xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 2f828fb2)
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 #include "core_priv.h"
53 
54 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
55 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
56 
57 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
58 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
59 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
60 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
61 
62 static struct list_head ib_mad_port_list;
63 static u32 ib_mad_client_id = 0;
64 
65 /* Port list lock */
66 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
67 
68 /* Forward declarations */
69 static int method_in_use(struct ib_mad_mgmt_method_table **method,
70 			 struct ib_mad_reg_req *mad_reg_req);
71 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
72 static struct ib_mad_agent_private *find_mad_agent(
73 					struct ib_mad_port_private *port_priv,
74 					const struct ib_mad_hdr *mad);
75 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
76 				    struct ib_mad_private *mad);
77 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
78 static void timeout_sends(struct work_struct *work);
79 static void local_completions(struct work_struct *work);
80 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
81 			      struct ib_mad_agent_private *agent_priv,
82 			      u8 mgmt_class);
83 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
84 			   struct ib_mad_agent_private *agent_priv);
85 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
86 			      struct ib_wc *wc);
87 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
88 
89 /*
90  * Returns a ib_mad_port_private structure or NULL for a device/port
91  * Assumes ib_mad_port_list_lock is being held
92  */
93 static inline struct ib_mad_port_private *
94 __ib_get_mad_port(struct ib_device *device, int port_num)
95 {
96 	struct ib_mad_port_private *entry;
97 
98 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
99 		if (entry->device == device && entry->port_num == port_num)
100 			return entry;
101 	}
102 	return NULL;
103 }
104 
105 /*
106  * Wrapper function to return a ib_mad_port_private structure or NULL
107  * for a device/port
108  */
109 static inline struct ib_mad_port_private *
110 ib_get_mad_port(struct ib_device *device, int port_num)
111 {
112 	struct ib_mad_port_private *entry;
113 	unsigned long flags;
114 
115 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
116 	entry = __ib_get_mad_port(device, port_num);
117 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
118 
119 	return entry;
120 }
121 
122 static inline u8 convert_mgmt_class(u8 mgmt_class)
123 {
124 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
125 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
126 		0 : mgmt_class;
127 }
128 
129 static int get_spl_qp_index(enum ib_qp_type qp_type)
130 {
131 	switch (qp_type)
132 	{
133 	case IB_QPT_SMI:
134 		return 0;
135 	case IB_QPT_GSI:
136 		return 1;
137 	default:
138 		return -1;
139 	}
140 }
141 
142 static int vendor_class_index(u8 mgmt_class)
143 {
144 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
145 }
146 
147 static int is_vendor_class(u8 mgmt_class)
148 {
149 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
150 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
151 		return 0;
152 	return 1;
153 }
154 
155 static int is_vendor_oui(char *oui)
156 {
157 	if (oui[0] || oui[1] || oui[2])
158 		return 1;
159 	return 0;
160 }
161 
162 static int is_vendor_method_in_use(
163 		struct ib_mad_mgmt_vendor_class *vendor_class,
164 		struct ib_mad_reg_req *mad_reg_req)
165 {
166 	struct ib_mad_mgmt_method_table *method;
167 	int i;
168 
169 	for (i = 0; i < MAX_MGMT_OUI; i++) {
170 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
171 			method = vendor_class->method_table[i];
172 			if (method) {
173 				if (method_in_use(&method, mad_reg_req))
174 					return 1;
175 				else
176 					break;
177 			}
178 		}
179 	}
180 	return 0;
181 }
182 
183 int ib_response_mad(const struct ib_mad_hdr *hdr)
184 {
185 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
186 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
187 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
188 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
189 }
190 EXPORT_SYMBOL(ib_response_mad);
191 
192 /*
193  * ib_register_mad_agent - Register to send/receive MADs
194  */
195 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
196 					   u8 port_num,
197 					   enum ib_qp_type qp_type,
198 					   struct ib_mad_reg_req *mad_reg_req,
199 					   u8 rmpp_version,
200 					   ib_mad_send_handler send_handler,
201 					   ib_mad_recv_handler recv_handler,
202 					   void *context,
203 					   u32 registration_flags)
204 {
205 	struct ib_mad_port_private *port_priv;
206 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
207 	struct ib_mad_agent_private *mad_agent_priv;
208 	struct ib_mad_reg_req *reg_req = NULL;
209 	struct ib_mad_mgmt_class_table *class;
210 	struct ib_mad_mgmt_vendor_class_table *vendor;
211 	struct ib_mad_mgmt_vendor_class *vendor_class;
212 	struct ib_mad_mgmt_method_table *method;
213 	int ret2, qpn;
214 	unsigned long flags;
215 	u8 mgmt_class, vclass;
216 
217 	/* Validate parameters */
218 	qpn = get_spl_qp_index(qp_type);
219 	if (qpn == -1) {
220 		dev_notice(&device->dev,
221 			   "ib_register_mad_agent: invalid QP Type %d\n",
222 			   qp_type);
223 		goto error1;
224 	}
225 
226 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
227 		dev_notice(&device->dev,
228 			   "ib_register_mad_agent: invalid RMPP Version %u\n",
229 			   rmpp_version);
230 		goto error1;
231 	}
232 
233 	/* Validate MAD registration request if supplied */
234 	if (mad_reg_req) {
235 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
236 			dev_notice(&device->dev,
237 				   "ib_register_mad_agent: invalid Class Version %u\n",
238 				   mad_reg_req->mgmt_class_version);
239 			goto error1;
240 		}
241 		if (!recv_handler) {
242 			dev_notice(&device->dev,
243 				   "ib_register_mad_agent: no recv_handler\n");
244 			goto error1;
245 		}
246 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
247 			/*
248 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
249 			 * one in this range currently allowed
250 			 */
251 			if (mad_reg_req->mgmt_class !=
252 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
253 				dev_notice(&device->dev,
254 					   "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
255 					   mad_reg_req->mgmt_class);
256 				goto error1;
257 			}
258 		} else if (mad_reg_req->mgmt_class == 0) {
259 			/*
260 			 * Class 0 is reserved in IBA and is used for
261 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
262 			 */
263 			dev_notice(&device->dev,
264 				   "ib_register_mad_agent: Invalid Mgmt Class 0\n");
265 			goto error1;
266 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
267 			/*
268 			 * If class is in "new" vendor range,
269 			 * ensure supplied OUI is not zero
270 			 */
271 			if (!is_vendor_oui(mad_reg_req->oui)) {
272 				dev_notice(&device->dev,
273 					   "ib_register_mad_agent: No OUI specified for class 0x%x\n",
274 					   mad_reg_req->mgmt_class);
275 				goto error1;
276 			}
277 		}
278 		/* Make sure class supplied is consistent with RMPP */
279 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
280 			if (rmpp_version) {
281 				dev_notice(&device->dev,
282 					   "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
283 					   mad_reg_req->mgmt_class);
284 				goto error1;
285 			}
286 		}
287 
288 		/* Make sure class supplied is consistent with QP type */
289 		if (qp_type == IB_QPT_SMI) {
290 			if ((mad_reg_req->mgmt_class !=
291 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
292 			    (mad_reg_req->mgmt_class !=
293 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
294 				dev_notice(&device->dev,
295 					   "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
296 					   mad_reg_req->mgmt_class);
297 				goto error1;
298 			}
299 		} else {
300 			if ((mad_reg_req->mgmt_class ==
301 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
302 			    (mad_reg_req->mgmt_class ==
303 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
304 				dev_notice(&device->dev,
305 					   "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
306 					   mad_reg_req->mgmt_class);
307 				goto error1;
308 			}
309 		}
310 	} else {
311 		/* No registration request supplied */
312 		if (!send_handler)
313 			goto error1;
314 		if (registration_flags & IB_MAD_USER_RMPP)
315 			goto error1;
316 	}
317 
318 	/* Validate device and port */
319 	port_priv = ib_get_mad_port(device, port_num);
320 	if (!port_priv) {
321 		dev_notice(&device->dev,
322 			   "ib_register_mad_agent: Invalid port %d\n",
323 			   port_num);
324 		ret = ERR_PTR(-ENODEV);
325 		goto error1;
326 	}
327 
328 	/* Verify the QP requested is supported.  For example, Ethernet devices
329 	 * will not have QP0 */
330 	if (!port_priv->qp_info[qpn].qp) {
331 		dev_notice(&device->dev,
332 			   "ib_register_mad_agent: QP %d not supported\n", qpn);
333 		ret = ERR_PTR(-EPROTONOSUPPORT);
334 		goto error1;
335 	}
336 
337 	/* Allocate structures */
338 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
339 	if (!mad_agent_priv) {
340 		ret = ERR_PTR(-ENOMEM);
341 		goto error1;
342 	}
343 
344 	if (mad_reg_req) {
345 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
346 		if (!reg_req) {
347 			ret = ERR_PTR(-ENOMEM);
348 			goto error3;
349 		}
350 	}
351 
352 	/* Now, fill in the various structures */
353 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
354 	mad_agent_priv->reg_req = reg_req;
355 	mad_agent_priv->agent.rmpp_version = rmpp_version;
356 	mad_agent_priv->agent.device = device;
357 	mad_agent_priv->agent.recv_handler = recv_handler;
358 	mad_agent_priv->agent.send_handler = send_handler;
359 	mad_agent_priv->agent.context = context;
360 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
361 	mad_agent_priv->agent.port_num = port_num;
362 	mad_agent_priv->agent.flags = registration_flags;
363 	spin_lock_init(&mad_agent_priv->lock);
364 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
365 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
366 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
367 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
368 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
369 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
370 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
371 	atomic_set(&mad_agent_priv->refcount, 1);
372 	init_completion(&mad_agent_priv->comp);
373 
374 	ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
375 	if (ret2) {
376 		ret = ERR_PTR(ret2);
377 		goto error4;
378 	}
379 
380 	spin_lock_irqsave(&port_priv->reg_lock, flags);
381 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
382 
383 	/*
384 	 * Make sure MAD registration (if supplied)
385 	 * is non overlapping with any existing ones
386 	 */
387 	if (mad_reg_req) {
388 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
389 		if (!is_vendor_class(mgmt_class)) {
390 			class = port_priv->version[mad_reg_req->
391 						   mgmt_class_version].class;
392 			if (class) {
393 				method = class->method_table[mgmt_class];
394 				if (method) {
395 					if (method_in_use(&method,
396 							   mad_reg_req))
397 						goto error5;
398 				}
399 			}
400 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
401 						  mgmt_class);
402 		} else {
403 			/* "New" vendor class range */
404 			vendor = port_priv->version[mad_reg_req->
405 						    mgmt_class_version].vendor;
406 			if (vendor) {
407 				vclass = vendor_class_index(mgmt_class);
408 				vendor_class = vendor->vendor_class[vclass];
409 				if (vendor_class) {
410 					if (is_vendor_method_in_use(
411 							vendor_class,
412 							mad_reg_req))
413 						goto error5;
414 				}
415 			}
416 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
417 		}
418 		if (ret2) {
419 			ret = ERR_PTR(ret2);
420 			goto error5;
421 		}
422 	}
423 
424 	/* Add mad agent into port's agent list */
425 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
426 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
427 
428 	return &mad_agent_priv->agent;
429 error5:
430 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
431 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
432 error4:
433 	kfree(reg_req);
434 error3:
435 	kfree(mad_agent_priv);
436 error1:
437 	return ret;
438 }
439 EXPORT_SYMBOL(ib_register_mad_agent);
440 
441 static inline int is_snooping_sends(int mad_snoop_flags)
442 {
443 	return (mad_snoop_flags &
444 		(/*IB_MAD_SNOOP_POSTED_SENDS |
445 		 IB_MAD_SNOOP_RMPP_SENDS |*/
446 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
447 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
448 }
449 
450 static inline int is_snooping_recvs(int mad_snoop_flags)
451 {
452 	return (mad_snoop_flags &
453 		(IB_MAD_SNOOP_RECVS /*|
454 		 IB_MAD_SNOOP_RMPP_RECVS*/));
455 }
456 
457 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
458 				struct ib_mad_snoop_private *mad_snoop_priv)
459 {
460 	struct ib_mad_snoop_private **new_snoop_table;
461 	unsigned long flags;
462 	int i;
463 
464 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
465 	/* Check for empty slot in array. */
466 	for (i = 0; i < qp_info->snoop_table_size; i++)
467 		if (!qp_info->snoop_table[i])
468 			break;
469 
470 	if (i == qp_info->snoop_table_size) {
471 		/* Grow table. */
472 		new_snoop_table = krealloc(qp_info->snoop_table,
473 					   sizeof mad_snoop_priv *
