xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision e917ba44)
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 
37 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
38 
39 #include <linux/dma-mapping.h>
40 #include <linux/slab.h>
41 #include <linux/module.h>
42 #include <rdma/ib_cache.h>
43 
44 #include "mad_priv.h"
45 #include "mad_rmpp.h"
46 #include "smi.h"
47 #include "agent.h"
48 
49 MODULE_LICENSE("Dual BSD/GPL");
50 MODULE_DESCRIPTION("kernel IB MAD API");
51 MODULE_AUTHOR("Hal Rosenstock");
52 MODULE_AUTHOR("Sean Hefty");
53 
54 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
55 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
56 
57 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
58 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
59 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
60 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
61 
62 static struct kmem_cache *ib_mad_cache;
63 
64 static struct list_head ib_mad_port_list;
65 static u32 ib_mad_client_id = 0;
66 
67 /* Port list lock */
68 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
69 
70 /* Forward declarations */
71 static int method_in_use(struct ib_mad_mgmt_method_table **method,
72 			 struct ib_mad_reg_req *mad_reg_req);
73 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
74 static struct ib_mad_agent_private *find_mad_agent(
75 					struct ib_mad_port_private *port_priv,
76 					struct ib_mad *mad);
77 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
78 				    struct ib_mad_private *mad);
79 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
80 static void timeout_sends(struct work_struct *work);
81 static void local_completions(struct work_struct *work);
82 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 			      struct ib_mad_agent_private *agent_priv,
84 			      u8 mgmt_class);
85 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
86 			   struct ib_mad_agent_private *agent_priv);
87 
88 /*
89  * Returns a ib_mad_port_private structure or NULL for a device/port
90  * Assumes ib_mad_port_list_lock is being held
91  */
92 static inline struct ib_mad_port_private *
93 __ib_get_mad_port(struct ib_device *device, int port_num)
94 {
95 	struct ib_mad_port_private *entry;
96 
97 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
98 		if (entry->device == device && entry->port_num == port_num)
99 			return entry;
100 	}
101 	return NULL;
102 }
103 
104 /*
105  * Wrapper function to return a ib_mad_port_private structure or NULL
106  * for a device/port
107  */
108 static inline struct ib_mad_port_private *
109 ib_get_mad_port(struct ib_device *device, int port_num)
110 {
111 	struct ib_mad_port_private *entry;
112 	unsigned long flags;
113 
114 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
115 	entry = __ib_get_mad_port(device, port_num);
116 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
117 
118 	return entry;
119 }
120 
121 static inline u8 convert_mgmt_class(u8 mgmt_class)
122 {
123 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
124 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
125 		0 : mgmt_class;
126 }
127 
128 static int get_spl_qp_index(enum ib_qp_type qp_type)
129 {
130 	switch (qp_type)
131 	{
132 	case IB_QPT_SMI:
133 		return 0;
134 	case IB_QPT_GSI:
135 		return 1;
136 	default:
137 		return -1;
138 	}
139 }
140 
141 static int vendor_class_index(u8 mgmt_class)
142 {
143 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
144 }
145 
146 static int is_vendor_class(u8 mgmt_class)
147 {
148 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
149 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
150 		return 0;
151 	return 1;
152 }
153 
154 static int is_vendor_oui(char *oui)
155 {
156 	if (oui[0] || oui[1] || oui[2])
157 		return 1;
158 	return 0;
159 }
160 
161 static int is_vendor_method_in_use(
162 		struct ib_mad_mgmt_vendor_class *vendor_class,
163 		struct ib_mad_reg_req *mad_reg_req)
164 {
165 	struct ib_mad_mgmt_method_table *method;
166 	int i;
167 
168 	for (i = 0; i < MAX_MGMT_OUI; i++) {
169 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
170 			method = vendor_class->method_table[i];
171 			if (method) {
172 				if (method_in_use(&method, mad_reg_req))
173 					return 1;
174 				else
175 					break;
176 			}
177 		}
178 	}
179 	return 0;
180 }
181 
182 int ib_response_mad(struct ib_mad *mad)
183 {
184 	return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
185 		(mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
186 		((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
187 		 (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
188 }
189 EXPORT_SYMBOL(ib_response_mad);
190 
191 /*
192  * ib_register_mad_agent - Register to send/receive MADs
193  */
194 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
195 					   u8 port_num,
196 					   enum ib_qp_type qp_type,
197 					   struct ib_mad_reg_req *mad_reg_req,
198 					   u8 rmpp_version,
199 					   ib_mad_send_handler send_handler,
200 					   ib_mad_recv_handler recv_handler,
201 					   void *context,
202 					   u32 registration_flags)
203 {
204 	struct ib_mad_port_private *port_priv;
205 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
206 	struct ib_mad_agent_private *mad_agent_priv;
207 	struct ib_mad_reg_req *reg_req = NULL;
208 	struct ib_mad_mgmt_class_table *class;
209 	struct ib_mad_mgmt_vendor_class_table *vendor;
210 	struct ib_mad_mgmt_vendor_class *vendor_class;
211 	struct ib_mad_mgmt_method_table *method;
212 	int ret2, qpn;
213 	unsigned long flags;
214 	u8 mgmt_class, vclass;
215 
216 	/* Validate parameters */
217 	qpn = get_spl_qp_index(qp_type);
218 	if (qpn == -1) {
219 		dev_notice(&device->dev,
220 			   "ib_register_mad_agent: invalid QP Type %d\n",
221 			   qp_type);
222 		goto error1;
223 	}
224 
225 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
226 		dev_notice(&device->dev,
227 			   "ib_register_mad_agent: invalid RMPP Version %u\n",
228 			   rmpp_version);
229 		goto error1;
230 	}
231 
232 	/* Validate MAD registration request if supplied */
233 	if (mad_reg_req) {
234 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
235 			dev_notice(&device->dev,
236 				   "ib_register_mad_agent: invalid Class Version %u\n",
237 				   mad_reg_req->mgmt_class_version);
238 			goto error1;
239 		}
240 		if (!recv_handler) {
241 			dev_notice(&device->dev,
242 				   "ib_register_mad_agent: no recv_handler\n");
243 			goto error1;
244 		}
245 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
246 			/*
247 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
248 			 * one in this range currently allowed
249 			 */
250 			if (mad_reg_req->mgmt_class !=
251 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
252 				dev_notice(&device->dev,
253 					   "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
254 					   mad_reg_req->mgmt_class);
255 				goto error1;
256 			}
257 		} else if (mad_reg_req->mgmt_class == 0) {
258 			/*
259 			 * Class 0 is reserved in IBA and is used for
260 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
261 			 */
262 			dev_notice(&device->dev,
263 				   "ib_register_mad_agent: Invalid Mgmt Class 0\n");
264 			goto error1;
265 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
266 			/*
267 			 * If class is in "new" vendor range,
268 			 * ensure supplied OUI is not zero
269 			 */
270 			if (!is_vendor_oui(mad_reg_req->oui)) {
271 				dev_notice(&device->dev,
272 					   "ib_register_mad_agent: No OUI specified for class 0x%x\n",
273 					   mad_reg_req->mgmt_class);
274 				goto error1;
275 			}
276 		}
277 		/* Make sure class supplied is consistent with RMPP */
278 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
279 			if (rmpp_version) {
280 				dev_notice(&device->dev,
281 					   "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
282 					   mad_reg_req->mgmt_class);
283 				goto error1;
284 			}
285 		}
286 
287 		/* Make sure class supplied is consistent with QP type */
288 		if (qp_type == IB_QPT_SMI) {
289 			if ((mad_reg_req->mgmt_class !=
290 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
291 			    (mad_reg_req->mgmt_class !=
292 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
293 				dev_notice(&device->dev,
294 					   "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
295 					   mad_reg_req->mgmt_class);
296 				goto error1;
297 			}
298 		} else {
299 			if ((mad_reg_req->mgmt_class ==
300 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
301 			    (mad_reg_req->mgmt_class ==
302 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
303 				dev_notice(&device->dev,
304 					   "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
305 					   mad_reg_req->mgmt_class);
306 				goto error1;
307 			}
308 		}
309 	} else {
310 		/* No registration request supplied */
311 		if (!send_handler)
312 			goto error1;
313 		if (registration_flags & IB_MAD_USER_RMPP)
314 			goto error1;
315 	}
316 
317 	/* Validate device and port */
318 	port_priv = ib_get_mad_port(device, port_num);
319 	if (!port_priv) {
320 		dev_notice(&device->dev, "ib_register_mad_agent: Invalid port\n");
321 		ret = ERR_PTR(-ENODEV);
322 		goto error1;
323 	}
324 
325 	/* Verify the QP requested is supported.  For example, Ethernet devices
326 	 * will not have QP0 */
327 	if (!port_priv->qp_info[qpn].qp) {
328 		dev_notice(&device->dev,
329 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
330 		ret = ERR_PTR(-EPROTONOSUPPORT);
331 		goto error1;
332 	}
333 
334 	/* Allocate structures */
335 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
336 	if (!mad_agent_priv) {
337 		ret = ERR_PTR(-ENOMEM);
338 		goto error1;
339 	}
340 
341 	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
342 						 IB_ACCESS_LOCAL_WRITE);
343 	if (IS_ERR(mad_agent_priv->agent.mr)) {
344 		ret = ERR_PTR(-ENOMEM);
345 		goto error2;
346 	}
347 
348 	if (mad_reg_req) {
349 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
350 		if (!reg_req) {
351 			ret = ERR_PTR(-ENOMEM);
352 			goto error3;
353 		}
354 	}
355 
356 	/* Now, fill in the various structures */
357 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
358 	mad_agent_priv->reg_req = reg_req;
359 	mad_agent_priv->agent.rmpp_version = rmpp_version;
360 	mad_agent_priv->agent.device = device;
361 	mad_agent_priv->agent.recv_handler = recv_handler;
362 	mad_agent_priv->agent.send_handler = send_handler;
363 	mad_agent_priv->agent.context = context;
364 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
365 	mad_agent_priv->agent.port_num = port_num;
366 	mad_agent_priv->agent.flags = registration_flags;
367 	spin_lock_init(&mad_agent_priv->lock);
368 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
369 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
370 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
371 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
372 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
373 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
374 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
375 	atomic_set(&mad_agent_priv->refcount, 1);
376 	init_completion(&mad_agent_priv->comp);
377 
378 	spin_lock_irqsave(&port_priv->reg_lock, flags);
379 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
380 
381 	/*
382 	 * Make sure MAD registration (if supplied)
383 	 * is non overlapping with any existing ones
384 	 */
385 	if (mad_reg_req) {
386 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
387 		if (!is_vendor_class(mgmt_class)) {
388 			class = port_priv->version[mad_reg_req->
389 						   mgmt_class_version].class;
390 			if (class) {
391 				method = class->method_table[mgmt_class];
392 				if (method) {
393 					if (method_in_use(&method,
394 							   mad_reg_req))
395 						goto error4;
396 				}
397 			}
398 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
399 						  mgmt_class);
400 		} else {
401 			/* "New" vendor class range */
402 			vendor = port_priv->version[mad_reg_req->
403 						    mgmt_class_version].vendor;
404 			if (vendor) {
405 				vclass = vendor_class_index(mgmt_class);
406 				vendor_class = vendor->vendor_class[vclass];
407 				if (vendor_class) {
408 					if (is_vendor_method_in_use(
409 							vendor_class,
410 							mad_reg_req))
411 						goto error4;
412 				}
413 			}
414 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
415 		}
416 		if (ret2) {
417 			ret = ERR_PTR(ret2);
418 			goto error4;
419 		}
420 	}
421 
422 	/* Add mad agent into port's agent list */
423 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
424 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
425 
426 	return &mad_agent_priv->agent;
427 
428 error4:
