xref: /openbmc/linux/drivers/infiniband/core/mad.c (revision c4c3c32d)
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) 2014,2018 Intel Corporation.  All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  */
37 
38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
39 
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/security.h>
44 #include <linux/xarray.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 #define CREATE_TRACE_POINTS
55 #include <trace/events/ib_mad.h>
56 
57 #ifdef CONFIG_TRACEPOINTS
58 static void create_mad_addr_info(struct ib_mad_send_wr_private *mad_send_wr,
59 			  struct ib_mad_qp_info *qp_info,
60 			  struct trace_event_raw_ib_mad_send_template *entry)
61 {
62 	struct ib_ud_wr *wr = &mad_send_wr->send_wr;
63 	struct rdma_ah_attr attr = {};
64 
65 	rdma_query_ah(wr->ah, &attr);
66 
67 	/* These are common */
68 	entry->sl = attr.sl;
69 	entry->rqpn = wr->remote_qpn;
70 	entry->rqkey = wr->remote_qkey;
71 	entry->dlid = rdma_ah_get_dlid(&attr);
72 }
73 #endif
74 
75 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
76 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
77 
78 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
79 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
80 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
81 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
82 
83 static DEFINE_XARRAY_ALLOC1(ib_mad_clients);
84 static u32 ib_mad_client_next;
85 static struct list_head ib_mad_port_list;
86 
87 /* Port list lock */
88 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
89 
90 /* Forward declarations */
91 static int method_in_use(struct ib_mad_mgmt_method_table **method,
92 			 struct ib_mad_reg_req *mad_reg_req);
93 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
94 static struct ib_mad_agent_private *find_mad_agent(
95 					struct ib_mad_port_private *port_priv,
96 					const struct ib_mad_hdr *mad);
97 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
98 				    struct ib_mad_private *mad);
99 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
100 static void timeout_sends(struct work_struct *work);
101 static void local_completions(struct work_struct *work);
102 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
103 			      struct ib_mad_agent_private *agent_priv,
104 			      u8 mgmt_class);
105 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
106 			   struct ib_mad_agent_private *agent_priv);
107 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
108 			      struct ib_wc *wc);
109 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
110 
111 /*
112  * Returns a ib_mad_port_private structure or NULL for a device/port
113  * Assumes ib_mad_port_list_lock is being held
114  */
115 static inline struct ib_mad_port_private *
116 __ib_get_mad_port(struct ib_device *device, u32 port_num)
117 {
118 	struct ib_mad_port_private *entry;
119 
120 	list_for_each_entry(entry, &ib_mad_port_list, port_list) {
121 		if (entry->device == device && entry->port_num == port_num)
122 			return entry;
123 	}
124 	return NULL;
125 }
126 
127 /*
128  * Wrapper function to return a ib_mad_port_private structure or NULL
129  * for a device/port
130  */
131 static inline struct ib_mad_port_private *
132 ib_get_mad_port(struct ib_device *device, u32 port_num)
133 {
134 	struct ib_mad_port_private *entry;
135 	unsigned long flags;
136 
137 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
138 	entry = __ib_get_mad_port(device, port_num);
139 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
140 
141 	return entry;
142 }
143 
144 static inline u8 convert_mgmt_class(u8 mgmt_class)
145 {
146 	/* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
147 	return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
148 		0 : mgmt_class;
149 }
150 
151 static int get_spl_qp_index(enum ib_qp_type qp_type)
152 {
153 	switch (qp_type) {
154 	case IB_QPT_SMI:
155 		return 0;
156 	case IB_QPT_GSI:
157 		return 1;
158 	default:
159 		return -1;
160 	}
161 }
162 
163 static int vendor_class_index(u8 mgmt_class)
164 {
165 	return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
166 }
167 
168 static int is_vendor_class(u8 mgmt_class)
169 {
170 	if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
171 	    (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
172 		return 0;
173 	return 1;
174 }
175 
176 static int is_vendor_oui(char *oui)
177 {
178 	if (oui[0] || oui[1] || oui[2])
179 		return 1;
180 	return 0;
181 }
182 
183 static int is_vendor_method_in_use(
184 		struct ib_mad_mgmt_vendor_class *vendor_class,
185 		struct ib_mad_reg_req *mad_reg_req)
186 {
187 	struct ib_mad_mgmt_method_table *method;
188 	int i;
189 
190 	for (i = 0; i < MAX_MGMT_OUI; i++) {
191 		if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
192 			method = vendor_class->method_table[i];
193 			if (method) {
194 				if (method_in_use(&method, mad_reg_req))
195 					return 1;
196 				else
197 					break;
198 			}
199 		}
200 	}
201 	return 0;
202 }
203 
204 int ib_response_mad(const struct ib_mad_hdr *hdr)
205 {
206 	return ((hdr->method & IB_MGMT_METHOD_RESP) ||
207 		(hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
208 		((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
209 		 (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
210 }
211 EXPORT_SYMBOL(ib_response_mad);
212 
213 /*
214  * ib_register_mad_agent - Register to send/receive MADs
215  *
216  * Context: Process context.
217  */
218 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
219 					   u32 port_num,
220 					   enum ib_qp_type qp_type,
221 					   struct ib_mad_reg_req *mad_reg_req,
222 					   u8 rmpp_version,
223 					   ib_mad_send_handler send_handler,
224 					   ib_mad_recv_handler recv_handler,
225 					   void *context,
226 					   u32 registration_flags)
227 {
228 	struct ib_mad_port_private *port_priv;
229 	struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
230 	struct ib_mad_agent_private *mad_agent_priv;
231 	struct ib_mad_reg_req *reg_req = NULL;
232 	struct ib_mad_mgmt_class_table *class;
233 	struct ib_mad_mgmt_vendor_class_table *vendor;
234 	struct ib_mad_mgmt_vendor_class *vendor_class;
235 	struct ib_mad_mgmt_method_table *method;
236 	int ret2, qpn;
237 	u8 mgmt_class, vclass;
238 
239 	if ((qp_type == IB_QPT_SMI && !rdma_cap_ib_smi(device, port_num)) ||
240 	    (qp_type == IB_QPT_GSI && !rdma_cap_ib_cm(device, port_num)))
241 		return ERR_PTR(-EPROTONOSUPPORT);
242 
243 	/* Validate parameters */
244 	qpn = get_spl_qp_index(qp_type);
245 	if (qpn == -1) {
246 		dev_dbg_ratelimited(&device->dev, "%s: invalid QP Type %d\n",
247 				    __func__, qp_type);
248 		goto error1;
249 	}
250 
251 	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
252 		dev_dbg_ratelimited(&device->dev,
253 				    "%s: invalid RMPP Version %u\n",
254 				    __func__, rmpp_version);
255 		goto error1;
256 	}
257 
258 	/* Validate MAD registration request if supplied */
259 	if (mad_reg_req) {
260 		if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
261 			dev_dbg_ratelimited(&device->dev,
262 					    "%s: invalid Class Version %u\n",
263 					    __func__,
264 					    mad_reg_req->mgmt_class_version);
265 			goto error1;
266 		}
267 		if (!recv_handler) {
268 			dev_dbg_ratelimited(&device->dev,
269 					    "%s: no recv_handler\n", __func__);
270 			goto error1;
271 		}
272 		if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
273 			/*
274 			 * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
275 			 * one in this range currently allowed
276 			 */
277 			if (mad_reg_req->mgmt_class !=
278 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
279 				dev_dbg_ratelimited(&device->dev,
280 					"%s: Invalid Mgmt Class 0x%x\n",
281 					__func__, mad_reg_req->mgmt_class);
282 				goto error1;
283 			}
284 		} else if (mad_reg_req->mgmt_class == 0) {
285 			/*
286 			 * Class 0 is reserved in IBA and is used for
287 			 * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
288 			 */
289 			dev_dbg_ratelimited(&device->dev,
290 					    "%s: Invalid Mgmt Class 0\n",
291 					    __func__);
292 			goto error1;
293 		} else if (is_vendor_class(mad_reg_req->mgmt_class)) {
294 			/*
295 			 * If class is in "new" vendor range,
296 			 * ensure supplied OUI is not zero
297 			 */
298 			if (!is_vendor_oui(mad_reg_req->oui)) {
299 				dev_dbg_ratelimited(&device->dev,
300 					"%s: No OUI specified for class 0x%x\n",
301 					__func__,
302 					mad_reg_req->mgmt_class);
303 				goto error1;
304 			}
305 		}
306 		/* Make sure class supplied is consistent with RMPP */
307 		if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
308 			if (rmpp_version) {
309 				dev_dbg_ratelimited(&device->dev,
310 					"%s: RMPP version for non-RMPP class 0x%x\n",
311 					__func__, mad_reg_req->mgmt_class);
312 				goto error1;
313 			}
314 		}
315 
316 		/* Make sure class supplied is consistent with QP type */
317 		if (qp_type == IB_QPT_SMI) {
318 			if ((mad_reg_req->mgmt_class !=
319 					IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
320 			    (mad_reg_req->mgmt_class !=
321 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
322 				dev_dbg_ratelimited(&device->dev,
323 					"%s: Invalid SM QP type: class 0x%x\n",
324 					__func__, mad_reg_req->mgmt_class);
325 				goto error1;
326 			}
327 		} else {
328 			if ((mad_reg_req->mgmt_class ==
329 					IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
330 			    (mad_reg_req->mgmt_class ==
331 					IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
332 				dev_dbg_ratelimited(&device->dev,
333 					"%s: Invalid GS QP type: class 0x%x\n",
334 					__func__, mad_reg_req->mgmt_class);
335 				goto error1;
336 			}
337 		}
338 	} else {
339 		/* No registration request supplied */
340 		if (!send_handler)
341 			goto error1;
342 		if (registration_flags & IB_MAD_USER_RMPP)
343 			goto error1;
344 	}
345 
346 	/* Validate device and port */
347 	port_priv = ib_get_mad_port(device, port_num);
348 	if (!port_priv) {
349 		dev_dbg_ratelimited(&device->dev, "%s: Invalid port %u\n",
350 				    __func__, port_num);
351 		ret = ERR_PTR(-ENODEV);
352 		goto error1;
353 	}
354 
355 	/* Verify the QP requested is supported. For example, Ethernet devices
356 	 * will not have QP0.
357 	 */
358 	if (!port_priv->qp_info[qpn].qp) {
359 		dev_dbg_ratelimited(&device->dev, "%s: QP %d not supported\n",
360 				    __func__, qpn);
361 		ret = ERR_PTR(-EPROTONOSUPPORT);
362 		goto error1;
363 	}
364 
365 	/* Allocate structures */
366 	mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
367 	if (!mad_agent_priv) {
368 		ret = ERR_PTR(-ENOMEM);
369 		goto error1;
370 	}
371 
372 	if (mad_reg_req) {
373 		reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
374 		if (!reg_req) {
375 			ret = ERR_PTR(-ENOMEM);
376 			goto error3;
377 		}
378 	}
379 
380 	/* Now, fill in the various structures */
381 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
382 	mad_agent_priv->reg_req = reg_req;
383 	mad_agent_priv->agent.rmpp_version = rmpp_version;
384 	mad_agent_priv->agent.device = device;
385 	mad_agent_priv->agent.recv_handler = recv_handler;
386 	mad_agent_priv->agent.send_handler = send_handler;
387 	mad_agent_priv->agent.context = context;
388 	mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
389 	mad_agent_priv->agent.port_num = port_num;
390 	mad_agent_priv->agent.flags = registration_flags;
391 	spin_lock_init(&mad_agent_priv->lock);
392 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
393 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
394 	INIT_LIST_HEAD(&mad_agent_priv->done_list);
395 	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
396 	INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
397 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
398 	INIT_WORK(&mad_agent_priv->local_work, local_completions);
399 	refcount_set(&mad_agent_priv->refcount, 1);
400 	init_completion(&mad_agent_priv->comp);
401 
402 	ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
403 	if (ret2) {
404 		ret = ERR_PTR(ret2);
405 		goto error4;
406 	}
407 
408 	/*
409 	 * The mlx4 driver uses the top byte to distinguish which virtual
410 	 * function generated the MAD, so we must avoid using it.