474 					   (qp_info->snoop_table_size + 1),
475 					   GFP_ATOMIC);
476 		if (!new_snoop_table) {
477 			i = -ENOMEM;
478 			goto out;
479 		}
480 
481 		qp_info->snoop_table = new_snoop_table;
482 		qp_info->snoop_table_size++;
483 	}
484 	qp_info->snoop_table[i] = mad_snoop_priv;
485 	atomic_inc(&qp_info->snoop_count);
486 out:
487 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
488 	return i;
489 }
490 
491 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
492 					   u8 port_num,
493 					   enum ib_qp_type qp_type,
494 					   int mad_snoop_flags,
495 					   ib_mad_snoop_handler snoop_handler,
496 					   ib_mad_recv_handler recv_handler,
497 					   void *context)
498 {
499 	struct ib_mad_port_private *port_priv;
500 	struct ib_mad_agent *ret;
501 	struct ib_mad_snoop_private *mad_snoop_priv;
502 	int qpn;
503 	int err;
504 
505 	/* Validate parameters */
506 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
507 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
508 		ret = ERR_PTR(-EINVAL);
509 		goto error1;
510 	}
511 	qpn = get_spl_qp_index(qp_type);
512 	if (qpn == -1) {
513 		ret = ERR_PTR(-EINVAL);
514 		goto error1;
515 	}
516 	port_priv = ib_get_mad_port(device, port_num);
517 	if (!port_priv) {
518 		ret = ERR_PTR(-ENODEV);
519 		goto error1;
520 	}
521 	/* Allocate structures */
522 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
523 	if (!mad_snoop_priv) {
524 		ret = ERR_PTR(-ENOMEM);
525 		goto error1;
526 	}
527 
528 	/* Now, fill in the various structures */
529 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
530 	mad_snoop_priv->agent.device = device;
531 	mad_snoop_priv->agent.recv_handler = recv_handler;
532 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
533 	mad_snoop_priv->agent.context = context;
534 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
535 	mad_snoop_priv->agent.port_num = port_num;
536 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
537 	init_completion(&mad_snoop_priv->comp);
538 
539 	err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
540 	if (err) {
541 		ret = ERR_PTR(err);
542 		goto error2;
543 	}
544 
545 	mad_snoop_priv->snoop_index = register_snoop_agent(
546 						&port_priv->qp_info[qpn],
547 						mad_snoop_priv);
548 	if (mad_snoop_priv->snoop_index < 0) {
549 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
550 		goto error3;
551 	}
552 
553 	atomic_set(&mad_snoop_priv->refcount, 1);
554 	return &mad_snoop_priv->agent;
555 error3:
556 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
557 error2:
558 	kfree(mad_snoop_priv);
559 error1:
560 	return ret;
561 }
562 EXPORT_SYMBOL(ib_register_mad_snoop);
563 
564 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
565 {
566 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
567 		complete(&mad_agent_priv->comp);
568 }
569 
570 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
571 {
572 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
573 		complete(&mad_snoop_priv->comp);
574 }
575 
576 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
577 {
578 	struct ib_mad_port_private *port_priv;
579 	unsigned long flags;
580 
581 	/* Note that we could still be handling received MADs */
582 
583 	/*
584 	 * Canceling all sends results in dropping received response
585 	 * MADs, preventing us from queuing additional work
586 	 */
587 	cancel_mads(mad_agent_priv);
588 	port_priv = mad_agent_priv->qp_info->port_priv;
589 	cancel_delayed_work(&mad_agent_priv->timed_work);
590 
591 	spin_lock_irqsave(&port_priv->reg_lock, flags);
592 	remove_mad_reg_req(mad_agent_priv);
593 	list_del(&mad_agent_priv->agent_list);
594 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
595 
596 	flush_workqueue(port_priv->wq);
597 	ib_cancel_rmpp_recvs(mad_agent_priv);
598 
599 	deref_mad_agent(mad_agent_priv);
600 	wait_for_completion(&mad_agent_priv->comp);
601 
602 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
603 
604 	kfree(mad_agent_priv->reg_req);
605 	kfree(mad_agent_priv);
606 }
607 
608 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
609 {
610 	struct ib_mad_qp_info *qp_info;
611 	unsigned long flags;
612 
613 	qp_info = mad_snoop_priv->qp_info;
614 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
615 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
616 	atomic_dec(&qp_info->snoop_count);
617 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
618 
619 	deref_snoop_agent(mad_snoop_priv);
620 	wait_for_completion(&mad_snoop_priv->comp);
621 
622 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
623 
624 	kfree(mad_snoop_priv);
625 }
626 
627 /*
628  * ib_unregister_mad_agent - Unregisters a client from using MAD services
629  */
630 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
631 {
632 	struct ib_mad_agent_private *mad_agent_priv;
633 	struct ib_mad_snoop_private *mad_snoop_priv;
634 
635 	/* If the TID is zero, the agent can only snoop. */
636 	if (mad_agent->hi_tid) {
637 		mad_agent_priv = container_of(mad_agent,
638 					      struct ib_mad_agent_private,
639 					      agent);
640 		unregister_mad_agent(mad_agent_priv);
641 	} else {
642 		mad_snoop_priv = container_of(mad_agent,
643 					      struct ib_mad_snoop_private,
644 					      agent);
645 		unregister_mad_snoop(mad_snoop_priv);
646 	}
647 }
648 EXPORT_SYMBOL(ib_unregister_mad_agent);
649 
650 static void dequeue_mad(struct ib_mad_list_head *mad_list)
651 {
652 	struct ib_mad_queue *mad_queue;
653 	unsigned long flags;
654 
655 	BUG_ON(!mad_list->mad_queue);
656 	mad_queue = mad_list->mad_queue;
657 	spin_lock_irqsave(&mad_queue->lock, flags);
658 	list_del(&mad_list->list);
659 	mad_queue->count--;
660 	spin_unlock_irqrestore(&mad_queue->lock, flags);
661 }
662 
663 static void snoop_send(struct ib_mad_qp_info *qp_info,
664 		       struct ib_mad_send_buf *send_buf,
665 		       struct ib_mad_send_wc *mad_send_wc,
666 		       int mad_snoop_flags)
667 {
668 	struct ib_mad_snoop_private *mad_snoop_priv;
669 	unsigned long flags;
670 	int i;
671 
672 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
673 	for (i = 0; i < qp_info->snoop_table_size; i++) {
674 		mad_snoop_priv = qp_info->snoop_table[i];
675 		if (!mad_snoop_priv ||
676 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
677 			continue;
678 
679 		atomic_inc(&mad_snoop_priv->refcount);
680 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
681 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
682 						    send_buf, mad_send_wc);
683 		deref_snoop_agent(mad_snoop_priv);
684 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
685 	}
686 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
687 }
688 
689 static void snoop_recv(struct ib_mad_qp_info *qp_info,
690 		       struct ib_mad_recv_wc *mad_recv_wc,
691 		       int mad_snoop_flags)
692 {
693 	struct ib_mad_snoop_private *mad_snoop_priv;
694 	unsigned long flags;
695 	int i;
696 
697 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
698 	for (i = 0; i < qp_info->snoop_table_size; i++) {
699 		mad_snoop_priv = qp_info->snoop_table[i];
700 		if (!mad_snoop_priv ||
701 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
702 			continue;
703 
704 		atomic_inc(&mad_snoop_priv->refcount);
705 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
706 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
707 						   mad_recv_wc);
708 		deref_snoop_agent(mad_snoop_priv);
709 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
710 	}
711 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
712 }
713 
714 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
715 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
716 {
717 	memset(wc, 0, sizeof *wc);
718 	wc->wr_cqe = cqe;
719 	wc->status = IB_WC_SUCCESS;
720 	wc->opcode = IB_WC_RECV;
721 	wc->pkey_index = pkey_index;
722 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
723 	wc->src_qp = IB_QP0;
724 	wc->qp = qp;
725 	wc->slid = slid;
726 	wc->sl = 0;
727 	wc->dlid_path_bits = 0;
728 	wc->port_num = port_num;
729 }
730 
731 static size_t mad_priv_size(const struct ib_mad_private *mp)
732 {
733 	return sizeof(struct ib_mad_private) + mp->mad_size;
734 }
735 
736 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
737 {
738 	size_t size = sizeof(struct ib_mad_private) + mad_size;
739 	struct ib_mad_private *ret = kzalloc(size, flags);
740 
741 	if (ret)
742 		ret->mad_size = mad_size;
743 
744 	return ret;
745 }
746 
747 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
748 {
749 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
750 }
751 
752 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
753 {
754 	return sizeof(struct ib_grh) + mp->mad_size;
755 }
756 
757 /*
758  * Return 0 if SMP is to be sent
759  * Return 1 if SMP was consumed locally (whether or not solicited)
760  * Return < 0 if error
761  */
762 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
763 				  struct ib_mad_send_wr_private *mad_send_wr)
764 {
765 	int ret = 0;
766 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
767 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
768 	unsigned long flags;
769 	struct ib_mad_local_private *local;
770 	struct ib_mad_private *mad_priv;
771 	struct ib_mad_port_private *port_priv;
772 	struct ib_mad_agent_private *recv_mad_agent = NULL;
773 	struct ib_device *device = mad_agent_priv->agent.device;
774 	u8 port_num;
775 	struct ib_wc mad_wc;
776 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
777 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
778 	u16 out_mad_pkey_index = 0;
779 	u16 drslid;
780 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
781 				    mad_agent_priv->qp_info->port_priv->port_num);
782 
783 	if (rdma_cap_ib_switch(device) &&
784 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
785 		port_num = send_wr->port_num;
786 	else
787 		port_num = mad_agent_priv->agent.port_num;
788 
789 	/*
790 	 * Directed route handling starts if the initial LID routed part of
791 	 * a request or the ending LID routed part of a response is empty.
792 	 * If we are at the start of the LID routed part, don't update the
793 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
794 	 */
795 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
796 		u32 opa_drslid;
797 
798 		if ((opa_get_smp_direction(opa_smp)
799 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
800 		     OPA_LID_PERMISSIVE &&
801 		     opa_smi_handle_dr_smp_send(opa_smp,
802 						rdma_cap_ib_switch(device),
803 						port_num) == IB_SMI_DISCARD) {
804 			ret = -EINVAL;
805 			dev_err(&device->dev, "OPA Invalid directed route\n");
806 			goto out;
807 		}
808 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
809 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
810 		    opa_drslid & 0xffff0000) {
811 			ret = -EINVAL;
812 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
813 			       opa_drslid);
814 			goto out;
815 		}
816 		drslid = (u16)(opa_drslid & 0x0000ffff);
817 
818 		/* Check to post send on QP or process locally */
819 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
820 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
821 			goto out;
822 	} else {
823 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
824 		     IB_LID_PERMISSIVE &&
825 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
826 		     IB_SMI_DISCARD) {
827 			ret = -EINVAL;
828 			dev_err(&device->dev, "Invalid directed route\n");
829 			goto out;
830 		}
831 		drslid = be16_to_cpu(smp->dr_slid);
832 
833 		/* Check to post send on QP or process locally */
834 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
835 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
836 			goto out;
837 	}
838 
839 	local = kmalloc(sizeof *local, GFP_ATOMIC);
840 	if (!local) {
841 		ret = -ENOMEM;
842 		goto out;
843 	}
844 	local->mad_priv = NULL;
845 	local->recv_mad_agent = NULL;
846 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
847 	if (!mad_priv) {
848 		ret = -ENOMEM;
849 		kfree(local);
850 		goto out;
851 	}
852 
853 	build_smp_wc(mad_agent_priv->agent.qp,
854 		     send_wr->wr.wr_cqe, drslid,
855 		     send_wr->pkey_index,
856 		     send_wr->port_num, &mad_wc);
857 
858 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
859 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
860 					+ mad_send_wr->send_buf.data_len
861 					+ sizeof(struct ib_grh);
862 	}
863 
864 	/* No GRH for DR SMP */
865 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
866 				  (const struct ib_mad_hdr *)smp, mad_size,
867 				  (struct ib_mad_hdr *)mad_priv->mad,
868 				  &mad_size, &out_mad_pkey_index);
869 	switch (ret)
870 	{
871 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
872 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
873 		    mad_agent_priv->agent.recv_handler) {
874 			local->mad_priv = mad_priv;
875 			local->recv_mad_agent = mad_agent_priv;
876 			/*
877 			 * Reference MAD agent until receive
878 			 * side of local completion handled
879 			 */
880 			atomic_inc(&mad_agent_priv->refcount);
881 		} else
882 			kfree(mad_priv);
883 		break;
884 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
885 		kfree(mad_priv);
886 		break;
887 	case IB_MAD_RESULT_SUCCESS:
888 		/* Treat like an incoming receive MAD */
889 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
890 					    mad_agent_priv->agent.port_num);
891 		if (port_priv) {
892 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
893 			recv_mad_agent = find_mad_agent(port_priv,
894 						        (const struct ib_mad_hdr *)mad_priv->mad);
895 		}
896 		if (!port_priv || !recv_mad_agent) {
897 			/*
898 			 * No receiving agent so drop packet and
899 			 * generate send completion.
