xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision a06c488d)
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 <rdma/ib_cache.h>
44 
45 #include "mad_priv.h"
46 #include "mad_rmpp.h"
47 #include "smi.h"
48 #include "opa_smi.h"
49 #include "agent.h"
50 
51 MODULE_LICENSE("Dual BSD/GPL");
52 MODULE_DESCRIPTION("kernel IB MAD API");
53 MODULE_AUTHOR("Hal Rosenstock");
54 MODULE_AUTHOR("Sean Hefty");
55 
56 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
57 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
58 
59 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
61 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
63 
64 static struct list_head ib_mad_port_list;
65 static u32 ib_mad_client_id = 0;
66 
67 /* Port list lock */
68 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
69 
70 /* Forward declarations */
71 static int method_in_use(struct ib_mad_mgmt_method_table **method,
72 			 struct ib_mad_reg_req *mad_reg_req);
73 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
74 static struct ib_mad_agent_private *find_mad_agent(
75 					struct ib_mad_port_private *port_priv,
76 					const struct ib_mad_hdr *mad);
77 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
78 				    struct ib_mad_private *mad);
79 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
80 static void timeout_sends(struct work_struct *work);
81 static void local_completions(struct work_struct *work);
82 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
83 			      struct ib_mad_agent_private *agent_priv,
84 			      u8 mgmt_class);
85 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
86 			   struct ib_mad_agent_private *agent_priv);
87 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
88 			      struct ib_wc *wc);
89 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
90 
91 /*
92  * Returns a ib_mad_port_private structure or NULL for a device/port
93  * Assumes ib_mad_port_list_lock is being held
94  */
95 static inline struct ib_mad_port_private *
96 __ib_get_mad_port(struct ib_device *device, int port_num)
97 {
98 	struct ib_mad_port_private *entry;
99 
100 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
101 		if (entry->device == device && entry->port_num == port_num)
102 			return entry;
103 	}
104 	return NULL;
105 }
106 
107 /*
108  * Wrapper function to return a ib_mad_port_private structure or NULL
109  * for a device/port
110  */
111 static inline struct ib_mad_port_private *
112 ib_get_mad_port(struct ib_device *device, int port_num)
113 {
114 	struct ib_mad_port_private *entry;
115 	unsigned long flags;
116 
117 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
118 	entry = __ib_get_mad_port(device, port_num);
119 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
120 
121 	return entry;
122 }
123 
124 static inline u8 convert_mgmt_class(u8 mgmt_class)
125 {
126 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
127 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
128 		0 : mgmt_class;
129 }
130 
131 static int get_spl_qp_index(enum ib_qp_type qp_type)
132 {
133 	switch (qp_type)
134 	{
135 	case IB_QPT_SMI:
136 		return 0;
137 	case IB_QPT_GSI:
138 		return 1;
139 	default:
140 		return -1;
141 	}
142 }
143 
144 static int vendor_class_index(u8 mgmt_class)
145 {
146 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
147 }
148 
149 static int is_vendor_class(u8 mgmt_class)
150 {
151 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
152 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
153 		return 0;
154 	return 1;
155 }
156 
157 static int is_vendor_oui(char *oui)
158 {
159 	if (oui[0] || oui[1] || oui[2])
160 		return 1;
161 	return 0;
162 }
163 
164 static int is_vendor_method_in_use(
165 		struct ib_mad_mgmt_vendor_class *vendor_class,
166 		struct ib_mad_reg_req *mad_reg_req)
167 {
168 	struct ib_mad_mgmt_method_table *method;
169 	int i;
170 
171 	for (i = 0; i < MAX_MGMT_OUI; i++) {
172 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
173 			method = vendor_class->method_table[i];
174 			if (method) {
175 				if (method_in_use(&method, mad_reg_req))
176 					return 1;
177 				else
178 					break;
179 			}
180 		}
181 	}
182 	return 0;
183 }
184 
185 int ib_response_mad(const struct ib_mad_hdr *hdr)
186 {
187 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
188 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
189 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
190 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
191 }
192 EXPORT_SYMBOL(ib_response_mad);
193 
194 /*
195  * ib_register_mad_agent - Register to send/receive MADs
196  */
197 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
198 					   u8 port_num,
199 					   enum ib_qp_type qp_type,
200 					   struct ib_mad_reg_req *mad_reg_req,
201 					   u8 rmpp_version,
202 					   ib_mad_send_handler send_handler,
203 					   ib_mad_recv_handler recv_handler,
204 					   void *context,
205 					   u32 registration_flags)
206 {
207 	struct ib_mad_port_private *port_priv;
208 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
209 	struct ib_mad_agent_private *mad_agent_priv;
210 	struct ib_mad_reg_req *reg_req = NULL;
211 	struct ib_mad_mgmt_class_table *class;
212 	struct ib_mad_mgmt_vendor_class_table *vendor;
213 	struct ib_mad_mgmt_vendor_class *vendor_class;
214 	struct ib_mad_mgmt_method_table *method;
215 	int ret2, qpn;
216 	unsigned long flags;
217 	u8 mgmt_class, vclass;
218 
219 	/* Validate parameters */
220 	qpn = get_spl_qp_index(qp_type);
221 	if (qpn == -1) {
222 		dev_notice(&device->dev,
223 			   "ib_register_mad_agent: invalid QP Type %d\n",
224 			   qp_type);
225 		goto error1;
226 	}
227 
228 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
229 		dev_notice(&device->dev,
230 			   "ib_register_mad_agent: invalid RMPP Version %u\n",
231 			   rmpp_version);
232 		goto error1;
233 	}
234 
235 	/* Validate MAD registration request if supplied */
236 	if (mad_reg_req) {
237 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
238 			dev_notice(&device->dev,
239 				   "ib_register_mad_agent: invalid Class Version %u\n",
240 				   mad_reg_req->mgmt_class_version);
241 			goto error1;
242 		}
243 		if (!recv_handler) {
244 			dev_notice(&device->dev,
245 				   "ib_register_mad_agent: no recv_handler\n");
246 			goto error1;
247 		}
248 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
249 			/*
250 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
251 			 * one in this range currently allowed
252 			 */
253 			if (mad_reg_req->mgmt_class !=
254 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
255 				dev_notice(&device->dev,
256 					   "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
257 					   mad_reg_req->mgmt_class);
258 				goto error1;
259 			}
260 		} else if (mad_reg_req->mgmt_class == 0) {
261 			/*
262 			 * Class 0 is reserved in IBA and is used for
263 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
264 			 */
265 			dev_notice(&device->dev,
266 				   "ib_register_mad_agent: Invalid Mgmt Class 0\n");
267 			goto error1;
268 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
269 			/*
270 			 * If class is in "new" vendor range,
271 			 * ensure supplied OUI is not zero
272 			 */
273 			if (!is_vendor_oui(mad_reg_req->oui)) {
274 				dev_notice(&device->dev,
275 					   "ib_register_mad_agent: No OUI specified for class 0x%x\n",
276 					   mad_reg_req->mgmt_class);
277 				goto error1;
278 			}
279 		}
280 		/* Make sure class supplied is consistent with RMPP */
281 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
282 			if (rmpp_version) {
283 				dev_notice(&device->dev,
284 					   "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
285 					   mad_reg_req->mgmt_class);
286 				goto error1;
287 			}
288 		}
289 
290 		/* Make sure class supplied is consistent with QP type */
291 		if (qp_type == IB_QPT_SMI) {
292 			if ((mad_reg_req->mgmt_class !=
293 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
294 			    (mad_reg_req->mgmt_class !=
295 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
296 				dev_notice(&device->dev,
297 					   "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
298 					   mad_reg_req->mgmt_class);
299 				goto error1;
300 			}
301 		} else {
302 			if ((mad_reg_req->mgmt_class ==
303 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
304 			    (mad_reg_req->mgmt_class ==
305 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
306 				dev_notice(&device->dev,
307 					   "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
308 					   mad_reg_req->mgmt_class);
309 				goto error1;
310 			}
311 		}
312 	} else {
313 		/* No registration request supplied */
314 		if (!send_handler)
315 			goto error1;
316 		if (registration_flags & IB_MAD_USER_RMPP)
317 			goto error1;
318 	}
319 
320 	/* Validate device and port */
321 	port_priv = ib_get_mad_port(device, port_num);
322 	if (!port_priv) {
323 		dev_notice(&device->dev, "ib_register_mad_agent: Invalid port\n");
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 	spin_lock_irqsave(&port_priv->reg_lock, flags);
375 	mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
376 
377 	/*
378 	 * Make sure MAD registration (if supplied)
379 	 * is non overlapping with any existing ones
380 	 */
381 	if (mad_reg_req) {
382 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
383 		if (!is_vendor_class(mgmt_class)) {
384 			class = port_priv->version[mad_reg_req->
385 						   mgmt_class_version].class;
386 			if (class) {
387 				method = class->method_table[mgmt_class];
388 				if (method) {
389 					if (method_in_use(&method,
390 							   mad_reg_req))
391 						goto error4;
392 				}
393 			}
394 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
395 						  mgmt_class);
396 		} else {
397 			/* "New" vendor class range */
398 			vendor = port_priv->version[mad_reg_req->
399 						    mgmt_class_version].vendor;
400 			if (vendor) {
401 				vclass = vendor_class_index(mgmt_class);
402 				vendor_class = vendor->vendor_class[vclass];
403 				if (vendor_class) {
404 					if (is_vendor_method_in_use(
405 							vendor_class,
406 							mad_reg_req))
407 						goto error4;
408 				}
409 			}
410 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
411 		}
412 		if (ret2) {
413 			ret = ERR_PTR(ret2);
414 			goto error4;
415 		}
416 	}
417 
418 	/* Add mad agent into port's agent list */
419 	list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
420 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
421 
422 	return &mad_agent_priv->agent;
423 
424 error4:
425 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
426 	kfree(reg_req);
427 error3:
428 	kfree(mad_agent_priv);
429 error1:
430 	return ret;
431 }
432 EXPORT_SYMBOL(ib_register_mad_agent);
433 
434 static inline int is_snooping_sends(int mad_snoop_flags)
435 {
436 	return (mad_snoop_flags &
437 		(/*IB_MAD_SNOOP_POSTED_SENDS |
438 		 IB_MAD_SNOOP_RMPP_SENDS |*/
439 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
440 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
441 }
442 
443 static inline int is_snooping_recvs(int mad_snoop_flags)
444 {
445 	return (mad_snoop_flags &
446 		(IB_MAD_SNOOP_RECVS /*|
447 		 IB_MAD_SNOOP_RMPP_RECVS*/));
448 }
449 
450 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
451 				struct ib_mad_snoop_private *mad_snoop_priv)
452 {
453 	struct ib_mad_snoop_private **new_snoop_table;
454 	unsigned long flags;
455 	int i;
456 
457 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
458 	/* Check for empty slot in array. */
459 	for (i = 0; i < qp_info->snoop_table_size; i++)
460 		if (!qp_info->snoop_table[i])
461 			break;
462 
463 	if (i == qp_info->snoop_table_size) {
464 		/* Grow table. */
465 		new_snoop_table = krealloc(qp_info->snoop_table,
466 					   sizeof mad_snoop_priv *
467 					   (qp_info->snoop_table_size + 1),
468 					   GFP_ATOMIC);
469 		if (!new_snoop_table) {
470 			i = -ENOMEM;
471 			goto out;
472 		}
473 
474 		qp_info->snoop_table = new_snoop_table;
475 		qp_info->snoop_table_size++;
476 	}
477 	qp_info->snoop_table[i] = mad_snoop_priv;
478 	atomic_inc(&qp_info->snoop_count);
479 out:
480 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
481 	return i;
482 }
483 
484 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
485 					   u8 port_num,
486 					   enum ib_qp_type qp_type,
487 					   int mad_snoop_flags,
488 					   ib_mad_snoop_handler snoop_handler,
489 					   ib_mad_recv_handler recv_handler,
490 					   void *context)
491 {
492 	struct ib_mad_port_private *port_priv;
493 	struct ib_mad_agent *ret;
494 	struct ib_mad_snoop_private *mad_snoop_priv;
495 	int qpn;
