xref: /titanic_50/usr/src/uts/common/io/ib/clients/of/sol_ofs/sol_cma.c (revision fffafeb2cc01732fd6a28ed530e4424094685ece)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 /*
27  * sol_cma is a part of sol_ofs misc module. This file
28  * provides interfaces for supporting the communication
29  * management API defined in "rdma_cm.h". In-Kernel
30  * consumers of the "rdma_cm.h" API should link sol_ofs
31  * misc module using :
32  *	-N misc/sol_ofs
33  * Solaris uCMA (sol_ucma) driver is the current consumer for
34  * sol_cma.
35  */
36 
37 /* Standard driver includes */
38 #include <sys/types.h>
39 #include <sys/modctl.h>
40 #include <sys/errno.h>
41 #include <sys/stat.h>
42 #include <sys/ddi.h>
43 #include <sys/sunddi.h>
44 #include <sys/modctl.h>
45 
46 #include <sys/ib/clients/of/ofed_kernel.h>
47 #include <sys/ib/clients/of/rdma/ib_addr.h>
48 
49 #include <sys/ib/clients/of/sol_ofs/sol_cma.h>
50 #include <sys/ib/clients/of/sol_ofs/sol_kverb_impl.h>
51 
52 /* Modload support */
53 static struct modlmisc sol_ofs_modmisc	= {
54 	&mod_miscops,
55 	"Solaris OFS Misc module"
56 };
57 
58 struct modlinkage sol_ofs_modlinkage = {
59 	MODREV_1,
60 	(void *)&sol_ofs_modmisc,
61 	NULL
62 };
63 
64 static ib_client_t	*sol_cma_ib_client;
65 sol_cma_glbl_listen_t	sol_cma_glbl_listen;
66 avl_tree_t		sol_cma_glbl_listen_tree;
67 
68 static void		sol_cma_add_dev(struct ib_device *);
69 static void		sol_cma_rem_dev(struct ib_device *);
70 
71 static llist_head_t	sol_cma_dev_list = LLIST_HEAD_INIT(sol_cma_dev_list);
72 kmutex_t		sol_cma_dev_mutex;
73 kmutex_t		sol_cma_glob_mutex;
74 
75 char	*sol_rdmacm_dbg_str = "sol_rdmacm";
76 char	*sol_ofs_dbg_str = "sol_ofs_mod";
77 
78 /*
79  * Local functions defines.
80  */
81 int sol_cma_req_cmid_cmp(const void *p1, const void *p2);
82 int sol_cma_cmid_cmp(const void *p1, const void *p2);
83 int sol_cma_svc_cmp(const void *, const void *);
84 
85 static struct rdma_cm_id *cma_alloc_chan(rdma_cm_event_handler,
86     void *, enum rdma_port_space);
87 static void cma_set_chan_state(sol_cma_chan_t *, cma_chan_state_t);
88 static int cma_cas_chan_state(sol_cma_chan_t *, cma_chan_state_t,
89     cma_chan_state_t);
90 static void cma_free_listen_list(struct rdma_cm_id *);
91 static void cma_destroy_id(struct rdma_cm_id *);
92 static void cma_handle_nomore_events(sol_cma_chan_t *);
93 
94 extern void sol_ofs_dprintf_init();
95 extern void sol_ofs_dprintf_fini();
96 
97 cma_chan_state_t cma_get_chan_state(sol_cma_chan_t *);
98 extern int ibcma_init_root_chan(sol_cma_chan_t *, sol_cma_glbl_listen_t *);
99 extern int ibcma_fini_root_chan(sol_cma_chan_t *);
100 extern void ibcma_copy_srv_hdl(sol_cma_chan_t *, sol_cma_glbl_listen_t *);
101 extern int ibcma_fini_ep_chan(sol_cma_chan_t *);
102 extern uint64_t ibcma_init_root_sid(sol_cma_chan_t *);
103 extern void rdma_ib_destroy_id(struct rdma_cm_id *);
104 extern int rdma_ib_bind_addr(struct rdma_cm_id *, struct sockaddr *);
105 extern int rdma_ib_resolve_addr(struct rdma_cm_id *, struct sockaddr *,
106     struct sockaddr *, int);
107 extern int rdma_ib_resolve_route(struct rdma_cm_id *, int);
108 extern int rdma_ib_init_qp_attr(struct rdma_cm_id *, struct ib_qp_attr *,
109     int *);
110 extern int rdma_ib_connect(struct rdma_cm_id *, struct rdma_conn_param *);
111 extern int rdma_ib_listen(struct rdma_cm_id *, int);
112 extern int rdma_ib_accept(struct rdma_cm_id *, struct rdma_conn_param *);
113 extern int rdma_ib_reject(struct rdma_cm_id *, const void *, uint8_t);
114 extern int rdma_ib_disconnect(struct rdma_cm_id *);
115 extern int rdma_ib_join_multicast(struct rdma_cm_id *, struct sockaddr *,
116     void *);
117 extern void rdma_ib_leave_multicast(struct rdma_cm_id *, struct sockaddr *);
118 
119 int
_init(void)120 _init(void)
121 {
122 	int		err;
123 
124 	sol_ofs_dprintf_init();
125 	SOL_OFS_DPRINTF_L5(sol_ofs_dbg_str, "_init()");
126 
127 	mutex_init(&sol_cma_glob_mutex, NULL, MUTEX_DRIVER, NULL);
128 	mutex_init(&sol_cma_dev_mutex, NULL, MUTEX_DRIVER, NULL);
129 	avl_create(&sol_cma_glbl_listen_tree,
130 	    sol_cma_svc_cmp, sizeof (sol_cma_glbl_listen_t),
131 	    offsetof(sol_cma_glbl_listen_t, cma_listen_node));
132 
133 	sol_cma_ib_client = kmem_zalloc(sizeof (ib_client_t), KM_NOSLEEP);
134 	if (!sol_cma_ib_client) {
135 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str,
136 		    "_init() - mem alloc failed");
137 		avl_destroy(&sol_cma_glbl_listen_tree);
138 		mutex_destroy(&sol_cma_dev_mutex);
139 		mutex_destroy(&sol_cma_glob_mutex);
140 		sol_ofs_dprintf_fini();
141 		return (ENOMEM);
142 	}
143 
144 	sol_cma_ib_client->name = "sol_ofs";
145 	sol_cma_ib_client->add = sol_cma_add_dev;
146 	sol_cma_ib_client->remove = sol_cma_rem_dev;
147 	sol_cma_ib_client->dip = NULL;
148 
149 	if ((err = ib_register_client(sol_cma_ib_client)) != 0) {
150 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str,
151 		    "_init() ib_register_client() failed with err %d",
152 		    err);
153 		kmem_free(sol_cma_ib_client, sizeof (ib_client_t));
154 		avl_destroy(&sol_cma_glbl_listen_tree);
155 		mutex_destroy(&sol_cma_dev_mutex);
156 		mutex_destroy(&sol_cma_glob_mutex);
157 		sol_ofs_dprintf_fini();
158 		return (err);
159 	}
160 
161 	if ((err = mod_install(&sol_ofs_modlinkage)) != 0) {
162 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str,
163 		    "_init() - mod_install() failed");
164 		ib_unregister_client(sol_cma_ib_client);
165 		kmem_free(sol_cma_ib_client, sizeof (ib_client_t));
166 		avl_destroy(&sol_cma_glbl_listen_tree);
167 		mutex_destroy(&sol_cma_dev_mutex);
168 		mutex_destroy(&sol_cma_glob_mutex);
169 		sol_ofs_dprintf_fini();
170 		return (err);
171 	}
172 
173 	SOL_OFS_DPRINTF_L5(sol_ofs_dbg_str, "_init() - ret");
174 	return (err);
175 }
176 
177 int
_fini(void)178 _fini(void)
179 {
180 	int		err;
181 
182 	SOL_OFS_DPRINTF_L5(sol_ofs_dbg_str, "_fini()");
183 
184 	if (avl_numnodes(&sol_cma_glbl_listen_tree)) {
185 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str, "_fini - "
186 		    "listen CMIDs still active");
187 		return (EBUSY);
188 	}
189 	if ((err = mod_remove(&sol_ofs_modlinkage)) != 0) {
190 		SOL_OFS_DPRINTF_L3(sol_ofs_dbg_str,
191 		    "_fini: mod_remove failed");
192 		return (err);
193 	}
194 
195 	ib_unregister_client(sol_cma_ib_client);
196 	kmem_free(sol_cma_ib_client, sizeof (ib_client_t));
197 	avl_destroy(&sol_cma_glbl_listen_tree);
198 	mutex_destroy(&sol_cma_dev_mutex);
199 	mutex_destroy(&sol_cma_glob_mutex);
200 	SOL_OFS_DPRINTF_L5(sol_ofs_dbg_str, "_fini() - ret");
201 	sol_ofs_dprintf_fini();
202 	return (err);
203 }
204 
205 int
_info(struct modinfo * modinfop)206 _info(struct modinfo *modinfop)
207 {
208 	return (mod_info(&sol_ofs_modlinkage, modinfop));
209 }
210 
211 typedef struct cma_device {
212 	kmutex_t		cma_mutex;
213 	/* Ptr in the global sol_cma_dev_list */
214 	llist_head_t		cma_list;
215 	/* List of listeners for this device */
216 	genlist_t		cma_epchan_list;
217 	struct ib_device	*cma_device;
218 	uint_t			cma_ref_count;
219 	enum {
220 		SOL_CMA_DEV_ADDED,
221 		SOL_CMA_DEV_REM_IN_PROGRESS
222 	} cma_dev_state;
223 } cma_device_t;
224 
225 static void
sol_cma_add_dev(struct ib_device * dev)226 sol_cma_add_dev(struct ib_device *dev)
227 {
228 	cma_device_t	*new_device;
229 
230 	new_device = kmem_zalloc(sizeof (cma_device_t), KM_NOSLEEP);
231 	if (!new_device) {
232 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str, "sol_cma_add_dev() "
233 		    "alloc failed!!");
234 		return;
235 	}
236 	mutex_init(&new_device->cma_mutex, NULL, MUTEX_DRIVER, NULL);
237 	llist_head_init(&new_device->cma_list, new_device);
238 	init_genlist(&new_device->cma_epchan_list);
239 	new_device->cma_device = dev;
240 
241 	ib_set_client_data(dev, sol_cma_ib_client, new_device);
242 
243 	mutex_enter(&sol_cma_dev_mutex);
244 	llist_add_tail(&new_device->cma_list, &sol_cma_dev_list);
245 	mutex_exit(&sol_cma_dev_mutex);
246 }
247 
248 static void
sol_cma_rem_dev(struct ib_device * dev)249 sol_cma_rem_dev(struct ib_device *dev)
250 {
251 	cma_device_t	*rem_device;
252 	genlist_entry_t	*entry;
253 
254 	SOL_OFS_DPRINTF_L5(sol_ofs_dbg_str, "sol_rem_dev(%p)", dev);
255 
256 	rem_device = (cma_device_t *)ib_get_client_data(dev, sol_cma_ib_client);
257 	if (!rem_device) {
258 		SOL_OFS_DPRINTF_L2(sol_ofs_dbg_str, "sol_cma_rem_dev() "
259 		    "NULL cma_dev!!");
260 		return;
261 	}
262 
263 	mutex_enter(&rem_device->cma_mutex);
264 	rem_device->cma_dev_state = SOL_CMA_DEV_REM_IN_PROGRESS;
265 	if (rem_device->cma_ref_count) {
266 		mutex_exit(&rem_device->cma_mutex);
267 		SOL_OFS_DPRINTF_L3(sol_ofs_dbg_str, "sol_cma_rem_dev() "
268 		    "BUSY cma_dev!!");
269 		return;
270 	}
271 	entry = remove_genlist_head(&rem_device->cma_epchan_list);
272 	while (entry) {
273 		sol_cma_chan_t	*ep_chanp;
274 
275 		ep_chanp = (sol_cma_chan_t *)entry->data;
276 		if (ibcma_fini_ep_chan(ep_chanp) == 0) {
277 			genlist_entry_t	*entry1;
278 			sol_cma_chan_t	*root_chanp;
279 
280 			ASSERT(ep_chanp->chan_listenp);
281 			entry1 = ep_chanp->chan_listenp->listen_ep_root_entry;
282 			root_chanp = (sol_cma_chan_t *)ep_chanp->listen_root;
283 			root_chanp->chan_listenp->listen_eps--;
284 			delete_genlist(&root_chanp->chan_listenp->listen_list,
285 			    entry1);
286 
287 			kmem_free(ep_chanp, sizeof (sol_cma_chan_t));
288 			kmem_free(entry, sizeof (genlist_entry_t));
289 		}
290 
291 		entry = remove_genlist_head(&rem_device->cma_epchan_list);
292 	}
293 	mutex_exit(&rem_device->cma_mutex);
294 
295 	mutex_enter(&sol_cma_dev_mutex);
296 	llist_del(&rem_device->cma_list);
297 	mutex_exit(&sol_cma_dev_mutex);
298 
299 	kmem_free(rem_device, sizeof (cma_device_t));
300 }
301 
302 struct ib_device *
sol_cma_acquire_device(ib_guid_t hca_guid)303 sol_cma_acquire_device(ib_guid_t hca_guid)
304 {
