xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 82e6fdd6)
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  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  */
37 
38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
39 
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/security.h>
44 #include <rdma/ib_cache.h>
45 
46 #include "mad_priv.h"
47 #include "core_priv.h"
48 #include "mad_rmpp.h"
49 #include "smi.h"
50 #include "opa_smi.h"
51 #include "agent.h"
52 
53 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
54 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
55 
56 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
57 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
58 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
59 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
60 
61 static struct list_head ib_mad_port_list;
62 static u32 ib_mad_client_id = 0;
63 
64 /* Port list lock */
65 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
66 
67 /* Forward declarations */
68 static int method_in_use(struct ib_mad_mgmt_method_table **method,
69 			 struct ib_mad_reg_req *mad_reg_req);
70 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
71 static struct ib_mad_agent_private *find_mad_agent(
72 					struct ib_mad_port_private *port_priv,
73 					const struct ib_mad_hdr *mad);
74 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
75 				    struct ib_mad_private *mad);
76 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
77 static void timeout_sends(struct work_struct *work);
78 static void local_completions(struct work_struct *work);
79 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
80 			      struct ib_mad_agent_private *agent_priv,
81 			      u8 mgmt_class);
82 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 			   struct ib_mad_agent_private *agent_priv);
84 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
85 			      struct ib_wc *wc);
86 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
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(const struct ib_mad_hdr *hdr)
183 {
184 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
185 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
186 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
187 		 (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,
321 			   "ib_register_mad_agent: Invalid port %d\n",
322 			   port_num);
323 		ret = ERR_PTR(-ENODEV);
324 		goto error1;
325 	}
326 
327 	/* Verify the QP requested is supported.  For example, Ethernet devices
328 	 * will not have QP0 */
329 	if (!port_priv->qp_info[qpn].qp) {
330 		dev_notice(&device->dev,
331 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
332 		ret = ERR_PTR(-EPROTONOSUPPORT);
333 		goto error1;
334 	}
335 
336 	/* Allocate structures */
337 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
338 	if (!mad_agent_priv) {
339 		ret = ERR_PTR(-ENOMEM);
340 		goto error1;
341 	}
342 
343 	if (mad_reg_req) {
344 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
345 		if (!reg_req) {
346 			ret = ERR_PTR(-ENOMEM);
347 			goto error3;
348 		}
349 	}
350 
351 	/* Now, fill in the various structures */
352 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
353 	mad_agent_priv->reg_req = reg_req;
354 	mad_agent_priv->agent.rmpp_version = rmpp_version;
355 	mad_agent_priv->agent.device = device;
356 	mad_agent_priv->agent.recv_handler = recv_handler;
357 	mad_agent_priv->agent.send_handler = send_handler;
358 	mad_agent_priv->agent.context = context;
359 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
360 	mad_agent_priv->agent.port_num = port_num;
361 	mad_agent_priv->agent.flags = registration_flags;
362 	spin_lock_init(&mad_agent_priv->lock);
363 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
364 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
365 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
366 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
367 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
368 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
369 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
370 	atomic_set(&mad_agent_priv->refcount, 1);
371 	init_completion(&mad_agent_priv->comp);
372 
373 	ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
374 	if (ret2) {
375 		ret = ERR_PTR(ret2);
376 		goto error4;
377 	}
378 
379 	spin_lock_irqsave(&port_priv->reg_lock, flags);
380 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
381 
382 	/*
383 	 * Make sure MAD registration (if supplied)
384 	 * is non overlapping with any existing ones
385 	 */
386 	if (mad_reg_req) {
387 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
388 		if (!is_vendor_class(mgmt_class)) {
389 			class = port_priv->version[mad_reg_req->
390 						   mgmt_class_version].class;
391 			if (class) {
392 				method = class->method_table[mgmt_class];
393 				if (method) {
394 					if (method_in_use(&method,
395 							   mad_reg_req))
396 						goto error5;
397 				}
398 			}
399 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
400 						  mgmt_class);
401 		} else {
402 			/* "New" vendor class range */
403 			vendor = port_priv->version[mad_reg_req->
404 						    mgmt_class_version].vendor;
405 			if (vendor) {
406 				vclass = vendor_class_index(mgmt_class);
407 				vendor_class = vendor->vendor_class[vclass];
408 				if (vendor_class) {
409 					if (is_vendor_method_in_use(
410 							vendor_class,
411 							mad_reg_req))
412 						goto error5;
413 				}
414 			}
415 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
416 		}
417 		if (ret2) {
418 			ret = ERR_PTR(ret2);
419 			goto error5;
420 		}
421 	}
422 
423 	/* Add mad agent into port's agent list */
424 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
425 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
426 
427 	return &mad_agent_priv->agent;
428 error5:
429 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
430 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
431 error4:
432 	kfree(reg_req);
433 error3:
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 	int err;
503 
504 	/* Validate parameters */
505 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
506 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
507 		ret = ERR_PTR(-EINVAL);
508 		goto error1;
509 	}
510 	qpn = get_spl_qp_index(qp_type);
511 	if (qpn == -1) {
512 		ret = ERR_PTR(-EINVAL);
513 		goto error1;
514 	}
515 	port_priv = ib_get_mad_port(device, port_num);
516 	if (!port_priv) {
517 		ret = ERR_PTR(-ENODEV);
518 		goto error1;
519 	}
520 	/* Allocate structures */
521 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
522 	if (!mad_snoop_priv) {
523 		ret = ERR_PTR(-ENOMEM);
524 		goto error1;
525 	}
526 
527 	/* Now, fill in the various structures */
528 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
529 	mad_snoop_priv->agent.device = device;
530 	mad_snoop_priv->agent.recv_handler = recv_handler;
531 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
532 	mad_snoop_priv->agent.context = context;
533 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
534 	mad_snoop_priv->agent.port_num = port_num;
535 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
536 	init_completion(&mad_snoop_priv->comp);
537 
538 	err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
539 	if (err) {
540 		ret = ERR_PTR(err);
541 		goto error2;
542 	}
543 
544 	mad_snoop_priv->snoop_index = register_snoop_agent(
545 						&port_priv->qp_info[qpn],
546 						mad_snoop_priv);
547 	if (mad_snoop_priv->snoop_index < 0) {
548 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
549 		goto error3;
550 	}
551 
552 	atomic_set(&mad_snoop_priv->refcount, 1);
553 	return &mad_snoop_priv->agent;
554 error3:
555 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
556 error2:
557 	kfree(mad_snoop_priv);
558 error1:
559 	return ret;
560 }
561 EXPORT_SYMBOL(ib_register_mad_snoop);
562 
563 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
564 {
565 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
566 		complete(&mad_agent_priv->comp);
567 }
568 
569 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
570 {
571 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
572 		complete(&mad_snoop_priv->comp);
573 }
574 
575 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
576 {
577 	struct ib_mad_port_private *port_priv;
578 	unsigned long flags;
579 
580 	/* Note that we could still be handling received MADs */
581 
582 	/*
583 	 * Canceling all sends results in dropping received response
584 	 * MADs, preventing us from queuing additional work
585 	 */
586 	cancel_mads(mad_agent_priv);
587 	port_priv = mad_agent_priv->qp_info->port_priv;
588 	cancel_delayed_work(&mad_agent_priv->timed_work);
589 
590 	spin_lock_irqsave(&port_priv->reg_lock, flags);
591 	remove_mad_reg_req(mad_agent_priv);
592 	list_del(&mad_agent_priv->agent_list);
593 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
594 
595 	flush_workqueue(port_priv->wq);
596 	ib_cancel_rmpp_recvs(mad_agent_priv);
597 
598 	deref_mad_agent(mad_agent_priv);
599 	wait_for_completion(&mad_agent_priv->comp);
600 
601 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
602 
603 	kfree(mad_agent_priv->reg_req);
604 	kfree(mad_agent_priv);
605 }
606 
607 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
608 {
609 	struct ib_mad_qp_info *qp_info;
610 	unsigned long flags;
611 
612 	qp_info = mad_snoop_priv->qp_info;
613 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
614 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
615 	atomic_dec(&qp_info->snoop_count);
616 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
617 
618 	deref_snoop_agent(mad_snoop_priv);
619 	wait_for_completion(&mad_snoop_priv->comp);
620 
621 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
622 
623 	kfree(mad_snoop_priv);
624 }
625 
626 /*
627  * ib_unregister_mad_agent - Unregisters a client from using MAD services
628  */
629 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
630 {
631 	struct ib_mad_agent_private *mad_agent_priv;
632 	struct ib_mad_snoop_private *mad_snoop_priv;
633 
634 	/* If the TID is zero, the agent can only snoop. */
635 	if (mad_agent->hi_tid) {
636 		mad_agent_priv = container_of(mad_agent,
637 					      struct ib_mad_agent_private,
638 					      agent);
639 		unregister_mad_agent(mad_agent_priv);
640 	} else {
641 		mad_snoop_priv = container_of(mad_agent,
642 					      struct ib_mad_snoop_private,
643 					      agent);
644 		unregister_mad_snoop(mad_snoop_priv);
645 	}
646 }
647 EXPORT_SYMBOL(ib_unregister_mad_agent);
648 
649 static void dequeue_mad(struct ib_mad_list_head *mad_list)
650 {
651 	struct ib_mad_queue *mad_queue;
652 	unsigned long flags;
653 
654 	BUG_ON(!mad_list->mad_queue);
655 	mad_queue = mad_list->mad_queue;
656 	spin_lock_irqsave(&mad_queue->lock, flags);
657 	list_del(&mad_list->list);
658 	mad_queue->count--;
659 	spin_unlock_irqrestore(&mad_queue->lock, flags);
660 }
661 
662 static void snoop_send(struct ib_mad_qp_info *qp_info,
663 		       struct ib_mad_send_buf *send_buf,
664 		       struct ib_mad_send_wc *mad_send_wc,
665 		       int mad_snoop_flags)
666 {
667 	struct ib_mad_snoop_private *mad_snoop_priv;
668 	unsigned long flags;
669 	int i;
670 
671 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
672 	for (i = 0; i < qp_info->snoop_table_size; i++) {
673 		mad_snoop_priv = qp_info->snoop_table[i];
674 		if (!mad_snoop_priv ||
675 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
676 			continue;
677 
678 		atomic_inc(&mad_snoop_priv->refcount);
679 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
680 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
681 						    send_buf, mad_send_wc);
682 		deref_snoop_agent(mad_snoop_priv);
683 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
684 	}
685 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
686 }
687 
688 static void snoop_recv(struct ib_mad_qp_info *qp_info,
689 		       struct ib_mad_recv_wc *mad_recv_wc,
690 		       int mad_snoop_flags)
691 {
692 	struct ib_mad_snoop_private *mad_snoop_priv;
693 	unsigned long flags;
694 	int i;
695 
696 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
697 	for (i = 0; i < qp_info->snoop_table_size; i++) {
698 		mad_snoop_priv = qp_info->snoop_table[i];
699 		if (!mad_snoop_priv ||
700 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
701 			continue;
702 
703 		atomic_inc(&mad_snoop_priv->refcount);
704 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
705 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
706 						   mad_recv_wc);
707 		deref_snoop_agent(mad_snoop_priv);
708 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
709 	}
710 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
711 }
712 
713 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
714 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
715 {
716 	memset(wc, 0, sizeof *wc);
717 	wc->wr_cqe = cqe;
718 	wc->status = IB_WC_SUCCESS;
719 	wc->opcode = IB_WC_RECV;
720 	wc->pkey_index = pkey_index;
721 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
722 	wc->src_qp = IB_QP0;
723 	wc->qp = qp;
724 	wc->slid = slid;
725 	wc->sl = 0;
726 	wc->dlid_path_bits = 0;
727 	wc->port_num = port_num;
728 }
729 
730 static size_t mad_priv_size(const struct ib_mad_private *mp)
731 {
732 	return sizeof(struct ib_mad_private) + mp->mad_size;
733 }
734 
735 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
736 {
737 	size_t size = sizeof(struct ib_mad_private) + mad_size;
738 	struct ib_mad_private *ret = kzalloc(size, flags);
739 
740 	if (ret)
741 		ret->mad_size = mad_size;
742 
743 	return ret;
744 }
745 
746 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
747 {
748 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
749 }
750 
751 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
752 {
753 	return sizeof(struct ib_grh) + mp->mad_size;
754 }
755 
756 /*
757  * Return 0 if SMP is to be sent
758  * Return 1 if SMP was consumed locally (whether or not solicited)
759  * Return < 0 if error
760  */
761 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
762 				  struct ib_mad_send_wr_private *mad_send_wr)
763 {
764 	int ret = 0;
765 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
766 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
767 	unsigned long flags;
768 	struct ib_mad_local_private *local;
769 	struct ib_mad_private *mad_priv;
770 	struct ib_mad_port_private *port_priv;
771 	struct ib_mad_agent_private *recv_mad_agent = NULL;
772 	struct ib_device *device = mad_agent_priv->agent.device;
773 	u8 port_num;
774 	struct ib_wc mad_wc;
775 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
776 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
777 	u16 out_mad_pkey_index = 0;
778 	u16 drslid;
779 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
780 				    mad_agent_priv->qp_info->port_priv->port_num);
781 
782 	if (rdma_cap_ib_switch(device) &&
783 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
784 		port_num = send_wr->port_num;
785 	else
786 		port_num = mad_agent_priv->agent.port_num;
787 
788 	/*
789 	 * Directed route handling starts if the initial LID routed part of
790 	 * a request or the ending LID routed part of a response is empty.