496 
497 	/* Validate parameters */
498 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
499 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
500 		ret = ERR_PTR(-EINVAL);
501 		goto error1;
502 	}
503 	qpn = get_spl_qp_index(qp_type);
504 	if (qpn == -1) {
505 		ret = ERR_PTR(-EINVAL);
506 		goto error1;
507 	}
508 	port_priv = ib_get_mad_port(device, port_num);
509 	if (!port_priv) {
510 		ret = ERR_PTR(-ENODEV);
511 		goto error1;
512 	}
513 	/* Allocate structures */
514 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
515 	if (!mad_snoop_priv) {
516 		ret = ERR_PTR(-ENOMEM);
517 		goto error1;
518 	}
519 
520 	/* Now, fill in the various structures */
521 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
522 	mad_snoop_priv->agent.device = device;
523 	mad_snoop_priv->agent.recv_handler = recv_handler;
524 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
525 	mad_snoop_priv->agent.context = context;
526 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
527 	mad_snoop_priv->agent.port_num = port_num;
528 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
529 	init_completion(&mad_snoop_priv->comp);
530 	mad_snoop_priv->snoop_index = register_snoop_agent(
531 						&port_priv->qp_info[qpn],
532 						mad_snoop_priv);
533 	if (mad_snoop_priv->snoop_index < 0) {
534 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
535 		goto error2;
536 	}
537 
538 	atomic_set(&mad_snoop_priv->refcount, 1);
539 	return &mad_snoop_priv->agent;
540 
541 error2:
542 	kfree(mad_snoop_priv);
543 error1:
544 	return ret;
545 }
546 EXPORT_SYMBOL(ib_register_mad_snoop);
547 
548 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
549 {
550 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
551 		complete(&mad_agent_priv->comp);
552 }
553 
554 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
555 {
556 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
557 		complete(&mad_snoop_priv->comp);
558 }
559 
560 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
561 {
562 	struct ib_mad_port_private *port_priv;
563 	unsigned long flags;
564 
565 	/* Note that we could still be handling received MADs */
566 
567 	/*
568 	 * Canceling all sends results in dropping received response
569 	 * MADs, preventing us from queuing additional work
570 	 */
571 	cancel_mads(mad_agent_priv);
572 	port_priv = mad_agent_priv->qp_info->port_priv;
573 	cancel_delayed_work(&mad_agent_priv->timed_work);
574 
575 	spin_lock_irqsave(&port_priv->reg_lock, flags);
576 	remove_mad_reg_req(mad_agent_priv);
577 	list_del(&mad_agent_priv->agent_list);
578 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
579 
580 	flush_workqueue(port_priv->wq);
581 	ib_cancel_rmpp_recvs(mad_agent_priv);
582 
583 	deref_mad_agent(mad_agent_priv);
584 	wait_for_completion(&mad_agent_priv->comp);
585 
586 	kfree(mad_agent_priv->reg_req);
587 	kfree(mad_agent_priv);
588 }
589 
590 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
591 {
592 	struct ib_mad_qp_info *qp_info;
593 	unsigned long flags;
594 
595 	qp_info = mad_snoop_priv->qp_info;
596 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
597 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
598 	atomic_dec(&qp_info->snoop_count);
599 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
600 
601 	deref_snoop_agent(mad_snoop_priv);
602 	wait_for_completion(&mad_snoop_priv->comp);
603 
604 	kfree(mad_snoop_priv);
605 }
606 
607 /*
608  * ib_unregister_mad_agent - Unregisters a client from using MAD services
609  */
610 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
611 {
612 	struct ib_mad_agent_private *mad_agent_priv;
613 	struct ib_mad_snoop_private *mad_snoop_priv;
614 
615 	/* If the TID is zero, the agent can only snoop. */
616 	if (mad_agent->hi_tid) {
617 		mad_agent_priv = container_of(mad_agent,
618 					      struct ib_mad_agent_private,
619 					      agent);
620 		unregister_mad_agent(mad_agent_priv);
621 	} else {
622 		mad_snoop_priv = container_of(mad_agent,
623 					      struct ib_mad_snoop_private,
624 					      agent);
625 		unregister_mad_snoop(mad_snoop_priv);
626 	}
627 	return 0;
628 }
629 EXPORT_SYMBOL(ib_unregister_mad_agent);
630 
631 static void dequeue_mad(struct ib_mad_list_head *mad_list)
632 {
633 	struct ib_mad_queue *mad_queue;
634 	unsigned long flags;
635 
636 	BUG_ON(!mad_list->mad_queue);
637 	mad_queue = mad_list->mad_queue;
638 	spin_lock_irqsave(&mad_queue->lock, flags);
639 	list_del(&mad_list->list);
640 	mad_queue->count--;
641 	spin_unlock_irqrestore(&mad_queue->lock, flags);
642 }
643 
644 static void snoop_send(struct ib_mad_qp_info *qp_info,
645 		       struct ib_mad_send_buf *send_buf,
646 		       struct ib_mad_send_wc *mad_send_wc,
647 		       int mad_snoop_flags)
648 {
649 	struct ib_mad_snoop_private *mad_snoop_priv;
650 	unsigned long flags;
651 	int i;
652 
653 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
654 	for (i = 0; i < qp_info->snoop_table_size; i++) {
655 		mad_snoop_priv = qp_info->snoop_table[i];
656 		if (!mad_snoop_priv ||
657 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
658 			continue;
659 
660 		atomic_inc(&mad_snoop_priv->refcount);
661 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
662 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
663 						    send_buf, mad_send_wc);
664 		deref_snoop_agent(mad_snoop_priv);
665 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
666 	}
667 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
668 }
669 
670 static void snoop_recv(struct ib_mad_qp_info *qp_info,
671 		       struct ib_mad_recv_wc *mad_recv_wc,
672 		       int mad_snoop_flags)
673 {
674 	struct ib_mad_snoop_private *mad_snoop_priv;
675 	unsigned long flags;
676 	int i;
677 
678 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
679 	for (i = 0; i < qp_info->snoop_table_size; i++) {
680 		mad_snoop_priv = qp_info->snoop_table[i];
681 		if (!mad_snoop_priv ||
682 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
683 			continue;
684 
685 		atomic_inc(&mad_snoop_priv->refcount);
686 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
687 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
688 						   mad_recv_wc);
689 		deref_snoop_agent(mad_snoop_priv);
690 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
691 	}
692 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
693 }
694 
695 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
696 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
697 {
698 	memset(wc, 0, sizeof *wc);
699 	wc->wr_cqe = cqe;
700 	wc->status = IB_WC_SUCCESS;
701 	wc->opcode = IB_WC_RECV;
702 	wc->pkey_index = pkey_index;
703 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
704 	wc->src_qp = IB_QP0;
705 	wc->qp = qp;
706 	wc->slid = slid;
707 	wc->sl = 0;
708 	wc->dlid_path_bits = 0;
709 	wc->port_num = port_num;
710 }
711 
712 static size_t mad_priv_size(const struct ib_mad_private *mp)
713 {
714 	return sizeof(struct ib_mad_private) + mp->mad_size;
715 }
716 
717 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
718 {
719 	size_t size = sizeof(struct ib_mad_private) + mad_size;
720 	struct ib_mad_private *ret = kzalloc(size, flags);
721 
722 	if (ret)
723 		ret->mad_size = mad_size;
724 
725 	return ret;
726 }
727 
728 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
729 {
730 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
731 }
732 
733 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
734 {
735 	return sizeof(struct ib_grh) + mp->mad_size;
736 }
737 
738 /*
739  * Return 0 if SMP is to be sent
740  * Return 1 if SMP was consumed locally (whether or not solicited)
741  * Return < 0 if error
742  */
743 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
744 				  struct ib_mad_send_wr_private *mad_send_wr)
745 {
746 	int ret = 0;
747 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
748 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
749 	unsigned long flags;
750 	struct ib_mad_local_private *local;
751 	struct ib_mad_private *mad_priv;
752 	struct ib_mad_port_private *port_priv;
753 	struct ib_mad_agent_private *recv_mad_agent = NULL;
754 	struct ib_device *device = mad_agent_priv->agent.device;
755 	u8 port_num;
756 	struct ib_wc mad_wc;
757 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
758 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
759 	u16 out_mad_pkey_index = 0;
760 	u16 drslid;
761 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
762 				    mad_agent_priv->qp_info->port_priv->port_num);
763 
764 	if (rdma_cap_ib_switch(device) &&
765 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
766 		port_num = send_wr->port_num;
767 	else
768 		port_num = mad_agent_priv->agent.port_num;
769 
770 	/*
771 	 * Directed route handling starts if the initial LID routed part of
772 	 * a request or the ending LID routed part of a response is empty.
773 	 * If we are at the start of the LID routed part, don't update the
774 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
775 	 */
776 	if (opa && smp->class_version == OPA_SMP_CLASS_VERSION) {
777 		u32 opa_drslid;
778 
779 		if ((opa_get_smp_direction(opa_smp)
780 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
781 		     OPA_LID_PERMISSIVE &&
782 		     opa_smi_handle_dr_smp_send(opa_smp,
783 						rdma_cap_ib_switch(device),
784 						port_num) == IB_SMI_DISCARD) {
785 			ret = -EINVAL;
786 			dev_err(&device->dev, "OPA Invalid directed route\n");
787 			goto out;
788 		}
789 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
790 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
791 		    opa_drslid & 0xffff0000) {
792 			ret = -EINVAL;
793 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
794 			       opa_drslid);
795 			goto out;
796 		}
797 		drslid = (u16)(opa_drslid & 0x0000ffff);
798 
799 		/* Check to post send on QP or process locally */
800 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
801 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
802 			goto out;
803 	} else {
804 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
805 		     IB_LID_PERMISSIVE &&
806 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
807 		     IB_SMI_DISCARD) {
808 			ret = -EINVAL;
809 			dev_err(&device->dev, "Invalid directed route\n");
810 			goto out;
811 		}
812 		drslid = be16_to_cpu(smp->dr_slid);
813 
814 		/* Check to post send on QP or process locally */
815 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
816 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
817 			goto out;
818 	}
819 
820 	local = kmalloc(sizeof *local, GFP_ATOMIC);
821 	if (!local) {
822 		ret = -ENOMEM;
823 		dev_err(&device->dev, "No memory for ib_mad_local_private\n");
824 		goto out;
825 	}
826 	local->mad_priv = NULL;
827 	local->recv_mad_agent = NULL;
828 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
829 	if (!mad_priv) {
830 		ret = -ENOMEM;
831 		dev_err(&device->dev, "No memory for local response MAD\n");
832 		kfree(local);
833 		goto out;
834 	}
835 
836 	build_smp_wc(mad_agent_priv->agent.qp,
837 		     send_wr->wr.wr_cqe, drslid,
838 		     send_wr->pkey_index,
839 		     send_wr->port_num, &mad_wc);
840 
841 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
842 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
843 					+ mad_send_wr->send_buf.data_len
844 					+ sizeof(struct ib_grh);
845 	}
846 
847 	/* No GRH for DR SMP */
848 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
849 				  (const struct ib_mad_hdr *)smp, mad_size,
850 				  (struct ib_mad_hdr *)mad_priv->mad,
851 				  &mad_size, &out_mad_pkey_index);
852 	switch (ret)
853 	{
854 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
855 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
856 		    mad_agent_priv->agent.recv_handler) {
857 			local->mad_priv = mad_priv;
858 			local->recv_mad_agent = mad_agent_priv;
859 			/*
860 			 * Reference MAD agent until receive
861 			 * side of local completion handled
862 			 */
863 			atomic_inc(&mad_agent_priv->refcount);
864 		} else
865 			kfree(mad_priv);
866 		break;
867 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
868 		kfree(mad_priv);
869 		break;
870 	case IB_MAD_RESULT_SUCCESS:
871 		/* Treat like an incoming receive MAD */
872 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
873 					    mad_agent_priv->agent.port_num);
874 		if (port_priv) {
875 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
876 			recv_mad_agent = find_mad_agent(port_priv,
877 						        (const struct ib_mad_hdr *)mad_priv->mad);
878 		}
879 		if (!port_priv || !recv_mad_agent) {
880 			/*
881 			 * No receiving agent so drop packet and
882 			 * generate send completion.
