xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 8a10bc9d)
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 <linux/module.h>
39 #include <rdma/ib_cache.h>
40 
41 #include "mad_priv.h"
42 #include "mad_rmpp.h"
43 #include "smi.h"
44 #include "agent.h"
45 
46 MODULE_LICENSE("Dual BSD/GPL");
47 MODULE_DESCRIPTION("kernel IB MAD API");
48 MODULE_AUTHOR("Hal Rosenstock");
49 MODULE_AUTHOR("Sean Hefty");
50 
51 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
52 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
53 
54 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
55 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
56 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
57 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
58 
59 static struct kmem_cache *ib_mad_cache;
60 
61 static struct list_head ib_mad_port_list;
62 static u32 ib_mad_client_id = 0;
63 
64 /* Port list lock */
65 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
66 
67 /* Forward declarations */
68 static int method_in_use(struct ib_mad_mgmt_method_table **method,
69 			 struct ib_mad_reg_req *mad_reg_req);
70 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
71 static struct ib_mad_agent_private *find_mad_agent(
72 					struct ib_mad_port_private *port_priv,
73 					struct ib_mad *mad);
74 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
75 				    struct ib_mad_private *mad);
76 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
77 static void timeout_sends(struct work_struct *work);
78 static void local_completions(struct work_struct *work);
79 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
80 			      struct ib_mad_agent_private *agent_priv,
81 			      u8 mgmt_class);
82 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 			   struct ib_mad_agent_private *agent_priv);
84 
85 /*
86  * Returns a ib_mad_port_private structure or NULL for a device/port
87  * Assumes ib_mad_port_list_lock is being held
88  */
89 static inline struct ib_mad_port_private *
90 __ib_get_mad_port(struct ib_device *device, int port_num)
91 {
92 	struct ib_mad_port_private *entry;
93 
94 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
95 		if (entry->device == device && entry->port_num == port_num)
96 			return entry;
97 	}
98 	return NULL;
99 }
100 
101 /*
102  * Wrapper function to return a ib_mad_port_private structure or NULL
103  * for a device/port
104  */
105 static inline struct ib_mad_port_private *
106 ib_get_mad_port(struct ib_device *device, int port_num)
107 {
108 	struct ib_mad_port_private *entry;
109 	unsigned long flags;
110 
111 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
112 	entry = __ib_get_mad_port(device, port_num);
113 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
114 
115 	return entry;
116 }
117 
118 static inline u8 convert_mgmt_class(u8 mgmt_class)
119 {
120 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
121 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
122 		0 : mgmt_class;
123 }
124 
125 static int get_spl_qp_index(enum ib_qp_type qp_type)
126 {
127 	switch (qp_type)
128 	{
129 	case IB_QPT_SMI:
130 		return 0;
131 	case IB_QPT_GSI:
132 		return 1;
133 	default:
134 		return -1;
135 	}
136 }
137 
138 static int vendor_class_index(u8 mgmt_class)
139 {
140 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
141 }
142 
143 static int is_vendor_class(u8 mgmt_class)
144 {
145 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
146 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
147 		return 0;
148 	return 1;
149 }
150 
151 static int is_vendor_oui(char *oui)
152 {
153 	if (oui[0] || oui[1] || oui[2])
154 		return 1;
155 	return 0;
156 }
157 
158 static int is_vendor_method_in_use(
159 		struct ib_mad_mgmt_vendor_class *vendor_class,
160 		struct ib_mad_reg_req *mad_reg_req)
161 {
162 	struct ib_mad_mgmt_method_table *method;
163 	int i;
164 
165 	for (i = 0; i < MAX_MGMT_OUI; i++) {
166 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
167 			method = vendor_class->method_table[i];
168 			if (method) {
169 				if (method_in_use(&method, mad_reg_req))
170 					return 1;
171 				else
172 					break;
173 			}
174 		}
175 	}
176 	return 0;
177 }
178 
179 int ib_response_mad(struct ib_mad *mad)
180 {
181 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
182 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
183 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
184 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
185 }
186 EXPORT_SYMBOL(ib_response_mad);
187 
188 /*
189  * ib_register_mad_agent - Register to send/receive MADs
190  */
191 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
192 					   u8 port_num,
193 					   enum ib_qp_type qp_type,
194 					   struct ib_mad_reg_req *mad_reg_req,
195 					   u8 rmpp_version,
196 					   ib_mad_send_handler send_handler,
197 					   ib_mad_recv_handler recv_handler,
198 					   void *context)
199 {
200 	struct ib_mad_port_private *port_priv;
201 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
202 	struct ib_mad_agent_private *mad_agent_priv;
203 	struct ib_mad_reg_req *reg_req = NULL;
204 	struct ib_mad_mgmt_class_table *class;
205 	struct ib_mad_mgmt_vendor_class_table *vendor;
206 	struct ib_mad_mgmt_vendor_class *vendor_class;
207 	struct ib_mad_mgmt_method_table *method;
208 	int ret2, qpn;
209 	unsigned long flags;
210 	u8 mgmt_class, vclass;
211 
212 	/* Validate parameters */
213 	qpn = get_spl_qp_index(qp_type);
214 	if (qpn == -1)
215 		goto error1;
216 
217 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
218 		goto error1;
219 
220 	/* Validate MAD registration request if supplied */
221 	if (mad_reg_req) {
222 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
223 			goto error1;
224 		if (!recv_handler)
225 			goto error1;
226 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
227 			/*
228 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
229 			 * one in this range currently allowed
230 			 */
231 			if (mad_reg_req->mgmt_class !=
232 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
233 				goto error1;
234 		} else if (mad_reg_req->mgmt_class == 0) {
235 			/*
236 			 * Class 0 is reserved in IBA and is used for
237 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
238 			 */
239 			goto error1;
240 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
241 			/*
242 			 * If class is in "new" vendor range,
243 			 * ensure supplied OUI is not zero
244 			 */
245 			if (!is_vendor_oui(mad_reg_req->oui))
246 				goto error1;
247 		}
248 		/* Make sure class supplied is consistent with RMPP */
249 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
250 			if (rmpp_version)
251 				goto error1;
252 		}
253 		/* Make sure class supplied is consistent with QP type */
254 		if (qp_type == IB_QPT_SMI) {
255 			if ((mad_reg_req->mgmt_class !=
256 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
257 			    (mad_reg_req->mgmt_class !=
258 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
259 				goto error1;
260 		} else {
261 			if ((mad_reg_req->mgmt_class ==
262 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
263 			    (mad_reg_req->mgmt_class ==
264 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
265 				goto error1;
266 		}
267 	} else {
268 		/* No registration request supplied */
269 		if (!send_handler)
270 			goto error1;
271 	}
272 
273 	/* Validate device and port */
274 	port_priv = ib_get_mad_port(device, port_num);
275 	if (!port_priv) {
276 		ret = ERR_PTR(-ENODEV);
277 		goto error1;
278 	}
279 
280 	/* Verify the QP requested is supported.  For example, Ethernet devices
281 	 * will not have QP0 */
282 	if (!port_priv->qp_info[qpn].qp) {
283 		ret = ERR_PTR(-EPROTONOSUPPORT);
284 		goto error1;
285 	}
286 
287 	/* Allocate structures */
288 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
289 	if (!mad_agent_priv) {
290 		ret = ERR_PTR(-ENOMEM);
291 		goto error1;
292 	}
293 
294 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
295 						 IB_ACCESS_LOCAL_WRITE);
296 	if (IS_ERR(mad_agent_priv->agent.mr)) {
297 		ret = ERR_PTR(-ENOMEM);
298 		goto error2;
299 	}
300 
301 	if (mad_reg_req) {
302 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
303 		if (!reg_req) {
304 			ret = ERR_PTR(-ENOMEM);
305 			goto error3;
306 		}
307 	}
308 
309 	/* Now, fill in the various structures */
310 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
311 	mad_agent_priv->reg_req = reg_req;
312 	mad_agent_priv->agent.rmpp_version = rmpp_version;
313 	mad_agent_priv->agent.device = device;
314 	mad_agent_priv->agent.recv_handler = recv_handler;
315 	mad_agent_priv->agent.send_handler = send_handler;
316 	mad_agent_priv->agent.context = context;
317 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
318 	mad_agent_priv->agent.port_num = port_num;
319 	spin_lock_init(&mad_agent_priv->lock);
320 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
321 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
322 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
323 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
324 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
325 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
326 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
327 	atomic_set(&mad_agent_priv->refcount, 1);
328 	init_completion(&mad_agent_priv->comp);
329 
330 	spin_lock_irqsave(&port_priv->reg_lock, flags);
331 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
332 
333 	/*
334 	 * Make sure MAD registration (if supplied)
335 	 * is non overlapping with any existing ones
336 	 */
337 	if (mad_reg_req) {
338 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
339 		if (!is_vendor_class(mgmt_class)) {
340 			class = port_priv->version[mad_reg_req->
341 						   mgmt_class_version].class;
342 			if (class) {
343 				method = class->method_table[mgmt_class];
344 				if (method) {
345 					if (method_in_use(&method,
346 							   mad_reg_req))
347 						goto error4;
348 				}
349 			}
350 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
351 						  mgmt_class);
352 		} else {
353 			/* "New" vendor class range */
354 			vendor = port_priv->version[mad_reg_req->
355 						    mgmt_class_version].vendor;
356 			if (vendor) {
357 				vclass = vendor_class_index(mgmt_class);
358 				vendor_class = vendor->vendor_class[vclass];
359 				if (vendor_class) {
360 					if (is_vendor_method_in_use(
361 							vendor_class,
362 							mad_reg_req))
363 						goto error4;
364 				}
365 			}
366 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
367 		}
368 		if (ret2) {
369 			ret = ERR_PTR(ret2);
370 			goto error4;
371 		}
372 	}
373 
374 	/* Add mad agent into port's agent list */
375 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
376 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
377 
378 	return &mad_agent_priv->agent;
379 
380 error4:
381 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
382 	kfree(reg_req);
383 error3:
384 	ib_dereg_mr(mad_agent_priv->agent.mr);
385 error2:
386 	kfree(mad_agent_priv);
387 error1:
388 	return ret;
389 }
390 EXPORT_SYMBOL(ib_register_mad_agent);
391 
392 static inline int is_snooping_sends(int mad_snoop_flags)
393 {
394 	return (mad_snoop_flags &
395 		(/*IB_MAD_SNOOP_POSTED_SENDS |
396 		 IB_MAD_SNOOP_RMPP_SENDS |*/
397 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
398 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
399 }
400 
401 static inline int is_snooping_recvs(int mad_snoop_flags)
402 {
403 	return (mad_snoop_flags &
404 		(IB_MAD_SNOOP_RECVS /*|
405 		 IB_MAD_SNOOP_RMPP_RECVS*/));
406 }
407 
408 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
409 				struct ib_mad_snoop_private *mad_snoop_priv)
410 {
411 	struct ib_mad_snoop_private **new_snoop_table;