791 	 * If we are at the start of the LID routed part, don't update the
792 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
793 	 */
794 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
795 		u32 opa_drslid;
796 
797 		if ((opa_get_smp_direction(opa_smp)
798 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
799 		     OPA_LID_PERMISSIVE &&
800 		     opa_smi_handle_dr_smp_send(opa_smp,
801 						rdma_cap_ib_switch(device),
802 						port_num) == IB_SMI_DISCARD) {
803 			ret = -EINVAL;
804 			dev_err(&device->dev, "OPA Invalid directed route\n");
805 			goto out;
806 		}
807 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
808 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
809 		    opa_drslid & 0xffff0000) {
810 			ret = -EINVAL;
811 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
812 			       opa_drslid);
813 			goto out;
814 		}
815 		drslid = (u16)(opa_drslid & 0x0000ffff);
816 
817 		/* Check to post send on QP or process locally */
818 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
819 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
820 			goto out;
821 	} else {
822 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
823 		     IB_LID_PERMISSIVE &&
824 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
825 		     IB_SMI_DISCARD) {
826 			ret = -EINVAL;
827 			dev_err(&device->dev, "Invalid directed route\n");
828 			goto out;
829 		}
830 		drslid = be16_to_cpu(smp->dr_slid);
831 
832 		/* Check to post send on QP or process locally */
833 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
834 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
835 			goto out;
836 	}
837 
838 	local = kmalloc(sizeof *local, GFP_ATOMIC);
839 	if (!local) {
840 		ret = -ENOMEM;
841 		goto out;
842 	}
843 	local->mad_priv = NULL;
844 	local->recv_mad_agent = NULL;
845 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
846 	if (!mad_priv) {
847 		ret = -ENOMEM;
848 		kfree(local);
849 		goto out;
850 	}
851 
852 	build_smp_wc(mad_agent_priv->agent.qp,
853 		     send_wr->wr.wr_cqe, drslid,
854 		     send_wr->pkey_index,
855 		     send_wr->port_num, &mad_wc);
856 
857 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
858 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
859 					+ mad_send_wr->send_buf.data_len
860 					+ sizeof(struct ib_grh);
861 	}
862 
863 	/* No GRH for DR SMP */
864 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
865 				  (const struct ib_mad_hdr *)smp, mad_size,
866 				  (struct ib_mad_hdr *)mad_priv->mad,
867 				  &mad_size, &out_mad_pkey_index);
868 	switch (ret)
869 	{
870 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
871 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
872 		    mad_agent_priv->agent.recv_handler) {
873 			local->mad_priv = mad_priv;
874 			local->recv_mad_agent = mad_agent_priv;
875 			/*
876 			 * Reference MAD agent until receive
877 			 * side of local completion handled
878 			 */
879 			atomic_inc(&mad_agent_priv->refcount);
880 		} else
881 			kfree(mad_priv);
882 		break;
883 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
884 		kfree(mad_priv);
885 		break;
886 	case IB_MAD_RESULT_SUCCESS:
887 		/* Treat like an incoming receive MAD */
888 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
889 					    mad_agent_priv->agent.port_num);
890 		if (port_priv) {
891 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
892 			recv_mad_agent = find_mad_agent(port_priv,
893 						        (const struct ib_mad_hdr *)mad_priv->mad);
894 		}
895 		if (!port_priv || !recv_mad_agent) {
896 			/*
897 			 * No receiving agent so drop packet and
898 			 * generate send completion.
899 			 */
900 			kfree(mad_priv);
901 			break;
902 		}
903 		local->mad_priv = mad_priv;
904 		local->recv_mad_agent = recv_mad_agent;
905 		break;
906 	default:
907 		kfree(mad_priv);
908 		kfree(local);
909 		ret = -EINVAL;
910 		goto out;
911 	}
912 
913 	local->mad_send_wr = mad_send_wr;
914 	if (opa) {
915 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
916 		local->return_wc_byte_len = mad_size;
917 	}
918 	/* Reference MAD agent until send side of local completion handled */
919 	atomic_inc(&mad_agent_priv->refcount);
920 	/* Queue local completion to local list */
921 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
922 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
923 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
924 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
925 		   &mad_agent_priv->local_work);
926 	ret = 1;
927 out:
928 	return ret;
929 }
930 
931 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
932 {
933 	int seg_size, pad;
934 
935 	seg_size = mad_size - hdr_len;
936 	if (data_len && seg_size) {
937 		pad = seg_size - data_len % seg_size;
938 		return pad == seg_size ? 0 : pad;
939 	} else
940 		return seg_size;
941 }
942 
943 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
944 {
945 	struct ib_rmpp_segment *s, *t;
946 
947 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
948 		list_del(&s->list);
949 		kfree(s);
950 	}
951 }
952 
953 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
954 				size_t mad_size, gfp_t gfp_mask)
955 {
956 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
957 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
958 	struct ib_rmpp_segment *seg = NULL;
959 	int left, seg_size, pad;
960 
961 	send_buf->seg_size = mad_size - send_buf->hdr_len;
962 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
963 	seg_size = send_buf->seg_size;
964 	pad = send_wr->pad;
965 
966 	/* Allocate data segments. */
967 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
968 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
969 		if (!seg) {
970 			free_send_rmpp_list(send_wr);
971 			return -ENOMEM;
972 		}
973 		seg->num = ++send_buf->seg_count;
974 		list_add_tail(&seg->list, &send_wr->rmpp_list);
975 	}
976 
977 	/* Zero any padding */
978 	if (pad)
979 		memset(seg->data + seg_size - pad, 0, pad);
980 
981 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
982 					  agent.rmpp_version;
983 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
984 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
985 
986 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
987 					struct ib_rmpp_segment, list);
988 	send_wr->last_ack_seg = send_wr->cur_seg;
989 	return 0;
990 }
991 
992 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
993 {
994 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
995 }
996 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
997 
998 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
999 					    u32 remote_qpn, u16 pkey_index,
1000 					    int rmpp_active,
1001 					    int hdr_len, int data_len,
1002 					    gfp_t gfp_mask,
1003 					    u8 base_version)
1004 {
1005 	struct ib_mad_agent_private *mad_agent_priv;
1006 	struct ib_mad_send_wr_private *mad_send_wr;
1007 	int pad, message_size, ret, size;
1008 	void *buf;
1009 	size_t mad_size;
1010 	bool opa;
1011 
1012 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1013 				      agent);
1014 
1015 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1016 
1017 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1018 		mad_size = sizeof(struct opa_mad);
1019 	else
1020 		mad_size = sizeof(struct ib_mad);
1021 
1022 	pad = get_pad_size(hdr_len, data_len, mad_size);
1023 	message_size = hdr_len + data_len + pad;
1024 
1025 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1026 		if (!rmpp_active && message_size > mad_size)
1027 			return ERR_PTR(-EINVAL);
1028 	} else
1029 		if (rmpp_active || message_size > mad_size)
1030 			return ERR_PTR(-EINVAL);
1031 
1032 	size = rmpp_active ? hdr_len : mad_size;
1033 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1034 	if (!buf)
1035 		return ERR_PTR(-ENOMEM);
1036 
1037 	mad_send_wr = buf + size;
1038 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1039 	mad_send_wr->send_buf.mad = buf;
1040 	mad_send_wr->send_buf.hdr_len = hdr_len;
1041 	mad_send_wr->send_buf.data_len = data_len;
1042 	mad_send_wr->pad = pad;
1043 
1044 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1045 	mad_send_wr->sg_list[0].length = hdr_len;
1046 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1047 
1048 	/* OPA MADs don't have to be the full 2048 bytes */
1049 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1050 	    data_len < mad_size - hdr_len)
1051 		mad_send_wr->sg_list[1].length = data_len;
1052 	else
1053 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1054 
1055 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1056 
1057 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1058 
1059 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1060 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1061 	mad_send_wr->send_wr.wr.num_sge = 2;
1062 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1063 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1064 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1065 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1066 	mad_send_wr->send_wr.pkey_index = pkey_index;
1067 
1068 	if (rmpp_active) {
1069 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1070 		if (ret) {
1071 			kfree(buf);
1072 			return ERR_PTR(ret);
1073 		}
1074 	}
1075 
1076 	mad_send_wr->send_buf.mad_agent = mad_agent;
1077 	atomic_inc(&mad_agent_priv->refcount);
1078 	return &mad_send_wr->send_buf;
1079 }
1080 EXPORT_SYMBOL(ib_create_send_mad);
1081 
1082 int ib_get_mad_data_offset(u8 mgmt_class)
1083 {
1084 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1085 		return IB_MGMT_SA_HDR;
1086 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1087 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1088 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1089 		return IB_MGMT_DEVICE_HDR;
1090 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1091 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1092 		return IB_MGMT_VENDOR_HDR;
1093 	else
1094 		return IB_MGMT_MAD_HDR;
1095 }
1096 EXPORT_SYMBOL(ib_get_mad_data_offset);
1097 
1098 int ib_is_mad_class_rmpp(u8 mgmt_class)
1099 {
1100 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1101 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1102 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1103 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1104 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1105 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1106 		return 1;
1107 	return 0;
1108 }
1109 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1110 
1111 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1112 {
1113 	struct ib_mad_send_wr_private *mad_send_wr;
1114 	struct list_head *list;
1115 
1116 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1117 				   send_buf);
1118 	list = &mad_send_wr->cur_seg->list;
1119 
1120 	if (mad_send_wr->cur_seg->num < seg_num) {
1121 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1122 			if (mad_send_wr->cur_seg->num == seg_num)
1123 				break;
1124 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1125 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1126 			if (mad_send_wr->cur_seg->num == seg_num)
1127 				break;
1128 	}
1129 	return mad_send_wr->cur_seg->data;
1130 }
1131 EXPORT_SYMBOL(ib_get_rmpp_segment);
1132 
1133 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1134 {
1135 	if (mad_send_wr->send_buf.seg_count)
1136 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1137 					   mad_send_wr->seg_num);
1138 	else
1139 		return mad_send_wr->send_buf.mad +
1140 		       mad_send_wr->send_buf.hdr_len;
1141 }
1142 
1143 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1144 {
1145 	struct ib_mad_agent_private *mad_agent_priv;
1146 	struct ib_mad_send_wr_private *mad_send_wr;
1147 
1148 	mad_agent_priv = container_of(send_buf->mad_agent,
1149 				      struct ib_mad_agent_private, agent);
1150 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1151 				   send_buf);
1152 
1153 	free_send_rmpp_list(mad_send_wr);
1154 	kfree(send_buf->mad);
1155 	deref_mad_agent(mad_agent_priv);
1156 }
1157 EXPORT_SYMBOL(ib_free_send_mad);
1158 
1159 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1160 {