883 			 */
884 			kfree(mad_priv);
885 			break;
886 		}
887 		local->mad_priv = mad_priv;
888 		local->recv_mad_agent = recv_mad_agent;
889 		break;
890 	default:
891 		kfree(mad_priv);
892 		kfree(local);
893 		ret = -EINVAL;
894 		goto out;
895 	}
896 
897 	local->mad_send_wr = mad_send_wr;
898 	if (opa) {
899 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
900 		local->return_wc_byte_len = mad_size;
901 	}
902 	/* Reference MAD agent until send side of local completion handled */
903 	atomic_inc(&mad_agent_priv->refcount);
904 	/* Queue local completion to local list */
905 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
906 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
907 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
908 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
909 		   &mad_agent_priv->local_work);
910 	ret = 1;
911 out:
912 	return ret;
913 }
914 
915 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
916 {
917 	int seg_size, pad;
918 
919 	seg_size = mad_size - hdr_len;
920 	if (data_len && seg_size) {
921 		pad = seg_size - data_len % seg_size;
922 		return pad == seg_size ? 0 : pad;
923 	} else
924 		return seg_size;
925 }
926 
927 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
928 {
929 	struct ib_rmpp_segment *s, *t;
930 
931 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
932 		list_del(&s->list);
933 		kfree(s);
934 	}
935 }
936 
937 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
938 				size_t mad_size, gfp_t gfp_mask)
939 {
940 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
941 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
942 	struct ib_rmpp_segment *seg = NULL;
943 	int left, seg_size, pad;
944 
945 	send_buf->seg_size = mad_size - send_buf->hdr_len;
946 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
947 	seg_size = send_buf->seg_size;
948 	pad = send_wr->pad;
949 
950 	/* Allocate data segments. */
951 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
952 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
953 		if (!seg) {
954 			dev_err(&send_buf->mad_agent->device->dev,
955 				"alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
956 				sizeof (*seg) + seg_size, gfp_mask);
957 			free_send_rmpp_list(send_wr);
958 			return -ENOMEM;
959 		}
960 		seg->num = ++send_buf->seg_count;
961 		list_add_tail(&seg->list, &send_wr->rmpp_list);
962 	}
963 
964 	/* Zero any padding */
965 	if (pad)
966 		memset(seg->data + seg_size - pad, 0, pad);
967 
968 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
969 					  agent.rmpp_version;
970 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
971 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
972 
973 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
974 					struct ib_rmpp_segment, list);
975 	send_wr->last_ack_seg = send_wr->cur_seg;
976 	return 0;
977 }
978 
979 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
980 {
981 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
982 }
983 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
984 
985 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
986 					    u32 remote_qpn, u16 pkey_index,
987 					    int rmpp_active,
988 					    int hdr_len, int data_len,
989 					    gfp_t gfp_mask,
990 					    u8 base_version)
991 {
992 	struct ib_mad_agent_private *mad_agent_priv;
993 	struct ib_mad_send_wr_private *mad_send_wr;
994 	int pad, message_size, ret, size;
995 	void *buf;
996 	size_t mad_size;
997 	bool opa;
998 
999 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1000 				      agent);
1001 
1002 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1003 
1004 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1005 		mad_size = sizeof(struct opa_mad);
1006 	else
1007 		mad_size = sizeof(struct ib_mad);
1008 
1009 	pad = get_pad_size(hdr_len, data_len, mad_size);
1010 	message_size = hdr_len + data_len + pad;
1011 
1012 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1013 		if (!rmpp_active && message_size > mad_size)
1014 			return ERR_PTR(-EINVAL);
1015 	} else
1016 		if (rmpp_active || message_size > mad_size)
1017 			return ERR_PTR(-EINVAL);
1018 
1019 	size = rmpp_active ? hdr_len : mad_size;
1020 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1021 	if (!buf)
1022 		return ERR_PTR(-ENOMEM);
1023 
1024 	mad_send_wr = buf + size;
1025 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1026 	mad_send_wr->send_buf.mad = buf;
1027 	mad_send_wr->send_buf.hdr_len = hdr_len;
1028 	mad_send_wr->send_buf.data_len = data_len;
1029 	mad_send_wr->pad = pad;
1030 
1031 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1032 	mad_send_wr->sg_list[0].length = hdr_len;
1033 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1034 
1035 	/* OPA MADs don't have to be the full 2048 bytes */
1036 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1037 	    data_len < mad_size - hdr_len)
1038 		mad_send_wr->sg_list[1].length = data_len;
1039 	else
1040 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1041 
1042 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1043 
1044 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1045 
1046 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1047 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1048 	mad_send_wr->send_wr.wr.num_sge = 2;
1049 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1050 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1051 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1052 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1053 	mad_send_wr->send_wr.pkey_index = pkey_index;
1054 
1055 	if (rmpp_active) {
1056 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1057 		if (ret) {
1058 			kfree(buf);
1059 			return ERR_PTR(ret);
1060 		}
1061 	}
1062 
1063 	mad_send_wr->send_buf.mad_agent = mad_agent;
1064 	atomic_inc(&mad_agent_priv->refcount);
1065 	return &mad_send_wr->send_buf;
1066 }
1067 EXPORT_SYMBOL(ib_create_send_mad);
1068 
1069 int ib_get_mad_data_offset(u8 mgmt_class)
1070 {
1071 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1072 		return IB_MGMT_SA_HDR;
1073 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1074 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1075 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1076 		return IB_MGMT_DEVICE_HDR;
1077 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1078 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1079 		return IB_MGMT_VENDOR_HDR;
1080 	else
1081 		return IB_MGMT_MAD_HDR;
1082 }
1083 EXPORT_SYMBOL(ib_get_mad_data_offset);
1084 
1085 int ib_is_mad_class_rmpp(u8 mgmt_class)
1086 {
1087 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1088 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1089 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1090 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1091 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1092 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1093 		return 1;
1094 	return 0;
1095 }
1096 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1097 
1098 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1099 {
1100 	struct ib_mad_send_wr_private *mad_send_wr;
1101 	struct list_head *list;
1102 
1103 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1104 				   send_buf);
1105 	list = &mad_send_wr->cur_seg->list;
1106 
1107 	if (mad_send_wr->cur_seg->num < seg_num) {
1108 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1109 			if (mad_send_wr->cur_seg->num == seg_num)
1110 				break;
1111 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1112 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1113 			if (mad_send_wr->cur_seg->num == seg_num)
1114 				break;
1115 	}
1116 	return mad_send_wr->cur_seg->data;
1117 }
1118 EXPORT_SYMBOL(ib_get_rmpp_segment);
1119 
1120 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1121 {
1122 	if (mad_send_wr->send_buf.seg_count)
1123 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1124 					   mad_send_wr->seg_num);
1125 	else
1126 		return mad_send_wr->send_buf.mad +
1127 		       mad_send_wr->send_buf.hdr_len;
1128 }
1129 
1130 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1131 {
1132 	struct ib_mad_agent_private *mad_agent_priv;
1133 	struct ib_mad_send_wr_private *mad_send_wr;
1134 
1135 	mad_agent_priv = container_of(send_buf->mad_agent,
1136 				      struct ib_mad_agent_private, agent);
1137 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1138 				   send_buf);
1139 
1140 	free_send_rmpp_list(mad_send_wr);
1141 	kfree(send_buf->mad);
1142 	deref_mad_agent(mad_agent_priv);
1143 }
1144 EXPORT_SYMBOL(ib_free_send_mad);
1145 
1146 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1147 {
1148 	struct ib_mad_qp_info *qp_info;
1149 	struct list_head *list;
1150 	struct ib_send_wr *bad_send_wr;
1151 	struct ib_mad_agent *mad_agent;
1152 	struct ib_sge *sge;
1153 	unsigned long flags;
1154 	int ret;
1155 
1156 	/* Set WR ID to find mad_send_wr upon completion */
1157 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1158 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1159 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1160 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1161 
1162 	mad_agent = mad_send_wr->send_buf.mad_agent;
1163 	sge = mad_send_wr->sg_list;
1164 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1165 					mad_send_wr->send_buf.mad,
1166 					sge[0].length,
1167 					DMA_TO_DEVICE);
1168 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1169 		return -ENOMEM;
1170 
1171 	mad_send_wr->header_mapping = sge[0].addr;
1172 
1173 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1174 					ib_get_payload(mad_send_wr),
1175 					sge[1].length,
1176 					DMA_TO_DEVICE);
1177 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1178 		ib_dma_unmap_single(mad_agent->device,
1179 				    mad_send_wr->header_mapping,
1180 				    sge[0].length, DMA_TO_DEVICE);
1181 		return -ENOMEM;
1182 	}
1183 	mad_send_wr->payload_mapping = sge[1].addr;
1184 
1185 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1186 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1187 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1188 				   &bad_send_wr);
1189 		list = &qp_info->send_queue.list;
1190 	} else {
1191 		ret = 0;
1192 		list = &qp_info->overflow_list;
1193 	}
1194 
1195 	if (!ret) {
1196 		qp_info->send_queue.count++;
1197 		list_add_tail(&mad_send_wr->mad_list.list, list);
1198 	}
1199 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1200 	if (ret) {
1201 		ib_dma_unmap_single(mad_agent->device,
1202 				    mad_send_wr->header_mapping,
1203 				    sge[0].length, DMA_TO_DEVICE);
1204 		ib_dma_unmap_single(mad_agent->device,
1205 				    mad_send_wr->payload_mapping,
1206 				    sge[1].length, DMA_TO_DEVICE);
1207 	}
1208 	return ret;
1209 }
1210 
1211 /*
1212  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1213  *  with the registered client
1214  */
1215 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1216 		     struct ib_mad_send_buf **bad_send_buf)
1217 {
1218 	struct ib_mad_agent_private *mad_agent_priv;
1219 	struct ib_mad_send_buf *next_send_buf;
1220 	struct ib_mad_send_wr_private *mad_send_wr;
1221 	unsigned long flags;
1222 	int ret = -EINVAL;
1223 
1224 	/* Walk list of send WRs and post each on send list */
1225 	for (; send_buf; send_buf = next_send_buf) {
1226 
1227 		mad_send_wr = container_of(send_buf,
1228 					   struct ib_mad_send_wr_private,
1229 					   send_buf);
1230 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1231 
1232 		if (!send_buf->mad_agent->send_handler ||
1233 		    (send_buf->timeout_ms &&
1234 		     !send_buf->mad_agent->recv_handler)) {
1235 			ret = -EINVAL;
1236 			goto error;
1237 		}
1238 
1239 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1240 			if (mad_agent_priv->agent.rmpp_version) {
1241 				ret = -EINVAL;
1242 				goto error;
1243 			}
1244 		}
1245 
1246 		/*
1247 		 * Save pointer to next work request to post in case the
1248 		 * current one completes, and the user modifies the work
1249 		 * request associated with the completion
1250 		 */
1251 		next_send_buf = send_buf->next;
1252 		mad_send_wr->send_wr.ah = send_buf->ah;
1253 
1254 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1255 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1256 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1257 						     mad_send_wr);
1258 			if (ret < 0)		/* error */
1259 				goto error;
1260 			else if (ret == 1)	/* locally consumed */
1261 				continue;
1262 		}
1263 
1264 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1265 		/* Timeout will be updated after send completes */
1266 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1267 		mad_send_wr->max_retries = send_buf->retries;
1268 		mad_send_wr->retries_left = send_buf->retries;
1269 		send_buf->retries = 0;
1270 		/* Reference for work request to QP + response */
1271 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1272 		mad_send_wr->status = IB_WC_SUCCESS;
1273 
1274 		/* Reference MAD agent until send completes */
1275 		atomic_inc(&mad_agent_priv->refcount);
1276 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1277 		list_add_tail(&mad_send_wr->agent_list,
1278 			      &mad_agent_priv->send_list);
1279 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1280 