411 	 */
412 	ret2 = xa_alloc_cyclic(&ib_mad_clients, &mad_agent_priv->agent.hi_tid,
413 			mad_agent_priv, XA_LIMIT(0, (1 << 24) - 1),
414 			&ib_mad_client_next, GFP_KERNEL);
415 	if (ret2 < 0) {
416 		ret = ERR_PTR(ret2);
417 		goto error5;
418 	}
419 
420 	/*
421 	 * Make sure MAD registration (if supplied)
422 	 * is non overlapping with any existing ones
423 	 */
424 	spin_lock_irq(&port_priv->reg_lock);
425 	if (mad_reg_req) {
426 		mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
427 		if (!is_vendor_class(mgmt_class)) {
428 			class = port_priv->version[mad_reg_req->
429 						   mgmt_class_version].class;
430 			if (class) {
431 				method = class->method_table[mgmt_class];
432 				if (method) {
433 					if (method_in_use(&method,
434 							   mad_reg_req))
435 						goto error6;
436 				}
437 			}
438 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
439 						  mgmt_class);
440 		} else {
441 			/* "New" vendor class range */
442 			vendor = port_priv->version[mad_reg_req->
443 						    mgmt_class_version].vendor;
444 			if (vendor) {
445 				vclass = vendor_class_index(mgmt_class);
446 				vendor_class = vendor->vendor_class[vclass];
447 				if (vendor_class) {
448 					if (is_vendor_method_in_use(
449 							vendor_class,
450 							mad_reg_req))
451 						goto error6;
452 				}
453 			}
454 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
455 		}
456 		if (ret2) {
457 			ret = ERR_PTR(ret2);
458 			goto error6;
459 		}
460 	}
461 	spin_unlock_irq(&port_priv->reg_lock);
462 
463 	trace_ib_mad_create_agent(mad_agent_priv);
464 	return &mad_agent_priv->agent;
465 error6:
466 	spin_unlock_irq(&port_priv->reg_lock);
467 	xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
468 error5:
469 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
470 error4:
471 	kfree(reg_req);
472 error3:
473 	kfree(mad_agent_priv);
474 error1:
475 	return ret;
476 }
477 EXPORT_SYMBOL(ib_register_mad_agent);
478 
479 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
480 {
481 	if (refcount_dec_and_test(&mad_agent_priv->refcount))
482 		complete(&mad_agent_priv->comp);
483 }
484 
485 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
486 {
487 	struct ib_mad_port_private *port_priv;
488 
489 	/* Note that we could still be handling received MADs */
490 	trace_ib_mad_unregister_agent(mad_agent_priv);
491 
492 	/*
493 	 * Canceling all sends results in dropping received response
494 	 * MADs, preventing us from queuing additional work
495 	 */
496 	cancel_mads(mad_agent_priv);
497 	port_priv = mad_agent_priv->qp_info->port_priv;
498 	cancel_delayed_work(&mad_agent_priv->timed_work);
499 
500 	spin_lock_irq(&port_priv->reg_lock);
501 	remove_mad_reg_req(mad_agent_priv);
502 	spin_unlock_irq(&port_priv->reg_lock);
503 	xa_erase(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
504 
505 	flush_workqueue(port_priv->wq);
506 
507 	deref_mad_agent(mad_agent_priv);
508 	wait_for_completion(&mad_agent_priv->comp);
509 	ib_cancel_rmpp_recvs(mad_agent_priv);
510 
511 	ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
512 
513 	kfree(mad_agent_priv->reg_req);
514 	kfree_rcu(mad_agent_priv, rcu);
515 }
516 
517 /*
518  * ib_unregister_mad_agent - Unregisters a client from using MAD services
519  *
520  * Context: Process context.
521  */
522 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
523 {
524 	struct ib_mad_agent_private *mad_agent_priv;
525 
526 	mad_agent_priv = container_of(mad_agent,
527 				      struct ib_mad_agent_private,
528 				      agent);
529 	unregister_mad_agent(mad_agent_priv);
530 }
531 EXPORT_SYMBOL(ib_unregister_mad_agent);
532 
533 static void dequeue_mad(struct ib_mad_list_head *mad_list)
534 {
535 	struct ib_mad_queue *mad_queue;
536 	unsigned long flags;
537 
538 	mad_queue = mad_list->mad_queue;
539 	spin_lock_irqsave(&mad_queue->lock, flags);
540 	list_del(&mad_list->list);
541 	mad_queue->count--;
542 	spin_unlock_irqrestore(&mad_queue->lock, flags);
543 }
544 
545 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
546 		u16 pkey_index, u32 port_num, struct ib_wc *wc)
547 {
548 	memset(wc, 0, sizeof *wc);
549 	wc->wr_cqe = cqe;
550 	wc->status = IB_WC_SUCCESS;
551 	wc->opcode = IB_WC_RECV;
552 	wc->pkey_index = pkey_index;
553 	wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
554 	wc->src_qp = IB_QP0;
555 	wc->qp = qp;
556 	wc->slid = slid;
557 	wc->sl = 0;
558 	wc->dlid_path_bits = 0;
559 	wc->port_num = port_num;
560 }
561 
562 static size_t mad_priv_size(const struct ib_mad_private *mp)
563 {
564 	return sizeof(struct ib_mad_private) + mp->mad_size;
565 }
566 
567 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
568 {
569 	size_t size = sizeof(struct ib_mad_private) + mad_size;
570 	struct ib_mad_private *ret = kzalloc(size, flags);
571 
572 	if (ret)
573 		ret->mad_size = mad_size;
574 
575 	return ret;
576 }
577 
578 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
579 {
580 	return rdma_max_mad_size(port_priv->device, port_priv->port_num);
581 }
582 
583 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
584 {
585 	return sizeof(struct ib_grh) + mp->mad_size;
586 }
587 
588 /*
589  * Return 0 if SMP is to be sent
590  * Return 1 if SMP was consumed locally (whether or not solicited)
591  * Return < 0 if error
592  */
593 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
594 				  struct ib_mad_send_wr_private *mad_send_wr)
595 {
596 	int ret = 0;
597 	struct ib_smp *smp = mad_send_wr->send_buf.mad;
598 	struct opa_smp *opa_smp = (struct opa_smp *)smp;
599 	unsigned long flags;
600 	struct ib_mad_local_private *local;
601 	struct ib_mad_private *mad_priv;
602 	struct ib_mad_port_private *port_priv;
603 	struct ib_mad_agent_private *recv_mad_agent = NULL;
604 	struct ib_device *device = mad_agent_priv->agent.device;
605 	u32 port_num;
606 	struct ib_wc mad_wc;
607 	struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
608 	size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
609 	u16 out_mad_pkey_index = 0;
610 	u16 drslid;
611 	bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
612 				    mad_agent_priv->qp_info->port_priv->port_num);
613 
614 	if (rdma_cap_ib_switch(device) &&
615 	    smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
616 		port_num = send_wr->port_num;
617 	else
618 		port_num = mad_agent_priv->agent.port_num;
619 
620 	/*
621 	 * Directed route handling starts if the initial LID routed part of
622 	 * a request or the ending LID routed part of a response is empty.
623 	 * If we are at the start of the LID routed part, don't update the
624 	 * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
625 	 */
626 	if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
627 		u32 opa_drslid;
628 
629 		trace_ib_mad_handle_out_opa_smi(opa_smp);
630 
631 		if ((opa_get_smp_direction(opa_smp)
632 		     ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
633 		     OPA_LID_PERMISSIVE &&
634 		     opa_smi_handle_dr_smp_send(opa_smp,
635 						rdma_cap_ib_switch(device),
636 						port_num) == IB_SMI_DISCARD) {
637 			ret = -EINVAL;
638 			dev_err(&device->dev, "OPA Invalid directed route\n");
639 			goto out;
640 		}
641 		opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
642 		if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
643 		    opa_drslid & 0xffff0000) {
644 			ret = -EINVAL;
645 			dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
646 			       opa_drslid);
647 			goto out;
648 		}
649 		drslid = (u16)(opa_drslid & 0x0000ffff);
650 
651 		/* Check to post send on QP or process locally */
652 		if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
653 		    opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
654 			goto out;
655 	} else {
656 		trace_ib_mad_handle_out_ib_smi(smp);
657 
658 		if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
659 		     IB_LID_PERMISSIVE &&
660 		     smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
661 		     IB_SMI_DISCARD) {
662 			ret = -EINVAL;
663 			dev_err(&device->dev, "Invalid directed route\n");
664 			goto out;
665 		}
666 		drslid = be16_to_cpu(smp->dr_slid);
667 
668 		/* Check to post send on QP or process locally */
669 		if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
670 		    smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
671 			goto out;
672 	}
673 
674 	local = kmalloc(sizeof *local, GFP_ATOMIC);
675 	if (!local) {
676 		ret = -ENOMEM;
677 		goto out;
678 	}
679 	local->mad_priv = NULL;
680 	local->recv_mad_agent = NULL;
681 	mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
682 	if (!mad_priv) {
683 		ret = -ENOMEM;
684 		kfree(local);
685 		goto out;
686 	}
687 
688 	build_smp_wc(mad_agent_priv->agent.qp,
689 		     send_wr->wr.wr_cqe, drslid,
690 		     send_wr->pkey_index,
691 		     send_wr->port_num, &mad_wc);
692 
693 	if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
694 		mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
695 					+ mad_send_wr->send_buf.data_len
696 					+ sizeof(struct ib_grh);
697 	}
698 
699 	/* No GRH for DR SMP */
700 	ret = device->ops.process_mad(device, 0, port_num, &mad_wc, NULL,
701 				      (const struct ib_mad *)smp,
702 				      (struct ib_mad *)mad_priv->mad, &mad_size,
703 				      &out_mad_pkey_index);
704 	switch (ret) {
705 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
706 		if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
707 		    mad_agent_priv->agent.recv_handler) {
708 			local->mad_priv = mad_priv;
709 			local->recv_mad_agent = mad_agent_priv;
710 			/*
711 			 * Reference MAD agent until receive
712 			 * side of local completion handled
713 			 */
714 			refcount_inc(&mad_agent_priv->refcount);
715 		} else
716 			kfree(mad_priv);
717 		break;
718 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
719 		kfree(mad_priv);
720 		break;
721 	case IB_MAD_RESULT_SUCCESS:
722 		/* Treat like an incoming receive MAD */
723 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
724 					    mad_agent_priv->agent.port_num);
725 		if (port_priv) {
726 			memcpy(mad_priv->mad, smp, mad_priv->mad_size);
727 			recv_mad_agent = find_mad_agent(port_priv,
728 						        (const struct ib_mad_hdr *)mad_priv->mad);
729 		}
730 		if (!port_priv || !recv_mad_agent) {
731 			/*
732 			 * No receiving agent so drop packet and
733 			 * generate send completion.