1161 	struct ib_mad_qp_info *qp_info;
1162 	struct list_head *list;
1163 	struct ib_send_wr *bad_send_wr;
1164 	struct ib_mad_agent *mad_agent;
1165 	struct ib_sge *sge;
1166 	unsigned long flags;
1167 	int ret;
1168 
1169 	/* Set WR ID to find mad_send_wr upon completion */
1170 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1171 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1172 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1173 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1174 
1175 	mad_agent = mad_send_wr->send_buf.mad_agent;
1176 	sge = mad_send_wr->sg_list;
1177 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1178 					mad_send_wr->send_buf.mad,
1179 					sge[0].length,
1180 					DMA_TO_DEVICE);
1181 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1182 		return -ENOMEM;
1183 
1184 	mad_send_wr->header_mapping = sge[0].addr;
1185 
1186 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1187 					ib_get_payload(mad_send_wr),
1188 					sge[1].length,
1189 					DMA_TO_DEVICE);
1190 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1191 		ib_dma_unmap_single(mad_agent->device,
1192 				    mad_send_wr->header_mapping,
1193 				    sge[0].length, DMA_TO_DEVICE);
1194 		return -ENOMEM;
1195 	}
1196 	mad_send_wr->payload_mapping = sge[1].addr;
1197 
1198 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1199 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1200 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1201 				   &bad_send_wr);
1202 		list = &qp_info->send_queue.list;
1203 	} else {
1204 		ret = 0;
1205 		list = &qp_info->overflow_list;
1206 	}
1207 
1208 	if (!ret) {
1209 		qp_info->send_queue.count++;
1210 		list_add_tail(&mad_send_wr->mad_list.list, list);
1211 	}
1212 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1213 	if (ret) {
1214 		ib_dma_unmap_single(mad_agent->device,
1215 				    mad_send_wr->header_mapping,
1216 				    sge[0].length, DMA_TO_DEVICE);
1217 		ib_dma_unmap_single(mad_agent->device,
1218 				    mad_send_wr->payload_mapping,
1219 				    sge[1].length, DMA_TO_DEVICE);
1220 	}
1221 	return ret;
1222 }
1223 
1224 /*
1225  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1226  *  with the registered client
1227  */
1228 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1229 		     struct ib_mad_send_buf **bad_send_buf)
1230 {
1231 	struct ib_mad_agent_private *mad_agent_priv;
1232 	struct ib_mad_send_buf *next_send_buf;
1233 	struct ib_mad_send_wr_private *mad_send_wr;
1234 	unsigned long flags;
1235 	int ret = -EINVAL;
1236 
1237 	/* Walk list of send WRs and post each on send list */
1238 	for (; send_buf; send_buf = next_send_buf) {
1239 		mad_send_wr = container_of(send_buf,
1240 					   struct ib_mad_send_wr_private,
1241 					   send_buf);
1242 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1243 
1244 		ret = ib_mad_enforce_security(mad_agent_priv,
1245 					      mad_send_wr->send_wr.pkey_index);
1246 		if (ret)
1247 			goto error;
1248 
1249 		if (!send_buf->mad_agent->send_handler ||
1250 		    (send_buf->timeout_ms &&
1251 		     !send_buf->mad_agent->recv_handler)) {
1252 			ret = -EINVAL;
1253 			goto error;
1254 		}
1255 
1256 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1257 			if (mad_agent_priv->agent.rmpp_version) {
1258 				ret = -EINVAL;
1259 				goto error;
1260 			}
1261 		}
1262 
1263 		/*
1264 		 * Save pointer to next work request to post in case the
1265 		 * current one completes, and the user modifies the work
1266 		 * request associated with the completion
1267 		 */
1268 		next_send_buf = send_buf->next;
1269 		mad_send_wr->send_wr.ah = send_buf->ah;
1270 
1271 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1272 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1273 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1274 						     mad_send_wr);
1275 			if (ret < 0)		/* error */
1276 				goto error;
1277 			else if (ret == 1)	/* locally consumed */
1278 				continue;
1279 		}
1280 
1281 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1282 		/* Timeout will be updated after send completes */
1283 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1284 		mad_send_wr->max_retries = send_buf->retries;
1285 		mad_send_wr->retries_left = send_buf->retries;
1286 		send_buf->retries = 0;
1287 		/* Reference for work request to QP + response */
1288 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1289 		mad_send_wr->status = IB_WC_SUCCESS;
1290 
1291 		/* Reference MAD agent until send completes */
1292 		atomic_inc(&mad_agent_priv->refcount);
1293 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1294 		list_add_tail(&mad_send_wr->agent_list,
1295 			      &mad_agent_priv->send_list);
1296 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1297 
1298 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1299 			ret = ib_send_rmpp_mad(mad_send_wr);
1300 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1301 				ret = ib_send_mad(mad_send_wr);
1302 		} else
1303 			ret = ib_send_mad(mad_send_wr);
1304 		if (ret < 0) {
1305 			/* Fail send request */
1306 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1307 			list_del(&mad_send_wr->agent_list);
1308 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1309 			atomic_dec(&mad_agent_priv->refcount);
1310 			goto error;
1311 		}
1312 	}
1313 	return 0;
1314 error:
1315 	if (bad_send_buf)
1316 		*bad_send_buf = send_buf;
1317 	return ret;
1318 }
1319 EXPORT_SYMBOL(ib_post_send_mad);
1320 
1321 /*
1322  * ib_free_recv_mad - Returns data buffers used to receive
1323  *  a MAD to the access layer
1324  */
1325 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1326 {
1327 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1328 	struct ib_mad_private_header *mad_priv_hdr;
1329 	struct ib_mad_private *priv;
1330 	struct list_head free_list;
1331 
1332 	INIT_LIST_HEAD(&free_list);
1333 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1334 
1335 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1336 					&free_list, list) {
1337 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1338 					   recv_buf);
1339 		mad_priv_hdr = container_of(mad_recv_wc,
1340 					    struct ib_mad_private_header,
1341 					    recv_wc);
1342 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1343 				    header);
1344 		kfree(priv);
1345 	}
1346 }
1347 EXPORT_SYMBOL(ib_free_recv_mad);
1348 
1349 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1350 					u8 rmpp_version,
1351 					ib_mad_send_handler send_handler,
1352 					ib_mad_recv_handler recv_handler,
1353 					void *context)
1354 {
1355 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1356 }
1357 EXPORT_SYMBOL(ib_redirect_mad_qp);
1358 
1359 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1360 		      struct ib_wc *wc)
1361 {
1362 	dev_err(&mad_agent->device->dev,
1363 		"ib_process_mad_wc() not implemented yet\n");
1364 	return 0;
1365 }
1366 EXPORT_SYMBOL(ib_process_mad_wc);
1367 
1368 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1369 			 struct ib_mad_reg_req *mad_reg_req)
1370 {
1371 	int i;
1372 
1373 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1374 		if ((*method)->agent[i]) {
1375 			pr_err("Method %d already in use\n", i);
1376 			return -EINVAL;
1377 		}
1378 	}
1379 	return 0;
1380 }
1381 
1382 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1383 {
1384 	/* Allocate management method table */
1385 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1386 	return (*method) ? 0 : (-ENOMEM);
1387 }
1388 
1389 /*
1390  * Check to see if there are any methods still in use
1391  */
1392 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1393 {
1394 	int i;
1395 
1396 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1397 		if (method->agent[i])
1398 			return 1;
1399 	return 0;
1400 }
1401 
1402 /*
1403  * Check to see if there are any method tables for this class still in use
1404  */
1405 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1406 {
1407 	int i;
1408 
1409 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1410 		if (class->method_table[i])
1411 			return 1;
1412 	return 0;
1413 }
1414 
1415 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1416 {
1417 	int i;
1418 
1419 	for (i = 0; i < MAX_MGMT_OUI; i++)
1420 		if (vendor_class->method_table[i])
1421 			return 1;
1422 	return 0;
1423 }
1424 
1425 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1426 			   const char *oui)
1427 {
1428 	int i;
1429 
1430 	for (i = 0; i < MAX_MGMT_OUI; i++)
1431 		/* Is there matching OUI for this vendor class ? */
1432 		if (!memcmp(vendor_class->oui[i], oui, 3))
1433 			return i;
1434 
1435 	return -1;
1436 }
1437 
1438 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1439 {
1440 	int i;
1441 
1442 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1443 		if (vendor->vendor_class[i])
1444 			return 1;
1445 
1446 	return 0;
1447 }
1448 
1449 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1450 				     struct ib_mad_agent_private *agent)
1451 {
1452 	int i;
1453 
1454 	/* Remove any methods for this mad agent */
1455 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1456 		if (method->agent[i] == agent) {
1457 			method->agent[i] = NULL;
1458 		}
1459 	}
1460 }
1461 
1462 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1463 			      struct ib_mad_agent_private *agent_priv,
1464 			      u8 mgmt_class)
1465 {
1466 	struct ib_mad_port_private *port_priv;
1467 	struct ib_mad_mgmt_class_table **class;
1468 	struct ib_mad_mgmt_method_table **method;
1469 	int i, ret;
1470 
1471 	port_priv = agent_priv->qp_info->port_priv;
1472 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1473 	if (!*class) {
1474 		/* Allocate management class table for "new" class version */
1475 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1476 		if (!*class) {
1477 			ret = -ENOMEM;
1478 			goto error1;
1479 		}
1480 
1481 		/* Allocate method table for this management class */
1482 		method = &(*class)->method_table[mgmt_class];
1483 		if ((ret = allocate_method_table(method)))
1484 			goto error2;
1485 	} else {
1486 		method = &(*class)->method_table[mgmt_class];
1487 		if (!*method) {
1488 			/* Allocate method table for this management class */
1489 			if ((ret = allocate_method_table(method)))
1490 				goto error1;
1491 		}
1492 	}
1493 
1494 	/* Now, make sure methods are not already in use */
1495 	if (method_in_use(method, mad_reg_req))
1496 		goto error3;
1497 
1498 	/* Finally, add in methods being registered */
1499 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1500 		(*method)->agent[i] = agent_priv;
1501 
1502 	return 0;
1503 
1504 error3:
1505 	/* Remove any methods for this mad agent */
1506 	remove_methods_mad_agent(*method, agent_priv);
1507 	/* Now, check to see if there are any methods in use */
1508 	if (!check_method_table(*method)) {
1509 		/* If not, release management method table */
1510 		kfree(*method);
1511 		*method = NULL;
1512 	}
1513 	ret = -EINVAL;
1514 	goto error1;
1515 error2:
1516 	kfree(*class);
1517 	*class = NULL;
1518 error1:
1519 	return ret;
1520 }
1521 
1522 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1523 			   struct ib_mad_agent_private *agent_priv)
1524 {
1525 	struct ib_mad_port_private *port_priv;
1526 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1527 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1528 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1529 	struct ib_mad_mgmt_method_table **method;
1530 	int i, ret = -ENOMEM;
1531 	u8 vclass;
1532 
1533 	/* "New" vendor (with OUI) class */
1534 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1535 	port_priv = agent_priv->qp_info->port_priv;
1536 	vendor_table = &port_priv->version[
1537 				mad_reg_req->mgmt_class_version].vendor;
1538 	if (!*vendor_table) {
1539 		/* Allocate mgmt vendor class table for "new" class version */
1540 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1541 		if (!vendor)
1542 			goto error1;
1543 
1544 		*vendor_table = vendor;
1545 	}
1546 	if (!(*vendor_table)->vendor_class[vclass]) {
1547 		/* Allocate table for this management vendor class */
1548 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1549 		if (!vendor_class)