900 			 */
901 			kfree(mad_priv);
902 			break;
903 		}
904 		local->mad_priv = mad_priv;
905 		local->recv_mad_agent = recv_mad_agent;
906 		break;
907 	default:
908 		kfree(mad_priv);
909 		kfree(local);
910 		ret = -EINVAL;
911 		goto out;
912 	}
913 
914 	local->mad_send_wr = mad_send_wr;
915 	if (opa) {
916 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
917 		local->return_wc_byte_len = mad_size;
918 	}
919 	/* Reference MAD agent until send side of local completion handled */
920 	atomic_inc(&mad_agent_priv->refcount);
921 	/* Queue local completion to local list */
922 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
923 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
924 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
925 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
926 		   &mad_agent_priv->local_work);
927 	ret = 1;
928 out:
929 	return ret;
930 }
931 
932 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
933 {
934 	int seg_size, pad;
935 
936 	seg_size = mad_size - hdr_len;
937 	if (data_len && seg_size) {
938 		pad = seg_size - data_len % seg_size;
939 		return pad == seg_size ? 0 : pad;
940 	} else
941 		return seg_size;
942 }
943 
944 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
945 {
946 	struct ib_rmpp_segment *s, *t;
947 
948 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
949 		list_del(&s->list);
950 		kfree(s);
951 	}
952 }
953 
954 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
955 				size_t mad_size, gfp_t gfp_mask)
956 {
957 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
958 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
959 	struct ib_rmpp_segment *seg = NULL;
960 	int left, seg_size, pad;
961 
962 	send_buf->seg_size = mad_size - send_buf->hdr_len;
963 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
964 	seg_size = send_buf->seg_size;
965 	pad = send_wr->pad;
966 
967 	/* Allocate data segments. */
968 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
969 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
970 		if (!seg) {
971 			free_send_rmpp_list(send_wr);
972 			return -ENOMEM;
973 		}
974 		seg->num = ++send_buf->seg_count;
975 		list_add_tail(&seg->list, &send_wr->rmpp_list);
976 	}
977 
978 	/* Zero any padding */
979 	if (pad)
980 		memset(seg->data + seg_size - pad, 0, pad);
981 
982 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
983 					  agent.rmpp_version;
984 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
985 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
986 
987 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
988 					struct ib_rmpp_segment, list);
989 	send_wr->last_ack_seg = send_wr->cur_seg;
990 	return 0;
991 }
992 
993 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
994 {
995 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
996 }
997 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
998 
999 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
1000 					    u32 remote_qpn, u16 pkey_index,
1001 					    int rmpp_active,
1002 					    int hdr_len, int data_len,
1003 					    gfp_t gfp_mask,
1004 					    u8 base_version)
1005 {
1006 	struct ib_mad_agent_private *mad_agent_priv;
1007 	struct ib_mad_send_wr_private *mad_send_wr;
1008 	int pad, message_size, ret, size;
1009 	void *buf;
1010 	size_t mad_size;
1011 	bool opa;
1012 
1013 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1014 				      agent);
1015 
1016 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1017 
1018 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1019 		mad_size = sizeof(struct opa_mad);
1020 	else
1021 		mad_size = sizeof(struct ib_mad);
1022 
1023 	pad = get_pad_size(hdr_len, data_len, mad_size);
1024 	message_size = hdr_len + data_len + pad;
1025 
1026 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1027 		if (!rmpp_active && message_size > mad_size)
1028 			return ERR_PTR(-EINVAL);
1029 	} else
1030 		if (rmpp_active || message_size > mad_size)
1031 			return ERR_PTR(-EINVAL);
1032 
1033 	size = rmpp_active ? hdr_len : mad_size;
1034 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1035 	if (!buf)
1036 		return ERR_PTR(-ENOMEM);
1037 
1038 	mad_send_wr = buf + size;
1039 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1040 	mad_send_wr->send_buf.mad = buf;
1041 	mad_send_wr->send_buf.hdr_len = hdr_len;
1042 	mad_send_wr->send_buf.data_len = data_len;
1043 	mad_send_wr->pad = pad;
1044 
1045 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1046 	mad_send_wr->sg_list[0].length = hdr_len;
1047 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1048 
1049 	/* OPA MADs don't have to be the full 2048 bytes */
1050 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1051 	    data_len < mad_size - hdr_len)
1052 		mad_send_wr->sg_list[1].length = data_len;
1053 	else
1054 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1055 
1056 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1057 
1058 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1059 
1060 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1061 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1062 	mad_send_wr->send_wr.wr.num_sge = 2;
1063 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1064 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1065 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1066 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1067 	mad_send_wr->send_wr.pkey_index = pkey_index;
1068 
1069 	if (rmpp_active) {
1070 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1071 		if (ret) {
1072 			kfree(buf);
1073 			return ERR_PTR(ret);
1074 		}
1075 	}
1076 
1077 	mad_send_wr->send_buf.mad_agent = mad_agent;
1078 	atomic_inc(&mad_agent_priv->refcount);
1079 	return &mad_send_wr->send_buf;
1080 }
1081 EXPORT_SYMBOL(ib_create_send_mad);
1082 
1083 int ib_get_mad_data_offset(u8 mgmt_class)
1084 {
1085 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1086 		return IB_MGMT_SA_HDR;
1087 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1088 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1089 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1090 		return IB_MGMT_DEVICE_HDR;
1091 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1092 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1093 		return IB_MGMT_VENDOR_HDR;
1094 	else
1095 		return IB_MGMT_MAD_HDR;
1096 }
1097 EXPORT_SYMBOL(ib_get_mad_data_offset);
1098 
1099 int ib_is_mad_class_rmpp(u8 mgmt_class)
1100 {
1101 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1102 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1103 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1104 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1105 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1106 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1107 		return 1;
1108 	return 0;
1109 }
1110 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1111 
1112 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1113 {
1114 	struct ib_mad_send_wr_private *mad_send_wr;
1115 	struct list_head *list;
1116 
1117 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1118 				   send_buf);
1119 	list = &mad_send_wr->cur_seg->list;
1120 
1121 	if (mad_send_wr->cur_seg->num < seg_num) {
1122 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1123 			if (mad_send_wr->cur_seg->num == seg_num)
1124 				break;
1125 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1126 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1127 			if (mad_send_wr->cur_seg->num == seg_num)
1128 				break;
1129 	}
1130 	return mad_send_wr->cur_seg->data;
1131 }
1132 EXPORT_SYMBOL(ib_get_rmpp_segment);
1133 
1134 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1135 {
1136 	if (mad_send_wr->send_buf.seg_count)
1137 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1138 					   mad_send_wr->seg_num);
1139 	else
1140 		return mad_send_wr->send_buf.mad +
1141 		       mad_send_wr->send_buf.hdr_len;
1142 }
1143 
1144 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1145 {
1146 	struct ib_mad_agent_private *mad_agent_priv;
1147 	struct ib_mad_send_wr_private *mad_send_wr;
1148 
1149 	mad_agent_priv = container_of(send_buf->mad_agent,
1150 				      struct ib_mad_agent_private, agent);
1151 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1152 				   send_buf);
1153 
1154 	free_send_rmpp_list(mad_send_wr);
1155 	kfree(send_buf->mad);
1156 	deref_mad_agent(mad_agent_priv);
1157 }
1158 EXPORT_SYMBOL(ib_free_send_mad);
1159 
1160 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1161 {
1162 	struct ib_mad_qp_info *qp_info;
1163 	struct list_head *list;
1164 	struct ib_send_wr *bad_send_wr;
1165 	struct ib_mad_agent *mad_agent;
1166 	struct ib_sge *sge;
1167 	unsigned long flags;
1168 	int ret;
1169 
1170 	/* Set WR ID to find mad_send_wr upon completion */
1171 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1172 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1173 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1174 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1175 
1176 	mad_agent = mad_send_wr->send_buf.mad_agent;
1177 	sge = mad_send_wr->sg_list;
1178 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1179 					mad_send_wr->send_buf.mad,
1180 					sge[0].length,
1181 					DMA_TO_DEVICE);
1182 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1183 		return -ENOMEM;
1184 
1185 	mad_send_wr->header_mapping = sge[0].addr;
1186 
1187 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1188 					ib_get_payload(mad_send_wr),
1189 					sge[1].length,
1190 					DMA_TO_DEVICE);
1191 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1192 		ib_dma_unmap_single(mad_agent->device,
1193 				    mad_send_wr->header_mapping,
1194 				    sge[0].length, DMA_TO_DEVICE);
1195 		return -ENOMEM;
1196 	}
1197 	mad_send_wr->payload_mapping = sge[1].addr;
1198 
1199 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1200 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1201 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1202 				   &bad_send_wr);
1203 		list = &qp_info->send_queue.list;
1204 	} else {
1205 		ret = 0;
1206 		list = &qp_info->overflow_list;
1207 	}
1208 
1209 	if (!ret) {
1210 		qp_info->send_queue.count++;
1211 		list_add_tail(&mad_send_wr->mad_list.list, list);
1212 	}
1213 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1214 	if (ret) {
1215 		ib_dma_unmap_single(mad_agent->device,
1216 				    mad_send_wr->header_mapping,
1217 				    sge[0].length, DMA_TO_DEVICE);
1218 		ib_dma_unmap_single(mad_agent->device,
1219 				    mad_send_wr->payload_mapping,
1220 				    sge[1].length, DMA_TO_DEVICE);
1221 	}
1222 	return ret;
1223 }
1224 
1225 /*
1226  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1227  *  with the registered client
1228  */
1229 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1230 		     struct ib_mad_send_buf **bad_send_buf)
1231 {
1232 	struct ib_mad_agent_private *mad_agent_priv;
1233 	struct ib_mad_send_buf *next_send_buf;
1234 	struct ib_mad_send_wr_private *mad_send_wr;
1235 	unsigned long flags;
1236 	int ret = -EINVAL;
1237 
1238 	/* Walk list of send WRs and post each on send list */
1239 	for (; send_buf; send_buf = next_send_buf) {
1240 		mad_send_wr = container_of(send_buf,
1241 					   struct ib_mad_send_wr_private,
1242 					   send_buf);
1243 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1244 
1245 		ret = ib_mad_enforce_security(mad_agent_priv,
1246 					      mad_send_wr->send_wr.pkey_index);
1247 		if (ret)
1248 			goto error;
1249 
1250 		if (!send_buf->mad_agent->send_handler ||
1251 		    (send_buf->timeout_ms &&
1252 		     !send_buf->mad_agent->recv_handler)) {
1253 			ret = -EINVAL;
1254 			goto error;
1255 		}
1256 
1257 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1258 			if (mad_agent_priv->agent.rmpp_version) {
1259 				ret = -EINVAL;
1260 				goto error;
1261 			}
1262 		}
1263 
1264 		/*
1265 		 * Save pointer to next work request to post in case the
1266 		 * current one completes, and the user modifies the work
1267 		 * request associated with the completion
1268 		 */
1269 		next_send_buf = send_buf->next;
1270 		mad_send_wr->send_wr.ah = send_buf->ah;
1271 
1272 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1273 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1274 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1275 						     mad_send_wr);
1276 			if (ret < 0)		/* error */
1277 				goto error;
1278 			else if (ret == 1)	/* locally consumed */
1279 				continue;
1280 		}
1281 
1282 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1283 		/* Timeout will be updated after send completes */
1284 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1285 		mad_send_wr->max_retries = send_buf->retries;
1286 		mad_send_wr->retries_left = send_buf->retries;
1287 		send_buf->retries = 0;
1288 		/* Reference for work request to QP + response */
1289 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1290 		mad_send_wr->status = IB_WC_SUCCESS;
1291 
1292 		/* Reference MAD agent until send completes */
1293 		atomic_inc(&mad_agent_priv->refcount);