412 	unsigned long flags;
413 	int i;
414 
415 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
416 	/* Check for empty slot in array. */
417 	for (i = 0; i < qp_info->snoop_table_size; i++)
418 		if (!qp_info->snoop_table[i])
419 			break;
420 
421 	if (i == qp_info->snoop_table_size) {
422 		/* Grow table. */
423 		new_snoop_table = krealloc(qp_info->snoop_table,
424 					   sizeof mad_snoop_priv *
425 					   (qp_info->snoop_table_size + 1),
426 					   GFP_ATOMIC);
427 		if (!new_snoop_table) {
428 			i = -ENOMEM;
429 			goto out;
430 		}
431 
432 		qp_info->snoop_table = new_snoop_table;
433 		qp_info->snoop_table_size++;
434 	}
435 	qp_info->snoop_table[i] = mad_snoop_priv;
436 	atomic_inc(&qp_info->snoop_count);
437 out:
438 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
439 	return i;
440 }
441 
442 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
443 					   u8 port_num,
444 					   enum ib_qp_type qp_type,
445 					   int mad_snoop_flags,
446 					   ib_mad_snoop_handler snoop_handler,
447 					   ib_mad_recv_handler recv_handler,
448 					   void *context)
449 {
450 	struct ib_mad_port_private *port_priv;
451 	struct ib_mad_agent *ret;
452 	struct ib_mad_snoop_private *mad_snoop_priv;
453 	int qpn;
454 
455 	/* Validate parameters */
456 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
457 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
458 		ret = ERR_PTR(-EINVAL);
459 		goto error1;
460 	}
461 	qpn = get_spl_qp_index(qp_type);
462 	if (qpn == -1) {
463 		ret = ERR_PTR(-EINVAL);
464 		goto error1;
465 	}
466 	port_priv = ib_get_mad_port(device, port_num);
467 	if (!port_priv) {
468 		ret = ERR_PTR(-ENODEV);
469 		goto error1;
470 	}
471 	/* Allocate structures */
472 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
473 	if (!mad_snoop_priv) {
474 		ret = ERR_PTR(-ENOMEM);
475 		goto error1;
476 	}
477 
478 	/* Now, fill in the various structures */
479 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
480 	mad_snoop_priv->agent.device = device;
481 	mad_snoop_priv->agent.recv_handler = recv_handler;
482 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
483 	mad_snoop_priv->agent.context = context;
484 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
485 	mad_snoop_priv->agent.port_num = port_num;
486 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
487 	init_completion(&mad_snoop_priv->comp);
488 	mad_snoop_priv->snoop_index = register_snoop_agent(
489 						&port_priv->qp_info[qpn],
490 						mad_snoop_priv);
491 	if (mad_snoop_priv->snoop_index < 0) {
492 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
493 		goto error2;
494 	}
495 
496 	atomic_set(&mad_snoop_priv->refcount, 1);
497 	return &mad_snoop_priv->agent;
498 
499 error2:
500 	kfree(mad_snoop_priv);
501 error1:
502 	return ret;
503 }
504 EXPORT_SYMBOL(ib_register_mad_snoop);
505 
506 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
507 {
508 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
509 		complete(&mad_agent_priv->comp);
510 }
511 
512 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
513 {
514 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
515 		complete(&mad_snoop_priv->comp);
516 }
517 
518 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
519 {
520 	struct ib_mad_port_private *port_priv;
521 	unsigned long flags;
522 
523 	/* Note that we could still be handling received MADs */
524 
525 	/*
526 	 * Canceling all sends results in dropping received response
527 	 * MADs, preventing us from queuing additional work
528 	 */
529 	cancel_mads(mad_agent_priv);
530 	port_priv = mad_agent_priv->qp_info->port_priv;
531 	cancel_delayed_work(&mad_agent_priv->timed_work);
532 
533 	spin_lock_irqsave(&port_priv->reg_lock, flags);
534 	remove_mad_reg_req(mad_agent_priv);
535 	list_del(&mad_agent_priv->agent_list);
536 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
537 
538 	flush_workqueue(port_priv->wq);
539 	ib_cancel_rmpp_recvs(mad_agent_priv);
540 
541 	deref_mad_agent(mad_agent_priv);
542 	wait_for_completion(&mad_agent_priv->comp);
543 
544 	kfree(mad_agent_priv->reg_req);
545 	ib_dereg_mr(mad_agent_priv->agent.mr);
546 	kfree(mad_agent_priv);
547 }
548 
549 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
550 {
551 	struct ib_mad_qp_info *qp_info;
552 	unsigned long flags;
553 
554 	qp_info = mad_snoop_priv->qp_info;
555 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
556 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
557 	atomic_dec(&qp_info->snoop_count);
558 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
559 
560 	deref_snoop_agent(mad_snoop_priv);
561 	wait_for_completion(&mad_snoop_priv->comp);
562 
563 	kfree(mad_snoop_priv);
564 }
565 
566 /*
567  * ib_unregister_mad_agent - Unregisters a client from using MAD services
568  */
569 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
570 {
571 	struct ib_mad_agent_private *mad_agent_priv;
572 	struct ib_mad_snoop_private *mad_snoop_priv;
573 
574 	/* If the TID is zero, the agent can only snoop. */
575 	if (mad_agent->hi_tid) {
576 		mad_agent_priv = container_of(mad_agent,
577 					      struct ib_mad_agent_private,
578 					      agent);
579 		unregister_mad_agent(mad_agent_priv);
580 	} else {
581 		mad_snoop_priv = container_of(mad_agent,
582 					      struct ib_mad_snoop_private,
583 					      agent);
584 		unregister_mad_snoop(mad_snoop_priv);
585 	}
586 	return 0;
587 }
588 EXPORT_SYMBOL(ib_unregister_mad_agent);
589 
590 static void dequeue_mad(struct ib_mad_list_head *mad_list)
591 {
592 	struct ib_mad_queue *mad_queue;
593 	unsigned long flags;
594 
595 	BUG_ON(!mad_list->mad_queue);
596 	mad_queue = mad_list->mad_queue;
597 	spin_lock_irqsave(&mad_queue->lock, flags);
598 	list_del(&mad_list->list);
599 	mad_queue->count--;
600 	spin_unlock_irqrestore(&mad_queue->lock, flags);
601 }
602 
603 static void snoop_send(struct ib_mad_qp_info *qp_info,
604 		       struct ib_mad_send_buf *send_buf,
605 		       struct ib_mad_send_wc *mad_send_wc,
606 		       int mad_snoop_flags)
607 {
608 	struct ib_mad_snoop_private *mad_snoop_priv;
609 	unsigned long flags;
610 	int i;
611 
612 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
613 	for (i = 0; i < qp_info->snoop_table_size; i++) {
614 		mad_snoop_priv = qp_info->snoop_table[i];
615 		if (!mad_snoop_priv ||
616 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
617 			continue;
618 
619 		atomic_inc(&mad_snoop_priv->refcount);
620 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
621 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
622 						    send_buf, mad_send_wc);
623 		deref_snoop_agent(mad_snoop_priv);
624 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
625 	}
626 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
627 }
628 
629 static void snoop_recv(struct ib_mad_qp_info *qp_info,
630 		       struct ib_mad_recv_wc *mad_recv_wc,
631 		       int mad_snoop_flags)
632 {
633 	struct ib_mad_snoop_private *mad_snoop_priv;
634 	unsigned long flags;
635 	int i;
636 
637 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
638 	for (i = 0; i < qp_info->snoop_table_size; i++) {
639 		mad_snoop_priv = qp_info->snoop_table[i];
640 		if (!mad_snoop_priv ||
641 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
642 			continue;
643 
644 		atomic_inc(&mad_snoop_priv->refcount);
645 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
646 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
647 						   mad_recv_wc);
648 		deref_snoop_agent(mad_snoop_priv);
649 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
650 	}
651 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
652 }
653 
654 static void build_smp_wc(struct ib_qp *qp,
655 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
656 			 struct ib_wc *wc)
657 {
658 	memset(wc, 0, sizeof *wc);
659 	wc->wr_id = wr_id;
660 	wc->status = IB_WC_SUCCESS;
661 	wc->opcode = IB_WC_RECV;
662 	wc->pkey_index = pkey_index;
663 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
664 	wc->src_qp = IB_QP0;
665 	wc->qp = qp;
666 	wc->slid = slid;
667 	wc->sl = 0;
668 	wc->dlid_path_bits = 0;
669 	wc->port_num = port_num;
670 }
671 
672 /*
673  * Return 0 if SMP is to be sent
674  * Return 1 if SMP was consumed locally (whether or not solicited)
675  * Return < 0 if error
676  */
677 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
678 				  struct ib_mad_send_wr_private *mad_send_wr)
679 {
680 	int ret = 0;
681 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
682 	unsigned long flags;
683 	struct ib_mad_local_private *local;
684 	struct ib_mad_private *mad_priv;
685 	struct ib_mad_port_private *port_priv;
686 	struct ib_mad_agent_private *recv_mad_agent = NULL;
687 	struct ib_device *device = mad_agent_priv->agent.device;
688 	u8 port_num;
689 	struct ib_wc mad_wc;
690 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
691 
692 	if (device->node_type == RDMA_NODE_IB_SWITCH &&
693 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
694 		port_num = send_wr->wr.ud.port_num;
695 	else
696 		port_num = mad_agent_priv->agent.port_num;
697 
698 	/*
699 	 * Directed route handling starts if the initial LID routed part of
700 	 * a request or the ending LID routed part of a response is empty.
701 	 * If we are at the start of the LID routed part, don't update the
702 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
703 	 */
704 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
705 	     IB_LID_PERMISSIVE &&
706 	     smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
707 	     IB_SMI_DISCARD) {
708 		ret = -EINVAL;
709 		printk(KERN_ERR PFX "Invalid directed route\n");
710 		goto out;
711 	}
712 
713 	/* Check to post send on QP or process locally */
714 	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
715 	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
716 		goto out;
717 
718 	local = kmalloc(sizeof *local, GFP_ATOMIC);
719 	if (!local) {
720 		ret = -ENOMEM;
721 		printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
722 		goto out;
723 	}
724 	local->mad_priv = NULL;
725 	local->recv_mad_agent = NULL;
726 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
727 	if (!mad_priv) {
728 		ret = -ENOMEM;
729 		printk(KERN_ERR PFX "No memory for local response MAD\n");
730 		kfree(local);
731 		goto out;
732 	}
733 
734 	build_smp_wc(mad_agent_priv->agent.qp,
735 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
736 		     send_wr->wr.ud.pkey_index,
737 		     send_wr->wr.ud.port_num, &mad_wc);
738 
739 	/* No GRH for DR SMP */
740 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
741 				  (struct ib_mad *)smp,
742 				  (struct ib_mad *)&mad_priv->mad);
743 	switch (ret)
744 	{
745 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
746 		if (ib_response_mad(&mad_priv->mad.mad) &&
747 		    mad_agent_priv->agent.recv_handler) {
748 			local->mad_priv = mad_priv;
749 			local->recv_mad_agent = mad_agent_priv;
750 			/*
751 			 * Reference MAD agent until receive
752 			 * side of local completion handled
753 			 */
754 			atomic_inc(&mad_agent_priv->refcount);
755 		} else
756 			kmem_cache_free(ib_mad_cache, mad_priv);
757 		break;
758 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
759 		kmem_cache_free(ib_mad_cache, mad_priv);
760 		break;
761 	case IB_MAD_RESULT_SUCCESS:
762 		/* Treat like an incoming receive MAD */
763 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
764 					    mad_agent_priv->agent.port_num);
765 		if (port_priv) {
766 			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
767 			recv_mad_agent = find_mad_agent(port_priv,
768 						        &mad_priv->mad.mad);
769 		}
770 		if (!port_priv || !recv_mad_agent) {
771 			/*
772 			 * No receiving agent so drop packet and
773 			 * generate send completion.