1550 			goto error2;
1551 
1552 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1553 	}
1554 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1555 		/* Is there matching OUI for this vendor class ? */
1556 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1557 			    mad_reg_req->oui, 3)) {
1558 			method = &(*vendor_table)->vendor_class[
1559 						vclass]->method_table[i];
1560 			BUG_ON(!*method);
1561 			goto check_in_use;
1562 		}
1563 	}
1564 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1565 		/* OUI slot available ? */
1566 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1567 				vclass]->oui[i])) {
1568 			method = &(*vendor_table)->vendor_class[
1569 				vclass]->method_table[i];
1570 			BUG_ON(*method);
1571 			/* Allocate method table for this OUI */
1572 			if ((ret = allocate_method_table(method)))
1573 				goto error3;
1574 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1575 			       mad_reg_req->oui, 3);
1576 			goto check_in_use;
1577 		}
1578 	}
1579 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1580 	goto error3;
1581 
1582 check_in_use:
1583 	/* Now, make sure methods are not already in use */
1584 	if (method_in_use(method, mad_reg_req))
1585 		goto error4;
1586 
1587 	/* Finally, add in methods being registered */
1588 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1589 		(*method)->agent[i] = agent_priv;
1590 
1591 	return 0;
1592 
1593 error4:
1594 	/* Remove any methods for this mad agent */
1595 	remove_methods_mad_agent(*method, agent_priv);
1596 	/* Now, check to see if there are any methods in use */
1597 	if (!check_method_table(*method)) {
1598 		/* If not, release management method table */
1599 		kfree(*method);
1600 		*method = NULL;
1601 	}
1602 	ret = -EINVAL;
1603 error3:
1604 	if (vendor_class) {
1605 		(*vendor_table)->vendor_class[vclass] = NULL;
1606 		kfree(vendor_class);
1607 	}
1608 error2:
1609 	if (vendor) {
1610 		*vendor_table = NULL;
1611 		kfree(vendor);
1612 	}
1613 error1:
1614 	return ret;
1615 }
1616 
1617 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1618 {
1619 	struct ib_mad_port_private *port_priv;
1620 	struct ib_mad_mgmt_class_table *class;
1621 	struct ib_mad_mgmt_method_table *method;
1622 	struct ib_mad_mgmt_vendor_class_table *vendor;
1623 	struct ib_mad_mgmt_vendor_class *vendor_class;
1624 	int index;
1625 	u8 mgmt_class;
1626 
1627 	/*
1628 	 * Was MAD registration request supplied
1629 	 * with original registration ?
1630 	 */
1631 	if (!agent_priv->reg_req) {
1632 		goto out;
1633 	}
1634 
1635 	port_priv = agent_priv->qp_info->port_priv;
1636 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1637 	class = port_priv->version[
1638 			agent_priv->reg_req->mgmt_class_version].class;
1639 	if (!class)
1640 		goto vendor_check;
1641 
1642 	method = class->method_table[mgmt_class];
1643 	if (method) {
1644 		/* Remove any methods for this mad agent */
1645 		remove_methods_mad_agent(method, agent_priv);
1646 		/* Now, check to see if there are any methods still in use */
1647 		if (!check_method_table(method)) {
1648 			/* If not, release management method table */
1649 			kfree(method);
1650 			class->method_table[mgmt_class] = NULL;
1651 			/* Any management classes left ? */
1652 			if (!check_class_table(class)) {
1653 				/* If not, release management class table */
1654 				kfree(class);
1655 				port_priv->version[
1656 					agent_priv->reg_req->
1657 					mgmt_class_version].class = NULL;
1658 			}
1659 		}
1660 	}
1661 
1662 vendor_check:
1663 	if (!is_vendor_class(mgmt_class))
1664 		goto out;
1665 
1666 	/* normalize mgmt_class to vendor range 2 */
1667 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1668 	vendor = port_priv->version[
1669 			agent_priv->reg_req->mgmt_class_version].vendor;
1670 
1671 	if (!vendor)
1672 		goto out;
1673 
1674 	vendor_class = vendor->vendor_class[mgmt_class];
1675 	if (vendor_class) {
1676 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1677 		if (index < 0)
1678 			goto out;
1679 		method = vendor_class->method_table[index];
1680 		if (method) {
1681 			/* Remove any methods for this mad agent */
1682 			remove_methods_mad_agent(method, agent_priv);
1683 			/*
1684 			 * Now, check to see if there are
1685 			 * any methods still in use
1686 			 */
1687 			if (!check_method_table(method)) {
1688 				/* If not, release management method table */
1689 				kfree(method);
1690 				vendor_class->method_table[index] = NULL;
1691 				memset(vendor_class->oui[index], 0, 3);
1692 				/* Any OUIs left ? */
1693 				if (!check_vendor_class(vendor_class)) {
1694 					/* If not, release vendor class table */
1695 					kfree(vendor_class);
1696 					vendor->vendor_class[mgmt_class] = NULL;
1697 					/* Any other vendor classes left ? */
1698 					if (!check_vendor_table(vendor)) {
1699 						kfree(vendor);
1700 						port_priv->version[
1701 							agent_priv->reg_req->
1702 							mgmt_class_version].
1703 							vendor = NULL;
1704 					}
1705 				}
1706 			}
1707 		}
1708 	}
1709 
1710 out:
1711 	return;
1712 }
1713 
1714 static struct ib_mad_agent_private *
1715 find_mad_agent(struct ib_mad_port_private *port_priv,
1716 	       const struct ib_mad_hdr *mad_hdr)
1717 {
1718 	struct ib_mad_agent_private *mad_agent = NULL;
1719 	unsigned long flags;
1720 
1721 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1722 	if (ib_response_mad(mad_hdr)) {
1723 		u32 hi_tid;
1724 		struct ib_mad_agent_private *entry;
1725 
1726 		/*
1727 		 * Routing is based on high 32 bits of transaction ID
1728 		 * of MAD.
1729 		 */
1730 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1731 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1732 			if (entry->agent.hi_tid == hi_tid) {
1733 				mad_agent = entry;
1734 				break;
1735 			}
1736 		}
1737 	} else {
1738 		struct ib_mad_mgmt_class_table *class;
1739 		struct ib_mad_mgmt_method_table *method;
1740 		struct ib_mad_mgmt_vendor_class_table *vendor;
1741 		struct ib_mad_mgmt_vendor_class *vendor_class;
1742 		const struct ib_vendor_mad *vendor_mad;
1743 		int index;
1744 
1745 		/*
1746 		 * Routing is based on version, class, and method
1747 		 * For "newer" vendor MADs, also based on OUI
1748 		 */
1749 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1750 			goto out;
1751 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1752 			class = port_priv->version[
1753 					mad_hdr->class_version].class;
1754 			if (!class)
1755 				goto out;
1756 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1757 			    ARRAY_SIZE(class->method_table))
1758 				goto out;
1759 			method = class->method_table[convert_mgmt_class(
1760 							mad_hdr->mgmt_class)];
1761 			if (method)
1762 				mad_agent = method->agent[mad_hdr->method &
1763 							  ~IB_MGMT_METHOD_RESP];
1764 		} else {
1765 			vendor = port_priv->version[
1766 					mad_hdr->class_version].vendor;
1767 			if (!vendor)
1768 				goto out;
1769 			vendor_class = vendor->vendor_class[vendor_class_index(
1770 						mad_hdr->mgmt_class)];
1771 			if (!vendor_class)
1772 				goto out;
1773 			/* Find matching OUI */
1774 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1775 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1776 			if (index == -1)
1777 				goto out;
1778 			method = vendor_class->method_table[index];
1779 			if (method) {
1780 				mad_agent = method->agent[mad_hdr->method &
1781 							  ~IB_MGMT_METHOD_RESP];
1782 			}
1783 		}
1784 	}
1785 
1786 	if (mad_agent) {
1787 		if (mad_agent->agent.recv_handler)
1788 			atomic_inc(&mad_agent->refcount);
1789 		else {
1790 			dev_notice(&port_priv->device->dev,
1791 				   "No receive handler for client %p on port %d\n",
1792 				   &mad_agent->agent, port_priv->port_num);
1793 			mad_agent = NULL;
1794 		}
1795 	}
1796 out:
1797 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1798 
1799 	return mad_agent;
1800 }
1801 
1802 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1803 			const struct ib_mad_qp_info *qp_info,
1804 			bool opa)
1805 {
1806 	int valid = 0;
1807 	u32 qp_num = qp_info->qp->qp_num;
1808 
1809 	/* Make sure MAD base version is understood */
1810 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1811 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1812 		pr_err("MAD received with unsupported base version %d %s\n",
1813 		       mad_hdr->base_version, opa ? "(opa)" : "");
1814 		goto out;
1815 	}
1816 
1817 	/* Filter SMI packets sent to other than QP0 */
1818 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1819 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1820 		if (qp_num == 0)
1821 			valid = 1;
1822 	} else {
1823 		/* CM attributes other than ClassPortInfo only use Send method */
1824 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1825 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1826 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1827 			goto out;
1828 		/* Filter GSI packets sent to QP0 */
1829 		if (qp_num != 0)
1830 			valid = 1;
1831 	}
1832 
1833 out:
1834 	return valid;
1835 }
1836 
1837 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1838 			    const struct ib_mad_hdr *mad_hdr)
1839 {
1840 	struct ib_rmpp_mad *rmpp_mad;
1841 
1842 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1843 	return !mad_agent_priv->agent.rmpp_version ||
1844 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1845 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1846 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1847 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1848 }
1849 
1850 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1851 				     const struct ib_mad_recv_wc *rwc)
1852 {
1853 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1854 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1855 }
1856 
1857 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1858 				   const struct ib_mad_send_wr_private *wr,
1859 				   const struct ib_mad_recv_wc *rwc )
1860 {
1861 	struct rdma_ah_attr attr;
1862 	u8 send_resp, rcv_resp;
1863 	union ib_gid sgid;
1864 	struct ib_device *device = mad_agent_priv->agent.device;
1865 	u8 port_num = mad_agent_priv->agent.port_num;
1866 	u8 lmc;
1867 	bool has_grh;
1868 
1869 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1870 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1871 
1872 	if (send_resp == rcv_resp)
1873 		/* both requests, or both responses. GIDs different */
1874 		return 0;
1875 
1876 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1877 		/* Assume not equal, to avoid false positives. */
1878 		return 0;
1879 
1880 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1881 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1882 		/* one has GID, other does not.  Assume different */
1883 		return 0;
1884 
1885 	if (!send_resp && rcv_resp) {
1886 		/* is request/response. */
1887 		if (!has_grh) {
1888 			if (ib_get_cached_lmc(device, port_num, &lmc))
1889 				return 0;
1890 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1891 					   rwc->wc->dlid_path_bits) &
1892 					  ((1 << lmc) - 1)));
1893 		} else {
1894 			const struct ib_global_route *grh =
1895 					rdma_ah_read_grh(&attr);
1896 
1897 			if (ib_get_cached_gid(device, port_num,
1898 					      grh->sgid_index, &sgid, NULL))
1899 				return 0;
1900 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1901 				       16);
1902 		}
1903 	}
1904 
1905 	if (!has_grh)
1906 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1907 	else
1908 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1909 			       rwc->recv_buf.grh->sgid.raw,
1910 			       16);
1911 }
1912 
1913 static inline int is_direct(u8 class)
1914 {
1915 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1916 }
1917 
1918 struct ib_mad_send_wr_private*
1919 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1920 		 const struct ib_mad_recv_wc *wc)
1921 {
1922 	struct ib_mad_send_wr_private *wr;
1923 	const struct ib_mad_hdr *mad_hdr;
1924 
1925 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1926 
1927 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1928 		if ((wr->tid == mad_hdr->tid) &&
1929 		    rcv_has_same_class(wr, wc) &&
1930 		    /*
1931 		     * Don't check GID for direct routed MADs.