1294 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1295 		list_add_tail(&mad_send_wr->agent_list,
1296 			      &mad_agent_priv->send_list);
1297 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1298 
1299 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1300 			ret = ib_send_rmpp_mad(mad_send_wr);
1301 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1302 				ret = ib_send_mad(mad_send_wr);
1303 		} else
1304 			ret = ib_send_mad(mad_send_wr);
1305 		if (ret < 0) {
1306 			/* Fail send request */
1307 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1308 			list_del(&mad_send_wr->agent_list);
1309 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1310 			atomic_dec(&mad_agent_priv->refcount);
1311 			goto error;
1312 		}
1313 	}
1314 	return 0;
1315 error:
1316 	if (bad_send_buf)
1317 		*bad_send_buf = send_buf;
1318 	return ret;
1319 }
1320 EXPORT_SYMBOL(ib_post_send_mad);
1321 
1322 /*
1323  * ib_free_recv_mad - Returns data buffers used to receive
1324  *  a MAD to the access layer
1325  */
1326 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1327 {
1328 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1329 	struct ib_mad_private_header *mad_priv_hdr;
1330 	struct ib_mad_private *priv;
1331 	struct list_head free_list;
1332 
1333 	INIT_LIST_HEAD(&free_list);
1334 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1335 
1336 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1337 					&free_list, list) {
1338 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1339 					   recv_buf);
1340 		mad_priv_hdr = container_of(mad_recv_wc,
1341 					    struct ib_mad_private_header,
1342 					    recv_wc);
1343 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1344 				    header);
1345 		kfree(priv);
1346 	}
1347 }
1348 EXPORT_SYMBOL(ib_free_recv_mad);
1349 
1350 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1351 					u8 rmpp_version,
1352 					ib_mad_send_handler send_handler,
1353 					ib_mad_recv_handler recv_handler,
1354 					void *context)
1355 {
1356 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1357 }
1358 EXPORT_SYMBOL(ib_redirect_mad_qp);
1359 
1360 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1361 		      struct ib_wc *wc)
1362 {
1363 	dev_err(&mad_agent->device->dev,
1364 		"ib_process_mad_wc() not implemented yet\n");
1365 	return 0;
1366 }
1367 EXPORT_SYMBOL(ib_process_mad_wc);
1368 
1369 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1370 			 struct ib_mad_reg_req *mad_reg_req)
1371 {
1372 	int i;
1373 
1374 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1375 		if ((*method)->agent[i]) {
1376 			pr_err("Method %d already in use\n", i);
1377 			return -EINVAL;
1378 		}
1379 	}
1380 	return 0;
1381 }
1382 
1383 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1384 {
1385 	/* Allocate management method table */
1386 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1387 	return (*method) ? 0 : (-ENOMEM);
1388 }
1389 
1390 /*
1391  * Check to see if there are any methods still in use
1392  */
1393 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1394 {
1395 	int i;
1396 
1397 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1398 		if (method->agent[i])
1399 			return 1;
1400 	return 0;
1401 }
1402 
1403 /*
1404  * Check to see if there are any method tables for this class still in use
1405  */
1406 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1407 {
1408 	int i;
1409 
1410 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1411 		if (class->method_table[i])
1412 			return 1;
1413 	return 0;
1414 }
1415 
1416 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1417 {
1418 	int i;
1419 
1420 	for (i = 0; i < MAX_MGMT_OUI; i++)
1421 		if (vendor_class->method_table[i])
1422 			return 1;
1423 	return 0;
1424 }
1425 
1426 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1427 			   const char *oui)
1428 {
1429 	int i;
1430 
1431 	for (i = 0; i < MAX_MGMT_OUI; i++)
1432 		/* Is there matching OUI for this vendor class ? */
1433 		if (!memcmp(vendor_class->oui[i], oui, 3))
1434 			return i;
1435 
1436 	return -1;
1437 }
1438 
1439 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1440 {
1441 	int i;
1442 
1443 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1444 		if (vendor->vendor_class[i])
1445 			return 1;
1446 
1447 	return 0;
1448 }
1449 
1450 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1451 				     struct ib_mad_agent_private *agent)
1452 {
1453 	int i;
1454 
1455 	/* Remove any methods for this mad agent */
1456 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1457 		if (method->agent[i] == agent) {
1458 			method->agent[i] = NULL;
1459 		}
1460 	}
1461 }
1462 
1463 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1464 			      struct ib_mad_agent_private *agent_priv,
1465 			      u8 mgmt_class)
1466 {
1467 	struct ib_mad_port_private *port_priv;
1468 	struct ib_mad_mgmt_class_table **class;
1469 	struct ib_mad_mgmt_method_table **method;
1470 	int i, ret;
1471 
1472 	port_priv = agent_priv->qp_info->port_priv;
1473 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1474 	if (!*class) {
1475 		/* Allocate management class table for "new" class version */
1476 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1477 		if (!*class) {
1478 			ret = -ENOMEM;
1479 			goto error1;
1480 		}
1481 
1482 		/* Allocate method table for this management class */
1483 		method = &(*class)->method_table[mgmt_class];
1484 		if ((ret = allocate_method_table(method)))
1485 			goto error2;
1486 	} else {
1487 		method = &(*class)->method_table[mgmt_class];
1488 		if (!*method) {
1489 			/* Allocate method table for this management class */
1490 			if ((ret = allocate_method_table(method)))
1491 				goto error1;
1492 		}
1493 	}
1494 
1495 	/* Now, make sure methods are not already in use */
1496 	if (method_in_use(method, mad_reg_req))
1497 		goto error3;
1498 
1499 	/* Finally, add in methods being registered */
1500 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1501 		(*method)->agent[i] = agent_priv;
1502 
1503 	return 0;
1504 
1505 error3:
1506 	/* Remove any methods for this mad agent */
1507 	remove_methods_mad_agent(*method, agent_priv);
1508 	/* Now, check to see if there are any methods in use */
1509 	if (!check_method_table(*method)) {
1510 		/* If not, release management method table */
1511 		kfree(*method);
1512 		*method = NULL;
1513 	}
1514 	ret = -EINVAL;
1515 	goto error1;
1516 error2:
1517 	kfree(*class);
1518 	*class = NULL;
1519 error1:
1520 	return ret;
1521 }
1522 
1523 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1524 			   struct ib_mad_agent_private *agent_priv)
1525 {
1526 	struct ib_mad_port_private *port_priv;
1527 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1528 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1529 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1530 	struct ib_mad_mgmt_method_table **method;
1531 	int i, ret = -ENOMEM;
1532 	u8 vclass;
1533 
1534 	/* "New" vendor (with OUI) class */
1535 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1536 	port_priv = agent_priv->qp_info->port_priv;
1537 	vendor_table = &port_priv->version[
1538 				mad_reg_req->mgmt_class_version].vendor;
1539 	if (!*vendor_table) {
1540 		/* Allocate mgmt vendor class table for "new" class version */
1541 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1542 		if (!vendor)
1543 			goto error1;
1544 
1545 		*vendor_table = vendor;
1546 	}
1547 	if (!(*vendor_table)->vendor_class[vclass]) {
1548 		/* Allocate table for this management vendor class */
1549 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1550 		if (!vendor_class)
1551 			goto error2;
1552 
1553 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1554 	}
1555 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1556 		/* Is there matching OUI for this vendor class ? */
1557 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1558 			    mad_reg_req->oui, 3)) {
1559 			method = &(*vendor_table)->vendor_class[
1560 						vclass]->method_table[i];
1561 			BUG_ON(!*method);
1562 			goto check_in_use;
1563 		}
1564 	}
1565 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1566 		/* OUI slot available ? */
1567 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1568 				vclass]->oui[i])) {
1569 			method = &(*vendor_table)->vendor_class[
1570 				vclass]->method_table[i];
1571 			BUG_ON(*method);
1572 			/* Allocate method table for this OUI */
1573 			if ((ret = allocate_method_table(method)))
1574 				goto error3;
1575 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1576 			       mad_reg_req->oui, 3);
1577 			goto check_in_use;
1578 		}
1579 	}
1580 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1581 	goto error3;
1582 
1583 check_in_use:
1584 	/* Now, make sure methods are not already in use */
1585 	if (method_in_use(method, mad_reg_req))
1586 		goto error4;
1587 
1588 	/* Finally, add in methods being registered */
1589 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1590 		(*method)->agent[i] = agent_priv;
1591 
1592 	return 0;
1593 
1594 error4:
1595 	/* Remove any methods for this mad agent */
1596 	remove_methods_mad_agent(*method, agent_priv);
1597 	/* Now, check to see if there are any methods in use */
1598 	if (!check_method_table(*method)) {
1599 		/* If not, release management method table */
1600 		kfree(*method);
1601 		*method = NULL;
1602 	}
1603 	ret = -EINVAL;
1604 error3:
1605 	if (vendor_class) {
1606 		(*vendor_table)->vendor_class[vclass] = NULL;
1607 		kfree(vendor_class);
1608 	}
1609 error2:
1610 	if (vendor) {
1611 		*vendor_table = NULL;
1612 		kfree(vendor);
1613 	}
1614 error1:
1615 	return ret;
1616 }
1617 
1618 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1619 {
1620 	struct ib_mad_port_private *port_priv;
1621 	struct ib_mad_mgmt_class_table *class;
1622 	struct ib_mad_mgmt_method_table *method;
1623 	struct ib_mad_mgmt_vendor_class_table *vendor;
1624 	struct ib_mad_mgmt_vendor_class *vendor_class;
1625 	int index;
1626 	u8 mgmt_class;
1627 
1628 	/*
1629 	 * Was MAD registration request supplied
1630 	 * with original registration ?
1631 	 */
1632 	if (!agent_priv->reg_req) {
1633 		goto out;
1634 	}
1635 
1636 	port_priv = agent_priv->qp_info->port_priv;
1637 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1638 	class = port_priv->version[
1639 			agent_priv->reg_req->mgmt_class_version].class;
1640 	if (!class)
1641 		goto vendor_check;
1642 
1643 	method = class->method_table[mgmt_class];
1644 	if (method) {
1645 		/* Remove any methods for this mad agent */
1646 		remove_methods_mad_agent(method, agent_priv);
1647 		/* Now, check to see if there are any methods still in use */
1648 		if (!check_method_table(method)) {
1649 			/* If not, release management method table */
1650 			kfree(method);
1651 			class->method_table[mgmt_class] = NULL;
1652 			/* Any management classes left ? */
1653 			if (!check_class_table(class)) {
1654 				/* If not, release management class table */
1655 				kfree(class);
1656 				port_priv->version[
1657 					agent_priv->reg_req->
1658 					mgmt_class_version].class = NULL;
1659 			}
1660 		}
1661 	}
1662 
1663 vendor_check:
1664 	if (!is_vendor_class(mgmt_class))
1665 		goto out;
1666 
1667 	/* normalize mgmt_class to vendor range 2 */
1668 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1669 	vendor = port_priv->version[
1670 			agent_priv->reg_req->mgmt_class_version].vendor;
1671 
1672 	if (!vendor)
1673 		goto out;
1674 
1675 	vendor_class = vendor->vendor_class[mgmt_class];
1676 	if (vendor_class) {
1677 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1678 		if (index < 0)
1679 			goto out;
1680 		method = vendor_class->method_table[index];
1681 		if (method) {
1682 			/* Remove any methods for this mad agent */
1683 			remove_methods_mad_agent(method, agent_priv);
1684 			/*
1685 			 * Now, check to see if there are
1686 			 * any methods still in use
1687 			 */
1688 			if (!check_method_table(method)) {
1689 				/* If not, release management method table */
1690 				kfree(method);
1691 				vendor_class->method_table[index] = NULL;
1692 				memset(vendor_class->oui[index], 0, 3);
1693 				/* Any OUIs left ? */
1694 				if (!check_vendor_class(vendor_class)) {
1695 					/* If not, release vendor class table */
1696 					kfree(vendor_class);
1697 					vendor->vendor_class[mgmt_class] = NULL;
1698 					/* Any other vendor classes left ? */
1699 					if (!check_vendor_table(vendor)) {
1700 						kfree(vendor);
1701 						port_priv->version[
1702 							agent_priv->reg_req->
1703 							mgmt_class_version].
1704 							vendor = NULL;
1705 					}
1706 				}
1707 			}
1708 		}
1709 	}
1710 
1711 out:
1712 	return;
1713 }
1714 
1715 static struct ib_mad_agent_private *
1716 find_mad_agent(struct ib_mad_port_private *port_priv,
1717 	       const struct ib_mad_hdr *mad_hdr)
1718 {
1719 	struct ib_mad_agent_private *mad_agent = NULL;
1720 	unsigned long flags;
1721 
1722 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1723 	if (ib_response_mad(mad_hdr)) {
1724 		u32 hi_tid;
1725 		struct ib_mad_agent_private *entry;
1726 
1727 		/*
1728 		 * Routing is based on high 32 bits of transaction ID
1729 		 * of MAD.