734 			 */
735 			kfree(mad_priv);
736 			break;
737 		}
738 		local->mad_priv = mad_priv;
739 		local->recv_mad_agent = recv_mad_agent;
740 		break;
741 	default:
742 		kfree(mad_priv);
743 		kfree(local);
744 		ret = -EINVAL;
745 		goto out;
746 	}
747 
748 	local->mad_send_wr = mad_send_wr;
749 	if (opa) {
750 		local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
751 		local->return_wc_byte_len = mad_size;
752 	}
753 	/* Reference MAD agent until send side of local completion handled */
754 	refcount_inc(&mad_agent_priv->refcount);
755 	/* Queue local completion to local list */
756 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
757 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
758 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
759 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
760 		   &mad_agent_priv->local_work);
761 	ret = 1;
762 out:
763 	return ret;
764 }
765 
766 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
767 {
768 	int seg_size, pad;
769 
770 	seg_size = mad_size - hdr_len;
771 	if (data_len && seg_size) {
772 		pad = seg_size - data_len % seg_size;
773 		return pad == seg_size ? 0 : pad;
774 	} else
775 		return seg_size;
776 }
777 
778 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
779 {
780 	struct ib_rmpp_segment *s, *t;
781 
782 	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
783 		list_del(&s->list);
784 		kfree(s);
785 	}
786 }
787 
788 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
789 				size_t mad_size, gfp_t gfp_mask)
790 {
791 	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
792 	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
793 	struct ib_rmpp_segment *seg = NULL;
794 	int left, seg_size, pad;
795 
796 	send_buf->seg_size = mad_size - send_buf->hdr_len;
797 	send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
798 	seg_size = send_buf->seg_size;
799 	pad = send_wr->pad;
800 
801 	/* Allocate data segments. */
802 	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
803 		seg = kmalloc(sizeof(*seg) + seg_size, gfp_mask);
804 		if (!seg) {
805 			free_send_rmpp_list(send_wr);
806 			return -ENOMEM;
807 		}
808 		seg->num = ++send_buf->seg_count;
809 		list_add_tail(&seg->list, &send_wr->rmpp_list);
810 	}
811 
812 	/* Zero any padding */
813 	if (pad)
814 		memset(seg->data + seg_size - pad, 0, pad);
815 
816 	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
817 					  agent.rmpp_version;
818 	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
819 	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
820 
821 	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
822 					struct ib_rmpp_segment, list);
823 	send_wr->last_ack_seg = send_wr->cur_seg;
824 	return 0;
825 }
826 
827 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
828 {
829 	return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
830 }
831 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
832 
833 struct ib_mad_send_buf *ib_create_send_mad(struct ib_mad_agent *mad_agent,
834 					   u32 remote_qpn, u16 pkey_index,
835 					   int rmpp_active, int hdr_len,
836 					   int data_len, gfp_t gfp_mask,
837 					   u8 base_version)
838 {
839 	struct ib_mad_agent_private *mad_agent_priv;
840 	struct ib_mad_send_wr_private *mad_send_wr;
841 	int pad, message_size, ret, size;
842 	void *buf;
843 	size_t mad_size;
844 	bool opa;
845 
846 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
847 				      agent);
848 
849 	opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
850 
851 	if (opa && base_version == OPA_MGMT_BASE_VERSION)
852 		mad_size = sizeof(struct opa_mad);
853 	else
854 		mad_size = sizeof(struct ib_mad);
855 
856 	pad = get_pad_size(hdr_len, data_len, mad_size);
857 	message_size = hdr_len + data_len + pad;
858 
859 	if (ib_mad_kernel_rmpp_agent(mad_agent)) {
860 		if (!rmpp_active && message_size > mad_size)
861 			return ERR_PTR(-EINVAL);
862 	} else
863 		if (rmpp_active || message_size > mad_size)
864 			return ERR_PTR(-EINVAL);
865 
866 	size = rmpp_active ? hdr_len : mad_size;
867 	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
868 	if (!buf)
869 		return ERR_PTR(-ENOMEM);
870 
871 	mad_send_wr = buf + size;
872 	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
873 	mad_send_wr->send_buf.mad = buf;
874 	mad_send_wr->send_buf.hdr_len = hdr_len;
875 	mad_send_wr->send_buf.data_len = data_len;
876 	mad_send_wr->pad = pad;
877 
878 	mad_send_wr->mad_agent_priv = mad_agent_priv;
879 	mad_send_wr->sg_list[0].length = hdr_len;
880 	mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
881 
882 	/* OPA MADs don't have to be the full 2048 bytes */
883 	if (opa && base_version == OPA_MGMT_BASE_VERSION &&
884 	    data_len < mad_size - hdr_len)
885 		mad_send_wr->sg_list[1].length = data_len;
886 	else
887 		mad_send_wr->sg_list[1].length = mad_size - hdr_len;
888 
889 	mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
890 
891 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
892 
893 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
894 	mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
895 	mad_send_wr->send_wr.wr.num_sge = 2;
896 	mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
897 	mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
898 	mad_send_wr->send_wr.remote_qpn = remote_qpn;
899 	mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
900 	mad_send_wr->send_wr.pkey_index = pkey_index;
901 
902 	if (rmpp_active) {
903 		ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
904 		if (ret) {
905 			kfree(buf);
906 			return ERR_PTR(ret);
907 		}
908 	}
909 
910 	mad_send_wr->send_buf.mad_agent = mad_agent;
911 	refcount_inc(&mad_agent_priv->refcount);
912 	return &mad_send_wr->send_buf;
913 }
914 EXPORT_SYMBOL(ib_create_send_mad);
915 
916 int ib_get_mad_data_offset(u8 mgmt_class)
917 {
918 	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
919 		return IB_MGMT_SA_HDR;
920 	else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
921 		 (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
922 		 (mgmt_class == IB_MGMT_CLASS_BIS))
923 		return IB_MGMT_DEVICE_HDR;
924 	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
925 		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
926 		return IB_MGMT_VENDOR_HDR;
927 	else
928 		return IB_MGMT_MAD_HDR;
929 }
930 EXPORT_SYMBOL(ib_get_mad_data_offset);
931 
932 int ib_is_mad_class_rmpp(u8 mgmt_class)
933 {
934 	if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
935 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
936 	    (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
937 	    (mgmt_class == IB_MGMT_CLASS_BIS) ||
938 	    ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
939 	     (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
940 		return 1;
941 	return 0;
942 }
943 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
944 
945 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
946 {
947 	struct ib_mad_send_wr_private *mad_send_wr;
948 	struct list_head *list;
949 
950 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
951 				   send_buf);
952 	list = &mad_send_wr->cur_seg->list;
953 
954 	if (mad_send_wr->cur_seg->num < seg_num) {
955 		list_for_each_entry(mad_send_wr->cur_seg, list, list)
956 			if (mad_send_wr->cur_seg->num == seg_num)
957 				break;
958 	} else if (mad_send_wr->cur_seg->num > seg_num) {
959 		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
960 			if (mad_send_wr->cur_seg->num == seg_num)
961 				break;
962 	}
963 	return mad_send_wr->cur_seg->data;
964 }
965 EXPORT_SYMBOL(ib_get_rmpp_segment);
966 
967 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
968 {
969 	if (mad_send_wr->send_buf.seg_count)
970 		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
971 					   mad_send_wr->seg_num);
972 	else
973 		return mad_send_wr->send_buf.mad +
974 		       mad_send_wr->send_buf.hdr_len;
975 }
976 
977 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
978 {
979 	struct ib_mad_agent_private *mad_agent_priv;
980 	struct ib_mad_send_wr_private *mad_send_wr;
981 
982 	mad_agent_priv = container_of(send_buf->mad_agent,
983 				      struct ib_mad_agent_private, agent);
984 	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
985 				   send_buf);
986 
987 	free_send_rmpp_list(mad_send_wr);
988 	kfree(send_buf->mad);
989 	deref_mad_agent(mad_agent_priv);
990 }
991 EXPORT_SYMBOL(ib_free_send_mad);
992 
993 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
994 {
995 	struct ib_mad_qp_info *qp_info;
996 	struct list_head *list;
997 	struct ib_mad_agent *mad_agent;
998 	struct ib_sge *sge;
999 	unsigned long flags;
1000 	int ret;
1001 
1002 	/* Set WR ID to find mad_send_wr upon completion */
1003 	qp_info = mad_send_wr->mad_agent_priv->qp_info;
1004 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1005 	mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1006 	mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1007 
1008 	mad_agent = mad_send_wr->send_buf.mad_agent;
1009 	sge = mad_send_wr->sg_list;
1010 	sge[0].addr = ib_dma_map_single(mad_agent->device,
1011 					mad_send_wr->send_buf.mad,
1012 					sge[0].length,
1013 					DMA_TO_DEVICE);
1014 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1015 		return -ENOMEM;
1016 
1017 	mad_send_wr->header_mapping = sge[0].addr;
1018 
1019 	sge[1].addr = ib_dma_map_single(mad_agent->device,
1020 					ib_get_payload(mad_send_wr),
1021 					sge[1].length,
1022 					DMA_TO_DEVICE);
1023 	if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1024 		ib_dma_unmap_single(mad_agent->device,
1025 				    mad_send_wr->header_mapping,
1026 				    sge[0].length, DMA_TO_DEVICE);
1027 		return -ENOMEM;
1028 	}
1029 	mad_send_wr->payload_mapping = sge[1].addr;
1030 
1031 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1032 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1033 		trace_ib_mad_ib_send_mad(mad_send_wr, qp_info);
1034 		ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1035 				   NULL);
1036 		list = &qp_info->send_queue.list;
1037 	} else {
1038 		ret = 0;
1039 		list = &qp_info->overflow_list;
1040 	}
1041 
1042 	if (!ret) {
1043 		qp_info->send_queue.count++;
1044 		list_add_tail(&mad_send_wr->mad_list.list, list);
1045 	}
1046 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1047 	if (ret) {
1048 		ib_dma_unmap_single(mad_agent->device,
1049 				    mad_send_wr->header_mapping,
1050 				    sge[0].length, DMA_TO_DEVICE);
1051 		ib_dma_unmap_single(mad_agent->device,
1052 				    mad_send_wr->payload_mapping,
1053 				    sge[1].length, DMA_TO_DEVICE);
1054 	}
1055 	return ret;
1056 }
1057 
1058 /*
1059  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1060  *  with the registered client
1061  */
1062 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1063 		     struct ib_mad_send_buf **bad_send_buf)
1064 {
1065 	struct ib_mad_agent_private *mad_agent_priv;
1066 	struct ib_mad_send_buf *next_send_buf;
1067 	struct ib_mad_send_wr_private *mad_send_wr;
1068 	unsigned long flags;
1069 	int ret = -EINVAL;
1070 
1071 	/* Walk list of send WRs and post each on send list */
1072 	for (; send_buf; send_buf = next_send_buf) {
1073 		mad_send_wr = container_of(send_buf,
1074 					   struct ib_mad_send_wr_private,
1075 					   send_buf);
1076 		mad_agent_priv = mad_send_wr->mad_agent_priv;
1077 
1078 		ret = ib_mad_enforce_security(mad_agent_priv,
1079 					      mad_send_wr->send_wr.pkey_index);
1080 		if (ret)
1081 			goto error;
1082 
1083 		if (!send_buf->mad_agent->send_handler ||
1084 		    (send_buf->timeout_ms &&
1085 		     !send_buf->mad_agent->recv_handler)) {
1086 			ret = -EINVAL;
1087 			goto error;
1088 		}
1089 
1090 		if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1091 			if (mad_agent_priv->agent.rmpp_version) {
1092 				ret = -EINVAL;
1093 				goto error;
1094 			}
1095 		}
1096 
1097 		/*
1098 		 * Save pointer to next work request to post in case the
1099 		 * current one completes, and the user modifies the work
1100 		 * request associated with the completion
1101 		 */
1102 		next_send_buf = send_buf->next;
1103 		mad_send_wr->send_wr.ah = send_buf->ah;
1104 
1105 		if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1106 		    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1107 			ret = handle_outgoing_dr_smp(mad_agent_priv,
1108 						     mad_send_wr);
1109 			if (ret < 0)		/* error */
1110 				goto error;
1111 			else if (ret == 1)	/* locally consumed */
1112 				continue;
1113 		}
1114 
1115 		mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1116 		/* Timeout will be updated after send completes */
1117 		mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1118 		mad_send_wr->max_retries = send_buf->retries;
1119 		mad_send_wr->retries_left = send_buf->retries;
1120 		send_buf->retries = 0;