1932 		     * These might have permissive LIDs.
1933 		     */
1934 		    (is_direct(mad_hdr->mgmt_class) ||
1935 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1936 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1937 	}
1938 
1939 	/*
1940 	 * It's possible to receive the response before we've
1941 	 * been notified that the send has completed
1942 	 */
1943 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1944 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1945 		    wr->tid == mad_hdr->tid &&
1946 		    wr->timeout &&
1947 		    rcv_has_same_class(wr, wc) &&
1948 		    /*
1949 		     * Don't check GID for direct routed MADs.
1950 		     * These might have permissive LIDs.
1951 		     */
1952 		    (is_direct(mad_hdr->mgmt_class) ||
1953 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1954 			/* Verify request has not been canceled */
1955 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1956 	}
1957 	return NULL;
1958 }
1959 
1960 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1961 {
1962 	mad_send_wr->timeout = 0;
1963 	if (mad_send_wr->refcount == 1)
1964 		list_move_tail(&mad_send_wr->agent_list,
1965 			      &mad_send_wr->mad_agent_priv->done_list);
1966 }
1967 
1968 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1969 				 struct ib_mad_recv_wc *mad_recv_wc)
1970 {
1971 	struct ib_mad_send_wr_private *mad_send_wr;
1972 	struct ib_mad_send_wc mad_send_wc;
1973 	unsigned long flags;
1974 	int ret;
1975 
1976 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1977 	ret = ib_mad_enforce_security(mad_agent_priv,
1978 				      mad_recv_wc->wc->pkey_index);
1979 	if (ret) {
1980 		ib_free_recv_mad(mad_recv_wc);
1981 		deref_mad_agent(mad_agent_priv);
1982 		return;
1983 	}
1984 
1985 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1986 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1987 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1988 						      mad_recv_wc);
1989 		if (!mad_recv_wc) {
1990 			deref_mad_agent(mad_agent_priv);
1991 			return;
1992 		}
1993 	}
1994 
1995 	/* Complete corresponding request */
1996 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1997 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1998 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1999 		if (!mad_send_wr) {
2000 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2001 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2002 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2003 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2004 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
2005 				/* user rmpp is in effect
2006 				 * and this is an active RMPP MAD
2007 				 */
2008 				mad_agent_priv->agent.recv_handler(
2009 						&mad_agent_priv->agent, NULL,
2010 						mad_recv_wc);
2011 				atomic_dec(&mad_agent_priv->refcount);
2012 			} else {
2013 				/* not user rmpp, revert to normal behavior and
2014 				 * drop the mad */
2015 				ib_free_recv_mad(mad_recv_wc);
2016 				deref_mad_agent(mad_agent_priv);
2017 				return;
2018 			}
2019 		} else {
2020 			ib_mark_mad_done(mad_send_wr);
2021 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2022 
2023 			/* Defined behavior is to complete response before request */
2024 			mad_agent_priv->agent.recv_handler(
2025 					&mad_agent_priv->agent,
2026 					&mad_send_wr->send_buf,
2027 					mad_recv_wc);
2028 			atomic_dec(&mad_agent_priv->refcount);
2029 
2030 			mad_send_wc.status = IB_WC_SUCCESS;
2031 			mad_send_wc.vendor_err = 0;
2032 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2033 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2034 		}
2035 	} else {
2036 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2037 						   mad_recv_wc);
2038 		deref_mad_agent(mad_agent_priv);
2039 	}
2040 
2041 	return;
2042 }
2043 
2044 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2045 				     const struct ib_mad_qp_info *qp_info,
2046 				     const struct ib_wc *wc,
2047 				     int port_num,
2048 				     struct ib_mad_private *recv,
2049 				     struct ib_mad_private *response)
2050 {
2051 	enum smi_forward_action retsmi;
2052 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2053 
2054 	if (smi_handle_dr_smp_recv(smp,
2055 				   rdma_cap_ib_switch(port_priv->device),
2056 				   port_num,
2057 				   port_priv->device->phys_port_cnt) ==
2058 				   IB_SMI_DISCARD)
2059 		return IB_SMI_DISCARD;
2060 
2061 	retsmi = smi_check_forward_dr_smp(smp);
2062 	if (retsmi == IB_SMI_LOCAL)
2063 		return IB_SMI_HANDLE;
2064 
2065 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2066 		if (smi_handle_dr_smp_send(smp,
2067 					   rdma_cap_ib_switch(port_priv->device),
2068 					   port_num) == IB_SMI_DISCARD)
2069 			return IB_SMI_DISCARD;
2070 
2071 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2072 			return IB_SMI_DISCARD;
2073 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2074 		/* forward case for switches */
2075 		memcpy(response, recv, mad_priv_size(response));
2076 		response->header.recv_wc.wc = &response->header.wc;
2077 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2078 		response->header.recv_wc.recv_buf.grh = &response->grh;
2079 
2080 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2081 				    &response->grh, wc,
2082 				    port_priv->device,
2083 				    smi_get_fwd_port(smp),
2084 				    qp_info->qp->qp_num,
2085 				    response->mad_size,
2086 				    false);
2087 
2088 		return IB_SMI_DISCARD;
2089 	}
2090 	return IB_SMI_HANDLE;
2091 }
2092 
2093 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2094 				    struct ib_mad_private *response,
2095 				    size_t *resp_len, bool opa)
2096 {
2097 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2098 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2099 
2100 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2101 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2102 		memcpy(response, recv, mad_priv_size(response));
2103 		response->header.recv_wc.wc = &response->header.wc;
2104 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2105 		response->header.recv_wc.recv_buf.grh = &response->grh;
2106 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2107 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2108 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2109 			resp_hdr->status |= IB_SMP_DIRECTION;
2110 
2111 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2112 			if (recv_hdr->mgmt_class ==
2113 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2114 			    recv_hdr->mgmt_class ==
2115 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2116 				*resp_len = opa_get_smp_header_size(
2117 							(struct opa_smp *)recv->mad);
2118 			else
2119 				*resp_len = sizeof(struct ib_mad_hdr);
2120 		}
2121 
2122 		return true;
2123 	} else {
2124 		return false;
2125 	}
2126 }
2127 
2128 static enum smi_action
2129 handle_opa_smi(struct ib_mad_port_private *port_priv,
2130 	       struct ib_mad_qp_info *qp_info,
2131 	       struct ib_wc *wc,
2132 	       int port_num,
2133 	       struct ib_mad_private *recv,
2134 	       struct ib_mad_private *response)
2135 {
2136 	enum smi_forward_action retsmi;
2137 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2138 
2139 	if (opa_smi_handle_dr_smp_recv(smp,
2140 				   rdma_cap_ib_switch(port_priv->device),
2141 				   port_num,
2142 				   port_priv->device->phys_port_cnt) ==
2143 				   IB_SMI_DISCARD)
2144 		return IB_SMI_DISCARD;
2145 
2146 	retsmi = opa_smi_check_forward_dr_smp(smp);
2147 	if (retsmi == IB_SMI_LOCAL)
2148 		return IB_SMI_HANDLE;
2149 
2150 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2151 		if (opa_smi_handle_dr_smp_send(smp,
2152 					   rdma_cap_ib_switch(port_priv->device),
2153 					   port_num) == IB_SMI_DISCARD)
2154 			return IB_SMI_DISCARD;
2155 
2156 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2157 		    IB_SMI_DISCARD)
2158 			return IB_SMI_DISCARD;
2159 
2160 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2161 		/* forward case for switches */
2162 		memcpy(response, recv, mad_priv_size(response));
2163 		response->header.recv_wc.wc = &response->header.wc;
2164 		response->header.recv_wc.recv_buf.opa_mad =
2165 				(struct opa_mad *)response->mad;
2166 		response->header.recv_wc.recv_buf.grh = &response->grh;
2167 
2168 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2169 				    &response->grh, wc,
2170 				    port_priv->device,
2171 				    opa_smi_get_fwd_port(smp),
2172 				    qp_info->qp->qp_num,
2173 				    recv->header.wc.byte_len,
2174 				    true);
2175 
2176 		return IB_SMI_DISCARD;
2177 	}
2178 
2179 	return IB_SMI_HANDLE;
2180 }
2181 
2182 static enum smi_action
2183 handle_smi(struct ib_mad_port_private *port_priv,
2184 	   struct ib_mad_qp_info *qp_info,
2185 	   struct ib_wc *wc,
2186 	   int port_num,
2187 	   struct ib_mad_private *recv,
2188 	   struct ib_mad_private *response,
2189 	   bool opa)
2190 {
2191 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2192 
2193 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2194 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2195 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2196 				      response);
2197 
2198 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2199 }
2200 
2201 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2202 {
2203 	struct ib_mad_port_private *port_priv = cq->cq_context;
2204 	struct ib_mad_list_head *mad_list =
2205 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2206 	struct ib_mad_qp_info *qp_info;
2207 	struct ib_mad_private_header *mad_priv_hdr;
2208 	struct ib_mad_private *recv, *response = NULL;
2209 	struct ib_mad_agent_private *mad_agent;
2210 	int port_num;
2211 	int ret = IB_MAD_RESULT_SUCCESS;
2212 	size_t mad_size;
2213 	u16 resp_mad_pkey_index = 0;
2214 	bool opa;
2215 
2216 	if (list_empty_careful(&port_priv->port_list))
2217 		return;
2218 
2219 	if (wc->status != IB_WC_SUCCESS) {
2220 		/*
2221 		 * Receive errors indicate that the QP has entered the error
2222 		 * state - error handling/shutdown code will cleanup
2223 		 */
2224 		return;
2225 	}
2226 
2227 	qp_info = mad_list->mad_queue->qp_info;
2228 	dequeue_mad(mad_list);
2229 
2230 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2231 			       qp_info->port_priv->port_num);
2232 
2233 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2234 				    mad_list);
2235 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2236 	ib_dma_unmap_single(port_priv->device,
2237 			    recv->header.mapping,
2238 			    mad_priv_dma_size(recv),
2239 			    DMA_FROM_DEVICE);
2240 
2241 	/* Setup MAD receive work completion from "normal" work completion */
2242 	recv->header.wc = *wc;
2243 	recv->header.recv_wc.wc = &recv->header.wc;
2244 
2245 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2246 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2247 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2248 	} else {
2249 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2250 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2251 	}
2252 
2253 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2254 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2255 
2256 	if (atomic_read(&qp_info->snoop_count))
2257 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2258 
2259 	/* Validate MAD */
2260 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2261 		goto out;
2262 
2263 	mad_size = recv->mad_size;
2264 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2265 	if (!response)
2266 		goto out;
2267 