1730 		 */
1731 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1732 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1733 			if (entry->agent.hi_tid == hi_tid) {
1734 				mad_agent = entry;
1735 				break;
1736 			}
1737 		}
1738 	} else {
1739 		struct ib_mad_mgmt_class_table *class;
1740 		struct ib_mad_mgmt_method_table *method;
1741 		struct ib_mad_mgmt_vendor_class_table *vendor;
1742 		struct ib_mad_mgmt_vendor_class *vendor_class;
1743 		const struct ib_vendor_mad *vendor_mad;
1744 		int index;
1745 
1746 		/*
1747 		 * Routing is based on version, class, and method
1748 		 * For "newer" vendor MADs, also based on OUI
1749 		 */
1750 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1751 			goto out;
1752 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1753 			class = port_priv->version[
1754 					mad_hdr->class_version].class;
1755 			if (!class)
1756 				goto out;
1757 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1758 			    ARRAY_SIZE(class->method_table))
1759 				goto out;
1760 			method = class->method_table[convert_mgmt_class(
1761 							mad_hdr->mgmt_class)];
1762 			if (method)
1763 				mad_agent = method->agent[mad_hdr->method &
1764 							  ~IB_MGMT_METHOD_RESP];
1765 		} else {
1766 			vendor = port_priv->version[
1767 					mad_hdr->class_version].vendor;
1768 			if (!vendor)
1769 				goto out;
1770 			vendor_class = vendor->vendor_class[vendor_class_index(
1771 						mad_hdr->mgmt_class)];
1772 			if (!vendor_class)
1773 				goto out;
1774 			/* Find matching OUI */
1775 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1776 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1777 			if (index == -1)
1778 				goto out;
1779 			method = vendor_class->method_table[index];
1780 			if (method) {
1781 				mad_agent = method->agent[mad_hdr->method &
1782 							  ~IB_MGMT_METHOD_RESP];
1783 			}
1784 		}
1785 	}
1786 
1787 	if (mad_agent) {
1788 		if (mad_agent->agent.recv_handler)
1789 			atomic_inc(&mad_agent->refcount);
1790 		else {
1791 			dev_notice(&port_priv->device->dev,
1792 				   "No receive handler for client %p on port %d\n",
1793 				   &mad_agent->agent, port_priv->port_num);
1794 			mad_agent = NULL;
1795 		}
1796 	}
1797 out:
1798 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1799 
1800 	return mad_agent;
1801 }
1802 
1803 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1804 			const struct ib_mad_qp_info *qp_info,
1805 			bool opa)
1806 {
1807 	int valid = 0;
1808 	u32 qp_num = qp_info->qp->qp_num;
1809 
1810 	/* Make sure MAD base version is understood */
1811 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1812 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1813 		pr_err("MAD received with unsupported base version %d %s\n",
1814 		       mad_hdr->base_version, opa ? "(opa)" : "");
1815 		goto out;
1816 	}
1817 
1818 	/* Filter SMI packets sent to other than QP0 */
1819 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1820 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1821 		if (qp_num == 0)
1822 			valid = 1;
1823 	} else {
1824 		/* CM attributes other than ClassPortInfo only use Send method */
1825 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1826 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1827 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1828 			goto out;
1829 		/* Filter GSI packets sent to QP0 */
1830 		if (qp_num != 0)
1831 			valid = 1;
1832 	}
1833 
1834 out:
1835 	return valid;
1836 }
1837 
1838 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1839 			    const struct ib_mad_hdr *mad_hdr)
1840 {
1841 	struct ib_rmpp_mad *rmpp_mad;
1842 
1843 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1844 	return !mad_agent_priv->agent.rmpp_version ||
1845 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1846 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1847 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1848 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1849 }
1850 
1851 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1852 				     const struct ib_mad_recv_wc *rwc)
1853 {
1854 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1855 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1856 }
1857 
1858 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1859 				   const struct ib_mad_send_wr_private *wr,
1860 				   const struct ib_mad_recv_wc *rwc )
1861 {
1862 	struct rdma_ah_attr attr;
1863 	u8 send_resp, rcv_resp;
1864 	union ib_gid sgid;
1865 	struct ib_device *device = mad_agent_priv->agent.device;
1866 	u8 port_num = mad_agent_priv->agent.port_num;
1867 	u8 lmc;
1868 	bool has_grh;
1869 
1870 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1871 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1872 
1873 	if (send_resp == rcv_resp)
1874 		/* both requests, or both responses. GIDs different */
1875 		return 0;
1876 
1877 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1878 		/* Assume not equal, to avoid false positives. */
1879 		return 0;
1880 
1881 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1882 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1883 		/* one has GID, other does not.  Assume different */
1884 		return 0;
1885 
1886 	if (!send_resp && rcv_resp) {
1887 		/* is request/response. */
1888 		if (!has_grh) {
1889 			if (ib_get_cached_lmc(device, port_num, &lmc))
1890 				return 0;
1891 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1892 					   rwc->wc->dlid_path_bits) &
1893 					  ((1 << lmc) - 1)));
1894 		} else {
1895 			const struct ib_global_route *grh =
1896 					rdma_ah_read_grh(&attr);
1897 
1898 			if (ib_get_cached_gid(device, port_num,
1899 					      grh->sgid_index, &sgid, NULL))
1900 				return 0;
1901 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1902 				       16);
1903 		}
1904 	}
1905 
1906 	if (!has_grh)
1907 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1908 	else
1909 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1910 			       rwc->recv_buf.grh->sgid.raw,
1911 			       16);
1912 }
1913 
1914 static inline int is_direct(u8 class)
1915 {
1916 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1917 }
1918 
1919 struct ib_mad_send_wr_private*
1920 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1921 		 const struct ib_mad_recv_wc *wc)
1922 {
1923 	struct ib_mad_send_wr_private *wr;
1924 	const struct ib_mad_hdr *mad_hdr;
1925 
1926 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1927 
1928 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1929 		if ((wr->tid == mad_hdr->tid) &&
1930 		    rcv_has_same_class(wr, wc) &&
1931 		    /*
1932 		     * Don't check GID for direct routed MADs.
1933 		     * These might have permissive LIDs.
1934 		     */
1935 		    (is_direct(mad_hdr->mgmt_class) ||
1936 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1937 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1938 	}
1939 
1940 	/*
1941 	 * It's possible to receive the response before we've
1942 	 * been notified that the send has completed
1943 	 */
1944 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1945 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1946 		    wr->tid == mad_hdr->tid &&
1947 		    wr->timeout &&
1948 		    rcv_has_same_class(wr, wc) &&
1949 		    /*
1950 		     * Don't check GID for direct routed MADs.
1951 		     * These might have permissive LIDs.
1952 		     */
1953 		    (is_direct(mad_hdr->mgmt_class) ||
1954 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1955 			/* Verify request has not been canceled */
1956 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1957 	}
1958 	return NULL;
1959 }
1960 
1961 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1962 {
1963 	mad_send_wr->timeout = 0;
1964 	if (mad_send_wr->refcount == 1)
1965 		list_move_tail(&mad_send_wr->agent_list,
1966 			      &mad_send_wr->mad_agent_priv->done_list);
1967 }
1968 
1969 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1970 				 struct ib_mad_recv_wc *mad_recv_wc)
1971 {
1972 	struct ib_mad_send_wr_private *mad_send_wr;
1973 	struct ib_mad_send_wc mad_send_wc;
1974 	unsigned long flags;
1975 	int ret;
1976 
1977 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1978 	ret = ib_mad_enforce_security(mad_agent_priv,
1979 				      mad_recv_wc->wc->pkey_index);
1980 	if (ret) {
1981 		ib_free_recv_mad(mad_recv_wc);
1982 		deref_mad_agent(mad_agent_priv);
1983 		return;
1984 	}
1985 
1986 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1987 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1988 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1989 						      mad_recv_wc);
1990 		if (!mad_recv_wc) {
1991 			deref_mad_agent(mad_agent_priv);
1992 			return;
1993 		}
1994 	}
1995 
1996 	/* Complete corresponding request */
1997 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1998 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1999 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2000 		if (!mad_send_wr) {
2001 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2002 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2003 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2004 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2005 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
2006 				/* user rmpp is in effect
2007 				 * and this is an active RMPP MAD
2008 				 */
2009 				mad_agent_priv->agent.recv_handler(
2010 						&mad_agent_priv->agent, NULL,
2011 						mad_recv_wc);
2012 				atomic_dec(&mad_agent_priv->refcount);
2013 			} else {
2014 				/* not user rmpp, revert to normal behavior and
2015 				 * drop the mad */
2016 				ib_free_recv_mad(mad_recv_wc);
2017 				deref_mad_agent(mad_agent_priv);
2018 				return;
2019 			}
2020 		} else {
2021 			ib_mark_mad_done(mad_send_wr);
2022 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2023 
2024 			/* Defined behavior is to complete response before request */
2025 			mad_agent_priv->agent.recv_handler(
2026 					&mad_agent_priv->agent,
2027 					&mad_send_wr->send_buf,
2028 					mad_recv_wc);
2029 			atomic_dec(&mad_agent_priv->refcount);
2030 
2031 			mad_send_wc.status = IB_WC_SUCCESS;
2032 			mad_send_wc.vendor_err = 0;
2033 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2034 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2035 		}
2036 	} else {
2037 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2038 						   mad_recv_wc);
2039 		deref_mad_agent(mad_agent_priv);
2040 	}
2041 
2042 	return;
2043 }
2044 
2045 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2046 				     const struct ib_mad_qp_info *qp_info,
2047 				     const struct ib_wc *wc,
2048 				     int port_num,
2049 				     struct ib_mad_private *recv,
2050 				     struct ib_mad_private *response)
2051 {
2052 	enum smi_forward_action retsmi;
2053 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2054 
2055 	if (smi_handle_dr_smp_recv(smp,
2056 				   rdma_cap_ib_switch(port_priv->device),
2057 				   port_num,
2058 				   port_priv->device->phys_port_cnt) ==
2059 				   IB_SMI_DISCARD)
2060 		return IB_SMI_DISCARD;
2061 
2062 	retsmi = smi_check_forward_dr_smp(smp);
2063 	if (retsmi == IB_SMI_LOCAL)
2064 		return IB_SMI_HANDLE;
2065 
2066 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2067 		if (smi_handle_dr_smp_send(smp,
2068 					   rdma_cap_ib_switch(port_priv->device),
2069 					   port_num) == IB_SMI_DISCARD)
2070 			return IB_SMI_DISCARD;
2071 
2072 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2073 			return IB_SMI_DISCARD;
2074 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2075 		/* forward case for switches */
2076 		memcpy(response, recv, mad_priv_size(response));
2077 		response->header.recv_wc.wc = &response->header.wc;
2078 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2079 		response->header.recv_wc.recv_buf.grh = &response->grh;
2080 
2081 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2082 				    &response->grh, wc,
2083 				    port_priv->device,
2084 				    smi_get_fwd_port(smp),
2085 				    qp_info->qp->qp_num,
2086 				    response->mad_size,
2087 				    false);
2088 
2089 		return IB_SMI_DISCARD;
2090 	}
2091 	return IB_SMI_HANDLE;
2092 }
2093 
2094 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2095 				    struct ib_mad_private *response,
2096 				    size_t *resp_len, bool opa)
2097 {
2098 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2099 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2100 
2101 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2102 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2103 		memcpy(response, recv, mad_priv_size(response));
2104 		response->header.recv_wc.wc = &response->header.wc;
2105 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2106 		response->header.recv_wc.recv_buf.grh = &response->grh;
2107 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2108 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2109 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2110 			resp_hdr->status |= IB_SMP_DIRECTION;
2111 
2112 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2113 			if (recv_hdr->mgmt_class ==
2114 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2115 			    recv_hdr->mgmt_class ==