429 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
430 	kfree(reg_req);
431 error3:
432 	ib_dereg_mr(mad_agent_priv->agent.mr);
433 error2:
434 	kfree(mad_agent_priv);
435 error1:
436 	return ret;
437 }
438 EXPORT_SYMBOL(ib_register_mad_agent);
439 
440 static inline int is_snooping_sends(int mad_snoop_flags)
441 {
442 	return (mad_snoop_flags &
443 		(/*IB_MAD_SNOOP_POSTED_SENDS |
444 		 IB_MAD_SNOOP_RMPP_SENDS |*/
445 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
446 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
447 }
448 
449 static inline int is_snooping_recvs(int mad_snoop_flags)
450 {
451 	return (mad_snoop_flags &
452 		(IB_MAD_SNOOP_RECVS /*|
453 		 IB_MAD_SNOOP_RMPP_RECVS*/));
454 }
455 
456 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
457 				struct ib_mad_snoop_private *mad_snoop_priv)
458 {
459 	struct ib_mad_snoop_private **new_snoop_table;
460 	unsigned long flags;
461 	int i;
462 
463 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
464 	/* Check for empty slot in array. */
465 	for (i = 0; i < qp_info->snoop_table_size; i++)
466 		if (!qp_info->snoop_table[i])
467 			break;
468 
469 	if (i == qp_info->snoop_table_size) {
470 		/* Grow table. */
471 		new_snoop_table = krealloc(qp_info->snoop_table,
472 					   sizeof mad_snoop_priv *
473 					   (qp_info->snoop_table_size + 1),
474 					   GFP_ATOMIC);
475 		if (!new_snoop_table) {
476 			i = -ENOMEM;
477 			goto out;
478 		}
479 
480 		qp_info->snoop_table = new_snoop_table;
481 		qp_info->snoop_table_size++;
482 	}
483 	qp_info->snoop_table[i] = mad_snoop_priv;
484 	atomic_inc(&qp_info->snoop_count);
485 out:
486 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
487 	return i;
488 }
489 
490 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
491 					   u8 port_num,
492 					   enum ib_qp_type qp_type,
493 					   int mad_snoop_flags,
494 					   ib_mad_snoop_handler snoop_handler,
495 					   ib_mad_recv_handler recv_handler,
496 					   void *context)
497 {
498 	struct ib_mad_port_private *port_priv;
499 	struct ib_mad_agent *ret;
500 	struct ib_mad_snoop_private *mad_snoop_priv;
501 	int qpn;
502 
503 	/* Validate parameters */
504 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
505 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
506 		ret = ERR_PTR(-EINVAL);
507 		goto error1;
508 	}
509 	qpn = get_spl_qp_index(qp_type);
510 	if (qpn == -1) {
511 		ret = ERR_PTR(-EINVAL);
512 		goto error1;
513 	}
514 	port_priv = ib_get_mad_port(device, port_num);
515 	if (!port_priv) {
516 		ret = ERR_PTR(-ENODEV);
517 		goto error1;
518 	}
519 	/* Allocate structures */
520 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
521 	if (!mad_snoop_priv) {
522 		ret = ERR_PTR(-ENOMEM);
523 		goto error1;
524 	}
525 
526 	/* Now, fill in the various structures */
527 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
528 	mad_snoop_priv->agent.device = device;
529 	mad_snoop_priv->agent.recv_handler = recv_handler;
530 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
531 	mad_snoop_priv->agent.context = context;
532 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
533 	mad_snoop_priv->agent.port_num = port_num;
534 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
535 	init_completion(&mad_snoop_priv->comp);
536 	mad_snoop_priv->snoop_index = register_snoop_agent(
537 						&port_priv->qp_info[qpn],
538 						mad_snoop_priv);
539 	if (mad_snoop_priv->snoop_index < 0) {
540 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
541 		goto error2;
542 	}
543 
544 	atomic_set(&mad_snoop_priv->refcount, 1);
545 	return &mad_snoop_priv->agent;
546 
547 error2:
548 	kfree(mad_snoop_priv);
549 error1:
550 	return ret;
551 }
552 EXPORT_SYMBOL(ib_register_mad_snoop);
553 
554 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
555 {
556 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
557 		complete(&mad_agent_priv->comp);
558 }
559 
560 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
561 {
562 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
563 		complete(&mad_snoop_priv->comp);
564 }
565 
566 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
567 {
568 	struct ib_mad_port_private *port_priv;
569 	unsigned long flags;
570 
571 	/* Note that we could still be handling received MADs */
572 
573 	/*
574 	 * Canceling all sends results in dropping received response
575 	 * MADs, preventing us from queuing additional work
576 	 */
577 	cancel_mads(mad_agent_priv);
578 	port_priv = mad_agent_priv->qp_info->port_priv;
579 	cancel_delayed_work(&mad_agent_priv->timed_work);
580 
581 	spin_lock_irqsave(&port_priv->reg_lock, flags);
582 	remove_mad_reg_req(mad_agent_priv);
583 	list_del(&mad_agent_priv->agent_list);
584 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
585 
586 	flush_workqueue(port_priv->wq);
587 	ib_cancel_rmpp_recvs(mad_agent_priv);
588 
589 	deref_mad_agent(mad_agent_priv);
590 	wait_for_completion(&mad_agent_priv->comp);
591 
592 	kfree(mad_agent_priv->reg_req);
593 	ib_dereg_mr(mad_agent_priv->agent.mr);
594 	kfree(mad_agent_priv);
595 }
596 
597 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
598 {
599 	struct ib_mad_qp_info *qp_info;
600 	unsigned long flags;
601 
602 	qp_info = mad_snoop_priv->qp_info;
603 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
604 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
605 	atomic_dec(&qp_info->snoop_count);
606 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
607 
608 	deref_snoop_agent(mad_snoop_priv);
609 	wait_for_completion(&mad_snoop_priv->comp);
610 
611 	kfree(mad_snoop_priv);
612 }
613 
614 /*
615  * ib_unregister_mad_agent - Unregisters a client from using MAD services
616  */
617 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
618 {
619 	struct ib_mad_agent_private *mad_agent_priv;
620 	struct ib_mad_snoop_private *mad_snoop_priv;
621 
622 	/* If the TID is zero, the agent can only snoop. */
623 	if (mad_agent->hi_tid) {
624 		mad_agent_priv = container_of(mad_agent,
625 					      struct ib_mad_agent_private,
626 					      agent);
627 		unregister_mad_agent(mad_agent_priv);
628 	} else {
629 		mad_snoop_priv = container_of(mad_agent,
630 					      struct ib_mad_snoop_private,
631 					      agent);
632 		unregister_mad_snoop(mad_snoop_priv);
633 	}
634 	return 0;
635 }
636 EXPORT_SYMBOL(ib_unregister_mad_agent);
637 
638 static void dequeue_mad(struct ib_mad_list_head *mad_list)
639 {
640 	struct ib_mad_queue *mad_queue;
641 	unsigned long flags;
642 
643 	BUG_ON(!mad_list->mad_queue);
644 	mad_queue = mad_list->mad_queue;
645 	spin_lock_irqsave(&mad_queue->lock, flags);
646 	list_del(&mad_list->list);
647 	mad_queue->count--;
648 	spin_unlock_irqrestore(&mad_queue->lock, flags);
649 }
650 
651 static void snoop_send(struct ib_mad_qp_info *qp_info,
652 		       struct ib_mad_send_buf *send_buf,
653 		       struct ib_mad_send_wc *mad_send_wc,
654 		       int mad_snoop_flags)
655 {
656 	struct ib_mad_snoop_private *mad_snoop_priv;
657 	unsigned long flags;
658 	int i;
659 
660 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
661 	for (i = 0; i < qp_info->snoop_table_size; i++) {
662 		mad_snoop_priv = qp_info->snoop_table[i];
663 		if (!mad_snoop_priv ||
664 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
665 			continue;
666 
667 		atomic_inc(&mad_snoop_priv->refcount);
668 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
669 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
670 						    send_buf, mad_send_wc);
671 		deref_snoop_agent(mad_snoop_priv);
672 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
673 	}
674 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
675 }
676 
677 static void snoop_recv(struct ib_mad_qp_info *qp_info,
678 		       struct ib_mad_recv_wc *mad_recv_wc,
679 		       int mad_snoop_flags)
680 {
681 	struct ib_mad_snoop_private *mad_snoop_priv;
682 	unsigned long flags;
683 	int i;
684 
685 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
686 	for (i = 0; i < qp_info->snoop_table_size; i++) {
687 		mad_snoop_priv = qp_info->snoop_table[i];
688 		if (!mad_snoop_priv ||
689 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
690 			continue;
691 
692 		atomic_inc(&mad_snoop_priv->refcount);
693 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
694 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
695 						   mad_recv_wc);
696 		deref_snoop_agent(mad_snoop_priv);
697 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
698 	}
699 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
700 }
701 
702 static void build_smp_wc(struct ib_qp *qp,
703 			 u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
704 			 struct ib_wc *wc)
705 {
706 	memset(wc, 0, sizeof *wc);
707 	wc->wr_id = wr_id;
708 	wc->status = IB_WC_SUCCESS;
709 	wc->opcode = IB_WC_RECV;
710 	wc->pkey_index = pkey_index;
711 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
712 	wc->src_qp = IB_QP0;
713 	wc->qp = qp;
714 	wc->slid = slid;
715 	wc->sl = 0;
716 	wc->dlid_path_bits = 0;
717 	wc->port_num = port_num;
718 }
719 
720 /*
721  * Return 0 if SMP is to be sent
722  * Return 1 if SMP was consumed locally (whether or not solicited)
723  * Return < 0 if error
724  */
725 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
726 				  struct ib_mad_send_wr_private *mad_send_wr)
727 {
728 	int ret = 0;
729 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
730 	unsigned long flags;
731 	struct ib_mad_local_private *local;
732 	struct ib_mad_private *mad_priv;
733 	struct ib_mad_port_private *port_priv;
734 	struct ib_mad_agent_private *recv_mad_agent = NULL;
735 	struct ib_device *device = mad_agent_priv->agent.device;
736 	u8 port_num;
737 	struct ib_wc mad_wc;
738 	struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
739 
740 	if (device->node_type == RDMA_NODE_IB_SWITCH &&
741 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
742 		port_num = send_wr->wr.ud.port_num;
743 	else
744 		port_num = mad_agent_priv->agent.port_num;
745 
746 	/*
747 	 * Directed route handling starts if the initial LID routed part of
748 	 * a request or the ending LID routed part of a response is empty.
749 	 * If we are at the start of the LID routed part, don't update the
750 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
751 	 */
752 	if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
753 	     IB_LID_PERMISSIVE &&
754 	     smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
755 	     IB_SMI_DISCARD) {
756 		ret = -EINVAL;
757 		dev_err(&device->dev, "Invalid directed route\n");
758 		goto out;
759 	}
760 
761 	/* Check to post send on QP or process locally */
762 	if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
763 	    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
764 		goto out;
765 
766 	local = kmalloc(sizeof *local, GFP_ATOMIC);
767 	if (!local) {
768 		ret = -ENOMEM;
769 		dev_err(&device->dev, "No memory for ib_mad_local_private\n");
770 		goto out;
771 	}
772 	local->mad_priv = NULL;
773 	local->recv_mad_agent = NULL;
774 	mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
775 	if (!mad_priv) {
776 		ret = -ENOMEM;
777 		dev_err(&device->dev, "No memory for local response MAD\n");
778 		kfree(local);
779 		goto out;
780 	}
781 
782 	build_smp_wc(mad_agent_priv->agent.qp,
783 		     send_wr->wr_id, be16_to_cpu(smp->dr_slid),
784 		     send_wr->wr.ud.pkey_index,
785 		     send_wr->wr.ud.port_num, &mad_wc);
786 
787 	/* No GRH for DR SMP */
788 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
789 				  (struct ib_mad *)smp,
790 				  (struct ib_mad *)&mad_priv->mad);
791 	switch (ret)
792 	{
793 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
794 		if (ib_response_mad(&mad_priv->mad.mad) &&
795 		    mad_agent_priv->agent.recv_handler) {
796 			local->mad_priv = mad_priv;
797 			local->recv_mad_agent = mad_agent_priv;
798 			/*
799 			 * Reference MAD agent until receive
800 			 * side of local completion handled
801 			 */
802 			atomic_inc(&mad_agent_priv->refcount);
803 		} else
804 			kmem_cache_free(ib_mad_cache, mad_priv);
805 		break;
806 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
807 		kmem_cache_free(ib_mad_cache, mad_priv);
808 		break;
809 	case IB_MAD_RESULT_SUCCESS:
810 		/* Treat like an incoming receive MAD */
811 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
812 					    mad_agent_priv->agent.port_num);
813 		if (port_priv) {
814 			memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
815 			recv_mad_agent = find_mad_agent(port_priv,
816 						        &mad_priv->mad.mad);
817 		}
818 		if (!port_priv || !recv_mad_agent) {
819 			/*
820 			 * No receiving agent so drop packet and
821 			 * generate send completion.