1121 		/* Reference for work request to QP + response */
1122 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1123 		mad_send_wr->status = IB_WC_SUCCESS;
1124 
1125 		/* Reference MAD agent until send completes */
1126 		refcount_inc(&mad_agent_priv->refcount);
1127 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1128 		list_add_tail(&mad_send_wr->agent_list,
1129 			      &mad_agent_priv->send_list);
1130 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1131 
1132 		if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1133 			ret = ib_send_rmpp_mad(mad_send_wr);
1134 			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1135 				ret = ib_send_mad(mad_send_wr);
1136 		} else
1137 			ret = ib_send_mad(mad_send_wr);
1138 		if (ret < 0) {
1139 			/* Fail send request */
1140 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
1141 			list_del(&mad_send_wr->agent_list);
1142 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1143 			deref_mad_agent(mad_agent_priv);
1144 			goto error;
1145 		}
1146 	}
1147 	return 0;
1148 error:
1149 	if (bad_send_buf)
1150 		*bad_send_buf = send_buf;
1151 	return ret;
1152 }
1153 EXPORT_SYMBOL(ib_post_send_mad);
1154 
1155 /*
1156  * ib_free_recv_mad - Returns data buffers used to receive
1157  *  a MAD to the access layer
1158  */
1159 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1160 {
1161 	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1162 	struct ib_mad_private_header *mad_priv_hdr;
1163 	struct ib_mad_private *priv;
1164 	struct list_head free_list;
1165 
1166 	INIT_LIST_HEAD(&free_list);
1167 	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1168 
1169 	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1170 					&free_list, list) {
1171 		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1172 					   recv_buf);
1173 		mad_priv_hdr = container_of(mad_recv_wc,
1174 					    struct ib_mad_private_header,
1175 					    recv_wc);
1176 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
1177 				    header);
1178 		kfree(priv);
1179 	}
1180 }
1181 EXPORT_SYMBOL(ib_free_recv_mad);
1182 
1183 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1184 			 struct ib_mad_reg_req *mad_reg_req)
1185 {
1186 	int i;
1187 
1188 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1189 		if ((*method)->agent[i]) {
1190 			pr_err("Method %d already in use\n", i);
1191 			return -EINVAL;
1192 		}
1193 	}
1194 	return 0;
1195 }
1196 
1197 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1198 {
1199 	/* Allocate management method table */
1200 	*method = kzalloc(sizeof **method, GFP_ATOMIC);
1201 	return (*method) ? 0 : (-ENOMEM);
1202 }
1203 
1204 /*
1205  * Check to see if there are any methods still in use
1206  */
1207 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1208 {
1209 	int i;
1210 
1211 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1212 		if (method->agent[i])
1213 			return 1;
1214 	return 0;
1215 }
1216 
1217 /*
1218  * Check to see if there are any method tables for this class still in use
1219  */
1220 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1221 {
1222 	int i;
1223 
1224 	for (i = 0; i < MAX_MGMT_CLASS; i++)
1225 		if (class->method_table[i])
1226 			return 1;
1227 	return 0;
1228 }
1229 
1230 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1231 {
1232 	int i;
1233 
1234 	for (i = 0; i < MAX_MGMT_OUI; i++)
1235 		if (vendor_class->method_table[i])
1236 			return 1;
1237 	return 0;
1238 }
1239 
1240 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1241 			   const char *oui)
1242 {
1243 	int i;
1244 
1245 	for (i = 0; i < MAX_MGMT_OUI; i++)
1246 		/* Is there matching OUI for this vendor class ? */
1247 		if (!memcmp(vendor_class->oui[i], oui, 3))
1248 			return i;
1249 
1250 	return -1;
1251 }
1252 
1253 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1254 {
1255 	int i;
1256 
1257 	for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1258 		if (vendor->vendor_class[i])
1259 			return 1;
1260 
1261 	return 0;
1262 }
1263 
1264 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1265 				     struct ib_mad_agent_private *agent)
1266 {
1267 	int i;
1268 
1269 	/* Remove any methods for this mad agent */
1270 	for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1271 		if (method->agent[i] == agent)
1272 			method->agent[i] = NULL;
1273 }
1274 
1275 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1276 			      struct ib_mad_agent_private *agent_priv,
1277 			      u8 mgmt_class)
1278 {
1279 	struct ib_mad_port_private *port_priv;
1280 	struct ib_mad_mgmt_class_table **class;
1281 	struct ib_mad_mgmt_method_table **method;
1282 	int i, ret;
1283 
1284 	port_priv = agent_priv->qp_info->port_priv;
1285 	class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1286 	if (!*class) {
1287 		/* Allocate management class table for "new" class version */
1288 		*class = kzalloc(sizeof **class, GFP_ATOMIC);
1289 		if (!*class) {
1290 			ret = -ENOMEM;
1291 			goto error1;
1292 		}
1293 
1294 		/* Allocate method table for this management class */
1295 		method = &(*class)->method_table[mgmt_class];
1296 		if ((ret = allocate_method_table(method)))
1297 			goto error2;
1298 	} else {
1299 		method = &(*class)->method_table[mgmt_class];
1300 		if (!*method) {
1301 			/* Allocate method table for this management class */
1302 			if ((ret = allocate_method_table(method)))
1303 				goto error1;
1304 		}
1305 	}
1306 
1307 	/* Now, make sure methods are not already in use */
1308 	if (method_in_use(method, mad_reg_req))
1309 		goto error3;
1310 
1311 	/* Finally, add in methods being registered */
1312 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1313 		(*method)->agent[i] = agent_priv;
1314 
1315 	return 0;
1316 
1317 error3:
1318 	/* Remove any methods for this mad agent */
1319 	remove_methods_mad_agent(*method, agent_priv);
1320 	/* Now, check to see if there are any methods in use */
1321 	if (!check_method_table(*method)) {
1322 		/* If not, release management method table */
1323 		kfree(*method);
1324 		*method = NULL;
1325 	}
1326 	ret = -EINVAL;
1327 	goto error1;
1328 error2:
1329 	kfree(*class);
1330 	*class = NULL;
1331 error1:
1332 	return ret;
1333 }
1334 
1335 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1336 			   struct ib_mad_agent_private *agent_priv)
1337 {
1338 	struct ib_mad_port_private *port_priv;
1339 	struct ib_mad_mgmt_vendor_class_table **vendor_table;
1340 	struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1341 	struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1342 	struct ib_mad_mgmt_method_table **method;
1343 	int i, ret = -ENOMEM;
1344 	u8 vclass;
1345 
1346 	/* "New" vendor (with OUI) class */
1347 	vclass = vendor_class_index(mad_reg_req->mgmt_class);
1348 	port_priv = agent_priv->qp_info->port_priv;
1349 	vendor_table = &port_priv->version[
1350 				mad_reg_req->mgmt_class_version].vendor;
1351 	if (!*vendor_table) {
1352 		/* Allocate mgmt vendor class table for "new" class version */
1353 		vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1354 		if (!vendor)
1355 			goto error1;
1356 
1357 		*vendor_table = vendor;
1358 	}
1359 	if (!(*vendor_table)->vendor_class[vclass]) {
1360 		/* Allocate table for this management vendor class */
1361 		vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1362 		if (!vendor_class)
1363 			goto error2;
1364 
1365 		(*vendor_table)->vendor_class[vclass] = vendor_class;
1366 	}
1367 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1368 		/* Is there matching OUI for this vendor class ? */
1369 		if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1370 			    mad_reg_req->oui, 3)) {
1371 			method = &(*vendor_table)->vendor_class[
1372 						vclass]->method_table[i];
1373 			if (!*method)
1374 				goto error3;
1375 			goto check_in_use;
1376 		}
1377 	}
1378 	for (i = 0; i < MAX_MGMT_OUI; i++) {
1379 		/* OUI slot available ? */
1380 		if (!is_vendor_oui((*vendor_table)->vendor_class[
1381 				vclass]->oui[i])) {
1382 			method = &(*vendor_table)->vendor_class[
1383 				vclass]->method_table[i];
1384 			/* Allocate method table for this OUI */
1385 			if (!*method) {
1386 				ret = allocate_method_table(method);
1387 				if (ret)
1388 					goto error3;
1389 			}
1390 			memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1391 			       mad_reg_req->oui, 3);
1392 			goto check_in_use;
1393 		}
1394 	}
1395 	dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1396 	goto error3;
1397 
1398 check_in_use:
1399 	/* Now, make sure methods are not already in use */
1400 	if (method_in_use(method, mad_reg_req))
1401 		goto error4;
1402 
1403 	/* Finally, add in methods being registered */
1404 	for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1405 		(*method)->agent[i] = agent_priv;
1406 
1407 	return 0;
1408 
1409 error4:
1410 	/* Remove any methods for this mad agent */
1411 	remove_methods_mad_agent(*method, agent_priv);
1412 	/* Now, check to see if there are any methods in use */
1413 	if (!check_method_table(*method)) {
1414 		/* If not, release management method table */
1415 		kfree(*method);
1416 		*method = NULL;
1417 	}
1418 	ret = -EINVAL;
1419 error3:
1420 	if (vendor_class) {
1421 		(*vendor_table)->vendor_class[vclass] = NULL;
1422 		kfree(vendor_class);
1423 	}
1424 error2:
1425 	if (vendor) {
1426 		*vendor_table = NULL;
1427 		kfree(vendor);
1428 	}
1429 error1:
1430 	return ret;
1431 }
1432 
1433 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1434 {
1435 	struct ib_mad_port_private *port_priv;
1436 	struct ib_mad_mgmt_class_table *class;
1437 	struct ib_mad_mgmt_method_table *method;
1438 	struct ib_mad_mgmt_vendor_class_table *vendor;
1439 	struct ib_mad_mgmt_vendor_class *vendor_class;
1440 	int index;
1441 	u8 mgmt_class;
1442 
1443 	/*
1444 	 * Was MAD registration request supplied
1445 	 * with original registration ?
1446 	 */
1447 	if (!agent_priv->reg_req)
1448 		goto out;
1449 
1450 	port_priv = agent_priv->qp_info->port_priv;
1451 	mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1452 	class = port_priv->version[
1453 			agent_priv->reg_req->mgmt_class_version].class;
1454 	if (!class)
1455 		goto vendor_check;
1456 
1457 	method = class->method_table[mgmt_class];
1458 	if (method) {
1459 		/* Remove any methods for this mad agent */
1460 		remove_methods_mad_agent(method, agent_priv);
1461 		/* Now, check to see if there are any methods still in use */
1462 		if (!check_method_table(method)) {
1463 			/* If not, release management method table */
1464 			kfree(method);
1465 			class->method_table[mgmt_class] = NULL;
1466 			/* Any management classes left ? */
1467 			if (!check_class_table(class)) {
1468 				/* If not, release management class table */
1469 				kfree(class);
1470 				port_priv->version[
1471 					agent_priv->reg_req->
1472 					mgmt_class_version].class = NULL;
1473 			}
1474 		}
1475 	}
1476 
1477 vendor_check:
1478 	if (!is_vendor_class(mgmt_class))
1479 		goto out;
1480 
1481 	/* normalize mgmt_class to vendor range 2 */
1482 	mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1483 	vendor = port_priv->version[
1484 			agent_priv->reg_req->mgmt_class_version].vendor;
1485 
1486 	if (!vendor)
1487 		goto out;
1488 
1489 	vendor_class = vendor->vendor_class[mgmt_class];
1490 	if (vendor_class) {
1491 		index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1492 		if (index < 0)
1493 			goto out;
1494 		method = vendor_class->method_table[index];
1495 		if (method) {
1496 			/* Remove any methods for this mad agent */
1497 			remove_methods_mad_agent(method, agent_priv);
1498 			/*
1499 			 * Now, check to see if there are
1500 			 * any methods still in use
1501 			 */
1502 			if (!check_method_table(method)) {
1503 				/* If not, release management method table */
1504 				kfree(method);
1505 				vendor_class->method_table[index] = NULL;
1506 				memset(vendor_class->oui[index], 0, 3);
1507 				/* Any OUIs left ? */
1508 				if (!check_vendor_class(vendor_class)) {
1509 					/* If not, release vendor class table */
1510 					kfree(vendor_class);
1511 					vendor->vendor_class[mgmt_class] = NULL;
1512 					/* Any other vendor classes left ? */
1513 					if (!check_vendor_table(vendor)) {
1514 						kfree(vendor);
1515 						port_priv->version[
1516 							agent_priv->reg_req->
1517 							mgmt_class_version].
1518 							vendor = NULL;
1519 					}
1520 				}
1521 			}
1522 		}
1523 	}
1524 
1525 out:
1526 	return;
1527 }
1528 
1529 static struct ib_mad_agent_private *
1530 find_mad_agent(struct ib_mad_port_private *port_priv,
1531 	       const struct ib_mad_hdr *mad_hdr)
1532 {
1533 	struct ib_mad_agent_private *mad_agent = NULL;
1534 	unsigned long flags;
1535 
1536 	if (ib_response_mad(mad_hdr)) {
1537 		u32 hi_tid;
1538 
1539 		/*
1540 		 * Routing is based on high 32 bits of transaction ID
1541 		 * of MAD.