2268 	if (rdma_cap_ib_switch(port_priv->device))
2269 		port_num = wc->port_num;
2270 	else
2271 		port_num = port_priv->port_num;
2272 
2273 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2274 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2275 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2276 			       response, opa)
2277 		    == IB_SMI_DISCARD)
2278 			goto out;
2279 	}
2280 
2281 	/* Give driver "right of first refusal" on incoming MAD */
2282 	if (port_priv->device->process_mad) {
2283 		ret = port_priv->device->process_mad(port_priv->device, 0,
2284 						     port_priv->port_num,
2285 						     wc, &recv->grh,
2286 						     (const struct ib_mad_hdr *)recv->mad,
2287 						     recv->mad_size,
2288 						     (struct ib_mad_hdr *)response->mad,
2289 						     &mad_size, &resp_mad_pkey_index);
2290 
2291 		if (opa)
2292 			wc->pkey_index = resp_mad_pkey_index;
2293 
2294 		if (ret & IB_MAD_RESULT_SUCCESS) {
2295 			if (ret & IB_MAD_RESULT_CONSUMED)
2296 				goto out;
2297 			if (ret & IB_MAD_RESULT_REPLY) {
2298 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2299 						    &recv->grh, wc,
2300 						    port_priv->device,
2301 						    port_num,
2302 						    qp_info->qp->qp_num,
2303 						    mad_size, opa);
2304 				goto out;
2305 			}
2306 		}
2307 	}
2308 
2309 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2310 	if (mad_agent) {
2311 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2312 		/*
2313 		 * recv is freed up in error cases in ib_mad_complete_recv
2314 		 * or via recv_handler in ib_mad_complete_recv()
2315 		 */
2316 		recv = NULL;
2317 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2318 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2319 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2320 				    port_priv->device, port_num,
2321 				    qp_info->qp->qp_num, mad_size, opa);
2322 	}
2323 
2324 out:
2325 	/* Post another receive request for this QP */
2326 	if (response) {
2327 		ib_mad_post_receive_mads(qp_info, response);
2328 		kfree(recv);
2329 	} else
2330 		ib_mad_post_receive_mads(qp_info, recv);
2331 }
2332 
2333 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2334 {
2335 	struct ib_mad_send_wr_private *mad_send_wr;
2336 	unsigned long delay;
2337 
2338 	if (list_empty(&mad_agent_priv->wait_list)) {
2339 		cancel_delayed_work(&mad_agent_priv->timed_work);
2340 	} else {
2341 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2342 					 struct ib_mad_send_wr_private,
2343 					 agent_list);
2344 
2345 		if (time_after(mad_agent_priv->timeout,
2346 			       mad_send_wr->timeout)) {
2347 			mad_agent_priv->timeout = mad_send_wr->timeout;
2348 			delay = mad_send_wr->timeout - jiffies;
2349 			if ((long)delay <= 0)
2350 				delay = 1;
2351 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2352 					 &mad_agent_priv->timed_work, delay);
2353 		}
2354 	}
2355 }
2356 
2357 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2358 {
2359 	struct ib_mad_agent_private *mad_agent_priv;
2360 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2361 	struct list_head *list_item;
2362 	unsigned long delay;
2363 
2364 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2365 	list_del(&mad_send_wr->agent_list);
2366 
2367 	delay = mad_send_wr->timeout;
2368 	mad_send_wr->timeout += jiffies;
2369 
2370 	if (delay) {
2371 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2372 			temp_mad_send_wr = list_entry(list_item,
2373 						struct ib_mad_send_wr_private,
2374 						agent_list);
2375 			if (time_after(mad_send_wr->timeout,
2376 				       temp_mad_send_wr->timeout))
2377 				break;
2378 		}
2379 	}
2380 	else
2381 		list_item = &mad_agent_priv->wait_list;
2382 	list_add(&mad_send_wr->agent_list, list_item);
2383 
2384 	/* Reschedule a work item if we have a shorter timeout */
2385 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2386 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2387 				 &mad_agent_priv->timed_work, delay);
2388 }
2389 
2390 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2391 			  int timeout_ms)
2392 {
2393 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2394 	wait_for_response(mad_send_wr);
2395 }
2396 
2397 /*
2398  * Process a send work completion
2399  */
2400 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2401 			     struct ib_mad_send_wc *mad_send_wc)
2402 {
2403 	struct ib_mad_agent_private	*mad_agent_priv;
2404 	unsigned long			flags;
2405 	int				ret;
2406 
2407 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2408 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2409 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2410 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2411 		if (ret == IB_RMPP_RESULT_CONSUMED)
2412 			goto done;
2413 	} else
2414 		ret = IB_RMPP_RESULT_UNHANDLED;
2415 
2416 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2417 	    mad_send_wr->status == IB_WC_SUCCESS) {
2418 		mad_send_wr->status = mad_send_wc->status;
2419 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2420 	}
2421 
2422 	if (--mad_send_wr->refcount > 0) {
2423 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2424 		    mad_send_wr->status == IB_WC_SUCCESS) {
2425 			wait_for_response(mad_send_wr);
2426 		}
2427 		goto done;
2428 	}
2429 
2430 	/* Remove send from MAD agent and notify client of completion */
2431 	list_del(&mad_send_wr->agent_list);
2432 	adjust_timeout(mad_agent_priv);
2433 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2434 
2435 	if (mad_send_wr->status != IB_WC_SUCCESS )
2436 		mad_send_wc->status = mad_send_wr->status;
2437 	if (ret == IB_RMPP_RESULT_INTERNAL)
2438 		ib_rmpp_send_handler(mad_send_wc);
2439 	else
2440 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2441 						   mad_send_wc);
2442 
2443 	/* Release reference on agent taken when sending */
2444 	deref_mad_agent(mad_agent_priv);
2445 	return;
2446 done:
2447 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2448 }
2449 
2450 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2451 {
2452 	struct ib_mad_port_private *port_priv = cq->cq_context;
2453 	struct ib_mad_list_head *mad_list =
2454 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2455 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2456 	struct ib_mad_qp_info		*qp_info;
2457 	struct ib_mad_queue		*send_queue;
2458 	struct ib_send_wr		*bad_send_wr;
2459 	struct ib_mad_send_wc		mad_send_wc;
2460 	unsigned long flags;
2461 	int ret;
2462 
2463 	if (list_empty_careful(&port_priv->port_list))
2464 		return;
2465 
2466 	if (wc->status != IB_WC_SUCCESS) {
2467 		if (!ib_mad_send_error(port_priv, wc))
2468 			return;
2469 	}
2470 
2471 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2472 				   mad_list);
2473 	send_queue = mad_list->mad_queue;
2474 	qp_info = send_queue->qp_info;
2475 
2476 retry:
2477 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2478 			    mad_send_wr->header_mapping,
2479 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2480 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2481 			    mad_send_wr->payload_mapping,
2482 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2483 	queued_send_wr = NULL;
2484 	spin_lock_irqsave(&send_queue->lock, flags);
2485 	list_del(&mad_list->list);
2486 
2487 	/* Move queued send to the send queue */
2488 	if (send_queue->count-- > send_queue->max_active) {
2489 		mad_list = container_of(qp_info->overflow_list.next,
2490 					struct ib_mad_list_head, list);
2491 		queued_send_wr = container_of(mad_list,
2492 					struct ib_mad_send_wr_private,
2493 					mad_list);
2494 		list_move_tail(&mad_list->list, &send_queue->list);
2495 	}
2496 	spin_unlock_irqrestore(&send_queue->lock, flags);
2497 
2498 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2499 	mad_send_wc.status = wc->status;
2500 	mad_send_wc.vendor_err = wc->vendor_err;
2501 	if (atomic_read(&qp_info->snoop_count))
2502 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2503 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2504 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2505 
2506 	if (queued_send_wr) {
2507 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2508 				   &bad_send_wr);
2509 		if (ret) {
2510 			dev_err(&port_priv->device->dev,
2511 				"ib_post_send failed: %d\n", ret);
2512 			mad_send_wr = queued_send_wr;
2513 			wc->status = IB_WC_LOC_QP_OP_ERR;
2514 			goto retry;
2515 		}
2516 	}
2517 }
2518 
2519 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2520 {
2521 	struct ib_mad_send_wr_private *mad_send_wr;
2522 	struct ib_mad_list_head *mad_list;
2523 	unsigned long flags;
2524 
2525 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2526 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2527 		mad_send_wr = container_of(mad_list,
2528 					   struct ib_mad_send_wr_private,
2529 					   mad_list);
2530 		mad_send_wr->retry = 1;
2531 	}
2532 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2533 }
2534 
2535 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2536 		struct ib_wc *wc)
2537 {
2538 	struct ib_mad_list_head *mad_list =
2539 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2540 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2541 	struct ib_mad_send_wr_private *mad_send_wr;
2542 	int ret;
2543 
2544 	/*
2545 	 * Send errors will transition the QP to SQE - move
2546 	 * QP to RTS and repost flushed work requests
2547 	 */
2548 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2549 				   mad_list);
2550 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2551 		if (mad_send_wr->retry) {
2552 			/* Repost send */
2553 			struct ib_send_wr *bad_send_wr;
2554 
2555 			mad_send_wr->retry = 0;
2556 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2557 					&bad_send_wr);
2558 			if (!ret)
2559 				return false;
2560 		}
2561 	} else {
2562 		struct ib_qp_attr *attr;
2563 
2564 		/* Transition QP to RTS and fail offending send */
2565 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2566 		if (attr) {
2567 			attr->qp_state = IB_QPS_RTS;
2568 			attr->cur_qp_state = IB_QPS_SQE;
2569 			ret = ib_modify_qp(qp_info->qp, attr,
2570 					   IB_QP_STATE | IB_QP_CUR_STATE);
2571 			kfree(attr);
2572 			if (ret)
2573 				dev_err(&port_priv->device->dev,
2574 					"%s - ib_modify_qp to RTS: %d\n",
2575 					__func__, ret);
2576 			else
2577 				mark_sends_for_retry(qp_info);
2578 		}
2579 	}
2580 
2581 	return true;
2582 }
2583 
2584 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2585 {
2586 	unsigned long flags;
2587 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2588 	struct ib_mad_send_wc mad_send_wc;
2589 	struct list_head cancel_list;
2590 
2591 	INIT_LIST_HEAD(&cancel_list);
2592 
2593 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2594 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2595 				 &mad_agent_priv->send_list, agent_list) {
2596 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2597 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2598 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2599 		}
2600 	}
2601 
2602 	/* Empty wait list to prevent receives from finding a request */
2603 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2604 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2605 
2606 	/* Report all cancelled requests */
2607 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2608 	mad_send_wc.vendor_err = 0;
2609 
2610 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2611 				 &cancel_list, agent_list) {
2612 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2613 		list_del(&mad_send_wr->agent_list);
