xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision 0c874100)
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/idr.h>
42 #include <linux/slab.h>
43 #include <linux/module.h>
44 #include <linux/security.h>
45 #include <rdma/ib_cache.h>
46 
47 #include "mad_priv.h"
48 #include "core_priv.h"
49 #include "mad_rmpp.h"
50 #include "smi.h"
51 #include "opa_smi.h"
52 #include "agent.h"
53 
54 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
55 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
56 
57 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
58 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
59 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
60 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
61 
62 /*
63  * The mlx4 driver uses the top byte to distinguish which virtual function
64  * generated the MAD, so we must avoid using it.
65  */
66 #define AGENT_ID_LIMIT		(1 << 24)
67 static DEFINE_IDR(ib_mad_clients);
68 static struct list_head ib_mad_port_list;
69 
70 /* Port list lock */
71 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
72 
73 /* Forward declarations */
74 static int method_in_use(struct ib_mad_mgmt_method_table **method,
75 			 struct ib_mad_reg_req *mad_reg_req);
76 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
77 static struct ib_mad_agent_private *find_mad_agent(
78 					struct ib_mad_port_private *port_priv,
79 					const struct ib_mad_hdr *mad);
80 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
81 				    struct ib_mad_private *mad);
82 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
83 static void timeout_sends(struct work_struct *work);
84 static void local_completions(struct work_struct *work);
85 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
86 			      struct ib_mad_agent_private *agent_priv,
87 			      u8 mgmt_class);
88 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
89 			   struct ib_mad_agent_private *agent_priv);
90 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
91 			      struct ib_wc *wc);
92 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
93 
94 /*
95  * Returns a ib_mad_port_private structure or NULL for a device/port
96  * Assumes ib_mad_port_list_lock is being held
97  */
98 static inline struct ib_mad_port_private *
99 __ib_get_mad_port(struct ib_device *device, int port_num)
100 {
101 	struct ib_mad_port_private *entry;
102 
103 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
104 		if (entry->device == device && entry->port_num == port_num)
105 			return entry;
106 	}
107 	return NULL;
108 }
109 
110 /*
111  * Wrapper function to return a ib_mad_port_private structure or NULL
112  * for a device/port
113  */
114 static inline struct ib_mad_port_private *
115 ib_get_mad_port(struct ib_device *device, int port_num)
116 {
117 	struct ib_mad_port_private *entry;
118 	unsigned long flags;
119 
120 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
121 	entry = __ib_get_mad_port(device, port_num);
122 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
123 
124 	return entry;
125 }
126 
127 static inline u8 convert_mgmt_class(u8 mgmt_class)
128 {
129 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
130 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
131 		0 : mgmt_class;
132 }
133 
134 static int get_spl_qp_index(enum ib_qp_type qp_type)
135 {
136 	switch (qp_type)
137 	{
138 	case IB_QPT_SMI:
139 		return 0;
140 	case IB_QPT_GSI:
141 		return 1;
142 	default:
143 		return -1;
144 	}
145 }
146 
147 static int vendor_class_index(u8 mgmt_class)
148 {
149 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
150 }
151 
152 static int is_vendor_class(u8 mgmt_class)
153 {
154 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
155 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
156 		return 0;
157 	return 1;
158 }
159 
160 static int is_vendor_oui(char *oui)
161 {
162 	if (oui[0] || oui[1] || oui[2])
163 		return 1;
164 	return 0;
165 }
166 
167 static int is_vendor_method_in_use(
168 		struct ib_mad_mgmt_vendor_class *vendor_class,
169 		struct ib_mad_reg_req *mad_reg_req)
170 {
171 	struct ib_mad_mgmt_method_table *method;
172 	int i;
173 
174 	for (i = 0; i < MAX_MGMT_OUI; i++) {
175 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
176 			method = vendor_class->method_table[i];
177 			if (method) {
178 				if (method_in_use(&method, mad_reg_req))
179 					return 1;
180 				else
181 					break;
182 			}
183 		}
184 	}
185 	return 0;
186 }
187 
188 int ib_response_mad(const struct ib_mad_hdr *hdr)
189 {
190 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
191 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
192 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
193 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
194 }
195 EXPORT_SYMBOL(ib_response_mad);
196 
197 /*
198  * ib_register_mad_agent - Register to send/receive MADs
199  *
200  * Context: Process context.
201  */
202 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
203 					   u8 port_num,
204 					   enum ib_qp_type qp_type,
205 					   struct ib_mad_reg_req *mad_reg_req,
206 					   u8 rmpp_version,
207 					   ib_mad_send_handler send_handler,
208 					   ib_mad_recv_handler recv_handler,
209 					   void *context,
210 					   u32 registration_flags)
211 {
212 	struct ib_mad_port_private *port_priv;
213 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
214 	struct ib_mad_agent_private *mad_agent_priv;
215 	struct ib_mad_reg_req *reg_req = NULL;
216 	struct ib_mad_mgmt_class_table *class;
217 	struct ib_mad_mgmt_vendor_class_table *vendor;
218 	struct ib_mad_mgmt_vendor_class *vendor_class;
219 	struct ib_mad_mgmt_method_table *method;
220 	int ret2, qpn;
221 	u8 mgmt_class, vclass;
222 
223 	if ((qp_type == IB_QPT_SMI && !rdma_cap_ib_smi(device, port_num)) ||
224 	    (qp_type == IB_QPT_GSI && !rdma_cap_ib_cm(device, port_num)))
225 		return ERR_PTR(-EPROTONOSUPPORT);
226 
227 	/* Validate parameters */
228 	qpn = get_spl_qp_index(qp_type);
229 	if (qpn == -1) {
230 		dev_dbg_ratelimited(&device->dev, "%s: invalid QP Type %d\n",
231 				    __func__, qp_type);
232 		goto error1;
233 	}
234 
235 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
236 		dev_dbg_ratelimited(&device->dev,
237 				    "%s: invalid RMPP Version %u\n",
238 				    __func__, rmpp_version);
239 		goto error1;
240 	}
241 
242 	/* Validate MAD registration request if supplied */
243 	if (mad_reg_req) {
244 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
245 			dev_dbg_ratelimited(&device->dev,
246 					    "%s: invalid Class Version %u\n",
247 					    __func__,
248 					    mad_reg_req->mgmt_class_version);
249 			goto error1;
250 		}
251 		if (!recv_handler) {
252 			dev_dbg_ratelimited(&device->dev,
253 					    "%s: no recv_handler\n", __func__);
254 			goto error1;
255 		}
256 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
257 			/*
258 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
259 			 * one in this range currently allowed
260 			 */
261 			if (mad_reg_req->mgmt_class !=
262 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
263 				dev_dbg_ratelimited(&device->dev,
264 					"%s: Invalid Mgmt Class 0x%x\n",
265 					__func__, mad_reg_req->mgmt_class);
266 				goto error1;
267 			}
268 		} else if (mad_reg_req->mgmt_class == 0) {
269 			/*
270 			 * Class 0 is reserved in IBA and is used for
271 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
272 			 */
273 			dev_dbg_ratelimited(&device->dev,
274 					    "%s: Invalid Mgmt Class 0\n",
275 					    __func__);
276 			goto error1;
277 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
278 			/*
279 			 * If class is in "new" vendor range,
280 			 * ensure supplied OUI is not zero
281 			 */
282 			if (!is_vendor_oui(mad_reg_req->oui)) {
283 				dev_dbg_ratelimited(&device->dev,
284 					"%s: No OUI specified for class 0x%x\n",
285 					__func__,
286 					mad_reg_req->mgmt_class);
287 				goto error1;
288 			}
289 		}
290 		/* Make sure class supplied is consistent with RMPP */
291 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
292 			if (rmpp_version) {
293 				dev_dbg_ratelimited(&device->dev,
294 					"%s: RMPP version for non-RMPP class 0x%x\n",
295 					__func__, mad_reg_req->mgmt_class);
296 				goto error1;
297 			}
298 		}
299 
300 		/* Make sure class supplied is consistent with QP type */
301 		if (qp_type == IB_QPT_SMI) {
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_dbg_ratelimited(&device->dev,
307 					"%s: Invalid SM QP type: class 0x%x\n",
308 					__func__, mad_reg_req->mgmt_class);
309 				goto error1;
310 			}
311 		} else {
312 			if ((mad_reg_req->mgmt_class ==
313 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
314 			    (mad_reg_req->mgmt_class ==
315 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
316 				dev_dbg_ratelimited(&device->dev,
317 					"%s: Invalid GS QP type: class 0x%x\n",
318 					__func__, mad_reg_req->mgmt_class);
319 				goto error1;
320 			}
321 		}
322 	} else {
323 		/* No registration request supplied */
324 		if (!send_handler)
325 			goto error1;
326 		if (registration_flags & IB_MAD_USER_RMPP)
327 			goto error1;
328 	}
329 
330 	/* Validate device and port */
331 	port_priv = ib_get_mad_port(device, port_num);
332 	if (!port_priv) {
333 		dev_dbg_ratelimited(&device->dev, "%s: Invalid port %d\n",
334 				    __func__, port_num);
335 		ret = ERR_PTR(-ENODEV);
336 		goto error1;
337 	}
338 
339 	/* Verify the QP requested is supported. For example, Ethernet devices
340 	 * will not have QP0.
