xref: /illumos-gate/usr/src/uts/common/rpc/rdma_subr.c (revision bfed486a) 
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  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Copyright (c) 2008, The Ohio State University. All rights reserved.
28  *
29  * Portions of this source code is developed by the team members of
30  * The Ohio State University's Network-Based Computing Laboratory (NBCL),
31  * headed by Professor Dhabaleswar K. (DK) Panda.
32  *
33  * Acknowledgements to contributions from developors:
34  *   Ranjit Noronha: noronha@cse.ohio-state.edu
35  *   Lei Chai      : chail@cse.ohio-state.edu
36  *   Weikuan Yu    : yuw@cse.ohio-state.edu
37  *
38  */
39 
40 #include <sys/systm.h>
41 #include <sys/kstat.h>
42 #include <sys/modctl.h>
43 #include <sys/sdt.h>
44 #include <rpc/rpc_rdma.h>
45 
46 #include <sys/ib/ibtl/ibti.h>
47 
48 uint_t rdma_minchunk = RDMA_MINCHUNK;
49 
50 /*
51  * Globals
52  */
53 int rdma_modloaded = 0;		/* flag to load RDMA plugin modules */
54 int rdma_dev_available = 0;	/* if any RDMA device is loaded */
55 kmutex_t rdma_modload_lock;	/* protects rdma_modloaded flag */
56 
57 rdma_svc_wait_t rdma_wait;
58 
59 rdma_registry_t	*rdma_mod_head = NULL;	/* head for RDMA modules */
60 krwlock_t	rdma_lock;		/* protects rdma_mod_head list */
61 ldi_ident_t rpcmod_li = NULL;	/* identifies us with ldi_ framework */
62 
63 kmem_cache_t *clist_cache = NULL;
64 
65 /*
66  * Statics
67  */
68 ldi_handle_t rpcib_handle = NULL;
69 
70 /*
71  * Externs
72  */
73 extern	kstat_named_t	*rdmarcstat_ptr;
74 extern	uint_t		rdmarcstat_ndata;
75 extern	kstat_named_t	*rdmarsstat_ptr;
76 extern	uint_t		rdmarsstat_ndata;
77 
78 void rdma_kstat_init();
79 
80 /*
81  * RDMATF module registration routine.
82  * This routine is expected to be called by the init routine in
83  * the plugin modules.
84  */
85 rdma_stat
86 rdma_register_mod(rdma_mod_t *mod)
87 {
88 	rdma_registry_t **mp, *m;
89 
90 	if (mod->rdma_version != RDMATF_VERS) {
91 		return (RDMA_BADVERS);
92 	}
93 
94 	rw_enter(&rdma_lock, RW_WRITER);
95 	/*
96 	 * Ensure not already registered
97 	 */
98 	mp = &rdma_mod_head;
99 	while (*mp != NULL) {
100 		if (strncmp((*mp)->r_mod->rdma_api, mod->rdma_api,
101 		    KNC_STRSIZE) == 0) {
102 			if ((*mp)->r_mod_state == RDMA_MOD_INACTIVE) {
103 				(*mp)->r_mod_state = RDMA_MOD_ACTIVE;
104 				(*mp)->r_mod->rdma_ops = mod->rdma_ops;
105 				(*mp)->r_mod->rdma_count = mod->rdma_count;
106 				goto announce_hca;
107 			}
108 			rw_exit(&rdma_lock);
109 			return (RDMA_REG_EXIST);
110 		}
111 		mp = &((*mp)->r_next);
112 	}
113 
114 	/*
115 	 * New one, create and add to registry
116 	 */
117 	m = kmem_alloc(sizeof (rdma_registry_t), KM_SLEEP);
118 	m->r_mod = kmem_alloc(sizeof (rdma_mod_t), KM_SLEEP);
119 	*m->r_mod = *mod;
120 	m->r_next = NULL;
121 	m->r_mod->rdma_api = kmem_zalloc(KNC_STRSIZE, KM_SLEEP);
122 	(void) strncpy(m->r_mod->rdma_api, mod->rdma_api, KNC_STRSIZE);
123 	m->r_mod->rdma_api[KNC_STRSIZE - 1] = '\0';
124 	m->r_mod_state = RDMA_MOD_ACTIVE;
125 	*mp = m;
126 
127 announce_hca:
128 	rw_exit(&rdma_lock);
129 	/*
130 	 * Start the nfs service on the rdma xprts.
131 	 * (this notification mechanism will need to change when we support
132 	 * multiple hcas and have support for multiple rdma plugins).