305 	llist_head_t	*entry;
306 	cma_device_t	*cma_devp;
307 
308 	mutex_enter(&sol_cma_dev_mutex);
309 	list_for_each(entry, &sol_cma_dev_list) {
310 		cma_devp = (cma_device_t *)entry->ptr;
311 
312 		if (cma_devp->cma_device->node_guid != hca_guid)
313 			continue;
314 
315 		mutex_enter(&cma_devp->cma_mutex);
316 		if (cma_devp->cma_dev_state == SOL_CMA_DEV_REM_IN_PROGRESS) {
317 			SOL_OFS_DPRINTF_L3(sol_ofs_dbg_str,
318 			    "sol_cma_acquire_dev() - Device getting removed!!");
319 			mutex_exit(&cma_devp->cma_mutex);
320 			mutex_exit(&sol_cma_dev_mutex);
321 			return (NULL);
322 		}
323 		cma_devp->cma_ref_count++;
324 		mutex_exit(&cma_devp->cma_mutex);
325 		mutex_exit(&sol_cma_dev_mutex);
326 		return (cma_devp->cma_device);
327 
328 	}
329 	mutex_exit(&sol_cma_dev_mutex);
330 	return (NULL);
331 }
332 
333 static void
sol_cma_release_device(struct rdma_cm_id * id)334 sol_cma_release_device(struct rdma_cm_id *id)
335 {
336 	ib_device_t	*device = id->device;
337 	llist_head_t	*entry;
338 	cma_device_t	*cma_devp;
339 
340 	mutex_enter(&sol_cma_dev_mutex);
341 	list_for_each(entry, &sol_cma_dev_list) {
342 		cma_devp = (cma_device_t *)entry->ptr;
343 
344 		if (cma_devp->cma_device != device)
345 			continue;
346 
347 		mutex_enter(&cma_devp->cma_mutex);
348 		cma_devp->cma_ref_count--;
349 		if (cma_devp->cma_dev_state == SOL_CMA_DEV_REM_IN_PROGRESS &&
350 		    cma_devp->cma_ref_count == 0) {
351 			SOL_OFS_DPRINTF_L3(sol_ofs_dbg_str,
352 			    "sol_cma_release_dev() - Device free removed!!");
353 			mutex_exit(&cma_devp->cma_mutex);
354 			llist_del(&cma_devp->cma_list);
355 			kmem_free(cma_devp, sizeof (cma_device_t));
356 			mutex_exit(&sol_cma_dev_mutex);
357 			return;
358 		}
359 		mutex_exit(&cma_devp->cma_mutex);
360 	}
361 	mutex_exit(&sol_cma_dev_mutex);
362 }
363 
364 void
sol_cma_add_hca_list(sol_cma_chan_t * ep_chanp,ib_guid_t hca_guid)365 sol_cma_add_hca_list(sol_cma_chan_t *ep_chanp, ib_guid_t hca_guid)
366 {
367 	llist_head_t	*entry;
368 	cma_device_t	*cma_devp;
369 
370 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "add_hca_list(%p, %llx)",
371 	    ep_chanp, hca_guid);
372 	mutex_enter(&sol_cma_dev_mutex);
373 	list_for_each(entry, &sol_cma_dev_list) {
374 		cma_devp = (cma_device_t *)entry->ptr;
375 
376 		if ((cma_devp->cma_device)->node_guid != hca_guid)
377 			continue;
378 
379 		mutex_enter(&cma_devp->cma_mutex);
380 		ep_chanp->chan_listenp->listen_ep_dev_entry =
381 		    add_genlist(&cma_devp->cma_epchan_list,
382 		    (uintptr_t)ep_chanp, NULL);
383 		ep_chanp->chan_listenp->listen_ep_device = cma_devp->cma_device;
384 		mutex_exit(&cma_devp->cma_mutex);
385 		mutex_exit(&sol_cma_dev_mutex);
386 		return;
387 	}
388 	mutex_exit(&sol_cma_dev_mutex);
389 	SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str, "add_hca_list(%p, %llx): "
390 	    "No matching HCA in list!!", ep_chanp, hca_guid);
391 }
392 
393 /*
394  * rdma_cm.h API functions.
395  */
396 struct rdma_cm_id *
rdma_create_id(rdma_cm_event_handler evt_hdlr,void * context,enum rdma_port_space ps)397 rdma_create_id(rdma_cm_event_handler evt_hdlr, void *context,
398     enum rdma_port_space ps)
399 {
400 	struct rdma_cm_id 	*rdma_idp;
401 
402 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_create_id(%p, %p, %x)",
403 	    evt_hdlr, context, ps);
404 
405 	if (ps != RDMA_PS_TCP && ps != RDMA_PS_UDP && ps != RDMA_PS_IPOIB) {
406 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
407 		    "rdma_create_id: unsupported protocol %x", ps);
408 		return (NULL);
409 	}
410 
411 	rdma_idp = cma_alloc_chan(evt_hdlr, context, ps);
412 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
413 	    "rdma_create_id : ret %p", rdma_idp);
414 
415 	return (rdma_idp);
416 }
417 
418 void
rdma_map_id2clnthdl(struct rdma_cm_id * rdma_idp,void * ib_client_hdl,void * iw_client_hdl)419 rdma_map_id2clnthdl(struct rdma_cm_id *rdma_idp, void *ib_client_hdl,
420     void *iw_client_hdl)
421 {
422 	sol_cma_chan_t	*chanp = (sol_cma_chan_t *)rdma_idp;
423 
424 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
425 	    "rdma_map_id2clnthdl(%p, %p, %p)",
426 	    rdma_idp, ib_client_hdl, iw_client_hdl);
427 	ASSERT(ib_client_hdl != NULL || iw_client_hdl != NULL);
428 	chanp->chan_ib_client_hdl = ib_client_hdl;
429 	chanp->chan_iw_client_hdl = iw_client_hdl;
430 }
431 
432 void
rdma_map_id2qphdl(struct rdma_cm_id * rdma_idp,void * qp_hdl)433 rdma_map_id2qphdl(struct rdma_cm_id *rdma_idp, void *qp_hdl)
434 {
435 	sol_cma_chan_t	*chanp = (sol_cma_chan_t *)rdma_idp;
436 
437 	ASSERT(rdma_idp);
438 	ASSERT(qp_hdl);
439 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_mapid2qphdl(%p, %p)",
440 	    rdma_idp, qp_hdl);
441 	chanp->chan_qp_hdl = qp_hdl;
442 }
443 
444 
445 void
rdma_destroy_id(struct rdma_cm_id * rdma_idp)446 rdma_destroy_id(struct rdma_cm_id *rdma_idp)
447 {
448 	sol_cma_chan_t		*chanp, *root_chanp;
449 	cma_chan_state_t	state;
450 	int			rc, is_root_cmid, do_wait, is_passive;
451 
452 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id(%p)", rdma_idp);
453 
454 	if (!rdma_idp)
455 		return;
456 
457 	is_root_cmid = do_wait = is_passive = 0;
458 
459 	chanp = (sol_cma_chan_t *)rdma_idp;
460 	root_chanp = (sol_cma_chan_t *)chanp->listen_root;
461 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id(%p), %p",
462 	    rdma_idp, root_chanp);
463 
464 	mutex_enter(&chanp->chan_mutex);
465 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_CMID_DESTROYED;
466 
467 	/*
468 	 * Wait in destroy of CMID when rdma_resolve_addr() / rdma_listen()
469 	 * rdma_resolve_route() API is in progress.
470 	 */
471 	while (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_API_PROGRESS)
472 		cv_wait(&chanp->chan_destroy_cv, &chanp->chan_mutex);
473 
474 	/* Wait if Event is been notified to consumer */
475 	while (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_EVENT_PROGRESS)
476 		cv_wait(&chanp->chan_destroy_cv, &chanp->chan_mutex);
477 
478 	if (rdma_idp->device)
479 		sol_cma_release_device(rdma_idp);
480 
481 	if (chanp->chan_listenp && chanp->chan_listenp->listen_is_root)
482 		is_root_cmid = 1;
483 	if (root_chanp == NULL && is_root_cmid == 0)
484 		is_passive = 1;
485 
486 	/*
487 	 * Skip Active side handling for passive CMIDs and listen CMID
488 	 * for which REQ CMIDs have not been created.
489 	 */
490 	if (is_passive || (is_root_cmid && chanp->chan_req_state !=
491 	    REQ_CMID_QUEUED)) {
492 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id: "
493 		    "Skipping passive %p, %x, %x", chanp->chan_listenp,
494 		    is_root_cmid, chanp->chan_req_state);
495 		goto skip_passive_handling;
496 	}
497 
498 	/*
499 	 * destroy_id() called for listening CMID and there are REQ
500 	 * CMIDs not yet notified. Reject such CMIDs and decrement
501 	 * the count.
502 	 */
503 	if (is_root_cmid && chanp->chan_req_cnt) {
504 		sol_cma_chan_t	*req_cmid_chan, *next_chan;
505 
506 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id: "
507 		    "not notified handling");
508 		for (req_cmid_chan = (sol_cma_chan_t *)avl_first(
509 		    &chanp->chan_req_avl_tree); req_cmid_chan &&
510 		    chanp->chan_req_cnt; req_cmid_chan = next_chan) {
511 			next_chan = AVL_NEXT(
512 			    &chanp->chan_req_avl_tree, req_cmid_chan);
513 			if (req_cmid_chan->chan_req_state ==
514 			    REQ_CMID_NOTIFIED) {
515 				avl_remove(&chanp->chan_req_avl_tree,
516 				    req_cmid_chan);
517 				chanp->chan_req_cnt--;
518 				chanp->chan_req_total_cnt--;
519 				mutex_exit(&chanp->chan_mutex);
520 				mutex_enter(&req_cmid_chan->chan_mutex);
521 				req_cmid_chan->chan_req_state =
522 				    REQ_CMID_SERVER_NONE;
523 				if (rdma_idp->ps == RDMA_PS_TCP)
524 					cma_set_chan_state(req_cmid_chan,
525 					    SOL_CMA_CHAN_DESTROY_PENDING);
526 				mutex_exit(&req_cmid_chan->chan_mutex);
527 				(void) rdma_disconnect(
528 				    (struct rdma_cm_id *)req_cmid_chan);
529 				mutex_enter(&chanp->chan_mutex);
530 				if (rdma_idp->ps == RDMA_PS_TCP) {
531 					mutex_enter(
532 					    &req_cmid_chan->chan_mutex);
533 					req_cmid_chan->listen_root =
534 					    rdma_idp;
535 					mutex_exit(
536 					    &req_cmid_chan->chan_mutex);
537 				} else {
538 					mutex_destroy(
539 					    &req_cmid_chan->chan_mutex);
540 					cv_destroy(
541 					    &req_cmid_chan->chan_destroy_cv);
542 					kmem_free(req_cmid_chan,
543 					    sizeof (sol_cma_chan_t));
544 				}
545 			}
546 		}
547 	}
548 
549 	/*
550 	 * destroy_id() called for :
551 	 * 	listening CMID and all REQ CMIDs destroy_id() called
552 	 *	REQ CMID and 1 more REQ CMID not yet destroyed.
553 	 * wait till the CMID is completly destroyed.
554 	 */
555 	if (is_root_cmid && chanp->chan_req_total_cnt == 0) {
556 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id: "
557 		    "root idp waiting");
558 		cma_set_chan_state(chanp, SOL_CMA_CHAN_DESTROY_WAIT);
559 		cv_wait(&chanp->chan_destroy_cv, &chanp->chan_mutex);
560 	}
561 	mutex_exit(&chanp->chan_mutex);
562 
563 	if (root_chanp)
564 		mutex_enter(&root_chanp->chan_mutex);
565 	mutex_enter(&chanp->chan_mutex);
566 #ifdef	DEBUG
567 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id: "
568 	    "root_idp %p, cnt %x, state %x", root_chanp,
569 	    root_chanp ? root_chanp->chan_req_total_cnt : 0,
570 	    root_chanp ? cma_get_chan_state(root_chanp) : 0);
571 #endif
572 
573 	if (root_chanp && root_chanp->chan_req_total_cnt == 1 &&
574 	    cma_get_chan_state(root_chanp) == SOL_CMA_CHAN_DESTROY_PENDING)
575 		do_wait = 1;
576 	if (root_chanp)
577 		mutex_exit(&root_chanp->chan_mutex);
578 
579 skip_passive_handling :
580 	state = cma_get_chan_state(chanp);
581 	if (is_root_cmid == 0 && state != SOL_CMA_CHAN_DISCONNECT &&
582 	    SOL_CMAID_CONNECTED(chanp)) {
583 		/*
584 		 * A connected CM ID has not been disconnected.