2116 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2117 				*resp_len = opa_get_smp_header_size(
2118 							(struct opa_smp *)recv->mad);
2119 			else
2120 				*resp_len = sizeof(struct ib_mad_hdr);
2121 		}
2122 
2123 		return true;
2124 	} else {
2125 		return false;
2126 	}
2127 }
2128 
2129 static enum smi_action
2130 handle_opa_smi(struct ib_mad_port_private *port_priv,
2131 	       struct ib_mad_qp_info *qp_info,
2132 	       struct ib_wc *wc,
2133 	       int port_num,
2134 	       struct ib_mad_private *recv,
2135 	       struct ib_mad_private *response)
2136 {
2137 	enum smi_forward_action retsmi;
2138 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2139 
2140 	if (opa_smi_handle_dr_smp_recv(smp,
2141 				   rdma_cap_ib_switch(port_priv->device),
2142 				   port_num,
2143 				   port_priv->device->phys_port_cnt) ==
2144 				   IB_SMI_DISCARD)
2145 		return IB_SMI_DISCARD;
2146 
2147 	retsmi = opa_smi_check_forward_dr_smp(smp);
2148 	if (retsmi == IB_SMI_LOCAL)
2149 		return IB_SMI_HANDLE;
2150 
2151 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2152 		if (opa_smi_handle_dr_smp_send(smp,
2153 					   rdma_cap_ib_switch(port_priv->device),
2154 					   port_num) == IB_SMI_DISCARD)
2155 			return IB_SMI_DISCARD;
2156 
2157 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2158 		    IB_SMI_DISCARD)
2159 			return IB_SMI_DISCARD;
2160 
2161 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2162 		/* forward case for switches */
2163 		memcpy(response, recv, mad_priv_size(response));
2164 		response->header.recv_wc.wc = &response->header.wc;
2165 		response->header.recv_wc.recv_buf.opa_mad =
2166 				(struct opa_mad *)response->mad;
2167 		response->header.recv_wc.recv_buf.grh = &response->grh;
2168 
2169 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2170 				    &response->grh, wc,
2171 				    port_priv->device,
2172 				    opa_smi_get_fwd_port(smp),
2173 				    qp_info->qp->qp_num,
2174 				    recv->header.wc.byte_len,
2175 				    true);
2176 
2177 		return IB_SMI_DISCARD;
2178 	}
2179 
2180 	return IB_SMI_HANDLE;
2181 }
2182 
2183 static enum smi_action
2184 handle_smi(struct ib_mad_port_private *port_priv,
2185 	   struct ib_mad_qp_info *qp_info,
2186 	   struct ib_wc *wc,
2187 	   int port_num,
2188 	   struct ib_mad_private *recv,
2189 	   struct ib_mad_private *response,
2190 	   bool opa)
2191 {
2192 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2193 
2194 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2195 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2196 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2197 				      response);
2198 
2199 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2200 }
2201 
2202 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2203 {
2204 	struct ib_mad_port_private *port_priv = cq->cq_context;
2205 	struct ib_mad_list_head *mad_list =
2206 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2207 	struct ib_mad_qp_info *qp_info;
2208 	struct ib_mad_private_header *mad_priv_hdr;
2209 	struct ib_mad_private *recv, *response = NULL;
2210 	struct ib_mad_agent_private *mad_agent;
2211 	int port_num;
2212 	int ret = IB_MAD_RESULT_SUCCESS;
2213 	size_t mad_size;
2214 	u16 resp_mad_pkey_index = 0;
2215 	bool opa;
2216 
2217 	if (list_empty_careful(&port_priv->port_list))
2218 		return;
2219 
2220 	if (wc->status != IB_WC_SUCCESS) {
2221 		/*
2222 		 * Receive errors indicate that the QP has entered the error
2223 		 * state - error handling/shutdown code will cleanup
2224 		 */
2225 		return;
2226 	}
2227 
2228 	qp_info = mad_list->mad_queue->qp_info;
2229 	dequeue_mad(mad_list);
2230 
2231 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2232 			       qp_info->port_priv->port_num);
2233 
2234 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2235 				    mad_list);
2236 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2237 	ib_dma_unmap_single(port_priv->device,
2238 			    recv->header.mapping,
2239 			    mad_priv_dma_size(recv),
2240 			    DMA_FROM_DEVICE);
2241 
2242 	/* Setup MAD receive work completion from "normal" work completion */
2243 	recv->header.wc = *wc;
2244 	recv->header.recv_wc.wc = &recv->header.wc;
2245 
2246 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2247 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2248 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2249 	} else {
2250 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2251 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2252 	}
2253 
2254 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2255 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2256 
2257 	if (atomic_read(&qp_info->snoop_count))
2258 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2259 
2260 	/* Validate MAD */
2261 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2262 		goto out;
2263 
2264 	mad_size = recv->mad_size;
2265 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2266 	if (!response)
2267 		goto out;
2268 
2269 	if (rdma_cap_ib_switch(port_priv->device))
2270 		port_num = wc->port_num;
2271 	else
2272 		port_num = port_priv->port_num;
2273 
2274 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2275 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2276 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2277 			       response, opa)
2278 		    == IB_SMI_DISCARD)
2279 			goto out;
2280 	}
2281 
2282 	/* Give driver "right of first refusal" on incoming MAD */
2283 	if (port_priv->device->process_mad) {
2284 		ret = port_priv->device->process_mad(port_priv->device, 0,
2285 						     port_priv->port_num,
2286 						     wc, &recv->grh,
2287 						     (const struct ib_mad_hdr *)recv->mad,
2288 						     recv->mad_size,
2289 						     (struct ib_mad_hdr *)response->mad,
2290 						     &mad_size, &resp_mad_pkey_index);
2291 
2292 		if (opa)
2293 			wc->pkey_index = resp_mad_pkey_index;
2294 
2295 		if (ret & IB_MAD_RESULT_SUCCESS) {
2296 			if (ret & IB_MAD_RESULT_CONSUMED)
2297 				goto out;
2298 			if (ret & IB_MAD_RESULT_REPLY) {
2299 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2300 						    &recv->grh, wc,
2301 						    port_priv->device,
2302 						    port_num,
2303 						    qp_info->qp->qp_num,
2304 						    mad_size, opa);
2305 				goto out;
2306 			}
2307 		}
2308 	}
2309 
2310 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2311 	if (mad_agent) {
2312 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2313 		/*
2314 		 * recv is freed up in error cases in ib_mad_complete_recv
2315 		 * or via recv_handler in ib_mad_complete_recv()
2316 		 */
2317 		recv = NULL;
2318 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2319 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2320 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2321 				    port_priv->device, port_num,
2322 				    qp_info->qp->qp_num, mad_size, opa);
2323 	}
2324 
2325 out:
2326 	/* Post another receive request for this QP */
2327 	if (response) {
2328 		ib_mad_post_receive_mads(qp_info, response);
2329 		kfree(recv);
2330 	} else
2331 		ib_mad_post_receive_mads(qp_info, recv);
2332 }
2333 
2334 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2335 {
2336 	struct ib_mad_send_wr_private *mad_send_wr;
2337 	unsigned long delay;
2338 
2339 	if (list_empty(&mad_agent_priv->wait_list)) {
2340 		cancel_delayed_work(&mad_agent_priv->timed_work);
2341 	} else {
2342 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2343 					 struct ib_mad_send_wr_private,
2344 					 agent_list);
2345 
2346 		if (time_after(mad_agent_priv->timeout,
2347 			       mad_send_wr->timeout)) {
2348 			mad_agent_priv->timeout = mad_send_wr->timeout;
2349 			delay = mad_send_wr->timeout - jiffies;
2350 			if ((long)delay <= 0)
2351 				delay = 1;
2352 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2353 					 &mad_agent_priv->timed_work, delay);
2354 		}
2355 	}
2356 }
2357 
2358 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2359 {
2360 	struct ib_mad_agent_private *mad_agent_priv;
2361 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2362 	struct list_head *list_item;
2363 	unsigned long delay;
2364 
2365 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2366 	list_del(&mad_send_wr->agent_list);
2367 
2368 	delay = mad_send_wr->timeout;
2369 	mad_send_wr->timeout += jiffies;
2370 
2371 	if (delay) {
2372 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2373 			temp_mad_send_wr = list_entry(list_item,
2374 						struct ib_mad_send_wr_private,
2375 						agent_list);
2376 			if (time_after(mad_send_wr->timeout,
2377 				       temp_mad_send_wr->timeout))
2378 				break;
2379 		}
2380 	}
2381 	else
2382 		list_item = &mad_agent_priv->wait_list;
2383 	list_add(&mad_send_wr->agent_list, list_item);
2384 
2385 	/* Reschedule a work item if we have a shorter timeout */
2386 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2387 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2388 				 &mad_agent_priv->timed_work, delay);
2389 }
2390 
2391 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2392 			  int timeout_ms)
2393 {
2394 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2395 	wait_for_response(mad_send_wr);
2396 }
2397 
2398 /*
2399  * Process a send work completion
2400  */
2401 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2402 			     struct ib_mad_send_wc *mad_send_wc)
2403 {
2404 	struct ib_mad_agent_private	*mad_agent_priv;
2405 	unsigned long			flags;
2406 	int				ret;
2407 
2408 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2409 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2410 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2411 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2412 		if (ret == IB_RMPP_RESULT_CONSUMED)
2413 			goto done;
2414 	} else
2415 		ret = IB_RMPP_RESULT_UNHANDLED;
2416 
2417 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2418 	    mad_send_wr->status == IB_WC_SUCCESS) {
2419 		mad_send_wr->status = mad_send_wc->status;
2420 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2421 	}
2422 
2423 	if (--mad_send_wr->refcount > 0) {
2424 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2425 		    mad_send_wr->status == IB_WC_SUCCESS) {
2426 			wait_for_response(mad_send_wr);
2427 		}
2428 		goto done;
2429 	}
2430 
2431 	/* Remove send from MAD agent and notify client of completion */
2432 	list_del(&mad_send_wr->agent_list);
2433 	adjust_timeout(mad_agent_priv);
2434 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2435 
2436 	if (mad_send_wr->status != IB_WC_SUCCESS )
2437 		mad_send_wc->status = mad_send_wr->status;
2438 	if (ret == IB_RMPP_RESULT_INTERNAL)
2439 		ib_rmpp_send_handler(mad_send_wc);
2440 	else
2441 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2442 						   mad_send_wc);
2443 
2444 	/* Release reference on agent taken when sending */
2445 	deref_mad_agent(mad_agent_priv);
2446 	return;
2447 done:
2448 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2449 }
2450 
2451 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2452 {
2453 	struct ib_mad_port_private *port_priv = cq->cq_context;
2454 	struct ib_mad_list_head *mad_list =
2455 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2456 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2457 	struct ib_mad_qp_info		*qp_info;
2458 	struct ib_mad_queue		*send_queue;
2459 	struct ib_send_wr		*bad_send_wr;
2460 	struct ib_mad_send_wc		mad_send_wc;
2461 	unsigned long flags;
2462 	int ret;
2463 
2464 	if (list_empty_careful(&port_priv->port_list))
2465 		return;
2466 
2467 	if (wc->status != IB_WC_SUCCESS) {
2468 		if (!ib_mad_send_error(port_priv, wc))
2469 			return;
2470 	}
2471 
2472 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2473 				   mad_list);
2474 	send_queue = mad_list->mad_queue;
2475 	qp_info = send_queue->qp_info;
2476 
2477 retry:
2478 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2479 			    mad_send_wr->header_mapping,
2480 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2481 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2482 			    mad_send_wr->payload_mapping,
2483 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2484 	queued_send_wr = NULL;
2485 	spin_lock_irqsave(&send_queue->lock, flags);
2486 	list_del(&mad_list->list);
2487 
2488 	/* Move queued send to the send queue */
2489 	if (send_queue->count-- > send_queue->max_active) {
2490 		mad_list = container_of(qp_info->overflow_list.next,
2491 					struct ib_mad_list_head, list);
2492 		queued_send_wr = container_of(mad_list,
2493 					struct ib_mad_send_wr_private,
2494 					mad_list);
2495 		list_move_tail(&mad_list->list, &send_queue->list);
2496 	}
2497 	spin_unlock_irqrestore(&send_queue->lock, flags);
2498 
2499 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2500 	mad_send_wc.status = wc->status;
2501 	mad_send_wc.vendor_err = wc->vendor_err;
2502 	if (atomic_read(&qp_info->snoop_count))
2503 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2504 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2505 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2506 
2507 	if (queued_send_wr) {
2508 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2509 				   &bad_send_wr);
2510 		if (ret) {
2511 			dev_err(&port_priv->device->dev,