822 			 */
823 			kmem_cache_free(ib_mad_cache, mad_priv);
824 			break;
825 		}
826 		local->mad_priv = mad_priv;
827 		local->recv_mad_agent = recv_mad_agent;
828 		break;
829 	default:
830 		kmem_cache_free(ib_mad_cache, mad_priv);
831 		kfree(local);
832 		ret = -EINVAL;
833 		goto out;
834 	}
835 
836 	local->mad_send_wr = mad_send_wr;
837 	/* Reference MAD agent until send side of local completion handled */
838 	atomic_inc(&mad_agent_priv->refcount);
839 	/* Queue local completion to local list */
840 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
841 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
842 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
843 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
844 		   &mad_agent_priv->local_work);
845 	ret = 1;
846 out:
847 	return ret;
848 }
849 
850 static int get_pad_size(int hdr_len, int data_len)
851 {
852 	int seg_size, pad;
853 
854 	seg_size = sizeof(struct ib_mad) - hdr_len;
855 	if (data_len && seg_size) {
856 		pad = seg_size - data_len % seg_size;
857 		return pad == seg_size ? 0 : pad;
858 	} else
859 		return seg_size;
860 }
861 
862 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
863 {
864 	struct ib_rmpp_segment *s, *t;
865 
866 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
867 		list_del(&s->list);
868 		kfree(s);
869 	}
870 }
871 
872 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
873 				gfp_t gfp_mask)
874 {
875 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
876 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
877 	struct ib_rmpp_segment *seg = NULL;
878 	int left, seg_size, pad;
879 
880 	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
881 	seg_size = send_buf->seg_size;
882 	pad = send_wr->pad;
883 
884 	/* Allocate data segments. */
885 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
886 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
887 		if (!seg) {
888 			dev_err(&send_buf->mad_agent->device->dev,
889 				"alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
890 				sizeof (*seg) + seg_size, gfp_mask);
891 			free_send_rmpp_list(send_wr);
892 			return -ENOMEM;
893 		}
894 		seg->num = ++send_buf->seg_count;
895 		list_add_tail(&seg->list, &send_wr->rmpp_list);
896 	}
897 
898 	/* Zero any padding */
899 	if (pad)
900 		memset(seg->data + seg_size - pad, 0, pad);
901 
902 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
903 					  agent.rmpp_version;
904 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
905 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
906 
907 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
908 					struct ib_rmpp_segment, list);
909 	send_wr->last_ack_seg = send_wr->cur_seg;
910 	return 0;
911 }
912 
913 int ib_mad_kernel_rmpp_agent(struct ib_mad_agent *agent)
914 {
915 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
916 }
917 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
918 
919 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
920 					    u32 remote_qpn, u16 pkey_index,
921 					    int rmpp_active,
922 					    int hdr_len, int data_len,
923 					    gfp_t gfp_mask)
924 {
925 	struct ib_mad_agent_private *mad_agent_priv;
926 	struct ib_mad_send_wr_private *mad_send_wr;
927 	int pad, message_size, ret, size;
928 	void *buf;
929 
930 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
931 				      agent);
932 	pad = get_pad_size(hdr_len, data_len);
933 	message_size = hdr_len + data_len + pad;
934 
935 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
936 		if (!rmpp_active && message_size > sizeof(struct ib_mad))
937 			return ERR_PTR(-EINVAL);
938 	} else
939 		if (rmpp_active || message_size > sizeof(struct ib_mad))
940 			return ERR_PTR(-EINVAL);
941 
942 	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
943 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
944 	if (!buf)
945 		return ERR_PTR(-ENOMEM);
946 
947 	mad_send_wr = buf + size;
948 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
949 	mad_send_wr->send_buf.mad = buf;
950 	mad_send_wr->send_buf.hdr_len = hdr_len;
951 	mad_send_wr->send_buf.data_len = data_len;
952 	mad_send_wr->pad = pad;
953 
954 	mad_send_wr->mad_agent_priv = mad_agent_priv;
955 	mad_send_wr->sg_list[0].length = hdr_len;
956 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
957 	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
958 	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
959 
960 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
961 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
962 	mad_send_wr->send_wr.num_sge = 2;
963 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
964 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
965 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
966 	mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
967 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
968 
969 	if (rmpp_active) {
970 		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
971 		if (ret) {
972 			kfree(buf);
973 			return ERR_PTR(ret);
974 		}
975 	}
976 
977 	mad_send_wr->send_buf.mad_agent = mad_agent;
978 	atomic_inc(&mad_agent_priv->refcount);
979 	return &mad_send_wr->send_buf;
980 }
981 EXPORT_SYMBOL(ib_create_send_mad);
982 
983 int ib_get_mad_data_offset(u8 mgmt_class)
984 {
985 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
986 		return IB_MGMT_SA_HDR;
987 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
988 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
989 		 (mgmt_class == IB_MGMT_CLASS_BIS))
990 		return IB_MGMT_DEVICE_HDR;
991 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
992 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
993 		return IB_MGMT_VENDOR_HDR;
994 	else
995 		return IB_MGMT_MAD_HDR;
996 }
997 EXPORT_SYMBOL(ib_get_mad_data_offset);
998 
999 int ib_is_mad_class_rmpp(u8 mgmt_class)
1000 {
1001 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1002 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1003 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1004 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1005 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1006 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1007 		return 1;
1008 	return 0;
1009 }
1010 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1011 
1012 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1013 {
1014 	struct ib_mad_send_wr_private *mad_send_wr;
1015 	struct list_head *list;
1016 
1017 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1018 				   send_buf);
1019 	list = &mad_send_wr->cur_seg->list;
1020 
1021 	if (mad_send_wr->cur_seg->num < seg_num) {
1022 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1023 			if (mad_send_wr->cur_seg->num == seg_num)
1024 				break;
1025 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1026 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1027 			if (mad_send_wr->cur_seg->num == seg_num)
1028 				break;
1029 	}
1030 	return mad_send_wr->cur_seg->data;
1031 }
1032 EXPORT_SYMBOL(ib_get_rmpp_segment);
1033 
1034 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1035 {
1036 	if (mad_send_wr->send_buf.seg_count)
1037 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1038 					   mad_send_wr->seg_num);
1039 	else
1040 		return mad_send_wr->send_buf.mad +
1041 		       mad_send_wr->send_buf.hdr_len;
1042 }
1043 
1044 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1045 {
1046 	struct ib_mad_agent_private *mad_agent_priv;
1047 	struct ib_mad_send_wr_private *mad_send_wr;
1048 
1049 	mad_agent_priv = container_of(send_buf->mad_agent,
1050 				      struct ib_mad_agent_private, agent);
1051 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1052 				   send_buf);
1053 
1054 	free_send_rmpp_list(mad_send_wr);
1055 	kfree(send_buf->mad);
1056 	deref_mad_agent(mad_agent_priv);
1057 }
1058 EXPORT_SYMBOL(ib_free_send_mad);
1059 
1060 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1061 {
1062 	struct ib_mad_qp_info *qp_info;
1063 	struct list_head *list;
1064 	struct ib_send_wr *bad_send_wr;
1065 	struct ib_mad_agent *mad_agent;
1066 	struct ib_sge *sge;
1067 	unsigned long flags;
1068 	int ret;
1069 
1070 	/* Set WR ID to find mad_send_wr upon completion */
1071 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1072 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1073 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1074 
1075 	mad_agent = mad_send_wr->send_buf.mad_agent;
1076 	sge = mad_send_wr->sg_list;
1077 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1078 					mad_send_wr->send_buf.mad,
1079 					sge[0].length,
1080 					DMA_TO_DEVICE);
1081 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1082 		return -ENOMEM;
1083 
1084 	mad_send_wr->header_mapping = sge[0].addr;
1085 
1086 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1087 					ib_get_payload(mad_send_wr),
1088 					sge[1].length,
1089 					DMA_TO_DEVICE);
1090 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1091 		ib_dma_unmap_single(mad_agent->device,
1092 				    mad_send_wr->header_mapping,
1093 				    sge[0].length, DMA_TO_DEVICE);
1094 		return -ENOMEM;
1095 	}
1096 	mad_send_wr->payload_mapping = sge[1].addr;
1097 
1098 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1099 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1100 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1101 				   &bad_send_wr);
1102 		list = &qp_info->send_queue.list;
1103 	} else {
1104 		ret = 0;
1105 		list = &qp_info->overflow_list;
1106 	}
1107 
1108 	if (!ret) {
1109 		qp_info->send_queue.count++;
1110 		list_add_tail(&mad_send_wr->mad_list.list, list);
1111 	}
1112 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1113 	if (ret) {
1114 		ib_dma_unmap_single(mad_agent->device,
1115 				    mad_send_wr->header_mapping,
1116 				    sge[0].length, DMA_TO_DEVICE);
1117 		ib_dma_unmap_single(mad_agent->device,
1118 				    mad_send_wr->payload_mapping,
1119 				    sge[1].length, DMA_TO_DEVICE);
1120 	}
1121 	return ret;
1122 }
1123 
1124 /*
1125  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1126  *  with the registered client
1127  */
1128 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1129 		     struct ib_mad_send_buf **bad_send_buf)
1130 {
1131 	struct ib_mad_agent_private *mad_agent_priv;
1132 	struct ib_mad_send_buf *next_send_buf;
1133 	struct ib_mad_send_wr_private *mad_send_wr;
1134 	unsigned long flags;
1135 	int ret = -EINVAL;
1136 
1137 	/* Walk list of send WRs and post each on send list */
1138 	for (; send_buf; send_buf = next_send_buf) {
1139 
1140 		mad_send_wr = container_of(send_buf,
1141 					   struct ib_mad_send_wr_private,
1142 					   send_buf);
1143 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1144 
1145 		if (!send_buf->mad_agent->send_handler ||
1146 		    (send_buf->timeout_ms &&
1147 		     !send_buf->mad_agent->recv_handler)) {
1148 			ret = -EINVAL;
1149 			goto error;
1150 		}
1151 
1152 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1153 			if (mad_agent_priv->agent.rmpp_version) {
1154 				ret = -EINVAL;
1155 				goto error;
1156 			}
1157 		}
1158 
1159 		/*
1160 		 * Save pointer to next work request to post in case the
1161 		 * current one completes, and the user modifies the work
1162 		 * request associated with the completion
1163 		 */
1164 		next_send_buf = send_buf->next;
1165 		mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1166 
1167 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1168 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1169 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1170 						     mad_send_wr);
1171 			if (ret < 0)		/* error */
1172 				goto error;
1173 			else if (ret == 1)	/* locally consumed */
1174 				continue;
1175 		}
1176 
1177 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1178 		/* Timeout will be updated after send completes */
1179 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1180 		mad_send_wr->max_retries = send_buf->retries;
1181 		mad_send_wr->retries_left = send_buf->retries;
1182 		send_buf->retries = 0;
1183 		/* Reference for work request to QP + response */
1184 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1185 		mad_send_wr->status = IB_WC_SUCCESS;
1186 
1187 		/* Reference MAD agent until send completes */
1188 		atomic_inc(&mad_agent_priv->refcount);
1189 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1190 		list_add_tail(&mad_send_wr->agent_list,
1191 			      &mad_agent_priv->send_list);
1192 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1193 
1194 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1195 			ret = ib_send_rmpp_mad(mad_send_wr);
1196 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1197 				ret = ib_send_mad(mad_send_wr);
1198 		} else
1199 			ret = ib_send_mad(mad_send_wr);
1200 		if (ret < 0) {
1201 			/* Fail send request */
1202 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1203 			list_del(&mad_send_wr->agent_list);