585 		 * Call rdma_disconnect() to disconnect it.
586 		 */
587 		mutex_exit(&chanp->chan_mutex);
588 		rc = rdma_disconnect(rdma_idp);
589 		if (rc) {
590 			SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
591 			    "rdma_destroy_id(%p)- disconnect failed!!",
592 			    rdma_idp);
593 			return;
594 		}
595 		mutex_enter(&chanp->chan_mutex);
596 		if (root_chanp && chanp->listen_root == NULL)
597 			chanp->listen_root = (struct rdma_cm_id *)root_chanp;
598 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
599 		    "rdma_destroy_id(chanp %p, connect %x, ps %x)",
600 		    chanp, chanp->chan_connect_flag, rdma_idp->ps);
601 		if (SOL_CMAID_CONNECTED(chanp)) {
602 			if (do_wait) {
603 				cma_set_chan_state(chanp,
604 				    SOL_CMA_CHAN_DESTROY_WAIT);
605 				cv_wait(&chanp->chan_destroy_cv,
606 				    &chanp->chan_mutex);
607 				mutex_exit(&chanp->chan_mutex);
608 				cma_destroy_id(rdma_idp);
609 			} else {
610 				cma_set_chan_state(chanp,
611 				    SOL_CMA_CHAN_DESTROY_PENDING);
612 				mutex_exit(&chanp->chan_mutex);
613 			}
614 		} else {
615 			/*
616 			 * No more callbacks are expected for this CMID.
617 			 * Free this CMID.
618 			 */
619 			mutex_exit(&chanp->chan_mutex);
620 			cma_destroy_id(rdma_idp);
621 		}
622 	} else if (is_root_cmid == 0 && state ==
623 	    SOL_CMA_CHAN_DISCONNECT && SOL_CMAID_CONNECTED(chanp)) {
624 		/*
625 		 * CM ID was connected and disconnect is process.
626 		 * Free of this CM ID is done for the DISCONNECT
627 		 * notification for this CMID.
628 		 */
629 		cma_set_chan_state(chanp, SOL_CMA_CHAN_DESTROY_PENDING);
630 		mutex_exit(&chanp->chan_mutex);
631 	} else if (state != SOL_CMA_CHAN_DESTROY_PENDING) {
632 		/* CM ID, not connected, just free it. */
633 		mutex_exit(&chanp->chan_mutex);
634 		cma_destroy_id(rdma_idp);
635 	} else
636 		mutex_exit(&chanp->chan_mutex);
637 
638 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_destroy_id: ret");
639 }
640 
641 /*
642  * State transitions for Address resolution :
643  *	Active Side (Client) :
644  *	1. CREATE_ID-->BIND_ADDR-->RESOLVE_ADDR-->RESOLVE_ROUTE
645  *
646  *	Passive Side (Server) :
647  *	2. CREATE_ID-->RESOLVE_ADDR-->RESOLVE_ROUTE
648  *	IF_ADDR_ANY can be passed as local address in RESOLVE_ADDR
649  */
650 int
rdma_bind_addr(struct rdma_cm_id * idp,struct sockaddr * addr)651 rdma_bind_addr(struct rdma_cm_id *idp, struct sockaddr *addr)
652 {
653 	sol_cma_chan_t		*chanp;
654 	struct rdma_addr	*addrp;
655 	int			ret;
656 
657 	ASSERT(idp);
658 	ASSERT(addr);
659 	chanp = (sol_cma_chan_t *)idp;
660 	addrp = &(idp->route.addr);
661 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_bind_addr(%p, %p)",
662 	    idp, addr);
663 
664 	mutex_enter(&chanp->chan_mutex);
665 	ret = cma_cas_chan_state(chanp, SOL_CMA_CHAN_IDLE, SOL_CMA_CHAN_BOUND);
666 	if (ret) {
667 		mutex_exit(&chanp->chan_mutex);
668 		return (ret);
669 	}
670 	/* Copy the local address to rdma_id structure */
671 	bcopy((void *)addr, (void *)&(addrp->src_addr),
672 	    sizeof (struct sockaddr));
673 	mutex_exit(&chanp->chan_mutex);
674 
675 	/*
676 	 * First call rdma_ib_bind_addr() to bind this address.
677 	 * Next call rdma_iw_bind_addr() to bind this address.
678 	 * For IF_ADDR_ANY, IB address is given priority over
679 	 * iWARP.
680 	 */
681 	if (chanp->chan_ib_client_hdl == NULL) {
682 		ofs_client_t	*ofs_clnt;
683 
684 		ofs_clnt = (ofs_client_t *)sol_cma_ib_client->clnt_hdl;
685 		chanp->chan_ib_client_hdl = ofs_clnt->ibt_hdl;
686 	}
687 	if (chanp->chan_ib_client_hdl && rdma_ib_bind_addr(idp, addr) == 0) {
688 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
689 		    "rdma_bind_addr: ret IB @");
690 		return (0);
691 #ifdef	IWARP_SUPPORT
692 	} else if (chanp->chan_iw_client_hdl && rdma_iw_bind_addr(idp, addr)
693 	    == 0) {
694 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
695 		    "rdma_bind_addr: ret iWARP @");
696 		return (0);
697 #endif	/* IWARP_SUPPORT */
698 	}
699 
700 	mutex_enter(&chanp->chan_mutex);
701 	cma_set_chan_state(chanp, SOL_CMA_CHAN_IDLE);
702 	mutex_exit(&chanp->chan_mutex);
703 	SOL_OFS_DPRINTF_L4(sol_rdmacm_dbg_str, "rdma_bind_addr: ret failure!");
704 	return (EINVAL);
705 }
706 
707 int
rdma_resolve_addr(struct rdma_cm_id * idp,struct sockaddr * src_addr,struct sockaddr * dst_addr,int timeout_ms)708 rdma_resolve_addr(struct rdma_cm_id *idp, struct sockaddr *src_addr,
709     struct sockaddr *dst_addr, int timeout_ms)
710 {
711 	sol_cma_chan_t		*chanp;
712 	struct rdma_addr	*addrp;
713 	cma_chan_state_t	state;
714 
715 	ASSERT(idp);
716 	chanp = (sol_cma_chan_t *)idp;
717 	addrp = &(idp->route.addr);
718 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_resolve_addr(%p, %p, "
719 	    "%p, %x)", idp, src_addr, dst_addr, timeout_ms);
720 
721 	mutex_enter(&chanp->chan_mutex);
722 	state = cma_get_chan_state(chanp);
723 	if (state != SOL_CMA_CHAN_IDLE && state != SOL_CMA_CHAN_BOUND) {
724 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
725 		    "rdma_resolve_addr : invalid chan state %x", state);
726 		mutex_exit(&chanp->chan_mutex);
727 		return (EINVAL);
728 	}
729 	if (chanp->chan_cmid_destroy_state &
730 	    SOL_CMA_CALLER_CMID_DESTROYED) {
731 		SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
732 		    "rdma_resolve_addr : CMID %p, destroy called", chanp);
733 		mutex_exit(&chanp->chan_mutex);
734 		return (EINVAL);
735 	}
736 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_API_PROGRESS;
737 
738 	if (chanp->chan_xport_type == SOL_CMA_XPORT_NONE) {
739 		bcopy((void *)src_addr, (void *)&(addrp->src_addr),
740 		    sizeof (struct sockaddr));
741 	}
742 	bcopy((void *)dst_addr, (void *)&(addrp->dst_addr),
743 	    sizeof (struct sockaddr));
744 	mutex_exit(&chanp->chan_mutex);
745 
746 	/*
747 	 * First resolve this as an @ corresponding to IB fabric
748 	 * if this fails, resolve this as an @ corresponding to iWARP
749 	 */
750 	if (chanp->chan_ib_client_hdl == NULL) {
751 		ofs_client_t	*ofs_clnt;
752 
753 		ofs_clnt = (ofs_client_t *)sol_cma_ib_client->clnt_hdl;
754 		chanp->chan_ib_client_hdl = ofs_clnt->ibt_hdl;
755 	}
756 	if (chanp->chan_ib_client_hdl && rdma_ib_resolve_addr(idp, src_addr,
757 	    dst_addr, timeout_ms) == 0) {
758 		SOL_OFS_DPRINTF_L4(sol_rdmacm_dbg_str,
759 		    "rdma_resolve_addr: ret IB @");
760 #ifdef IWARP_SUPPORT
761 	} else if (chanp->chan_iw_client_hdl && rdma_iw_resolve_addr(idp,
762 	    src_addr, dst_addr, timeout_ms) == 0) {
763 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
764 		    "rdma_resolve_addr: ret iWARP @");
765 #endif	/* IWARP_SUPPORT */
766 	} else {
767 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
768 		    "rdma_resolve_addr: Invalid @");
769 		return (EINVAL);
770 	}
771 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_resolve_addr: ret 0");
772 	return (0);
773 }
774 
775 static void cma_generate_event_sync(struct rdma_cm_id *,
776     enum rdma_cm_event_type, int, struct rdma_conn_param *,
777     struct rdma_ud_param *);
778 
779 void
cma_resolve_addr_callback(sol_cma_chan_t * chanp,int rc)780 cma_resolve_addr_callback(sol_cma_chan_t *chanp, int rc)
781 {
782 	enum rdma_cm_event_type	event;
783 
784 	mutex_enter(&chanp->chan_mutex);
785 	if (chanp->chan_cmid_destroy_state &
786 	    SOL_CMA_CALLER_CMID_DESTROYED) {
787 		SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
788 		    "cma_resolve_addr : CMID %p, destroy called", chanp);
789 		chanp->chan_cmid_destroy_state &=
790 		    ~SOL_CMA_CALLER_API_PROGRESS;
791 		cv_broadcast(&chanp->chan_destroy_cv);
792 		mutex_exit(&chanp->chan_mutex);
793 		return;
794 	}
795 	if (rc == 0) {
796 		cma_set_chan_state(chanp, SOL_CMA_CHAN_ADDR_RESLVD);
797 		event = RDMA_CM_EVENT_ADDR_RESOLVED;
798 	} else
799 		event = RDMA_CM_EVENT_ADDR_ERROR;
800 
801 	/*
802 	 * Generate RDMA_CM_EVENT_ADDR_RESOLVED event
803 	 * This will result in RDMA_USER_CM_CMD_RESOLVE_ROUTE in
804 	 * userland.
805 	 */
806 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_EVENT_PROGRESS;
807 	mutex_exit(&chanp->chan_mutex);
808 	cma_generate_event_sync((struct rdma_cm_id *)chanp, event, 0,
809 	    NULL, NULL);
810 
811 	mutex_enter(&chanp->chan_mutex);
812 	chanp->chan_cmid_destroy_state &= ~SOL_CMA_CALLER_API_PROGRESS;
813 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED)
814 		cv_broadcast(&chanp->chan_destroy_cv);
815 	mutex_exit(&chanp->chan_mutex);
816 }
817 
818 int
rdma_resolve_route(struct rdma_cm_id * idp,int timeout_ms)819 rdma_resolve_route(struct rdma_cm_id *idp, int timeout_ms)
820 {
821 	sol_cma_chan_t		*chanp;
822 
823 	ASSERT(idp);
824 	chanp = (sol_cma_chan_t *)idp;
825 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "resolve_route(%p, %x)", idp,
826 	    timeout_ms);
827 
828 	mutex_enter(&chanp->chan_mutex);
829 	if (cma_cas_chan_state(chanp, SOL_CMA_CHAN_ADDR_RESLVD,
830 	    SOL_CMA_CHAN_ROUTE_RESLVD) != 0) {
831 		mutex_exit(&chanp->chan_mutex);
832 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
833 		    "resolve_route: Invalid state");
834 		return (EINVAL);
835 	}
836 	if (chanp->chan_cmid_destroy_state &
837 	    SOL_CMA_CALLER_CMID_DESTROYED) {
838 		SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
839 		    "rdma_resolve_route : CMID %p, destroy called", chanp);
840 		mutex_exit(&chanp->chan_mutex);
841 		return (EINVAL);
842 	}
843 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_API_PROGRESS;
844 	mutex_exit(&chanp->chan_mutex);
845 
846 	/*
847 	 * Generate RDMA_CM_EVENT_ROUTE_RESOLVED event
848 	 * This will result in RDMA_USER_CM_CMD_RESOLVE_ROUTE in
849 	 * userland
850 	 */
851 	cma_generate_event(idp, RDMA_CM_EVENT_ROUTE_RESOLVED, 0,
852 	    NULL, NULL);
853 
854 	mutex_enter(&chanp->chan_mutex);
855 	chanp->chan_cmid_destroy_state &= ~SOL_CMA_CALLER_API_PROGRESS;
856 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED)
857 		cv_broadcast(&chanp->chan_destroy_cv);
858 	mutex_exit(&chanp->chan_mutex);
859 
860 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "resolve_route: ret 0");
861 	return (0);
862 }
863 
864 /*
865  * Connect or Listen request should be send after Route is resolved
866  *
867  *	Active Side (Client) :
868  *	1. (State ROUTE_RESOLVED)-->CONNECT-->ACCEPT/REJECT-->DISCONNECT
869  *	       -->DESTROY_ID-->close(9E)
870  *	2. Same as (1), DESTROY_ID without DISCONNECT
871  *	3. Same as (1), close(9e) without DESTROY_ID.