2512 				"ib_post_send failed: %d\n", ret);
2513 			mad_send_wr = queued_send_wr;
2514 			wc->status = IB_WC_LOC_QP_OP_ERR;
2515 			goto retry;
2516 		}
2517 	}
2518 }
2519 
2520 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2521 {
2522 	struct ib_mad_send_wr_private *mad_send_wr;
2523 	struct ib_mad_list_head *mad_list;
2524 	unsigned long flags;
2525 
2526 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2527 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2528 		mad_send_wr = container_of(mad_list,
2529 					   struct ib_mad_send_wr_private,
2530 					   mad_list);
2531 		mad_send_wr->retry = 1;
2532 	}
2533 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2534 }
2535 
2536 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2537 		struct ib_wc *wc)
2538 {
2539 	struct ib_mad_list_head *mad_list =
2540 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2541 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2542 	struct ib_mad_send_wr_private *mad_send_wr;
2543 	int ret;
2544 
2545 	/*
2546 	 * Send errors will transition the QP to SQE - move
2547 	 * QP to RTS and repost flushed work requests
2548 	 */
2549 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2550 				   mad_list);
2551 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2552 		if (mad_send_wr->retry) {
2553 			/* Repost send */
2554 			struct ib_send_wr *bad_send_wr;
2555 
2556 			mad_send_wr->retry = 0;
2557 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2558 					&bad_send_wr);
2559 			if (!ret)
2560 				return false;
2561 		}
2562 	} else {
2563 		struct ib_qp_attr *attr;
2564 
2565 		/* Transition QP to RTS and fail offending send */
2566 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2567 		if (attr) {
2568 			attr->qp_state = IB_QPS_RTS;
2569 			attr->cur_qp_state = IB_QPS_SQE;
2570 			ret = ib_modify_qp(qp_info->qp, attr,
2571 					   IB_QP_STATE | IB_QP_CUR_STATE);
2572 			kfree(attr);
2573 			if (ret)
2574 				dev_err(&port_priv->device->dev,
2575 					"%s - ib_modify_qp to RTS: %d\n",
2576 					__func__, ret);
2577 			else
2578 				mark_sends_for_retry(qp_info);
2579 		}
2580 	}
2581 
2582 	return true;
2583 }
2584 
2585 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2586 {
2587 	unsigned long flags;
2588 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2589 	struct ib_mad_send_wc mad_send_wc;
2590 	struct list_head cancel_list;
2591 
2592 	INIT_LIST_HEAD(&cancel_list);
2593 
2594 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2595 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2596 				 &mad_agent_priv->send_list, agent_list) {
2597 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2598 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2599 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2600 		}
2601 	}
2602 
2603 	/* Empty wait list to prevent receives from finding a request */
2604 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2605 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2606 
2607 	/* Report all cancelled requests */
2608 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2609 	mad_send_wc.vendor_err = 0;
2610 
2611 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2612 				 &cancel_list, agent_list) {
2613 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2614 		list_del(&mad_send_wr->agent_list);
2615 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2616 						   &mad_send_wc);
2617 		atomic_dec(&mad_agent_priv->refcount);
2618 	}
2619 }
2620 
2621 static struct ib_mad_send_wr_private*
2622 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2623 	     struct ib_mad_send_buf *send_buf)
2624 {
2625 	struct ib_mad_send_wr_private *mad_send_wr;
2626 
2627 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2628 			    agent_list) {
2629 		if (&mad_send_wr->send_buf == send_buf)
2630 			return mad_send_wr;
2631 	}
2632 
2633 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2634 			    agent_list) {
2635 		if (is_rmpp_data_mad(mad_agent_priv,
2636 				     mad_send_wr->send_buf.mad) &&
2637 		    &mad_send_wr->send_buf == send_buf)
2638 			return mad_send_wr;
2639 	}
2640 	return NULL;
2641 }
2642 
2643 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2644 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2645 {
2646 	struct ib_mad_agent_private *mad_agent_priv;
2647 	struct ib_mad_send_wr_private *mad_send_wr;
2648 	unsigned long flags;
2649 	int active;
2650 
2651 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2652 				      agent);
2653 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2654 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2655 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2656 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2657 		return -EINVAL;
2658 	}
2659 
2660 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2661 	if (!timeout_ms) {
2662 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2663 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2664 	}
2665 
2666 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2667 	if (active)
2668 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2669 	else
2670 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2671 
2672 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2673 	return 0;
2674 }
2675 EXPORT_SYMBOL(ib_modify_mad);
2676 
2677 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2678 		   struct ib_mad_send_buf *send_buf)
2679 {
2680 	ib_modify_mad(mad_agent, send_buf, 0);
2681 }
2682 EXPORT_SYMBOL(ib_cancel_mad);
2683 
2684 static void local_completions(struct work_struct *work)
2685 {
2686 	struct ib_mad_agent_private *mad_agent_priv;
2687 	struct ib_mad_local_private *local;
2688 	struct ib_mad_agent_private *recv_mad_agent;
2689 	unsigned long flags;
2690 	int free_mad;
2691 	struct ib_wc wc;
2692 	struct ib_mad_send_wc mad_send_wc;
2693 	bool opa;
2694 
2695 	mad_agent_priv =
2696 		container_of(work, struct ib_mad_agent_private, local_work);
2697 
2698 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2699 			       mad_agent_priv->qp_info->port_priv->port_num);
2700 
2701 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2702 	while (!list_empty(&mad_agent_priv->local_list)) {
2703 		local = list_entry(mad_agent_priv->local_list.next,
2704 				   struct ib_mad_local_private,
2705 				   completion_list);
2706 		list_del(&local->completion_list);
2707 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2708 		free_mad = 0;
2709 		if (local->mad_priv) {
2710 			u8 base_version;
2711 			recv_mad_agent = local->recv_mad_agent;
2712 			if (!recv_mad_agent) {
2713 				dev_err(&mad_agent_priv->agent.device->dev,
2714 					"No receive MAD agent for local completion\n");
2715 				free_mad = 1;
2716 				goto local_send_completion;
2717 			}
2718 
2719 			/*
2720 			 * Defined behavior is to complete response
2721 			 * before request
2722 			 */
2723 			build_smp_wc(recv_mad_agent->agent.qp,
2724 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2725 				     be16_to_cpu(IB_LID_PERMISSIVE),
2726 				     local->mad_send_wr->send_wr.pkey_index,
2727 				     recv_mad_agent->agent.port_num, &wc);
2728 
2729 			local->mad_priv->header.recv_wc.wc = &wc;
2730 
2731 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2732 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2733 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2734 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2735 			} else {
2736 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2737 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2738 			}
2739 
2740 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2741 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2742 				 &local->mad_priv->header.recv_wc.rmpp_list);
2743 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2744 			local->mad_priv->header.recv_wc.recv_buf.mad =
2745 						(struct ib_mad *)local->mad_priv->mad;
2746 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2747 				snoop_recv(recv_mad_agent->qp_info,
2748 					  &local->mad_priv->header.recv_wc,
2749 					   IB_MAD_SNOOP_RECVS);
2750 			recv_mad_agent->agent.recv_handler(
2751 						&recv_mad_agent->agent,
2752 						&local->mad_send_wr->send_buf,
2753 						&local->mad_priv->header.recv_wc);
2754 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2755 			atomic_dec(&recv_mad_agent->refcount);
2756 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2757 		}
2758 
2759 local_send_completion:
2760 		/* Complete send */
2761 		mad_send_wc.status = IB_WC_SUCCESS;
2762 		mad_send_wc.vendor_err = 0;
2763 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2764 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2765 			snoop_send(mad_agent_priv->qp_info,
2766 				   &local->mad_send_wr->send_buf,
2767 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2768 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2769 						   &mad_send_wc);
2770 
2771 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2772 		atomic_dec(&mad_agent_priv->refcount);
2773 		if (free_mad)
2774 			kfree(local->mad_priv);
2775 		kfree(local);
2776 	}
2777 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2778 }
2779 
2780 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2781 {
2782 	int ret;
2783 
2784 	if (!mad_send_wr->retries_left)
2785 		return -ETIMEDOUT;
2786 
2787 	mad_send_wr->retries_left--;
2788 	mad_send_wr->send_buf.retries++;
2789 
2790 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2791 
2792 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2793 		ret = ib_retry_rmpp(mad_send_wr);
2794 		switch (ret) {
2795 		case IB_RMPP_RESULT_UNHANDLED:
2796 			ret = ib_send_mad(mad_send_wr);
2797 			break;
2798 		case IB_RMPP_RESULT_CONSUMED:
2799 			ret = 0;
2800 			break;
2801 		default:
2802 			ret = -ECOMM;
2803 			break;
2804 		}
2805 	} else
2806 		ret = ib_send_mad(mad_send_wr);
2807 
2808 	if (!ret) {
2809 		mad_send_wr->refcount++;
2810 		list_add_tail(&mad_send_wr->agent_list,
2811 			      &mad_send_wr->mad_agent_priv->send_list);
2812 	}
2813 	return ret;
2814 }
2815 
2816 static void timeout_sends(struct work_struct *work)
2817 {
2818 	struct ib_mad_agent_private *mad_agent_priv;
2819 	struct ib_mad_send_wr_private *mad_send_wr;
2820 	struct ib_mad_send_wc mad_send_wc;
2821 	unsigned long flags, delay;
2822 
2823 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2824 				      timed_work.work);
2825 	mad_send_wc.vendor_err = 0;
2826 
2827 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2828 	while (!list_empty(&mad_agent_priv->wait_list)) {
2829 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2830 					 struct ib_mad_send_wr_private,
2831 					 agent_list);
2832 
2833 		if (time_after(mad_send_wr->timeout, jiffies)) {
2834 			delay = mad_send_wr->timeout - jiffies;
2835 			if ((long)delay <= 0)
2836 				delay = 1;
2837 			queue_delayed_work(mad_agent_priv->qp_info->
2838 					   port_priv->wq,
2839 					   &mad_agent_priv->timed_work, delay);
2840 			break;
2841 		}
2842 
2843 		list_del(&mad_send_wr->agent_list);
2844 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2845 		    !retry_send(mad_send_wr))
2846 			continue;
2847 
2848 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2849 
2850 		if (mad_send_wr->status == IB_WC_SUCCESS)
2851 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2852 		else
2853 			mad_send_wc.status = mad_send_wr->status;
2854 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2855 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2856 						   &mad_send_wc);
2857 
2858 		atomic_dec(&mad_agent_priv->refcount);
2859 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2860 	}
2861 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2862 }
2863 
2864 /*
2865  * Allocate receive MADs and post receive WRs for them
2866  */
2867 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2868 				    struct ib_mad_private *mad)
2869 {
2870 	unsigned long flags;
2871 	int post, ret;
2872 	struct ib_mad_private *mad_priv;
2873 	struct ib_sge sg_list;
2874 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2875 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2876 
2877 	/* Initialize common scatter list fields */
2878 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2879 
2880 	/* Initialize common receive WR fields */
2881 	recv_wr.next = NULL;
2882 	recv_wr.sg_list = &sg_list;
2883 	recv_wr.num_sge = 1;
2884 
2885 	do {
2886 		/* Allocate and map receive buffer */
2887 		if (mad) {
2888 			mad_priv = mad;
2889 			mad = NULL;
2890 		} else {
2891 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2892 						     GFP_ATOMIC);
2893 			if (!mad_priv) {
2894 				ret = -ENOMEM;
2895 				break;
2896 			}
2897 		}
2898 		sg_list.length = mad_priv_dma_size(mad_priv);
2899 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2900 						 &mad_priv->grh,
2901 						 mad_priv_dma_size(mad_priv),
2902 						 DMA_FROM_DEVICE);
2903 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2904 						  sg_list.addr))) {
2905 			ret = -ENOMEM;
2906 			break;
2907 		}
2908 		mad_priv->header.mapping = sg_list.addr;
2909 		mad_priv->header.mad_list.mad_queue = recv_queue;