341 	 */
342 	if (!port_priv->qp_info[qpn].qp) {
343 		dev_dbg_ratelimited(&device->dev, "%s: QP %d not supported\n",
344 				    __func__, qpn);
345 		ret = ERR_PTR(-EPROTONOSUPPORT);
346 		goto error1;
347 	}
348 
349 	/* Allocate structures */
350 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
351 	if (!mad_agent_priv) {
352 		ret = ERR_PTR(-ENOMEM);
353 		goto error1;
354 	}
355 
356 	if (mad_reg_req) {
357 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
358 		if (!reg_req) {
359 			ret = ERR_PTR(-ENOMEM);
360 			goto error3;
361 		}
362 	}
363 
364 	/* Now, fill in the various structures */
365 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
366 	mad_agent_priv->reg_req = reg_req;
367 	mad_agent_priv->agent.rmpp_version = rmpp_version;
368 	mad_agent_priv->agent.device = device;
369 	mad_agent_priv->agent.recv_handler = recv_handler;
370 	mad_agent_priv->agent.send_handler = send_handler;
371 	mad_agent_priv->agent.context = context;
372 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
373 	mad_agent_priv->agent.port_num = port_num;
374 	mad_agent_priv->agent.flags = registration_flags;
375 	spin_lock_init(&mad_agent_priv->lock);
376 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
377 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
378 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
379 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
380 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
381 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
382 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
383 	atomic_set(&mad_agent_priv->refcount, 1);
384 	init_completion(&mad_agent_priv->comp);
385 
386 	ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
387 	if (ret2) {
388 		ret = ERR_PTR(ret2);
389 		goto error4;
390 	}
391 
392 	idr_preload(GFP_KERNEL);
393 	idr_lock(&ib_mad_clients);
394 	ret2 = idr_alloc_cyclic(&ib_mad_clients, mad_agent_priv, 0,
395 			AGENT_ID_LIMIT, GFP_ATOMIC);
396 	idr_unlock(&ib_mad_clients);
397 	idr_preload_end();
398 
399 	if (ret2 < 0) {
400 		ret = ERR_PTR(ret2);
401 		goto error5;
402 	}
403 	mad_agent_priv->agent.hi_tid = ret2;
404 
405 	/*
406 	 * Make sure MAD registration (if supplied)
407 	 * is non overlapping with any existing ones
408 	 */
409 	spin_lock_irq(&port_priv->reg_lock);
410 	if (mad_reg_req) {
411 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
412 		if (!is_vendor_class(mgmt_class)) {
413 			class = port_priv->version[mad_reg_req->
414 						   mgmt_class_version].class;
415 			if (class) {
416 				method = class->method_table[mgmt_class];
417 				if (method) {
418 					if (method_in_use(&method,
419 							   mad_reg_req))
420 						goto error6;
421 				}
422 			}
423 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
424 						  mgmt_class);
425 		} else {
426 			/* "New" vendor class range */
427 			vendor = port_priv->version[mad_reg_req->
428 						    mgmt_class_version].vendor;
429 			if (vendor) {
430 				vclass = vendor_class_index(mgmt_class);
431 				vendor_class = vendor->vendor_class[vclass];
432 				if (vendor_class) {
433 					if (is_vendor_method_in_use(
434 							vendor_class,
435 							mad_reg_req))
436 						goto error6;
437 				}
438 			}
439 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
440 		}
441 		if (ret2) {
442 			ret = ERR_PTR(ret2);
443 			goto error6;
444 		}
445 	}
446 	spin_unlock_irq(&port_priv->reg_lock);
447 
448 	return &mad_agent_priv->agent;
449 error6:
450 	spin_unlock_irq(&port_priv->reg_lock);
451 	idr_lock(&ib_mad_clients);
452 	idr_remove(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
453 	idr_unlock(&ib_mad_clients);
454 error5:
455 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
456 error4:
457 	kfree(reg_req);
458 error3:
459 	kfree(mad_agent_priv);
460 error1:
461 	return ret;
462 }
463 EXPORT_SYMBOL(ib_register_mad_agent);
464 
465 static inline int is_snooping_sends(int mad_snoop_flags)
466 {
467 	return (mad_snoop_flags &
468 		(/*IB_MAD_SNOOP_POSTED_SENDS |
469 		 IB_MAD_SNOOP_RMPP_SENDS |*/
470 		 IB_MAD_SNOOP_SEND_COMPLETIONS /*|
471 		 IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
472 }
473 
474 static inline int is_snooping_recvs(int mad_snoop_flags)
475 {
476 	return (mad_snoop_flags &
477 		(IB_MAD_SNOOP_RECVS /*|
478 		 IB_MAD_SNOOP_RMPP_RECVS*/));
479 }
480 
481 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
482 				struct ib_mad_snoop_private *mad_snoop_priv)
483 {
484 	struct ib_mad_snoop_private **new_snoop_table;
485 	unsigned long flags;
486 	int i;
487 
488 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
489 	/* Check for empty slot in array. */
490 	for (i = 0; i < qp_info->snoop_table_size; i++)
491 		if (!qp_info->snoop_table[i])
492 			break;
493 
494 	if (i == qp_info->snoop_table_size) {
495 		/* Grow table. */
496 		new_snoop_table = krealloc(qp_info->snoop_table,
497 					   sizeof mad_snoop_priv *
498 					   (qp_info->snoop_table_size + 1),
499 					   GFP_ATOMIC);
500 		if (!new_snoop_table) {
501 			i = -ENOMEM;
502 			goto out;
503 		}
504 
505 		qp_info->snoop_table = new_snoop_table;
506 		qp_info->snoop_table_size++;
507 	}
508 	qp_info->snoop_table[i] = mad_snoop_priv;
509 	atomic_inc(&qp_info->snoop_count);
510 out:
511 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
512 	return i;
513 }
514 
515 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
516 					   u8 port_num,
517 					   enum ib_qp_type qp_type,
518 					   int mad_snoop_flags,
519 					   ib_mad_snoop_handler snoop_handler,
520 					   ib_mad_recv_handler recv_handler,
521 					   void *context)
522 {
523 	struct ib_mad_port_private *port_priv;
524 	struct ib_mad_agent *ret;
525 	struct ib_mad_snoop_private *mad_snoop_priv;
526 	int qpn;
527 	int err;
528 
529 	/* Validate parameters */
530 	if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
531 	    (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
532 		ret = ERR_PTR(-EINVAL);
533 		goto error1;
534 	}
535 	qpn = get_spl_qp_index(qp_type);
536 	if (qpn == -1) {
537 		ret = ERR_PTR(-EINVAL);
538 		goto error1;
539 	}
540 	port_priv = ib_get_mad_port(device, port_num);
541 	if (!port_priv) {
542 		ret = ERR_PTR(-ENODEV);
543 		goto error1;
544 	}
545 	/* Allocate structures */
546 	mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
547 	if (!mad_snoop_priv) {
548 		ret = ERR_PTR(-ENOMEM);
549 		goto error1;
550 	}
551 
552 	/* Now, fill in the various structures */
553 	mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
554 	mad_snoop_priv->agent.device = device;
555 	mad_snoop_priv->agent.recv_handler = recv_handler;
556 	mad_snoop_priv->agent.snoop_handler = snoop_handler;
557 	mad_snoop_priv->agent.context = context;
558 	mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
559 	mad_snoop_priv->agent.port_num = port_num;
560 	mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
561 	init_completion(&mad_snoop_priv->comp);
562 
563 	err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
564 	if (err) {
565 		ret = ERR_PTR(err);
566 		goto error2;
567 	}
568 
569 	mad_snoop_priv->snoop_index = register_snoop_agent(
570 						&port_priv->qp_info[qpn],
571 						mad_snoop_priv);
572 	if (mad_snoop_priv->snoop_index < 0) {
573 		ret = ERR_PTR(mad_snoop_priv->snoop_index);
574 		goto error3;
575 	}
576 
577 	atomic_set(&mad_snoop_priv->refcount, 1);
578 	return &mad_snoop_priv->agent;
579 error3:
580 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
581 error2:
582 	kfree(mad_snoop_priv);
583 error1:
584 	return ret;
585 }
586 EXPORT_SYMBOL(ib_register_mad_snoop);
587 
588 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
589 {
590 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
591 		complete(&mad_agent_priv->comp);
592 }
593 
594 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
595 {
596 	if (atomic_dec_and_test(&mad_snoop_priv->refcount))
597 		complete(&mad_snoop_priv->comp);
598 }
599 
600 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
601 {
602 	struct ib_mad_port_private *port_priv;
603 
604 	/* Note that we could still be handling received MADs */
605 
606 	/*
607 	 * Canceling all sends results in dropping received response
608 	 * MADs, preventing us from queuing additional work
609 	 */
610 	cancel_mads(mad_agent_priv);
611 	port_priv = mad_agent_priv->qp_info->port_priv;
612 	cancel_delayed_work(&mad_agent_priv->timed_work);
613 
614 	spin_lock_irq(&port_priv->reg_lock);
615 	remove_mad_reg_req(mad_agent_priv);
616 	spin_unlock_irq(&port_priv->reg_lock);
617 	idr_lock(&ib_mad_clients);
618 	idr_remove(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
619 	idr_unlock(&ib_mad_clients);
620 
621 	flush_workqueue(port_priv->wq);
622 	ib_cancel_rmpp_recvs(mad_agent_priv);
623 
624 	deref_mad_agent(mad_agent_priv);
625 	wait_for_completion(&mad_agent_priv->comp);
626 
627 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
628 
629 	kfree(mad_agent_priv->reg_req);
630 	kfree_rcu(mad_agent_priv, rcu);
631 }
632 
633 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
634 {
635 	struct ib_mad_qp_info *qp_info;
636 	unsigned long flags;
637 
638 	qp_info = mad_snoop_priv->qp_info;
639 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
640 	qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
641 	atomic_dec(&qp_info->snoop_count);
642 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
643 
644 	deref_snoop_agent(mad_snoop_priv);
645 	wait_for_completion(&mad_snoop_priv->comp);
646 
647 	ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
648 
649 	kfree(mad_snoop_priv);
650 }
651 
652 /*
653  * ib_unregister_mad_agent - Unregisters a client from using MAD services
654  *
655  * Context: Process context.
656  */
657 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
658 {
659 	struct ib_mad_agent_private *mad_agent_priv;
660 	struct ib_mad_snoop_private *mad_snoop_priv;
661 
662 	/* If the TID is zero, the agent can only snoop. */
663 	if (mad_agent->hi_tid) {
664 		mad_agent_priv = container_of(mad_agent,
665 					      struct ib_mad_agent_private,
666 					      agent);
667 		unregister_mad_agent(mad_agent_priv);
668 	} else {
669 		mad_snoop_priv = container_of(mad_agent,
670 					      struct ib_mad_snoop_private,
671 					      agent);
672 		unregister_mad_snoop(mad_snoop_priv);
673 	}
674 }
675 EXPORT_SYMBOL(ib_unregister_mad_agent);
676 
677 static void dequeue_mad(struct ib_mad_list_head *mad_list)
678 {
679 	struct ib_mad_queue *mad_queue;
680 	unsigned long flags;
681 
682 	mad_queue = mad_list->mad_queue;
683 	spin_lock_irqsave(&mad_queue->lock, flags);
684 	list_del(&mad_list->list);
685 	mad_queue->count--;
686 	spin_unlock_irqrestore(&mad_queue->lock, flags);
687 }
688 
689 static void snoop_send(struct ib_mad_qp_info *qp_info,
690 		       struct ib_mad_send_buf *send_buf,
691 		       struct ib_mad_send_wc *mad_send_wc,
692 		       int mad_snoop_flags)
693 {
694 	struct ib_mad_snoop_private *mad_snoop_priv;
695 	unsigned long flags;
696 	int i;
697 
698 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
699 	for (i = 0; i < qp_info->snoop_table_size; i++) {
700 		mad_snoop_priv = qp_info->snoop_table[i];
701 		if (!mad_snoop_priv ||
702 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
703 			continue;
704 
705 		atomic_inc(&mad_snoop_priv->refcount);
706 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
707 		mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
708 						    send_buf, mad_send_wc);
709 		deref_snoop_agent(mad_snoop_priv);
710 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
711 	}
712 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
713 }
714 
715 static void snoop_recv(struct ib_mad_qp_info *qp_info,
716 		       struct ib_mad_recv_wc *mad_recv_wc,
717 		       int mad_snoop_flags)
718 {
719 	struct ib_mad_snoop_private *mad_snoop_priv;
720 	unsigned long flags;
721 	int i;
722 
723 	spin_lock_irqsave(&qp_info->snoop_lock, flags);
724 	for (i = 0; i < qp_info->snoop_table_size; i++) {
725 		mad_snoop_priv = qp_info->snoop_table[i];
726 		if (!mad_snoop_priv ||
727 		    !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
728 			continue;
729 
730 		atomic_inc(&mad_snoop_priv->refcount);
731 		spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
732 		mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
733 						   mad_recv_wc);
734 		deref_snoop_agent(mad_snoop_priv);
735 		spin_lock_irqsave(&qp_info->snoop_lock, flags);
736 	}
737 	spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
738 }
739 
740 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
741 		u16 pkey_index, u8 port_num, struct ib_wc *wc)
742 {
743 	memset(wc, 0, sizeof *wc);
744 	wc->wr_cqe = cqe;
745 	wc->status = IB_WC_SUCCESS;
746 	wc->opcode = IB_WC_RECV;
747 	wc->pkey_index = pkey_index;
748 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
749 	wc->src_qp = IB_QP0;
750 	wc->qp = qp;
751 	wc->slid = slid;
752 	wc->sl = 0;
753 	wc->dlid_path_bits = 0;
754 	wc->port_num = port_num;
755 }
756 
757 static size_t mad_priv_size(const struct ib_mad_private *mp)
758 {
759 	return sizeof(struct ib_mad_private) + mp->mad_size;
760 }
761 
762 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
763 {
764 	size_t size = sizeof(struct ib_mad_private) + mad_size;
765 	struct ib_mad_private *ret = kzalloc(size, flags);
766 
767 	if (ret)
768 		ret->mad_size = mad_size;
769 
770 	return ret;
771 }
772 
773 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
774 {
775 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
776 }
777 
778 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
779 {
780 	return sizeof(struct ib_grh) + mp->mad_size;
781 }
782 
783 /*
784  * Return 0 if SMP is to be sent
785  * Return 1 if SMP was consumed locally (whether or not solicited)
786  * Return < 0 if error
787  */
788 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
789 				  struct ib_mad_send_wr_private *mad_send_wr)
790 {
791 	int ret = 0;
792 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
793 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
794 	unsigned long flags;
795 	struct ib_mad_local_private *local;
796 	struct ib_mad_private *mad_priv;
797 	struct ib_mad_port_private *port_priv;
798 	struct ib_mad_agent_private *recv_mad_agent = NULL;
799 	struct ib_device *device = mad_agent_priv->agent.device;
800 	u8 port_num;
801 	struct ib_wc mad_wc;
802 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
803 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
804 	u16 out_mad_pkey_index = 0;
805 	u16 drslid;
806 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
807 				    mad_agent_priv->qp_info->port_priv->port_num);
808 
809 	if (rdma_cap_ib_switch(device) &&
810 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
811 		port_num = send_wr->port_num;
812 	else
813 		port_num = mad_agent_priv->agent.port_num;
814 
815 	/*
816 	 * Directed route handling starts if the initial LID routed part of
817 	 * a request or the ending LID routed part of a response is empty.