1542 		 */
1543 		hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1544 		rcu_read_lock();
1545 		mad_agent = xa_load(&ib_mad_clients, hi_tid);
1546 		if (mad_agent && !refcount_inc_not_zero(&mad_agent->refcount))
1547 			mad_agent = NULL;
1548 		rcu_read_unlock();
1549 	} else {
1550 		struct ib_mad_mgmt_class_table *class;
1551 		struct ib_mad_mgmt_method_table *method;
1552 		struct ib_mad_mgmt_vendor_class_table *vendor;
1553 		struct ib_mad_mgmt_vendor_class *vendor_class;
1554 		const struct ib_vendor_mad *vendor_mad;
1555 		int index;
1556 
1557 		spin_lock_irqsave(&port_priv->reg_lock, flags);
1558 		/*
1559 		 * Routing is based on version, class, and method
1560 		 * For "newer" vendor MADs, also based on OUI
1561 		 */
1562 		if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1563 			goto out;
1564 		if (!is_vendor_class(mad_hdr->mgmt_class)) {
1565 			class = port_priv->version[
1566 					mad_hdr->class_version].class;
1567 			if (!class)
1568 				goto out;
1569 			if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1570 			    ARRAY_SIZE(class->method_table))
1571 				goto out;
1572 			method = class->method_table[convert_mgmt_class(
1573 							mad_hdr->mgmt_class)];
1574 			if (method)
1575 				mad_agent = method->agent[mad_hdr->method &
1576 							  ~IB_MGMT_METHOD_RESP];
1577 		} else {
1578 			vendor = port_priv->version[
1579 					mad_hdr->class_version].vendor;
1580 			if (!vendor)
1581 				goto out;
1582 			vendor_class = vendor->vendor_class[vendor_class_index(
1583 						mad_hdr->mgmt_class)];
1584 			if (!vendor_class)
1585 				goto out;
1586 			/* Find matching OUI */
1587 			vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1588 			index = find_vendor_oui(vendor_class, vendor_mad->oui);
1589 			if (index == -1)
1590 				goto out;
1591 			method = vendor_class->method_table[index];
1592 			if (method) {
1593 				mad_agent = method->agent[mad_hdr->method &
1594 							  ~IB_MGMT_METHOD_RESP];
1595 			}
1596 		}
1597 		if (mad_agent)
1598 			refcount_inc(&mad_agent->refcount);
1599 out:
1600 		spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1601 	}
1602 
1603 	if (mad_agent && !mad_agent->agent.recv_handler) {
1604 		dev_notice(&port_priv->device->dev,
1605 			   "No receive handler for client %p on port %u\n",
1606 			   &mad_agent->agent, port_priv->port_num);
1607 		deref_mad_agent(mad_agent);
1608 		mad_agent = NULL;
1609 	}
1610 
1611 	return mad_agent;
1612 }
1613 
1614 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1615 			const struct ib_mad_qp_info *qp_info,
1616 			bool opa)
1617 {
1618 	int valid = 0;
1619 	u32 qp_num = qp_info->qp->qp_num;
1620 
1621 	/* Make sure MAD base version is understood */
1622 	if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1623 	    (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1624 		pr_err("MAD received with unsupported base version %u %s\n",
1625 		       mad_hdr->base_version, opa ? "(opa)" : "");
1626 		goto out;
1627 	}
1628 
1629 	/* Filter SMI packets sent to other than QP0 */
1630 	if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1631 	    (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1632 		if (qp_num == 0)
1633 			valid = 1;
1634 	} else {
1635 		/* CM attributes other than ClassPortInfo only use Send method */
1636 		if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1637 		    (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1638 		    (mad_hdr->method != IB_MGMT_METHOD_SEND))
1639 			goto out;
1640 		/* Filter GSI packets sent to QP0 */
1641 		if (qp_num != 0)
1642 			valid = 1;
1643 	}
1644 
1645 out:
1646 	return valid;
1647 }
1648 
1649 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1650 			    const struct ib_mad_hdr *mad_hdr)
1651 {
1652 	struct ib_rmpp_mad *rmpp_mad;
1653 
1654 	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1655 	return !mad_agent_priv->agent.rmpp_version ||
1656 		!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1657 		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1658 				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
1659 		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1660 }
1661 
1662 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1663 				     const struct ib_mad_recv_wc *rwc)
1664 {
1665 	return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1666 		rwc->recv_buf.mad->mad_hdr.mgmt_class;
1667 }
1668 
1669 static inline int
1670 rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1671 		 const struct ib_mad_send_wr_private *wr,
1672 		 const struct ib_mad_recv_wc *rwc)
1673 {
1674 	struct rdma_ah_attr attr;
1675 	u8 send_resp, rcv_resp;
1676 	union ib_gid sgid;
1677 	struct ib_device *device = mad_agent_priv->agent.device;
1678 	u32 port_num = mad_agent_priv->agent.port_num;
1679 	u8 lmc;
1680 	bool has_grh;
1681 
1682 	send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1683 	rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1684 
1685 	if (send_resp == rcv_resp)
1686 		/* both requests, or both responses. GIDs different */
1687 		return 0;
1688 
1689 	if (rdma_query_ah(wr->send_buf.ah, &attr))
1690 		/* Assume not equal, to avoid false positives. */
1691 		return 0;
1692 
1693 	has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1694 	if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1695 		/* one has GID, other does not.  Assume different */
1696 		return 0;
1697 
1698 	if (!send_resp && rcv_resp) {
1699 		/* is request/response. */
1700 		if (!has_grh) {
1701 			if (ib_get_cached_lmc(device, port_num, &lmc))
1702 				return 0;
1703 			return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1704 					   rwc->wc->dlid_path_bits) &
1705 					  ((1 << lmc) - 1)));
1706 		} else {
1707 			const struct ib_global_route *grh =
1708 					rdma_ah_read_grh(&attr);
1709 
1710 			if (rdma_query_gid(device, port_num,
1711 					   grh->sgid_index, &sgid))
1712 				return 0;
1713 			return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1714 				       16);
1715 		}
1716 	}
1717 
1718 	if (!has_grh)
1719 		return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1720 	else
1721 		return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1722 			       rwc->recv_buf.grh->sgid.raw,
1723 			       16);
1724 }
1725 
1726 static inline int is_direct(u8 class)
1727 {
1728 	return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1729 }
1730 
1731 struct ib_mad_send_wr_private*
1732 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1733 		 const struct ib_mad_recv_wc *wc)
1734 {
1735 	struct ib_mad_send_wr_private *wr;
1736 	const struct ib_mad_hdr *mad_hdr;
1737 
1738 	mad_hdr = &wc->recv_buf.mad->mad_hdr;
1739 
1740 	list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1741 		if ((wr->tid == mad_hdr->tid) &&
1742 		    rcv_has_same_class(wr, wc) &&
1743 		    /*
1744 		     * Don't check GID for direct routed MADs.
1745 		     * These might have permissive LIDs.
1746 		     */
1747 		    (is_direct(mad_hdr->mgmt_class) ||
1748 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1749 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1750 	}
1751 
1752 	/*
1753 	 * It's possible to receive the response before we've
1754 	 * been notified that the send has completed
1755 	 */
1756 	list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1757 		if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1758 		    wr->tid == mad_hdr->tid &&
1759 		    wr->timeout &&
1760 		    rcv_has_same_class(wr, wc) &&
1761 		    /*
1762 		     * Don't check GID for direct routed MADs.
1763 		     * These might have permissive LIDs.
1764 		     */
1765 		    (is_direct(mad_hdr->mgmt_class) ||
1766 		     rcv_has_same_gid(mad_agent_priv, wr, wc)))
1767 			/* Verify request has not been canceled */
1768 			return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1769 	}
1770 	return NULL;
1771 }
1772 
1773 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1774 {
1775 	mad_send_wr->timeout = 0;
1776 	if (mad_send_wr->refcount == 1)
1777 		list_move_tail(&mad_send_wr->agent_list,
1778 			      &mad_send_wr->mad_agent_priv->done_list);
1779 }
1780 
1781 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1782 				 struct ib_mad_recv_wc *mad_recv_wc)
1783 {
1784 	struct ib_mad_send_wr_private *mad_send_wr;
1785 	struct ib_mad_send_wc mad_send_wc;
1786 	unsigned long flags;
1787 	int ret;
1788 
1789 	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1790 	ret = ib_mad_enforce_security(mad_agent_priv,
1791 				      mad_recv_wc->wc->pkey_index);
1792 	if (ret) {
1793 		ib_free_recv_mad(mad_recv_wc);
1794 		deref_mad_agent(mad_agent_priv);
1795 		return;
1796 	}
1797 
1798 	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1799 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1800 		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1801 						      mad_recv_wc);
1802 		if (!mad_recv_wc) {
1803 			deref_mad_agent(mad_agent_priv);
1804 			return;
1805 		}
1806 	}
1807 
1808 	/* Complete corresponding request */
1809 	if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1810 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
1811 		mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1812 		if (!mad_send_wr) {
1813 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1814 			if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1815 			   && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1816 			   && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1817 					& IB_MGMT_RMPP_FLAG_ACTIVE)) {
1818 				/* user rmpp is in effect
1819 				 * and this is an active RMPP MAD
1820 				 */
1821 				mad_agent_priv->agent.recv_handler(
1822 						&mad_agent_priv->agent, NULL,
1823 						mad_recv_wc);
1824 				deref_mad_agent(mad_agent_priv);
1825 			} else {
1826 				/* not user rmpp, revert to normal behavior and
1827 				 * drop the mad
1828 				 */
1829 				ib_free_recv_mad(mad_recv_wc);
1830 				deref_mad_agent(mad_agent_priv);
1831 				return;
1832 			}
1833 		} else {
1834 			ib_mark_mad_done(mad_send_wr);
1835 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1836 
1837 			/* Defined behavior is to complete response before request */
1838 			mad_agent_priv->agent.recv_handler(
1839 					&mad_agent_priv->agent,
1840 					&mad_send_wr->send_buf,
1841 					mad_recv_wc);
1842 			deref_mad_agent(mad_agent_priv);
1843 
1844 			mad_send_wc.status = IB_WC_SUCCESS;
1845 			mad_send_wc.vendor_err = 0;
1846 			mad_send_wc.send_buf = &mad_send_wr->send_buf;
1847 			ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1848 		}
1849 	} else {
1850 		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
1851 						   mad_recv_wc);
1852 		deref_mad_agent(mad_agent_priv);
1853 	}
1854 }
1855 
1856 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
1857 				     const struct ib_mad_qp_info *qp_info,
1858 				     const struct ib_wc *wc,
1859 				     u32 port_num,
1860 				     struct ib_mad_private *recv,
1861 				     struct ib_mad_private *response)
1862 {
1863 	enum smi_forward_action retsmi;
1864 	struct ib_smp *smp = (struct ib_smp *)recv->mad;
1865 
1866 	trace_ib_mad_handle_ib_smi(smp);
1867 
1868 	if (smi_handle_dr_smp_recv(smp,
1869 				   rdma_cap_ib_switch(port_priv->device),
1870 				   port_num,
1871 				   port_priv->device->phys_port_cnt) ==
1872 				   IB_SMI_DISCARD)
1873 		return IB_SMI_DISCARD;
1874 
1875 	retsmi = smi_check_forward_dr_smp(smp);
1876 	if (retsmi == IB_SMI_LOCAL)
1877 		return IB_SMI_HANDLE;
1878 
1879 	if (retsmi == IB_SMI_SEND) { /* don't forward */
1880 		if (smi_handle_dr_smp_send(smp,
1881 					   rdma_cap_ib_switch(port_priv->device),
1882 					   port_num) == IB_SMI_DISCARD)
1883 			return IB_SMI_DISCARD;
1884 
1885 		if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
1886 			return IB_SMI_DISCARD;
1887 	} else if (rdma_cap_ib_switch(port_priv->device)) {
1888 		/* forward case for switches */
1889 		memcpy(response, recv, mad_priv_size(response));
1890 		response->header.recv_wc.wc = &response->header.wc;
1891 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
1892 		response->header.recv_wc.recv_buf.grh = &response->grh;
1893 
1894 		agent_send_response((const struct ib_mad_hdr *)response->mad,
1895 				    &response->grh, wc,
1896 				    port_priv->device,
1897 				    smi_get_fwd_port(smp),
1898 				    qp_info->qp->qp_num,
1899 				    response->mad_size,
1900 				    false);
1901 
1902 		return IB_SMI_DISCARD;
1903 	}
1904 	return IB_SMI_HANDLE;
1905 }
1906 
1907 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
1908 				    struct ib_mad_private *response,
1909 				    size_t *resp_len, bool opa)
1910 {
1911 	const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
1912 	struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
1913 
1914 	if (recv_hdr->method == IB_MGMT_METHOD_GET ||
1915 	    recv_hdr->method == IB_MGMT_METHOD_SET) {
1916 		memcpy(response, recv, mad_priv_size(response));
1917 		response->header.recv_wc.wc = &response->header.wc;
1918 		response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
1919 		response->header.recv_wc.recv_buf.grh = &response->grh;
1920 		resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
1921 		resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
