
3002 static struct ib_client mad_client = {
3003 	.name   = "mad",
3004 	.add = ib_mad_init_device,
3005 	.remove = ib_mad_remove_device
3006 };
3007 
3008 static int __init ib_mad_init_module(void)
3009 {
3010 	int ret;
3011 
3012 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3013 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3014 
3015 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3016 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3017 
3018 	ib_mad_cache = kmem_cache_create("ib_mad",
3019 					 sizeof(struct ib_mad_private),
3020 					 0,
3021 					 SLAB_HWCACHE_ALIGN,
3022 					 NULL);
3023 	if (!ib_mad_cache) {
3024 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
3025 		ret = -ENOMEM;
3026 		goto error1;
3027 	}
3028 
3029 	INIT_LIST_HEAD(&ib_mad_port_list);
3030 
3031 	if (ib_register_client(&mad_client)) {
3032 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3033 		ret = -EINVAL;
3034 		goto error2;
3035 	}
3036 
3037 	return 0;
3038 
3039 error2:
3040 	kmem_cache_destroy(ib_mad_cache);
3041 error1:
3042 	return ret;
3043 }
3044 
3045 static void __exit ib_mad_cleanup_module(void)
3046 {
3047 	ib_unregister_client(&mad_client);
3048 	kmem_cache_destroy(ib_mad_cache);
3049 }
3050 
3051 module_init(ib_mad_init_module);
3052 module_exit(ib_mad_cleanup_module);
3053