1204 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1205 			atomic_dec(&mad_agent_priv->refcount);
1206 			goto error;
1207 		}
1208 	}
1209 	return 0;
1210 error:
1211 	if (bad_send_buf)
1212 		*bad_send_buf = send_buf;
1213 	return ret;
1214 }
1215 EXPORT_SYMBOL(ib_post_send_mad);
1216 
1217 /*
1218  * ib_free_recv_mad - Returns data buffers used to receive
1219  *  a MAD to the access layer
1220  */
1221 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1222 {
1223 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1224 	struct ib_mad_private_header *mad_priv_hdr;
1225 	struct ib_mad_private *priv;
1226 	struct list_head free_list;
1227 
1228 	INIT_LIST_HEAD(&free_list);
1229 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1230 
1231 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1232 					&free_list, list) {
1233 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1234 					   recv_buf);
1235 		mad_priv_hdr = container_of(mad_recv_wc,
1236 					    struct ib_mad_private_header,
1237 					    recv_wc);
1238 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1239 				    header);
1240 		kmem_cache_free(ib_mad_cache, priv);
1241 	}
1242 }
1243 EXPORT_SYMBOL(ib_free_recv_mad);
1244 
1245 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1246 					u8 rmpp_version,
1247 					ib_mad_send_handler send_handler,
1248 					ib_mad_recv_handler recv_handler,
1249 					void *context)
1250 {
1251 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1252 }
1253 EXPORT_SYMBOL(ib_redirect_mad_qp);
1254 
1255 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1256 		      struct ib_wc *wc)
1257 {
1258 	dev_err(&mad_agent->device->dev,
1259 		"ib_process_mad_wc() not implemented yet\n");
1260 	return 0;
1261 }
1262 EXPORT_SYMBOL(ib_process_mad_wc);
1263 
1264 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1265 			 struct ib_mad_reg_req *mad_reg_req)
1266 {
1267 	int i;
1268 
1269 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1270 		if ((*method)->agent[i]) {
1271 			pr_err("Method %d already in use\n", i);
1272 			return -EINVAL;
1273 		}
1274 	}
1275 	return 0;
1276 }
1277 
1278 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1279 {
1280 	/* Allocate management method table */
1281 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1282 	if (!*method) {
1283 		pr_err("No memory for ib_mad_mgmt_method_table\n");
1284 		return -ENOMEM;
1285 	}
1286 
1287 	return 0;
1288 }
1289 
1290 /*
1291  * Check to see if there are any methods still in use
1292  */
1293 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1294 {
1295 	int i;
1296 
1297 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1298 		if (method->agent[i])
1299 			return 1;
1300 	return 0;
1301 }
1302 
1303 /*
1304  * Check to see if there are any method tables for this class still in use
1305  */
1306 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1307 {
1308 	int i;
1309 
1310 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1311 		if (class->method_table[i])
1312 			return 1;
1313 	return 0;
1314 }
1315 
1316 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1317 {
1318 	int i;
1319 
1320 	for (i = 0; i < MAX_MGMT_OUI; i++)
1321 		if (vendor_class->method_table[i])
1322 			return 1;
1323 	return 0;
1324 }
1325 
1326 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1327 			   char *oui)
1328 {
1329 	int i;
1330 
1331 	for (i = 0; i < MAX_MGMT_OUI; i++)
1332 		/* Is there matching OUI for this vendor class ? */
1333 		if (!memcmp(vendor_class->oui[i], oui, 3))
1334 			return i;
1335 
1336 	return -1;
1337 }
1338 
1339 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1340 {
1341 	int i;
1342 
1343 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1344 		if (vendor->vendor_class[i])
1345 			return 1;
1346 
1347 	return 0;
1348 }
1349 
1350 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1351 				     struct ib_mad_agent_private *agent)
1352 {
1353 	int i;
1354 
1355 	/* Remove any methods for this mad agent */
1356 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1357 		if (method->agent[i] == agent) {
1358 			method->agent[i] = NULL;
1359 		}
1360 	}
1361 }
1362 
1363 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1364 			      struct ib_mad_agent_private *agent_priv,
1365 			      u8 mgmt_class)
1366 {
1367 	struct ib_mad_port_private *port_priv;
1368 	struct ib_mad_mgmt_class_table **class;
1369 	struct ib_mad_mgmt_method_table **method;
1370 	int i, ret;
1371 
1372 	port_priv = agent_priv->qp_info->port_priv;
1373 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1374 	if (!*class) {
1375 		/* Allocate management class table for "new" class version */
1376 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1377 		if (!*class) {
1378 			dev_err(&agent_priv->agent.device->dev,
1379 				"No memory for ib_mad_mgmt_class_table\n");
1380 			ret = -ENOMEM;
1381 			goto error1;
1382 		}
1383 
1384 		/* Allocate method table for this management class */
1385 		method = &(*class)->method_table[mgmt_class];
1386 		if ((ret = allocate_method_table(method)))
1387 			goto error2;
1388 	} else {
1389 		method = &(*class)->method_table[mgmt_class];
1390 		if (!*method) {
1391 			/* Allocate method table for this management class */
1392 			if ((ret = allocate_method_table(method)))
1393 				goto error1;
1394 		}
1395 	}
1396 
1397 	/* Now, make sure methods are not already in use */
1398 	if (method_in_use(method, mad_reg_req))
1399 		goto error3;
1400 
1401 	/* Finally, add in methods being registered */
1402 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1403 		(*method)->agent[i] = agent_priv;
1404 
1405 	return 0;
1406 
1407 error3:
1408 	/* Remove any methods for this mad agent */
1409 	remove_methods_mad_agent(*method, agent_priv);
1410 	/* Now, check to see if there are any methods in use */
1411 	if (!check_method_table(*method)) {
1412 		/* If not, release management method table */
1413 		kfree(*method);
1414 		*method = NULL;
1415 	}
1416 	ret = -EINVAL;
1417 	goto error1;
1418 error2:
1419 	kfree(*class);
1420 	*class = NULL;
1421 error1:
1422 	return ret;
1423 }
1424 
1425 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1426 			   struct ib_mad_agent_private *agent_priv)
1427 {
1428 	struct ib_mad_port_private *port_priv;
1429 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1430 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1431 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1432 	struct ib_mad_mgmt_method_table **method;
1433 	int i, ret = -ENOMEM;
1434 	u8 vclass;
1435 
1436 	/* "New" vendor (with OUI) class */
1437 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1438 	port_priv = agent_priv->qp_info->port_priv;
1439 	vendor_table = &port_priv->version[
1440 				mad_reg_req->mgmt_class_version].vendor;
1441 	if (!*vendor_table) {
1442 		/* Allocate mgmt vendor class table for "new" class version */
1443 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1444 		if (!vendor) {
1445 			dev_err(&agent_priv->agent.device->dev,
1446 				"No memory for ib_mad_mgmt_vendor_class_table\n");
1447 			goto error1;
1448 		}
1449 
1450 		*vendor_table = vendor;
1451 	}
1452 	if (!(*vendor_table)->vendor_class[vclass]) {
1453 		/* Allocate table for this management vendor class */
1454 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1455 		if (!vendor_class) {
1456 			dev_err(&agent_priv->agent.device->dev,
1457 				"No memory for ib_mad_mgmt_vendor_class\n");
1458 			goto error2;
1459 		}
1460 
1461 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1462 	}
1463 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1464 		/* Is there matching OUI for this vendor class ? */
1465 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1466 			    mad_reg_req->oui, 3)) {
1467 			method = &(*vendor_table)->vendor_class[
1468 						vclass]->method_table[i];
1469 			BUG_ON(!*method);
1470 			goto check_in_use;
1471 		}
1472 	}
1473 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1474 		/* OUI slot available ? */
1475 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1476 				vclass]->oui[i])) {
1477 			method = &(*vendor_table)->vendor_class[
1478 				vclass]->method_table[i];
1479 			BUG_ON(*method);
1480 			/* Allocate method table for this OUI */
1481 			if ((ret = allocate_method_table(method)))
1482 				goto error3;
1483 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1484 			       mad_reg_req->oui, 3);
1485 			goto check_in_use;
1486 		}
1487 	}
1488 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1489 	goto error3;
1490 
1491 check_in_use:
1492 	/* Now, make sure methods are not already in use */
1493 	if (method_in_use(method, mad_reg_req))
1494 		goto error4;
1495 
1496 	/* Finally, add in methods being registered */
1497 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1498 		(*method)->agent[i] = agent_priv;
1499 
1500 	return 0;
1501 
1502 error4:
1503 	/* Remove any methods for this mad agent */
1504 	remove_methods_mad_agent(*method, agent_priv);
1505 	/* Now, check to see if there are any methods in use */
1506 	if (!check_method_table(*method)) {
1507 		/* If not, release management method table */
1508 		kfree(*method);
1509 		*method = NULL;
1510 	}
1511 	ret = -EINVAL;
1512 error3:
1513 	if (vendor_class) {
1514 		(*vendor_table)->vendor_class[vclass] = NULL;
1515 		kfree(vendor_class);
1516 	}
1517 error2:
1518 	if (vendor) {
1519 		*vendor_table = NULL;
1520 		kfree(vendor);
1521 	}
1522 error1:
1523 	return ret;
1524 }
1525 
1526 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1527 {
1528 	struct ib_mad_port_private *port_priv;
1529 	struct ib_mad_mgmt_class_table *class;
1530 	struct ib_mad_mgmt_method_table *method;
1531 	struct ib_mad_mgmt_vendor_class_table *vendor;
1532 	struct ib_mad_mgmt_vendor_class *vendor_class;
1533 	int index;
1534 	u8 mgmt_class;
1535 
1536 	/*
1537 	 * Was MAD registration request supplied
1538 	 * with original registration ?
1539 	 */
1540 	if (!agent_priv->reg_req) {
1541 		goto out;
1542 	}
1543 
1544 	port_priv = agent_priv->qp_info->port_priv;
1545 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1546 	class = port_priv->version[
1547 			agent_priv->reg_req->mgmt_class_version].class;
1548 	if (!class)
1549 		goto vendor_check;
1550 
1551 	method = class->method_table[mgmt_class];
1552 	if (method) {
1553 		/* Remove any methods for this mad agent */
1554 		remove_methods_mad_agent(method, agent_priv);
1555 		/* Now, check to see if there are any methods still in use */
1556 		if (!check_method_table(method)) {
1557 			/* If not, release management method table */
1558 			 kfree(method);
1559 			 class->method_table[mgmt_class] = NULL;
1560 			 /* Any management classes left ? */
1561 			if (!check_class_table(class)) {
1562 				/* If not, release management class table */
1563 				kfree(class);
1564 				port_priv->version[
1565 					agent_priv->reg_req->
1566 					mgmt_class_version].class = NULL;
1567 			}
1568 		}
1569 	}
1570 
1571 vendor_check:
1572 	if (!is_vendor_class(mgmt_class))
1573 		goto out;
1574 
1575 	/* normalize mgmt_class to vendor range 2 */
1576 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1577 	vendor = port_priv->version[
1578 			agent_priv->reg_req->mgmt_class_version].vendor;
1579 
1580 	if (!vendor)
1581 		goto out;
1582 
1583 	vendor_class = vendor->vendor_class[mgmt_class];
1584 	if (vendor_class) {
1585 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1586 		if (index < 0)
1587 			goto out;
1588 		method = vendor_class->method_table[index];
1589 		if (method) {
1590 			/* Remove any methods for this mad agent */
1591 			remove_methods_mad_agent(method, agent_priv);
1592 			/*
1593 			 * Now, check to see if there are
1594 			 * any methods still in use
1595 			 */
1596 			if (!check_method_table(method)) {
1597 				/* If not, release management method table */
1598 				kfree(method);
1599 				vendor_class->method_table[index] = NULL;
1600 				memset(vendor_class->oui[index], 0, 3);
1601 				/* Any OUIs left ? */
1602 				if (!check_vendor_class(vendor_class)) {
1603 					/* If not, release vendor class table */
1604 					kfree(vendor_class);
1605 					vendor->vendor_class[mgmt_class] = NULL;
1606 					/* Any other vendor classes left ? */
1607 					if (!check_vendor_table(vendor)) {
1608 						kfree(vendor);
1609 						port_priv->version[
1610 							agent_priv->reg_req->
1611 							mgmt_class_version].
1612 							vendor = NULL;
1613 					}
1614 				}
1615 			}
1616 		}
1617 	}
1618 
1619 out:
1620 	return;
1621 }
1622 
1623 static struct ib_mad_agent_private *
1624 find_mad_agent(struct ib_mad_port_private *port_priv,
1625 	       struct ib_mad *mad)
1626 {
1627 	struct ib_mad_agent_private *mad_agent = NULL;
1628 	unsigned long flags;
1629 
1630 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1631 	if (ib_response_mad(mad)) {
1632 		u32 hi_tid;
1633 		struct ib_mad_agent_private *entry;
1634 
1635 		/*
1636 		 * Routing is based on high 32 bits of transaction ID
1637 		 * of MAD.