774 			 */
775 			kmem_cache_free(ib_mad_cache, mad_priv);
776 			break;
777 		}
778 		local->mad_priv = mad_priv;
779 		local->recv_mad_agent = recv_mad_agent;
780 		break;
781 	default:
782 		kmem_cache_free(ib_mad_cache, mad_priv);
783 		kfree(local);
784 		ret = -EINVAL;
785 		goto out;
786 	}
787 
788 	local->mad_send_wr = mad_send_wr;
789 	/* Reference MAD agent until send side of local completion handled */
790 	atomic_inc(&mad_agent_priv->refcount);
791 	/* Queue local completion to local list */
792 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
793 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
794 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
795 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
796 		   &mad_agent_priv->local_work);
797 	ret = 1;
798 out:
799 	return ret;
800 }
801 
802 static int get_pad_size(int hdr_len, int data_len)
803 {
804 	int seg_size, pad;
805 
806 	seg_size = sizeof(struct ib_mad) - hdr_len;
807 	if (data_len && seg_size) {
808 		pad = seg_size - data_len % seg_size;
809 		return pad == seg_size ? 0 : pad;
810 	} else
811 		return seg_size;
812 }
813 
814 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
815 {
816 	struct ib_rmpp_segment *s, *t;
817 
818 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
819 		list_del(&s->list);
820 		kfree(s);
821 	}
822 }
823 
824 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
825 				gfp_t gfp_mask)
826 {
827 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
828 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
829 	struct ib_rmpp_segment *seg = NULL;
830 	int left, seg_size, pad;
831 
832 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
833 	seg_size = send_buf->seg_size;
834 	pad = send_wr->pad;
835 
836 	/* Allocate data segments. */
837 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
838 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
839 		if (!seg) {
840 			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
841 			       "alloc failed for len %zd, gfp %#x\n",
842 			       sizeof (*seg) + seg_size, gfp_mask);
843 			free_send_rmpp_list(send_wr);
844 			return -ENOMEM;
845 		}
846 		seg->num = ++send_buf->seg_count;
847 		list_add_tail(&seg->list, &send_wr->rmpp_list);
848 	}
849 
850 	/* Zero any padding */
851 	if (pad)
852 		memset(seg->data + seg_size - pad, 0, pad);
853 
854 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
855 					  agent.rmpp_version;
856 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
857 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
858 
859 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
860 					struct ib_rmpp_segment, list);
861 	send_wr->last_ack_seg = send_wr->cur_seg;
862 	return 0;
863 }
864 
865 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
866 					    u32 remote_qpn, u16 pkey_index,
867 					    int rmpp_active,
868 					    int hdr_len, int data_len,
869 					    gfp_t gfp_mask)
870 {
871 	struct ib_mad_agent_private *mad_agent_priv;
872 	struct ib_mad_send_wr_private *mad_send_wr;
873 	int pad, message_size, ret, size;
874 	void *buf;
875 
876 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
877 				      agent);
878 	pad = get_pad_size(hdr_len, data_len);
879 	message_size = hdr_len + data_len + pad;
880 
881 	if ((!mad_agent->rmpp_version &&
882 	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
883 	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
884 		return ERR_PTR(-EINVAL);
885 
886 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
887 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
888 	if (!buf)
889 		return ERR_PTR(-ENOMEM);
890 
891 	mad_send_wr = buf + size;
892 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
893 	mad_send_wr->send_buf.mad = buf;
894 	mad_send_wr->send_buf.hdr_len = hdr_len;
895 	mad_send_wr->send_buf.data_len = data_len;
896 	mad_send_wr->pad = pad;
897 
898 	mad_send_wr->mad_agent_priv = mad_agent_priv;
899 	mad_send_wr->sg_list[0].length = hdr_len;
900 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
901 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
902 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
903 
904 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
905 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
906 	mad_send_wr->send_wr.num_sge = 2;
907 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
908 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
909 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
910 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
911 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
912 
913 	if (rmpp_active) {
914 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
915 		if (ret) {
916 			kfree(buf);
917 			return ERR_PTR(ret);
918 		}
919 	}
920 
921 	mad_send_wr->send_buf.mad_agent = mad_agent;
922 	atomic_inc(&mad_agent_priv->refcount);
923 	return &mad_send_wr->send_buf;
924 }
925 EXPORT_SYMBOL(ib_create_send_mad);
926 
927 int ib_get_mad_data_offset(u8 mgmt_class)
928 {
929 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
930 		return IB_MGMT_SA_HDR;
931 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
932 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
933 		 (mgmt_class == IB_MGMT_CLASS_BIS))
934 		return IB_MGMT_DEVICE_HDR;
935 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
936 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
937 		return IB_MGMT_VENDOR_HDR;
938 	else
939 		return IB_MGMT_MAD_HDR;
940 }
941 EXPORT_SYMBOL(ib_get_mad_data_offset);
942 
943 int ib_is_mad_class_rmpp(u8 mgmt_class)
944 {
945 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
946 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
947 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
948 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
949 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
950 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
951 		return 1;
952 	return 0;
953 }
954 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
955 
956 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
957 {
958 	struct ib_mad_send_wr_private *mad_send_wr;
959 	struct list_head *list;
960 
961 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
962 				   send_buf);
963 	list = &mad_send_wr->cur_seg->list;
964 
965 	if (mad_send_wr->cur_seg->num < seg_num) {
966 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
967 			if (mad_send_wr->cur_seg->num == seg_num)
968 				break;
969 	} else if (mad_send_wr->cur_seg->num > seg_num) {
970 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
971 			if (mad_send_wr->cur_seg->num == seg_num)
972 				break;
973 	}
974 	return mad_send_wr->cur_seg->data;
975 }
976 EXPORT_SYMBOL(ib_get_rmpp_segment);
977 
978 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
979 {
980 	if (mad_send_wr->send_buf.seg_count)
981 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
982 					   mad_send_wr->seg_num);
983 	else
984 		return mad_send_wr->send_buf.mad +
985 		       mad_send_wr->send_buf.hdr_len;
986 }
987 
988 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
989 {
990 	struct ib_mad_agent_private *mad_agent_priv;
991 	struct ib_mad_send_wr_private *mad_send_wr;
992 
993 	mad_agent_priv = container_of(send_buf->mad_agent,
994 				      struct ib_mad_agent_private, agent);
995 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
996 				   send_buf);
997 
998 	free_send_rmpp_list(mad_send_wr);
999 	kfree(send_buf->mad);
1000 	deref_mad_agent(mad_agent_priv);
1001 }
1002 EXPORT_SYMBOL(ib_free_send_mad);
1003 
1004 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1005 {
1006 	struct ib_mad_qp_info *qp_info;
1007 	struct list_head *list;
1008 	struct ib_send_wr *bad_send_wr;
1009 	struct ib_mad_agent *mad_agent;
1010 	struct ib_sge *sge;
1011 	unsigned long flags;
1012 	int ret;
1013 
1014 	/* Set WR ID to find mad_send_wr upon completion */
1015 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1016 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1017 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1018 
1019 	mad_agent = mad_send_wr->send_buf.mad_agent;
1020 	sge = mad_send_wr->sg_list;
1021 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1022 					mad_send_wr->send_buf.mad,
1023 					sge[0].length,
1024 					DMA_TO_DEVICE);
1025 	mad_send_wr->header_mapping = sge[0].addr;
1026 
1027 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1028 					ib_get_payload(mad_send_wr),
1029 					sge[1].length,
1030 					DMA_TO_DEVICE);
1031 	mad_send_wr->payload_mapping = sge[1].addr;
1032 
1033 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1034 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1035 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1036 				   &bad_send_wr);
1037 		list = &qp_info->send_queue.list;
1038 	} else {
1039 		ret = 0;
1040 		list = &qp_info->overflow_list;
1041 	}
1042 
1043 	if (!ret) {
1044 		qp_info->send_queue.count++;
1045 		list_add_tail(&mad_send_wr->mad_list.list, list);
1046 	}
1047 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1048 	if (ret) {
1049 		ib_dma_unmap_single(mad_agent->device,
1050 				    mad_send_wr->header_mapping,
1051 				    sge[0].length, DMA_TO_DEVICE);
1052 		ib_dma_unmap_single(mad_agent->device,
1053 				    mad_send_wr->payload_mapping,
1054 				    sge[1].length, DMA_TO_DEVICE);
1055 	}
1056 	return ret;
1057 }
1058 
1059 /*
1060  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1061  *  with the registered client
1062  */
1063 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1064 		     struct ib_mad_send_buf **bad_send_buf)
1065 {
1066 	struct ib_mad_agent_private *mad_agent_priv;
1067 	struct ib_mad_send_buf *next_send_buf;
1068 	struct ib_mad_send_wr_private *mad_send_wr;
1069 	unsigned long flags;
1070 	int ret = -EINVAL;
1071 
1072 	/* Walk list of send WRs and post each on send list */
1073 	for (; send_buf; send_buf = next_send_buf) {
1074 
1075 		mad_send_wr = container_of(send_buf,
1076 					   struct ib_mad_send_wr_private,
1077 					   send_buf);
1078 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1079 
1080 		if (!send_buf->mad_agent->send_handler ||
1081 		    (send_buf->timeout_ms &&
1082 		     !send_buf->mad_agent->recv_handler)) {
1083 			ret = -EINVAL;
1084 			goto error;
1085 		}
1086 
1087 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1088 			if (mad_agent_priv->agent.rmpp_version) {
1089 				ret = -EINVAL;
1090 				goto error;
1091 			}
1092 		}
1093 
1094 		/*
1095 		 * Save pointer to next work request to post in case the
1096 		 * current one completes, and the user modifies the work
1097 		 * request associated with the completion
1098 		 */
1099 		next_send_buf = send_buf->next;
1100 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1101 
1102 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1103 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1104 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1105 						     mad_send_wr);
1106 			if (ret < 0)		/* error */
1107 				goto error;
1108 			else if (ret == 1)	/* locally consumed */
1109 				continue;
1110 		}
1111 
1112 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1113 		/* Timeout will be updated after send completes */
1114 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1115 		mad_send_wr->max_retries = send_buf->retries;
1116 		mad_send_wr->retries_left = send_buf->retries;
1117 		send_buf->retries = 0;
1118 		/* Reference for work request to QP + response */
1119 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1120 		mad_send_wr->status = IB_WC_SUCCESS;
1121 
1122 		/* Reference MAD agent until send completes */
1123 		atomic_inc(&mad_agent_priv->refcount);
1124 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1125 		list_add_tail(&mad_send_wr->agent_list,
1126 			      &mad_agent_priv->send_list);
1127 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1128 
1129 		if (mad_agent_priv->agent.rmpp_version) {
1130 			ret = ib_send_rmpp_mad(mad_send_wr);
1131 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1132 				ret = ib_send_mad(mad_send_wr);
1133 		} else
1134 			ret = ib_send_mad(mad_send_wr);
1135 		if (ret < 0) {
1136 			/* Fail send request */
1137 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1138 			list_del(&mad_send_wr->agent_list);
1139 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1140 			atomic_dec(&mad_agent_priv->refcount);
1141 			goto error;
1142 		}
1143 	}
1144 	return 0;
1145 error:
1146 	if (bad_send_buf)
1147 		*bad_send_buf = send_buf;
1148 	return ret;
1149 }
1150 EXPORT_SYMBOL(ib_post_send_mad);
1151 
1152 /*