133 	 */
134 	mutex_enter(&rdma_wait.svc_lock);
135 	rdma_wait.svc_stat = RDMA_HCA_ATTACH;
136 	cv_signal(&rdma_wait.svc_cv);
137 	mutex_exit(&rdma_wait.svc_lock);
138 
139 	return (RDMA_SUCCESS);
140 }
141 
142 /*
143  * RDMATF module unregistration routine.
144  * This routine is expected to be called by the fini routine in
145  * the plugin modules.
146  */
147 rdma_stat
148 rdma_unregister_mod(rdma_mod_t *mod)
149 {
150 	rdma_registry_t **m, *mmod = NULL;
151 
152 	rw_enter(&rdma_lock, RW_WRITER);
153 
154 	m = &rdma_mod_head;
155 	while (*m != NULL) {
156 		if (strncmp((*m)->r_mod->rdma_api, mod->rdma_api,
157 		    KNC_STRSIZE) != 0) {
158 			m = &((*m)->r_next);
159 			continue;
160 		}
161 		/*
162 		 * Check if any device attached, if so return error
163 		 */
164 		if (mod->rdma_count != 0) {
165 			rw_exit(&rdma_lock);
166 			return (RDMA_FAILED);
167 		}
168 		/*
169 		 * Found entry. Mark it inactive.
170 		 */
171 		mmod = *m;
172 		mmod->r_mod->rdma_count = 0;
173 		mmod->r_mod_state = RDMA_MOD_INACTIVE;
174 		break;
175 	}
176 
177 	rdma_modloaded = 0;
178 	rdma_dev_available = 0;
179 	rw_exit(&rdma_lock);
180 
181 	/*
182 	 * Stop the nfs service running on the rdma xprts.
183 	 * (this notification mechanism will need to change when we support
184 	 * multiple hcas and have support for multiple rdma plugins).
185 	 */
186 	mutex_enter(&rdma_wait.svc_lock);
187 	rdma_wait.svc_stat = RDMA_HCA_DETACH;
188 	cv_signal(&rdma_wait.svc_cv);
189 	mutex_exit(&rdma_wait.svc_lock);
190 
191 	/*
192 	 * Not found.
193 	 */
194 	return (RDMA_SUCCESS);
195 }
196 
197 struct clist *
198 clist_alloc(void)
199 {
200 	struct clist *clp;
201 
202 	clp = kmem_cache_alloc(clist_cache, KM_SLEEP);
203 
204 	bzero(clp, sizeof (*clp));
205 
206 	return (clp);
207 }
208 
209 /*
210  * Creates a new chunk list entry, and
211  * adds it to the end of a chunk list.
212  */
213 void
214 clist_add(struct clist **clp, uint32_t xdroff, int len,
215 	struct mrc *shandle, caddr_t saddr,
216 	struct mrc *dhandle, caddr_t daddr)
217 {
218 	struct clist *cl;
219 
220 	/* Find the end of the list */
221 
222 	while (*clp != NULL)
223 		clp = &((*clp)->c_next);
224 
225 	cl = clist_alloc();
226 	cl->c_xdroff = xdroff;
227 	cl->c_len = len;
228 	cl->w.c_saddr = (uint64_t)(uintptr_t)saddr;
229 	if (shandle)
230 		cl->c_smemhandle = *shandle;
231 	cl->u.c_daddr = (uint64_t)(uintptr_t)daddr;
232 	if (dhandle)
233 		cl->c_dmemhandle = *dhandle;
234 	cl->c_next = NULL;
235 
236 	*clp = cl;
237 }
238 
239 rdma_stat
240 clist_register(CONN *conn, struct clist *cl, clist_dstsrc dstsrc)
241 {
242 	struct clist *c;
243 	int status;
244 
245 	for (c = cl; c; c = c->c_next) {
246 		if (c->c_len <= 0)
247 			continue;
248 		switch (dstsrc) {
249 		case CLIST_REG_SOURCE:
250 			status = RDMA_REGMEMSYNC(conn,
251 			    (caddr_t)(struct as *)cl->c_adspc,
252 			    (caddr_t)(uintptr_t)c->w.c_saddr3, c->c_len,
253 			    &c->c_smemhandle, (void **)&c->c_ssynchandle,
254 			    (void *)c->rb_longbuf.rb_private);
255 			break;
256 		case CLIST_REG_DST:
257 			status = RDMA_REGMEMSYNC(conn,
258 			    (caddr_t)(struct as *)cl->c_adspc,
259 			    (caddr_t)(uintptr_t)c->u.c_daddr3, c->c_len,