1281 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1282 			ret = ib_send_rmpp_mad(mad_send_wr);
1283 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1284 				ret = ib_send_mad(mad_send_wr);
1285 		} else
1286 			ret = ib_send_mad(mad_send_wr);
1287 		if (ret < 0) {
1288 			/* Fail send request */
1289 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1290 			list_del(&mad_send_wr->agent_list);
1291 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1292 			atomic_dec(&mad_agent_priv->refcount);
1293 			goto error;
1294 		}
1295 	}
1296 	return 0;
1297 error:
1298 	if (bad_send_buf)
1299 		*bad_send_buf = send_buf;
1300 	return ret;
1301 }
1302 EXPORT_SYMBOL(ib_post_send_mad);
1303 
1304 /*
1305  * ib_free_recv_mad - Returns data buffers used to receive
1306  *  a MAD to the access layer
1307  */
1308 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1309 {
1310 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1311 	struct ib_mad_private_header *mad_priv_hdr;
1312 	struct ib_mad_private *priv;
1313 	struct list_head free_list;
1314 
1315 	INIT_LIST_HEAD(&free_list);
1316 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1317 
1318 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1319 					&free_list, list) {
1320 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1321 					   recv_buf);
1322 		mad_priv_hdr = container_of(mad_recv_wc,
1323 					    struct ib_mad_private_header,
1324 					    recv_wc);
1325 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1326 				    header);
1327 		kfree(priv);
1328 	}
1329 }
1330 EXPORT_SYMBOL(ib_free_recv_mad);
1331 
1332 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1333 					u8 rmpp_version,
1334 					ib_mad_send_handler send_handler,
1335 					ib_mad_recv_handler recv_handler,
1336 					void *context)
1337 {
1338 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1339 }
1340 EXPORT_SYMBOL(ib_redirect_mad_qp);
1341 
1342 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1343 		      struct ib_wc *wc)
1344 {
1345 	dev_err(&mad_agent->device->dev,
1346 		"ib_process_mad_wc() not implemented yet\n");
1347 	return 0;
1348 }
1349 EXPORT_SYMBOL(ib_process_mad_wc);
1350 
1351 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1352 			 struct ib_mad_reg_req *mad_reg_req)
1353 {
1354 	int i;
1355 
1356 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1357 		if ((*method)->agent[i]) {
1358 			pr_err("Method %d already in use\n", i);
1359 			return -EINVAL;
1360 		}
1361 	}
1362 	return 0;
1363 }
1364 
1365 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1366 {
1367 	/* Allocate management method table */
1368 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1369 	if (!*method) {
1370 		pr_err("No memory for ib_mad_mgmt_method_table\n");
1371 		return -ENOMEM;
1372 	}
1373 
1374 	return 0;
1375 }
1376 
1377 /*
1378  * Check to see if there are any methods still in use
1379  */
1380 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1381 {
1382 	int i;
1383 
1384 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1385 		if (method->agent[i])
1386 			return 1;
1387 	return 0;
1388 }
1389 
1390 /*
1391  * Check to see if there are any method tables for this class still in use
1392  */
1393 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1394 {
1395 	int i;
1396 
1397 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1398 		if (class->method_table[i])
1399 			return 1;
1400 	return 0;
1401 }
1402 
1403 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1404 {
1405 	int i;
1406 
1407 	for (i = 0; i < MAX_MGMT_OUI; i++)
1408 		if (vendor_class->method_table[i])
1409 			return 1;
1410 	return 0;
1411 }
1412 
1413 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1414 			   const char *oui)
1415 {
1416 	int i;
1417 
1418 	for (i = 0; i < MAX_MGMT_OUI; i++)
1419 		/* Is there matching OUI for this vendor class ? */
1420 		if (!memcmp(vendor_class->oui[i], oui, 3))
1421 			return i;
1422 
1423 	return -1;
1424 }
1425 
1426 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1427 {
1428 	int i;
1429 
1430 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1431 		if (vendor->vendor_class[i])
1432 			return 1;
1433 
1434 	return 0;
1435 }
1436 
1437 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1438 				     struct ib_mad_agent_private *agent)
1439 {
1440 	int i;
1441 
1442 	/* Remove any methods for this mad agent */
1443 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1444 		if (method->agent[i] == agent) {
1445 			method->agent[i] = NULL;
1446 		}
1447 	}
1448 }
1449 
1450 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1451 			      struct ib_mad_agent_private *agent_priv,
1452 			      u8 mgmt_class)
1453 {
1454 	struct ib_mad_port_private *port_priv;
1455 	struct ib_mad_mgmt_class_table **class;
1456 	struct ib_mad_mgmt_method_table **method;
1457 	int i, ret;
1458 
1459 	port_priv = agent_priv->qp_info->port_priv;
1460 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1461 	if (!*class) {
1462 		/* Allocate management class table for "new" class version */
1463 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1464 		if (!*class) {
1465 			dev_err(&agent_priv->agent.device->dev,
1466 				"No memory for ib_mad_mgmt_class_table\n");
1467 			ret = -ENOMEM;
1468 			goto error1;
1469 		}
1470 
1471 		/* Allocate method table for this management class */
1472 		method = &(*class)->method_table[mgmt_class];
1473 		if ((ret = allocate_method_table(method)))
1474 			goto error2;
1475 	} else {
1476 		method = &(*class)->method_table[mgmt_class];
1477 		if (!*method) {
1478 			/* Allocate method table for this management class */
1479 			if ((ret = allocate_method_table(method)))
1480 				goto error1;
1481 		}
1482 	}
1483 
1484 	/* Now, make sure methods are not already in use */
1485 	if (method_in_use(method, mad_reg_req))
1486 		goto error3;
1487 
1488 	/* Finally, add in methods being registered */
1489 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1490 		(*method)->agent[i] = agent_priv;
1491 
1492 	return 0;
1493 
1494 error3:
1495 	/* Remove any methods for this mad agent */
1496 	remove_methods_mad_agent(*method, agent_priv);
1497 	/* Now, check to see if there are any methods in use */
1498 	if (!check_method_table(*method)) {
1499 		/* If not, release management method table */
1500 		kfree(*method);
1501 		*method = NULL;
1502 	}
1503 	ret = -EINVAL;
1504 	goto error1;
1505 error2:
1506 	kfree(*class);
1507 	*class = NULL;
1508 error1:
1509 	return ret;
1510 }
1511 
1512 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1513 			   struct ib_mad_agent_private *agent_priv)
1514 {
1515 	struct ib_mad_port_private *port_priv;
1516 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1517 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1518 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1519 	struct ib_mad_mgmt_method_table **method;
1520 	int i, ret = -ENOMEM;
1521 	u8 vclass;
1522 
1523 	/* "New" vendor (with OUI) class */
1524 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1525 	port_priv = agent_priv->qp_info->port_priv;
1526 	vendor_table = &port_priv->version[
1527 				mad_reg_req->mgmt_class_version].vendor;
1528 	if (!*vendor_table) {
1529 		/* Allocate mgmt vendor class table for "new" class version */
1530 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1531 		if (!vendor) {
1532 			dev_err(&agent_priv->agent.device->dev,
1533 				"No memory for ib_mad_mgmt_vendor_class_table\n");
1534 			goto error1;
1535 		}
1536 
1537 		*vendor_table = vendor;
1538 	}
1539 	if (!(*vendor_table)->vendor_class[vclass]) {
1540 		/* Allocate table for this management vendor class */
1541 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1542 		if (!vendor_class) {
1543 			dev_err(&agent_priv->agent.device->dev,
1544 				"No memory for ib_mad_mgmt_vendor_class\n");
1545 			goto error2;
1546 		}
1547 
1548 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1549 	}
1550 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1551 		/* Is there matching OUI for this vendor class ? */
1552 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1553 			    mad_reg_req->oui, 3)) {
1554 			method = &(*vendor_table)->vendor_class[
1555 						vclass]->method_table[i];
1556 			BUG_ON(!*method);
1557 			goto check_in_use;
1558 		}
1559 	}
1560 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1561 		/* OUI slot available ? */
1562 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1563 				vclass]->oui[i])) {
1564 			method = &(*vendor_table)->vendor_class[
1565 				vclass]->method_table[i];
1566 			BUG_ON(*method);
1567 			/* Allocate method table for this OUI */
1568 			if ((ret = allocate_method_table(method)))
1569 				goto error3;
1570 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1571 			       mad_reg_req->oui, 3);
1572 			goto check_in_use;
1573 		}
1574 	}
1575 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1576 	goto error3;
1577 
1578 check_in_use:
1579 	/* Now, make sure methods are not already in use */
1580 	if (method_in_use(method, mad_reg_req))
1581 		goto error4;
1582 
1583 	/* Finally, add in methods being registered */
1584 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1585 		(*method)->agent[i] = agent_priv;
1586 
1587 	return 0;
1588 
1589 error4:
1590 	/* Remove any methods for this mad agent */
1591 	remove_methods_mad_agent(*method, agent_priv);
1592 	/* Now, check to see if there are any methods in use */
1593 	if (!check_method_table(*method)) {
1594 		/* If not, release management method table */
1595 		kfree(*method);
1596 		*method = NULL;
1597 	}
1598 	ret = -EINVAL;
1599 error3:
1600 	if (vendor_class) {
1601 		(*vendor_table)->vendor_class[vclass] = NULL;
1602 		kfree(vendor_class);
1603 	}
1604 error2:
1605 	if (vendor) {
1606 		*vendor_table = NULL;
1607 		kfree(vendor);
1608 	}
1609 error1:
1610 	return ret;
1611 }
1612 
1613 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1614 {
1615 	struct ib_mad_port_private *port_priv;
1616 	struct ib_mad_mgmt_class_table *class;
1617 	struct ib_mad_mgmt_method_table *method;
1618 	struct ib_mad_mgmt_vendor_class_table *vendor;
1619 	struct ib_mad_mgmt_vendor_class *vendor_class;
1620 	int index;
1621 	u8 mgmt_class;
1622 
1623 	/*
1624 	 * Was MAD registration request supplied
1625 	 * with original registration ?
1626 	 */
1627 	if (!agent_priv->reg_req) {
1628 		goto out;
1629 	}
1630 
1631 	port_priv = agent_priv->qp_info->port_priv;
1632 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1633 	class = port_priv->version[
1634 			agent_priv->reg_req->mgmt_class_version].class;
1635 	if (!class)
1636 		goto vendor_check;
1637 
1638 	method = class->method_table[mgmt_class];
1639 	if (method) {
1640 		/* Remove any methods for this mad agent */
1641 		remove_methods_mad_agent(method, agent_priv);
1642 		/* Now, check to see if there are any methods still in use */
1643 		if (!check_method_table(method)) {
1644 			/* If not, release management method table */
1645 			 kfree(method);
1646 			 class->method_table[mgmt_class] = NULL;
1647 			 /* Any management classes left ? */
1648 			if (!check_class_table(class)) {
1649 				/* If not, release management class table */
1650 				kfree(class);
1651 				port_priv->version[
1652 					agent_priv->reg_req->
1653 					mgmt_class_version].class = NULL;
1654 			}
1655 		}
1656 	}
1657 
1658 vendor_check:
1659 	if (!is_vendor_class(mgmt_class))
1660 		goto out;
1661 
1662 	/* normalize mgmt_class to vendor range 2 */
1663 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1664 	vendor = port_priv->version[
1665 			agent_priv->reg_req->mgmt_class_version].vendor;
1666 
1667 	if (!vendor)
1668 		goto out;
1669 
1670 	vendor_class = vendor->vendor_class[mgmt_class];
1671 	if (vendor_class) {
1672 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1673 		if (index < 0)
1674 			goto out;
1675 		method = vendor_class->method_table[index];
1676 		if (method) {
1677 			/* Remove any methods for this mad agent */
1678 			remove_methods_mad_agent(method, agent_priv);
1679 			/*
1680 			 * Now, check to see if there are
1681 			 * any methods still in use
1682 			 */
1683 			if (!check_method_table(method)) {
1684 				/* If not, release management method table */
1685 				kfree(method);
1686 				vendor_class->method_table[index] = NULL;
1687 				memset(vendor_class->oui[index], 0, 3);
1688 				/* Any OUIs left ? */
1689 				if (!check_vendor_class(vendor_class)) {
1690 					/* If not, release vendor class table */
1691 					kfree(vendor_class);
1692 					vendor->vendor_class[mgmt_class] = NULL;
1693 					/* Any other vendor classes left ? */
1694 					if (!check_vendor_table(vendor)) {
1695 						kfree(vendor);
1696 						port_priv->version[
1697 							agent_priv->reg_req->
1698 							mgmt_class_version].
1699 							vendor = NULL;
1700 					}
1701 				}
1702 			}
1703 		}
1704 	}
1705 
1706 out:
1707 	return;
1708 }
1709 
1710 static struct ib_mad_agent_private *
1711 find_mad_agent(struct ib_mad_port_private *port_priv,
1712 	       const struct ib_mad_hdr *mad_hdr)
1713 {
1714 	struct ib_mad_agent_private *mad_agent = NULL;
1715 	unsigned long flags;
1716 
1717 	spin_lock_irqsave(&port_priv->reg_lock, flags);
1718 	if (ib_response_mad(mad_hdr)) {
1719 		u32 hi_tid;
1720 		struct ib_mad_agent_private *entry;
1721 
1722 		/*
1723 		 * Routing is based on high 32 bits of transaction ID
1724 		 * of MAD.