2614 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2615 						   &mad_send_wc);
2616 		atomic_dec(&mad_agent_priv->refcount);
2617 	}
2618 }
2619 
2620 static struct ib_mad_send_wr_private*
2621 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2622 	     struct ib_mad_send_buf *send_buf)
2623 {
2624 	struct ib_mad_send_wr_private *mad_send_wr;
2625 
2626 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2627 			    agent_list) {
2628 		if (&mad_send_wr->send_buf == send_buf)
2629 			return mad_send_wr;
2630 	}
2631 
2632 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2633 			    agent_list) {
2634 		if (is_rmpp_data_mad(mad_agent_priv,
2635 				     mad_send_wr->send_buf.mad) &&
2636 		    &mad_send_wr->send_buf == send_buf)
2637 			return mad_send_wr;
2638 	}
2639 	return NULL;
2640 }
2641 
2642 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2643 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2644 {
2645 	struct ib_mad_agent_private *mad_agent_priv;
2646 	struct ib_mad_send_wr_private *mad_send_wr;
2647 	unsigned long flags;
2648 	int active;
2649 
2650 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2651 				      agent);
2652 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2653 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2654 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2655 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2656 		return -EINVAL;
2657 	}
2658 
2659 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2660 	if (!timeout_ms) {
2661 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2662 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2663 	}
2664 
2665 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2666 	if (active)
2667 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2668 	else
2669 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2670 
2671 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2672 	return 0;
2673 }
2674 EXPORT_SYMBOL(ib_modify_mad);
2675 
2676 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2677 		   struct ib_mad_send_buf *send_buf)
2678 {
2679 	ib_modify_mad(mad_agent, send_buf, 0);
2680 }
2681 EXPORT_SYMBOL(ib_cancel_mad);
2682 
2683 static void local_completions(struct work_struct *work)
2684 {
2685 	struct ib_mad_agent_private *mad_agent_priv;
2686 	struct ib_mad_local_private *local;
2687 	struct ib_mad_agent_private *recv_mad_agent;
2688 	unsigned long flags;
2689 	int free_mad;
2690 	struct ib_wc wc;
2691 	struct ib_mad_send_wc mad_send_wc;
2692 	bool opa;
2693 
2694 	mad_agent_priv =
2695 		container_of(work, struct ib_mad_agent_private, local_work);
2696 
2697 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2698 			       mad_agent_priv->qp_info->port_priv->port_num);
2699 
2700 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2701 	while (!list_empty(&mad_agent_priv->local_list)) {
2702 		local = list_entry(mad_agent_priv->local_list.next,
2703 				   struct ib_mad_local_private,
2704 				   completion_list);
2705 		list_del(&local->completion_list);
2706 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2707 		free_mad = 0;
2708 		if (local->mad_priv) {
2709 			u8 base_version;
2710 			recv_mad_agent = local->recv_mad_agent;
2711 			if (!recv_mad_agent) {
2712 				dev_err(&mad_agent_priv->agent.device->dev,
2713 					"No receive MAD agent for local completion\n");
2714 				free_mad = 1;
2715 				goto local_send_completion;
2716 			}
2717 
2718 			/*
2719 			 * Defined behavior is to complete response
2720 			 * before request
2721 			 */
2722 			build_smp_wc(recv_mad_agent->agent.qp,
2723 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2724 				     be16_to_cpu(IB_LID_PERMISSIVE),
2725 				     local->mad_send_wr->send_wr.pkey_index,
2726 				     recv_mad_agent->agent.port_num, &wc);
2727 
2728 			local->mad_priv->header.recv_wc.wc = &wc;
2729 
2730 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2731 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2732 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2733 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2734 			} else {
2735 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2736 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2737 			}
2738 
2739 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2740 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2741 				 &local->mad_priv->header.recv_wc.rmpp_list);
2742 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2743 			local->mad_priv->header.recv_wc.recv_buf.mad =
2744 						(struct ib_mad *)local->mad_priv->mad;
2745 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2746 				snoop_recv(recv_mad_agent->qp_info,
2747 					  &local->mad_priv->header.recv_wc,
2748 					   IB_MAD_SNOOP_RECVS);
2749 			recv_mad_agent->agent.recv_handler(
2750 						&recv_mad_agent->agent,
2751 						&local->mad_send_wr->send_buf,
2752 						&local->mad_priv->header.recv_wc);
2753 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2754 			atomic_dec(&recv_mad_agent->refcount);
2755 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2756 		}
2757 
2758 local_send_completion:
2759 		/* Complete send */
2760 		mad_send_wc.status = IB_WC_SUCCESS;
2761 		mad_send_wc.vendor_err = 0;
2762 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2763 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2764 			snoop_send(mad_agent_priv->qp_info,
2765 				   &local->mad_send_wr->send_buf,
2766 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2767 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2768 						   &mad_send_wc);
2769 
2770 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2771 		atomic_dec(&mad_agent_priv->refcount);
2772 		if (free_mad)
2773 			kfree(local->mad_priv);
2774 		kfree(local);
2775 	}
2776 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2777 }
2778 
2779 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2780 {
2781 	int ret;
2782 
2783 	if (!mad_send_wr->retries_left)
2784 		return -ETIMEDOUT;
2785 
2786 	mad_send_wr->retries_left--;
2787 	mad_send_wr->send_buf.retries++;
2788 
2789 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2790 
2791 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2792 		ret = ib_retry_rmpp(mad_send_wr);
2793 		switch (ret) {
2794 		case IB_RMPP_RESULT_UNHANDLED:
2795 			ret = ib_send_mad(mad_send_wr);
2796 			break;
2797 		case IB_RMPP_RESULT_CONSUMED:
2798 			ret = 0;
2799 			break;
2800 		default:
2801 			ret = -ECOMM;
2802 			break;
2803 		}
2804 	} else
2805 		ret = ib_send_mad(mad_send_wr);
2806 
2807 	if (!ret) {
2808 		mad_send_wr->refcount++;
2809 		list_add_tail(&mad_send_wr->agent_list,
2810 			      &mad_send_wr->mad_agent_priv->send_list);
2811 	}
2812 	return ret;
2813 }
2814 
2815 static void timeout_sends(struct work_struct *work)
2816 {
2817 	struct ib_mad_agent_private *mad_agent_priv;
2818 	struct ib_mad_send_wr_private *mad_send_wr;
2819 	struct ib_mad_send_wc mad_send_wc;
2820 	unsigned long flags, delay;
2821 
2822 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2823 				      timed_work.work);
2824 	mad_send_wc.vendor_err = 0;
2825 
2826 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2827 	while (!list_empty(&mad_agent_priv->wait_list)) {
2828 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2829 					 struct ib_mad_send_wr_private,
2830 					 agent_list);
2831 
2832 		if (time_after(mad_send_wr->timeout, jiffies)) {
2833 			delay = mad_send_wr->timeout - jiffies;
2834 			if ((long)delay <= 0)
2835 				delay = 1;
2836 			queue_delayed_work(mad_agent_priv->qp_info->
2837 					   port_priv->wq,
2838 					   &mad_agent_priv->timed_work, delay);
2839 			break;
2840 		}
2841 
2842 		list_del(&mad_send_wr->agent_list);
2843 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2844 		    !retry_send(mad_send_wr))
2845 			continue;
2846 
2847 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2848 
2849 		if (mad_send_wr->status == IB_WC_SUCCESS)
2850 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2851 		else
2852 			mad_send_wc.status = mad_send_wr->status;
2853 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2854 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2855 						   &mad_send_wc);
2856 
2857 		atomic_dec(&mad_agent_priv->refcount);
2858 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2859 	}
2860 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2861 }
2862 
2863 /*
2864  * Allocate receive MADs and post receive WRs for them
2865  */
2866 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2867 				    struct ib_mad_private *mad)
2868 {
2869 	unsigned long flags;
2870 	int post, ret;
2871 	struct ib_mad_private *mad_priv;
2872 	struct ib_sge sg_list;
2873 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2874 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2875 
2876 	/* Initialize common scatter list fields */
2877 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2878 
2879 	/* Initialize common receive WR fields */
2880 	recv_wr.next = NULL;
2881 	recv_wr.sg_list = &sg_list;
2882 	recv_wr.num_sge = 1;
2883 
2884 	do {
2885 		/* Allocate and map receive buffer */
2886 		if (mad) {
2887 			mad_priv = mad;
2888 			mad = NULL;
2889 		} else {
2890 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2891 						     GFP_ATOMIC);
2892 			if (!mad_priv) {
2893 				ret = -ENOMEM;
2894 				break;
2895 			}
2896 		}
2897 		sg_list.length = mad_priv_dma_size(mad_priv);
2898 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2899 						 &mad_priv->grh,
2900 						 mad_priv_dma_size(mad_priv),
2901 						 DMA_FROM_DEVICE);
2902 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2903 						  sg_list.addr))) {
2904 			ret = -ENOMEM;
2905 			break;
2906 		}
2907 		mad_priv->header.mapping = sg_list.addr;
2908 		mad_priv->header.mad_list.mad_queue = recv_queue;
2909 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2910 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2911 
2912 		/* Post receive WR */
2913 		spin_lock_irqsave(&recv_queue->lock, flags);
2914 		post = (++recv_queue->count < recv_queue->max_active);
2915 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2916 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2917 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2918 		if (ret) {
2919 			spin_lock_irqsave(&recv_queue->lock, flags);
2920 			list_del(&mad_priv->header.mad_list.list);
2921 			recv_queue->count--;
2922 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2923 			ib_dma_unmap_single(qp_info->port_priv->device,
2924 					    mad_priv->header.mapping,
2925 					    mad_priv_dma_size(mad_priv),
2926 					    DMA_FROM_DEVICE);
2927 			kfree(mad_priv);
2928 			dev_err(&qp_info->port_priv->device->dev,
2929 				"ib_post_recv failed: %d\n", ret);
2930 			break;
2931 		}
2932 	} while (post);
2933 
2934 	return ret;
2935 }
2936 
2937 /*
2938  * Return all the posted receive MADs
2939  */
2940 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2941 {
2942 	struct ib_mad_private_header *mad_priv_hdr;
2943 	struct ib_mad_private *recv;
2944 	struct ib_mad_list_head *mad_list;
2945 
2946 	if (!qp_info->qp)
2947 		return;
2948 
2949 	while (!list_empty(&qp_info->recv_queue.list)) {
2950 
2951 		mad_list = list_entry(qp_info->recv_queue.list.next,
2952 				      struct ib_mad_list_head, list);
2953 		mad_priv_hdr = container_of(mad_list,
2954 					    struct ib_mad_private_header,
2955 					    mad_list);
2956 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2957 				    header);
2958 
2959 		/* Remove from posted receive MAD list */