1922 		if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1923 			resp_hdr->status |= IB_SMP_DIRECTION;
1924 
1925 		if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
1926 			if (recv_hdr->mgmt_class ==
1927 			    IB_MGMT_CLASS_SUBN_LID_ROUTED ||
1928 			    recv_hdr->mgmt_class ==
1929 			    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1930 				*resp_len = opa_get_smp_header_size(
1931 							(struct opa_smp *)recv->mad);
1932 			else
1933 				*resp_len = sizeof(struct ib_mad_hdr);
1934 		}
1935 
1936 		return true;
1937 	} else {
1938 		return false;
1939 	}
1940 }
1941 
1942 static enum smi_action
1943 handle_opa_smi(struct ib_mad_port_private *port_priv,
1944 	       struct ib_mad_qp_info *qp_info,
1945 	       struct ib_wc *wc,
1946 	       u32 port_num,
1947 	       struct ib_mad_private *recv,
1948 	       struct ib_mad_private *response)
1949 {
1950 	enum smi_forward_action retsmi;
1951 	struct opa_smp *smp = (struct opa_smp *)recv->mad;
1952 
1953 	trace_ib_mad_handle_opa_smi(smp);
1954 
1955 	if (opa_smi_handle_dr_smp_recv(smp,
1956 				   rdma_cap_ib_switch(port_priv->device),
1957 				   port_num,
1958 				   port_priv->device->phys_port_cnt) ==
1959 				   IB_SMI_DISCARD)
1960 		return IB_SMI_DISCARD;
1961 
1962 	retsmi = opa_smi_check_forward_dr_smp(smp);
1963 	if (retsmi == IB_SMI_LOCAL)
1964 		return IB_SMI_HANDLE;
1965 
1966 	if (retsmi == IB_SMI_SEND) { /* don't forward */
1967 		if (opa_smi_handle_dr_smp_send(smp,
1968 					   rdma_cap_ib_switch(port_priv->device),
1969 					   port_num) == IB_SMI_DISCARD)
1970 			return IB_SMI_DISCARD;
1971 
1972 		if (opa_smi_check_local_smp(smp, port_priv->device) ==
1973 		    IB_SMI_DISCARD)
1974 			return IB_SMI_DISCARD;
1975 
1976 	} else if (rdma_cap_ib_switch(port_priv->device)) {
1977 		/* forward case for switches */
1978 		memcpy(response, recv, mad_priv_size(response));
1979 		response->header.recv_wc.wc = &response->header.wc;
1980 		response->header.recv_wc.recv_buf.opa_mad =
1981 				(struct opa_mad *)response->mad;
1982 		response->header.recv_wc.recv_buf.grh = &response->grh;
1983 
1984 		agent_send_response((const struct ib_mad_hdr *)response->mad,
1985 				    &response->grh, wc,
1986 				    port_priv->device,
1987 				    opa_smi_get_fwd_port(smp),
1988 				    qp_info->qp->qp_num,
1989 				    recv->header.wc.byte_len,
1990 				    true);
1991 
1992 		return IB_SMI_DISCARD;
1993 	}
1994 
1995 	return IB_SMI_HANDLE;
1996 }
1997 
1998 static enum smi_action
1999 handle_smi(struct ib_mad_port_private *port_priv,
2000 	   struct ib_mad_qp_info *qp_info,
2001 	   struct ib_wc *wc,
2002 	   u32 port_num,
2003 	   struct ib_mad_private *recv,
2004 	   struct ib_mad_private *response,
2005 	   bool opa)
2006 {
2007 	struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2008 
2009 	if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2010 	    mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2011 		return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2012 				      response);
2013 
2014 	return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2015 }
2016 
2017 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2018 {
2019 	struct ib_mad_port_private *port_priv = cq->cq_context;
2020 	struct ib_mad_list_head *mad_list =
2021 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2022 	struct ib_mad_qp_info *qp_info;
2023 	struct ib_mad_private_header *mad_priv_hdr;
2024 	struct ib_mad_private *recv, *response = NULL;
2025 	struct ib_mad_agent_private *mad_agent;
2026 	u32 port_num;
2027 	int ret = IB_MAD_RESULT_SUCCESS;
2028 	size_t mad_size;
2029 	u16 resp_mad_pkey_index = 0;
2030 	bool opa;
2031 
2032 	if (list_empty_careful(&port_priv->port_list))
2033 		return;
2034 
2035 	if (wc->status != IB_WC_SUCCESS) {
2036 		/*
2037 		 * Receive errors indicate that the QP has entered the error
2038 		 * state - error handling/shutdown code will cleanup
2039 		 */
2040 		return;
2041 	}
2042 
2043 	qp_info = mad_list->mad_queue->qp_info;
2044 	dequeue_mad(mad_list);
2045 
2046 	opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2047 			       qp_info->port_priv->port_num);
2048 
2049 	mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2050 				    mad_list);
2051 	recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2052 	ib_dma_unmap_single(port_priv->device,
2053 			    recv->header.mapping,
2054 			    mad_priv_dma_size(recv),
2055 			    DMA_FROM_DEVICE);
2056 
2057 	/* Setup MAD receive work completion from "normal" work completion */
2058 	recv->header.wc = *wc;
2059 	recv->header.recv_wc.wc = &recv->header.wc;
2060 
2061 	if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2062 		recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2063 		recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2064 	} else {
2065 		recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2066 		recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2067 	}
2068 
2069 	recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2070 	recv->header.recv_wc.recv_buf.grh = &recv->grh;
2071 
2072 	/* Validate MAD */
2073 	if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2074 		goto out;
2075 
2076 	trace_ib_mad_recv_done_handler(qp_info, wc,
2077 				       (struct ib_mad_hdr *)recv->mad);
2078 
2079 	mad_size = recv->mad_size;
2080 	response = alloc_mad_private(mad_size, GFP_KERNEL);
2081 	if (!response)
2082 		goto out;
2083 
2084 	if (rdma_cap_ib_switch(port_priv->device))
2085 		port_num = wc->port_num;
2086 	else
2087 		port_num = port_priv->port_num;
2088 
2089 	if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2090 	    IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2091 		if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2092 			       response, opa)
2093 		    == IB_SMI_DISCARD)
2094 			goto out;
2095 	}
2096 
2097 	/* Give driver "right of first refusal" on incoming MAD */
2098 	if (port_priv->device->ops.process_mad) {
2099 		ret = port_priv->device->ops.process_mad(
2100 			port_priv->device, 0, port_priv->port_num, wc,
2101 			&recv->grh, (const struct ib_mad *)recv->mad,
2102 			(struct ib_mad *)response->mad, &mad_size,
2103 			&resp_mad_pkey_index);
2104 
2105 		if (opa)
2106 			wc->pkey_index = resp_mad_pkey_index;
2107 
2108 		if (ret & IB_MAD_RESULT_SUCCESS) {
2109 			if (ret & IB_MAD_RESULT_CONSUMED)
2110 				goto out;
2111 			if (ret & IB_MAD_RESULT_REPLY) {
2112 				agent_send_response((const struct ib_mad_hdr *)response->mad,
2113 						    &recv->grh, wc,
2114 						    port_priv->device,
2115 						    port_num,
2116 						    qp_info->qp->qp_num,
2117 						    mad_size, opa);
2118 				goto out;
2119 			}
2120 		}
2121 	}
2122 
2123 	mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2124 	if (mad_agent) {
2125 		trace_ib_mad_recv_done_agent(mad_agent);
2126 		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2127 		/*
2128 		 * recv is freed up in error cases in ib_mad_complete_recv
2129 		 * or via recv_handler in ib_mad_complete_recv()
2130 		 */
2131 		recv = NULL;
2132 	} else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2133 		   generate_unmatched_resp(recv, response, &mad_size, opa)) {
2134 		agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2135 				    port_priv->device, port_num,
2136 				    qp_info->qp->qp_num, mad_size, opa);
2137 	}
2138 
2139 out:
2140 	/* Post another receive request for this QP */
2141 	if (response) {
2142 		ib_mad_post_receive_mads(qp_info, response);
2143 		kfree(recv);
2144 	} else
2145 		ib_mad_post_receive_mads(qp_info, recv);
2146 }
2147 
2148 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2149 {
2150 	struct ib_mad_send_wr_private *mad_send_wr;
2151 	unsigned long delay;
2152 
2153 	if (list_empty(&mad_agent_priv->wait_list)) {
2154 		cancel_delayed_work(&mad_agent_priv->timed_work);
2155 	} else {
2156 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2157 					 struct ib_mad_send_wr_private,
2158 					 agent_list);
2159 
2160 		if (time_after(mad_agent_priv->timeout,
2161 			       mad_send_wr->timeout)) {
2162 			mad_agent_priv->timeout = mad_send_wr->timeout;
2163 			delay = mad_send_wr->timeout - jiffies;
2164 			if ((long)delay <= 0)
2165 				delay = 1;
2166 			mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2167 					 &mad_agent_priv->timed_work, delay);
2168 		}
2169 	}
2170 }
2171 
2172 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2173 {
2174 	struct ib_mad_agent_private *mad_agent_priv;
2175 	struct ib_mad_send_wr_private *temp_mad_send_wr;
2176 	struct list_head *list_item;
2177 	unsigned long delay;
2178 
2179 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2180 	list_del(&mad_send_wr->agent_list);
2181 
2182 	delay = mad_send_wr->timeout;
2183 	mad_send_wr->timeout += jiffies;
2184 
2185 	if (delay) {
2186 		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2187 			temp_mad_send_wr = list_entry(list_item,
2188 						struct ib_mad_send_wr_private,
2189 						agent_list);
2190 			if (time_after(mad_send_wr->timeout,
2191 				       temp_mad_send_wr->timeout))
2192 				break;
2193 		}
2194 	} else {
2195 		list_item = &mad_agent_priv->wait_list;
2196 	}
2197 
2198 	list_add(&mad_send_wr->agent_list, list_item);
2199 
2200 	/* Reschedule a work item if we have a shorter timeout */
2201 	if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2202 		mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2203 				 &mad_agent_priv->timed_work, delay);
2204 }
2205 
2206 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2207 			  unsigned long timeout_ms)
2208 {
2209 	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2210 	wait_for_response(mad_send_wr);
2211 }
2212 
2213 /*
2214  * Process a send work completion
2215  */
2216 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2217 			     struct ib_mad_send_wc *mad_send_wc)
2218 {
2219 	struct ib_mad_agent_private	*mad_agent_priv;
2220 	unsigned long			flags;
2221 	int				ret;
2222 
2223 	mad_agent_priv = mad_send_wr->mad_agent_priv;
2224 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2225 	if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2226 		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2227 		if (ret == IB_RMPP_RESULT_CONSUMED)
2228 			goto done;
2229 	} else
2230 		ret = IB_RMPP_RESULT_UNHANDLED;
2231 
2232 	if (mad_send_wc->status != IB_WC_SUCCESS &&
2233 	    mad_send_wr->status == IB_WC_SUCCESS) {
2234 		mad_send_wr->status = mad_send_wc->status;
2235 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2236 	}
2237 
2238 	if (--mad_send_wr->refcount > 0) {
2239 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2240 		    mad_send_wr->status == IB_WC_SUCCESS) {
2241 			wait_for_response(mad_send_wr);
2242 		}
2243 		goto done;
2244 	}
2245 
2246 	/* Remove send from MAD agent and notify client of completion */
2247 	list_del(&mad_send_wr->agent_list);
2248 	adjust_timeout(mad_agent_priv);
2249 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2250 
2251 	if (mad_send_wr->status != IB_WC_SUCCESS)
2252 		mad_send_wc->status = mad_send_wr->status;
2253 	if (ret == IB_RMPP_RESULT_INTERNAL)
2254 		ib_rmpp_send_handler(mad_send_wc);
2255 	else
2256 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2257 						   mad_send_wc);
2258 
2259 	/* Release reference on agent taken when sending */
2260 	deref_mad_agent(mad_agent_priv);
2261 	return;
2262 done:
2263 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2264 }
2265 
2266 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2267 {
2268 	struct ib_mad_port_private *port_priv = cq->cq_context;
2269 	struct ib_mad_list_head *mad_list =
2270 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2271 	struct ib_mad_send_wr_private	*mad_send_wr, *queued_send_wr;
2272 	struct ib_mad_qp_info		*qp_info;
2273 	struct ib_mad_queue		*send_queue;
2274 	struct ib_mad_send_wc		mad_send_wc;
2275 	unsigned long flags;
2276 	int ret;
2277 
2278 	if (list_empty_careful(&port_priv->port_list))
2279 		return;
2280 
2281 	if (wc->status != IB_WC_SUCCESS) {
2282 		if (!ib_mad_send_error(port_priv, wc))
2283 			return;
2284 	}
2285 
2286 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2287 				   mad_list);
2288 	send_queue = mad_list->mad_queue;
2289 	qp_info = send_queue->qp_info;
2290 
2291 	trace_ib_mad_send_done_agent(mad_send_wr->mad_agent_priv);
2292 	trace_ib_mad_send_done_handler(mad_send_wr, wc);
2293 
2294 retry:
2295 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2296 			    mad_send_wr->header_mapping,
2297 			    mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2298 	ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2299 			    mad_send_wr->payload_mapping,
2300 			    mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2301 	queued_send_wr = NULL;
2302 	spin_lock_irqsave(&send_queue->lock, flags);
2303 	list_del(&mad_list->list);
2304 
2305 	/* Move queued send to the send queue */
2306 	if (send_queue->count-- > send_queue->max_active) {
2307 		mad_list = container_of(qp_info->overflow_list.next,
2308 					struct ib_mad_list_head, list);
2309 		queued_send_wr = container_of(mad_list,
2310 					struct ib_mad_send_wr_private,
2311 					mad_list);
2312 		list_move_tail(&mad_list->list, &send_queue->list);
2313 	}