1725 		 */
1726 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1727 		list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1728 			if (entry->agent.hi_tid == hi_tid) {
1729 				mad_agent = entry;
1730 				break;
1731 			}
1732 		}
1733 	} else {
1734 		struct ib_mad_mgmt_class_table *class;
1735 		struct ib_mad_mgmt_method_table *method;
1736 		struct ib_mad_mgmt_vendor_class_table *vendor;
1737 		struct ib_mad_mgmt_vendor_class *vendor_class;
1738 		const struct ib_vendor_mad *vendor_mad;
1739 		int index;
1740 
1741 		/*
1742 		 * Routing is based on version, class, and method
1743 		 * For "newer" vendor MADs, also based on OUI
1744 		 */
1745 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1746 			goto out;
1747 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1748 			class = port_priv->version[
1749 					mad_hdr->class_version].class;
1750 			if (!class)
1751 				goto out;
1752 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1753 			    IB_MGMT_MAX_METHODS)
1754 				goto out;
1755 			method = class->method_table[convert_mgmt_class(
1756 							mad_hdr->mgmt_class)];
1757 			if (method)
1758 				mad_agent = method->agent[mad_hdr->method &
1759 							  ~IB_MGMT_METHOD_RESP];
1760 		} else {
1761 			vendor = port_priv->version[
1762 					mad_hdr->class_version].vendor;
1763 			if (!vendor)
1764 				goto out;
1765 			vendor_class = vendor->vendor_class[vendor_class_index(
1766 						mad_hdr->mgmt_class)];
1767 			if (!vendor_class)
1768 				goto out;
1769 			/* Find matching OUI */
1770 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1771 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1772 			if (index == -1)
1773 				goto out;
1774 			method = vendor_class->method_table[index];
1775 			if (method) {
1776 				mad_agent = method->agent[mad_hdr->method &
1777 							  ~IB_MGMT_METHOD_RESP];
1778 			}
1779 		}
1780 	}
1781 
1782 	if (mad_agent) {
1783 		if (mad_agent->agent.recv_handler)
1784 			atomic_inc(&mad_agent->refcount);
1785 		else {
1786 			dev_notice(&port_priv->device->dev,
1787 				   "No receive handler for client %p on port %d\n",
1788 				   &mad_agent->agent, port_priv->port_num);
1789 			mad_agent = NULL;
1790 		}
1791 	}
1792 out:
1793 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1794 
1795 	return mad_agent;
1796 }
1797 
1798 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1799 			const struct ib_mad_qp_info *qp_info,
1800 			bool opa)
1801 {
1802 	int valid = 0;
1803 	u32 qp_num = qp_info->qp->qp_num;
1804 
1805 	/* Make sure MAD base version is understood */
1806 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1807 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1808 		pr_err("MAD received with unsupported base version %d %s\n",
1809 		       mad_hdr->base_version, opa ? "(opa)" : "");
1810 		goto out;
1811 	}
1812 
1813 	/* Filter SMI packets sent to other than QP0 */
1814 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1815 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1816 		if (qp_num == 0)
1817 			valid = 1;
1818 	} else {
1819 		/* CM attributes other than ClassPortInfo only use Send method */
1820 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1821 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1822 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1823 			goto out;
1824 		/* Filter GSI packets sent to QP0 */
1825 		if (qp_num != 0)
1826 			valid = 1;
1827 	}
1828 
1829 out:
1830 	return valid;
1831 }
1832 
1833 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1834 			    const struct ib_mad_hdr *mad_hdr)
1835 {
1836 	struct ib_rmpp_mad *rmpp_mad;
1837 
1838 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1839 	return !mad_agent_priv->agent.rmpp_version ||
1840 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1841 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1842 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1843 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1844 }
1845 
1846 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1847 				     const struct ib_mad_recv_wc *rwc)
1848 {
1849 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1850 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1851 }
1852 
1853 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1854 				   const struct ib_mad_send_wr_private *wr,
1855 				   const struct ib_mad_recv_wc *rwc )
1856 {
1857 	struct ib_ah_attr attr;
1858 	u8 send_resp, rcv_resp;
1859 	union ib_gid sgid;
1860 	struct ib_device *device = mad_agent_priv->agent.device;
1861 	u8 port_num = mad_agent_priv->agent.port_num;
1862 	u8 lmc;
1863 
1864 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1865 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1866 
1867 	if (send_resp == rcv_resp)
1868 		/* both requests, or both responses. GIDs different */
1869 		return 0;
1870 
1871 	if (ib_query_ah(wr->send_buf.ah, &attr))
1872 		/* Assume not equal, to avoid false positives. */
1873 		return 0;
1874 
1875 	if (!!(attr.ah_flags & IB_AH_GRH) !=
1876 	    !!(rwc->wc->wc_flags & IB_WC_GRH))
1877 		/* one has GID, other does not.  Assume different */
1878 		return 0;
1879 
1880 	if (!send_resp && rcv_resp) {
1881 		/* is request/response. */
1882 		if (!(attr.ah_flags & IB_AH_GRH)) {
1883 			if (ib_get_cached_lmc(device, port_num, &lmc))
1884 				return 0;
1885 			return (!lmc || !((attr.src_path_bits ^
1886 					   rwc->wc->dlid_path_bits) &
1887 					  ((1 << lmc) - 1)));
1888 		} else {
1889 			if (ib_get_cached_gid(device, port_num,
1890 					      attr.grh.sgid_index, &sgid, NULL))
1891 				return 0;
1892 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1893 				       16);
1894 		}
1895 	}
1896 
1897 	if (!(attr.ah_flags & IB_AH_GRH))
1898 		return attr.dlid == rwc->wc->slid;
1899 	else
1900 		return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1901 			       16);
1902 }
1903 
1904 static inline int is_direct(u8 class)
1905 {
1906 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1907 }
1908 
1909 struct ib_mad_send_wr_private*
1910 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1911 		 const struct ib_mad_recv_wc *wc)
1912 {
1913 	struct ib_mad_send_wr_private *wr;
1914 	const struct ib_mad_hdr *mad_hdr;
1915 
1916 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1917 
1918 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1919 		if ((wr->tid == mad_hdr->tid) &&
1920 		    rcv_has_same_class(wr, wc) &&
1921 		    /*
1922 		     * Don't check GID for direct routed MADs.
1923 		     * These might have permissive LIDs.
1924 		     */
1925 		    (is_direct(mad_hdr->mgmt_class) ||
1926 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1927 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1928 	}
1929 
1930 	/*
1931 	 * It's possible to receive the response before we've
1932 	 * been notified that the send has completed
1933 	 */
1934 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1935 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1936 		    wr->tid == mad_hdr->tid &&
1937 		    wr->timeout &&
1938 		    rcv_has_same_class(wr, wc) &&
1939 		    /*
1940 		     * Don't check GID for direct routed MADs.
1941 		     * These might have permissive LIDs.
1942 		     */
1943 		    (is_direct(mad_hdr->mgmt_class) ||
1944 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1945 			/* Verify request has not been canceled */
1946 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1947 	}
1948 	return NULL;
1949 }
1950 
1951 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1952 {
1953 	mad_send_wr->timeout = 0;
1954 	if (mad_send_wr->refcount == 1)
1955 		list_move_tail(&mad_send_wr->agent_list,
1956 			      &mad_send_wr->mad_agent_priv->done_list);
1957 }
1958 
1959 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1960 				 struct ib_mad_recv_wc *mad_recv_wc)
1961 {
1962 	struct ib_mad_send_wr_private *mad_send_wr;
1963 	struct ib_mad_send_wc mad_send_wc;
1964 	unsigned long flags;
1965 
1966 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1967 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1968 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1969 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1970 						      mad_recv_wc);
1971 		if (!mad_recv_wc) {
1972 			deref_mad_agent(mad_agent_priv);
1973 			return;
1974 		}
1975 	}
1976 
1977 	/* Complete corresponding request */
1978 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1979 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1980 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1981 		if (!mad_send_wr) {
1982 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1983 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1984 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1985 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1986 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1987 				/* user rmpp is in effect
1988 				 * and this is an active RMPP MAD
1989 				 */
1990 				mad_agent_priv->agent.recv_handler(
1991 						&mad_agent_priv->agent, NULL,
1992 						mad_recv_wc);
1993 				atomic_dec(&mad_agent_priv->refcount);
1994 			} else {
1995 				/* not user rmpp, revert to normal behavior and
1996 				 * drop the mad */
1997 				ib_free_recv_mad(mad_recv_wc);
1998 				deref_mad_agent(mad_agent_priv);
1999 				return;
2000 			}
2001 		} else {
2002 			ib_mark_mad_done(mad_send_wr);
2003 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2004 
2005 			/* Defined behavior is to complete response before request */
2006 			mad_agent_priv->agent.recv_handler(
2007 					&mad_agent_priv->agent,
2008 					&mad_send_wr->send_buf,
2009 					mad_recv_wc);
2010 			atomic_dec(&mad_agent_priv->refcount);
2011 
2012 			mad_send_wc.status = IB_WC_SUCCESS;
2013 			mad_send_wc.vendor_err = 0;
2014 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2015 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2016 		}
2017 	} else {
2018 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2019 						   mad_recv_wc);
2020 		deref_mad_agent(mad_agent_priv);
2021 	}
2022 }
2023 
2024 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2025 				     const struct ib_mad_qp_info *qp_info,
2026 				     const struct ib_wc *wc,
2027 				     int port_num,
2028 				     struct ib_mad_private *recv,
2029 				     struct ib_mad_private *response)
2030 {
2031 	enum smi_forward_action retsmi;
2032 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2033 
2034 	if (smi_handle_dr_smp_recv(smp,
2035 				   rdma_cap_ib_switch(port_priv->device),
2036 				   port_num,
2037 				   port_priv->device->phys_port_cnt) ==
2038 				   IB_SMI_DISCARD)
2039 		return IB_SMI_DISCARD;
2040 
2041 	retsmi = smi_check_forward_dr_smp(smp);
2042 	if (retsmi == IB_SMI_LOCAL)
2043 		return IB_SMI_HANDLE;
2044 
2045 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2046 		if (smi_handle_dr_smp_send(smp,
2047 					   rdma_cap_ib_switch(port_priv->device),
2048 					   port_num) == IB_SMI_DISCARD)
2049 			return IB_SMI_DISCARD;
2050 
2051 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2052 			return IB_SMI_DISCARD;
2053 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2054 		/* forward case for switches */
2055 		memcpy(response, recv, mad_priv_size(response));
2056 		response->header.recv_wc.wc = &response->header.wc;
2057 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2058 		response->header.recv_wc.recv_buf.grh = &response->grh;
2059 
2060 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2061 				    &response->grh, wc,
2062 				    port_priv->device,
2063 				    smi_get_fwd_port(smp),
2064 				    qp_info->qp->qp_num,
2065 				    response->mad_size,
2066 				    false);
2067 
2068 		return IB_SMI_DISCARD;
2069 	}
2070 	return IB_SMI_HANDLE;
2071 }
2072 
2073 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2074 				    struct ib_mad_private *response,
2075 				    size_t *resp_len, bool opa)
2076 {
2077 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2078 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2079 
2080 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2081 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2082 		memcpy(response, recv, mad_priv_size(response));
2083 		response->header.recv_wc.wc = &response->header.wc;
2084 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2085 		response->header.recv_wc.recv_buf.grh = &response->grh;
2086 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2087 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2088 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2089 			resp_hdr->status |= IB_SMP_DIRECTION;
2090 
2091 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2092 			if (recv_hdr->mgmt_class ==
2093 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2094 			    recv_hdr->mgmt_class ==
2095 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2096 				*resp_len = opa_get_smp_header_size(
2097 							(struct opa_smp *)recv->mad);
2098 			else
2099 				*resp_len = sizeof(struct ib_mad_hdr);
2100 		}
2101 
2102 		return true;
2103 	} else {
2104 		return false;
2105 	}
2106 }
2107 
2108 static enum smi_action
2109 handle_opa_smi(struct ib_mad_port_private *port_priv,
2110 	       struct ib_mad_qp_info *qp_info,
2111 	       struct ib_wc *wc,