872  *
873  *	Passive Side (Server) :
874  *	4. (State ROUTE_RESOLVED)-->LISTEN->DISCONNECT
875  *		-->DESTROY_ID-->close(9E)
876  *	5. Same as (4), DESTROY_ID without DISCONNECT
877  *	6. Same as (4), close(9e) without DESTROY_ID.
878  */
879 int
rdma_connect(struct rdma_cm_id * idp,struct rdma_conn_param * conn_param)880 rdma_connect(struct rdma_cm_id *idp, struct rdma_conn_param *conn_param)
881 {
882 	sol_cma_chan_t		*chanp;
883 	int			ret = EINVAL;
884 
885 	ASSERT(idp);
886 	chanp = (sol_cma_chan_t *)idp;
887 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_connect(%p, %p)", idp,
888 	    conn_param);
889 
890 	mutex_enter(&chanp->chan_mutex);
891 	if (chanp->chan_xport_type == SOL_CMA_XPORT_NONE) {
892 		mutex_exit(&chanp->chan_mutex);
893 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
894 		    "rdma_connect, Invalid Xport");
895 		return (EINVAL);
896 	}
897 	if (cma_cas_chan_state(chanp, SOL_CMA_CHAN_ROUTE_RESLVD,
898 	    SOL_CMA_CHAN_CONNECT)) {
899 		mutex_exit(&chanp->chan_mutex);
900 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
901 		    "rdma_connect, Invalid state");
902 		return (EINVAL);
903 	}
904 
905 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB) {
906 		ret = rdma_ib_connect(idp, conn_param);
907 #ifdef	IWARP_SUPPORT
908 	} else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP) {
909 		ret = rdma_iw_connect(idp, conn_param);
910 #endif	/* IWARP_SUPPORT */
911 	}
912 	mutex_exit(&chanp->chan_mutex);
913 
914 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_connect: ret %x", ret);
915 	return (ret);
916 }
917 
918 static int cma_init_listen_root(sol_cma_chan_t *);
919 static void cma_fini_listen_root(sol_cma_chan_t *);
920 
921 int
rdma_listen(struct rdma_cm_id * idp,int bklog)922 rdma_listen(struct rdma_cm_id *idp, int bklog)
923 {
924 	sol_cma_chan_t		*chanp;
925 	int			ret = 0;
926 	genlist_entry_t		*entry;
927 	cma_chan_state_t	state;
928 
929 	ASSERT(idp);
930 	chanp = (sol_cma_chan_t *)idp;
931 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_listen(%p, %x)",
932 	    idp, bklog);
933 
934 	mutex_enter(&chanp->chan_mutex);
935 	state = cma_get_chan_state(chanp);
936 	if (state == SOL_CMA_CHAN_IDLE) {
937 		mutex_exit(&chanp->chan_mutex);
938 		return (EINVAL);
939 	}
940 	cma_set_chan_state(chanp, SOL_CMA_CHAN_LISTEN);
941 
942 	if (chanp->chan_cmid_destroy_state &
943 	    SOL_CMA_CALLER_CMID_DESTROYED) {
944 		SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
945 		    "rdma_listen : CMID %p, destroy called", chanp);
946 		mutex_exit(&chanp->chan_mutex);
947 		return (EINVAL);
948 	}
949 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_API_PROGRESS;
950 
951 	ASSERT(chanp->chan_listenp == NULL);
952 
953 	chanp->chan_listenp = kmem_zalloc(sizeof (sol_cma_listen_info_t),
954 	    KM_SLEEP);
955 	init_genlist(&(CHAN_LISTEN_LIST(chanp)));
956 	(chanp->chan_listenp)->listen_is_root = 1;
957 	ret = cma_init_listen_root(chanp);
958 	if (ret) {
959 		chanp->chan_listenp = NULL;
960 		mutex_exit(&chanp->chan_mutex);
961 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str, "rdma_listen: "
962 		    "cma_init_listen_root: failed");
963 		kmem_free(chanp->chan_listenp,
964 		    sizeof (sol_cma_listen_info_t));
965 		return (EINVAL);
966 	}
967 
968 	if (chanp->chan_xport_type == SOL_CMA_XPORT_NONE) {
969 		ibcma_append_listen_list(idp);
970 #ifdef IWARP_SUPPORT
971 		iwcma_append_listen_list(idp);
972 #endif
973 	} else if (chanp->chan_xport_type == SOL_CMA_XPORT_IB) {
974 		ibcma_append_listen_list(idp);
975 #ifdef	IWARP_SUPPORT
976 	} else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP) {
977 		iwcma_append_listen_list(idp);
978 #endif	/* IWARP_SUPPORT */
979 	}
980 
981 	if (genlist_empty(&(CHAN_LISTEN_LIST(chanp)))) {
982 		cma_fini_listen_root(chanp);
983 		kmem_free((void *)chanp->chan_listenp,
984 		    sizeof (sol_cma_listen_info_t));
985 		chanp->chan_listenp = NULL;
986 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str, "rdma_listen: "
987 		    "No listeners");
988 		mutex_exit(&chanp->chan_mutex);
989 		return (0);
990 	}
991 
992 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED) {
993 		chanp->chan_cmid_destroy_state &=
994 		    ~SOL_CMA_CALLER_API_PROGRESS;
995 		cv_broadcast(&chanp->chan_destroy_cv);
996 	}
997 
998 	genlist_for_each(entry, &(CHAN_LISTEN_LIST(chanp))) {
999 		struct rdma_cm_id	*ep_idp;
1000 		sol_cma_chan_t		*ep_chanp;
1001 
1002 		ep_idp = (struct rdma_cm_id *)entry->data;
1003 		ep_chanp = (sol_cma_chan_t *)ep_idp;
1004 		if (ep_chanp->chan_xport_type == SOL_CMA_XPORT_IB)
1005 			ret = rdma_ib_listen(ep_idp, bklog);
1006 #ifdef IWARP_SUPPORT
1007 		if (ep_chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1008 			ret = rdma_iw_listen(ep_idp, bklog);
1009 #endif
1010 		if (ret)
1011 			break;
1012 	}
1013 
1014 	chanp->chan_cmid_destroy_state &= ~SOL_CMA_CALLER_API_PROGRESS;
1015 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED)
1016 		cv_broadcast(&chanp->chan_destroy_cv);
1017 	mutex_exit(&chanp->chan_mutex);
1018 
1019 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_listen: ret %x", ret);
1020 	return (ret);
1021 }
1022 
1023 int
rdma_accept(struct rdma_cm_id * idp,struct rdma_conn_param * conn_param)1024 rdma_accept(struct rdma_cm_id *idp, struct rdma_conn_param *conn_param)
1025 {
1026 	struct rdma_cm_id	*root_idp;
1027 	sol_cma_chan_t		*root_chanp, *chanp;
1028 	int			ret = EINVAL;
1029 
1030 	ASSERT(idp);
1031 	chanp = (sol_cma_chan_t *)idp;
1032 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_accept(%p, %p)",
1033 	    idp, conn_param);
1034 
1035 	mutex_enter(&chanp->chan_mutex);
1036 	if (cma_cas_chan_state(chanp, SOL_CMA_CHAN_LISTEN,
1037 	    SOL_CMA_CHAN_ACCEPT) && cma_cas_chan_state(chanp,
1038 	    SOL_CMA_CHAN_CONNECT, SOL_CMA_CHAN_ACCEPT)) {
1039 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1040 		    "rdma_accept, Invalid state");
1041 		mutex_exit(&chanp->chan_mutex);
1042 		return (EINVAL);
1043 	}
1044 	mutex_exit(&chanp->chan_mutex);
1045 
1046 	root_idp = CHAN_LISTEN_ROOT(chanp);
1047 	root_chanp = (sol_cma_chan_t *)root_idp;
1048 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "accept: root_idp %p",
1049 	    root_idp);
1050 
1051 	/* For TCP, delete from REQ AVL & insert to ACPT AVL */
1052 	if (root_idp && root_idp->ps == RDMA_PS_TCP) {
1053 		void		*find_ret;
1054 		avl_index_t	where;
1055 
1056 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "accept: root_idp %p"
1057 		    "REQ AVL remove %p", root_chanp, idp);
1058 		mutex_enter(&root_chanp->chan_mutex);
1059 		mutex_enter(&chanp->chan_mutex);
1060 
1061 		/*
1062 		 * This CMID has been deleted, maybe because of timeout.
1063 		 * Return EINVAL.
1064 		 */
1065 		if (chanp->chan_req_state != REQ_CMID_NOTIFIED) {
1066 			mutex_exit(&chanp->chan_mutex);
1067 			mutex_exit(&root_chanp->chan_mutex);
1068 			SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
1069 			    "accept: root_idp %p chanp %p, not in REQ "
1070 			    "AVL tree",  root_chanp, chanp);
1071 			return (EINVAL);
1072 		}
1073 		ASSERT(cma_get_req_idp(root_idp, chanp->chan_session_id));
1074 		avl_remove(&root_chanp->chan_req_avl_tree, idp);
1075 
1076 
1077 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1078 		    "Add to ACPT AVL of %p IDP, idp %p, qp_hdl %p",
1079 		    root_idp, idp, chanp->chan_qp_hdl);
1080 		find_ret = avl_find(&root_chanp->chan_acpt_avl_tree,
1081 		    (void *)chanp->chan_qp_hdl, &where);
1082 		if (find_ret) {
1083 			chanp->chan_req_state = REQ_CMID_SERVER_NONE;
1084 			mutex_exit(&chanp->chan_mutex);
1085 			mutex_exit(&root_chanp->chan_mutex);
1086 			SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1087 			    "DUPLICATE ENTRY in ACPT AVL : root %p, "
1088 			    "idp %p, qp_hdl %p",
1089 			    root_idp, idp, chanp->chan_qp_hdl);
1090 			return (EINVAL);
1091 		}
1092 		avl_insert(&root_chanp->chan_acpt_avl_tree,
1093 		    (void *)idp, where);
1094 		chanp->chan_req_state = REQ_CMID_ACCEPTED;
1095 		mutex_exit(&chanp->chan_mutex);
1096 		mutex_exit(&root_chanp->chan_mutex);
1097 	}
1098 
1099 	if (root_idp && IS_UDP_CMID(root_idp)) {
1100 		cma_chan_state_t	chan_state;
1101 
1102 		/*
1103 		 * Accepting the connect request, no more events for this
1104 		 * connection.