818 	 * If we are at the start of the LID routed part, don't update the
819 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
820 	 */
821 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
822 		u32 opa_drslid;
823 
824 		if ((opa_get_smp_direction(opa_smp)
825 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
826 		     OPA_LID_PERMISSIVE &&
827 		     opa_smi_handle_dr_smp_send(opa_smp,
828 						rdma_cap_ib_switch(device),
829 						port_num) == IB_SMI_DISCARD) {
830 			ret = -EINVAL;
831 			dev_err(&device->dev, "OPA Invalid directed route\n");
832 			goto out;
833 		}
834 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
835 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
836 		    opa_drslid & 0xffff0000) {
837 			ret = -EINVAL;
838 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
839 			       opa_drslid);
840 			goto out;
841 		}
842 		drslid = (u16)(opa_drslid & 0x0000ffff);
843 
844 		/* Check to post send on QP or process locally */
845 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
846 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
847 			goto out;
848 	} else {
849 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
850 		     IB_LID_PERMISSIVE &&
851 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
852 		     IB_SMI_DISCARD) {
853 			ret = -EINVAL;
854 			dev_err(&device->dev, "Invalid directed route\n");
855 			goto out;
856 		}
857 		drslid = be16_to_cpu(smp->dr_slid);
858 
859 		/* Check to post send on QP or process locally */
860 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
861 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
862 			goto out;
863 	}
864 
865 	local = kmalloc(sizeof *local, GFP_ATOMIC);
866 	if (!local) {
867 		ret = -ENOMEM;
868 		goto out;
869 	}
870 	local->mad_priv = NULL;
871 	local->recv_mad_agent = NULL;
872 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
873 	if (!mad_priv) {
874 		ret = -ENOMEM;
875 		kfree(local);
876 		goto out;
877 	}
878 
879 	build_smp_wc(mad_agent_priv->agent.qp,
880 		     send_wr->wr.wr_cqe, drslid,
881 		     send_wr->pkey_index,
882 		     send_wr->port_num, &mad_wc);
883 
884 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
885 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
886 					+ mad_send_wr->send_buf.data_len
887 					+ sizeof(struct ib_grh);
888 	}
889 
890 	/* No GRH for DR SMP */
891 	ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
892 				  (const struct ib_mad_hdr *)smp, mad_size,
893 				  (struct ib_mad_hdr *)mad_priv->mad,
894 				  &mad_size, &out_mad_pkey_index);
895 	switch (ret)
896 	{
897 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
898 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
899 		    mad_agent_priv->agent.recv_handler) {
900 			local->mad_priv = mad_priv;
901 			local->recv_mad_agent = mad_agent_priv;
902 			/*
903 			 * Reference MAD agent until receive
904 			 * side of local completion handled
905 			 */
906 			atomic_inc(&mad_agent_priv->refcount);
907 		} else
908 			kfree(mad_priv);
909 		break;
910 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
911 		kfree(mad_priv);
912 		break;
913 	case IB_MAD_RESULT_SUCCESS:
914 		/* Treat like an incoming receive MAD */
915 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
916 					    mad_agent_priv->agent.port_num);
917 		if (port_priv) {
918 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
919 			recv_mad_agent = find_mad_agent(port_priv,
920 						        (const struct ib_mad_hdr *)mad_priv->mad);
921 		}
922 		if (!port_priv || !recv_mad_agent) {
923 			/*
924 			 * No receiving agent so drop packet and
925 			 * generate send completion.
926 			 */
927 			kfree(mad_priv);
928 			break;
929 		}
930 		local->mad_priv = mad_priv;
931 		local->recv_mad_agent = recv_mad_agent;
932 		break;
933 	default:
934 		kfree(mad_priv);
935 		kfree(local);
936 		ret = -EINVAL;
937 		goto out;
938 	}
939 
940 	local->mad_send_wr = mad_send_wr;
941 	if (opa) {
942 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
943 		local->return_wc_byte_len = mad_size;
944 	}
945 	/* Reference MAD agent until send side of local completion handled */
946 	atomic_inc(&mad_agent_priv->refcount);
947 	/* Queue local completion to local list */
948 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
949 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
950 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
951 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
952 		   &mad_agent_priv->local_work);
953 	ret = 1;
954 out:
955 	return ret;
956 }
957 
958 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
959 {
960 	int seg_size, pad;
961 
962 	seg_size = mad_size - hdr_len;
963 	if (data_len && seg_size) {
964 		pad = seg_size - data_len % seg_size;
965 		return pad == seg_size ? 0 : pad;
966 	} else
967 		return seg_size;
968 }
969 
970 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
971 {
972 	struct ib_rmpp_segment *s, *t;
973 
974 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
975 		list_del(&s->list);
976 		kfree(s);
977 	}
978 }
979 
980 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
981 				size_t mad_size, gfp_t gfp_mask)
982 {
983 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
984 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
985 	struct ib_rmpp_segment *seg = NULL;
986 	int left, seg_size, pad;
987 
988 	send_buf->seg_size = mad_size - send_buf->hdr_len;
989 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
990 	seg_size = send_buf->seg_size;
991 	pad = send_wr->pad;
992 
993 	/* Allocate data segments. */
994 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
995 		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
996 		if (!seg) {
997 			free_send_rmpp_list(send_wr);
998 			return -ENOMEM;
999 		}
1000 		seg->num = ++send_buf->seg_count;
1001 		list_add_tail(&seg->list, &send_wr->rmpp_list);
1002 	}
1003 
1004 	/* Zero any padding */
1005 	if (pad)
1006 		memset(seg->data + seg_size - pad, 0, pad);
1007 
1008 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
1009 					  agent.rmpp_version;
1010 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
1011 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
1012 
1013 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
1014 					struct ib_rmpp_segment, list);
1015 	send_wr->last_ack_seg = send_wr->cur_seg;
1016 	return 0;
1017 }
1018 
1019 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
1020 {
1021 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
1022 }
1023 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
1024 
1025 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
1026 					    u32 remote_qpn, u16 pkey_index,
1027 					    int rmpp_active,
1028 					    int hdr_len, int data_len,
1029 					    gfp_t gfp_mask,
1030 					    u8 base_version)
1031 {
1032 	struct ib_mad_agent_private *mad_agent_priv;
1033 	struct ib_mad_send_wr_private *mad_send_wr;
1034 	int pad, message_size, ret, size;
1035 	void *buf;
1036 	size_t mad_size;
1037 	bool opa;
1038 
1039 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1040 				      agent);
1041 
1042 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1043 
1044 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
1045 		mad_size = sizeof(struct opa_mad);
1046 	else
1047 		mad_size = sizeof(struct ib_mad);
1048 
1049 	pad = get_pad_size(hdr_len, data_len, mad_size);
1050 	message_size = hdr_len + data_len + pad;
1051 
1052 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1053 		if (!rmpp_active && message_size > mad_size)
1054 			return ERR_PTR(-EINVAL);
1055 	} else
1056 		if (rmpp_active || message_size > mad_size)
1057 			return ERR_PTR(-EINVAL);
1058 
1059 	size = rmpp_active ? hdr_len : mad_size;
1060 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1061 	if (!buf)
1062 		return ERR_PTR(-ENOMEM);
1063 
1064 	mad_send_wr = buf + size;
1065 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1066 	mad_send_wr->send_buf.mad = buf;
1067 	mad_send_wr->send_buf.hdr_len = hdr_len;
1068 	mad_send_wr->send_buf.data_len = data_len;
1069 	mad_send_wr->pad = pad;
1070 
1071 	mad_send_wr->mad_agent_priv = mad_agent_priv;
1072 	mad_send_wr->sg_list[0].length = hdr_len;
1073 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1074 
1075 	/* OPA MADs don't have to be the full 2048 bytes */
1076 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1077 	    data_len < mad_size - hdr_len)
1078 		mad_send_wr->sg_list[1].length = data_len;
1079 	else
1080 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1081 
1082 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1083 
1084 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1085 
1086 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1087 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1088 	mad_send_wr->send_wr.wr.num_sge = 2;
1089 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1090 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1091 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
1092 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1093 	mad_send_wr->send_wr.pkey_index = pkey_index;
1094 
1095 	if (rmpp_active) {
1096 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1097 		if (ret) {
1098 			kfree(buf);
1099 			return ERR_PTR(ret);
1100 		}
1101 	}
1102 
1103 	mad_send_wr->send_buf.mad_agent = mad_agent;
1104 	atomic_inc(&mad_agent_priv->refcount);
1105 	return &mad_send_wr->send_buf;
1106 }
1107 EXPORT_SYMBOL(ib_create_send_mad);
1108 
1109 int ib_get_mad_data_offset(u8 mgmt_class)
1110 {
1111 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1112 		return IB_MGMT_SA_HDR;
1113 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1114 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1115 		 (mgmt_class == IB_MGMT_CLASS_BIS))
1116 		return IB_MGMT_DEVICE_HDR;
1117 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1118 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1119 		return IB_MGMT_VENDOR_HDR;
1120 	else
1121 		return IB_MGMT_MAD_HDR;
1122 }
1123 EXPORT_SYMBOL(ib_get_mad_data_offset);
1124 
1125 int ib_is_mad_class_rmpp(u8 mgmt_class)
1126 {
1127 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1128 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1129 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1130 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
1131 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1132 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1133 		return 1;
1134 	return 0;
1135 }
1136 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1137 
1138 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1139 {
1140 	struct ib_mad_send_wr_private *mad_send_wr;
1141 	struct list_head *list;
1142 
1143 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1144 				   send_buf);
1145 	list = &mad_send_wr->cur_seg->list;
1146 
1147 	if (mad_send_wr->cur_seg->num < seg_num) {
1148 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
1149 			if (mad_send_wr->cur_seg->num == seg_num)
1150 				break;
1151 	} else if (mad_send_wr->cur_seg->num > seg_num) {
1152 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1153 			if (mad_send_wr->cur_seg->num == seg_num)
1154 				break;
1155 	}
1156 	return mad_send_wr->cur_seg->data;
1157 }
1158 EXPORT_SYMBOL(ib_get_rmpp_segment);
1159 
1160 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1161 {
1162 	if (mad_send_wr->send_buf.seg_count)
1163 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1164 					   mad_send_wr->seg_num);
1165 	else
1166 		return mad_send_wr->send_buf.mad +
1167 		       mad_send_wr->send_buf.hdr_len;
1168 }
1169 
1170 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1171 {
1172 	struct ib_mad_agent_private *mad_agent_priv;
1173 	struct ib_mad_send_wr_private *mad_send_wr;
1174 
1175 	mad_agent_priv = container_of(send_buf->mad_agent,
1176 				      struct ib_mad_agent_private, agent);
1177 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1178 				   send_buf);
1179 
1180 	free_send_rmpp_list(mad_send_wr);
1181 	kfree(send_buf->mad);
1182 	deref_mad_agent(mad_agent_priv);
1183 }
1184 EXPORT_SYMBOL(ib_free_send_mad);
1185 
1186 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1187 {
1188 	struct ib_mad_qp_info *qp_info;
1189 	struct list_head *list;
1190 	struct ib_mad_agent *mad_agent;
1191 	struct ib_sge *sge;
1192 	unsigned long flags;
1193 	int ret;
1194 
1195 	/* Set WR ID to find mad_send_wr upon completion */
1196 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1197 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1198 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1199 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1200 
1201 	mad_agent = mad_send_wr->send_buf.mad_agent;
1202 	sge = mad_send_wr->sg_list;
1203 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1204 					mad_send_wr->send_buf.mad,
1205 					sge[0].length,
1206 					DMA_TO_DEVICE);
1207 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1208 		return -ENOMEM;
1209 
1210 	mad_send_wr->header_mapping = sge[0].addr;
1211 
1212 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1213 					ib_get_payload(mad_send_wr),
1214 					sge[1].length,
1215 					DMA_TO_DEVICE);
1216 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1217 		ib_dma_unmap_single(mad_agent->device,
1218 				    mad_send_wr->header_mapping,
1219 				    sge[0].length, DMA_TO_DEVICE);
1220 		return -ENOMEM;
1221 	}
1222 	mad_send_wr->payload_mapping = sge[1].addr;
1223 
1224 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1225 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1226 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1227 				   NULL);
1228 		list = &qp_info->send_queue.list;
1229 	} else {
1230 		ret = 0;
1231 		list = &qp_info->overflow_list;
1232 	}
1233 
1234 	if (!ret) {
1235 		qp_info->send_queue.count++;
1236 		list_add_tail(&mad_send_wr->mad_list.list, list);
1237 	}
1238 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1239 	if (ret) {
1240 		ib_dma_unmap_single(mad_agent->device,
1241 				    mad_send_wr->header_mapping,
1242 				    sge[0].length, DMA_TO_DEVICE);
1243 		ib_dma_unmap_single(mad_agent->device,
1244 				    mad_send_wr->payload_mapping,
1245 				    sge[1].length, DMA_TO_DEVICE);
1246 	}
1247 	return ret;
1248 }
1249 
1250 /*
1251  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1252  *  with the registered client
1253  */
1254 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1255 		     struct ib_mad_send_buf **bad_send_buf)
1256 {
1257 	struct ib_mad_agent_private *mad_agent_priv;
1258 	struct ib_mad_send_buf *next_send_buf;
1259 	struct ib_mad_send_wr_private *mad_send_wr;