1153  * ib_free_recv_mad - Returns data buffers used to receive
1154  *  a MAD to the access layer
1155  */
1156 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1157 {
1158 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1159 	struct ib_mad_private_header *mad_priv_hdr;
1160 	struct ib_mad_private *priv;
1161 	struct list_head free_list;
1162 
1163 	INIT_LIST_HEAD(&free_list);
1164 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1165 
1166 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1167 					&free_list, list) {
1168 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1169 					   recv_buf);
1170 		mad_priv_hdr = container_of(mad_recv_wc,
1171 					    struct ib_mad_private_header,
1172 					    recv_wc);
1173 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1174 				    header);
1175 		kmem_cache_free(ib_mad_cache, priv);
1176 	}
1177 }
1178 EXPORT_SYMBOL(ib_free_recv_mad);
1179 
1180 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1181 					u8 rmpp_version,
1182 					ib_mad_send_handler send_handler,
1183 					ib_mad_recv_handler recv_handler,
1184 					void *context)
1185 {
1186 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1187 }
1188 EXPORT_SYMBOL(ib_redirect_mad_qp);
1189 
1190 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1191 		      struct ib_wc *wc)
1192 {
1193 	printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1194 	return 0;
1195 }
1196 EXPORT_SYMBOL(ib_process_mad_wc);
1197 
1198 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1199 			 struct ib_mad_reg_req *mad_reg_req)
1200 {
1201 	int i;
1202 
1203 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1204 		if ((*method)->agent[i]) {
1205 			printk(KERN_ERR PFX "Method %d already in use\n", i);
1206 			return -EINVAL;
1207 		}
1208 	}
1209 	return 0;
1210 }
1211 
1212 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1213 {
1214 	/* Allocate management method table */
1215 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1216 	if (!*method) {
1217 		printk(KERN_ERR PFX "No memory for "
1218 		       "ib_mad_mgmt_method_table\n");
1219 		return -ENOMEM;
1220 	}
1221 
1222 	return 0;
1223 }
1224 
1225 /*
1226  * Check to see if there are any methods still in use
1227  */
1228 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1229 {
1230 	int i;
1231 
1232 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1233 		if (method->agent[i])
1234 			return 1;
1235 	return 0;
1236 }
1237 
1238 /*
1239  * Check to see if there are any method tables for this class still in use
1240  */
1241 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1242 {
1243 	int i;
1244 
1245 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1246 		if (class->method_table[i])
1247 			return 1;
1248 	return 0;
1249 }
1250 
1251 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1252 {
1253 	int i;
1254 
1255 	for (i = 0; i < MAX_MGMT_OUI; i++)
1256 		if (vendor_class->method_table[i])
1257 			return 1;
1258 	return 0;
1259 }
1260 
1261 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1262 			   char *oui)
1263 {
1264 	int i;
1265 
1266 	for (i = 0; i < MAX_MGMT_OUI; i++)
1267 		/* Is there matching OUI for this vendor class ? */
1268 		if (!memcmp(vendor_class->oui[i], oui, 3))
1269 			return i;
1270 
1271 	return -1;
1272 }
1273 
1274 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1275 {
1276 	int i;
1277 
1278 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1279 		if (vendor->vendor_class[i])
1280 			return 1;
1281 
1282 	return 0;
1283 }
1284 
1285 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1286 				     struct ib_mad_agent_private *agent)
1287 {
1288 	int i;
1289 
1290 	/* Remove any methods for this mad agent */
1291 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1292 		if (method->agent[i] == agent) {
1293 			method->agent[i] = NULL;
1294 		}
1295 	}
1296 }
1297 
1298 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1299 			      struct ib_mad_agent_private *agent_priv,
1300 			      u8 mgmt_class)
1301 {
1302 	struct ib_mad_port_private *port_priv;
1303 	struct ib_mad_mgmt_class_table **class;
1304 	struct ib_mad_mgmt_method_table **method;
1305 	int i, ret;
1306 
1307 	port_priv = agent_priv->qp_info->port_priv;
1308 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1309 	if (!*class) {
1310 		/* Allocate management class table for "new" class version */
1311 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1312 		if (!*class) {
1313 			printk(KERN_ERR PFX "No memory for "
1314 			       "ib_mad_mgmt_class_table\n");
1315 			ret = -ENOMEM;
1316 			goto error1;
1317 		}
1318 
1319 		/* Allocate method table for this management class */
1320 		method = &(*class)->method_table[mgmt_class];
1321 		if ((ret = allocate_method_table(method)))
1322 			goto error2;
1323 	} else {
1324 		method = &(*class)->method_table[mgmt_class];
1325 		if (!*method) {
1326 			/* Allocate method table for this management class */
1327 			if ((ret = allocate_method_table(method)))
1328 				goto error1;
1329 		}
1330 	}
1331 
1332 	/* Now, make sure methods are not already in use */
1333 	if (method_in_use(method, mad_reg_req))
1334 		goto error3;
1335 
1336 	/* Finally, add in methods being registered */
1337 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1338 		(*method)->agent[i] = agent_priv;
1339 
1340 	return 0;
1341 
1342 error3:
1343 	/* Remove any methods for this mad agent */
1344 	remove_methods_mad_agent(*method, agent_priv);
1345 	/* Now, check to see if there are any methods in use */
1346 	if (!check_method_table(*method)) {
1347 		/* If not, release management method table */
1348 		kfree(*method);
1349 		*method = NULL;
1350 	}
1351 	ret = -EINVAL;
1352 	goto error1;
1353 error2:
1354 	kfree(*class);
1355 	*class = NULL;
1356 error1:
1357 	return ret;
1358 }
1359 
1360 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1361 			   struct ib_mad_agent_private *agent_priv)
1362 {
1363 	struct ib_mad_port_private *port_priv;
1364 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1365 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1366 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1367 	struct ib_mad_mgmt_method_table **method;
1368 	int i, ret = -ENOMEM;
1369 	u8 vclass;
1370 
1371 	/* "New" vendor (with OUI) class */
1372 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1373 	port_priv = agent_priv->qp_info->port_priv;
1374 	vendor_table = &port_priv->version[
1375 				mad_reg_req->mgmt_class_version].vendor;
1376 	if (!*vendor_table) {
1377 		/* Allocate mgmt vendor class table for "new" class version */
1378 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1379 		if (!vendor) {
1380 			printk(KERN_ERR PFX "No memory for "
1381 			       "ib_mad_mgmt_vendor_class_table\n");
1382 			goto error1;
1383 		}
1384 
1385 		*vendor_table = vendor;
1386 	}
1387 	if (!(*vendor_table)->vendor_class[vclass]) {
1388 		/* Allocate table for this management vendor class */
1389 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1390 		if (!vendor_class) {
1391 			printk(KERN_ERR PFX "No memory for "
1392 			       "ib_mad_mgmt_vendor_class\n");
1393 			goto error2;
1394 		}
1395 
1396 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1397 	}
1398 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1399 		/* Is there matching OUI for this vendor class ? */
1400 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1401 			    mad_reg_req->oui, 3)) {
1402 			method = &(*vendor_table)->vendor_class[
1403 						vclass]->method_table[i];
1404 			BUG_ON(!*method);
1405 			goto check_in_use;
1406 		}
1407 	}
1408 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1409 		/* OUI slot available ? */
1410 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1411 				vclass]->oui[i])) {
1412 			method = &(*vendor_table)->vendor_class[
1413 				vclass]->method_table[i];
1414 			BUG_ON(*method);
1415 			/* Allocate method table for this OUI */
1416 			if ((ret = allocate_method_table(method)))
1417 				goto error3;
1418 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1419 			       mad_reg_req->oui, 3);
1420 			goto check_in_use;
1421 		}
1422 	}
1423 	printk(KERN_ERR PFX "All OUI slots in use\n");
1424 	goto error3;
1425 
1426 check_in_use:
1427 	/* Now, make sure methods are not already in use */
1428 	if (method_in_use(method, mad_reg_req))
1429 		goto error4;
1430 
1431 	/* Finally, add in methods being registered */
1432 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1433 		(*method)->agent[i] = agent_priv;
1434 
1435 	return 0;
1436 
1437 error4:
1438 	/* Remove any methods for this mad agent */
1439 	remove_methods_mad_agent(*method, agent_priv);
1440 	/* Now, check to see if there are any methods in use */
1441 	if (!check_method_table(*method)) {
1442 		/* If not, release management method table */
1443 		kfree(*method);
1444 		*method = NULL;
1445 	}
1446 	ret = -EINVAL;
1447 error3:
1448 	if (vendor_class) {
1449 		(*vendor_table)->vendor_class[vclass] = NULL;
1450 		kfree(vendor_class);
1451 	}
1452 error2:
1453 	if (vendor) {
1454 		*vendor_table = NULL;
1455 		kfree(vendor);
1456 	}
1457 error1:
1458 	return ret;
1459 }
1460 
1461 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1462 {
1463 	struct ib_mad_port_private *port_priv;
1464 	struct ib_mad_mgmt_class_table *class;
1465 	struct ib_mad_mgmt_method_table *method;
1466 	struct ib_mad_mgmt_vendor_class_table *vendor;
1467 	struct ib_mad_mgmt_vendor_class *vendor_class;
1468 	int index;
1469 	u8 mgmt_class;
1470 
1471 	/*
1472 	 * Was MAD registration request supplied
1473 	 * with original registration ?
1474 	 */
1475 	if (!agent_priv->reg_req) {
1476 		goto out;
1477 	}
1478 
1479 	port_priv = agent_priv->qp_info->port_priv;
1480 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1481 	class = port_priv->version[
1482 			agent_priv->reg_req->mgmt_class_version].class;
1483 	if (!class)
1484 		goto vendor_check;
1485 
1486 	method = class->method_table[mgmt_class];
1487 	if (method) {
1488 		/* Remove any methods for this mad agent */
1489 		remove_methods_mad_agent(method, agent_priv);
1490 		/* Now, check to see if there are any methods still in use */
1491 		if (!check_method_table(method)) {
1492 			/* If not, release management method table */
1493 			 kfree(method);
1494 			 class->method_table[mgmt_class] = NULL;
1495 			 /* Any management classes left ? */
1496 			if (!check_class_table(class)) {
1497 				/* If not, release management class table */
1498 				kfree(class);
1499 				port_priv->version[
1500 					agent_priv->reg_req->
1501 					mgmt_class_version].class = NULL;
1502 			}
1503 		}
1504 	}
1505 
1506 vendor_check:
1507 	if (!is_vendor_class(mgmt_class))
1508 		goto out;
1509 
1510 	/* normalize mgmt_class to vendor range 2 */
1511 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1512 	vendor = port_priv->version[
1513 			agent_priv->reg_req->mgmt_class_version].vendor;
1514 
1515 	if (!vendor)
1516 		goto out;
1517 
1518 	vendor_class = vendor->vendor_class[mgmt_class];
1519 	if (vendor_class) {
1520 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1521 		if (index < 0)
1522 			goto out;
1523 		method = vendor_class->method_table[index];
1524 		if (method) {
1525 			/* Remove any methods for this mad agent */
1526 			remove_methods_mad_agent(method, agent_priv);
1527 			/*
1528 			 * Now, check to see if there are
1529 			 * any methods still in use
1530 			 */
1531 			if (!check_method_table(method)) {
1532 				/* If not, release management method table */
1533 				kfree(method);
1534 				vendor_class->method_table[index] = NULL;
1535 				memset(vendor_class->oui[index], 0, 3);
1536 				/* Any OUIs left ? */
1537 				if (!check_vendor_class(vendor_class)) {
1538 					/* If not, release vendor class table */
1539 					kfree(vendor_class);
1540 					vendor->vendor_class[mgmt_class] = NULL;
1541 					/* Any other vendor classes left ? */
1542 					if (!check_vendor_table(vendor)) {
1543 						kfree(vendor);
1544 						port_priv->version[
1545 							agent_priv->reg_req->
1546 							mgmt_class_version].
1547 							vendor = NULL;
1548 					}
1549 				}
1550 			}
1551 		}
1552 	}
1553 
1554 out:
1555 	return;
1556 }
1557 
1558 static struct ib_mad_agent_private *
1559 find_mad_agent(struct ib_mad_port_private *port_priv,
1560 	       struct ib_mad *mad)
1561 {
1562 	struct ib_mad_agent_private *mad_agent = NULL;
1563 	unsigned long flags;
1564 
1565 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1566 	if (ib_response_mad(mad)) {
1567 		u32 hi_tid;
1568 		struct ib_mad_agent_private *entry;
1569 
1570 		/*
1571 		 * Routing is based on high 32 bits of transaction ID
1572 		 * of MAD.