1105 		 */
1106 		cma_handle_nomore_events(chanp);
1107 		mutex_enter(&chanp->chan_mutex);
1108 		chan_state = cma_get_chan_state(chanp);
1109 		mutex_exit(&chanp->chan_mutex);
1110 		/* If rdma_destroy_id() was called, destroy CMID */
1111 		if (chan_state == SOL_CMA_CHAN_DESTROY_PENDING) {
1112 			cma_destroy_id((struct rdma_cm_id *)chanp);
1113 			return (EINVAL);
1114 		}
1115 	}
1116 
1117 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB)
1118 		ret = rdma_ib_accept(idp, conn_param);
1119 #ifdef	IWARP_SUPPORT
1120 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1121 		ret = rdma_iw_accept(idp, conn_param);
1122 #endif	/* IWARP_SUPPORT */
1123 
1124 	if (ret && root_idp && idp->ps == RDMA_PS_TCP) {
1125 		void		*find_ret;
1126 		avl_index_t	where;
1127 
1128 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1129 		    "Delete from REQ AVL of %p IDP, idp %p",
1130 		    root_idp, idp);
1131 		mutex_enter(&root_chanp->chan_mutex);
1132 		mutex_enter(&chanp->chan_mutex);
1133 		if (chanp->chan_req_state == REQ_CMID_ACCEPTED) {
1134 			ASSERT(cma_get_acpt_idp(root_idp,
1135 			    chanp->chan_qp_hdl));
1136 			avl_remove(&root_chanp->chan_acpt_avl_tree,
1137 			    idp);
1138 			find_ret = avl_find(&root_chanp->chan_req_avl_tree,
1139 			    (void *)chanp->chan_qp_hdl, &where);
1140 			if (find_ret) {
1141 				chanp->chan_req_state = REQ_CMID_SERVER_NONE;
1142 				mutex_exit(&chanp->chan_mutex);
1143 				mutex_exit(&root_chanp->chan_mutex);
1144 				SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1145 				    "DUPLICATE ENTRY in REQ AVL : root %p, "
1146 				    "idp %p, session_id %p",
1147 				    root_idp, idp, chanp->chan_session_id);
1148 				return (EINVAL);
1149 			}
1150 			avl_insert(&root_chanp->chan_req_avl_tree, idp, where);
1151 			chanp->chan_req_state = REQ_CMID_NOTIFIED;
1152 		}
1153 		mutex_exit(&chanp->chan_mutex);
1154 		mutex_exit(&root_chanp->chan_mutex);
1155 	}
1156 
1157 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_accept: ret %x", ret);
1158 	return (ret);
1159 }
1160 
1161 int
rdma_notify(struct rdma_cm_id * idp,enum ib_event_type evt)1162 rdma_notify(struct rdma_cm_id *idp, enum ib_event_type evt)
1163 {
1164 	sol_cma_chan_t		*chanp;
1165 
1166 	ASSERT(idp);
1167 	chanp = (sol_cma_chan_t *)idp;
1168 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_notify(%p, %x)", idp, evt);
1169 
1170 	mutex_enter(&chanp->chan_mutex);
1171 	if (cma_cas_chan_state(chanp, SOL_CMA_CHAN_ROUTE_RESLVD,
1172 	    SOL_CMA_CHAN_EVENT_NOTIFIED)) {
1173 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1174 		    "rdma_notify, Invalid state");
1175 		mutex_exit(&chanp->chan_mutex);
1176 		return (EINVAL);
1177 	}
1178 	mutex_exit(&chanp->chan_mutex);
1179 
1180 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_notify: ret 0");
1181 	return (0);
1182 }
1183 
1184 int
rdma_reject(struct rdma_cm_id * idp,const void * priv_data,uint8_t priv_data_len)1185 rdma_reject(struct rdma_cm_id *idp, const void *priv_data,
1186     uint8_t priv_data_len)
1187 {
1188 	struct rdma_cm_id	*root_idp;
1189 	sol_cma_chan_t		*root_chanp, *chanp;
1190 	int			ret = EINVAL;
1191 
1192 	ASSERT(idp);
1193 	chanp = (sol_cma_chan_t *)idp;
1194 	root_idp = CHAN_LISTEN_ROOT(chanp);
1195 	root_chanp = (sol_cma_chan_t *)root_idp;
1196 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_reject(%p, %p)", idp,
1197 	    priv_data, priv_data_len);
1198 
1199 	mutex_enter(&chanp->chan_mutex);
1200 	if (cma_cas_chan_state(chanp, SOL_CMA_CHAN_LISTEN,
1201 	    SOL_CMA_CHAN_REJECT)) {
1202 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1203 		    "rdma_accept, Invalid state");
1204 		mutex_exit(&chanp->chan_mutex);
1205 		return (EINVAL);
1206 	}
1207 	mutex_exit(&chanp->chan_mutex);
1208 
1209 	if (root_idp) {
1210 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "reject: root_idp %p"
1211 		    "REQ AVL remove %p", root_chanp, idp);
1212 
1213 		/*
1214 		 * Remove from REQ AVL tree. If this CMID has been deleted,
1215 		 * it maybe because of timeout. Return EINVAL.
1216 		 */
1217 		mutex_enter(&root_chanp->chan_mutex);
1218 		mutex_enter(&chanp->chan_mutex);
1219 		if (chanp->chan_req_state != REQ_CMID_NOTIFIED &&
1220 		    chanp->chan_req_state != REQ_CMID_QUEUED) {
1221 			mutex_exit(&chanp->chan_mutex);
1222 			mutex_exit(&root_chanp->chan_mutex);
1223 			SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
1224 			    "reject: root_idp %p chanp %p, not in REQ "
1225 			    "AVL tree",  root_chanp, chanp);
1226 			return (EINVAL);
1227 		}
1228 		ASSERT(cma_get_req_idp(root_idp, chanp->chan_session_id));
1229 		avl_remove(&root_chanp->chan_req_avl_tree, idp);
1230 		chanp->chan_req_state = REQ_CMID_SERVER_NONE;
1231 		mutex_exit(&chanp->chan_mutex);
1232 		mutex_exit(&root_chanp->chan_mutex);
1233 	}
1234 
1235 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB)
1236 		ret = rdma_ib_reject(idp, priv_data, priv_data_len);
1237 #ifdef	IWARP_SUPPORT
1238 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1239 		ret = rdma_iw_reject(idp, priv_data, priv_data_len);
1240 #endif	/* IWARP_SUPPORT */
1241 
1242 
1243 	if (!ret && root_idp) {
1244 		cma_chan_state_t	chan_state;
1245 
1246 		/*
1247 		 * Rejecting connect request, no more events for this
1248 		 * connection.
1249 		 */
1250 		cma_handle_nomore_events(chanp);
1251 		mutex_enter(&chanp->chan_mutex);
1252 		chan_state = cma_get_chan_state(chanp);
1253 		mutex_exit(&chanp->chan_mutex);
1254 		/* If rdma_destroy_id() was called, destroy CMID */
1255 		if (chan_state == SOL_CMA_CHAN_DESTROY_PENDING)
1256 			cma_destroy_id((struct rdma_cm_id *)chanp);
1257 	} else if (ret && root_idp) {
1258 		avl_index_t	where;
1259 
1260 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1261 		    "reject fail: Add to Req AVL of %p IDP, idp %p,"
1262 		    "session_id %p", root_idp, idp,
1263 		    chanp->chan_session_id);
1264 		mutex_enter(&root_chanp->chan_mutex);
1265 		mutex_enter(&chanp->chan_mutex);
1266 		if (chanp->chan_req_state == REQ_CMID_SERVER_NONE) {
1267 			if (avl_find(&root_chanp->chan_req_avl_tree,
1268 			    (void *)chanp->chan_session_id, &where)) {
1269 				mutex_exit(&chanp->chan_mutex);
1270 				mutex_exit(&root_chanp->chan_mutex);
1271 				SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1272 				    "DUPLICATE ENTRY in REQ AVL : root %p, "
1273 				    "idp %p, session_id %p",
1274 				    root_idp, idp, chanp->chan_session_id);
1275 				return (EINVAL);
1276 			}
1277 			avl_insert(&root_chanp->chan_req_avl_tree,
1278 			    (void *)idp, where);
1279 			chanp->chan_req_state = REQ_CMID_NOTIFIED;
1280 		}
1281 		mutex_exit(&chanp->chan_mutex);
1282 		mutex_exit(&root_chanp->chan_mutex);
1283 	}
1284 
1285 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_reject: ret %x", ret);
1286 	return (ret);
1287 }
1288 
1289 int
rdma_disconnect(struct rdma_cm_id * idp)1290 rdma_disconnect(struct rdma_cm_id *idp)
1291 {
1292 	sol_cma_chan_t		*chanp;
1293 	int			ret = EINVAL;
1294 	cma_chan_state_t	state;
1295 
1296 	chanp = (sol_cma_chan_t *)idp;
1297 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_disconnect(%p)", idp);
1298 
1299 	if (!idp)
1300 		return (0);
1301 
1302 	mutex_enter(&chanp->chan_mutex);
1303 	if (!(SOL_CMAID_CONNECTED(chanp))) {
1304 		SOL_OFS_DPRINTF_L3(sol_rdmacm_dbg_str,
1305 		    "rdma_disconnect(%p) - Not connected!!", idp);
1306 		mutex_exit(&chanp->chan_mutex);
1307 		return (EINVAL);
1308 	}
1309 	state = cma_get_chan_state(chanp);
1310 	cma_set_chan_state(chanp, SOL_CMA_CHAN_DISCONNECT);
1311 	mutex_exit(&chanp->chan_mutex);
1312 
1313 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB) {
1314 		ret = rdma_ib_disconnect(idp);
1315 #ifdef	IWARP_SUPPORT
1316 	} else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP) {
1317 		ret = rdma_iw_disconnect(idp);
1318 #endif	/* IWARP_SUPPORT */
1319 	}
1320 
1321 	if (ret) {
1322 		mutex_enter(&chanp->chan_mutex);
1323 		cma_set_chan_state(chanp, state);
1324 		mutex_exit(&chanp->chan_mutex);
1325 		return (ret);
1326 	}
1327 
1328 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_disconnect: ret %x", ret);
1329 	return (ret);
1330 }
1331 
1332 int
rdma_init_qp_attr(struct rdma_cm_id * idp,struct ib_qp_attr * qpattr,int * qp_attr_mask)1333 rdma_init_qp_attr(struct rdma_cm_id *idp, struct ib_qp_attr *qpattr,
1334     int *qp_attr_mask)
1335 {
1336 	sol_cma_chan_t		*chanp;
1337 	int			ret = EINVAL;
1338 
1339 	ASSERT(idp);
1340 	chanp = (sol_cma_chan_t *)idp;
1341 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_init_qp_attr(%p, %p, %p)",
1342 	    idp, qpattr, qp_attr_mask);
1343 
1344 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB) {
1345 		ret = rdma_ib_init_qp_attr(idp, qpattr, qp_attr_mask);
1346 #ifdef	IWARP_SUPPORT
1347 	} else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1348 		ret = rdma_iw_init_qp_attr(idp, qpattr, qp_attr_mask);
1349 #endif	/* IWARP_SUPPORT */
1350 	} else {
1351 		ret = EINVAL;
1352 	}
1353 
1354 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1355 	    "rdma_init_qp_attr: ret %x", ret);
1356 
1357 	return (ret);
1358 }
1359 
1360 int
1361 rdma_join_multicast(struct rdma_cm_id *idp, struct sockaddr *addr,
1362     void *context)
1363 {
1364 	sol_cma_chan_t		*chanp;
1365 	int			ret = ENODEV;
1366 	cma_chan_state_t	state;
1367 
1368 	ASSERT(idp);
1369 	chanp = (sol_cma_chan_t *)idp;
1370 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1371 	    "rdma_join_multicast(%p, %p, %p)",
1372 	    idp, addr, context);
1373 
1374 	mutex_enter(&chanp->chan_mutex);
1375 	state = cma_get_chan_state(chanp);
1376 	if (state != SOL_CMA_CHAN_BOUND &&
1377 	    state != SOL_CMA_CHAN_ROUTE_RESLVD &&
1378 	    state != SOL_CMA_CHAN_ADDR_RESLVD) {
1379 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1380 		    "rdma_join_multicast, Invalid state");
1381 		mutex_exit(&chanp->chan_mutex);
1382 		return (EINVAL);
1383 	}
1384 
1385 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB)
1386 		ret = rdma_ib_join_multicast(idp, addr, context);
1387 #ifdef	IWARP_SUPPORT
1388 	/* No support for Multicast on iWARP */
1389 	else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1390 		ret = ENOTSUP;
1391 #endif	/* IWARP_SUPPORT */
1392 	mutex_exit(&chanp->chan_mutex);
1393 
1394 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1395 	    "rdma_join_multicast: ret %x", ret);
1396 	return (ret);
1397 }
1398 
1399 void
1400 rdma_leave_multicast(struct rdma_cm_id *idp, struct sockaddr *addr)
1401 {
1402 	sol_cma_chan_t		*chanp;
1403 	cma_chan_state_t	state;
1404 
1405 	ASSERT(idp);
1406 	chanp = (sol_cma_chan_t *)idp;
1407 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_leave_multicast(%p, %p)",
1408 	    idp, addr);
1409 
1410 	mutex_enter(&chanp->chan_mutex);
1411 	state = cma_get_chan_state(chanp);
1412 	if (state != SOL_CMA_CHAN_BOUND &&
1413 	    state != SOL_CMA_CHAN_ROUTE_RESLVD &&
1414 	    state != SOL_CMA_CHAN_ADDR_RESLVD) {
1415 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1416 		    "rdma_leave_multicast, Invalid state");
1417 		mutex_exit(&chanp->chan_mutex);
1418 		return;
1419 	}
1420 
1421 	if (chanp->chan_xport_type == SOL_CMA_XPORT_IB)
1422 		rdma_ib_leave_multicast(idp, addr);
1423 #ifdef	IWARP_SUPPORT
1424 	/* No support for Multicast on iWARP */
1425 	else if (chanp->chan_xport_type == SOL_CMA_XPORT_IWARP)
1426 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1427 		    "rdma_leave_multicast, iWARP");
1428 #endif	/* IWARP_SUPPORT */
1429 	mutex_exit(&chanp->chan_mutex);
1430 
1431 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "rdma_join_multicast: ret");
1432 }
1433 
1434 /*
1435  * Functions to compare to rdma_cm_id *, used by AVL tree
1436  * routines.