2112 	       int port_num,
2113 	       struct ib_mad_private *recv,
2114 	       struct ib_mad_private *response)
2115 {
2116 	enum smi_forward_action retsmi;
2117 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2118 
2119 	if (opa_smi_handle_dr_smp_recv(smp,
2120 				   rdma_cap_ib_switch(port_priv->device),
2121 				   port_num,
2122 				   port_priv->device->phys_port_cnt) ==
2123 				   IB_SMI_DISCARD)
2124 		return IB_SMI_DISCARD;
2125 
2126 	retsmi = opa_smi_check_forward_dr_smp(smp);
2127 	if (retsmi == IB_SMI_LOCAL)
2128 		return IB_SMI_HANDLE;
2129 
2130 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2131 		if (opa_smi_handle_dr_smp_send(smp,
2132 					   rdma_cap_ib_switch(port_priv->device),
2133 					   port_num) == IB_SMI_DISCARD)
2134 			return IB_SMI_DISCARD;
2135 
2136 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2137 		    IB_SMI_DISCARD)
2138 			return IB_SMI_DISCARD;
2139 
2140 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2141 		/* forward case for switches */
2142 		memcpy(response, recv, mad_priv_size(response));
2143 		response->header.recv_wc.wc = &response->header.wc;
2144 		response->header.recv_wc.recv_buf.opa_mad =
2145 				(struct opa_mad *)response->mad;
2146 		response->header.recv_wc.recv_buf.grh = &response->grh;
2147 
2148 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2149 				    &response->grh, wc,
2150 				    port_priv->device,
2151 				    opa_smi_get_fwd_port(smp),
2152 				    qp_info->qp->qp_num,
2153 				    recv->header.wc.byte_len,
2154 				    true);
2155 
2156 		return IB_SMI_DISCARD;
2157 	}
2158 
2159 	return IB_SMI_HANDLE;
2160 }
2161 
2162 static enum smi_action
2163 handle_smi(struct ib_mad_port_private *port_priv,
2164 	   struct ib_mad_qp_info *qp_info,
2165 	   struct ib_wc *wc,
2166 	   int port_num,
2167 	   struct ib_mad_private *recv,
2168 	   struct ib_mad_private *response,
2169 	   bool opa)
2170 {
2171 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2172 
2173 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2174 	    mad_hdr->class_version == OPA_SMI_CLASS_VERSION)
2175 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2176 				      response);
2177 
2178 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2179 }
2180 
2181 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2182 {
2183 	struct ib_mad_port_private *port_priv = cq->cq_context;
2184 	struct ib_mad_list_head *mad_list =
2185 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2186 	struct ib_mad_qp_info *qp_info;
2187 	struct ib_mad_private_header *mad_priv_hdr;
2188 	struct ib_mad_private *recv, *response = NULL;
2189 	struct ib_mad_agent_private *mad_agent;
2190 	int port_num;
2191 	int ret = IB_MAD_RESULT_SUCCESS;
2192 	size_t mad_size;
2193 	u16 resp_mad_pkey_index = 0;
2194 	bool opa;
2195 
2196 	if (list_empty_careful(&port_priv->port_list))
2197 		return;
2198 
2199 	if (wc->status != IB_WC_SUCCESS) {
2200 		/*
2201 		 * Receive errors indicate that the QP has entered the error
2202 		 * state - error handling/shutdown code will cleanup
2203 		 */
2204 		return;
2205 	}
2206 
2207 	qp_info = mad_list->mad_queue->qp_info;
2208 	dequeue_mad(mad_list);
2209 
2210 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2211 			       qp_info->port_priv->port_num);
2212 
2213 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2214 				    mad_list);
2215 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2216 	ib_dma_unmap_single(port_priv->device,
2217 			    recv->header.mapping,
2218 			    mad_priv_dma_size(recv),
2219 			    DMA_FROM_DEVICE);
2220 
2221 	/* Setup MAD receive work completion from "normal" work completion */
2222 	recv->header.wc = *wc;
2223 	recv->header.recv_wc.wc = &recv->header.wc;
2224 
2225 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2226 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2227 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2228 	} else {
2229 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2230 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2231 	}
2232 
2233 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2234 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2235 
2236 	if (atomic_read(&qp_info->snoop_count))
2237 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2238 
2239 	/* Validate MAD */
2240 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2241 		goto out;
2242 
2243 	mad_size = recv->mad_size;
2244 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2245 	if (!response) {
2246 		dev_err(&port_priv->device->dev,
2247 			"%s: no memory for response buffer\n", __func__);
2248 		goto out;
2249 	}
2250 
2251 	if (rdma_cap_ib_switch(port_priv->device))
2252 		port_num = wc->port_num;
2253 	else
2254 		port_num = port_priv->port_num;
2255 
2256 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2257 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2258 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2259 			       response, opa)
2260 		    == IB_SMI_DISCARD)
2261 			goto out;
2262 	}
2263 
2264 	/* Give driver "right of first refusal" on incoming MAD */
2265 	if (port_priv->device->process_mad) {
2266 		ret = port_priv->device->process_mad(port_priv->device, 0,
2267 						     port_priv->port_num,
2268 						     wc, &recv->grh,
2269 						     (const struct ib_mad_hdr *)recv->mad,
2270 						     recv->mad_size,
2271 						     (struct ib_mad_hdr *)response->mad,
2272 						     &mad_size, &resp_mad_pkey_index);
2273 
2274 		if (opa)
2275 			wc->pkey_index = resp_mad_pkey_index;
2276 
2277 		if (ret & IB_MAD_RESULT_SUCCESS) {
2278 			if (ret & IB_MAD_RESULT_CONSUMED)
2279 				goto out;
2280 			if (ret & IB_MAD_RESULT_REPLY) {
2281 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2282 						    &recv->grh, wc,
2283 						    port_priv->device,
2284 						    port_num,
2285 						    qp_info->qp->qp_num,
2286 						    mad_size, opa);
2287 				goto out;
2288 			}
2289 		}
2290 	}
2291 
2292 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2293 	if (mad_agent) {
2294 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2295 		/*
2296 		 * recv is freed up in error cases in ib_mad_complete_recv
2297 		 * or via recv_handler in ib_mad_complete_recv()
2298 		 */
2299 		recv = NULL;
2300 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2301 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2302 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2303 				    port_priv->device, port_num,
2304 				    qp_info->qp->qp_num, mad_size, opa);
2305 	}
2306 
2307 out:
2308 	/* Post another receive request for this QP */
2309 	if (response) {
2310 		ib_mad_post_receive_mads(qp_info, response);
2311 		kfree(recv);
2312 	} else
2313 		ib_mad_post_receive_mads(qp_info, recv);
2314 }
2315 
2316 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2317 {
2318 	struct ib_mad_send_wr_private *mad_send_wr;
2319 	unsigned long delay;
2320 
2321 	if (list_empty(&mad_agent_priv->wait_list)) {
2322 		cancel_delayed_work(&mad_agent_priv->timed_work);
2323 	} else {
2324 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2325 					 struct ib_mad_send_wr_private,
2326 					 agent_list);
2327 
2328 		if (time_after(mad_agent_priv->timeout,
2329 			       mad_send_wr->timeout)) {
2330 			mad_agent_priv->timeout = mad_send_wr->timeout;
2331 			delay = mad_send_wr->timeout - jiffies;
2332 			if ((long)delay <= 0)
2333 				delay = 1;
2334 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2335 					 &mad_agent_priv->timed_work, delay);
2336 		}
2337 	}
2338 }
2339 
2340 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2341 {
2342 	struct ib_mad_agent_private *mad_agent_priv;
2343 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2344 	struct list_head *list_item;
2345 	unsigned long delay;
2346 
2347 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2348 	list_del(&mad_send_wr->agent_list);
2349 
2350 	delay = mad_send_wr->timeout;
2351 	mad_send_wr->timeout += jiffies;
2352 
2353 	if (delay) {
2354 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2355 			temp_mad_send_wr = list_entry(list_item,
2356 						struct ib_mad_send_wr_private,
2357 						agent_list);
2358 			if (time_after(mad_send_wr->timeout,
2359 				       temp_mad_send_wr->timeout))
2360 				break;
2361 		}
2362 	}
2363 	else
2364 		list_item = &mad_agent_priv->wait_list;
2365 	list_add(&mad_send_wr->agent_list, list_item);
2366 
2367 	/* Reschedule a work item if we have a shorter timeout */
2368 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2369 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2370 				 &mad_agent_priv->timed_work, delay);
2371 }
2372 
2373 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2374 			  int timeout_ms)
2375 {
2376 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2377 	wait_for_response(mad_send_wr);
2378 }
2379 
2380 /*
2381  * Process a send work completion
2382  */
2383 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2384 			     struct ib_mad_send_wc *mad_send_wc)
2385 {
2386 	struct ib_mad_agent_private	*mad_agent_priv;
2387 	unsigned long			flags;
2388 	int				ret;
2389 
2390 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2391 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2392 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2393 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2394 		if (ret == IB_RMPP_RESULT_CONSUMED)
2395 			goto done;
2396 	} else
2397 		ret = IB_RMPP_RESULT_UNHANDLED;
2398 
2399 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2400 	    mad_send_wr->status == IB_WC_SUCCESS) {
2401 		mad_send_wr->status = mad_send_wc->status;
2402 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2403 	}
2404 
2405 	if (--mad_send_wr->refcount > 0) {
2406 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2407 		    mad_send_wr->status == IB_WC_SUCCESS) {
2408 			wait_for_response(mad_send_wr);
2409 		}
2410 		goto done;
2411 	}
2412 
2413 	/* Remove send from MAD agent and notify client of completion */
2414 	list_del(&mad_send_wr->agent_list);
2415 	adjust_timeout(mad_agent_priv);
2416 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2417 
2418 	if (mad_send_wr->status != IB_WC_SUCCESS )
2419 		mad_send_wc->status = mad_send_wr->status;
2420 	if (ret == IB_RMPP_RESULT_INTERNAL)
2421 		ib_rmpp_send_handler(mad_send_wc);
2422 	else
2423 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2424 						   mad_send_wc);
2425 
2426 	/* Release reference on agent taken when sending */
2427 	deref_mad_agent(mad_agent_priv);
2428 	return;
2429 done:
2430 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2431 }
2432 
2433 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2434 {
2435 	struct ib_mad_port_private *port_priv = cq->cq_context;
2436 	struct ib_mad_list_head *mad_list =
2437 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2438 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2439 	struct ib_mad_qp_info		*qp_info;
2440 	struct ib_mad_queue		*send_queue;
2441 	struct ib_send_wr		*bad_send_wr;
2442 	struct ib_mad_send_wc		mad_send_wc;
2443 	unsigned long flags;
2444 	int ret;
2445 
2446 	if (list_empty_careful(&port_priv->port_list))
2447 		return;
2448 
2449 	if (wc->status != IB_WC_SUCCESS) {
2450 		if (!ib_mad_send_error(port_priv, wc))
2451 			return;
2452 	}
2453 
2454 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2455 				   mad_list);
2456 	send_queue = mad_list->mad_queue;
2457 	qp_info = send_queue->qp_info;
2458 
2459 retry:
2460 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2461 			    mad_send_wr->header_mapping,
2462 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2463 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2464 			    mad_send_wr->payload_mapping,
2465 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2466 	queued_send_wr = NULL;
2467 	spin_lock_irqsave(&send_queue->lock, flags);
2468 	list_del(&mad_list->list);
2469 
2470 	/* Move queued send to the send queue */
2471 	if (send_queue->count-- > send_queue->max_active) {
2472 		mad_list = container_of(qp_info->overflow_list.next,
2473 					struct ib_mad_list_head, list);
2474 		queued_send_wr = container_of(mad_list,
2475 					struct ib_mad_send_wr_private,
2476 					mad_list);
2477 		list_move_tail(&mad_list->list, &send_queue->list);
2478 	}
2479 	spin_unlock_irqrestore(&send_queue->lock, flags);
2480 
2481 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2482 	mad_send_wc.status = wc->status;
2483 	mad_send_wc.vendor_err = wc->vendor_err;
2484 	if (atomic_read(&qp_info->snoop_count))
2485 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2486 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2487 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2488 
2489 	if (queued_send_wr) {
2490 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2491 				   &bad_send_wr);
2492 		if (ret) {
2493 			dev_err(&port_priv->device->dev,
2494 				"ib_post_send failed: %d\n", ret);
2495 			mad_send_wr = queued_send_wr;
2496 			wc->status = IB_WC_LOC_QP_OP_ERR;
2497 			goto retry;
2498 		}
2499 	}
2500 }
2501 
2502 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2503 {
2504 	struct ib_mad_send_wr_private *mad_send_wr;
2505 	struct ib_mad_list_head *mad_list;