1573 		 */
1574 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1575 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1576 			if (entry->agent.hi_tid == hi_tid) {
1577 				mad_agent = entry;
1578 				break;
1579 			}
1580 		}
1581 	} else {
1582 		struct ib_mad_mgmt_class_table *class;
1583 		struct ib_mad_mgmt_method_table *method;
1584 		struct ib_mad_mgmt_vendor_class_table *vendor;
1585 		struct ib_mad_mgmt_vendor_class *vendor_class;
1586 		struct ib_vendor_mad *vendor_mad;
1587 		int index;
1588 
1589 		/*
1590 		 * Routing is based on version, class, and method
1591 		 * For "newer" vendor MADs, also based on OUI
1592 		 */
1593 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1594 			goto out;
1595 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1596 			class = port_priv->version[
1597 					mad->mad_hdr.class_version].class;
1598 			if (!class)
1599 				goto out;
1600 			if (convert_mgmt_class(mad->mad_hdr.mgmt_class) >=
1601 			    IB_MGMT_MAX_METHODS)
1602 				goto out;
1603 			method = class->method_table[convert_mgmt_class(
1604 							mad->mad_hdr.mgmt_class)];
1605 			if (method)
1606 				mad_agent = method->agent[mad->mad_hdr.method &
1607 							  ~IB_MGMT_METHOD_RESP];
1608 		} else {
1609 			vendor = port_priv->version[
1610 					mad->mad_hdr.class_version].vendor;
1611 			if (!vendor)
1612 				goto out;
1613 			vendor_class = vendor->vendor_class[vendor_class_index(
1614 						mad->mad_hdr.mgmt_class)];
1615 			if (!vendor_class)
1616 				goto out;
1617 			/* Find matching OUI */
1618 			vendor_mad = (struct ib_vendor_mad *)mad;
1619 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1620 			if (index == -1)
1621 				goto out;
1622 			method = vendor_class->method_table[index];
1623 			if (method) {
1624 				mad_agent = method->agent[mad->mad_hdr.method &
1625 							  ~IB_MGMT_METHOD_RESP];
1626 			}
1627 		}
1628 	}
1629 
1630 	if (mad_agent) {
1631 		if (mad_agent->agent.recv_handler)
1632 			atomic_inc(&mad_agent->refcount);
1633 		else {
1634 			printk(KERN_NOTICE PFX "No receive handler for client "
1635 			       "%p on port %d\n",
1636 			       &mad_agent->agent, port_priv->port_num);
1637 			mad_agent = NULL;
1638 		}
1639 	}
1640 out:
1641 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1642 
1643 	return mad_agent;
1644 }
1645 
1646 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1647 {
1648 	int valid = 0;
1649 
1650 	/* Make sure MAD base version is understood */
1651 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1652 		printk(KERN_ERR PFX "MAD received with unsupported base "
1653 		       "version %d\n", mad->mad_hdr.base_version);
1654 		goto out;
1655 	}
1656 
1657 	/* Filter SMI packets sent to other than QP0 */
1658 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1659 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1660 		if (qp_num == 0)
1661 			valid = 1;
1662 	} else {
1663 		/* Filter GSI packets sent to QP0 */
1664 		if (qp_num != 0)
1665 			valid = 1;
1666 	}
1667 
1668 out:
1669 	return valid;
1670 }
1671 
1672 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1673 		       struct ib_mad_hdr *mad_hdr)
1674 {
1675 	struct ib_rmpp_mad *rmpp_mad;
1676 
1677 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1678 	return !mad_agent_priv->agent.rmpp_version ||
1679 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1680 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1681 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1682 }
1683 
1684 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1685 				     struct ib_mad_recv_wc *rwc)
1686 {
1687 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1688 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1689 }
1690 
1691 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1692 				   struct ib_mad_send_wr_private *wr,
1693 				   struct ib_mad_recv_wc *rwc )
1694 {
1695 	struct ib_ah_attr attr;
1696 	u8 send_resp, rcv_resp;
1697 	union ib_gid sgid;
1698 	struct ib_device *device = mad_agent_priv->agent.device;
1699 	u8 port_num = mad_agent_priv->agent.port_num;
1700 	u8 lmc;
1701 
1702 	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1703 	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1704 
1705 	if (send_resp == rcv_resp)
1706 		/* both requests, or both responses. GIDs different */
1707 		return 0;
1708 
1709 	if (ib_query_ah(wr->send_buf.ah, &attr))
1710 		/* Assume not equal, to avoid false positives. */
1711 		return 0;
1712 
1713 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1714 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1715 		/* one has GID, other does not.  Assume different */
1716 		return 0;
1717 
1718 	if (!send_resp && rcv_resp) {
1719 		/* is request/response. */
1720 		if (!(attr.ah_flags & IB_AH_GRH)) {
1721 			if (ib_get_cached_lmc(device, port_num, &lmc))
1722 				return 0;
1723 			return (!lmc || !((attr.src_path_bits ^
1724 					   rwc->wc->dlid_path_bits) &
1725 					  ((1 << lmc) - 1)));
1726 		} else {
1727 			if (ib_get_cached_gid(device, port_num,
1728 					      attr.grh.sgid_index, &sgid))
1729 				return 0;
1730 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1731 				       16);
1732 		}
1733 	}
1734 
1735 	if (!(attr.ah_flags & IB_AH_GRH))
1736 		return attr.dlid == rwc->wc->slid;
1737 	else
1738 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1739 			       16);
1740 }
1741 
1742 static inline int is_direct(u8 class)
1743 {
1744 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1745 }
1746 
1747 struct ib_mad_send_wr_private*
1748 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1749 		 struct ib_mad_recv_wc *wc)
1750 {
1751 	struct ib_mad_send_wr_private *wr;
1752 	struct ib_mad *mad;
1753 
1754 	mad = (struct ib_mad *)wc->recv_buf.mad;
1755 
1756 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1757 		if ((wr->tid == mad->mad_hdr.tid) &&
1758 		    rcv_has_same_class(wr, wc) &&
1759 		    /*
1760 		     * Don't check GID for direct routed MADs.
1761 		     * These might have permissive LIDs.
1762 		     */
1763 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1764 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1765 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1766 	}
1767 
1768 	/*
1769 	 * It's possible to receive the response before we've
1770 	 * been notified that the send has completed
1771 	 */
1772 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1773 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1774 		    wr->tid == mad->mad_hdr.tid &&
1775 		    wr->timeout &&
1776 		    rcv_has_same_class(wr, wc) &&
1777 		    /*
1778 		     * Don't check GID for direct routed MADs.
1779 		     * These might have permissive LIDs.
1780 		     */
1781 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1782 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1783 			/* Verify request has not been canceled */
1784 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1785 	}
1786 	return NULL;
1787 }
1788 
1789 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1790 {
1791 	mad_send_wr->timeout = 0;
1792 	if (mad_send_wr->refcount == 1)
1793 		list_move_tail(&mad_send_wr->agent_list,
1794 			      &mad_send_wr->mad_agent_priv->done_list);
1795 }
1796 
1797 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1798 				 struct ib_mad_recv_wc *mad_recv_wc)
1799 {
1800 	struct ib_mad_send_wr_private *mad_send_wr;
1801 	struct ib_mad_send_wc mad_send_wc;
1802 	unsigned long flags;
1803 
1804 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1805 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1806 	if (mad_agent_priv->agent.rmpp_version) {
1807 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1808 						      mad_recv_wc);
1809 		if (!mad_recv_wc) {
1810 			deref_mad_agent(mad_agent_priv);
1811 			return;
1812 		}
1813 	}
1814 
1815 	/* Complete corresponding request */
1816 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1817 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1818 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1819 		if (!mad_send_wr) {
1820 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1821 			ib_free_recv_mad(mad_recv_wc);
1822 			deref_mad_agent(mad_agent_priv);
1823 			return;
1824 		}
1825 		ib_mark_mad_done(mad_send_wr);
1826 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1827 
1828 		/* Defined behavior is to complete response before request */
1829 		mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1830 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1831 						   mad_recv_wc);
1832 		atomic_dec(&mad_agent_priv->refcount);
1833 
1834 		mad_send_wc.status = IB_WC_SUCCESS;
1835 		mad_send_wc.vendor_err = 0;
1836 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
1837 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1838 	} else {
1839 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1840 						   mad_recv_wc);
1841 		deref_mad_agent(mad_agent_priv);
1842 	}
1843 }
1844 
1845 static bool generate_unmatched_resp(struct ib_mad_private *recv,
1846 				    struct ib_mad_private *response)
1847 {
1848 	if (recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_GET ||
1849 	    recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_SET) {
1850 		memcpy(response, recv, sizeof *response);
1851 		response->header.recv_wc.wc = &response->header.wc;
1852 		response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1853 		response->header.recv_wc.recv_buf.grh = &response->grh;
1854 		response->mad.mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
1855 		response->mad.mad.mad_hdr.status =
1856 			cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
1857 		if (recv->mad.mad.mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1858 			response->mad.mad.mad_hdr.status |= IB_SMP_DIRECTION;
1859 
1860 		return true;
1861 	} else {
1862 		return false;
1863 	}
1864 }
1865 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1866 				     struct ib_wc *wc)
1867 {
1868 	struct ib_mad_qp_info *qp_info;
1869 	struct ib_mad_private_header *mad_priv_hdr;
1870 	struct ib_mad_private *recv, *response = NULL;
1871 	struct ib_mad_list_head *mad_list;
1872 	struct ib_mad_agent_private *mad_agent;
1873 	int port_num;
1874 	int ret = IB_MAD_RESULT_SUCCESS;
1875 
1876 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1877 	qp_info = mad_list->mad_queue->qp_info;
1878 	dequeue_mad(mad_list);
1879 
1880 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1881 				    mad_list);
1882 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1883 	ib_dma_unmap_single(port_priv->device,
1884 			    recv->header.mapping,
1885 			    sizeof(struct ib_mad_private) -
1886 			      sizeof(struct ib_mad_private_header),
1887 			    DMA_FROM_DEVICE);
1888 
1889 	/* Setup MAD receive work completion from "normal" work completion */
1890 	recv->header.wc = *wc;
1891 	recv->header.recv_wc.wc = &recv->header.wc;
1892 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1893 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1894 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1895 
1896 	if (atomic_read(&qp_info->snoop_count))
1897 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1898 
1899 	/* Validate MAD */
1900 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1901 		goto out;
1902 
1903 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1904 	if (!response) {
1905 		printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1906 		       "for response buffer\n");
1907 		goto out;
1908 	}
1909 
1910 	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1911 		port_num = wc->port_num;
1912 	else
1913 		port_num = port_priv->port_num;
1914 
1915 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1916 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1917 		enum smi_forward_action retsmi;
1918 
1919 		if (smi_handle_dr_smp_recv(&recv->mad.smp,
1920 					   port_priv->device->node_type,
1921 					   port_num,
1922 					   port_priv->device->phys_port_cnt) ==
1923 					   IB_SMI_DISCARD)
1924 			goto out;
1925 
1926 		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1927 		if (retsmi == IB_SMI_LOCAL)
1928 			goto local;
1929 
1930 		if (retsmi == IB_SMI_SEND) { /* don't forward */
1931 			if (smi_handle_dr_smp_send(&recv->mad.smp,
1932 						   port_priv->device->node_type,