1638 		 */
1639 		hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1640 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1641 			if (entry->agent.hi_tid == hi_tid) {
1642 				mad_agent = entry;
1643 				break;
1644 			}
1645 		}
1646 	} else {
1647 		struct ib_mad_mgmt_class_table *class;
1648 		struct ib_mad_mgmt_method_table *method;
1649 		struct ib_mad_mgmt_vendor_class_table *vendor;
1650 		struct ib_mad_mgmt_vendor_class *vendor_class;
1651 		struct ib_vendor_mad *vendor_mad;
1652 		int index;
1653 
1654 		/*
1655 		 * Routing is based on version, class, and method
1656 		 * For "newer" vendor MADs, also based on OUI
1657 		 */
1658 		if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1659 			goto out;
1660 		if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1661 			class = port_priv->version[
1662 					mad->mad_hdr.class_version].class;
1663 			if (!class)
1664 				goto out;
1665 			if (convert_mgmt_class(mad->mad_hdr.mgmt_class) >=
1666 			    IB_MGMT_MAX_METHODS)
1667 				goto out;
1668 			method = class->method_table[convert_mgmt_class(
1669 							mad->mad_hdr.mgmt_class)];
1670 			if (method)
1671 				mad_agent = method->agent[mad->mad_hdr.method &
1672 							  ~IB_MGMT_METHOD_RESP];
1673 		} else {
1674 			vendor = port_priv->version[
1675 					mad->mad_hdr.class_version].vendor;
1676 			if (!vendor)
1677 				goto out;
1678 			vendor_class = vendor->vendor_class[vendor_class_index(
1679 						mad->mad_hdr.mgmt_class)];
1680 			if (!vendor_class)
1681 				goto out;
1682 			/* Find matching OUI */
1683 			vendor_mad = (struct ib_vendor_mad *)mad;
1684 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1685 			if (index == -1)
1686 				goto out;
1687 			method = vendor_class->method_table[index];
1688 			if (method) {
1689 				mad_agent = method->agent[mad->mad_hdr.method &
1690 							  ~IB_MGMT_METHOD_RESP];
1691 			}
1692 		}
1693 	}
1694 
1695 	if (mad_agent) {
1696 		if (mad_agent->agent.recv_handler)
1697 			atomic_inc(&mad_agent->refcount);
1698 		else {
1699 			dev_notice(&port_priv->device->dev,
1700 				   "No receive handler for client %p on port %d\n",
1701 				   &mad_agent->agent, port_priv->port_num);
1702 			mad_agent = NULL;
1703 		}
1704 	}
1705 out:
1706 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1707 
1708 	return mad_agent;
1709 }
1710 
1711 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1712 {
1713 	int valid = 0;
1714 
1715 	/* Make sure MAD base version is understood */
1716 	if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1717 		pr_err("MAD received with unsupported base version %d\n",
1718 			mad->mad_hdr.base_version);
1719 		goto out;
1720 	}
1721 
1722 	/* Filter SMI packets sent to other than QP0 */
1723 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1724 	    (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1725 		if (qp_num == 0)
1726 			valid = 1;
1727 	} else {
1728 		/* Filter GSI packets sent to QP0 */
1729 		if (qp_num != 0)
1730 			valid = 1;
1731 	}
1732 
1733 out:
1734 	return valid;
1735 }
1736 
1737 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1738 		       struct ib_mad_hdr *mad_hdr)
1739 {
1740 	struct ib_rmpp_mad *rmpp_mad;
1741 
1742 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1743 	return !mad_agent_priv->agent.rmpp_version ||
1744 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1745 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1746 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1747 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1748 }
1749 
1750 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1751 				     struct ib_mad_recv_wc *rwc)
1752 {
1753 	return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1754 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1755 }
1756 
1757 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1758 				   struct ib_mad_send_wr_private *wr,
1759 				   struct ib_mad_recv_wc *rwc )
1760 {
1761 	struct ib_ah_attr attr;
1762 	u8 send_resp, rcv_resp;
1763 	union ib_gid sgid;
1764 	struct ib_device *device = mad_agent_priv->agent.device;
1765 	u8 port_num = mad_agent_priv->agent.port_num;
1766 	u8 lmc;
1767 
1768 	send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1769 	rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1770 
1771 	if (send_resp == rcv_resp)
1772 		/* both requests, or both responses. GIDs different */
1773 		return 0;
1774 
1775 	if (ib_query_ah(wr->send_buf.ah, &attr))
1776 		/* Assume not equal, to avoid false positives. */
1777 		return 0;
1778 
1779 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1780 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1781 		/* one has GID, other does not.  Assume different */
1782 		return 0;
1783 
1784 	if (!send_resp && rcv_resp) {
1785 		/* is request/response. */
1786 		if (!(attr.ah_flags & IB_AH_GRH)) {
1787 			if (ib_get_cached_lmc(device, port_num, &lmc))
1788 				return 0;
1789 			return (!lmc || !((attr.src_path_bits ^
1790 					   rwc->wc->dlid_path_bits) &
1791 					  ((1 << lmc) - 1)));
1792 		} else {
1793 			if (ib_get_cached_gid(device, port_num,
1794 					      attr.grh.sgid_index, &sgid))
1795 				return 0;
1796 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1797 				       16);
1798 		}
1799 	}
1800 
1801 	if (!(attr.ah_flags & IB_AH_GRH))
1802 		return attr.dlid == rwc->wc->slid;
1803 	else
1804 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1805 			       16);
1806 }
1807 
1808 static inline int is_direct(u8 class)
1809 {
1810 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1811 }
1812 
1813 struct ib_mad_send_wr_private*
1814 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1815 		 struct ib_mad_recv_wc *wc)
1816 {
1817 	struct ib_mad_send_wr_private *wr;
1818 	struct ib_mad *mad;
1819 
1820 	mad = (struct ib_mad *)wc->recv_buf.mad;
1821 
1822 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1823 		if ((wr->tid == mad->mad_hdr.tid) &&
1824 		    rcv_has_same_class(wr, wc) &&
1825 		    /*
1826 		     * Don't check GID for direct routed MADs.
1827 		     * These might have permissive LIDs.
1828 		     */
1829 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1830 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1831 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1832 	}
1833 
1834 	/*
1835 	 * It's possible to receive the response before we've
1836 	 * been notified that the send has completed
1837 	 */
1838 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1839 		if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1840 		    wr->tid == mad->mad_hdr.tid &&
1841 		    wr->timeout &&
1842 		    rcv_has_same_class(wr, wc) &&
1843 		    /*
1844 		     * Don't check GID for direct routed MADs.
1845 		     * These might have permissive LIDs.
1846 		     */
1847 		    (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1848 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1849 			/* Verify request has not been canceled */
1850 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1851 	}
1852 	return NULL;
1853 }
1854 
1855 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1856 {
1857 	mad_send_wr->timeout = 0;
1858 	if (mad_send_wr->refcount == 1)
1859 		list_move_tail(&mad_send_wr->agent_list,
1860 			      &mad_send_wr->mad_agent_priv->done_list);
1861 }
1862 
1863 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1864 				 struct ib_mad_recv_wc *mad_recv_wc)
1865 {
1866 	struct ib_mad_send_wr_private *mad_send_wr;
1867 	struct ib_mad_send_wc mad_send_wc;
1868 	unsigned long flags;
1869 
1870 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1871 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1872 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1873 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1874 						      mad_recv_wc);
1875 		if (!mad_recv_wc) {
1876 			deref_mad_agent(mad_agent_priv);
1877 			return;
1878 		}
1879 	}
1880 
1881 	/* Complete corresponding request */
1882 	if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1883 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1884 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1885 		if (!mad_send_wr) {
1886 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1887 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1888 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1889 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1890 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1891 				/* user rmpp is in effect
1892 				 * and this is an active RMPP MAD
1893 				 */
1894 				mad_recv_wc->wc->wr_id = 0;
1895 				mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1896 								   mad_recv_wc);
1897 				atomic_dec(&mad_agent_priv->refcount);
1898 			} else {
1899 				/* not user rmpp, revert to normal behavior and
1900 				 * drop the mad */
1901 				ib_free_recv_mad(mad_recv_wc);
1902 				deref_mad_agent(mad_agent_priv);
1903 				return;
1904 			}
1905 		} else {
1906 			ib_mark_mad_done(mad_send_wr);
1907 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1908 
1909 			/* Defined behavior is to complete response before request */
1910 			mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1911 			mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1912 							   mad_recv_wc);
1913 			atomic_dec(&mad_agent_priv->refcount);
1914 
1915 			mad_send_wc.status = IB_WC_SUCCESS;
1916 			mad_send_wc.vendor_err = 0;
1917 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
1918 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1919 		}
1920 	} else {
1921 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1922 						   mad_recv_wc);
1923 		deref_mad_agent(mad_agent_priv);
1924 	}
1925 }
1926 
1927 static bool generate_unmatched_resp(struct ib_mad_private *recv,
1928 				    struct ib_mad_private *response)
1929 {
1930 	if (recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_GET ||
1931 	    recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_SET) {
1932 		memcpy(response, recv, sizeof *response);
1933 		response->header.recv_wc.wc = &response->header.wc;
1934 		response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1935 		response->header.recv_wc.recv_buf.grh = &response->grh;
1936 		response->mad.mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
1937 		response->mad.mad.mad_hdr.status =
1938 			cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
1939 		if (recv->mad.mad.mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1940 			response->mad.mad.mad_hdr.status |= IB_SMP_DIRECTION;
1941 
1942 		return true;
1943 	} else {
1944 		return false;
1945 	}
1946 }
1947 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1948 				     struct ib_wc *wc)
1949 {
1950 	struct ib_mad_qp_info *qp_info;
1951 	struct ib_mad_private_header *mad_priv_hdr;
1952 	struct ib_mad_private *recv, *response = NULL;
1953 	struct ib_mad_list_head *mad_list;
1954 	struct ib_mad_agent_private *mad_agent;
1955 	int port_num;
1956 	int ret = IB_MAD_RESULT_SUCCESS;
1957 
1958 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1959 	qp_info = mad_list->mad_queue->qp_info;
1960 	dequeue_mad(mad_list);
1961 
1962 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1963 				    mad_list);
1964 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1965 	ib_dma_unmap_single(port_priv->device,
1966 			    recv->header.mapping,
1967 			    sizeof(struct ib_mad_private) -
1968 			      sizeof(struct ib_mad_private_header),
1969 			    DMA_FROM_DEVICE);
1970 
1971 	/* Setup MAD receive work completion from "normal" work completion */
1972 	recv->header.wc = *wc;
1973 	recv->header.recv_wc.wc = &recv->header.wc;
1974 	recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1975 	recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1976 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
1977 
1978 	if (atomic_read(&qp_info->snoop_count))
1979 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1980 
1981 	/* Validate MAD */
1982 	if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1983 		goto out;
1984 
1985 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1986 	if (!response) {
1987 		dev_err(&port_priv->device->dev,
1988 			"ib_mad_recv_done_handler no memory for response buffer\n");
1989 		goto out;
1990 	}
1991 
1992 	if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1993 		port_num = wc->port_num;
1994 	else
1995 		port_num = port_priv->port_num;
1996 
1997 	if (recv->mad.mad.mad_hdr.mgmt_class ==
1998 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1999 		enum smi_forward_action retsmi;
2000 