1260 	unsigned long flags;
1261 	int ret = -EINVAL;
1262 
1263 	/* Walk list of send WRs and post each on send list */
1264 	for (; send_buf; send_buf = next_send_buf) {
1265 		mad_send_wr = container_of(send_buf,
1266 					   struct ib_mad_send_wr_private,
1267 					   send_buf);
1268 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1269 
1270 		ret = ib_mad_enforce_security(mad_agent_priv,
1271 					      mad_send_wr->send_wr.pkey_index);
1272 		if (ret)
1273 			goto error;
1274 
1275 		if (!send_buf->mad_agent->send_handler ||
1276 		    (send_buf->timeout_ms &&
1277 		     !send_buf->mad_agent->recv_handler)) {
1278 			ret = -EINVAL;
1279 			goto error;
1280 		}
1281 
1282 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1283 			if (mad_agent_priv->agent.rmpp_version) {
1284 				ret = -EINVAL;
1285 				goto error;
1286 			}
1287 		}
1288 
1289 		/*
1290 		 * Save pointer to next work request to post in case the
1291 		 * current one completes, and the user modifies the work
1292 		 * request associated with the completion
1293 		 */
1294 		next_send_buf = send_buf->next;
1295 		mad_send_wr->send_wr.ah = send_buf->ah;
1296 
1297 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1298 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1299 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1300 						     mad_send_wr);
1301 			if (ret < 0)		/* error */
1302 				goto error;
1303 			else if (ret == 1)	/* locally consumed */
1304 				continue;
1305 		}
1306 
1307 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1308 		/* Timeout will be updated after send completes */
1309 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1310 		mad_send_wr->max_retries = send_buf->retries;
1311 		mad_send_wr->retries_left = send_buf->retries;
1312 		send_buf->retries = 0;
1313 		/* Reference for work request to QP + response */
1314 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1315 		mad_send_wr->status = IB_WC_SUCCESS;
1316 
1317 		/* Reference MAD agent until send completes */
1318 		atomic_inc(&mad_agent_priv->refcount);
1319 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1320 		list_add_tail(&mad_send_wr->agent_list,
1321 			      &mad_agent_priv->send_list);
1322 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1323 
1324 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1325 			ret = ib_send_rmpp_mad(mad_send_wr);
1326 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1327 				ret = ib_send_mad(mad_send_wr);
1328 		} else
1329 			ret = ib_send_mad(mad_send_wr);
1330 		if (ret < 0) {
1331 			/* Fail send request */
1332 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1333 			list_del(&mad_send_wr->agent_list);
1334 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1335 			atomic_dec(&mad_agent_priv->refcount);
1336 			goto error;
1337 		}
1338 	}
1339 	return 0;
1340 error:
1341 	if (bad_send_buf)
1342 		*bad_send_buf = send_buf;
1343 	return ret;
1344 }
1345 EXPORT_SYMBOL(ib_post_send_mad);
1346 
1347 /*
1348  * ib_free_recv_mad - Returns data buffers used to receive
1349  *  a MAD to the access layer
1350  */
1351 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1352 {
1353 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1354 	struct ib_mad_private_header *mad_priv_hdr;
1355 	struct ib_mad_private *priv;
1356 	struct list_head free_list;
1357 
1358 	INIT_LIST_HEAD(&free_list);
1359 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1360 
1361 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1362 					&free_list, list) {
1363 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1364 					   recv_buf);
1365 		mad_priv_hdr = container_of(mad_recv_wc,
1366 					    struct ib_mad_private_header,
1367 					    recv_wc);
1368 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1369 				    header);
1370 		kfree(priv);
1371 	}
1372 }
1373 EXPORT_SYMBOL(ib_free_recv_mad);
1374 
1375 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1376 					u8 rmpp_version,
1377 					ib_mad_send_handler send_handler,
1378 					ib_mad_recv_handler recv_handler,
1379 					void *context)
1380 {
1381 	return ERR_PTR(-EINVAL);	/* XXX: for now */
1382 }
1383 EXPORT_SYMBOL(ib_redirect_mad_qp);
1384 
1385 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1386 		      struct ib_wc *wc)
1387 {
1388 	dev_err(&mad_agent->device->dev,
1389 		"ib_process_mad_wc() not implemented yet\n");
1390 	return 0;
1391 }
1392 EXPORT_SYMBOL(ib_process_mad_wc);
1393 
1394 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1395 			 struct ib_mad_reg_req *mad_reg_req)
1396 {
1397 	int i;
1398 
1399 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1400 		if ((*method)->agent[i]) {
1401 			pr_err("Method %d already in use\n", i);
1402 			return -EINVAL;
1403 		}
1404 	}
1405 	return 0;
1406 }
1407 
1408 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1409 {
1410 	/* Allocate management method table */
1411 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1412 	return (*method) ? 0 : (-ENOMEM);
1413 }
1414 
1415 /*
1416  * Check to see if there are any methods still in use
1417  */
1418 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1419 {
1420 	int i;
1421 
1422 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1423 		if (method->agent[i])
1424 			return 1;
1425 	return 0;
1426 }
1427 
1428 /*
1429  * Check to see if there are any method tables for this class still in use
1430  */
1431 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1432 {
1433 	int i;
1434 
1435 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1436 		if (class->method_table[i])
1437 			return 1;
1438 	return 0;
1439 }
1440 
1441 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1442 {
1443 	int i;
1444 
1445 	for (i = 0; i < MAX_MGMT_OUI; i++)
1446 		if (vendor_class->method_table[i])
1447 			return 1;
1448 	return 0;
1449 }
1450 
1451 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1452 			   const char *oui)
1453 {
1454 	int i;
1455 
1456 	for (i = 0; i < MAX_MGMT_OUI; i++)
1457 		/* Is there matching OUI for this vendor class ? */
1458 		if (!memcmp(vendor_class->oui[i], oui, 3))
1459 			return i;
1460 
1461 	return -1;
1462 }
1463 
1464 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1465 {
1466 	int i;
1467 
1468 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1469 		if (vendor->vendor_class[i])
1470 			return 1;
1471 
1472 	return 0;
1473 }
1474 
1475 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1476 				     struct ib_mad_agent_private *agent)
1477 {
1478 	int i;
1479 
1480 	/* Remove any methods for this mad agent */
1481 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1482 		if (method->agent[i] == agent) {
1483 			method->agent[i] = NULL;
1484 		}
1485 	}
1486 }
1487 
1488 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1489 			      struct ib_mad_agent_private *agent_priv,
1490 			      u8 mgmt_class)
1491 {
1492 	struct ib_mad_port_private *port_priv;
1493 	struct ib_mad_mgmt_class_table **class;
1494 	struct ib_mad_mgmt_method_table **method;
1495 	int i, ret;
1496 
1497 	port_priv = agent_priv->qp_info->port_priv;
1498 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1499 	if (!*class) {
1500 		/* Allocate management class table for "new" class version */
1501 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1502 		if (!*class) {
1503 			ret = -ENOMEM;
1504 			goto error1;
1505 		}
1506 
1507 		/* Allocate method table for this management class */
1508 		method = &(*class)->method_table[mgmt_class];
1509 		if ((ret = allocate_method_table(method)))
1510 			goto error2;
1511 	} else {
1512 		method = &(*class)->method_table[mgmt_class];
1513 		if (!*method) {
1514 			/* Allocate method table for this management class */
1515 			if ((ret = allocate_method_table(method)))
1516 				goto error1;
1517 		}
1518 	}
1519 
1520 	/* Now, make sure methods are not already in use */
1521 	if (method_in_use(method, mad_reg_req))
1522 		goto error3;
1523 
1524 	/* Finally, add in methods being registered */
1525 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1526 		(*method)->agent[i] = agent_priv;
1527 
1528 	return 0;
1529 
1530 error3:
1531 	/* Remove any methods for this mad agent */
1532 	remove_methods_mad_agent(*method, agent_priv);
1533 	/* Now, check to see if there are any methods in use */
1534 	if (!check_method_table(*method)) {
1535 		/* If not, release management method table */
1536 		kfree(*method);
1537 		*method = NULL;
1538 	}
1539 	ret = -EINVAL;
1540 	goto error1;
1541 error2:
1542 	kfree(*class);
1543 	*class = NULL;
1544 error1:
1545 	return ret;
1546 }
1547 
1548 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1549 			   struct ib_mad_agent_private *agent_priv)
1550 {
1551 	struct ib_mad_port_private *port_priv;
1552 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1553 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1554 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1555 	struct ib_mad_mgmt_method_table **method;
1556 	int i, ret = -ENOMEM;
1557 	u8 vclass;
1558 
1559 	/* "New" vendor (with OUI) class */
1560 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1561 	port_priv = agent_priv->qp_info->port_priv;
1562 	vendor_table = &port_priv->version[
1563 				mad_reg_req->mgmt_class_version].vendor;
1564 	if (!*vendor_table) {
1565 		/* Allocate mgmt vendor class table for "new" class version */
1566 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1567 		if (!vendor)
1568 			goto error1;
1569 
1570 		*vendor_table = vendor;
1571 	}
1572 	if (!(*vendor_table)->vendor_class[vclass]) {
1573 		/* Allocate table for this management vendor class */
1574 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1575 		if (!vendor_class)
1576 			goto error2;
1577 
1578 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1579 	}
1580 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1581 		/* Is there matching OUI for this vendor class ? */
1582 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1583 			    mad_reg_req->oui, 3)) {
1584 			method = &(*vendor_table)->vendor_class[
1585 						vclass]->method_table[i];
1586 			if (!*method)
1587 				goto error3;
1588 			goto check_in_use;
1589 		}
1590 	}
1591 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1592 		/* OUI slot available ? */
1593 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1594 				vclass]->oui[i])) {
1595 			method = &(*vendor_table)->vendor_class[
1596 				vclass]->method_table[i];
1597 			/* Allocate method table for this OUI */
1598 			if (!*method) {
1599 				ret = allocate_method_table(method);
1600 				if (ret)
1601 					goto error3;
1602 			}
1603 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1604 			       mad_reg_req->oui, 3);
1605 			goto check_in_use;
1606 		}
1607 	}
1608 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1609 	goto error3;
1610 
1611 check_in_use:
1612 	/* Now, make sure methods are not already in use */
1613 	if (method_in_use(method, mad_reg_req))
1614 		goto error4;
1615 
1616 	/* Finally, add in methods being registered */
1617 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1618 		(*method)->agent[i] = agent_priv;
1619 
1620 	return 0;
1621 
1622 error4:
1623 	/* Remove any methods for this mad agent */
1624 	remove_methods_mad_agent(*method, agent_priv);
1625 	/* Now, check to see if there are any methods in use */
1626 	if (!check_method_table(*method)) {
1627 		/* If not, release management method table */
1628 		kfree(*method);
1629 		*method = NULL;
1630 	}
1631 	ret = -EINVAL;
1632 error3:
1633 	if (vendor_class) {
1634 		(*vendor_table)->vendor_class[vclass] = NULL;
1635 		kfree(vendor_class);
1636 	}
1637 error2:
1638 	if (vendor) {
1639 		*vendor_table = NULL;
1640 		kfree(vendor);
1641 	}
1642 error1:
1643 	return ret;
1644 }
1645 
1646 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1647 {
1648 	struct ib_mad_port_private *port_priv;
1649 	struct ib_mad_mgmt_class_table *class;
1650 	struct ib_mad_mgmt_method_table *method;
1651 	struct ib_mad_mgmt_vendor_class_table *vendor;
1652 	struct ib_mad_mgmt_vendor_class *vendor_class;
1653 	int index;
1654 	u8 mgmt_class;
1655 
1656 	/*
1657 	 * Was MAD registration request supplied
1658 	 * with original registration ?
1659 	 */
1660 	if (!agent_priv->reg_req) {
1661 		goto out;
1662 	}
1663 
1664 	port_priv = agent_priv->qp_info->port_priv;
1665 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1666 	class = port_priv->version[
1667 			agent_priv->reg_req->mgmt_class_version].class;
1668 	if (!class)
1669 		goto vendor_check;
1670 
1671 	method = class->method_table[mgmt_class];
1672 	if (method) {
1673 		/* Remove any methods for this mad agent */
1674 		remove_methods_mad_agent(method, agent_priv);
1675 		/* Now, check to see if there are any methods still in use */
1676 		if (!check_method_table(method)) {
1677 			/* If not, release management method table */
1678 			kfree(method);
1679 			class->method_table[mgmt_class] = NULL;
1680 			/* Any management classes left ? */
1681 			if (!check_class_table(class)) {
1682 				/* If not, release management class table */
1683 				kfree(class);
1684 				port_priv->version[
1685 					agent_priv->reg_req->
1686 					mgmt_class_version].class = NULL;
1687 			}
1688 		}
1689 	}
1690 
1691 vendor_check:
1692 	if (!is_vendor_class(mgmt_class))
1693 		goto out;
1694 
1695 	/* normalize mgmt_class to vendor range 2 */
1696 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1697 	vendor = port_priv->version[
1698 			agent_priv->reg_req->mgmt_class_version].vendor;
1699 
1700 	if (!vendor)
1701 		goto out;
1702 
1703 	vendor_class = vendor->vendor_class[mgmt_class];
1704 	if (vendor_class) {
1705 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1706 		if (index < 0)
1707 			goto out;
1708 		method = vendor_class->method_table[index];
1709 		if (method) {
1710 			/* Remove any methods for this mad agent */
1711 			remove_methods_mad_agent(method, agent_priv);
1712 			/*
1713 			 * Now, check to see if there are
1714 			 * any methods still in use
1715 			 */
1716 			if (!check_method_table(method)) {
1717 				/* If not, release management method table */
1718 				kfree(method);
1719 				vendor_class->method_table[index] = NULL;
1720 				memset(vendor_class->oui[index], 0, 3);
1721 				/* Any OUIs left ? */
1722 				if (!check_vendor_class(vendor_class)) {
1723 					/* If not, release vendor class table */
1724 					kfree(vendor_class);
1725 					vendor->vendor_class[mgmt_class] = NULL;
1726 					/* Any other vendor classes left ? */
1727 					if (!check_vendor_table(vendor)) {
1728 						kfree(vendor);
1729 						port_priv->version[
1730 							agent_priv->reg_req->
1731 							mgmt_class_version].