1437  */
1438 int
1439 sol_cma_req_cmid_cmp(const void *p1, const void *p2)
1440 {
1441 	sol_cma_chan_t		*chanp;
1442 
1443 	chanp = (sol_cma_chan_t *)p2;
1444 	if (chanp->chan_session_id > p1)
1445 		return (+1);
1446 	else if (chanp->chan_session_id < p1)
1447 		return (-1);
1448 	else
1449 		return (0);
1450 }
1451 
1452 int
1453 sol_cma_cmid_cmp(const void *p1, const void *p2)
1454 {
1455 	sol_cma_chan_t		*chanp;
1456 
1457 	chanp = (sol_cma_chan_t *)p2;
1458 	if (chanp->chan_qp_hdl > p1)
1459 		return (+1);
1460 	else if (chanp->chan_qp_hdl < p1)
1461 		return (-1);
1462 	else
1463 		return (0);
1464 }
1465 
1466 /*
1467  * Function to compare two sol_cma_glbl_listen_t *, used by
1468  * AVL tree routines.
1469  */
1470 int
1471 sol_cma_svc_cmp(const void *p1, const void *p2)
1472 {
1473 	sol_cma_glbl_listen_t	*listenp;
1474 	uint64_t		sid;
1475 
1476 	sid = *(uint64_t *)p1;
1477 	listenp = (sol_cma_glbl_listen_t *)p2;
1478 	if (listenp->cma_listen_chan_sid > sid)
1479 		return (+1);
1480 	else if (listenp->cma_listen_chan_sid < sid)
1481 		return (-1);
1482 	else
1483 		return (0);
1484 }
1485 
1486 static int
1487 cma_init_listen_root(sol_cma_chan_t *chanp)
1488 {
1489 	sol_cma_glbl_listen_t	*cma_listenp;
1490 	sol_cma_listen_info_t	*chan_listenp;
1491 	int			rc = 0;
1492 	avl_index_t		where = 0;
1493 	uint64_t		listen_sid;
1494 
1495 	ASSERT(chanp);
1496 	ASSERT(chanp->chan_listenp);
1497 	chan_listenp = chanp->chan_listenp;
1498 
1499 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1500 	    "cma_init_listen_root(%p)", chanp);
1501 
1502 	/*
1503 	 * First search for matching global listen_info for this SID.
1504 	 * If found with the same client handle, reuse the service
1505 	 * handle, if matching SID is found with different client
1506 	 * handle, return EINVAL.
1507 	 */
1508 	listen_sid = ibcma_init_root_sid(chanp);
1509 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1510 	    "cma_init_listen_root: search SID 0x%llx",
1511 	    listen_sid);
1512 
1513 	mutex_enter(&sol_cma_glob_mutex);
1514 	cma_listenp = avl_find(&sol_cma_glbl_listen_tree,
1515 	    (void *) &listen_sid, &where);
1516 	if (cma_listenp && cma_listenp->cma_listen_clnt_hdl ==
1517 	    chanp->chan_ib_client_hdl) {
1518 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1519 		    "cma_init_listen_root: matching listenp %p SID 0x%llx",
1520 		    cma_listenp, listen_sid);
1521 		chan_listenp->listen_entry = add_genlist(
1522 		    &cma_listenp->cma_listen_chan_list,
1523 		    (uintptr_t)chanp, NULL);
1524 		chan_listenp->chan_glbl_listen_info = cma_listenp;
1525 		ibcma_copy_srv_hdl(chanp, cma_listenp);
1526 		mutex_exit(&sol_cma_glob_mutex);
1527 		return (0);
1528 	} else if (cma_listenp) {
1529 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1530 		    "cma_init_listen_root: listenp %p, SID 0x%llx match, "
1531 		    "client hdl prev %p, new %p mismatch",
1532 		    cma_listenp, listen_sid,
1533 		    cma_listenp->cma_listen_clnt_hdl,
1534 		    chanp->chan_ib_client_hdl);
1535 		mutex_exit(&sol_cma_glob_mutex);
1536 		return (EINVAL);
1537 	}
1538 
1539 	cma_listenp = kmem_zalloc(sizeof (sol_cma_glbl_listen_t), KM_SLEEP);
1540 	init_genlist(&cma_listenp->cma_listen_chan_list);
1541 	chan_listenp->listen_entry = add_genlist(
1542 	    &cma_listenp->cma_listen_chan_list, (uintptr_t)chanp, NULL);
1543 	chan_listenp->chan_glbl_listen_info = cma_listenp;
1544 	cma_listenp->cma_listen_clnt_hdl = chanp->chan_ib_client_hdl;
1545 	cma_listenp->cma_listen_chan_sid = listen_sid;
1546 
1547 	rc = ibcma_init_root_chan(chanp, cma_listenp);
1548 	if (rc) {
1549 		mutex_exit(&sol_cma_glob_mutex);
1550 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1551 		    "cma_init_listen_root: ibcma_init_root_chan failed!!");
1552 		delete_genlist(&cma_listenp->cma_listen_chan_list,
1553 		    chan_listenp->listen_entry);
1554 		kmem_free(cma_listenp, sizeof (sol_cma_glbl_listen_t));
1555 		return (rc);
1556 	}
1557 	avl_insert(&sol_cma_glbl_listen_tree, cma_listenp, where);
1558 	mutex_exit(&sol_cma_glob_mutex);
1559 	return (0);
1560 }
1561 
1562 static void
1563 cma_fini_listen_root(sol_cma_chan_t *chanp)
1564 {
1565 	sol_cma_glbl_listen_t	*cma_listenp;
1566 	sol_cma_listen_info_t	*chan_listenp;
1567 
1568 	ASSERT(chanp);
1569 	ASSERT(chanp->chan_listenp);
1570 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "cma_fini_listen_root(%p)",
1571 	    chanp);
1572 	chan_listenp = chanp->chan_listenp;
1573 	cma_listenp = chan_listenp->chan_glbl_listen_info;
1574 	ASSERT(cma_listenp);
1575 	mutex_enter(&sol_cma_glob_mutex);
1576 	delete_genlist(&cma_listenp->cma_listen_chan_list,
1577 	    chan_listenp->listen_entry);
1578 	if (genlist_empty(&cma_listenp->cma_listen_chan_list)) {
1579 		if (ibcma_fini_root_chan(chanp) == 0) {
1580 			avl_remove(&sol_cma_glbl_listen_tree,
1581 			    cma_listenp);
1582 			kmem_free(cma_listenp,
1583 			    sizeof (sol_cma_glbl_listen_t));
1584 		} else
1585 			SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1586 			    "cma_fini_listen_root: "
1587 			    "ibcma_fini_root_chan failed");
1588 	}
1589 
1590 	mutex_exit(&sol_cma_glob_mutex);
1591 }
1592 
1593 typedef struct cma_event_async_arg {
1594 	struct rdma_cm_id	*idp;
1595 	enum rdma_cm_event_type	event;
1596 	int			status;
1597 	union {
1598 		struct rdma_conn_param	conn;
1599 		struct rdma_ud_param	param;
1600 	} un;
1601 	struct rdma_conn_param	*conn_param;
1602 	struct rdma_ud_param	*ud_paramp;
1603 } cma_event_async_arg_t;
1604 
1605 static void cma_generate_event_sync(struct rdma_cm_id *,
1606     enum rdma_cm_event_type, int, struct rdma_conn_param *,
1607     struct rdma_ud_param *);
1608 
1609 void
1610 cma_generate_event_thr(void *arg)
1611 {
1612 	cma_event_async_arg_t	*event_arg = (cma_event_async_arg_t *)arg;
1613 
1614 	cma_generate_event_sync(event_arg->idp, event_arg->event,
1615 	    event_arg->status, event_arg->conn_param,
1616 	    event_arg->ud_paramp);
1617 
1618 	if (event_arg->conn_param && event_arg->conn_param->private_data_len)
1619 		kmem_free((void *)event_arg->conn_param->private_data,
1620 		    event_arg->conn_param->private_data_len);
1621 	if (event_arg->ud_paramp && event_arg->ud_paramp->private_data_len)
1622 		kmem_free((void *)event_arg->ud_paramp->private_data,
1623 		    event_arg->ud_paramp->private_data_len);
1624 	kmem_free(arg, sizeof (cma_event_async_arg_t));
1625 }
1626 
1627 void
1628 cma_generate_event(struct rdma_cm_id *idp, enum rdma_cm_event_type event,
1629     int status, struct rdma_conn_param *conn_param,
1630     struct rdma_ud_param *ud_paramp)
1631 {
1632 	cma_event_async_arg_t	*event_arg;
1633 	sol_cma_chan_t		*chanp = (sol_cma_chan_t *)idp;
1634 
1635 	/*
1636 	 * Set SOL_CMA_CALLER_EVENT_PROGRESS to indicate event
1637 	 * notification is in progress, so that races between
1638 	 * rdma_destroy_id() and event notification is taken care.
1639 	 *
1640 	 * If rdma_destroy_id() has been called for this CMID, call
1641 	 * cma_generate_event_sync() which skips notification to the
1642 	 * consumer and handles the event.
1643 	 */
1644 	mutex_enter(&chanp->chan_mutex);
1645 	chanp->chan_cmid_destroy_state |= SOL_CMA_CALLER_EVENT_PROGRESS;
1646 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED) {
1647 		mutex_exit(&chanp->chan_mutex);
1648 		cma_generate_event_sync(idp, event, status, conn_param,
1649 		    ud_paramp);
1650 		return;
1651 	}
1652 	mutex_exit(&chanp->chan_mutex);
1653 
1654 	event_arg = kmem_zalloc(sizeof (cma_event_async_arg_t), KM_SLEEP);
1655 	event_arg->idp = idp;
1656 	event_arg->event = event;
1657 	event_arg->status = status;
1658 	event_arg->conn_param = NULL;
1659 	event_arg->ud_paramp = NULL;
1660 	if (conn_param && conn_param->private_data_len) {
1661 		bcopy(conn_param, &(event_arg->un.conn),
1662 		    sizeof (struct rdma_conn_param));
1663 		event_arg->conn_param = &(event_arg->un.conn);
1664 		event_arg->conn_param->private_data = kmem_zalloc(
1665 		    conn_param->private_data_len, KM_SLEEP);
1666 		bcopy(conn_param->private_data,
1667 		    (void *)event_arg->conn_param->private_data,
1668 		    conn_param->private_data_len);
1669 	} else if (conn_param && conn_param->private_data_len == 0) {
1670 		bcopy(conn_param, &(event_arg->un.conn),
1671 		    sizeof (struct rdma_conn_param));
1672 	} else if (ud_paramp) {
1673 		bcopy(ud_paramp, &(event_arg->un.param),
1674 		    sizeof (struct rdma_ud_param));
1675 		event_arg->ud_paramp = &(event_arg->un.param);
1676 		if (ud_paramp->private_data_len) {
1677 			event_arg->ud_paramp->private_data = kmem_zalloc(
1678 			    ud_paramp->private_data_len, KM_SLEEP);
1679 			bcopy(ud_paramp->private_data,
1680 			    (void *)event_arg->ud_paramp->private_data,
1681 			    ud_paramp->private_data_len);
1682 		} else if (ud_paramp->private_data) {
1683 			event_arg->ud_paramp->private_data =
1684 			    ud_paramp->private_data;
1685 		}
1686 	}
1687 
1688 	if (taskq_dispatch(system_taskq, cma_generate_event_thr,
1689 	    (void *)event_arg, TQ_SLEEP) == 0) {
1690 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1691 		    "generate_event_async: taskq_dispatch() failed!!");
1692 		mutex_enter(&chanp->chan_mutex);
1693 		chanp->chan_cmid_destroy_state &=
1694 		    ~SOL_CMA_CALLER_EVENT_PROGRESS;
1695 		if (chanp->chan_cmid_destroy_state &
1696 		    SOL_CMA_CALLER_CMID_DESTROYED)
1697 			cv_broadcast(&chanp->chan_destroy_cv);
1698 		mutex_exit(&chanp->chan_mutex);
1699 	}
1700 }
1701 
1702 static void
1703 cma_generate_event_sync(struct rdma_cm_id *idp, enum rdma_cm_event_type event,
1704     int status, struct rdma_conn_param *conn_param,
1705     struct rdma_ud_param *ud_paramp)
1706 {
1707 	struct rdma_cm_event	cm_event;
1708 	sol_cma_chan_t		*chanp = (sol_cma_chan_t *)idp;
1709 	struct rdma_cm_id	*root_idp = NULL;
1710 	sol_cma_chan_t		*root_chanp;
1711 	int			ret;
1712 	cma_chan_state_t	chan_state;
1713 
1714 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "generate_event_sync(%p, %x, "
1715 	    "%x, %p, %p", idp, event, status, conn_param, ud_paramp);
1716 
1717 	bzero(&cm_event, sizeof (cm_event));
1718 	cm_event.event = event;
1719 	cm_event.status = status;
1720 	if (conn_param)
1721 		bcopy((void *)conn_param, (void *)(&(cm_event.param.conn)),
1722 		    sizeof (struct rdma_conn_param));
1723 	else if (ud_paramp)
1724 		bcopy((void *)ud_paramp, (void *)(&(cm_event.param.ud)),
1725 		    sizeof (struct rdma_ud_param));
1726 
1727 	/*
1728 	 * If the consumer has destroyed the context for this CMID -
1729 	 * do not notify, skip to handling the sol_ofs specific
1730 	 * handling of the event.