2314 	spin_unlock_irqrestore(&send_queue->lock, flags);
2315 
2316 	mad_send_wc.send_buf = &mad_send_wr->send_buf;
2317 	mad_send_wc.status = wc->status;
2318 	mad_send_wc.vendor_err = wc->vendor_err;
2319 	ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2320 
2321 	if (queued_send_wr) {
2322 		trace_ib_mad_send_done_resend(queued_send_wr, qp_info);
2323 		ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2324 				   NULL);
2325 		if (ret) {
2326 			dev_err(&port_priv->device->dev,
2327 				"ib_post_send failed: %d\n", ret);
2328 			mad_send_wr = queued_send_wr;
2329 			wc->status = IB_WC_LOC_QP_OP_ERR;
2330 			goto retry;
2331 		}
2332 	}
2333 }
2334 
2335 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2336 {
2337 	struct ib_mad_send_wr_private *mad_send_wr;
2338 	struct ib_mad_list_head *mad_list;
2339 	unsigned long flags;
2340 
2341 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2342 	list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2343 		mad_send_wr = container_of(mad_list,
2344 					   struct ib_mad_send_wr_private,
2345 					   mad_list);
2346 		mad_send_wr->retry = 1;
2347 	}
2348 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2349 }
2350 
2351 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2352 		struct ib_wc *wc)
2353 {
2354 	struct ib_mad_list_head *mad_list =
2355 		container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2356 	struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2357 	struct ib_mad_send_wr_private *mad_send_wr;
2358 	int ret;
2359 
2360 	/*
2361 	 * Send errors will transition the QP to SQE - move
2362 	 * QP to RTS and repost flushed work requests
2363 	 */
2364 	mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2365 				   mad_list);
2366 	if (wc->status == IB_WC_WR_FLUSH_ERR) {
2367 		if (mad_send_wr->retry) {
2368 			/* Repost send */
2369 			mad_send_wr->retry = 0;
2370 			trace_ib_mad_error_handler(mad_send_wr, qp_info);
2371 			ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2372 					   NULL);
2373 			if (!ret)
2374 				return false;
2375 		}
2376 	} else {
2377 		struct ib_qp_attr *attr;
2378 
2379 		/* Transition QP to RTS and fail offending send */
2380 		attr = kmalloc(sizeof *attr, GFP_KERNEL);
2381 		if (attr) {
2382 			attr->qp_state = IB_QPS_RTS;
2383 			attr->cur_qp_state = IB_QPS_SQE;
2384 			ret = ib_modify_qp(qp_info->qp, attr,
2385 					   IB_QP_STATE | IB_QP_CUR_STATE);
2386 			kfree(attr);
2387 			if (ret)
2388 				dev_err(&port_priv->device->dev,
2389 					"%s - ib_modify_qp to RTS: %d\n",
2390 					__func__, ret);
2391 			else
2392 				mark_sends_for_retry(qp_info);
2393 		}
2394 	}
2395 
2396 	return true;
2397 }
2398 
2399 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2400 {
2401 	unsigned long flags;
2402 	struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2403 	struct ib_mad_send_wc mad_send_wc;
2404 	struct list_head cancel_list;
2405 
2406 	INIT_LIST_HEAD(&cancel_list);
2407 
2408 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2409 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2410 				 &mad_agent_priv->send_list, agent_list) {
2411 		if (mad_send_wr->status == IB_WC_SUCCESS) {
2412 			mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2413 			mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2414 		}
2415 	}
2416 
2417 	/* Empty wait list to prevent receives from finding a request */
2418 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2419 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2420 
2421 	/* Report all cancelled requests */
2422 	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2423 	mad_send_wc.vendor_err = 0;
2424 
2425 	list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2426 				 &cancel_list, agent_list) {
2427 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2428 		list_del(&mad_send_wr->agent_list);
2429 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2430 						   &mad_send_wc);
2431 		deref_mad_agent(mad_agent_priv);
2432 	}
2433 }
2434 
2435 static struct ib_mad_send_wr_private*
2436 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2437 	     struct ib_mad_send_buf *send_buf)
2438 {
2439 	struct ib_mad_send_wr_private *mad_send_wr;
2440 
2441 	list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2442 			    agent_list) {
2443 		if (&mad_send_wr->send_buf == send_buf)
2444 			return mad_send_wr;
2445 	}
2446 
2447 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2448 			    agent_list) {
2449 		if (is_rmpp_data_mad(mad_agent_priv,
2450 				     mad_send_wr->send_buf.mad) &&
2451 		    &mad_send_wr->send_buf == send_buf)
2452 			return mad_send_wr;
2453 	}
2454 	return NULL;
2455 }
2456 
2457 int ib_modify_mad(struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2458 {
2459 	struct ib_mad_agent_private *mad_agent_priv;
2460 	struct ib_mad_send_wr_private *mad_send_wr;
2461 	unsigned long flags;
2462 	int active;
2463 
2464 	if (!send_buf)
2465 		return -EINVAL;
2466 
2467 	mad_agent_priv = container_of(send_buf->mad_agent,
2468 				      struct ib_mad_agent_private, agent);
2469 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2470 	mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2471 	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2472 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2473 		return -EINVAL;
2474 	}
2475 
2476 	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2477 	if (!timeout_ms) {
2478 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2479 		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2480 	}
2481 
2482 	mad_send_wr->send_buf.timeout_ms = timeout_ms;
2483 	if (active)
2484 		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2485 	else
2486 		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2487 
2488 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2489 	return 0;
2490 }
2491 EXPORT_SYMBOL(ib_modify_mad);
2492 
2493 static void local_completions(struct work_struct *work)
2494 {
2495 	struct ib_mad_agent_private *mad_agent_priv;
2496 	struct ib_mad_local_private *local;
2497 	struct ib_mad_agent_private *recv_mad_agent;
2498 	unsigned long flags;
2499 	int free_mad;
2500 	struct ib_wc wc;
2501 	struct ib_mad_send_wc mad_send_wc;
2502 	bool opa;
2503 
2504 	mad_agent_priv =
2505 		container_of(work, struct ib_mad_agent_private, local_work);
2506 
2507 	opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2508 			       mad_agent_priv->qp_info->port_priv->port_num);
2509 
2510 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2511 	while (!list_empty(&mad_agent_priv->local_list)) {
2512 		local = list_entry(mad_agent_priv->local_list.next,
2513 				   struct ib_mad_local_private,
2514 				   completion_list);
2515 		list_del(&local->completion_list);
2516 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2517 		free_mad = 0;
2518 		if (local->mad_priv) {
2519 			u8 base_version;
2520 			recv_mad_agent = local->recv_mad_agent;
2521 			if (!recv_mad_agent) {
2522 				dev_err(&mad_agent_priv->agent.device->dev,
2523 					"No receive MAD agent for local completion\n");
2524 				free_mad = 1;
2525 				goto local_send_completion;
2526 			}
2527 
2528 			/*
2529 			 * Defined behavior is to complete response
2530 			 * before request
2531 			 */
2532 			build_smp_wc(recv_mad_agent->agent.qp,
2533 				     local->mad_send_wr->send_wr.wr.wr_cqe,
2534 				     be16_to_cpu(IB_LID_PERMISSIVE),
2535 				     local->mad_send_wr->send_wr.pkey_index,
2536 				     recv_mad_agent->agent.port_num, &wc);
2537 
2538 			local->mad_priv->header.recv_wc.wc = &wc;
2539 
2540 			base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2541 			if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2542 				local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2543 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2544 			} else {
2545 				local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2546 				local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2547 			}
2548 
2549 			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2550 			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2551 				 &local->mad_priv->header.recv_wc.rmpp_list);
2552 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2553 			local->mad_priv->header.recv_wc.recv_buf.mad =
2554 						(struct ib_mad *)local->mad_priv->mad;
2555 			recv_mad_agent->agent.recv_handler(
2556 						&recv_mad_agent->agent,
2557 						&local->mad_send_wr->send_buf,
2558 						&local->mad_priv->header.recv_wc);
2559 			spin_lock_irqsave(&recv_mad_agent->lock, flags);
2560 			deref_mad_agent(recv_mad_agent);
2561 			spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2562 		}
2563 
2564 local_send_completion:
2565 		/* Complete send */
2566 		mad_send_wc.status = IB_WC_SUCCESS;
2567 		mad_send_wc.vendor_err = 0;
2568 		mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2569 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2570 						   &mad_send_wc);
2571 
2572 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2573 		deref_mad_agent(mad_agent_priv);
2574 		if (free_mad)
2575 			kfree(local->mad_priv);
2576 		kfree(local);
2577 	}
2578 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2579 }
2580 
2581 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2582 {
2583 	int ret;
2584 
2585 	if (!mad_send_wr->retries_left)
2586 		return -ETIMEDOUT;
2587 
2588 	mad_send_wr->retries_left--;
2589 	mad_send_wr->send_buf.retries++;
2590 
2591 	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2592 
2593 	if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2594 		ret = ib_retry_rmpp(mad_send_wr);
2595 		switch (ret) {
2596 		case IB_RMPP_RESULT_UNHANDLED:
2597 			ret = ib_send_mad(mad_send_wr);
2598 			break;
2599 		case IB_RMPP_RESULT_CONSUMED:
2600 			ret = 0;
2601 			break;
2602 		default:
2603 			ret = -ECOMM;
2604 			break;
2605 		}
2606 	} else
2607 		ret = ib_send_mad(mad_send_wr);
2608 
2609 	if (!ret) {
2610 		mad_send_wr->refcount++;
2611 		list_add_tail(&mad_send_wr->agent_list,
2612 			      &mad_send_wr->mad_agent_priv->send_list);
2613 	}
2614 	return ret;
2615 }
2616 
2617 static void timeout_sends(struct work_struct *work)
2618 {
2619 	struct ib_mad_agent_private *mad_agent_priv;
2620 	struct ib_mad_send_wr_private *mad_send_wr;
2621 	struct ib_mad_send_wc mad_send_wc;
2622 	unsigned long flags, delay;
2623 
2624 	mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2625 				      timed_work.work);
2626 	mad_send_wc.vendor_err = 0;
2627 
2628 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
2629 	while (!list_empty(&mad_agent_priv->wait_list)) {
2630 		mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2631 					 struct ib_mad_send_wr_private,
2632 					 agent_list);
2633 
2634 		if (time_after(mad_send_wr->timeout, jiffies)) {
2635 			delay = mad_send_wr->timeout - jiffies;
2636 			if ((long)delay <= 0)
2637 				delay = 1;
2638 			queue_delayed_work(mad_agent_priv->qp_info->
2639 					   port_priv->wq,
2640 					   &mad_agent_priv->timed_work, delay);
2641 			break;
2642 		}
2643 
2644 		list_del(&mad_send_wr->agent_list);
2645 		if (mad_send_wr->status == IB_WC_SUCCESS &&
2646 		    !retry_send(mad_send_wr))
2647 			continue;
2648 
2649 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2650 
2651 		if (mad_send_wr->status == IB_WC_SUCCESS)
2652 			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2653 		else
2654 			mad_send_wc.status = mad_send_wr->status;
2655 		mad_send_wc.send_buf = &mad_send_wr->send_buf;
2656 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2657 						   &mad_send_wc);
2658 
2659 		deref_mad_agent(mad_agent_priv);
2660 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
2661 	}
2662 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2663 }
2664 
2665 /*
2666  * Allocate receive MADs and post receive WRs for them
2667  */
2668 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2669 				    struct ib_mad_private *mad)
2670 {
2671 	unsigned long flags;
2672 	int post, ret;
2673 	struct ib_mad_private *mad_priv;
2674 	struct ib_sge sg_list;
2675 	struct ib_recv_wr recv_wr;
2676 	struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2677 
2678 	/* Initialize common scatter list fields */
2679 	sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2680 
2681 	/* Initialize common receive WR fields */
2682 	recv_wr.next = NULL;
2683 	recv_wr.sg_list = &sg_list;
2684 	recv_wr.num_sge = 1;
2685 
2686 	do {
2687 		/* Allocate and map receive buffer */
2688 		if (mad) {
2689 			mad_priv = mad;
2690 			mad = NULL;
2691 		} else {
2692 			mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2693 						     GFP_ATOMIC);
2694 			if (!mad_priv) {
2695 				ret = -ENOMEM;
2696 				break;
2697 			}
2698 		}
2699 		sg_list.length = mad_priv_dma_size(mad_priv);
2700 		sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2701 						 &mad_priv->grh,
2702 						 mad_priv_dma_size(mad_priv),
2703 						 DMA_FROM_DEVICE);
2704 		if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2705 						  sg_list.addr))) {
2706 			kfree(mad_priv);
2707 			ret = -ENOMEM;
2708 			break;
2709 		}
2710 		mad_priv->header.mapping = sg_list.addr;
2711 		mad_priv->header.mad_list.mad_queue = recv_queue;