2001 		if (smi_handle_dr_smp_recv(&recv->mad.smp,
2002 					   port_priv->device->node_type,
2003 					   port_num,
2004 					   port_priv->device->phys_port_cnt) ==
2005 					   IB_SMI_DISCARD)
2006 			goto out;
2007 
2008 		retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
2009 		if (retsmi == IB_SMI_LOCAL)
2010 			goto local;
2011 
2012 		if (retsmi == IB_SMI_SEND) { /* don't forward */
2013 			if (smi_handle_dr_smp_send(&recv->mad.smp,
2014 						   port_priv->device->node_type,
2015 						   port_num) == IB_SMI_DISCARD)
2016 				goto out;
2017 
2018 			if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
2019 				goto out;
2020 		} else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
2021 			/* forward case for switches */
2022 			memcpy(response, recv, sizeof(*response));
2023 			response->header.recv_wc.wc = &response->header.wc;
2024 			response->header.recv_wc.recv_buf.mad = &response->mad.mad;
2025 			response->header.recv_wc.recv_buf.grh = &response->grh;
2026 
2027 			agent_send_response(&response->mad.mad,
2028 					    &response->grh, wc,
2029 					    port_priv->device,
2030 					    smi_get_fwd_port(&recv->mad.smp),
2031 					    qp_info->qp->qp_num);
2032 
2033 			goto out;
2034 		}
2035 	}
2036 
2037 local:
2038 	/* Give driver "right of first refusal" on incoming MAD */
2039 	if (port_priv->device->process_mad) {
2040 		ret = port_priv->device->process_mad(port_priv->device, 0,
2041 						     port_priv->port_num,
2042 						     wc, &recv->grh,
2043 						     &recv->mad.mad,
2044 						     &response->mad.mad);
2045 		if (ret & IB_MAD_RESULT_SUCCESS) {
2046 			if (ret & IB_MAD_RESULT_CONSUMED)
2047 				goto out;
2048 			if (ret & IB_MAD_RESULT_REPLY) {
2049 				agent_send_response(&response->mad.mad,
2050 						    &recv->grh, wc,
2051 						    port_priv->device,
2052 						    port_num,
2053 						    qp_info->qp->qp_num);
2054 				goto out;
2055 			}
2056 		}
2057 	}
2058 
2059 	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
2060 	if (mad_agent) {
2061 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2062 		/*
2063 		 * recv is freed up in error cases in ib_mad_complete_recv
2064 		 * or via recv_handler in ib_mad_complete_recv()
2065 		 */
2066 		recv = NULL;
2067 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2068 		   generate_unmatched_resp(recv, response)) {
2069 		agent_send_response(&response->mad.mad, &recv->grh, wc,
2070 				    port_priv->device, port_num, qp_info->qp->qp_num);
2071 	}
2072 
2073 out:
2074 	/* Post another receive request for this QP */
2075 	if (response) {
2076 		ib_mad_post_receive_mads(qp_info, response);
2077 		if (recv)
2078 			kmem_cache_free(ib_mad_cache, recv);
2079 	} else
2080 		ib_mad_post_receive_mads(qp_info, recv);
2081 }
2082 
2083 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2084 {
2085 	struct ib_mad_send_wr_private *mad_send_wr;
2086 	unsigned long delay;
2087 
2088 	if (list_empty(&mad_agent_priv->wait_list)) {
2089 		cancel_delayed_work(&mad_agent_priv->timed_work);
2090 	} else {
2091 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2092 					 struct ib_mad_send_wr_private,
2093 					 agent_list);
2094 
2095 		if (time_after(mad_agent_priv->timeout,
2096 			       mad_send_wr->timeout)) {
2097 			mad_agent_priv->timeout = mad_send_wr->timeout;
2098 			delay = mad_send_wr->timeout - jiffies;
2099 			if ((long)delay <= 0)
2100 				delay = 1;
2101 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2102 					 &mad_agent_priv->timed_work, delay);
2103 		}
2104 	}
2105 }
2106 
2107 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2108 {
2109 	struct ib_mad_agent_private *mad_agent_priv;
2110 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2111 	struct list_head *list_item;
2112 	unsigned long delay;
2113 
2114 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2115 	list_del(&mad_send_wr->agent_list);
2116 
2117 	delay = mad_send_wr->timeout;
2118 	mad_send_wr->timeout += jiffies;
2119 
2120 	if (delay) {
2121 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2122 			temp_mad_send_wr = list_entry(list_item,
2123 						struct ib_mad_send_wr_private,
2124 						agent_list);
2125 			if (time_after(mad_send_wr->timeout,
2126 				       temp_mad_send_wr->timeout))
2127 				break;
2128 		}
2129 	}
2130 	else
2131 		list_item = &mad_agent_priv->wait_list;
2132 	list_add(&mad_send_wr->agent_list, list_item);
2133 
2134 	/* Reschedule a work item if we have a shorter timeout */
2135 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2136 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2137 				 &mad_agent_priv->timed_work, delay);
2138 }
2139 
2140 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2141 			  int timeout_ms)
2142 {
2143 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2144 	wait_for_response(mad_send_wr);
2145 }
2146 
2147 /*
2148  * Process a send work completion
2149  */
2150 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2151 			     struct ib_mad_send_wc *mad_send_wc)
2152 {
2153 	struct ib_mad_agent_private	*mad_agent_priv;
2154 	unsigned long			flags;
2155 	int				ret;
2156 
2157 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2158 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2159 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2160 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2161 		if (ret == IB_RMPP_RESULT_CONSUMED)
2162 			goto done;
2163 	} else
2164 		ret = IB_RMPP_RESULT_UNHANDLED;
2165 
2166 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2167 	    mad_send_wr->status == IB_WC_SUCCESS) {
2168 		mad_send_wr->status = mad_send_wc->status;
2169 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2170 	}
2171 
2172 	if (--mad_send_wr->refcount > 0) {
2173 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2174 		    mad_send_wr->status == IB_WC_SUCCESS) {
2175 			wait_for_response(mad_send_wr);
2176 		}
2177 		goto done;
2178 	}
2179 
2180 	/* Remove send from MAD agent and notify client of completion */
2181 	list_del(&mad_send_wr->agent_list);
2182 	adjust_timeout(mad_agent_priv);
2183 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2184 
2185 	if (mad_send_wr->status != IB_WC_SUCCESS )
2186 		mad_send_wc->status = mad_send_wr->status;
2187 	if (ret == IB_RMPP_RESULT_INTERNAL)
2188 		ib_rmpp_send_handler(mad_send_wc);
2189 	else
2190 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2191 						   mad_send_wc);
2192 
2193 	/* Release reference on agent taken when sending */
2194 	deref_mad_agent(mad_agent_priv);
2195 	return;
2196 done:
2197 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2198 }
2199 
2200 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2201 				     struct ib_wc *wc)
2202 {
2203 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2204 	struct ib_mad_list_head		*mad_list;
2205 	struct ib_mad_qp_info		*qp_info;
2206 	struct ib_mad_queue		*send_queue;
2207 	struct ib_send_wr		*bad_send_wr;
2208 	struct ib_mad_send_wc		mad_send_wc;
2209 	unsigned long flags;
2210 	int ret;
2211 
2212 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2213 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2214 				   mad_list);
2215 	send_queue = mad_list->mad_queue;
2216 	qp_info = send_queue->qp_info;
2217 
2218 retry:
2219 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2220 			    mad_send_wr->header_mapping,
2221 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2222 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2223 			    mad_send_wr->payload_mapping,
2224 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2225 	queued_send_wr = NULL;
2226 	spin_lock_irqsave(&send_queue->lock, flags);
2227 	list_del(&mad_list->list);
2228 
2229 	/* Move queued send to the send queue */
2230 	if (send_queue->count-- > send_queue->max_active) {
2231 		mad_list = container_of(qp_info->overflow_list.next,
2232 					struct ib_mad_list_head, list);
2233 		queued_send_wr = container_of(mad_list,
2234 					struct ib_mad_send_wr_private,
2235 					mad_list);
2236 		list_move_tail(&mad_list->list, &send_queue->list);
2237 	}
2238 	spin_unlock_irqrestore(&send_queue->lock, flags);
2239 
2240 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2241 	mad_send_wc.status = wc->status;
2242 	mad_send_wc.vendor_err = wc->vendor_err;
2243 	if (atomic_read(&qp_info->snoop_count))
2244 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2245 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2246 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2247 
2248 	if (queued_send_wr) {
2249 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2250 				   &bad_send_wr);
2251 		if (ret) {
2252 			dev_err(&port_priv->device->dev,
2253 				"ib_post_send failed: %d\n", ret);
2254 			mad_send_wr = queued_send_wr;
2255 			wc->status = IB_WC_LOC_QP_OP_ERR;
2256 			goto retry;
2257 		}
2258 	}
2259 }
2260 
2261 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2262 {
2263 	struct ib_mad_send_wr_private *mad_send_wr;
2264 	struct ib_mad_list_head *mad_list;
2265 	unsigned long flags;
2266 
2267 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2268 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2269 		mad_send_wr = container_of(mad_list,
2270 					   struct ib_mad_send_wr_private,
2271 					   mad_list);
2272 		mad_send_wr->retry = 1;
2273 	}
2274 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2275 }
2276 
2277 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2278 			      struct ib_wc *wc)
2279 {
2280 	struct ib_mad_list_head *mad_list;
2281 	struct ib_mad_qp_info *qp_info;
2282 	struct ib_mad_send_wr_private *mad_send_wr;
2283 	int ret;
2284 
2285 	/* Determine if failure was a send or receive */
2286 	mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2287 	qp_info = mad_list->mad_queue->qp_info;
2288 	if (mad_list->mad_queue == &qp_info->recv_queue)
2289 		/*
2290 		 * Receive errors indicate that the QP has entered the error
2291 		 * state - error handling/shutdown code will cleanup
2292 		 */
2293 		return;
2294 
2295 	/*
2296 	 * Send errors will transition the QP to SQE - move
2297 	 * QP to RTS and repost flushed work requests
2298 	 */
2299 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2300 				   mad_list);
2301 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2302 		if (mad_send_wr->retry) {
2303 			/* Repost send */
2304 			struct ib_send_wr *bad_send_wr;
2305 
2306 			mad_send_wr->retry = 0;
2307 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2308 					&bad_send_wr);
2309 			if (ret)
2310 				ib_mad_send_done_handler(port_priv, wc);
2311 		} else
2312 			ib_mad_send_done_handler(port_priv, wc);
2313 	} else {
2314 		struct ib_qp_attr *attr;
2315 
2316 		/* Transition QP to RTS and fail offending send */
2317 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2318 		if (attr) {
2319 			attr->qp_state = IB_QPS_RTS;
2320 			attr->cur_qp_state = IB_QPS_SQE;
2321 			ret = ib_modify_qp(qp_info->qp, attr,
2322 					   IB_QP_STATE | IB_QP_CUR_STATE);
2323 			kfree(attr);
2324 			if (ret)
2325 				dev_err(&port_priv->device->dev,
2326 					"mad_error_handler - ib_modify_qp to RTS : %d\n",
2327 					ret);
2328 			else
2329 				mark_sends_for_retry(qp_info);
2330 		}
2331 		ib_mad_send_done_handler(port_priv, wc);
2332 	}
2333 }
2334 
2335 /*
2336  * IB MAD completion callback
2337  */
2338 static void ib_mad_completion_handler(struct work_struct *work)
2339 {
2340 	struct ib_mad_port_private *port_priv;
2341 	struct ib_wc wc;
2342 
2343 	port_priv = container_of(work, struct ib_mad_port_private, work);
2344 	ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2345 
2346 	while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2347 		if (wc.status == IB_WC_SUCCESS) {
2348 			switch (wc.opcode) {
2349 			case IB_WC_SEND:
2350 				ib_mad_send_done_handler(port_priv, &wc);
2351 				break;
2352 			case IB_WC_RECV:
2353 				ib_mad_recv_done_handler(port_priv, &wc);
2354 				break;
2355 			default:
2356 				BUG_ON(1);
2357 				break;
2358 			}
2359 		} else
2360 			mad_error_handler(port_priv, &wc);
2361 	}
2362 }
2363 
2364 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2365 {
2366 	unsigned long flags;
2367 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2368 	struct ib_mad_send_wc mad_send_wc;
2369 	struct list_head cancel_list;
2370 
2371 	INIT_LIST_HEAD(&cancel_list);
2372 
2373 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2374 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2375 				 &mad_agent_priv->send_list, agent_list) {