1933 						   port_num) == IB_SMI_DISCARD)
1934 				goto out;
1935 
1936 			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1937 				goto out;
1938 		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1939 			/* forward case for switches */
1940 			memcpy(response, recv, sizeof(*response));
1941 			response->header.recv_wc.wc = &response->header.wc;
1942 			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1943 			response->header.recv_wc.recv_buf.grh = &response->grh;
1944 
1945 			agent_send_response(&response->mad.mad,
1946 					    &response->grh, wc,
1947 					    port_priv->device,
1948 					    smi_get_fwd_port(&recv->mad.smp),
1949 					    qp_info->qp->qp_num);
1950 
1951 			goto out;
1952 		}
1953 	}
1954 
1955 local:
1956 	/* Give driver "right of first refusal" on incoming MAD */
1957 	if (port_priv->device->process_mad) {
1958 		ret = port_priv->device->process_mad(port_priv->device, 0,
1959 						     port_priv->port_num,
1960 						     wc, &recv->grh,
1961 						     &recv->mad.mad,
1962 						     &response->mad.mad);
1963 		if (ret & IB_MAD_RESULT_SUCCESS) {
1964 			if (ret & IB_MAD_RESULT_CONSUMED)
1965 				goto out;
1966 			if (ret & IB_MAD_RESULT_REPLY) {
1967 				agent_send_response(&response->mad.mad,
1968 						    &recv->grh, wc,
1969 						    port_priv->device,
1970 						    port_num,
1971 						    qp_info->qp->qp_num);
1972 				goto out;
1973 			}
1974 		}
1975 	}
1976 
1977 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1978 	if (mad_agent) {
1979 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1980 		/*
1981 		 * recv is freed up in error cases in ib_mad_complete_recv
1982 		 * or via recv_handler in ib_mad_complete_recv()
1983 		 */
1984 		recv = NULL;
1985 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
1986 		   generate_unmatched_resp(recv, response)) {
1987 		agent_send_response(&response->mad.mad, &recv->grh, wc,
1988 				    port_priv->device, port_num, qp_info->qp->qp_num);
1989 	}
1990 
1991 out:
1992 	/* Post another receive request for this QP */
1993 	if (response) {
1994 		ib_mad_post_receive_mads(qp_info, response);
1995 		if (recv)
1996 			kmem_cache_free(ib_mad_cache, recv);
1997 	} else
1998 		ib_mad_post_receive_mads(qp_info, recv);
1999 }
2000 
2001 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2002 {
2003 	struct ib_mad_send_wr_private *mad_send_wr;
2004 	unsigned long delay;
2005 
2006 	if (list_empty(&mad_agent_priv->wait_list)) {
2007 		cancel_delayed_work(&mad_agent_priv->timed_work);
2008 	} else {
2009 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2010 					 struct ib_mad_send_wr_private,
2011 					 agent_list);
2012 
2013 		if (time_after(mad_agent_priv->timeout,
2014 			       mad_send_wr->timeout)) {
2015 			mad_agent_priv->timeout = mad_send_wr->timeout;
2016 			delay = mad_send_wr->timeout - jiffies;
2017 			if ((long)delay <= 0)
2018 				delay = 1;
2019 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2020 					 &mad_agent_priv->timed_work, delay);
2021 		}
2022 	}
2023 }
2024 
2025 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2026 {
2027 	struct ib_mad_agent_private *mad_agent_priv;
2028 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2029 	struct list_head *list_item;
2030 	unsigned long delay;
2031 
2032 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2033 	list_del(&mad_send_wr->agent_list);
2034 
2035 	delay = mad_send_wr->timeout;
2036 	mad_send_wr->timeout += jiffies;
2037 
2038 	if (delay) {
2039 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2040 			temp_mad_send_wr = list_entry(list_item,
2041 						struct ib_mad_send_wr_private,
2042 						agent_list);
2043 			if (time_after(mad_send_wr->timeout,
2044 				       temp_mad_send_wr->timeout))
2045 				break;
2046 		}
2047 	}
2048 	else
2049 		list_item = &mad_agent_priv->wait_list;
2050 	list_add(&mad_send_wr->agent_list, list_item);
2051 
2052 	/* Reschedule a work item if we have a shorter timeout */
2053 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2054 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2055 				 &mad_agent_priv->timed_work, delay);
2056 }
2057 
2058 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2059 			  int timeout_ms)
2060 {
2061 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2062 	wait_for_response(mad_send_wr);
2063 }
2064 
2065 /*
2066  * Process a send work completion
2067  */
2068 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2069 			     struct ib_mad_send_wc *mad_send_wc)
2070 {
2071 	struct ib_mad_agent_private	*mad_agent_priv;
2072 	unsigned long			flags;
2073 	int				ret;
2074 
2075 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2076 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2077 	if (mad_agent_priv->agent.rmpp_version) {
2078 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2079 		if (ret == IB_RMPP_RESULT_CONSUMED)
2080 			goto done;
2081 	} else
2082 		ret = IB_RMPP_RESULT_UNHANDLED;
2083 
2084 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2085 	    mad_send_wr->status == IB_WC_SUCCESS) {
2086 		mad_send_wr->status = mad_send_wc->status;
2087 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2088 	}
2089 
2090 	if (--mad_send_wr->refcount > 0) {
2091 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2092 		    mad_send_wr->status == IB_WC_SUCCESS) {
2093 			wait_for_response(mad_send_wr);
2094 		}
2095 		goto done;
2096 	}
2097 
2098 	/* Remove send from MAD agent and notify client of completion */
2099 	list_del(&mad_send_wr->agent_list);
2100 	adjust_timeout(mad_agent_priv);
2101 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2102 
2103 	if (mad_send_wr->status != IB_WC_SUCCESS )
2104 		mad_send_wc->status = mad_send_wr->status;
2105 	if (ret == IB_RMPP_RESULT_INTERNAL)
2106 		ib_rmpp_send_handler(mad_send_wc);
2107 	else
2108 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2109 						   mad_send_wc);
2110 
2111 	/* Release reference on agent taken when sending */
2112 	deref_mad_agent(mad_agent_priv);
2113 	return;
2114 done:
2115 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2116 }
2117 
2118 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2119 				     struct ib_wc *wc)
2120 {
2121 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2122 	struct ib_mad_list_head		*mad_list;
2123 	struct ib_mad_qp_info		*qp_info;
2124 	struct ib_mad_queue		*send_queue;
2125 	struct ib_send_wr		*bad_send_wr;
2126 	struct ib_mad_send_wc		mad_send_wc;
2127 	unsigned long flags;
2128 	int ret;
2129 
2130 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2131 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2132 				   mad_list);
2133 	send_queue = mad_list->mad_queue;
2134 	qp_info = send_queue->qp_info;
2135 
2136 retry:
2137 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2138 			    mad_send_wr->header_mapping,
2139 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2140 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2141 			    mad_send_wr->payload_mapping,
2142 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2143 	queued_send_wr = NULL;
2144 	spin_lock_irqsave(&send_queue->lock, flags);
2145 	list_del(&mad_list->list);
2146 
2147 	/* Move queued send to the send queue */
2148 	if (send_queue->count-- > send_queue->max_active) {
2149 		mad_list = container_of(qp_info->overflow_list.next,
2150 					struct ib_mad_list_head, list);
2151 		queued_send_wr = container_of(mad_list,
2152 					struct ib_mad_send_wr_private,
2153 					mad_list);
2154 		list_move_tail(&mad_list->list, &send_queue->list);
2155 	}
2156 	spin_unlock_irqrestore(&send_queue->lock, flags);
2157 
2158 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2159 	mad_send_wc.status = wc->status;
2160 	mad_send_wc.vendor_err = wc->vendor_err;
2161 	if (atomic_read(&qp_info->snoop_count))
2162 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2163 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2164 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2165 
2166 	if (queued_send_wr) {
2167 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2168 				   &bad_send_wr);
2169 		if (ret) {
2170 			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2171 			mad_send_wr = queued_send_wr;
2172 			wc->status = IB_WC_LOC_QP_OP_ERR;
2173 			goto retry;
2174 		}
2175 	}
2176 }
2177 
2178 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2179 {
2180 	struct ib_mad_send_wr_private *mad_send_wr;
2181 	struct ib_mad_list_head *mad_list;
2182 	unsigned long flags;
2183 
2184 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2185 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2186 		mad_send_wr = container_of(mad_list,
2187 					   struct ib_mad_send_wr_private,
2188 					   mad_list);
2189 		mad_send_wr->retry = 1;
2190 	}
2191 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2192 }
2193 
2194 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2195 			      struct ib_wc *wc)
2196 {
2197 	struct ib_mad_list_head *mad_list;
2198 	struct ib_mad_qp_info *qp_info;
2199 	struct ib_mad_send_wr_private *mad_send_wr;
2200 	int ret;
2201 
2202 	/* Determine if failure was a send or receive */
2203 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2204 	qp_info = mad_list->mad_queue->qp_info;
2205 	if (mad_list->mad_queue == &qp_info->recv_queue)
2206 		/*
2207 		 * Receive errors indicate that the QP has entered the error
2208 		 * state - error handling/shutdown code will cleanup
2209 		 */
2210 		return;
2211 
2212 	/*
2213 	 * Send errors will transition the QP to SQE - move
2214 	 * QP to RTS and repost flushed work requests
2215 	 */
2216 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2217 				   mad_list);
2218 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2219 		if (mad_send_wr->retry) {
2220 			/* Repost send */
2221 			struct ib_send_wr *bad_send_wr;
2222 
2223 			mad_send_wr->retry = 0;
2224 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2225 					&bad_send_wr);
2226 			if (ret)
2227 				ib_mad_send_done_handler(port_priv, wc);
2228 		} else
2229 			ib_mad_send_done_handler(port_priv, wc);
2230 	} else {
2231 		struct ib_qp_attr *attr;
2232 
2233 		/* Transition QP to RTS and fail offending send */
2234 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2235 		if (attr) {
2236 			attr->qp_state = IB_QPS_RTS;
2237 			attr->cur_qp_state = IB_QPS_SQE;
2238 			ret = ib_modify_qp(qp_info->qp, attr,
2239 					   IB_QP_STATE | IB_QP_CUR_STATE);
2240 			kfree(attr);
2241 			if (ret)
2242 				printk(KERN_ERR PFX "mad_error_handler - "
2243 				       "ib_modify_qp to RTS : %d\n", ret);
2244 			else
2245 				mark_sends_for_retry(qp_info);
2246 		}
2247 		ib_mad_send_done_handler(port_priv, wc);
2248 	}
2249 }
2250 
2251 /*
2252  * IB MAD completion callback
2253  */
2254 static void ib_mad_completion_handler(struct work_struct *work)
2255 {
2256 	struct ib_mad_port_private *port_priv;
2257 	struct ib_wc wc;
2258 
2259 	port_priv = container_of(work, struct ib_mad_port_private, work);
2260 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2261 
2262 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2263 		if (wc.status == IB_WC_SUCCESS) {
2264 			switch (wc.opcode) {
2265 			case IB_WC_SEND:
2266 				ib_mad_send_done_handler(port_priv, &wc);
2267 				break;
2268 			case IB_WC_RECV:
2269 				ib_mad_recv_done_handler(port_priv, &wc);
2270 				break;
2271 			default:
2272 				BUG_ON(1);
2273 				break;
2274 			}
2275 		} else
2276 			mad_error_handler(port_priv, &wc);
2277 	}
2278 }
2279 
2280 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2281 {
2282 	unsigned long flags;
2283 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2284 	struct ib_mad_send_wc mad_send_wc;
2285 	struct list_head cancel_list;
2286 
2287 	INIT_LIST_HEAD(&cancel_list);
2288 
2289 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2290 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2291 				 &mad_agent_priv->send_list, agent_list) {
2292 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2293 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2294 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2295 		}
2296 	}
2297 
2298 	/* Empty wait list to prevent receives from finding a request */
2299 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2300 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2301 