1732 							vendor = NULL;
1733 					}
1734 				}
1735 			}
1736 		}
1737 	}
1738 
1739 out:
1740 	return;
1741 }
1742 
1743 static struct ib_mad_agent_private *
1744 find_mad_agent(struct ib_mad_port_private *port_priv,
1745 	       const struct ib_mad_hdr *mad_hdr)
1746 {
1747 	struct ib_mad_agent_private *mad_agent = NULL;
1748 	unsigned long flags;
1749 
1750 	if (ib_response_mad(mad_hdr)) {
1751 		u32 hi_tid;
1752 
1753 		/*
1754 		 * Routing is based on high 32 bits of transaction ID
1755 		 * of MAD.
1756 		 */
1757 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1758 		rcu_read_lock();
1759 		mad_agent = idr_find(&ib_mad_clients, hi_tid);
1760 		if (mad_agent && !atomic_inc_not_zero(&mad_agent->refcount))
1761 			mad_agent = NULL;
1762 		rcu_read_unlock();
1763 	} else {
1764 		struct ib_mad_mgmt_class_table *class;
1765 		struct ib_mad_mgmt_method_table *method;
1766 		struct ib_mad_mgmt_vendor_class_table *vendor;
1767 		struct ib_mad_mgmt_vendor_class *vendor_class;
1768 		const struct ib_vendor_mad *vendor_mad;
1769 		int index;
1770 
1771 		spin_lock_irqsave(&port_priv->reg_lock, flags);
1772 		/*
1773 		 * Routing is based on version, class, and method
1774 		 * For "newer" vendor MADs, also based on OUI
1775 		 */
1776 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1777 			goto out;
1778 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1779 			class = port_priv->version[
1780 					mad_hdr->class_version].class;
1781 			if (!class)
1782 				goto out;
1783 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1784 			    ARRAY_SIZE(class->method_table))
1785 				goto out;
1786 			method = class->method_table[convert_mgmt_class(
1787 							mad_hdr->mgmt_class)];
1788 			if (method)
1789 				mad_agent = method->agent[mad_hdr->method &
1790 							  ~IB_MGMT_METHOD_RESP];
1791 		} else {
1792 			vendor = port_priv->version[
1793 					mad_hdr->class_version].vendor;
1794 			if (!vendor)
1795 				goto out;
1796 			vendor_class = vendor->vendor_class[vendor_class_index(
1797 						mad_hdr->mgmt_class)];
1798 			if (!vendor_class)
1799 				goto out;
1800 			/* Find matching OUI */
1801 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1802 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1803 			if (index == -1)
1804 				goto out;
1805 			method = vendor_class->method_table[index];
1806 			if (method) {
1807 				mad_agent = method->agent[mad_hdr->method &
1808 							  ~IB_MGMT_METHOD_RESP];
1809 			}
1810 		}
1811 		if (mad_agent)
1812 			atomic_inc(&mad_agent->refcount);
1813 out:
1814 		spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1815 	}
1816 
1817 	if (mad_agent && !mad_agent->agent.recv_handler) {
1818 		dev_notice(&port_priv->device->dev,
1819 			   "No receive handler for client %p on port %d\n",
1820 			   &mad_agent->agent, port_priv->port_num);
1821 		deref_mad_agent(mad_agent);
1822 		mad_agent = NULL;
1823 	}
1824 
1825 	return mad_agent;
1826 }
1827 
1828 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1829 			const struct ib_mad_qp_info *qp_info,
1830 			bool opa)
1831 {
1832 	int valid = 0;
1833 	u32 qp_num = qp_info->qp->qp_num;
1834 
1835 	/* Make sure MAD base version is understood */
1836 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1837 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1838 		pr_err("MAD received with unsupported base version %d %s\n",
1839 		       mad_hdr->base_version, opa ? "(opa)" : "");
1840 		goto out;
1841 	}
1842 
1843 	/* Filter SMI packets sent to other than QP0 */
1844 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1845 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1846 		if (qp_num == 0)
1847 			valid = 1;
1848 	} else {
1849 		/* CM attributes other than ClassPortInfo only use Send method */
1850 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1851 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1852 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1853 			goto out;
1854 		/* Filter GSI packets sent to QP0 */
1855 		if (qp_num != 0)
1856 			valid = 1;
1857 	}
1858 
1859 out:
1860 	return valid;
1861 }
1862 
1863 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1864 			    const struct ib_mad_hdr *mad_hdr)
1865 {
1866 	struct ib_rmpp_mad *rmpp_mad;
1867 
1868 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1869 	return !mad_agent_priv->agent.rmpp_version ||
1870 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1871 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1872 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1873 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1874 }
1875 
1876 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1877 				     const struct ib_mad_recv_wc *rwc)
1878 {
1879 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1880 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1881 }
1882 
1883 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1884 				   const struct ib_mad_send_wr_private *wr,
1885 				   const struct ib_mad_recv_wc *rwc )
1886 {
1887 	struct rdma_ah_attr attr;
1888 	u8 send_resp, rcv_resp;
1889 	union ib_gid sgid;
1890 	struct ib_device *device = mad_agent_priv->agent.device;
1891 	u8 port_num = mad_agent_priv->agent.port_num;
1892 	u8 lmc;
1893 	bool has_grh;
1894 
1895 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1896 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1897 
1898 	if (send_resp == rcv_resp)
1899 		/* both requests, or both responses. GIDs different */
1900 		return 0;
1901 
1902 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1903 		/* Assume not equal, to avoid false positives. */
1904 		return 0;
1905 
1906 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1907 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1908 		/* one has GID, other does not.  Assume different */
1909 		return 0;
1910 
1911 	if (!send_resp && rcv_resp) {
1912 		/* is request/response. */
1913 		if (!has_grh) {
1914 			if (ib_get_cached_lmc(device, port_num, &lmc))
1915 				return 0;
1916 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1917 					   rwc->wc->dlid_path_bits) &
1918 					  ((1 << lmc) - 1)));
1919 		} else {
1920 			const struct ib_global_route *grh =
1921 					rdma_ah_read_grh(&attr);
1922 
1923 			if (rdma_query_gid(device, port_num,
1924 					   grh->sgid_index, &sgid))
1925 				return 0;
1926 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1927 				       16);
1928 		}
1929 	}
1930 
1931 	if (!has_grh)
1932 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1933 	else
1934 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1935 			       rwc->recv_buf.grh->sgid.raw,
1936 			       16);
1937 }
1938 
1939 static inline int is_direct(u8 class)
1940 {
1941 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1942 }
1943 
1944 struct ib_mad_send_wr_private*
1945 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1946 		 const struct ib_mad_recv_wc *wc)
1947 {
1948 	struct ib_mad_send_wr_private *wr;
1949 	const struct ib_mad_hdr *mad_hdr;
1950 
1951 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1952 
1953 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1954 		if ((wr->tid == mad_hdr->tid) &&
1955 		    rcv_has_same_class(wr, wc) &&
1956 		    /*
1957 		     * Don't check GID for direct routed MADs.
1958 		     * These might have permissive LIDs.
1959 		     */
1960 		    (is_direct(mad_hdr->mgmt_class) ||
1961 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1962 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1963 	}
1964 
1965 	/*
1966 	 * It's possible to receive the response before we've
1967 	 * been notified that the send has completed
1968 	 */
1969 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1970 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1971 		    wr->tid == mad_hdr->tid &&
1972 		    wr->timeout &&
1973 		    rcv_has_same_class(wr, wc) &&
1974 		    /*
1975 		     * Don't check GID for direct routed MADs.
1976 		     * These might have permissive LIDs.
1977 		     */
1978 		    (is_direct(mad_hdr->mgmt_class) ||
1979 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1980 			/* Verify request has not been canceled */
1981 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1982 	}
1983 	return NULL;
1984 }
1985 
1986 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1987 {
1988 	mad_send_wr->timeout = 0;
1989 	if (mad_send_wr->refcount == 1)
1990 		list_move_tail(&mad_send_wr->agent_list,
1991 			      &mad_send_wr->mad_agent_priv->done_list);
1992 }
1993 
1994 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1995 				 struct ib_mad_recv_wc *mad_recv_wc)
1996 {
1997 	struct ib_mad_send_wr_private *mad_send_wr;
1998 	struct ib_mad_send_wc mad_send_wc;
1999 	unsigned long flags;
2000 	int ret;
2001 
2002 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
2003 	ret = ib_mad_enforce_security(mad_agent_priv,
2004 				      mad_recv_wc->wc->pkey_index);
2005 	if (ret) {
2006 		ib_free_recv_mad(mad_recv_wc);
2007 		deref_mad_agent(mad_agent_priv);
2008 		return;
2009 	}
2010 
2011 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
2012 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2013 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
2014 						      mad_recv_wc);
2015 		if (!mad_recv_wc) {
2016 			deref_mad_agent(mad_agent_priv);
2017 			return;
2018 		}
2019 	}
2020 
2021 	/* Complete corresponding request */
2022 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
2023 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2024 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2025 		if (!mad_send_wr) {
2026 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2027 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2028 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2029 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2030 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
2031 				/* user rmpp is in effect
2032 				 * and this is an active RMPP MAD
2033 				 */
2034 				mad_agent_priv->agent.recv_handler(
2035 						&mad_agent_priv->agent, NULL,
2036 						mad_recv_wc);
2037 				atomic_dec(&mad_agent_priv->refcount);
2038 			} else {
2039 				/* not user rmpp, revert to normal behavior and
2040 				 * drop the mad */
2041 				ib_free_recv_mad(mad_recv_wc);
2042 				deref_mad_agent(mad_agent_priv);
2043 				return;
2044 			}
2045 		} else {
2046 			ib_mark_mad_done(mad_send_wr);
2047 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2048 
2049 			/* Defined behavior is to complete response before request */
2050 			mad_agent_priv->agent.recv_handler(
2051 					&mad_agent_priv->agent,
2052 					&mad_send_wr->send_buf,
2053 					mad_recv_wc);
2054 			atomic_dec(&mad_agent_priv->refcount);
2055 
2056 			mad_send_wc.status = IB_WC_SUCCESS;
2057 			mad_send_wc.vendor_err = 0;
2058 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
2059 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2060 		}
2061 	} else {
2062 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2063 						   mad_recv_wc);
2064 		deref_mad_agent(mad_agent_priv);
2065 	}
2066 
2067 	return;
2068 }
2069 
2070 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2071 				     const struct ib_mad_qp_info *qp_info,
2072 				     const struct ib_wc *wc,
2073 				     int port_num,
2074 				     struct ib_mad_private *recv,
2075 				     struct ib_mad_private *response)
2076 {
2077 	enum smi_forward_action retsmi;
2078 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
2079 
2080 	if (smi_handle_dr_smp_recv(smp,
2081 				   rdma_cap_ib_switch(port_priv->device),
2082 				   port_num,
2083 				   port_priv->device->phys_port_cnt) ==
2084 				   IB_SMI_DISCARD)
2085 		return IB_SMI_DISCARD;
2086 
2087 	retsmi = smi_check_forward_dr_smp(smp);
2088 	if (retsmi == IB_SMI_LOCAL)
2089 		return IB_SMI_HANDLE;
2090 
2091 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2092 		if (smi_handle_dr_smp_send(smp,
2093 					   rdma_cap_ib_switch(port_priv->device),
2094 					   port_num) == IB_SMI_DISCARD)
2095 			return IB_SMI_DISCARD;
2096 
2097 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2098 			return IB_SMI_DISCARD;
2099 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2100 		/* forward case for switches */
2101 		memcpy(response, recv, mad_priv_size(response));
2102 		response->header.recv_wc.wc = &response->header.wc;
2103 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2104 		response->header.recv_wc.recv_buf.grh = &response->grh;
2105 
2106 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2107 				    &response->grh, wc,
2108 				    port_priv->device,
2109 				    smi_get_fwd_port(smp),
2110 				    qp_info->qp->qp_num,
2111 				    response->mad_size,
2112 				    false);
2113 
2114 		return IB_SMI_DISCARD;
2115 	}
2116 	return IB_SMI_HANDLE;
2117 }
2118 
2119 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2120 				    struct ib_mad_private *response,
2121 				    size_t *resp_len, bool opa)
2122 {
2123 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2124 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2125 
2126 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2127 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
2128 		memcpy(response, recv, mad_priv_size(response));
2129 		response->header.recv_wc.wc = &response->header.wc;
2130 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2131 		response->header.recv_wc.recv_buf.grh = &response->grh;