1731 	 */
1732 	mutex_enter(&chanp->chan_mutex);
1733 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED) {
1734 		mutex_exit(&chanp->chan_mutex);
1735 		goto ofs_consume_event;
1736 	}
1737 	mutex_exit(&chanp->chan_mutex);
1738 
1739 	root_idp = CHAN_LISTEN_ROOT(chanp);
1740 	root_chanp = (sol_cma_chan_t *)root_idp;
1741 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "gen_event: root_idp %p",
1742 	    root_idp);
1743 
1744 	if (event == RDMA_CM_EVENT_CONNECT_REQUEST) {
1745 		/*
1746 		 * Update chan_req_state for the REQ CMID. Decrement
1747 		 * count of REQ CMIDs not notifed to consumer.
1748 		 */
1749 		ASSERT(root_idp);
1750 		mutex_enter(&root_chanp->chan_mutex);
1751 		root_chanp->chan_req_cnt--;
1752 #ifdef	DEBUG
1753 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str,
1754 		    "Dec req_cnt of %p IDP, idp %p, req_cnt %x",
1755 		    root_idp, idp, root_chanp->chan_req_cnt);
1756 #endif
1757 		mutex_exit(&root_chanp->chan_mutex);
1758 	}
1759 
1760 	/* Pass the event to the client */
1761 	ret = (idp->event_handler) (idp, &cm_event);
1762 
1763 	if (ret) {
1764 		/*
1765 		 * If the consumer returned failure :
1766 		 * 	CONNECT_REQUEST :
1767 		 * 	1. rdma_disconnect() to disconnect connection.
1768 		 * 	2. wakeup destroy, if destroy has been called
1769 		 * 		for this CMID
1770 		 * 	3. Destroy CMID if rdma_destroy has not been
1771 		 * 		called.
1772 		 * 	DISCONNECTED :
1773 		 * 	1. call cma_handle_nomore_events() to cleanup
1774 		 * 	Other Events :
1775 		 * 	1. Client is expected to destroy the CMID.
1776 		 */
1777 		if (event == RDMA_CM_EVENT_CONNECT_REQUEST) {
1778 			SOL_OFS_DPRINTF_L4(sol_rdmacm_dbg_str,
1779 			    "cma_generate_event_async: consumer failed %d "
1780 			    "event", event);
1781 			if (rdma_disconnect(idp)) {
1782 				SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1783 				    "generate_event_async: rdma_disconnect "
1784 				    "failed");
1785 			}
1786 			mutex_enter(&chanp->chan_mutex);
1787 			ASSERT(SOL_IS_SERVER_CMID(chanp));
1788 			chanp->chan_req_state = REQ_CMID_SERVER_NONE;
1789 			chanp->chan_cmid_destroy_state &=
1790 			    ~SOL_CMA_CALLER_EVENT_PROGRESS;
1791 			if (chanp->chan_cmid_destroy_state &
1792 			    SOL_CMA_CALLER_CMID_DESTROYED) {
1793 				cv_broadcast(&chanp->chan_destroy_cv);
1794 				mutex_exit(&chanp->chan_mutex);
1795 			} else {
1796 				mutex_exit(&chanp->chan_mutex);
1797 				rdma_destroy_id(idp);
1798 			}
1799 		} else if (event == RDMA_CM_EVENT_DISCONNECTED) {
1800 			SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1801 			    "generate_event_async: consumer failed %d event",
1802 			    event);
1803 			cma_handle_nomore_events(chanp);
1804 			mutex_enter(&chanp->chan_mutex);
1805 			chan_state = cma_get_chan_state(chanp);
1806 			chanp->chan_cmid_destroy_state &=
1807 			    ~SOL_CMA_CALLER_EVENT_PROGRESS;
1808 			if (chanp->chan_cmid_destroy_state &
1809 			    SOL_CMA_CALLER_CMID_DESTROYED) {
1810 				cv_broadcast(&chanp->chan_destroy_cv);
1811 				mutex_exit(&chanp->chan_mutex);
1812 			} else if (chan_state == SOL_CMA_CHAN_DESTROY_PENDING) {
1813 				/* rdma_destroy_id() called: destroy CMID */
1814 				mutex_exit(&chanp->chan_mutex);
1815 				cma_destroy_id((struct rdma_cm_id *)chanp);
1816 			} else
1817 				mutex_exit(&chanp->chan_mutex);
1818 		} else {
1819 			SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
1820 			    "generate_event_async: consumer failed %d event",
1821 			    event);
1822 		}
1823 
1824 		return;
1825 	}
1826 ofs_consume_event:
1827 	if (event == RDMA_CM_EVENT_DISCONNECTED) {
1828 		cma_chan_state_t	chan_state;
1829 
1830 		cma_handle_nomore_events(chanp);
1831 		mutex_enter(&chanp->chan_mutex);
1832 		chan_state = cma_get_chan_state(chanp);
1833 		chanp->chan_cmid_destroy_state &=
1834 		    ~SOL_CMA_CALLER_EVENT_PROGRESS;
1835 		if (chanp->chan_cmid_destroy_state &
1836 		    SOL_CMA_CALLER_CMID_DESTROYED) {
1837 			cv_broadcast(&chanp->chan_destroy_cv);
1838 			mutex_exit(&chanp->chan_mutex);
1839 		} else if (chan_state == SOL_CMA_CHAN_DESTROY_PENDING) {
1840 			/* If rdma_destroy_id() was called, destroy CMID */
1841 			mutex_exit(&chanp->chan_mutex);
1842 			cma_destroy_id((struct rdma_cm_id *)chanp);
1843 		} else
1844 			mutex_exit(&chanp->chan_mutex);
1845 		return;
1846 	} else if (IS_UDP_CMID(idp) && event == RDMA_CM_EVENT_UNREACHABLE) {
1847 		/*
1848 		 * If rdma_destroy_id() was called, destroy CMID
1849 		 * If not chan_connect_flag/ chan_req_state has already been
1850 		 * set to indicate that it can be deleted.
1851 		 */
1852 		mutex_enter(&chanp->chan_mutex);
1853 		chan_state = cma_get_chan_state(chanp);
1854 		chanp->chan_cmid_destroy_state &=
1855 		    ~SOL_CMA_CALLER_EVENT_PROGRESS;
1856 		if (chanp->chan_cmid_destroy_state &
1857 		    SOL_CMA_CALLER_CMID_DESTROYED) {
1858 			cv_broadcast(&chanp->chan_destroy_cv);
1859 			mutex_exit(&chanp->chan_mutex);
1860 		} else if (chan_state == SOL_CMA_CHAN_DESTROY_PENDING) {
1861 			mutex_exit(&chanp->chan_mutex);
1862 			cma_destroy_id(idp);
1863 		} else
1864 			mutex_exit(&chanp->chan_mutex);
1865 		return;
1866 	}
1867 
1868 	mutex_enter(&chanp->chan_mutex);
1869 	chanp->chan_cmid_destroy_state &= ~SOL_CMA_CALLER_EVENT_PROGRESS;
1870 	if (chanp->chan_cmid_destroy_state & SOL_CMA_CALLER_CMID_DESTROYED)
1871 		cv_broadcast(&chanp->chan_destroy_cv);
1872 	mutex_exit(&chanp->chan_mutex);
1873 }
1874 
1875 /* Local Static functions */
1876 static struct rdma_cm_id *
1877 cma_alloc_chan(rdma_cm_event_handler evt_hdlr, void *context,
1878     enum rdma_port_space ps)
1879 {
1880 	struct rdma_cm_id	*rdma_idp;
1881 	sol_cma_chan_t		*chanp;
1882 
1883 	chanp = kmem_zalloc(sizeof (sol_cma_chan_t), KM_SLEEP);
1884 	mutex_init(&chanp->chan_mutex, NULL, MUTEX_DRIVER, NULL);
1885 	cv_init(&chanp->chan_destroy_cv, NULL, CV_DRIVER, NULL);
1886 	rdma_idp = &(chanp->chan_rdma_cm);
1887 	rdma_idp->context = context;
1888 	rdma_idp->ps = ps;
1889 	rdma_idp->event_handler = evt_hdlr;
1890 	mutex_enter(&chanp->chan_mutex);
1891 	cma_set_chan_state(chanp, SOL_CMA_CHAN_IDLE);
1892 	avl_create(&chanp->chan_req_avl_tree, sol_cma_req_cmid_cmp,
1893 	    sizeof (sol_cma_chan_t),
1894 	    offsetof(sol_cma_chan_t, chan_req_avl_node));
1895 	avl_create(&chanp->chan_acpt_avl_tree, sol_cma_cmid_cmp,
1896 	    sizeof (sol_cma_chan_t),
1897 	    offsetof(sol_cma_chan_t, chan_acpt_avl_node));
1898 	mutex_exit(&chanp->chan_mutex);
1899 
1900 	return (rdma_idp);
1901 }
1902 
1903 /* Change the state of sol_cma_chan_t */
1904 static void
1905 cma_set_chan_state(sol_cma_chan_t *chanp, cma_chan_state_t newstate)
1906 {
1907 	ASSERT(MUTEX_HELD(&chanp->chan_mutex));
1908 	chanp->chan_state = newstate;
1909 }
1910 
1911 cma_chan_state_t
1912 cma_get_chan_state(sol_cma_chan_t *chanp)
1913 {
1914 	ASSERT(MUTEX_HELD(&chanp->chan_mutex));
1915 	return (chanp->chan_state);
1916 }
1917 
1918 /* Check & Swap the state of sol_ucma_chan_t */
1919 static int
1920 cma_cas_chan_state(sol_cma_chan_t *chanp, cma_chan_state_t prevstate,
1921     cma_chan_state_t newstate)
1922 {
1923 	int	ret = 0;
1924 
1925 	ASSERT(MUTEX_HELD(&chanp->chan_mutex));
1926 	if (chanp->chan_state != prevstate)
1927 		ret = -1;
1928 	else
1929 		chanp->chan_state = newstate;
1930 
1931 	return (ret);
1932 }
1933 
1934 static void
1935 cma_free_listen_list(struct rdma_cm_id *idp)
1936 {
1937 	genlist_entry_t	*entry;
1938 	sol_cma_chan_t	*chanp = (sol_cma_chan_t *)idp;
1939 
1940 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "cma_free_listen_list(%p)", idp);
1941 	mutex_enter(&chanp->chan_mutex);
1942 	entry = remove_genlist_head(&(CHAN_LISTEN_LIST(chanp)));
1943 	mutex_exit(&chanp->chan_mutex);
1944 	while (entry) {
1945 		sol_cma_chan_t	*ep_chanp;
1946 
1947 		ep_chanp = (sol_cma_chan_t *)entry->data;
1948 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "fini_ep_chan: %p",
1949 		    ep_chanp);
1950 		if (ibcma_fini_ep_chan(ep_chanp) == 0) {
1951 			genlist_entry_t		*entry1;
1952 			struct ib_device	*device;
1953 			cma_device_t		*cma_device;
1954 
1955 			ASSERT(ep_chanp->chan_listenp);
1956 			mutex_enter(&ep_chanp->chan_mutex);
1957 			entry1 = ep_chanp->chan_listenp->listen_ep_dev_entry;
1958 			device = ep_chanp->chan_listenp->listen_ep_device;
1959 			ASSERT(device);
1960 			cma_device = device->data;
1961 			delete_genlist(&cma_device->cma_epchan_list,
1962 			    entry1);
1963 			sol_cma_release_device(
1964 			    (struct rdma_cm_id *)ep_chanp);
1965 			mutex_exit(&ep_chanp->chan_mutex);
1966 			if (ep_chanp->chan_listenp)
1967 				kmem_free(ep_chanp->chan_listenp,
1968 				    sizeof (sol_cma_listen_info_t));
1969 
1970 			mutex_destroy(&ep_chanp->chan_mutex);
1971 			cv_destroy(&ep_chanp->chan_destroy_cv);
1972 			kmem_free(ep_chanp, sizeof (sol_cma_chan_t));
1973 			kmem_free(entry, sizeof (genlist_entry_t));
1974 		}
1975 
1976 		mutex_enter(&chanp->chan_mutex);
1977 		entry = remove_genlist_head(&(CHAN_LISTEN_LIST(chanp)));
1978 		mutex_exit(&chanp->chan_mutex);
1979 	}
1980 }
1981 
1982 /*
1983  * Destroy a listening CMID when :
1984  *	a. All CONNECTION REQUEST recieved have been rejected
1985  *	   or closed.