2302 	/* Report all cancelled requests */
2303 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2304 	mad_send_wc.vendor_err = 0;
2305 
2306 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2307 				 &cancel_list, agent_list) {
2308 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2309 		list_del(&mad_send_wr->agent_list);
2310 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2311 						   &mad_send_wc);
2312 		atomic_dec(&mad_agent_priv->refcount);
2313 	}
2314 }
2315 
2316 static struct ib_mad_send_wr_private*
2317 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2318 	     struct ib_mad_send_buf *send_buf)
2319 {
2320 	struct ib_mad_send_wr_private *mad_send_wr;
2321 
2322 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2323 			    agent_list) {
2324 		if (&mad_send_wr->send_buf == send_buf)
2325 			return mad_send_wr;
2326 	}
2327 
2328 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2329 			    agent_list) {
2330 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2331 		    &mad_send_wr->send_buf == send_buf)
2332 			return mad_send_wr;
2333 	}
2334 	return NULL;
2335 }
2336 
2337 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2338 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2339 {
2340 	struct ib_mad_agent_private *mad_agent_priv;
2341 	struct ib_mad_send_wr_private *mad_send_wr;
2342 	unsigned long flags;
2343 	int active;
2344 
2345 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2346 				      agent);
2347 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2348 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2349 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2350 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2351 		return -EINVAL;
2352 	}
2353 
2354 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2355 	if (!timeout_ms) {
2356 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2357 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2358 	}
2359 
2360 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2361 	if (active)
2362 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2363 	else
2364 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2365 
2366 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2367 	return 0;
2368 }
2369 EXPORT_SYMBOL(ib_modify_mad);
2370 
2371 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2372 		   struct ib_mad_send_buf *send_buf)
2373 {
2374 	ib_modify_mad(mad_agent, send_buf, 0);
2375 }
2376 EXPORT_SYMBOL(ib_cancel_mad);
2377 
2378 static void local_completions(struct work_struct *work)
2379 {
2380 	struct ib_mad_agent_private *mad_agent_priv;
2381 	struct ib_mad_local_private *local;
2382 	struct ib_mad_agent_private *recv_mad_agent;
2383 	unsigned long flags;
2384 	int free_mad;
2385 	struct ib_wc wc;
2386 	struct ib_mad_send_wc mad_send_wc;
2387 
2388 	mad_agent_priv =
2389 		container_of(work, struct ib_mad_agent_private, local_work);
2390 
2391 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2392 	while (!list_empty(&mad_agent_priv->local_list)) {
2393 		local = list_entry(mad_agent_priv->local_list.next,
2394 				   struct ib_mad_local_private,
2395 				   completion_list);
2396 		list_del(&local->completion_list);
2397 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2398 		free_mad = 0;
2399 		if (local->mad_priv) {
2400 			recv_mad_agent = local->recv_mad_agent;
2401 			if (!recv_mad_agent) {
2402 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2403 				free_mad = 1;
2404 				goto local_send_completion;
2405 			}
2406 
2407 			/*
2408 			 * Defined behavior is to complete response
2409 			 * before request
2410 			 */
2411 			build_smp_wc(recv_mad_agent->agent.qp,
2412 				     (unsigned long) local->mad_send_wr,
2413 				     be16_to_cpu(IB_LID_PERMISSIVE),
2414 				     0, recv_mad_agent->agent.port_num, &wc);
2415 
2416 			local->mad_priv->header.recv_wc.wc = &wc;
2417 			local->mad_priv->header.recv_wc.mad_len =
2418 						sizeof(struct ib_mad);
2419 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2420 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2421 				 &local->mad_priv->header.recv_wc.rmpp_list);
2422 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2423 			local->mad_priv->header.recv_wc.recv_buf.mad =
2424 						&local->mad_priv->mad.mad;
2425 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2426 				snoop_recv(recv_mad_agent->qp_info,
2427 					  &local->mad_priv->header.recv_wc,
2428 					   IB_MAD_SNOOP_RECVS);
2429 			recv_mad_agent->agent.recv_handler(
2430 						&recv_mad_agent->agent,
2431 						&local->mad_priv->header.recv_wc);
2432 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2433 			atomic_dec(&recv_mad_agent->refcount);
2434 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2435 		}
2436 
2437 local_send_completion:
2438 		/* Complete send */
2439 		mad_send_wc.status = IB_WC_SUCCESS;
2440 		mad_send_wc.vendor_err = 0;
2441 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2442 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2443 			snoop_send(mad_agent_priv->qp_info,
2444 				   &local->mad_send_wr->send_buf,
2445 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2446 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2447 						   &mad_send_wc);
2448 
2449 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2450 		atomic_dec(&mad_agent_priv->refcount);
2451 		if (free_mad)
2452 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2453 		kfree(local);
2454 	}
2455 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2456 }
2457 
2458 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2459 {
2460 	int ret;
2461 
2462 	if (!mad_send_wr->retries_left)
2463 		return -ETIMEDOUT;
2464 
2465 	mad_send_wr->retries_left--;
2466 	mad_send_wr->send_buf.retries++;
2467 
2468 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2469 
2470 	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2471 		ret = ib_retry_rmpp(mad_send_wr);
2472 		switch (ret) {
2473 		case IB_RMPP_RESULT_UNHANDLED:
2474 			ret = ib_send_mad(mad_send_wr);
2475 			break;
2476 		case IB_RMPP_RESULT_CONSUMED:
2477 			ret = 0;
2478 			break;
2479 		default:
2480 			ret = -ECOMM;
2481 			break;
2482 		}
2483 	} else
2484 		ret = ib_send_mad(mad_send_wr);
2485 
2486 	if (!ret) {
2487 		mad_send_wr->refcount++;
2488 		list_add_tail(&mad_send_wr->agent_list,
2489 			      &mad_send_wr->mad_agent_priv->send_list);
2490 	}
2491 	return ret;
2492 }
2493 
2494 static void timeout_sends(struct work_struct *work)
2495 {
2496 	struct ib_mad_agent_private *mad_agent_priv;
2497 	struct ib_mad_send_wr_private *mad_send_wr;
2498 	struct ib_mad_send_wc mad_send_wc;
2499 	unsigned long flags, delay;
2500 
2501 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2502 				      timed_work.work);
2503 	mad_send_wc.vendor_err = 0;
2504 
2505 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2506 	while (!list_empty(&mad_agent_priv->wait_list)) {
2507 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2508 					 struct ib_mad_send_wr_private,
2509 					 agent_list);
2510 
2511 		if (time_after(mad_send_wr->timeout, jiffies)) {
2512 			delay = mad_send_wr->timeout - jiffies;
2513 			if ((long)delay <= 0)
2514 				delay = 1;
2515 			queue_delayed_work(mad_agent_priv->qp_info->
2516 					   port_priv->wq,
2517 					   &mad_agent_priv->timed_work, delay);
2518 			break;
2519 		}
2520 
2521 		list_del(&mad_send_wr->agent_list);
2522 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2523 		    !retry_send(mad_send_wr))
2524 			continue;
2525 
2526 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2527 
2528 		if (mad_send_wr->status == IB_WC_SUCCESS)
2529 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2530 		else
2531 			mad_send_wc.status = mad_send_wr->status;
2532 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2533 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2534 						   &mad_send_wc);
2535 
2536 		atomic_dec(&mad_agent_priv->refcount);
2537 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2538 	}
2539 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2540 }
2541 
2542 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2543 {
2544 	struct ib_mad_port_private *port_priv = cq->cq_context;
2545 	unsigned long flags;
2546 
2547 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2548 	if (!list_empty(&port_priv->port_list))
2549 		queue_work(port_priv->wq, &port_priv->work);
2550 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2551 }
2552 
2553 /*
2554  * Allocate receive MADs and post receive WRs for them
2555  */
2556 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2557 				    struct ib_mad_private *mad)
2558 {
2559 	unsigned long flags;
2560 	int post, ret;
2561 	struct ib_mad_private *mad_priv;
2562 	struct ib_sge sg_list;
2563 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2564 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2565 
2566 	/* Initialize common scatter list fields */
2567 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2568 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2569 
2570 	/* Initialize common receive WR fields */
2571 	recv_wr.next = NULL;
2572 	recv_wr.sg_list = &sg_list;
2573 	recv_wr.num_sge = 1;
2574 
2575 	do {
2576 		/* Allocate and map receive buffer */
2577 		if (mad) {
2578 			mad_priv = mad;
2579 			mad = NULL;
2580 		} else {
2581 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2582 			if (!mad_priv) {
2583 				printk(KERN_ERR PFX "No memory for receive buffer\n");
2584 				ret = -ENOMEM;
2585 				break;
2586 			}
2587 		}
2588 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2589 						 &mad_priv->grh,
2590 						 sizeof *mad_priv -
2591 						   sizeof mad_priv->header,
2592 						 DMA_FROM_DEVICE);
2593 		mad_priv->header.mapping = sg_list.addr;
2594 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2595 		mad_priv->header.mad_list.mad_queue = recv_queue;
2596 
2597 		/* Post receive WR */
2598 		spin_lock_irqsave(&recv_queue->lock, flags);
2599 		post = (++recv_queue->count < recv_queue->max_active);
2600 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2601 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2602 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2603 		if (ret) {
2604 			spin_lock_irqsave(&recv_queue->lock, flags);
2605 			list_del(&mad_priv->header.mad_list.list);
2606 			recv_queue->count--;
2607 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2608 			ib_dma_unmap_single(qp_info->port_priv->device,
2609 					    mad_priv->header.mapping,
2610 					    sizeof *mad_priv -
2611 					      sizeof mad_priv->header,
2612 					    DMA_FROM_DEVICE);
2613 			kmem_cache_free(ib_mad_cache, mad_priv);
2614 			printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2615 			break;
2616 		}
2617 	} while (post);
2618 
2619 	return ret;
2620 }
2621 
2622 /*
2623  * Return all the posted receive MADs
2624  */
2625 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2626 {
2627 	struct ib_mad_private_header *mad_priv_hdr;
2628 	struct ib_mad_private *recv;
2629 	struct ib_mad_list_head *mad_list;
2630 
2631 	if (!qp_info->qp)
2632 		return;
2633 
2634 	while (!list_empty(&qp_info->recv_queue.list)) {
2635 
2636 		mad_list = list_entry(qp_info->recv_queue.list.next,
2637 				      struct ib_mad_list_head, list);
2638 		mad_priv_hdr = container_of(mad_list,
2639 					    struct ib_mad_private_header,
2640 					    mad_list);
2641 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2642 				    header);
2643 
2644 		/* Remove from posted receive MAD list */
2645 		list_del(&mad_list->list);
2646 
2647 		ib_dma_unmap_single(qp_info->port_priv->device,
2648 				    recv->header.mapping,
2649 				    sizeof(struct ib_mad_private) -
2650 				      sizeof(struct ib_mad_private_header),
2651 				    DMA_FROM_DEVICE);
2652 		kmem_cache_free(ib_mad_cache, recv);
2653 	}
2654 
2655 	qp_info->recv_queue.count = 0;
2656 }
2657 
2658 /*
2659  * Start the port
2660  */
2661 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2662 {
2663 	int ret, i;
2664 	struct ib_qp_attr *attr;
2665 	struct ib_qp *qp;
2666 	u16 pkey_index;
2667 
2668 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2669 	if (!attr) {
2670 		printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2671 		return -ENOMEM;
2672 	}
2673 