2910 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2911 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2912 
2913 		/* Post receive WR */
2914 		spin_lock_irqsave(&recv_queue->lock, flags);
2915 		post = (++recv_queue->count < recv_queue->max_active);
2916 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2917 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2918 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2919 		if (ret) {
2920 			spin_lock_irqsave(&recv_queue->lock, flags);
2921 			list_del(&mad_priv->header.mad_list.list);
2922 			recv_queue->count--;
2923 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2924 			ib_dma_unmap_single(qp_info->port_priv->device,
2925 					    mad_priv->header.mapping,
2926 					    mad_priv_dma_size(mad_priv),
2927 					    DMA_FROM_DEVICE);
2928 			kfree(mad_priv);
2929 			dev_err(&qp_info->port_priv->device->dev,
2930 				"ib_post_recv failed: %d\n", ret);
2931 			break;
2932 		}
2933 	} while (post);
2934 
2935 	return ret;
2936 }
2937 
2938 /*
2939  * Return all the posted receive MADs
2940  */
2941 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2942 {
2943 	struct ib_mad_private_header *mad_priv_hdr;
2944 	struct ib_mad_private *recv;
2945 	struct ib_mad_list_head *mad_list;
2946 
2947 	if (!qp_info->qp)
2948 		return;
2949 
2950 	while (!list_empty(&qp_info->recv_queue.list)) {
2951 
2952 		mad_list = list_entry(qp_info->recv_queue.list.next,
2953 				      struct ib_mad_list_head, list);
2954 		mad_priv_hdr = container_of(mad_list,
2955 					    struct ib_mad_private_header,
2956 					    mad_list);
2957 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2958 				    header);
2959 
2960 		/* Remove from posted receive MAD list */
2961 		list_del(&mad_list->list);
2962 
2963 		ib_dma_unmap_single(qp_info->port_priv->device,
2964 				    recv->header.mapping,
2965 				    mad_priv_dma_size(recv),
2966 				    DMA_FROM_DEVICE);
2967 		kfree(recv);
2968 	}
2969 
2970 	qp_info->recv_queue.count = 0;
2971 }
2972 
2973 /*
2974  * Start the port
2975  */
2976 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2977 {
2978 	int ret, i;
2979 	struct ib_qp_attr *attr;
2980 	struct ib_qp *qp;
2981 	u16 pkey_index;
2982 
2983 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2984 	if (!attr)
2985 		return -ENOMEM;
2986 
2987 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2988 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2989 	if (ret)
2990 		pkey_index = 0;
2991 
2992 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2993 		qp = port_priv->qp_info[i].qp;
2994 		if (!qp)
2995 			continue;
2996 
2997 		/*
2998 		 * PKey index for QP1 is irrelevant but
2999 		 * one is needed for the Reset to Init transition
3000 		 */
3001 		attr->qp_state = IB_QPS_INIT;
3002 		attr->pkey_index = pkey_index;
3003 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3004 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3005 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
3006 		if (ret) {
3007 			dev_err(&port_priv->device->dev,
3008 				"Couldn't change QP%d state to INIT: %d\n",
3009 				i, ret);
3010 			goto out;
3011 		}
3012 
3013 		attr->qp_state = IB_QPS_RTR;
3014 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3015 		if (ret) {
3016 			dev_err(&port_priv->device->dev,
3017 				"Couldn't change QP%d state to RTR: %d\n",
3018 				i, ret);
3019 			goto out;
3020 		}
3021 
3022 		attr->qp_state = IB_QPS_RTS;
3023 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3024 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3025 		if (ret) {
3026 			dev_err(&port_priv->device->dev,
3027 				"Couldn't change QP%d state to RTS: %d\n",
3028 				i, ret);
3029 			goto out;
3030 		}
3031 	}
3032 
3033 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3034 	if (ret) {
3035 		dev_err(&port_priv->device->dev,
3036 			"Failed to request completion notification: %d\n",
3037 			ret);
3038 		goto out;
3039 	}
3040 
3041 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3042 		if (!port_priv->qp_info[i].qp)
3043 			continue;
3044 
3045 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3046 		if (ret) {
3047 			dev_err(&port_priv->device->dev,
3048 				"Couldn't post receive WRs\n");
3049 			goto out;
3050 		}
3051 	}
3052 out:
3053 	kfree(attr);
3054 	return ret;
3055 }
3056 
3057 static void qp_event_handler(struct ib_event *event, void *qp_context)
3058 {
3059 	struct ib_mad_qp_info	*qp_info = qp_context;
3060 
3061 	/* It's worse than that! He's dead, Jim! */
3062 	dev_err(&qp_info->port_priv->device->dev,
3063 		"Fatal error (%d) on MAD QP (%d)\n",
3064 		event->event, qp_info->qp->qp_num);
3065 }
3066 
3067 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3068 			   struct ib_mad_queue *mad_queue)
3069 {
3070 	mad_queue->qp_info = qp_info;
3071 	mad_queue->count = 0;
3072 	spin_lock_init(&mad_queue->lock);
3073 	INIT_LIST_HEAD(&mad_queue->list);
3074 }
3075 
3076 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3077 			struct ib_mad_qp_info *qp_info)
3078 {
3079 	qp_info->port_priv = port_priv;
3080 	init_mad_queue(qp_info, &qp_info->send_queue);
3081 	init_mad_queue(qp_info, &qp_info->recv_queue);
3082 	INIT_LIST_HEAD(&qp_info->overflow_list);
3083 	spin_lock_init(&qp_info->snoop_lock);
3084 	qp_info->snoop_table = NULL;
3085 	qp_info->snoop_table_size = 0;
3086 	atomic_set(&qp_info->snoop_count, 0);
3087 }
3088 
3089 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3090 			 enum ib_qp_type qp_type)
3091 {
3092 	struct ib_qp_init_attr	qp_init_attr;
3093 	int ret;
3094 
3095 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3096 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3097 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3098 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3099 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3100 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3101 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3102 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3103 	qp_init_attr.qp_type = qp_type;
3104 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3105 	qp_init_attr.qp_context = qp_info;
3106 	qp_init_attr.event_handler = qp_event_handler;
3107 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3108 	if (IS_ERR(qp_info->qp)) {
3109 		dev_err(&qp_info->port_priv->device->dev,
3110 			"Couldn't create ib_mad QP%d\n",
3111 			get_spl_qp_index(qp_type));
3112 		ret = PTR_ERR(qp_info->qp);
3113 		goto error;
3114 	}
3115 	/* Use minimum queue sizes unless the CQ is resized */
3116 	qp_info->send_queue.max_active = mad_sendq_size;
3117 	qp_info->recv_queue.max_active = mad_recvq_size;
3118 	return 0;
3119 
3120 error:
3121 	return ret;
3122 }
3123 
3124 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3125 {
3126 	if (!qp_info->qp)
3127 		return;
3128 
3129 	ib_destroy_qp(qp_info->qp);
3130 	kfree(qp_info->snoop_table);
3131 }
3132 
3133 /*
3134  * Open the port
3135  * Create the QP, PD, MR, and CQ if needed
3136  */
3137 static int ib_mad_port_open(struct ib_device *device,
3138 			    int port_num)
3139 {
3140 	int ret, cq_size;
3141 	struct ib_mad_port_private *port_priv;
3142 	unsigned long flags;
3143 	char name[sizeof "ib_mad123"];
3144 	int has_smi;
3145 
3146 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3147 		return -EFAULT;
3148 
3149 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3150 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3151 		return -EFAULT;
3152 
3153 	/* Create new device info */
3154 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3155 	if (!port_priv)
3156 		return -ENOMEM;
3157 
3158 	port_priv->device = device;
3159 	port_priv->port_num = port_num;
3160 	spin_lock_init(&port_priv->reg_lock);
3161 	INIT_LIST_HEAD(&port_priv->agent_list);
3162 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3163 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3164 
3165 	cq_size = mad_sendq_size + mad_recvq_size;
3166 	has_smi = rdma_cap_ib_smi(device, port_num);
3167 	if (has_smi)
3168 		cq_size *= 2;
3169 
3170 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3171 			IB_POLL_WORKQUEUE);
3172 	if (IS_ERR(port_priv->cq)) {
3173 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3174 		ret = PTR_ERR(port_priv->cq);
3175 		goto error3;
3176 	}
3177 
3178 	port_priv->pd = ib_alloc_pd(device, 0);
3179 	if (IS_ERR(port_priv->pd)) {
3180 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3181 		ret = PTR_ERR(port_priv->pd);
3182 		goto error4;
3183 	}
3184 
3185 	if (has_smi) {
3186 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3187 		if (ret)
3188 			goto error6;
3189 	}
3190 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3191 	if (ret)
3192 		goto error7;
3193 
3194 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3195 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3196 	if (!port_priv->wq) {
3197 		ret = -ENOMEM;
3198 		goto error8;
3199 	}
3200 
3201 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3202 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3203 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3204 
3205 	ret = ib_mad_port_start(port_priv);
3206 	if (ret) {
3207 		dev_err(&device->dev, "Couldn't start port\n");
3208 		goto error9;
3209 	}
3210 
3211 	return 0;
3212 
3213 error9:
3214 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3215 	list_del_init(&port_priv->port_list);
3216 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3217 
3218 	destroy_workqueue(port_priv->wq);
3219 error8:
3220 	destroy_mad_qp(&port_priv->qp_info[1]);
3221 error7:
3222 	destroy_mad_qp(&port_priv->qp_info[0]);
3223 error6:
3224 	ib_dealloc_pd(port_priv->pd);
3225 error4:
3226 	ib_free_cq(port_priv->cq);
3227 	cleanup_recv_queue(&port_priv->qp_info[1]);
3228 	cleanup_recv_queue(&port_priv->qp_info[0]);
3229 error3:
3230 	kfree(port_priv);
3231 
3232 	return ret;
3233 }
3234 
3235 /*
3236  * Close the port
3237  * If there are no classes using the port, free the port
3238  * resources (CQ, MR, PD, QP) and remove the port's info structure
3239  */
3240 static int ib_mad_port_close(struct ib_device *device, int port_num)
3241 {
3242 	struct ib_mad_port_private *port_priv;
3243 	unsigned long flags;
3244 
3245 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3246 	port_priv = __ib_get_mad_port(device, port_num);
3247 	if (port_priv == NULL) {
3248 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3249 		dev_err(&device->dev, "Port %d not found\n", port_num);
3250 		return -ENODEV;
3251 	}
3252 	list_del_init(&port_priv->port_list);
3253 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3254 
3255 	destroy_workqueue(port_priv->wq);
3256 	destroy_mad_qp(&port_priv->qp_info[1]);
3257 	destroy_mad_qp(&port_priv->qp_info[0]);
3258 	ib_dealloc_pd(port_priv->pd);
3259 	ib_free_cq(port_priv->cq);
3260 	cleanup_recv_queue(&port_priv->qp_info[1]);
3261 	cleanup_recv_queue(&port_priv->qp_info[0]);
3262 	/* XXX: Handle deallocation of MAD registration tables */
3263 
3264 	kfree(port_priv);
3265 
3266 	return 0;
3267 }
3268 
3269 static void ib_mad_init_device(struct ib_device *device)
3270 {
3271 	int start, i;
3272 
3273 	start = rdma_start_port(device);
3274 
3275 	for (i = start; i <= rdma_end_port(device); i++) {
3276 		if (!rdma_cap_ib_mad(device, i))
3277 			continue;
3278 
3279 		if (ib_mad_port_open(device, i)) {
3280 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3281 			goto error;
3282 		}
3283 		if (ib_agent_port_open(device, i)) {
3284 			dev_err(&device->dev,
3285 				"Couldn't open port %d for agents\n", i);
3286 			goto error_agent;
3287 		}
3288 	}
3289 	return;
3290 
3291 error_agent:
3292 	if (ib_mad_port_close(device, i))
3293 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3294 
3295 error:
3296 	while (--i >= start) {
3297 		if (!rdma_cap_ib_mad(device, i))
3298 			continue;
3299 
3300 		if (ib_agent_port_close(device, i))
3301 			dev_err(&device->dev,
3302 				"Couldn't close port %d for agents\n", i);
3303 		if (ib_mad_port_close(device, i))
3304 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3305 	}
3306 }
3307 
3308 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3309 {
3310 	int i;
3311 
3312 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3313 		if (!rdma_cap_ib_mad(device, i))
3314 			continue;
3315 
3316 		if (ib_agent_port_close(device, i))
3317 			dev_err(&device->dev,
3318 				"Couldn't close port %d for agents\n", i);
3319 		if (ib_mad_port_close(device, i))
3320 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3321 	}
3322 }
3323 
3324 static struct ib_client mad_client = {
3325 	.name   = "mad",
3326 	.add = ib_mad_init_device,
3327 	.remove = ib_mad_remove_device
3328 };
3329 
3330 int ib_mad_init(void)
3331 {
3332 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3333 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3334 
3335 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3336 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3337 
3338 	INIT_LIST_HEAD(&ib_mad_port_list);
3339 
3340 	if (ib_register_client(&mad_client)) {
3341 		pr_err("Couldn't register ib_mad client\n");
3342 		return -EINVAL;
3343 	}
3344 
3345 	return 0;
3346 }
3347 
3348 void ib_mad_cleanup(void)
3349 {
3350 	ib_unregister_client(&mad_client);
3351 }
3352