260 			    &c->c_dmemhandle, (void **)&c->c_dsynchandle,
261 			    (void *)c->rb_longbuf.rb_private);
262 			break;
263 		default:
264 			return (RDMA_INVAL);
265 		}
266 		if (status != RDMA_SUCCESS) {
267 			(void) clist_deregister(conn, cl, dstsrc);
268 			return (status);
269 		}
270 	}
271 
272 	return (RDMA_SUCCESS);
273 }
274 
275 rdma_stat
276 clist_deregister(CONN *conn, struct clist *cl, clist_dstsrc dstsrc)
277 {
278 	struct clist *c;
279 
280 	for (c = cl; c; c = c->c_next) {
281 		switch (dstsrc) {
282 		case CLIST_REG_SOURCE:
283 			if (c->c_smemhandle.mrc_rmr != 0) {
284 				(void) RDMA_DEREGMEMSYNC(conn,
285 				    (caddr_t)(uintptr_t)c->w.c_saddr3,
286 				    c->c_smemhandle,
287 				    (void *)(uintptr_t)c->c_ssynchandle,
288 				    (void *)c->rb_longbuf.rb_private);
289 				c->c_smemhandle.mrc_rmr = 0;
290 				c->c_ssynchandle = NULL;
291 			}
292 			break;
293 		case CLIST_REG_DST:
294 			if (c->c_dmemhandle.mrc_rmr != 0) {
295 				(void) RDMA_DEREGMEMSYNC(conn,
296 				    (caddr_t)(uintptr_t)c->u.c_daddr3,
297 				    c->c_dmemhandle,
298 				    (void *)(uintptr_t)c->c_dsynchandle,
299 				    (void *)c->rb_longbuf.rb_private);
300 				c->c_dmemhandle.mrc_rmr = 0;
301 				c->c_dsynchandle = NULL;
302 			}
303 			break;
304 		default:
305 			return (RDMA_INVAL);
306 		}
307 	}
308 
309 	return (RDMA_SUCCESS);
310 }
311 
312 rdma_stat
313 clist_syncmem(CONN *conn, struct clist *cl, clist_dstsrc dstsrc)
314 {
315 	struct clist *c;
316 	rdma_stat status;
317 
318 	c = cl;
319 	switch (dstsrc) {
320 	case CLIST_REG_SOURCE:
321 		while (c != NULL) {
322 			if (c->c_ssynchandle) {
323 				status = RDMA_SYNCMEM(conn,
324 				    (void *)(uintptr_t)c->c_ssynchandle,
325 				    (caddr_t)(uintptr_t)c->w.c_saddr3,
326 				    c->c_len, 0);
327 				if (status != RDMA_SUCCESS)
328 					return (status);
329 			}
330 			c = c->c_next;
331 		}
332 		break;
333 	case CLIST_REG_DST:
334 		while (c != NULL) {
335 			if (c->c_ssynchandle) {
336 				status = RDMA_SYNCMEM(conn,
337 				    (void *)(uintptr_t)c->c_dsynchandle,
338 				    (caddr_t)(uintptr_t)c->u.c_daddr3,
339 				    c->c_len, 1);
340 				if (status != RDMA_SUCCESS)
341 					return (status);
342 			}
343 			c = c->c_next;
344 		}
345 		break;
346 	default:
347 		return (RDMA_INVAL);
348 	}
349 
350 	return (RDMA_SUCCESS);
351 }
352 
353 /*
354  * Frees up entries in chunk list
355  */
356 void
357 clist_free(struct clist *cl)
358 {
359 	struct clist *c = cl;
360 
361 	while (c != NULL) {
362 		cl = cl->c_next;
363 		kmem_cache_free(clist_cache, c);
364 		c = cl;
365 	}
366 }
367 
368 rdma_stat
369 rdma_clnt_postrecv(CONN *conn, uint32_t xid)
370 {
371 	struct clist *cl = NULL;
372 	rdma_stat retval;
373 	rdma_buf_t rbuf = {0};
374 
375 	rbuf.type = RECV_BUFFER;
376 	if (RDMA_BUF_ALLOC(conn, &rbuf)) {
377 		return (RDMA_NORESOURCE);
378 	}
379 
380 	clist_add(&cl, 0, rbuf.len, &rbuf.handle, rbuf.addr,
381 	    NULL, NULL);
382 	retval = RDMA_CLNT_RECVBUF(conn, cl, xid);
383 	clist_free(cl);
384 
385 	return (retval);
386 }
387 
388 rdma_stat
389 rdma_clnt_postrecv_remove(CONN *conn, uint32_t xid)
390 {
391 	return (RDMA_CLNT_RECVBUF_REMOVE(conn, xid));
392 }
393 
394 rdma_stat