2132 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2133 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2134 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2135 			resp_hdr->status |= IB_SMP_DIRECTION;
2136 
2137 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2138 			if (recv_hdr->mgmt_class ==
2139 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2140 			    recv_hdr->mgmt_class ==
2141 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2142 				*resp_len = opa_get_smp_header_size(
2143 							(struct opa_smp *)recv->mad);
2144 			else
2145 				*resp_len = sizeof(struct ib_mad_hdr);
2146 		}
2147 
2148 		return true;
2149 	} else {
2150 		return false;
2151 	}
2152 }
2153 
2154 static enum smi_action
2155 handle_opa_smi(struct ib_mad_port_private *port_priv,
2156 	       struct ib_mad_qp_info *qp_info,
2157 	       struct ib_wc *wc,
2158 	       int port_num,
2159 	       struct ib_mad_private *recv,
2160 	       struct ib_mad_private *response)
2161 {
2162 	enum smi_forward_action retsmi;
2163 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
2164 
2165 	if (opa_smi_handle_dr_smp_recv(smp,
2166 				   rdma_cap_ib_switch(port_priv->device),
2167 				   port_num,
2168 				   port_priv->device->phys_port_cnt) ==
2169 				   IB_SMI_DISCARD)
2170 		return IB_SMI_DISCARD;
2171 
2172 	retsmi = opa_smi_check_forward_dr_smp(smp);
2173 	if (retsmi == IB_SMI_LOCAL)
2174 		return IB_SMI_HANDLE;
2175 
2176 	if (retsmi == IB_SMI_SEND) { /* don't forward */
2177 		if (opa_smi_handle_dr_smp_send(smp,
2178 					   rdma_cap_ib_switch(port_priv->device),
2179 					   port_num) == IB_SMI_DISCARD)
2180 			return IB_SMI_DISCARD;
2181 
2182 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
2183 		    IB_SMI_DISCARD)
2184 			return IB_SMI_DISCARD;
2185 
2186 	} else if (rdma_cap_ib_switch(port_priv->device)) {
2187 		/* forward case for switches */
2188 		memcpy(response, recv, mad_priv_size(response));
2189 		response->header.recv_wc.wc = &response->header.wc;
2190 		response->header.recv_wc.recv_buf.opa_mad =
2191 				(struct opa_mad *)response->mad;
2192 		response->header.recv_wc.recv_buf.grh = &response->grh;
2193 
2194 		agent_send_response((const struct ib_mad_hdr *)response->mad,
2195 				    &response->grh, wc,
2196 				    port_priv->device,
2197 				    opa_smi_get_fwd_port(smp),
2198 				    qp_info->qp->qp_num,
2199 				    recv->header.wc.byte_len,
2200 				    true);
2201 
2202 		return IB_SMI_DISCARD;
2203 	}
2204 
2205 	return IB_SMI_HANDLE;
2206 }
2207 
2208 static enum smi_action
2209 handle_smi(struct ib_mad_port_private *port_priv,
2210 	   struct ib_mad_qp_info *qp_info,
2211 	   struct ib_wc *wc,
2212 	   int port_num,
2213 	   struct ib_mad_private *recv,
2214 	   struct ib_mad_private *response,
2215 	   bool opa)
2216 {
2217 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2218 
2219 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2220 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2221 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2222 				      response);
2223 
2224 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2225 }
2226 
2227 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2228 {
2229 	struct ib_mad_port_private *port_priv = cq->cq_context;
2230 	struct ib_mad_list_head *mad_list =
2231 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2232 	struct ib_mad_qp_info *qp_info;
2233 	struct ib_mad_private_header *mad_priv_hdr;
2234 	struct ib_mad_private *recv, *response = NULL;
2235 	struct ib_mad_agent_private *mad_agent;
2236 	int port_num;
2237 	int ret = IB_MAD_RESULT_SUCCESS;
2238 	size_t mad_size;
2239 	u16 resp_mad_pkey_index = 0;
2240 	bool opa;
2241 
2242 	if (list_empty_careful(&port_priv->port_list))
2243 		return;
2244 
2245 	if (wc->status != IB_WC_SUCCESS) {
2246 		/*
2247 		 * Receive errors indicate that the QP has entered the error
2248 		 * state - error handling/shutdown code will cleanup
2249 		 */
2250 		return;
2251 	}
2252 
2253 	qp_info = mad_list->mad_queue->qp_info;
2254 	dequeue_mad(mad_list);
2255 
2256 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2257 			       qp_info->port_priv->port_num);
2258 
2259 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2260 				    mad_list);
2261 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2262 	ib_dma_unmap_single(port_priv->device,
2263 			    recv->header.mapping,
2264 			    mad_priv_dma_size(recv),
2265 			    DMA_FROM_DEVICE);
2266 
2267 	/* Setup MAD receive work completion from "normal" work completion */
2268 	recv->header.wc = *wc;
2269 	recv->header.recv_wc.wc = &recv->header.wc;
2270 
2271 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2272 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2273 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2274 	} else {
2275 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2276 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2277 	}
2278 
2279 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2280 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2281 
2282 	if (atomic_read(&qp_info->snoop_count))
2283 		snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2284 
2285 	/* Validate MAD */
2286 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2287 		goto out;
2288 
2289 	mad_size = recv->mad_size;
2290 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2291 	if (!response)
2292 		goto out;
2293 
2294 	if (rdma_cap_ib_switch(port_priv->device))
2295 		port_num = wc->port_num;
2296 	else
2297 		port_num = port_priv->port_num;
2298 
2299 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2300 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2301 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2302 			       response, opa)
2303 		    == IB_SMI_DISCARD)
2304 			goto out;
2305 	}
2306 
2307 	/* Give driver "right of first refusal" on incoming MAD */
2308 	if (port_priv->device->process_mad) {
2309 		ret = port_priv->device->process_mad(port_priv->device, 0,
2310 						     port_priv->port_num,
2311 						     wc, &recv->grh,
2312 						     (const struct ib_mad_hdr *)recv->mad,
2313 						     recv->mad_size,
2314 						     (struct ib_mad_hdr *)response->mad,
2315 						     &mad_size, &resp_mad_pkey_index);
2316 
2317 		if (opa)
2318 			wc->pkey_index = resp_mad_pkey_index;
2319 
2320 		if (ret & IB_MAD_RESULT_SUCCESS) {
2321 			if (ret & IB_MAD_RESULT_CONSUMED)
2322 				goto out;
2323 			if (ret & IB_MAD_RESULT_REPLY) {
2324 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2325 						    &recv->grh, wc,
2326 						    port_priv->device,
2327 						    port_num,
2328 						    qp_info->qp->qp_num,
2329 						    mad_size, opa);
2330 				goto out;
2331 			}
2332 		}
2333 	}
2334 
2335 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2336 	if (mad_agent) {
2337 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2338 		/*
2339 		 * recv is freed up in error cases in ib_mad_complete_recv
2340 		 * or via recv_handler in ib_mad_complete_recv()
2341 		 */
2342 		recv = NULL;
2343 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2344 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2345 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2346 				    port_priv->device, port_num,
2347 				    qp_info->qp->qp_num, mad_size, opa);
2348 	}
2349 
2350 out:
2351 	/* Post another receive request for this QP */
2352 	if (response) {
2353 		ib_mad_post_receive_mads(qp_info, response);
2354 		kfree(recv);
2355 	} else
2356 		ib_mad_post_receive_mads(qp_info, recv);
2357 }
2358 
2359 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2360 {
2361 	struct ib_mad_send_wr_private *mad_send_wr;
2362 	unsigned long delay;
2363 
2364 	if (list_empty(&mad_agent_priv->wait_list)) {
2365 		cancel_delayed_work(&mad_agent_priv->timed_work);
2366 	} else {
2367 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2368 					 struct ib_mad_send_wr_private,
2369 					 agent_list);
2370 
2371 		if (time_after(mad_agent_priv->timeout,
2372 			       mad_send_wr->timeout)) {
2373 			mad_agent_priv->timeout = mad_send_wr->timeout;
2374 			delay = mad_send_wr->timeout - jiffies;
2375 			if ((long)delay <= 0)
2376 				delay = 1;
2377 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2378 					 &mad_agent_priv->timed_work, delay);
2379 		}
2380 	}
2381 }
2382 
2383 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2384 {
2385 	struct ib_mad_agent_private *mad_agent_priv;
2386 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2387 	struct list_head *list_item;
2388 	unsigned long delay;
2389 
2390 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2391 	list_del(&mad_send_wr->agent_list);
2392 
2393 	delay = mad_send_wr->timeout;
2394 	mad_send_wr->timeout += jiffies;
2395 
2396 	if (delay) {
2397 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2398 			temp_mad_send_wr = list_entry(list_item,
2399 						struct ib_mad_send_wr_private,
2400 						agent_list);
2401 			if (time_after(mad_send_wr->timeout,
2402 				       temp_mad_send_wr->timeout))
2403 				break;
2404 		}
2405 	}
2406 	else
2407 		list_item = &mad_agent_priv->wait_list;
2408 	list_add(&mad_send_wr->agent_list, list_item);
2409 
2410 	/* Reschedule a work item if we have a shorter timeout */
2411 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2412 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2413 				 &mad_agent_priv->timed_work, delay);
2414 }
2415 
2416 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2417 			  unsigned long timeout_ms)
2418 {
2419 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2420 	wait_for_response(mad_send_wr);
2421 }
2422 
2423 /*
2424  * Process a send work completion
2425  */
2426 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2427 			     struct ib_mad_send_wc *mad_send_wc)
2428 {
2429 	struct ib_mad_agent_private	*mad_agent_priv;
2430 	unsigned long			flags;
2431 	int				ret;
2432 
2433 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2434 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2435 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2436 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2437 		if (ret == IB_RMPP_RESULT_CONSUMED)
2438 			goto done;
2439 	} else
2440 		ret = IB_RMPP_RESULT_UNHANDLED;
2441 
2442 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2443 	    mad_send_wr->status == IB_WC_SUCCESS) {
2444 		mad_send_wr->status = mad_send_wc->status;
2445 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2446 	}
2447 
2448 	if (--mad_send_wr->refcount > 0) {
2449 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2450 		    mad_send_wr->status == IB_WC_SUCCESS) {
2451 			wait_for_response(mad_send_wr);
2452 		}
2453 		goto done;
2454 	}
2455 
2456 	/* Remove send from MAD agent and notify client of completion */
2457 	list_del(&mad_send_wr->agent_list);
2458 	adjust_timeout(mad_agent_priv);
2459 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2460 
2461 	if (mad_send_wr->status != IB_WC_SUCCESS )
2462 		mad_send_wc->status = mad_send_wr->status;
2463 	if (ret == IB_RMPP_RESULT_INTERNAL)
2464 		ib_rmpp_send_handler(mad_send_wc);
2465 	else
2466 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2467 						   mad_send_wc);
2468 
2469 	/* Release reference on agent taken when sending */
2470 	deref_mad_agent(mad_agent_priv);
2471 	return;
2472 done:
2473 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2474 }
2475 
2476 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2477 {
2478 	struct ib_mad_port_private *port_priv = cq->cq_context;
2479 	struct ib_mad_list_head *mad_list =
2480 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2481 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2482 	struct ib_mad_qp_info		*qp_info;
2483 	struct ib_mad_queue		*send_queue;
2484 	struct ib_mad_send_wc		mad_send_wc;
2485 	unsigned long flags;
2486 	int ret;
2487 
2488 	if (list_empty_careful(&port_priv->port_list))
2489 		return;
2490 
2491 	if (wc->status != IB_WC_SUCCESS) {
2492 		if (!ib_mad_send_error(port_priv, wc))
2493 			return;
2494 	}
2495 
2496 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2497 				   mad_list);
2498 	send_queue = mad_list->mad_queue;
2499 	qp_info = send_queue->qp_info;
2500 
2501 retry:
2502 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2503 			    mad_send_wr->header_mapping,
2504 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2505 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2506 			    mad_send_wr->payload_mapping,
2507 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2508 	queued_send_wr = NULL;
2509 	spin_lock_irqsave(&send_queue->lock, flags);
2510 	list_del(&mad_list->list);
2511 
2512 	/* Move queued send to the send queue */
2513 	if (send_queue->count-- > send_queue->max_active) {
2514 		mad_list = container_of(qp_info->overflow_list.next,
2515 					struct ib_mad_list_head, list);
2516 		queued_send_wr = container_of(mad_list,
2517 					struct ib_mad_send_wr_private,
2518 					mad_list);
2519 		list_move_tail(&mad_list->list, &send_queue->list);
2520 	}
2521 	spin_unlock_irqrestore(&send_queue->lock, flags);
2522 
2523 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2524 	mad_send_wc.status = wc->status;
2525 	mad_send_wc.vendor_err = wc->vendor_err;
2526 	if (atomic_read(&qp_info->snoop_count))