2376 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2377 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2378 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2379 		}
2380 	}
2381 
2382 	/* Empty wait list to prevent receives from finding a request */
2383 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2384 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2385 
2386 	/* Report all cancelled requests */
2387 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2388 	mad_send_wc.vendor_err = 0;
2389 
2390 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2391 				 &cancel_list, agent_list) {
2392 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2393 		list_del(&mad_send_wr->agent_list);
2394 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2395 						   &mad_send_wc);
2396 		atomic_dec(&mad_agent_priv->refcount);
2397 	}
2398 }
2399 
2400 static struct ib_mad_send_wr_private*
2401 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2402 	     struct ib_mad_send_buf *send_buf)
2403 {
2404 	struct ib_mad_send_wr_private *mad_send_wr;
2405 
2406 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2407 			    agent_list) {
2408 		if (&mad_send_wr->send_buf == send_buf)
2409 			return mad_send_wr;
2410 	}
2411 
2412 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2413 			    agent_list) {
2414 		if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2415 		    &mad_send_wr->send_buf == send_buf)
2416 			return mad_send_wr;
2417 	}
2418 	return NULL;
2419 }
2420 
2421 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2422 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2423 {
2424 	struct ib_mad_agent_private *mad_agent_priv;
2425 	struct ib_mad_send_wr_private *mad_send_wr;
2426 	unsigned long flags;
2427 	int active;
2428 
2429 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2430 				      agent);
2431 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2432 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2433 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2434 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2435 		return -EINVAL;
2436 	}
2437 
2438 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2439 	if (!timeout_ms) {
2440 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2441 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2442 	}
2443 
2444 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2445 	if (active)
2446 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2447 	else
2448 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2449 
2450 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2451 	return 0;
2452 }
2453 EXPORT_SYMBOL(ib_modify_mad);
2454 
2455 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2456 		   struct ib_mad_send_buf *send_buf)
2457 {
2458 	ib_modify_mad(mad_agent, send_buf, 0);
2459 }
2460 EXPORT_SYMBOL(ib_cancel_mad);
2461 
2462 static void local_completions(struct work_struct *work)
2463 {
2464 	struct ib_mad_agent_private *mad_agent_priv;
2465 	struct ib_mad_local_private *local;
2466 	struct ib_mad_agent_private *recv_mad_agent;
2467 	unsigned long flags;
2468 	int free_mad;
2469 	struct ib_wc wc;
2470 	struct ib_mad_send_wc mad_send_wc;
2471 
2472 	mad_agent_priv =
2473 		container_of(work, struct ib_mad_agent_private, local_work);
2474 
2475 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2476 	while (!list_empty(&mad_agent_priv->local_list)) {
2477 		local = list_entry(mad_agent_priv->local_list.next,
2478 				   struct ib_mad_local_private,
2479 				   completion_list);
2480 		list_del(&local->completion_list);
2481 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2482 		free_mad = 0;
2483 		if (local->mad_priv) {
2484 			recv_mad_agent = local->recv_mad_agent;
2485 			if (!recv_mad_agent) {
2486 				dev_err(&mad_agent_priv->agent.device->dev,
2487 					"No receive MAD agent for local completion\n");
2488 				free_mad = 1;
2489 				goto local_send_completion;
2490 			}
2491 
2492 			/*
2493 			 * Defined behavior is to complete response
2494 			 * before request
2495 			 */
2496 			build_smp_wc(recv_mad_agent->agent.qp,
2497 				     (unsigned long) local->mad_send_wr,
2498 				     be16_to_cpu(IB_LID_PERMISSIVE),
2499 				     0, recv_mad_agent->agent.port_num, &wc);
2500 
2501 			local->mad_priv->header.recv_wc.wc = &wc;
2502 			local->mad_priv->header.recv_wc.mad_len =
2503 						sizeof(struct ib_mad);
2504 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2505 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2506 				 &local->mad_priv->header.recv_wc.rmpp_list);
2507 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2508 			local->mad_priv->header.recv_wc.recv_buf.mad =
2509 						&local->mad_priv->mad.mad;
2510 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2511 				snoop_recv(recv_mad_agent->qp_info,
2512 					  &local->mad_priv->header.recv_wc,
2513 					   IB_MAD_SNOOP_RECVS);
2514 			recv_mad_agent->agent.recv_handler(
2515 						&recv_mad_agent->agent,
2516 						&local->mad_priv->header.recv_wc);
2517 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2518 			atomic_dec(&recv_mad_agent->refcount);
2519 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2520 		}
2521 
2522 local_send_completion:
2523 		/* Complete send */
2524 		mad_send_wc.status = IB_WC_SUCCESS;
2525 		mad_send_wc.vendor_err = 0;
2526 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2527 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2528 			snoop_send(mad_agent_priv->qp_info,
2529 				   &local->mad_send_wr->send_buf,
2530 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2531 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2532 						   &mad_send_wc);
2533 
2534 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2535 		atomic_dec(&mad_agent_priv->refcount);
2536 		if (free_mad)
2537 			kmem_cache_free(ib_mad_cache, local->mad_priv);
2538 		kfree(local);
2539 	}
2540 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2541 }
2542 
2543 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2544 {
2545 	int ret;
2546 
2547 	if (!mad_send_wr->retries_left)
2548 		return -ETIMEDOUT;
2549 
2550 	mad_send_wr->retries_left--;
2551 	mad_send_wr->send_buf.retries++;
2552 
2553 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2554 
2555 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2556 		ret = ib_retry_rmpp(mad_send_wr);
2557 		switch (ret) {
2558 		case IB_RMPP_RESULT_UNHANDLED:
2559 			ret = ib_send_mad(mad_send_wr);
2560 			break;
2561 		case IB_RMPP_RESULT_CONSUMED:
2562 			ret = 0;
2563 			break;
2564 		default:
2565 			ret = -ECOMM;
2566 			break;
2567 		}
2568 	} else
2569 		ret = ib_send_mad(mad_send_wr);
2570 
2571 	if (!ret) {
2572 		mad_send_wr->refcount++;
2573 		list_add_tail(&mad_send_wr->agent_list,
2574 			      &mad_send_wr->mad_agent_priv->send_list);
2575 	}
2576 	return ret;
2577 }
2578 
2579 static void timeout_sends(struct work_struct *work)
2580 {
2581 	struct ib_mad_agent_private *mad_agent_priv;
2582 	struct ib_mad_send_wr_private *mad_send_wr;
2583 	struct ib_mad_send_wc mad_send_wc;
2584 	unsigned long flags, delay;
2585 
2586 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2587 				      timed_work.work);
2588 	mad_send_wc.vendor_err = 0;
2589 
2590 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2591 	while (!list_empty(&mad_agent_priv->wait_list)) {
2592 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2593 					 struct ib_mad_send_wr_private,
2594 					 agent_list);
2595 
2596 		if (time_after(mad_send_wr->timeout, jiffies)) {
2597 			delay = mad_send_wr->timeout - jiffies;
2598 			if ((long)delay <= 0)
2599 				delay = 1;
2600 			queue_delayed_work(mad_agent_priv->qp_info->
2601 					   port_priv->wq,
2602 					   &mad_agent_priv->timed_work, delay);
2603 			break;
2604 		}
2605 
2606 		list_del(&mad_send_wr->agent_list);
2607 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2608 		    !retry_send(mad_send_wr))
2609 			continue;
2610 
2611 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2612 
2613 		if (mad_send_wr->status == IB_WC_SUCCESS)
2614 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2615 		else
2616 			mad_send_wc.status = mad_send_wr->status;
2617 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2618 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2619 						   &mad_send_wc);
2620 
2621 		atomic_dec(&mad_agent_priv->refcount);
2622 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2623 	}
2624 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2625 }
2626 
2627 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2628 {
2629 	struct ib_mad_port_private *port_priv = cq->cq_context;
2630 	unsigned long flags;
2631 
2632 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2633 	if (!list_empty(&port_priv->port_list))
2634 		queue_work(port_priv->wq, &port_priv->work);
2635 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2636 }
2637 
2638 /*
2639  * Allocate receive MADs and post receive WRs for them
2640  */
2641 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2642 				    struct ib_mad_private *mad)
2643 {
2644 	unsigned long flags;
2645 	int post, ret;
2646 	struct ib_mad_private *mad_priv;
2647 	struct ib_sge sg_list;
2648 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2649 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2650 
2651 	/* Initialize common scatter list fields */
2652 	sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2653 	sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2654 
2655 	/* Initialize common receive WR fields */
2656 	recv_wr.next = NULL;
2657 	recv_wr.sg_list = &sg_list;
2658 	recv_wr.num_sge = 1;
2659 
2660 	do {
2661 		/* Allocate and map receive buffer */
2662 		if (mad) {
2663 			mad_priv = mad;
2664 			mad = NULL;
2665 		} else {
2666 			mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2667 			if (!mad_priv) {
2668 				dev_err(&qp_info->port_priv->device->dev,
2669 					"No memory for receive buffer\n");
2670 				ret = -ENOMEM;
2671 				break;
2672 			}
2673 		}
2674 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2675 						 &mad_priv->grh,
2676 						 sizeof *mad_priv -
2677 						   sizeof mad_priv->header,
2678 						 DMA_FROM_DEVICE);
2679 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2680 						  sg_list.addr))) {
2681 			ret = -ENOMEM;
2682 			break;
2683 		}
2684 		mad_priv->header.mapping = sg_list.addr;
2685 		recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2686 		mad_priv->header.mad_list.mad_queue = recv_queue;
2687 
2688 		/* Post receive WR */
2689 		spin_lock_irqsave(&recv_queue->lock, flags);
2690 		post = (++recv_queue->count < recv_queue->max_active);
2691 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2692 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2693 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2694 		if (ret) {
2695 			spin_lock_irqsave(&recv_queue->lock, flags);
2696 			list_del(&mad_priv->header.mad_list.list);
2697 			recv_queue->count--;
2698 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2699 			ib_dma_unmap_single(qp_info->port_priv->device,
2700 					    mad_priv->header.mapping,
2701 					    sizeof *mad_priv -
2702 					      sizeof mad_priv->header,
2703 					    DMA_FROM_DEVICE);
2704 			kmem_cache_free(ib_mad_cache, mad_priv);
2705 			dev_err(&qp_info->port_priv->device->dev,
2706 				"ib_post_recv failed: %d\n", ret);
2707 			break;
2708 		}
2709 	} while (post);
2710 
2711 	return ret;
2712 }
2713 
2714 /*
2715  * Return all the posted receive MADs
2716  */
2717 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2718 {
2719 	struct ib_mad_private_header *mad_priv_hdr;
2720 	struct ib_mad_private *recv;
2721 	struct ib_mad_list_head *mad_list;
2722 
2723 	if (!qp_info->qp)
2724 		return;
2725 
2726 	while (!list_empty(&qp_info->recv_queue.list)) {
2727 
2728 		mad_list = list_entry(qp_info->recv_queue.list.next,
2729 				      struct ib_mad_list_head, list);
2730 		mad_priv_hdr = container_of(mad_list,
2731 					    struct ib_mad_private_header,
2732 					    mad_list);
2733 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2734 				    header);
2735 
2736 		/* Remove from posted receive MAD list */
2737 		list_del(&mad_list->list);
2738 
2739 		ib_dma_unmap_single(qp_info->port_priv->device,
2740 				    recv->header.mapping,
2741 				    sizeof(struct ib_mad_private) -
2742 				      sizeof(struct ib_mad_private_header),
2743 				    DMA_FROM_DEVICE);
2744 		kmem_cache_free(ib_mad_cache, recv);
2745 	}
2746 
2747 	qp_info->recv_queue.count = 0;
2748 }
2749 
2750 /*
2751  * Start the port
2752  */