2506 	unsigned long flags;
2507 
2508 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2509 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2510 		mad_send_wr = container_of(mad_list,
2511 					   struct ib_mad_send_wr_private,
2512 					   mad_list);
2513 		mad_send_wr->retry = 1;
2514 	}
2515 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2516 }
2517 
2518 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2519 		struct ib_wc *wc)
2520 {
2521 	struct ib_mad_list_head *mad_list =
2522 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2523 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2524 	struct ib_mad_send_wr_private *mad_send_wr;
2525 	int ret;
2526 
2527 	/*
2528 	 * Send errors will transition the QP to SQE - move
2529 	 * QP to RTS and repost flushed work requests
2530 	 */
2531 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2532 				   mad_list);
2533 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2534 		if (mad_send_wr->retry) {
2535 			/* Repost send */
2536 			struct ib_send_wr *bad_send_wr;
2537 
2538 			mad_send_wr->retry = 0;
2539 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2540 					&bad_send_wr);
2541 			if (!ret)
2542 				return false;
2543 		}
2544 	} else {
2545 		struct ib_qp_attr *attr;
2546 
2547 		/* Transition QP to RTS and fail offending send */
2548 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2549 		if (attr) {
2550 			attr->qp_state = IB_QPS_RTS;
2551 			attr->cur_qp_state = IB_QPS_SQE;
2552 			ret = ib_modify_qp(qp_info->qp, attr,
2553 					   IB_QP_STATE | IB_QP_CUR_STATE);
2554 			kfree(attr);
2555 			if (ret)
2556 				dev_err(&port_priv->device->dev,
2557 					"%s - ib_modify_qp to RTS: %d\n",
2558 					__func__, ret);
2559 			else
2560 				mark_sends_for_retry(qp_info);
2561 		}
2562 	}
2563 
2564 	return true;
2565 }
2566 
2567 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2568 {
2569 	unsigned long flags;
2570 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2571 	struct ib_mad_send_wc mad_send_wc;
2572 	struct list_head cancel_list;
2573 
2574 	INIT_LIST_HEAD(&cancel_list);
2575 
2576 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2577 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2578 				 &mad_agent_priv->send_list, agent_list) {
2579 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2580 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2581 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2582 		}
2583 	}
2584 
2585 	/* Empty wait list to prevent receives from finding a request */
2586 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2587 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2588 
2589 	/* Report all cancelled requests */
2590 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2591 	mad_send_wc.vendor_err = 0;
2592 
2593 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2594 				 &cancel_list, agent_list) {
2595 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2596 		list_del(&mad_send_wr->agent_list);
2597 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2598 						   &mad_send_wc);
2599 		atomic_dec(&mad_agent_priv->refcount);
2600 	}
2601 }
2602 
2603 static struct ib_mad_send_wr_private*
2604 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2605 	     struct ib_mad_send_buf *send_buf)
2606 {
2607 	struct ib_mad_send_wr_private *mad_send_wr;
2608 
2609 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2610 			    agent_list) {
2611 		if (&mad_send_wr->send_buf == send_buf)
2612 			return mad_send_wr;
2613 	}
2614 
2615 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2616 			    agent_list) {
2617 		if (is_rmpp_data_mad(mad_agent_priv,
2618 				     mad_send_wr->send_buf.mad) &&
2619 		    &mad_send_wr->send_buf == send_buf)
2620 			return mad_send_wr;
2621 	}
2622 	return NULL;
2623 }
2624 
2625 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2626 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2627 {
2628 	struct ib_mad_agent_private *mad_agent_priv;
2629 	struct ib_mad_send_wr_private *mad_send_wr;
2630 	unsigned long flags;
2631 	int active;
2632 
2633 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2634 				      agent);
2635 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2636 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2637 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2638 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2639 		return -EINVAL;
2640 	}
2641 
2642 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2643 	if (!timeout_ms) {
2644 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2645 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2646 	}
2647 
2648 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2649 	if (active)
2650 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2651 	else
2652 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2653 
2654 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2655 	return 0;
2656 }
2657 EXPORT_SYMBOL(ib_modify_mad);
2658 
2659 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2660 		   struct ib_mad_send_buf *send_buf)
2661 {
2662 	ib_modify_mad(mad_agent, send_buf, 0);
2663 }
2664 EXPORT_SYMBOL(ib_cancel_mad);
2665 
2666 static void local_completions(struct work_struct *work)
2667 {
2668 	struct ib_mad_agent_private *mad_agent_priv;
2669 	struct ib_mad_local_private *local;
2670 	struct ib_mad_agent_private *recv_mad_agent;
2671 	unsigned long flags;
2672 	int free_mad;
2673 	struct ib_wc wc;
2674 	struct ib_mad_send_wc mad_send_wc;
2675 	bool opa;
2676 
2677 	mad_agent_priv =
2678 		container_of(work, struct ib_mad_agent_private, local_work);
2679 
2680 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2681 			       mad_agent_priv->qp_info->port_priv->port_num);
2682 
2683 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2684 	while (!list_empty(&mad_agent_priv->local_list)) {
2685 		local = list_entry(mad_agent_priv->local_list.next,
2686 				   struct ib_mad_local_private,
2687 				   completion_list);
2688 		list_del(&local->completion_list);
2689 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2690 		free_mad = 0;
2691 		if (local->mad_priv) {
2692 			u8 base_version;
2693 			recv_mad_agent = local->recv_mad_agent;
2694 			if (!recv_mad_agent) {
2695 				dev_err(&mad_agent_priv->agent.device->dev,
2696 					"No receive MAD agent for local completion\n");
2697 				free_mad = 1;
2698 				goto local_send_completion;
2699 			}
2700 
2701 			/*
2702 			 * Defined behavior is to complete response
2703 			 * before request
2704 			 */
2705 			build_smp_wc(recv_mad_agent->agent.qp,
2706 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2707 				     be16_to_cpu(IB_LID_PERMISSIVE),
2708 				     local->mad_send_wr->send_wr.pkey_index,
2709 				     recv_mad_agent->agent.port_num, &wc);
2710 
2711 			local->mad_priv->header.recv_wc.wc = &wc;
2712 
2713 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2714 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2715 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2716 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2717 			} else {
2718 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2719 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2720 			}
2721 
2722 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2723 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2724 				 &local->mad_priv->header.recv_wc.rmpp_list);
2725 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2726 			local->mad_priv->header.recv_wc.recv_buf.mad =
2727 						(struct ib_mad *)local->mad_priv->mad;
2728 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2729 				snoop_recv(recv_mad_agent->qp_info,
2730 					  &local->mad_priv->header.recv_wc,
2731 					   IB_MAD_SNOOP_RECVS);
2732 			recv_mad_agent->agent.recv_handler(
2733 						&recv_mad_agent->agent,
2734 						&local->mad_send_wr->send_buf,
2735 						&local->mad_priv->header.recv_wc);
2736 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2737 			atomic_dec(&recv_mad_agent->refcount);
2738 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2739 		}
2740 
2741 local_send_completion:
2742 		/* Complete send */
2743 		mad_send_wc.status = IB_WC_SUCCESS;
2744 		mad_send_wc.vendor_err = 0;
2745 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2746 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2747 			snoop_send(mad_agent_priv->qp_info,
2748 				   &local->mad_send_wr->send_buf,
2749 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2750 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2751 						   &mad_send_wc);
2752 
2753 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2754 		atomic_dec(&mad_agent_priv->refcount);
2755 		if (free_mad)
2756 			kfree(local->mad_priv);
2757 		kfree(local);
2758 	}
2759 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2760 }
2761 
2762 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2763 {
2764 	int ret;
2765 
2766 	if (!mad_send_wr->retries_left)
2767 		return -ETIMEDOUT;
2768 
2769 	mad_send_wr->retries_left--;
2770 	mad_send_wr->send_buf.retries++;
2771 
2772 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2773 
2774 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2775 		ret = ib_retry_rmpp(mad_send_wr);
2776 		switch (ret) {
2777 		case IB_RMPP_RESULT_UNHANDLED:
2778 			ret = ib_send_mad(mad_send_wr);
2779 			break;
2780 		case IB_RMPP_RESULT_CONSUMED:
2781 			ret = 0;
2782 			break;
2783 		default:
2784 			ret = -ECOMM;
2785 			break;
2786 		}
2787 	} else
2788 		ret = ib_send_mad(mad_send_wr);
2789 
2790 	if (!ret) {
2791 		mad_send_wr->refcount++;
2792 		list_add_tail(&mad_send_wr->agent_list,
2793 			      &mad_send_wr->mad_agent_priv->send_list);
2794 	}
2795 	return ret;
2796 }
2797 
2798 static void timeout_sends(struct work_struct *work)
2799 {
2800 	struct ib_mad_agent_private *mad_agent_priv;
2801 	struct ib_mad_send_wr_private *mad_send_wr;
2802 	struct ib_mad_send_wc mad_send_wc;
2803 	unsigned long flags, delay;
2804 
2805 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2806 				      timed_work.work);
2807 	mad_send_wc.vendor_err = 0;
2808 
2809 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2810 	while (!list_empty(&mad_agent_priv->wait_list)) {
2811 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2812 					 struct ib_mad_send_wr_private,
2813 					 agent_list);
2814 
2815 		if (time_after(mad_send_wr->timeout, jiffies)) {
2816 			delay = mad_send_wr->timeout - jiffies;
2817 			if ((long)delay <= 0)
2818 				delay = 1;
2819 			queue_delayed_work(mad_agent_priv->qp_info->
2820 					   port_priv->wq,
2821 					   &mad_agent_priv->timed_work, delay);
2822 			break;
2823 		}
2824 
2825 		list_del(&mad_send_wr->agent_list);
2826 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2827 		    !retry_send(mad_send_wr))
2828 			continue;
2829 
2830 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2831 
2832 		if (mad_send_wr->status == IB_WC_SUCCESS)
2833 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2834 		else
2835 			mad_send_wc.status = mad_send_wr->status;
2836 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2837 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2838 						   &mad_send_wc);
2839 
2840 		atomic_dec(&mad_agent_priv->refcount);
2841 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2842 	}
2843 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2844 }
2845 
2846 /*
2847  * Allocate receive MADs and post receive WRs for them
2848  */
2849 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2850 				    struct ib_mad_private *mad)
2851 {
2852 	unsigned long flags;
2853 	int post, ret;
2854 	struct ib_mad_private *mad_priv;
2855 	struct ib_sge sg_list;
2856 	struct ib_recv_wr recv_wr, *bad_recv_wr;
2857 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2858 
2859 	/* Initialize common scatter list fields */
2860 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2861 
2862 	/* Initialize common receive WR fields */
2863 	recv_wr.next = NULL;
2864 	recv_wr.sg_list = &sg_list;
2865 	recv_wr.num_sge = 1;
2866 
2867 	do {
2868 		/* Allocate and map receive buffer */
2869 		if (mad) {
2870 			mad_priv = mad;
2871 			mad = NULL;
2872 		} else {
2873 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2874 						     GFP_ATOMIC);
2875 			if (!mad_priv) {
2876 				dev_err(&qp_info->port_priv->device->dev,
2877 					"No memory for receive buffer\n");
2878 				ret = -ENOMEM;
2879 				break;
2880 			}
2881 		}
2882 		sg_list.length = mad_priv_dma_size(mad_priv);
2883 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2884 						 &mad_priv->grh,
2885 						 mad_priv_dma_size(mad_priv),
2886 						 DMA_FROM_DEVICE);
2887 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2888 						  sg_list.addr))) {
2889 			ret = -ENOMEM;
2890 			break;
2891 		}
2892 		mad_priv->header.mapping = sg_list.addr;
2893 		mad_priv->header.mad_list.mad_queue = recv_queue;
2894 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2895 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2896 
2897 		/* Post receive WR */
2898 		spin_lock_irqsave(&recv_queue->lock, flags);
2899 		post = (++recv_queue->count < recv_queue->max_active);