2527 		snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2528 			   IB_MAD_SNOOP_SEND_COMPLETIONS);
2529 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2530 
2531 	if (queued_send_wr) {
2532 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2533 				   NULL);
2534 		if (ret) {
2535 			dev_err(&port_priv->device->dev,
2536 				"ib_post_send failed: %d\n", ret);
2537 			mad_send_wr = queued_send_wr;
2538 			wc->status = IB_WC_LOC_QP_OP_ERR;
2539 			goto retry;
2540 		}
2541 	}
2542 }
2543 
2544 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2545 {
2546 	struct ib_mad_send_wr_private *mad_send_wr;
2547 	struct ib_mad_list_head *mad_list;
2548 	unsigned long flags;
2549 
2550 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2551 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2552 		mad_send_wr = container_of(mad_list,
2553 					   struct ib_mad_send_wr_private,
2554 					   mad_list);
2555 		mad_send_wr->retry = 1;
2556 	}
2557 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2558 }
2559 
2560 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2561 		struct ib_wc *wc)
2562 {
2563 	struct ib_mad_list_head *mad_list =
2564 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2565 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2566 	struct ib_mad_send_wr_private *mad_send_wr;
2567 	int ret;
2568 
2569 	/*
2570 	 * Send errors will transition the QP to SQE - move
2571 	 * QP to RTS and repost flushed work requests
2572 	 */
2573 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2574 				   mad_list);
2575 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2576 		if (mad_send_wr->retry) {
2577 			/* Repost send */
2578 			mad_send_wr->retry = 0;
2579 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2580 					   NULL);
2581 			if (!ret)
2582 				return false;
2583 		}
2584 	} else {
2585 		struct ib_qp_attr *attr;
2586 
2587 		/* Transition QP to RTS and fail offending send */
2588 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2589 		if (attr) {
2590 			attr->qp_state = IB_QPS_RTS;
2591 			attr->cur_qp_state = IB_QPS_SQE;
2592 			ret = ib_modify_qp(qp_info->qp, attr,
2593 					   IB_QP_STATE | IB_QP_CUR_STATE);
2594 			kfree(attr);
2595 			if (ret)
2596 				dev_err(&port_priv->device->dev,
2597 					"%s - ib_modify_qp to RTS: %d\n",
2598 					__func__, ret);
2599 			else
2600 				mark_sends_for_retry(qp_info);
2601 		}
2602 	}
2603 
2604 	return true;
2605 }
2606 
2607 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2608 {
2609 	unsigned long flags;
2610 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2611 	struct ib_mad_send_wc mad_send_wc;
2612 	struct list_head cancel_list;
2613 
2614 	INIT_LIST_HEAD(&cancel_list);
2615 
2616 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2617 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2618 				 &mad_agent_priv->send_list, agent_list) {
2619 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2620 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2621 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2622 		}
2623 	}
2624 
2625 	/* Empty wait list to prevent receives from finding a request */
2626 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2627 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2628 
2629 	/* Report all cancelled requests */
2630 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2631 	mad_send_wc.vendor_err = 0;
2632 
2633 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2634 				 &cancel_list, agent_list) {
2635 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2636 		list_del(&mad_send_wr->agent_list);
2637 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2638 						   &mad_send_wc);
2639 		atomic_dec(&mad_agent_priv->refcount);
2640 	}
2641 }
2642 
2643 static struct ib_mad_send_wr_private*
2644 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2645 	     struct ib_mad_send_buf *send_buf)
2646 {
2647 	struct ib_mad_send_wr_private *mad_send_wr;
2648 
2649 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2650 			    agent_list) {
2651 		if (&mad_send_wr->send_buf == send_buf)
2652 			return mad_send_wr;
2653 	}
2654 
2655 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2656 			    agent_list) {
2657 		if (is_rmpp_data_mad(mad_agent_priv,
2658 				     mad_send_wr->send_buf.mad) &&
2659 		    &mad_send_wr->send_buf == send_buf)
2660 			return mad_send_wr;
2661 	}
2662 	return NULL;
2663 }
2664 
2665 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2666 		  struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2667 {
2668 	struct ib_mad_agent_private *mad_agent_priv;
2669 	struct ib_mad_send_wr_private *mad_send_wr;
2670 	unsigned long flags;
2671 	int active;
2672 
2673 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2674 				      agent);
2675 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2676 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2677 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2678 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2679 		return -EINVAL;
2680 	}
2681 
2682 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2683 	if (!timeout_ms) {
2684 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2685 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2686 	}
2687 
2688 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2689 	if (active)
2690 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2691 	else
2692 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2693 
2694 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2695 	return 0;
2696 }
2697 EXPORT_SYMBOL(ib_modify_mad);
2698 
2699 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2700 		   struct ib_mad_send_buf *send_buf)
2701 {
2702 	ib_modify_mad(mad_agent, send_buf, 0);
2703 }
2704 EXPORT_SYMBOL(ib_cancel_mad);
2705 
2706 static void local_completions(struct work_struct *work)
2707 {
2708 	struct ib_mad_agent_private *mad_agent_priv;
2709 	struct ib_mad_local_private *local;
2710 	struct ib_mad_agent_private *recv_mad_agent;
2711 	unsigned long flags;
2712 	int free_mad;
2713 	struct ib_wc wc;
2714 	struct ib_mad_send_wc mad_send_wc;
2715 	bool opa;
2716 
2717 	mad_agent_priv =
2718 		container_of(work, struct ib_mad_agent_private, local_work);
2719 
2720 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2721 			       mad_agent_priv->qp_info->port_priv->port_num);
2722 
2723 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2724 	while (!list_empty(&mad_agent_priv->local_list)) {
2725 		local = list_entry(mad_agent_priv->local_list.next,
2726 				   struct ib_mad_local_private,
2727 				   completion_list);
2728 		list_del(&local->completion_list);
2729 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2730 		free_mad = 0;
2731 		if (local->mad_priv) {
2732 			u8 base_version;
2733 			recv_mad_agent = local->recv_mad_agent;
2734 			if (!recv_mad_agent) {
2735 				dev_err(&mad_agent_priv->agent.device->dev,
2736 					"No receive MAD agent for local completion\n");
2737 				free_mad = 1;
2738 				goto local_send_completion;
2739 			}
2740 
2741 			/*
2742 			 * Defined behavior is to complete response
2743 			 * before request
2744 			 */
2745 			build_smp_wc(recv_mad_agent->agent.qp,
2746 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2747 				     be16_to_cpu(IB_LID_PERMISSIVE),
2748 				     local->mad_send_wr->send_wr.pkey_index,
2749 				     recv_mad_agent->agent.port_num, &wc);
2750 
2751 			local->mad_priv->header.recv_wc.wc = &wc;
2752 
2753 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2754 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2755 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2756 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2757 			} else {
2758 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2759 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2760 			}
2761 
2762 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2763 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2764 				 &local->mad_priv->header.recv_wc.rmpp_list);
2765 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2766 			local->mad_priv->header.recv_wc.recv_buf.mad =
2767 						(struct ib_mad *)local->mad_priv->mad;
2768 			if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2769 				snoop_recv(recv_mad_agent->qp_info,
2770 					  &local->mad_priv->header.recv_wc,
2771 					   IB_MAD_SNOOP_RECVS);
2772 			recv_mad_agent->agent.recv_handler(
2773 						&recv_mad_agent->agent,
2774 						&local->mad_send_wr->send_buf,
2775 						&local->mad_priv->header.recv_wc);
2776 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2777 			atomic_dec(&recv_mad_agent->refcount);
2778 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2779 		}
2780 
2781 local_send_completion:
2782 		/* Complete send */
2783 		mad_send_wc.status = IB_WC_SUCCESS;
2784 		mad_send_wc.vendor_err = 0;
2785 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2786 		if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2787 			snoop_send(mad_agent_priv->qp_info,
2788 				   &local->mad_send_wr->send_buf,
2789 				   &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2790 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2791 						   &mad_send_wc);
2792 
2793 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2794 		atomic_dec(&mad_agent_priv->refcount);
2795 		if (free_mad)
2796 			kfree(local->mad_priv);
2797 		kfree(local);
2798 	}
2799 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2800 }
2801 
2802 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2803 {
2804 	int ret;
2805 
2806 	if (!mad_send_wr->retries_left)
2807 		return -ETIMEDOUT;
2808 
2809 	mad_send_wr->retries_left--;
2810 	mad_send_wr->send_buf.retries++;
2811 
2812 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2813 
2814 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2815 		ret = ib_retry_rmpp(mad_send_wr);
2816 		switch (ret) {
2817 		case IB_RMPP_RESULT_UNHANDLED:
2818 			ret = ib_send_mad(mad_send_wr);
2819 			break;
2820 		case IB_RMPP_RESULT_CONSUMED:
2821 			ret = 0;
2822 			break;
2823 		default:
2824 			ret = -ECOMM;
2825 			break;
2826 		}
2827 	} else
2828 		ret = ib_send_mad(mad_send_wr);
2829 
2830 	if (!ret) {
2831 		mad_send_wr->refcount++;
2832 		list_add_tail(&mad_send_wr->agent_list,
2833 			      &mad_send_wr->mad_agent_priv->send_list);
2834 	}
2835 	return ret;
2836 }
2837 
2838 static void timeout_sends(struct work_struct *work)
2839 {
2840 	struct ib_mad_agent_private *mad_agent_priv;
2841 	struct ib_mad_send_wr_private *mad_send_wr;
2842 	struct ib_mad_send_wc mad_send_wc;
2843 	unsigned long flags, delay;
2844 
2845 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2846 				      timed_work.work);
2847 	mad_send_wc.vendor_err = 0;
2848 
2849 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2850 	while (!list_empty(&mad_agent_priv->wait_list)) {
2851 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2852 					 struct ib_mad_send_wr_private,
2853 					 agent_list);
2854 
2855 		if (time_after(mad_send_wr->timeout, jiffies)) {
2856 			delay = mad_send_wr->timeout - jiffies;
2857 			if ((long)delay <= 0)
2858 				delay = 1;
2859 			queue_delayed_work(mad_agent_priv->qp_info->
2860 					   port_priv->wq,
2861 					   &mad_agent_priv->timed_work, delay);
2862 			break;
2863 		}
2864 
2865 		list_del(&mad_send_wr->agent_list);
2866 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2867 		    !retry_send(mad_send_wr))
2868 			continue;
2869 
2870 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2871 
2872 		if (mad_send_wr->status == IB_WC_SUCCESS)
2873 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2874 		else
2875 			mad_send_wc.status = mad_send_wr->status;
2876 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2877 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2878 						   &mad_send_wc);
2879 
2880 		atomic_dec(&mad_agent_priv->refcount);
2881 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2882 	}
2883 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2884 }
2885 
2886 /*
2887  * Allocate receive MADs and post receive WRs for them
2888  */
2889 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2890 				    struct ib_mad_private *mad)
2891 {
2892 	unsigned long flags;
2893 	int post, ret;
2894 	struct ib_mad_private *mad_priv;
2895 	struct ib_sge sg_list;
2896 	struct ib_recv_wr recv_wr;
2897 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2898 
2899 	/* Initialize common scatter list fields */
2900 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2901 
2902 	/* Initialize common receive WR fields */
2903 	recv_wr.next = NULL;
2904 	recv_wr.sg_list = &sg_list;
2905 	recv_wr.num_sge = 1;
2906 
2907 	do {
2908 		/* Allocate and map receive buffer */
2909 		if (mad) {
2910 			mad_priv = mad;
2911 			mad = NULL;
2912 		} else {
2913 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2914 						     GFP_ATOMIC);
2915 			if (!mad_priv) {
2916 				ret = -ENOMEM;
2917 				break;
2918 			}
2919 		}
2920 		sg_list.length = mad_priv_dma_size(mad_priv);
2921 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2922 						 &mad_priv->grh,
2923 						 mad_priv_dma_size(mad_priv),
2924 						 DMA_FROM_DEVICE);
2925 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2926 						  sg_list.addr))) {
2927 			ret = -ENOMEM;
2928 			break;
2929 		}
2930 		mad_priv->header.mapping = sg_list.addr;
2931 		mad_priv->header.mad_list.mad_queue = recv_queue;