2712 		mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2713 		recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2714 
2715 		/* Post receive WR */
2716 		spin_lock_irqsave(&recv_queue->lock, flags);
2717 		post = (++recv_queue->count < recv_queue->max_active);
2718 		list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2719 		spin_unlock_irqrestore(&recv_queue->lock, flags);
2720 		ret = ib_post_recv(qp_info->qp, &recv_wr, NULL);
2721 		if (ret) {
2722 			spin_lock_irqsave(&recv_queue->lock, flags);
2723 			list_del(&mad_priv->header.mad_list.list);
2724 			recv_queue->count--;
2725 			spin_unlock_irqrestore(&recv_queue->lock, flags);
2726 			ib_dma_unmap_single(qp_info->port_priv->device,
2727 					    mad_priv->header.mapping,
2728 					    mad_priv_dma_size(mad_priv),
2729 					    DMA_FROM_DEVICE);
2730 			kfree(mad_priv);
2731 			dev_err(&qp_info->port_priv->device->dev,
2732 				"ib_post_recv failed: %d\n", ret);
2733 			break;
2734 		}
2735 	} while (post);
2736 
2737 	return ret;
2738 }
2739 
2740 /*
2741  * Return all the posted receive MADs
2742  */
2743 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2744 {
2745 	struct ib_mad_private_header *mad_priv_hdr;
2746 	struct ib_mad_private *recv;
2747 	struct ib_mad_list_head *mad_list;
2748 
2749 	if (!qp_info->qp)
2750 		return;
2751 
2752 	while (!list_empty(&qp_info->recv_queue.list)) {
2753 
2754 		mad_list = list_entry(qp_info->recv_queue.list.next,
2755 				      struct ib_mad_list_head, list);
2756 		mad_priv_hdr = container_of(mad_list,
2757 					    struct ib_mad_private_header,
2758 					    mad_list);
2759 		recv = container_of(mad_priv_hdr, struct ib_mad_private,
2760 				    header);
2761 
2762 		/* Remove from posted receive MAD list */
2763 		list_del(&mad_list->list);
2764 
2765 		ib_dma_unmap_single(qp_info->port_priv->device,
2766 				    recv->header.mapping,
2767 				    mad_priv_dma_size(recv),
2768 				    DMA_FROM_DEVICE);
2769 		kfree(recv);
2770 	}
2771 
2772 	qp_info->recv_queue.count = 0;
2773 }
2774 
2775 /*
2776  * Start the port
2777  */
2778 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2779 {
2780 	int ret, i;
2781 	struct ib_qp_attr *attr;
2782 	struct ib_qp *qp;
2783 	u16 pkey_index;
2784 
2785 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
2786 	if (!attr)
2787 		return -ENOMEM;
2788 
2789 	ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2790 			   IB_DEFAULT_PKEY_FULL, &pkey_index);
2791 	if (ret)
2792 		pkey_index = 0;
2793 
2794 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2795 		qp = port_priv->qp_info[i].qp;
2796 		if (!qp)
2797 			continue;
2798 
2799 		/*
2800 		 * PKey index for QP1 is irrelevant but
2801 		 * one is needed for the Reset to Init transition
2802 		 */
2803 		attr->qp_state = IB_QPS_INIT;
2804 		attr->pkey_index = pkey_index;
2805 		attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2806 		ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2807 					     IB_QP_PKEY_INDEX | IB_QP_QKEY);
2808 		if (ret) {
2809 			dev_err(&port_priv->device->dev,
2810 				"Couldn't change QP%d state to INIT: %d\n",
2811 				i, ret);
2812 			goto out;
2813 		}
2814 
2815 		attr->qp_state = IB_QPS_RTR;
2816 		ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2817 		if (ret) {
2818 			dev_err(&port_priv->device->dev,
2819 				"Couldn't change QP%d state to RTR: %d\n",
2820 				i, ret);
2821 			goto out;
2822 		}
2823 
2824 		attr->qp_state = IB_QPS_RTS;
2825 		attr->sq_psn = IB_MAD_SEND_Q_PSN;
2826 		ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2827 		if (ret) {
2828 			dev_err(&port_priv->device->dev,
2829 				"Couldn't change QP%d state to RTS: %d\n",
2830 				i, ret);
2831 			goto out;
2832 		}
2833 	}
2834 
2835 	ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2836 	if (ret) {
2837 		dev_err(&port_priv->device->dev,
2838 			"Failed to request completion notification: %d\n",
2839 			ret);
2840 		goto out;
2841 	}
2842 
2843 	for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2844 		if (!port_priv->qp_info[i].qp)
2845 			continue;
2846 
2847 		ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2848 		if (ret) {
2849 			dev_err(&port_priv->device->dev,
2850 				"Couldn't post receive WRs\n");
2851 			goto out;
2852 		}
2853 	}
2854 out:
2855 	kfree(attr);
2856 	return ret;
2857 }
2858 
2859 static void qp_event_handler(struct ib_event *event, void *qp_context)
2860 {
2861 	struct ib_mad_qp_info	*qp_info = qp_context;
2862 
2863 	/* It's worse than that! He's dead, Jim! */
2864 	dev_err(&qp_info->port_priv->device->dev,
2865 		"Fatal error (%d) on MAD QP (%u)\n",
2866 		event->event, qp_info->qp->qp_num);
2867 }
2868 
2869 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2870 			   struct ib_mad_queue *mad_queue)
2871 {
2872 	mad_queue->qp_info = qp_info;
2873 	mad_queue->count = 0;
2874 	spin_lock_init(&mad_queue->lock);
2875 	INIT_LIST_HEAD(&mad_queue->list);
2876 }
2877 
2878 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2879 			struct ib_mad_qp_info *qp_info)
2880 {
2881 	qp_info->port_priv = port_priv;
2882 	init_mad_queue(qp_info, &qp_info->send_queue);
2883 	init_mad_queue(qp_info, &qp_info->recv_queue);
2884 	INIT_LIST_HEAD(&qp_info->overflow_list);
2885 }
2886 
2887 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2888 			 enum ib_qp_type qp_type)
2889 {
2890 	struct ib_qp_init_attr	qp_init_attr;
2891 	int ret;
2892 
2893 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
2894 	qp_init_attr.send_cq = qp_info->port_priv->cq;
2895 	qp_init_attr.recv_cq = qp_info->port_priv->cq;
2896 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2897 	qp_init_attr.cap.max_send_wr = mad_sendq_size;
2898 	qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2899 	qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2900 	qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2901 	qp_init_attr.qp_type = qp_type;
2902 	qp_init_attr.port_num = qp_info->port_priv->port_num;
2903 	qp_init_attr.qp_context = qp_info;
2904 	qp_init_attr.event_handler = qp_event_handler;
2905 	qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2906 	if (IS_ERR(qp_info->qp)) {
2907 		dev_err(&qp_info->port_priv->device->dev,
2908 			"Couldn't create ib_mad QP%d\n",
2909 			get_spl_qp_index(qp_type));
2910 		ret = PTR_ERR(qp_info->qp);
2911 		goto error;
2912 	}
2913 	/* Use minimum queue sizes unless the CQ is resized */
2914 	qp_info->send_queue.max_active = mad_sendq_size;
2915 	qp_info->recv_queue.max_active = mad_recvq_size;
2916 	return 0;
2917 
2918 error:
2919 	return ret;
2920 }
2921 
2922 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2923 {
2924 	if (!qp_info->qp)
2925 		return;
2926 
2927 	ib_destroy_qp(qp_info->qp);
2928 }
2929 
2930 /*
2931  * Open the port
2932  * Create the QP, PD, MR, and CQ if needed
2933  */
2934 static int ib_mad_port_open(struct ib_device *device,
2935 			    u32 port_num)
2936 {
2937 	int ret, cq_size;
2938 	struct ib_mad_port_private *port_priv;
2939 	unsigned long flags;
2940 	char name[sizeof "ib_mad123"];
2941 	int has_smi;
2942 
2943 	if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
2944 		return -EFAULT;
2945 
2946 	if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
2947 		    rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
2948 		return -EFAULT;
2949 
2950 	/* Create new device info */
2951 	port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2952 	if (!port_priv)
2953 		return -ENOMEM;
2954 
2955 	port_priv->device = device;
2956 	port_priv->port_num = port_num;
2957 	spin_lock_init(&port_priv->reg_lock);
2958 	init_mad_qp(port_priv, &port_priv->qp_info[0]);
2959 	init_mad_qp(port_priv, &port_priv->qp_info[1]);
2960 
2961 	cq_size = mad_sendq_size + mad_recvq_size;
2962 	has_smi = rdma_cap_ib_smi(device, port_num);
2963 	if (has_smi)
2964 		cq_size *= 2;
2965 
2966 	port_priv->pd = ib_alloc_pd(device, 0);
2967 	if (IS_ERR(port_priv->pd)) {
2968 		dev_err(&device->dev, "Couldn't create ib_mad PD\n");
2969 		ret = PTR_ERR(port_priv->pd);
2970 		goto error3;
2971 	}
2972 
2973 	port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
2974 			IB_POLL_UNBOUND_WORKQUEUE);
2975 	if (IS_ERR(port_priv->cq)) {
2976 		dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
2977 		ret = PTR_ERR(port_priv->cq);
2978 		goto error4;
2979 	}
2980 
2981 	if (has_smi) {
2982 		ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2983 		if (ret)
2984 			goto error6;
2985 	}
2986 	ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2987 	if (ret)
2988 		goto error7;
2989 
2990 	snprintf(name, sizeof(name), "ib_mad%u", port_num);
2991 	port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2992 	if (!port_priv->wq) {
2993 		ret = -ENOMEM;
2994 		goto error8;
2995 	}
2996 
2997 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2998 	list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2999 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3000 
3001 	ret = ib_mad_port_start(port_priv);
3002 	if (ret) {
3003 		dev_err(&device->dev, "Couldn't start port\n");
3004 		goto error9;
3005 	}
3006 
3007 	return 0;
3008 
3009 error9:
3010 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3011 	list_del_init(&port_priv->port_list);
3012 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3013 
3014 	destroy_workqueue(port_priv->wq);
3015 error8:
3016 	destroy_mad_qp(&port_priv->qp_info[1]);
3017 error7:
3018 	destroy_mad_qp(&port_priv->qp_info[0]);
3019 error6:
3020 	ib_free_cq(port_priv->cq);
3021 	cleanup_recv_queue(&port_priv->qp_info[1]);
3022 	cleanup_recv_queue(&port_priv->qp_info[0]);
3023 error4:
3024 	ib_dealloc_pd(port_priv->pd);
3025 error3:
3026 	kfree(port_priv);
3027 
3028 	return ret;
3029 }
3030 
3031 /*
3032  * Close the port
3033  * If there are no classes using the port, free the port
3034  * resources (CQ, MR, PD, QP) and remove the port's info structure
3035  */
3036 static int ib_mad_port_close(struct ib_device *device, u32 port_num)
3037 {
3038 	struct ib_mad_port_private *port_priv;
3039 	unsigned long flags;
3040 
3041 	spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3042 	port_priv = __ib_get_mad_port(device, port_num);
3043 	if (port_priv == NULL) {
3044 		spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3045 		dev_err(&device->dev, "Port %u not found\n", port_num);
3046 		return -ENODEV;
3047 	}
3048 	list_del_init(&port_priv->port_list);
3049 	spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3050 
3051 	destroy_workqueue(port_priv->wq);
3052 	destroy_mad_qp(&port_priv->qp_info[1]);
3053 	destroy_mad_qp(&port_priv->qp_info[0]);
3054 	ib_free_cq(port_priv->cq);
3055 	ib_dealloc_pd(port_priv->pd);
3056 	cleanup_recv_queue(&port_priv->qp_info[1]);
3057 	cleanup_recv_queue(&port_priv->qp_info[0]);
3058 	/* XXX: Handle deallocation of MAD registration tables */
3059 
3060 	kfree(port_priv);
3061 
3062 	return 0;
3063 }
3064 
3065 static int ib_mad_init_device(struct ib_device *device)
3066 {
3067 	int start, i;
3068 	unsigned int count = 0;
3069 	int ret;
3070 
3071 	start = rdma_start_port(device);
3072 
3073 	for (i = start; i <= rdma_end_port(device); i++) {
3074 		if (!rdma_cap_ib_mad(device, i))
3075 			continue;
3076 
3077 		ret = ib_mad_port_open(device, i);
3078 		if (ret) {
3079 			dev_err(&device->dev, "Couldn't open port %d\n", i);
3080 			goto error;
3081 		}
3082 		ret = ib_agent_port_open(device, i);
3083 		if (ret) {
3084 			dev_err(&device->dev,
3085 				"Couldn't open port %d for agents\n", i);
3086 			goto error_agent;
3087 		}
3088 		count++;
3089 	}
3090 	if (!count)
3091 		return -EOPNOTSUPP;
3092 
3093 	return 0;
3094 
3095 error_agent:
3096 	if (ib_mad_port_close(device, i))
3097 		dev_err(&device->dev, "Couldn't close port %d\n", i);
3098 
3099 error:
3100 	while (--i >= start) {
3101 		if (!rdma_cap_ib_mad(device, i))
3102 			continue;
3103 
3104 		if (ib_agent_port_close(device, i))
3105 			dev_err(&device->dev,
3106 				"Couldn't close port %d for agents\n", i);
3107 		if (ib_mad_port_close(device, i))
3108 			dev_err(&device->dev, "Couldn't close port %d\n", i);
3109 	}
3110 	return ret;
3111 }
3112 
3113 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3114 {
3115 	unsigned int i;
3116 
3117 	rdma_for_each_port (device, i) {
3118 		if (!rdma_cap_ib_mad(device, i))
3119 			continue;
3120 
3121 		if (ib_agent_port_close(device, i))
3122 			dev_err(&device->dev,
3123 				"Couldn't close port %u for agents\n", i);
3124 		if (ib_mad_port_close(device, i))
3125 			dev_err(&device->dev, "Couldn't close port %u\n", i);
3126 	}
3127 }
3128 
3129 static struct ib_client mad_client = {
3130 	.name   = "mad",
3131 	.add = ib_mad_init_device,
3132 	.remove = ib_mad_remove_device
3133 };
3134 
3135 int ib_mad_init(void)
3136 {
3137 	mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3138 	mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3139 
3140 	mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3141 	mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3142 
3143 	INIT_LIST_HEAD(&ib_mad_port_list);
3144 
3145 	if (ib_register_client(&mad_client)) {
3146 		pr_err("Couldn't register ib_mad client\n");
3147 		return -EINVAL;
3148 	}
3149 
3150 	return 0;
3151 }
3152 
3153 void ib_mad_cleanup(void)
3154 {
3155 	ib_unregister_client(&mad_client);
3156 }
3157