2753 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2754 {
2755 	int ret, i;
2756 	struct ib_qp_attr *attr;
2757 	struct ib_qp *qp;
2758 	u16 pkey_index;
2759 
2760 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2761 	if (!attr) {
2762 		dev_err(&port_priv->device->dev,
2763 			"Couldn't kmalloc ib_qp_attr\n");
2764 		return -ENOMEM;
2765 	}
2766 
2767 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2768 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2769 	if (ret)
2770 		pkey_index = 0;
2771 
2772 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2773 		qp = port_priv->qp_info[i].qp;
2774 		if (!qp)
2775 			continue;
2776 
2777 		/*
2778 		 * PKey index for QP1 is irrelevant but
2779 		 * one is needed for the Reset to Init transition
2780 		 */
2781 		attr->qp_state = IB_QPS_INIT;
2782 		attr->pkey_index = pkey_index;
2783 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2784 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2785 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2786 		if (ret) {
2787 			dev_err(&port_priv->device->dev,
2788 				"Couldn't change QP%d state to INIT: %d\n",
2789 				i, ret);
2790 			goto out;
2791 		}
2792 
2793 		attr->qp_state = IB_QPS_RTR;
2794 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2795 		if (ret) {
2796 			dev_err(&port_priv->device->dev,
2797 				"Couldn't change QP%d state to RTR: %d\n",
2798 				i, ret);
2799 			goto out;
2800 		}
2801 
2802 		attr->qp_state = IB_QPS_RTS;
2803 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2804 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2805 		if (ret) {
2806 			dev_err(&port_priv->device->dev,
2807 				"Couldn't change QP%d state to RTS: %d\n",
2808 				i, ret);
2809 			goto out;
2810 		}
2811 	}
2812 
2813 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2814 	if (ret) {
2815 		dev_err(&port_priv->device->dev,
2816 			"Failed to request completion notification: %d\n",
2817 			ret);
2818 		goto out;
2819 	}
2820 
2821 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2822 		if (!port_priv->qp_info[i].qp)
2823 			continue;
2824 
2825 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2826 		if (ret) {
2827 			dev_err(&port_priv->device->dev,
2828 				"Couldn't post receive WRs\n");
2829 			goto out;
2830 		}
2831 	}
2832 out:
2833 	kfree(attr);
2834 	return ret;
2835 }
2836 
2837 static void qp_event_handler(struct ib_event *event, void *qp_context)
2838 {
2839 	struct ib_mad_qp_info	*qp_info = qp_context;
2840 
2841 	/* It's worse than that! He's dead, Jim! */
2842 	dev_err(&qp_info->port_priv->device->dev,
2843 		"Fatal error (%d) on MAD QP (%d)\n",
2844 		event->event, qp_info->qp->qp_num);
2845 }
2846 
2847 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2848 			   struct ib_mad_queue *mad_queue)
2849 {
2850 	mad_queue->qp_info = qp_info;
2851 	mad_queue->count = 0;
2852 	spin_lock_init(&mad_queue->lock);
2853 	INIT_LIST_HEAD(&mad_queue->list);
2854 }
2855 
2856 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2857 			struct ib_mad_qp_info *qp_info)
2858 {
2859 	qp_info->port_priv = port_priv;
2860 	init_mad_queue(qp_info, &qp_info->send_queue);
2861 	init_mad_queue(qp_info, &qp_info->recv_queue);
2862 	INIT_LIST_HEAD(&qp_info->overflow_list);
2863 	spin_lock_init(&qp_info->snoop_lock);
2864 	qp_info->snoop_table = NULL;
2865 	qp_info->snoop_table_size = 0;
2866 	atomic_set(&qp_info->snoop_count, 0);
2867 }
2868 
2869 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2870 			 enum ib_qp_type qp_type)
2871 {
2872 	struct ib_qp_init_attr	qp_init_attr;
2873 	int ret;
2874 
2875 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2876 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2877 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2878 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2879 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
2880 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2881 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2882 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2883 	qp_init_attr.qp_type = qp_type;
2884 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2885 	qp_init_attr.qp_context = qp_info;
2886 	qp_init_attr.event_handler = qp_event_handler;
2887 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2888 	if (IS_ERR(qp_info->qp)) {
2889 		dev_err(&qp_info->port_priv->device->dev,
2890 			"Couldn't create ib_mad QP%d\n",
2891 			get_spl_qp_index(qp_type));
2892 		ret = PTR_ERR(qp_info->qp);
2893 		goto error;
2894 	}
2895 	/* Use minimum queue sizes unless the CQ is resized */
2896 	qp_info->send_queue.max_active = mad_sendq_size;
2897 	qp_info->recv_queue.max_active = mad_recvq_size;
2898 	return 0;
2899 
2900 error:
2901 	return ret;
2902 }
2903 
2904 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2905 {
2906 	if (!qp_info->qp)
2907 		return;
2908 
2909 	ib_destroy_qp(qp_info->qp);
2910 	kfree(qp_info->snoop_table);
2911 }
2912 
2913 /*
2914  * Open the port
2915  * Create the QP, PD, MR, and CQ if needed
2916  */
2917 static int ib_mad_port_open(struct ib_device *device,
2918 			    int port_num)
2919 {
2920 	int ret, cq_size;
2921 	struct ib_mad_port_private *port_priv;
2922 	unsigned long flags;
2923 	char name[sizeof "ib_mad123"];
2924 	int has_smi;
2925 
2926 	/* Create new device info */
2927 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2928 	if (!port_priv) {
2929 		dev_err(&device->dev, "No memory for ib_mad_port_private\n");
2930 		return -ENOMEM;
2931 	}
2932 
2933 	port_priv->device = device;
2934 	port_priv->port_num = port_num;
2935 	spin_lock_init(&port_priv->reg_lock);
2936 	INIT_LIST_HEAD(&port_priv->agent_list);
2937 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2938 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2939 
2940 	cq_size = mad_sendq_size + mad_recvq_size;
2941 	has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2942 	if (has_smi)
2943 		cq_size *= 2;
2944 
2945 	port_priv->cq = ib_create_cq(port_priv->device,
2946 				     ib_mad_thread_completion_handler,
2947 				     NULL, port_priv, cq_size, 0);
2948 	if (IS_ERR(port_priv->cq)) {
2949 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
2950 		ret = PTR_ERR(port_priv->cq);
2951 		goto error3;
2952 	}
2953 
2954 	port_priv->pd = ib_alloc_pd(device);
2955 	if (IS_ERR(port_priv->pd)) {
2956 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
2957 		ret = PTR_ERR(port_priv->pd);
2958 		goto error4;
2959 	}
2960 
2961 	port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2962 	if (IS_ERR(port_priv->mr)) {
2963 		dev_err(&device->dev, "Couldn't get ib_mad DMA MR\n");
2964 		ret = PTR_ERR(port_priv->mr);
2965 		goto error5;
2966 	}
2967 
2968 	if (has_smi) {
2969 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2970 		if (ret)
2971 			goto error6;
2972 	}
2973 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2974 	if (ret)
2975 		goto error7;
2976 
2977 	snprintf(name, sizeof name, "ib_mad%d", port_num);
2978 	port_priv->wq = create_singlethread_workqueue(name);
2979 	if (!port_priv->wq) {
2980 		ret = -ENOMEM;
2981 		goto error8;
2982 	}
2983 	INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2984 
2985 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2986 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2987 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2988 
2989 	ret = ib_mad_port_start(port_priv);
2990 	if (ret) {
2991 		dev_err(&device->dev, "Couldn't start port\n");
2992 		goto error9;
2993 	}
2994 
2995 	return 0;
2996 
2997 error9:
2998 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2999 	list_del_init(&port_priv->port_list);
3000 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3001 
3002 	destroy_workqueue(port_priv->wq);
3003 error8:
3004 	destroy_mad_qp(&port_priv->qp_info[1]);
3005 error7:
3006 	destroy_mad_qp(&port_priv->qp_info[0]);
3007 error6:
3008 	ib_dereg_mr(port_priv->mr);
3009 error5:
3010 	ib_dealloc_pd(port_priv->pd);
3011 error4:
3012 	ib_destroy_cq(port_priv->cq);
3013 	cleanup_recv_queue(&port_priv->qp_info[1]);
3014 	cleanup_recv_queue(&port_priv->qp_info[0]);
3015 error3:
3016 	kfree(port_priv);
3017 
3018 	return ret;
3019 }
3020 
3021 /*
3022  * Close the port
3023  * If there are no classes using the port, free the port
3024  * resources (CQ, MR, PD, QP) and remove the port's info structure
3025  */
3026 static int ib_mad_port_close(struct ib_device *device, int port_num)
3027 {
3028 	struct ib_mad_port_private *port_priv;
3029 	unsigned long flags;
3030 
3031 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3032 	port_priv = __ib_get_mad_port(device, port_num);
3033 	if (port_priv == NULL) {
3034 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3035 		dev_err(&device->dev, "Port %d not found\n", port_num);
3036 		return -ENODEV;
3037 	}
3038 	list_del_init(&port_priv->port_list);
3039 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3040 
3041 	destroy_workqueue(port_priv->wq);
3042 	destroy_mad_qp(&port_priv->qp_info[1]);
3043 	destroy_mad_qp(&port_priv->qp_info[0]);
3044 	ib_dereg_mr(port_priv->mr);
3045 	ib_dealloc_pd(port_priv->pd);
3046 	ib_destroy_cq(port_priv->cq);
3047 	cleanup_recv_queue(&port_priv->qp_info[1]);
3048 	cleanup_recv_queue(&port_priv->qp_info[0]);
3049 	/* XXX: Handle deallocation of MAD registration tables */
3050 
3051 	kfree(port_priv);
3052 
3053 	return 0;
3054 }
3055 
3056 static void ib_mad_init_device(struct ib_device *device)
3057 {
3058 	int start, end, i;
3059 
3060 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3061 		return;
3062 
3063 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
3064 		start = 0;
3065 		end   = 0;
3066 	} else {
3067 		start = 1;
3068 		end   = device->phys_port_cnt;
3069 	}
3070 
3071 	for (i = start; i <= end; i++) {
3072 		if (ib_mad_port_open(device, i)) {
3073 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3074 			goto error;
3075 		}
3076 		if (ib_agent_port_open(device, i)) {
3077 			dev_err(&device->dev,
3078 				"Couldn't open port %d for agents\n", i);
3079 			goto error_agent;
3080 		}
3081 	}
3082 	return;
3083 
3084 error_agent:
3085 	if (ib_mad_port_close(device, i))
3086 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3087 
3088 error:
3089 	i--;
3090 
3091 	while (i >= start) {
3092 		if (ib_agent_port_close(device, i))
3093 			dev_err(&device->dev,
3094 				"Couldn't close port %d for agents\n", i);
3095 		if (ib_mad_port_close(device, i))
3096 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3097 		i--;
3098 	}
3099 }
3100 
3101 static void ib_mad_remove_device(struct ib_device *device)
3102 {
3103 	int i, num_ports, cur_port;
3104 
3105 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3106 		return;
3107 
3108 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
3109 		num_ports = 1;
3110 		cur_port = 0;
3111 	} else {
3112 		num_ports = device->phys_port_cnt;
3113 		cur_port = 1;
3114 	}
3115 	for (i = 0; i < num_ports; i++, cur_port++) {
3116 		if (ib_agent_port_close(device, cur_port))
3117 			dev_err(&device->dev,
3118 				"Couldn't close port %d for agents\n",
3119 				cur_port);
3120 		if (ib_mad_port_close(device, cur_port))
3121 			dev_err(&device->dev, "Couldn't close port %d\n",
3122 				cur_port);
3123 	}
3124 }
3125 
3126 static struct ib_client mad_client = {
3127 	.name   = "mad",
3128 	.add = ib_mad_init_device,
3129 	.remove = ib_mad_remove_device
3130 };
3131 
3132 static int __init ib_mad_init_module(void)
3133 {
3134 	int ret;
3135 
3136 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3137 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3138 
3139 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3140 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3141 
3142 	ib_mad_cache = kmem_cache_create("ib_mad",
3143 					 sizeof(struct ib_mad_private),
3144 					 0,
3145 					 SLAB_HWCACHE_ALIGN,
3146 					 NULL);
3147 	if (!ib_mad_cache) {
3148 		pr_err("Couldn't create ib_mad cache\n");
3149 		ret = -ENOMEM;
3150 		goto error1;
3151 	}
3152 
3153 	INIT_LIST_HEAD(&ib_mad_port_list);
3154 
3155 	if (ib_register_client(&mad_client)) {
3156 		pr_err("Couldn't register ib_mad client\n");
3157 		ret = -EINVAL;
3158 		goto error2;
3159 	}
3160 
3161 	return 0;
3162 
3163 error2:
3164 	kmem_cache_destroy(ib_mad_cache);
3165 error1:
3166 	return ret;
3167 }
3168 
3169 static void __exit ib_mad_cleanup_module(void)
3170 {
3171 	ib_unregister_client(&mad_client);
3172 	kmem_cache_destroy(ib_mad_cache);
3173 }
3174 
3175 module_init(ib_mad_init_module);
3176 module_exit(ib_mad_cleanup_module);
3177