2960 		list_del(&mad_list->list);
2961 
2962 		ib_dma_unmap_single(qp_info->port_priv->device,
2963 				    recv->header.mapping,
2964 				    mad_priv_dma_size(recv),
2965 				    DMA_FROM_DEVICE);
2966 		kfree(recv);
2967 	}
2968 
2969 	qp_info->recv_queue.count = 0;
2970 }
2971 
2972 /*
2973  * Start the port
2974  */
2975 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2976 {
2977 	int ret, i;
2978 	struct ib_qp_attr *attr;
2979 	struct ib_qp *qp;
2980 	u16 pkey_index;
2981 
2982 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2983 	if (!attr)
2984 		return -ENOMEM;
2985 
2986 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2987 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2988 	if (ret)
2989 		pkey_index = 0;
2990 
2991 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2992 		qp = port_priv->qp_info[i].qp;
2993 		if (!qp)
2994 			continue;
2995 
2996 		/*
2997 		 * PKey index for QP1 is irrelevant but
2998 		 * one is needed for the Reset to Init transition
2999 		 */
3000 		attr->qp_state = IB_QPS_INIT;
3001 		attr->pkey_index = pkey_index;
3002 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3003 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3004 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
3005 		if (ret) {
3006 			dev_err(&port_priv->device->dev,
3007 				"Couldn't change QP%d state to INIT: %d\n",
3008 				i, ret);
3009 			goto out;
3010 		}
3011 
3012 		attr->qp_state = IB_QPS_RTR;
3013 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3014 		if (ret) {
3015 			dev_err(&port_priv->device->dev,
3016 				"Couldn't change QP%d state to RTR: %d\n",
3017 				i, ret);
3018 			goto out;
3019 		}
3020 
3021 		attr->qp_state = IB_QPS_RTS;
3022 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3023 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3024 		if (ret) {
3025 			dev_err(&port_priv->device->dev,
3026 				"Couldn't change QP%d state to RTS: %d\n",
3027 				i, ret);
3028 			goto out;
3029 		}
3030 	}
3031 
3032 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3033 	if (ret) {
3034 		dev_err(&port_priv->device->dev,
3035 			"Failed to request completion notification: %d\n",
3036 			ret);
3037 		goto out;
3038 	}
3039 
3040 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3041 		if (!port_priv->qp_info[i].qp)
3042 			continue;
3043 
3044 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3045 		if (ret) {
3046 			dev_err(&port_priv->device->dev,
3047 				"Couldn't post receive WRs\n");
3048 			goto out;
3049 		}
3050 	}
3051 out:
3052 	kfree(attr);
3053 	return ret;
3054 }
3055 
3056 static void qp_event_handler(struct ib_event *event, void *qp_context)
3057 {
3058 	struct ib_mad_qp_info	*qp_info = qp_context;
3059 
3060 	/* It's worse than that! He's dead, Jim! */
3061 	dev_err(&qp_info->port_priv->device->dev,
3062 		"Fatal error (%d) on MAD QP (%d)\n",
3063 		event->event, qp_info->qp->qp_num);
3064 }
3065 
3066 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3067 			   struct ib_mad_queue *mad_queue)
3068 {
3069 	mad_queue->qp_info = qp_info;
3070 	mad_queue->count = 0;
3071 	spin_lock_init(&mad_queue->lock);
3072 	INIT_LIST_HEAD(&mad_queue->list);
3073 }
3074 
3075 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3076 			struct ib_mad_qp_info *qp_info)
3077 {
3078 	qp_info->port_priv = port_priv;
3079 	init_mad_queue(qp_info, &qp_info->send_queue);
3080 	init_mad_queue(qp_info, &qp_info->recv_queue);
3081 	INIT_LIST_HEAD(&qp_info->overflow_list);
3082 	spin_lock_init(&qp_info->snoop_lock);
3083 	qp_info->snoop_table = NULL;
3084 	qp_info->snoop_table_size = 0;
3085 	atomic_set(&qp_info->snoop_count, 0);
3086 }
3087 
3088 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3089 			 enum ib_qp_type qp_type)
3090 {
3091 	struct ib_qp_init_attr	qp_init_attr;
3092 	int ret;
3093 
3094 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3095 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3096 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3097 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3098 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3099 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3100 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3101 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3102 	qp_init_attr.qp_type = qp_type;
3103 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3104 	qp_init_attr.qp_context = qp_info;
3105 	qp_init_attr.event_handler = qp_event_handler;
3106 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3107 	if (IS_ERR(qp_info->qp)) {
3108 		dev_err(&qp_info->port_priv->device->dev,
3109 			"Couldn't create ib_mad QP%d\n",
3110 			get_spl_qp_index(qp_type));
3111 		ret = PTR_ERR(qp_info->qp);
3112 		goto error;
3113 	}
3114 	/* Use minimum queue sizes unless the CQ is resized */
3115 	qp_info->send_queue.max_active = mad_sendq_size;
3116 	qp_info->recv_queue.max_active = mad_recvq_size;
3117 	return 0;
3118 
3119 error:
3120 	return ret;
3121 }
3122 
3123 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3124 {
3125 	if (!qp_info->qp)
3126 		return;
3127 
3128 	ib_destroy_qp(qp_info->qp);
3129 	kfree(qp_info->snoop_table);
3130 }
3131 
3132 /*
3133  * Open the port
3134  * Create the QP, PD, MR, and CQ if needed
3135  */
3136 static int ib_mad_port_open(struct ib_device *device,
3137 			    int port_num)
3138 {
3139 	int ret, cq_size;
3140 	struct ib_mad_port_private *port_priv;
3141 	unsigned long flags;
3142 	char name[sizeof "ib_mad123"];
3143 	int has_smi;
3144 
3145 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3146 		return -EFAULT;
3147 
3148 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3149 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3150 		return -EFAULT;
3151 
3152 	/* Create new device info */
3153 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3154 	if (!port_priv)
3155 		return -ENOMEM;
3156 
3157 	port_priv->device = device;
3158 	port_priv->port_num = port_num;
3159 	spin_lock_init(&port_priv->reg_lock);
3160 	INIT_LIST_HEAD(&port_priv->agent_list);
3161 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3162 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3163 
3164 	cq_size = mad_sendq_size + mad_recvq_size;
3165 	has_smi = rdma_cap_ib_smi(device, port_num);
3166 	if (has_smi)
3167 		cq_size *= 2;
3168 
3169 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3170 			IB_POLL_WORKQUEUE);
3171 	if (IS_ERR(port_priv->cq)) {
3172 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3173 		ret = PTR_ERR(port_priv->cq);
3174 		goto error3;
3175 	}
3176 
3177 	port_priv->pd = ib_alloc_pd(device, 0);
3178 	if (IS_ERR(port_priv->pd)) {
3179 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3180 		ret = PTR_ERR(port_priv->pd);
3181 		goto error4;
3182 	}
3183 
3184 	if (has_smi) {
3185 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3186 		if (ret)
3187 			goto error6;
3188 	}
3189 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3190 	if (ret)
3191 		goto error7;
3192 
3193 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3194 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3195 	if (!port_priv->wq) {
3196 		ret = -ENOMEM;
3197 		goto error8;
3198 	}
3199 
3200 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3201 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3202 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3203 
3204 	ret = ib_mad_port_start(port_priv);
3205 	if (ret) {
3206 		dev_err(&device->dev, "Couldn't start port\n");
3207 		goto error9;
3208 	}
3209 
3210 	return 0;
3211 
3212 error9:
3213 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3214 	list_del_init(&port_priv->port_list);
3215 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3216 
3217 	destroy_workqueue(port_priv->wq);
3218 error8:
3219 	destroy_mad_qp(&port_priv->qp_info[1]);
3220 error7:
3221 	destroy_mad_qp(&port_priv->qp_info[0]);
3222 error6:
3223 	ib_dealloc_pd(port_priv->pd);
3224 error4:
3225 	ib_free_cq(port_priv->cq);
3226 	cleanup_recv_queue(&port_priv->qp_info[1]);
3227 	cleanup_recv_queue(&port_priv->qp_info[0]);
3228 error3:
3229 	kfree(port_priv);
3230 
3231 	return ret;
3232 }
3233 
3234 /*
3235  * Close the port
3236  * If there are no classes using the port, free the port
3237  * resources (CQ, MR, PD, QP) and remove the port's info structure
3238  */
3239 static int ib_mad_port_close(struct ib_device *device, int port_num)
3240 {
3241 	struct ib_mad_port_private *port_priv;
3242 	unsigned long flags;
3243 
3244 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3245 	port_priv = __ib_get_mad_port(device, port_num);
3246 	if (port_priv == NULL) {
3247 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3248 		dev_err(&device->dev, "Port %d not found\n", port_num);
3249 		return -ENODEV;
3250 	}
3251 	list_del_init(&port_priv->port_list);
3252 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3253 
3254 	destroy_workqueue(port_priv->wq);
3255 	destroy_mad_qp(&port_priv->qp_info[1]);
3256 	destroy_mad_qp(&port_priv->qp_info[0]);
3257 	ib_dealloc_pd(port_priv->pd);
3258 	ib_free_cq(port_priv->cq);
3259 	cleanup_recv_queue(&port_priv->qp_info[1]);
3260 	cleanup_recv_queue(&port_priv->qp_info[0]);
3261 	/* XXX: Handle deallocation of MAD registration tables */
3262 
3263 	kfree(port_priv);
3264 
3265 	return 0;
3266 }
3267 
3268 static void ib_mad_init_device(struct ib_device *device)
3269 {
3270 	int start, i;
3271 
3272 	start = rdma_start_port(device);
3273 
3274 	for (i = start; i <= rdma_end_port(device); i++) {
3275 		if (!rdma_cap_ib_mad(device, i))
3276 			continue;
3277 
3278 		if (ib_mad_port_open(device, i)) {
3279 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3280 			goto error;
3281 		}
3282 		if (ib_agent_port_open(device, i)) {
3283 			dev_err(&device->dev,
3284 				"Couldn't open port %d for agents\n", i);
3285 			goto error_agent;
3286 		}
3287 	}
3288 	return;
3289 
3290 error_agent:
3291 	if (ib_mad_port_close(device, i))
3292 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3293 
3294 error:
3295 	while (--i >= start) {
3296 		if (!rdma_cap_ib_mad(device, i))
3297 			continue;
3298 
3299 		if (ib_agent_port_close(device, i))
3300 			dev_err(&device->dev,
3301 				"Couldn't close port %d for agents\n", i);
3302 		if (ib_mad_port_close(device, i))
3303 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3304 	}
3305 }
3306 
3307 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3308 {
3309 	int i;
3310 
3311 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3312 		if (!rdma_cap_ib_mad(device, i))
3313 			continue;
3314 
3315 		if (ib_agent_port_close(device, i))
3316 			dev_err(&device->dev,
3317 				"Couldn't close port %d for agents\n", i);
3318 		if (ib_mad_port_close(device, i))
3319 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3320 	}
3321 }
3322 
3323 static struct ib_client mad_client = {
3324 	.name   = "mad",
3325 	.add = ib_mad_init_device,
3326 	.remove = ib_mad_remove_device
3327 };
3328 
3329 int ib_mad_init(void)
3330 {
3331 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3332 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3333 
3334 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3335 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3336 
3337 	INIT_LIST_HEAD(&ib_mad_port_list);
3338 
3339 	if (ib_register_client(&mad_client)) {
3340 		pr_err("Couldn't register ib_mad client\n");
3341 		return -EINVAL;
3342 	}
3343 
3344 	return 0;
3345 }
3346 
3347 void ib_mad_cleanup(void)
3348 {
3349 	ib_unregister_client(&mad_client);
3350 }
3351