2932 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2933 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2934 
2935 		/* Post receive WR */
2936 		spin_lock_irqsave(&recv_queue->lock, flags);
2937 		post = (++recv_queue->count < recv_queue->max_active);
2938 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2939 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2940 		ret = ib_post_recv(qp_info->qp, &recv_wr, NULL);
2941 		if (ret) {
2942 			spin_lock_irqsave(&recv_queue->lock, flags);
2943 			list_del(&mad_priv->header.mad_list.list);
2944 			recv_queue->count--;
2945 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2946 			ib_dma_unmap_single(qp_info->port_priv->device,
2947 					    mad_priv->header.mapping,
2948 					    mad_priv_dma_size(mad_priv),
2949 					    DMA_FROM_DEVICE);
2950 			kfree(mad_priv);
2951 			dev_err(&qp_info->port_priv->device->dev,
2952 				"ib_post_recv failed: %d\n", ret);
2953 			break;
2954 		}
2955 	} while (post);
2956 
2957 	return ret;
2958 }
2959 
2960 /*
2961  * Return all the posted receive MADs
2962  */
2963 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2964 {
2965 	struct ib_mad_private_header *mad_priv_hdr;
2966 	struct ib_mad_private *recv;
2967 	struct ib_mad_list_head *mad_list;
2968 
2969 	if (!qp_info->qp)
2970 		return;
2971 
2972 	while (!list_empty(&qp_info->recv_queue.list)) {
2973 
2974 		mad_list = list_entry(qp_info->recv_queue.list.next,
2975 				      struct ib_mad_list_head, list);
2976 		mad_priv_hdr = container_of(mad_list,
2977 					    struct ib_mad_private_header,
2978 					    mad_list);
2979 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2980 				    header);
2981 
2982 		/* Remove from posted receive MAD list */
2983 		list_del(&mad_list->list);
2984 
2985 		ib_dma_unmap_single(qp_info->port_priv->device,
2986 				    recv->header.mapping,
2987 				    mad_priv_dma_size(recv),
2988 				    DMA_FROM_DEVICE);
2989 		kfree(recv);
2990 	}
2991 
2992 	qp_info->recv_queue.count = 0;
2993 }
2994 
2995 /*
2996  * Start the port
2997  */
2998 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2999 {
3000 	int ret, i;
3001 	struct ib_qp_attr *attr;
3002 	struct ib_qp *qp;
3003 	u16 pkey_index;
3004 
3005 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
3006 	if (!attr)
3007 		return -ENOMEM;
3008 
3009 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
3010 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
3011 	if (ret)
3012 		pkey_index = 0;
3013 
3014 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3015 		qp = port_priv->qp_info[i].qp;
3016 		if (!qp)
3017 			continue;
3018 
3019 		/*
3020 		 * PKey index for QP1 is irrelevant but
3021 		 * one is needed for the Reset to Init transition
3022 		 */
3023 		attr->qp_state = IB_QPS_INIT;
3024 		attr->pkey_index = pkey_index;
3025 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3026 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3027 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
3028 		if (ret) {
3029 			dev_err(&port_priv->device->dev,
3030 				"Couldn't change QP%d state to INIT: %d\n",
3031 				i, ret);
3032 			goto out;
3033 		}
3034 
3035 		attr->qp_state = IB_QPS_RTR;
3036 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3037 		if (ret) {
3038 			dev_err(&port_priv->device->dev,
3039 				"Couldn't change QP%d state to RTR: %d\n",
3040 				i, ret);
3041 			goto out;
3042 		}
3043 
3044 		attr->qp_state = IB_QPS_RTS;
3045 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
3046 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3047 		if (ret) {
3048 			dev_err(&port_priv->device->dev,
3049 				"Couldn't change QP%d state to RTS: %d\n",
3050 				i, ret);
3051 			goto out;
3052 		}
3053 	}
3054 
3055 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3056 	if (ret) {
3057 		dev_err(&port_priv->device->dev,
3058 			"Failed to request completion notification: %d\n",
3059 			ret);
3060 		goto out;
3061 	}
3062 
3063 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3064 		if (!port_priv->qp_info[i].qp)
3065 			continue;
3066 
3067 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3068 		if (ret) {
3069 			dev_err(&port_priv->device->dev,
3070 				"Couldn't post receive WRs\n");
3071 			goto out;
3072 		}
3073 	}
3074 out:
3075 	kfree(attr);
3076 	return ret;
3077 }
3078 
3079 static void qp_event_handler(struct ib_event *event, void *qp_context)
3080 {
3081 	struct ib_mad_qp_info	*qp_info = qp_context;
3082 
3083 	/* It's worse than that! He's dead, Jim! */
3084 	dev_err(&qp_info->port_priv->device->dev,
3085 		"Fatal error (%d) on MAD QP (%d)\n",
3086 		event->event, qp_info->qp->qp_num);
3087 }
3088 
3089 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3090 			   struct ib_mad_queue *mad_queue)
3091 {
3092 	mad_queue->qp_info = qp_info;
3093 	mad_queue->count = 0;
3094 	spin_lock_init(&mad_queue->lock);
3095 	INIT_LIST_HEAD(&mad_queue->list);
3096 }
3097 
3098 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3099 			struct ib_mad_qp_info *qp_info)
3100 {
3101 	qp_info->port_priv = port_priv;
3102 	init_mad_queue(qp_info, &qp_info->send_queue);
3103 	init_mad_queue(qp_info, &qp_info->recv_queue);
3104 	INIT_LIST_HEAD(&qp_info->overflow_list);
3105 	spin_lock_init(&qp_info->snoop_lock);
3106 	qp_info->snoop_table = NULL;
3107 	qp_info->snoop_table_size = 0;
3108 	atomic_set(&qp_info->snoop_count, 0);
3109 }
3110 
3111 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3112 			 enum ib_qp_type qp_type)
3113 {
3114 	struct ib_qp_init_attr	qp_init_attr;
3115 	int ret;
3116 
3117 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
3118 	qp_init_attr.send_cq = qp_info->port_priv->cq;
3119 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
3120 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3121 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
3122 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3123 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3124 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3125 	qp_init_attr.qp_type = qp_type;
3126 	qp_init_attr.port_num = qp_info->port_priv->port_num;
3127 	qp_init_attr.qp_context = qp_info;
3128 	qp_init_attr.event_handler = qp_event_handler;
3129 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3130 	if (IS_ERR(qp_info->qp)) {
3131 		dev_err(&qp_info->port_priv->device->dev,
3132 			"Couldn't create ib_mad QP%d\n",
3133 			get_spl_qp_index(qp_type));
3134 		ret = PTR_ERR(qp_info->qp);
3135 		goto error;
3136 	}
3137 	/* Use minimum queue sizes unless the CQ is resized */
3138 	qp_info->send_queue.max_active = mad_sendq_size;
3139 	qp_info->recv_queue.max_active = mad_recvq_size;
3140 	return 0;
3141 
3142 error:
3143 	return ret;
3144 }
3145 
3146 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3147 {
3148 	if (!qp_info->qp)
3149 		return;
3150 
3151 	ib_destroy_qp(qp_info->qp);
3152 	kfree(qp_info->snoop_table);
3153 }
3154 
3155 /*
3156  * Open the port
3157  * Create the QP, PD, MR, and CQ if needed
3158  */
3159 static int ib_mad_port_open(struct ib_device *device,
3160 			    int port_num)
3161 {
3162 	int ret, cq_size;
3163 	struct ib_mad_port_private *port_priv;
3164 	unsigned long flags;
3165 	char name[sizeof "ib_mad123"];
3166 	int has_smi;
3167 
3168 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3169 		return -EFAULT;
3170 
3171 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3172 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3173 		return -EFAULT;
3174 
3175 	/* Create new device info */
3176 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3177 	if (!port_priv)
3178 		return -ENOMEM;
3179 
3180 	port_priv->device = device;
3181 	port_priv->port_num = port_num;
3182 	spin_lock_init(&port_priv->reg_lock);
3183 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
3184 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
3185 
3186 	cq_size = mad_sendq_size + mad_recvq_size;
3187 	has_smi = rdma_cap_ib_smi(device, port_num);
3188 	if (has_smi)
3189 		cq_size *= 2;
3190 
3191 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3192 			IB_POLL_UNBOUND_WORKQUEUE);
3193 	if (IS_ERR(port_priv->cq)) {
3194 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3195 		ret = PTR_ERR(port_priv->cq);
3196 		goto error3;
3197 	}
3198 
3199 	port_priv->pd = ib_alloc_pd(device, 0);
3200 	if (IS_ERR(port_priv->pd)) {
3201 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3202 		ret = PTR_ERR(port_priv->pd);
3203 		goto error4;
3204 	}
3205 
3206 	if (has_smi) {
3207 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3208 		if (ret)
3209 			goto error6;
3210 	}
3211 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3212 	if (ret)
3213 		goto error7;
3214 
3215 	snprintf(name, sizeof name, "ib_mad%d", port_num);
3216 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3217 	if (!port_priv->wq) {
3218 		ret = -ENOMEM;
3219 		goto error8;
3220 	}
3221 
3222 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3223 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3224 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3225 
3226 	ret = ib_mad_port_start(port_priv);
3227 	if (ret) {
3228 		dev_err(&device->dev, "Couldn't start port\n");
3229 		goto error9;
3230 	}
3231 
3232 	return 0;
3233 
3234 error9:
3235 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3236 	list_del_init(&port_priv->port_list);
3237 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3238 
3239 	destroy_workqueue(port_priv->wq);
3240 error8:
3241 	destroy_mad_qp(&port_priv->qp_info[1]);
3242 error7:
3243 	destroy_mad_qp(&port_priv->qp_info[0]);
3244 error6:
3245 	ib_dealloc_pd(port_priv->pd);
3246 error4:
3247 	ib_free_cq(port_priv->cq);
3248 	cleanup_recv_queue(&port_priv->qp_info[1]);
3249 	cleanup_recv_queue(&port_priv->qp_info[0]);
3250 error3:
3251 	kfree(port_priv);
3252 
3253 	return ret;
3254 }
3255 
3256 /*
3257  * Close the port
3258  * If there are no classes using the port, free the port
3259  * resources (CQ, MR, PD, QP) and remove the port's info structure
3260  */
3261 static int ib_mad_port_close(struct ib_device *device, int port_num)
3262 {
3263 	struct ib_mad_port_private *port_priv;
3264 	unsigned long flags;
3265 
3266 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3267 	port_priv = __ib_get_mad_port(device, port_num);
3268 	if (port_priv == NULL) {
3269 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3270 		dev_err(&device->dev, "Port %d not found\n", port_num);
3271 		return -ENODEV;
3272 	}
3273 	list_del_init(&port_priv->port_list);
3274 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3275 
3276 	destroy_workqueue(port_priv->wq);
3277 	destroy_mad_qp(&port_priv->qp_info[1]);
3278 	destroy_mad_qp(&port_priv->qp_info[0]);
3279 	ib_dealloc_pd(port_priv->pd);
3280 	ib_free_cq(port_priv->cq);
3281 	cleanup_recv_queue(&port_priv->qp_info[1]);
3282 	cleanup_recv_queue(&port_priv->qp_info[0]);
3283 	/* XXX: Handle deallocation of MAD registration tables */
3284 
3285 	kfree(port_priv);
3286 
3287 	return 0;
3288 }
3289 
3290 static void ib_mad_init_device(struct ib_device *device)
3291 {
3292 	int start, i;
3293 
3294 	start = rdma_start_port(device);
3295 
3296 	for (i = start; i <= rdma_end_port(device); i++) {
3297 		if (!rdma_cap_ib_mad(device, i))
3298 			continue;
3299 
3300 		if (ib_mad_port_open(device, i)) {
3301 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3302 			goto error;
3303 		}
3304 		if (ib_agent_port_open(device, i)) {
3305 			dev_err(&device->dev,
3306 				"Couldn't open port %d for agents\n", i);
3307 			goto error_agent;
3308 		}
3309 	}
3310 	return;
3311 
3312 error_agent:
3313 	if (ib_mad_port_close(device, i))
3314 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3315 
3316 error:
3317 	while (--i >= start) {
3318 		if (!rdma_cap_ib_mad(device, i))
3319 			continue;
3320 
3321 		if (ib_agent_port_close(device, i))
3322 			dev_err(&device->dev,
3323 				"Couldn't close port %d for agents\n", i);
3324 		if (ib_mad_port_close(device, i))
3325 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3326 	}
3327 }
3328 
3329 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3330 {
3331 	int i;
3332 
3333 	for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3334 		if (!rdma_cap_ib_mad(device, i))
3335 			continue;
3336 
3337 		if (ib_agent_port_close(device, i))
3338 			dev_err(&device->dev,
3339 				"Couldn't close port %d for agents\n", i);
3340 		if (ib_mad_port_close(device, i))
3341 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3342 	}
3343 }
3344 
3345 static struct ib_client mad_client = {
3346 	.name   = "mad",
3347 	.add = ib_mad_init_device,
3348 	.remove = ib_mad_remove_device
3349 };
3350 
3351 int ib_mad_init(void)
3352 {
3353 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3354 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3355 
3356 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3357 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3358 
3359 	INIT_LIST_HEAD(&ib_mad_port_list);
3360 
3361 	/* Client ID 0 is used for snoop-only clients */
3362 	idr_alloc(&ib_mad_clients, NULL, 0, 0, GFP_KERNEL);
3363 
3364 	if (ib_register_client(&mad_client)) {
3365 		pr_err("Couldn't register ib_mad client\n");
3366 		return -EINVAL;
3367 	}
3368 
3369 	return 0;
3370 }
3371 
3372 void ib_mad_cleanup(void)
3373 {
3374 	ib_unregister_client(&mad_client);
3375 }
3376