2900 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2901 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2902 		ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2903 		if (ret) {
2904 			spin_lock_irqsave(&recv_queue->lock, flags);
2905 			list_del(&mad_priv->header.mad_list.list);
2906 			recv_queue->count--;
2907 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2908 			ib_dma_unmap_single(qp_info->port_priv->device,
2909 					    mad_priv->header.mapping,
2910 					    mad_priv_dma_size(mad_priv),
2911 					    DMA_FROM_DEVICE);
2912 			kfree(mad_priv);
2913 			dev_err(&qp_info->port_priv->device->dev,
2914 				"ib_post_recv failed: %d\n", ret);
2915 			break;
2916 		}
2917 	} while (post);
2918 
2919 	return ret;
2920 }
2921 
2922 /*
2923  * Return all the posted receive MADs
2924  */
2925 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2926 {
2927 	struct ib_mad_private_header *mad_priv_hdr;
2928 	struct ib_mad_private *recv;
2929 	struct ib_mad_list_head *mad_list;
2930 
2931 	if (!qp_info->qp)
2932 		return;
2933 
2934 	while (!list_empty(&qp_info->recv_queue.list)) {
2935 
2936 		mad_list = list_entry(qp_info->recv_queue.list.next,
2937 				      struct ib_mad_list_head, list);
2938 		mad_priv_hdr = container_of(mad_list,
2939 					    struct ib_mad_private_header,
2940 					    mad_list);
2941 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2942 				    header);
2943 
2944 		/* Remove from posted receive MAD list */
2945 		list_del(&mad_list->list);
2946 
2947 		ib_dma_unmap_single(qp_info->port_priv->device,
2948 				    recv->header.mapping,
2949 				    mad_priv_dma_size(recv),
2950 				    DMA_FROM_DEVICE);
2951 		kfree(recv);
2952 	}
2953 
2954 	qp_info->recv_queue.count = 0;
2955 }
2956 
2957 /*
2958  * Start the port
2959  */
2960 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2961 {
2962 	int ret, i;
2963 	struct ib_qp_attr *attr;
2964 	struct ib_qp *qp;
2965 	u16 pkey_index;
2966 
2967 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2968 	if (!attr) {
2969 		dev_err(&port_priv->device->dev,
2970 			"Couldn't kmalloc ib_qp_attr\n");
2971 		return -ENOMEM;
2972 	}
2973 
2974 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2975 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2976 	if (ret)
2977 		pkey_index = 0;
2978 
2979 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2980 		qp = port_priv->qp_info[i].qp;
2981 		if (!qp)
2982 			continue;
2983 
2984 		/*
2985 		 * PKey index for QP1 is irrelevant but
2986 		 * one is needed for the Reset to Init transition
2987 		 */
2988 		attr->qp_state = IB_QPS_INIT;
2989 		attr->pkey_index = pkey_index;
2990 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2991 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2992 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2993 		if (ret) {
2994 			dev_err(&port_priv->device->dev,
2995 				"Couldn't change QP%d state to INIT: %d\n",
2996 				i, ret);
2997 			goto out;
2998 		}
2999 
3000 		attr->qp_state = IB_QPS_RTR;
3001 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3002 		if (ret) {
3003 			dev_err(&port_priv->device->dev,
3004 				"Couldn't change QP%d state to RTR: %d\n",
3005 				i, ret);
3006 			goto out;
3007 		}
3008 
3009 		attr->qp_state = IB_QPS_RTS;
3010 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3011 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3012 		if (ret) {
3013 			dev_err(&port_priv->device->dev,
3014 				"Couldn't change QP%d state to RTS: %d\n",
3015 				i, ret);
3016 			goto out;
3017 		}
3018 	}
3019 
3020 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3021 	if (ret) {
3022 		dev_err(&port_priv->device->dev,
3023 			"Failed to request completion notification: %d\n",
3024 			ret);
3025 		goto out;
3026 	}
3027 
3028 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3029 		if (!port_priv->qp_info[i].qp)
3030 			continue;
3031 
3032 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3033 		if (ret) {
3034 			dev_err(&port_priv->device->dev,
3035 				"Couldn't post receive WRs\n");
3036 			goto out;
3037 		}
3038 	}
3039 out:
3040 	kfree(attr);
3041 	return ret;
3042 }
3043 
3044 static void qp_event_handler(struct ib_event *event, void *qp_context)
3045 {
3046 	struct ib_mad_qp_info	*qp_info = qp_context;
3047 
3048 	/* It's worse than that! He's dead, Jim! */
3049 	dev_err(&qp_info->port_priv->device->dev,
3050 		"Fatal error (%d) on MAD QP (%d)\n",
3051 		event->event, qp_info->qp->qp_num);
3052 }
3053 
3054 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3055 			   struct ib_mad_queue *mad_queue)
3056 {
3057 	mad_queue->qp_info = qp_info;
3058 	mad_queue->count = 0;
3059 	spin_lock_init(&mad_queue->lock);
3060 	INIT_LIST_HEAD(&mad_queue->list);
3061 }
3062 
3063 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3064 			struct ib_mad_qp_info *qp_info)
3065 {
3066 	qp_info->port_priv = port_priv;
3067 	init_mad_queue(qp_info, &qp_info->send_queue);
3068 	init_mad_queue(qp_info, &qp_info->recv_queue);
3069 	INIT_LIST_HEAD(&qp_info->overflow_list);
3070 	spin_lock_init(&qp_info->snoop_lock);
3071 	qp_info->snoop_table = NULL;
3072 	qp_info->snoop_table_size = 0;
3073 	atomic_set(&qp_info->snoop_count, 0);
3074 }
3075 
3076 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3077 			 enum ib_qp_type qp_type)
3078 {
3079 	struct ib_qp_init_attr	qp_init_attr;
3080 	int ret;
3081 
3082 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3083 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3084 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3085 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3086 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3087 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3088 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3089 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3090 	qp_init_attr.qp_type = qp_type;
3091 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3092 	qp_init_attr.qp_context = qp_info;
3093 	qp_init_attr.event_handler = qp_event_handler;
3094 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3095 	if (IS_ERR(qp_info->qp)) {
3096 		dev_err(&qp_info->port_priv->device->dev,
3097 			"Couldn't create ib_mad QP%d\n",
3098 			get_spl_qp_index(qp_type));
3099 		ret = PTR_ERR(qp_info->qp);
3100 		goto error;
3101 	}
3102 	/* Use minimum queue sizes unless the CQ is resized */
3103 	qp_info->send_queue.max_active = mad_sendq_size;
3104 	qp_info->recv_queue.max_active = mad_recvq_size;
3105 	return 0;
3106 
3107 error:
3108 	return ret;
3109 }
3110 
3111 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3112 {
3113 	if (!qp_info->qp)
3114 		return;
3115 
3116 	ib_destroy_qp(qp_info->qp);
3117 	kfree(qp_info->snoop_table);
3118 }
3119 
3120 /*
3121  * Open the port
3122  * Create the QP, PD, MR, and CQ if needed
3123  */
3124 static int ib_mad_port_open(struct ib_device *device,
3125 			    int port_num)
3126 {
3127 	int ret, cq_size;
3128 	struct ib_mad_port_private *port_priv;
3129 	unsigned long flags;
3130 	char name[sizeof "ib_mad123"];
3131 	int has_smi;
3132 
3133 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3134 		return -EFAULT;
3135 
3136 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3137 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3138 		return -EFAULT;
3139 
3140 	/* Create new device info */
3141 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3142 	if (!port_priv) {
3143 		dev_err(&device->dev, "No memory for ib_mad_port_private\n");
3144 		return -ENOMEM;
3145 	}
3146 
3147 	port_priv->device = device;
3148 	port_priv->port_num = port_num;
3149 	spin_lock_init(&port_priv->reg_lock);
3150 	INIT_LIST_HEAD(&port_priv->agent_list);
3151 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3152 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3153 
3154 	cq_size = mad_sendq_size + mad_recvq_size;
3155 	has_smi = rdma_cap_ib_smi(device, port_num);
3156 	if (has_smi)
3157 		cq_size *= 2;
3158 
3159 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3160 			IB_POLL_WORKQUEUE);
3161 	if (IS_ERR(port_priv->cq)) {
3162 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3163 		ret = PTR_ERR(port_priv->cq);
3164 		goto error3;
3165 	}
3166 
3167 	port_priv->pd = ib_alloc_pd(device);
3168 	if (IS_ERR(port_priv->pd)) {
3169 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3170 		ret = PTR_ERR(port_priv->pd);
3171 		goto error4;
3172 	}
3173 
3174 	if (has_smi) {
3175 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3176 		if (ret)
3177 			goto error6;
3178 	}
3179 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3180 	if (ret)
3181 		goto error7;
3182 
3183 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3184 	port_priv->wq = create_singlethread_workqueue(name);
3185 	if (!port_priv->wq) {
3186 		ret = -ENOMEM;
3187 		goto error8;
3188 	}
3189 
3190 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3191 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3192 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3193 
3194 	ret = ib_mad_port_start(port_priv);
3195 	if (ret) {
3196 		dev_err(&device->dev, "Couldn't start port\n");
3197 		goto error9;
3198 	}
3199 
3200 	return 0;
3201 
3202 error9:
3203 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3204 	list_del_init(&port_priv->port_list);
3205 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3206 
3207 	destroy_workqueue(port_priv->wq);
3208 error8:
3209 	destroy_mad_qp(&port_priv->qp_info[1]);
3210 error7:
3211 	destroy_mad_qp(&port_priv->qp_info[0]);
3212 error6:
3213 	ib_dealloc_pd(port_priv->pd);
3214 error4:
3215 	ib_free_cq(port_priv->cq);
3216 	cleanup_recv_queue(&port_priv->qp_info[1]);
3217 	cleanup_recv_queue(&port_priv->qp_info[0]);
3218 error3:
3219 	kfree(port_priv);
3220 
3221 	return ret;
3222 }
3223 
3224 /*
3225  * Close the port
3226  * If there are no classes using the port, free the port
3227  * resources (CQ, MR, PD, QP) and remove the port's info structure
3228  */
3229 static int ib_mad_port_close(struct ib_device *device, int port_num)
3230 {
3231 	struct ib_mad_port_private *port_priv;
3232 	unsigned long flags;
3233 
3234 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3235 	port_priv = __ib_get_mad_port(device, port_num);
3236 	if (port_priv == NULL) {
3237 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3238 		dev_err(&device->dev, "Port %d not found\n", port_num);
3239 		return -ENODEV;
3240 	}
3241 	list_del_init(&port_priv->port_list);
3242 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3243 
3244 	destroy_workqueue(port_priv->wq);
3245 	destroy_mad_qp(&port_priv->qp_info[1]);
3246 	destroy_mad_qp(&port_priv->qp_info[0]);
3247 	ib_dealloc_pd(port_priv->pd);
3248 	ib_free_cq(port_priv->cq);
3249 	cleanup_recv_queue(&port_priv->qp_info[1]);
3250 	cleanup_recv_queue(&port_priv->qp_info[0]);
3251 	/* XXX: Handle deallocation of MAD registration tables */
3252 
3253 	kfree(port_priv);
3254 
3255 	return 0;
3256 }
3257 
3258 static void ib_mad_init_device(struct ib_device *device)
3259 {
3260 	int start, i;
3261 
3262 	start = rdma_start_port(device);
3263 
3264 	for (i = start; i <= rdma_end_port(device); i++) {
3265 		if (!rdma_cap_ib_mad(device, i))
3266 			continue;
3267 
3268 		if (ib_mad_port_open(device, i)) {
3269 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3270 			goto error;
3271 		}
3272 		if (ib_agent_port_open(device, i)) {
3273 			dev_err(&device->dev,
3274 				"Couldn't open port %d for agents\n", i);
3275 			goto error_agent;
3276 		}
3277 	}
3278 	return;
3279 
3280 error_agent:
3281 	if (ib_mad_port_close(device, i))
3282 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3283 
3284 error:
3285 	while (--i >= start) {
3286 		if (!rdma_cap_ib_mad(device, i))
3287 			continue;
3288 
3289 		if (ib_agent_port_close(device, i))
3290 			dev_err(&device->dev,
3291 				"Couldn't close port %d for agents\n", i);
3292 		if (ib_mad_port_close(device, i))
3293 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3294 	}
3295 }
3296 
3297 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3298 {
3299 	int i;
3300 
3301 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3302 		if (!rdma_cap_ib_mad(device, i))
3303 			continue;
3304 
3305 		if (ib_agent_port_close(device, i))
3306 			dev_err(&device->dev,
3307 				"Couldn't close port %d for agents\n", i);
3308 		if (ib_mad_port_close(device, i))
3309 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3310 	}
3311 }
3312 
3313 static struct ib_client mad_client = {
3314 	.name   = "mad",
3315 	.add = ib_mad_init_device,
3316 	.remove = ib_mad_remove_device
3317 };
3318 
3319 static int __init ib_mad_init_module(void)
3320 {
3321 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3322 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3323 
3324 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3325 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3326 
3327 	INIT_LIST_HEAD(&ib_mad_port_list);
3328 
3329 	if (ib_register_client(&mad_client)) {
3330 		pr_err("Couldn't register ib_mad client\n");
3331 		return -EINVAL;
3332 	}
3333 
3334 	return 0;
3335 }
3336 
3337 static void __exit ib_mad_cleanup_module(void)
3338 {
3339 	ib_unregister_client(&mad_client);
3340 }
3341 
3342 module_init(ib_mad_init_module);
3343 module_exit(ib_mad_cleanup_module);
3344