2674 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2675 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2676 	if (ret)
2677 		pkey_index = 0;
2678 
2679 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2680 		qp = port_priv->qp_info[i].qp;
2681 		if (!qp)
2682 			continue;
2683 
2684 		/*
2685 		 * PKey index for QP1 is irrelevant but
2686 		 * one is needed for the Reset to Init transition
2687 		 */
2688 		attr->qp_state = IB_QPS_INIT;
2689 		attr->pkey_index = pkey_index;
2690 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2691 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2692 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2693 		if (ret) {
2694 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2695 			       "INIT: %d\n", i, ret);
2696 			goto out;
2697 		}
2698 
2699 		attr->qp_state = IB_QPS_RTR;
2700 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2701 		if (ret) {
2702 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2703 			       "RTR: %d\n", i, ret);
2704 			goto out;
2705 		}
2706 
2707 		attr->qp_state = IB_QPS_RTS;
2708 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2709 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2710 		if (ret) {
2711 			printk(KERN_ERR PFX "Couldn't change QP%d state to "
2712 			       "RTS: %d\n", i, ret);
2713 			goto out;
2714 		}
2715 	}
2716 
2717 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2718 	if (ret) {
2719 		printk(KERN_ERR PFX "Failed to request completion "
2720 		       "notification: %d\n", ret);
2721 		goto out;
2722 	}
2723 
2724 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2725 		if (!port_priv->qp_info[i].qp)
2726 			continue;
2727 
2728 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2729 		if (ret) {
2730 			printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2731 			goto out;
2732 		}
2733 	}
2734 out:
2735 	kfree(attr);
2736 	return ret;
2737 }
2738 
2739 static void qp_event_handler(struct ib_event *event, void *qp_context)
2740 {
2741 	struct ib_mad_qp_info	*qp_info = qp_context;
2742 
2743 	/* It's worse than that! He's dead, Jim! */
2744 	printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2745 		event->event, qp_info->qp->qp_num);
2746 }
2747 
2748 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2749 			   struct ib_mad_queue *mad_queue)
2750 {
2751 	mad_queue->qp_info = qp_info;
2752 	mad_queue->count = 0;
2753 	spin_lock_init(&mad_queue->lock);
2754 	INIT_LIST_HEAD(&mad_queue->list);
2755 }
2756 
2757 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2758 			struct ib_mad_qp_info *qp_info)
2759 {
2760 	qp_info->port_priv = port_priv;
2761 	init_mad_queue(qp_info, &qp_info->send_queue);
2762 	init_mad_queue(qp_info, &qp_info->recv_queue);
2763 	INIT_LIST_HEAD(&qp_info->overflow_list);
2764 	spin_lock_init(&qp_info->snoop_lock);
2765 	qp_info->snoop_table = NULL;
2766 	qp_info->snoop_table_size = 0;
2767 	atomic_set(&qp_info->snoop_count, 0);
2768 }
2769 
2770 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2771 			 enum ib_qp_type qp_type)
2772 {
2773 	struct ib_qp_init_attr	qp_init_attr;
2774 	int ret;
2775 
2776 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2777 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2778 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2779 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2780 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
2781 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2782 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2783 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2784 	qp_init_attr.qp_type = qp_type;
2785 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2786 	qp_init_attr.qp_context = qp_info;
2787 	qp_init_attr.event_handler = qp_event_handler;
2788 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2789 	if (IS_ERR(qp_info->qp)) {
2790 		printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2791 		       get_spl_qp_index(qp_type));
2792 		ret = PTR_ERR(qp_info->qp);
2793 		goto error;
2794 	}
2795 	/* Use minimum queue sizes unless the CQ is resized */
2796 	qp_info->send_queue.max_active = mad_sendq_size;
2797 	qp_info->recv_queue.max_active = mad_recvq_size;
2798 	return 0;
2799 
2800 error:
2801 	return ret;
2802 }
2803 
2804 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2805 {
2806 	if (!qp_info->qp)
2807 		return;
2808 
2809 	ib_destroy_qp(qp_info->qp);
2810 	kfree(qp_info->snoop_table);
2811 }
2812 
2813 /*
2814  * Open the port
2815  * Create the QP, PD, MR, and CQ if needed
2816  */
2817 static int ib_mad_port_open(struct ib_device *device,
2818 			    int port_num)
2819 {
2820 	int ret, cq_size;
2821 	struct ib_mad_port_private *port_priv;
2822 	unsigned long flags;
2823 	char name[sizeof "ib_mad123"];
2824 	int has_smi;
2825 
2826 	/* Create new device info */
2827 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2828 	if (!port_priv) {
2829 		printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2830 		return -ENOMEM;
2831 	}
2832 
2833 	port_priv->device = device;
2834 	port_priv->port_num = port_num;
2835 	spin_lock_init(&port_priv->reg_lock);
2836 	INIT_LIST_HEAD(&port_priv->agent_list);
2837 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2838 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2839 
2840 	cq_size = mad_sendq_size + mad_recvq_size;
2841 	has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2842 	if (has_smi)
2843 		cq_size *= 2;
2844 
2845 	port_priv->cq = ib_create_cq(port_priv->device,
2846 				     ib_mad_thread_completion_handler,
2847 				     NULL, port_priv, cq_size, 0);
2848 	if (IS_ERR(port_priv->cq)) {
2849 		printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2850 		ret = PTR_ERR(port_priv->cq);
2851 		goto error3;
2852 	}
2853 
2854 	port_priv->pd = ib_alloc_pd(device);
2855 	if (IS_ERR(port_priv->pd)) {
2856 		printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2857 		ret = PTR_ERR(port_priv->pd);
2858 		goto error4;
2859 	}
2860 
2861 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2862 	if (IS_ERR(port_priv->mr)) {
2863 		printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2864 		ret = PTR_ERR(port_priv->mr);
2865 		goto error5;
2866 	}
2867 
2868 	if (has_smi) {
2869 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2870 		if (ret)
2871 			goto error6;
2872 	}
2873 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2874 	if (ret)
2875 		goto error7;
2876 
2877 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2878 	port_priv->wq = create_singlethread_workqueue(name);
2879 	if (!port_priv->wq) {
2880 		ret = -ENOMEM;
2881 		goto error8;
2882 	}
2883 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2884 
2885 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2886 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2887 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2888 
2889 	ret = ib_mad_port_start(port_priv);
2890 	if (ret) {
2891 		printk(KERN_ERR PFX "Couldn't start port\n");
2892 		goto error9;
2893 	}
2894 
2895 	return 0;
2896 
2897 error9:
2898 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2899 	list_del_init(&port_priv->port_list);
2900 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2901 
2902 	destroy_workqueue(port_priv->wq);
2903 error8:
2904 	destroy_mad_qp(&port_priv->qp_info[1]);
2905 error7:
2906 	destroy_mad_qp(&port_priv->qp_info[0]);
2907 error6:
2908 	ib_dereg_mr(port_priv->mr);
2909 error5:
2910 	ib_dealloc_pd(port_priv->pd);
2911 error4:
2912 	ib_destroy_cq(port_priv->cq);
2913 	cleanup_recv_queue(&port_priv->qp_info[1]);
2914 	cleanup_recv_queue(&port_priv->qp_info[0]);
2915 error3:
2916 	kfree(port_priv);
2917 
2918 	return ret;
2919 }
2920 
2921 /*
2922  * Close the port
2923  * If there are no classes using the port, free the port
2924  * resources (CQ, MR, PD, QP) and remove the port's info structure
2925  */
2926 static int ib_mad_port_close(struct ib_device *device, int port_num)
2927 {
2928 	struct ib_mad_port_private *port_priv;
2929 	unsigned long flags;
2930 
2931 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2932 	port_priv = __ib_get_mad_port(device, port_num);
2933 	if (port_priv == NULL) {
2934 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2935 		printk(KERN_ERR PFX "Port %d not found\n", port_num);
2936 		return -ENODEV;
2937 	}
2938 	list_del_init(&port_priv->port_list);
2939 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2940 
2941 	destroy_workqueue(port_priv->wq);
2942 	destroy_mad_qp(&port_priv->qp_info[1]);
2943 	destroy_mad_qp(&port_priv->qp_info[0]);
2944 	ib_dereg_mr(port_priv->mr);
2945 	ib_dealloc_pd(port_priv->pd);
2946 	ib_destroy_cq(port_priv->cq);
2947 	cleanup_recv_queue(&port_priv->qp_info[1]);
2948 	cleanup_recv_queue(&port_priv->qp_info[0]);
2949 	/* XXX: Handle deallocation of MAD registration tables */
2950 
2951 	kfree(port_priv);
2952 
2953 	return 0;
2954 }
2955 
2956 static void ib_mad_init_device(struct ib_device *device)
2957 {
2958 	int start, end, i;
2959 
2960 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2961 		return;
2962 
2963 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2964 		start = 0;
2965 		end   = 0;
2966 	} else {
2967 		start = 1;
2968 		end   = device->phys_port_cnt;
2969 	}
2970 
2971 	for (i = start; i <= end; i++) {
2972 		if (ib_mad_port_open(device, i)) {
2973 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2974 			       device->name, i);
2975 			goto error;
2976 		}
2977 		if (ib_agent_port_open(device, i)) {
2978 			printk(KERN_ERR PFX "Couldn't open %s port %d "
2979 			       "for agents\n",
2980 			       device->name, i);
2981 			goto error_agent;
2982 		}
2983 	}
2984 	return;
2985 
2986 error_agent:
2987 	if (ib_mad_port_close(device, i))
2988 		printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2989 		       device->name, i);
2990 
2991 error:
2992 	i--;
2993 
2994 	while (i >= start) {
2995 		if (ib_agent_port_close(device, i))
2996 			printk(KERN_ERR PFX "Couldn't close %s port %d "
2997 			       "for agents\n",
2998 			       device->name, i);
2999 		if (ib_mad_port_close(device, i))
3000 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
3001 			       device->name, i);
3002 		i--;
3003 	}
3004 }
3005 
3006 static void ib_mad_remove_device(struct ib_device *device)
3007 {
3008 	int i, num_ports, cur_port;
3009 
3010 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3011 		return;
3012 
3013 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
3014 		num_ports = 1;
3015 		cur_port = 0;
3016 	} else {
3017 		num_ports = device->phys_port_cnt;
3018 		cur_port = 1;
3019 	}
3020 	for (i = 0; i < num_ports; i++, cur_port++) {
3021 		if (ib_agent_port_close(device, cur_port))
3022 			printk(KERN_ERR PFX "Couldn't close %s port %d "
3023 			       "for agents\n",
3024 			       device->name, cur_port);
3025 		if (ib_mad_port_close(device, cur_port))
3026 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
3027 			       device->name, cur_port);
3028 	}
3029 }
3030 
3031 static struct ib_client mad_client = {
3032 	.name   = "mad",
3033 	.add = ib_mad_init_device,
3034 	.remove = ib_mad_remove_device
3035 };
3036 
3037 static int __init ib_mad_init_module(void)
3038 {
3039 	int ret;
3040 
3041 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3042 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3043 
3044 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3045 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3046 
3047 	ib_mad_cache = kmem_cache_create("ib_mad",
3048 					 sizeof(struct ib_mad_private),
3049 					 0,
3050 					 SLAB_HWCACHE_ALIGN,
3051 					 NULL);
3052 	if (!ib_mad_cache) {
3053 		printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
3054 		ret = -ENOMEM;
3055 		goto error1;
3056 	}
3057 
3058 	INIT_LIST_HEAD(&ib_mad_port_list);
3059 
3060 	if (ib_register_client(&mad_client)) {
3061 		printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3062 		ret = -EINVAL;
3063 		goto error2;
3064 	}
3065 
3066 	return 0;
3067 
3068 error2:
3069 	kmem_cache_destroy(ib_mad_cache);
3070 error1:
3071 	return ret;
3072 }
3073 
3074 static void __exit ib_mad_cleanup_module(void)
3075 {
3076 	ib_unregister_client(&mad_client);
3077 	kmem_cache_destroy(ib_mad_cache);
3078 }
3079 
3080 module_init(ib_mad_init_module);
3081 module_exit(ib_mad_cleanup_module);
3082