1986  *	b. No CONNECTION REQUEST recieved.
1987  * Do not destroy a listening CMID when :
1988  *	a. CONNECTION REQUEST has been recieved and not been
1989  *	   accepted from the passive / server side.
1990  *	b. CONNECTION REQUEST has been recieved and has been
1991  *	   accepted from the passive server side.
1992  *	Mark the listening CMID as destroy pending.
1993  *
1994  * For CMIDs created for rdma_connect() or created for a
1995  * CONNECT request, destroy the CMID only when :
1996  *       CONNECTION has been closed or rejected.
1997  *
1998  *       Mark the CMID as destroy pending.
1999  *
2000  * When a connection is rejected or closed :
2001  *	Check if flag indicates - destroy pending,
2002  *	cma_destroy_id() is called, this also does
2003  *
2004  *	If there is a listening CMID assosiated with it,
2005  *	   call cma_destroy_if(listen_cmid);
2006  */
2007 void
2008 cma_destroy_id(struct rdma_cm_id *idp)
2009 {
2010 	sol_cma_chan_t		*chanp = (sol_cma_chan_t *)idp;
2011 	cma_chan_state_t	state;
2012 	ulong_t			acpt_nodes, req_nodes;
2013 
2014 	mutex_enter(&chanp->chan_mutex);
2015 	acpt_nodes = avl_numnodes(&chanp->chan_acpt_avl_tree);
2016 	req_nodes = avl_numnodes(&chanp->chan_req_avl_tree);
2017 	state = cma_get_chan_state(chanp);
2018 	SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "cma_destroy_id(%p)- "
2019 	    "est CMIDs %ld, req CMID %ld, listen_root %p, state %x, %x",
2020 	    idp, acpt_nodes, req_nodes, chanp->listen_root,
2021 	    state, chanp->chan_req_state);
2022 
2023 	/*
2024 	 * If there are either REQ recieved or Established CMIDs just return.
2025 	 * rdma_destroy() for these CMIDs can be called by client later.
2026 	 */
2027 	if (acpt_nodes || req_nodes) {
2028 		cma_set_chan_state(chanp, SOL_CMA_CHAN_DESTROY_PENDING);
2029 		mutex_exit(&chanp->chan_mutex);
2030 		return;
2031 	}
2032 	cma_set_chan_state(chanp, SOL_CMA_CHAN_DESTROYING);
2033 	avl_destroy(&chanp->chan_req_avl_tree);
2034 	avl_destroy(&chanp->chan_acpt_avl_tree);
2035 
2036 	mutex_exit(&chanp->chan_mutex);
2037 	if (idp->route.path_rec) {
2038 		kmem_free(idp->route.path_rec,
2039 		    sizeof (struct ib_sa_path_rec) * idp->route.num_paths);
2040 		idp->route.path_rec = NULL;
2041 	}
2042 
2043 	switch (chanp->chan_xport_type) {
2044 	case SOL_CMA_XPORT_NONE :
2045 		break;
2046 	case SOL_CMA_XPORT_IB :
2047 		rdma_ib_destroy_id(idp);
2048 		break;
2049 #ifdef	IWARP_SUPPORT
2050 	case SOL_CMA_XPORT_IWARP :
2051 		rdma_iw_destroy_id(idp);
2052 		break;
2053 #endif	/* IWARP_SUPPORT */
2054 	default :
2055 		SOL_OFS_DPRINTF_L2(sol_rdmacm_dbg_str,
2056 		    "cma_destroy_id: Unsupported xport type %x",
2057 		    chanp->chan_xport_type);
2058 		break;
2059 	}
2060 
2061 	/*
2062 	 * Flush out & Free all listeners wrt to this ID
2063 	 * No locking is required as this code is executed
2064 	 * all REQ CMIDs have been destroyed. listen_list
2065 	 * will therefore not be modified during this loop.
2066 	 */
2067 	if (chanp->chan_listenp) {
2068 		cma_free_listen_list(idp);
2069 		cma_fini_listen_root(chanp);
2070 		kmem_free((void *)chanp->chan_listenp,
2071 		    sizeof (sol_cma_listen_info_t));
2072 		chanp->chan_listenp = NULL;
2073 	}
2074 
2075 	if (chanp->listen_root) {
2076 		struct rdma_cm_id	*root_idp;
2077 		sol_cma_chan_t		*root_chanp;
2078 
2079 		root_idp = chanp->listen_root;
2080 		root_chanp = (sol_cma_chan_t *)root_idp;
2081 		mutex_enter(&root_chanp->chan_mutex);
2082 		state = cma_get_chan_state(root_chanp);
2083 		acpt_nodes = avl_numnodes(&root_chanp->chan_acpt_avl_tree);
2084 		req_nodes = avl_numnodes(&root_chanp->chan_req_avl_tree);
2085 		mutex_exit(&root_chanp->chan_mutex);
2086 		SOL_OFS_DPRINTF_L5(sol_rdmacm_dbg_str, "cma_destroy_id(%p)-"
2087 		    " root idp %p, state %x, acpt_nodes %ld, req_nodes %ld",
2088 		    idp, root_idp, state, acpt_nodes, req_nodes);
2089 
2090 		if (state == SOL_CMA_CHAN_DESTROY_PENDING &&
2091 		    req_nodes == 0UL && acpt_nodes == 0UL) {
2092 			mutex_enter(&root_chanp->chan_mutex);
2093 			root_chanp->chan_req_state = REQ_CMID_SERVER_NONE;
2094 			mutex_exit(&root_chanp->chan_mutex);
2095 			cma_destroy_id(root_idp);
2096 		} else if (state == SOL_CMA_CHAN_DESTROY_WAIT &&
2097 		    req_nodes == 0UL && acpt_nodes == 0UL) {
2098 			mutex_enter(&root_chanp->chan_mutex);
2099 			cma_set_chan_state(root_chanp,
2100 			    SOL_CMA_CHAN_DESTROY_PENDING);
2101 			root_chanp->chan_req_state = REQ_CMID_SERVER_NONE;
2102 			cv_broadcast(&root_chanp->chan_destroy_cv);
2103 			mutex_exit(&root_chanp->chan_mutex);
2104 		}
2105 	}
2106 
2107 	mutex_destroy(&chanp->chan_mutex);
2108 	cv_destroy(&chanp->chan_destroy_cv);
2109 	kmem_free(chanp, sizeof (sol_cma_chan_t));
2110 }
2111 
2112 /*
2113  * Server TCP disconnect for an established channel.
2114  *	If destroy_id() has been called for the listening
2115  *	CMID and there are no more CMIDs with pending
2116  *	events corresponding to the listening CMID, free
2117  *	the listening CMID.
2118  *
2119  */
2120 static void
2121 cma_handle_nomore_events(sol_cma_chan_t *chanp)
2122 {
2123 	struct rdma_cm_id	*idp, *root_idp;
2124 	sol_cma_chan_t		*root_chanp;
2125 	cma_chan_state_t	state;
2126 	ulong_t			req_nodes, acpt_nodes;
2127 
2128 	idp = (struct rdma_cm_id *)chanp;
2129 	root_idp = CHAN_LISTEN_ROOT(chanp);
2130 	root_chanp = (sol_cma_chan_t *)root_idp;
2131 	if (!root_chanp)
2132 		return;
2133 
2134 	mutex_enter(&root_chanp->chan_mutex);
2135 	mutex_enter(&chanp->chan_mutex);
2136 	CHAN_LISTEN_ROOT(chanp) = NULL;
2137 	root_chanp->chan_req_total_cnt--;
2138 
2139 	/*
2140 	 * Removal of CMID from the AVL trees should already have been done
2141 	 * by now. Below code mainly as a  safety net.
2142 	 */
2143 	if (chanp->chan_req_state == REQ_CMID_ACCEPTED) {
2144 		ASSERT(chanp->chan_qp_hdl);
2145 		ASSERT(cma_get_acpt_idp(root_idp,
2146 		    chanp->chan_qp_hdl));
2147 		avl_remove(&root_chanp->chan_acpt_avl_tree, idp);
2148 		chanp->chan_req_state = REQ_CMID_SERVER_NONE;
2149 	}
2150 	if (REQ_CMID_IN_REQ_AVL_TREE(chanp)) {
2151 		ASSERT(chanp->chan_session_id);
2152 		ASSERT(cma_get_req_idp(root_idp,
2153 		    chanp->chan_session_id));
2154 		avl_remove(&root_chanp->chan_req_avl_tree, idp);
2155 		chanp->chan_req_state = REQ_CMID_SERVER_NONE;
2156 	}
2157 
2158 	state = cma_get_chan_state(root_chanp);
2159 	req_nodes = avl_numnodes(&root_chanp->chan_req_avl_tree);
2160 	acpt_nodes = avl_numnodes(&root_chanp->chan_acpt_avl_tree);
2161 	mutex_exit(&chanp->chan_mutex);
2162 	mutex_exit(&root_chanp->chan_mutex);
2163 	if (state == SOL_CMA_CHAN_DESTROY_PENDING && req_nodes == 0UL &&
2164 	    acpt_nodes == 0UL)
2165 		cma_destroy_id(root_idp);
2166 }
2167 
2168 extern int ib_modify_qp(struct ib_qp *, struct ib_qp_attr *, int);
2169 extern int rdma_init_qp_attr(struct rdma_cm_id *, struct ib_qp_attr *,
2170     int *);
2171 
2172 static int
2173 cma_init_ud_qp(sol_cma_chan_t *chanp, struct ib_qp *qp)
2174 {
2175 	struct ib_qp_attr qp_attr;
2176 	int qp_attr_mask, ret;
2177 
2178 	qp_attr.qp_state = IB_QPS_INIT;
2179 	ret = rdma_init_qp_attr(&chanp->chan_rdma_cm, &qp_attr, &qp_attr_mask);
2180 	if (ret)
2181 		return (ret);
2182 
2183 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
2184 	if (ret)
2185 		return (ret);
2186 
2187 	qp_attr.qp_state = IB_QPS_RTR;
2188 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
2189 	if (ret)
2190 		return (ret);
2191 
2192 	qp_attr.qp_state = IB_QPS_RTS;
2193 	qp_attr.sq_psn = 0;
2194 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
2195 
2196 	return (ret);
2197 }
2198 
2199 static int
2200 cma_init_conn_qp(sol_cma_chan_t *chanp, struct ib_qp *qp)
2201 {
2202 	struct ib_qp_attr qp_attr;
2203 	int qp_attr_mask, ret;
2204 
2205 	qp_attr.qp_state = IB_QPS_INIT;
2206 	ret = rdma_init_qp_attr(&chanp->chan_rdma_cm, &qp_attr, &qp_attr_mask);
2207 	if (ret)
2208 		return (ret);
2209 
2210 	return (ib_modify_qp(qp, &qp_attr, qp_attr_mask));
2211 }
2212 
2213 static inline int
2214 cma_is_ud_ps(enum rdma_port_space ps)
2215 {
2216 	return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
2217 }
2218 
2219 int
2220 rdma_create_qp(struct rdma_cm_id *idp, struct ib_pd *pd,
2221     struct ib_qp_init_attr *qp_init_attr)
2222 {
2223 	sol_cma_chan_t	*chanp;
2224 	struct ib_qp	*qp;
2225 	int		ret;
2226 	ofs_client_t	*dev_ofs_client;
2227 
2228 	ASSERT(idp);
2229 	chanp = (sol_cma_chan_t *)idp;
2230 	if (idp->device->node_guid != pd->device->node_guid)
2231 		return (-EINVAL);
2232 
2233 	dev_ofs_client = (ofs_client_t *)pd->device->clnt_hdl;
2234 	rdma_map_id2clnthdl(idp, dev_ofs_client->ibt_hdl, NULL);
2235 
2236 	qp = ib_create_qp(pd, qp_init_attr);
2237 	if ((uintptr_t)qp >= (uintptr_t)-0xFFF) {
2238 		return ((intptr_t)qp);
2239 	}
2240 	rdma_map_id2qphdl(idp, (void *)qp->ibt_qp);
2241 
2242 	if (cma_is_ud_ps(idp->ps)) {
2243 		ret = cma_init_ud_qp(chanp, qp);
2244 	} else {
2245 		ret = cma_init_conn_qp(chanp, qp);
2246 	}
2247 
2248 	if (ret) {
2249 		goto err;
2250 	}
2251 
2252 	idp->qp = qp;
2253 	chanp->chan_qp_num = qp->qp_num;
2254 	chanp->chan_is_srq = (qp->srq != NULL);
2255 	return (0);
2256 err:
2257 	(void) ib_destroy_qp(qp);
2258 	return (ret);
2259 }
2260 
2261 void
2262 rdma_destroy_qp(struct rdma_cm_id *idp)
2263 {
2264 	ASSERT(idp);
2265 	(void) ib_destroy_qp(idp->qp);
2266 	idp->qp = NULL;
2267 }
2268