395 rdma_svc_postrecv(CONN *conn)
396 {
397 	struct clist *cl = NULL;
398 	rdma_stat retval;
399 	rdma_buf_t rbuf = {0};
400 
401 	rbuf.type = RECV_BUFFER;
402 	if (RDMA_BUF_ALLOC(conn, &rbuf)) {
403 		retval = RDMA_NORESOURCE;
404 	} else {
405 		clist_add(&cl, 0, rbuf.len, &rbuf.handle, rbuf.addr,
406 		    NULL, NULL);
407 		retval = RDMA_SVC_RECVBUF(conn, cl);
408 		clist_free(cl);
409 	}
410 	return (retval);
411 }
412 
413 rdma_stat
414 rdma_buf_alloc(CONN *conn, rdma_buf_t *rbuf)
415 {
416 	return (RDMA_BUF_ALLOC(conn, rbuf));
417 }
418 
419 void
420 rdma_buf_free(CONN *conn, rdma_buf_t *rbuf)
421 {
422 	if (!rbuf || rbuf->addr == NULL) {
423 		return;
424 	}
425 	RDMA_BUF_FREE(conn, rbuf);
426 	bzero(rbuf, sizeof (rdma_buf_t));
427 }
428 
429 /*
430  * Caller is holding rdma_modload_lock mutex
431  */
432 int
433 rdma_modload()
434 {
435 	int status;
436 	ASSERT(MUTEX_HELD(&rdma_modload_lock));
437 	/*
438 	 * Load all available RDMA plugins which right now is only IB plugin.
439 	 * If no IB hardware is present, then quit right away.
440 	 * ENODEV -- For no device on the system
441 	 * EPROTONOSUPPORT -- For module not avilable either due to failure to
442 	 * load or some other reason.
443 	 */
444 	rdma_modloaded = 1;
445 	if (ibt_hw_is_present() == 0) {
446 		rdma_dev_available = 0;
447 		return (ENODEV);
448 	}
449 
450 	rdma_dev_available = 1;
451 	if (rpcmod_li == NULL)
452 		return (EPROTONOSUPPORT);
453 
454 	status = ldi_open_by_name("/devices/ib/rpcib@0:rpcib",
455 	    FREAD | FWRITE, kcred,
456 	    &rpcib_handle, rpcmod_li);
457 
458 	if (status != 0)
459 		return (EPROTONOSUPPORT);
460 
461 
462 	/*
463 	 * We will need to reload the plugin module after it was unregistered
464 	 * but the resources below need to allocated only the first time.
465 	 */
466 	if (!clist_cache) {
467 		clist_cache = kmem_cache_create("rdma_clist",
468 		    sizeof (struct clist), _POINTER_ALIGNMENT, NULL,
469 		    NULL, NULL, NULL, 0, 0);
470 		rdma_kstat_init();
471 	}
472 
473 	(void) ldi_close(rpcib_handle, FREAD|FWRITE, kcred);
474 
475 	return (0);
476 }
477 
478 void
479 rdma_kstat_init(void)
480 {
481 	kstat_t *ksp;
482 
483 	/*
484 	 * The RDMA framework doesn't know how to deal with Zones, and is
485 	 * only available in the global zone.
486 	 */
487 	ASSERT(INGLOBALZONE(curproc));
488 	ksp = kstat_create_zone("unix", 0, "rpc_rdma_client", "rpc",
489 	    KSTAT_TYPE_NAMED, rdmarcstat_ndata,
490 	    KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_WRITABLE, GLOBAL_ZONEID);
491 	if (ksp) {
492 		ksp->ks_data = (void *) rdmarcstat_ptr;
493 		kstat_install(ksp);
494 	}
495 
496 	ksp = kstat_create_zone("unix", 0, "rpc_rdma_server", "rpc",
497 	    KSTAT_TYPE_NAMED, rdmarsstat_ndata,
498 	    KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_WRITABLE, GLOBAL_ZONEID);
499 	if (ksp) {
500 		ksp->ks_data = (void *) rdmarsstat_ptr;
501 		kstat_install(ksp);
502 	}
503 }
504 
505 rdma_stat
506 rdma_kwait(void)
507 {
508 	int ret;
509 	rdma_stat stat;
510 
511 	mutex_enter(&rdma_wait.svc_lock);
512 
513 	ret = cv_wait_sig(&rdma_wait.svc_cv, &rdma_wait.svc_lock);
514 
515 	/*
516 	 * If signalled by a hca attach/detach, pass the right
517 	 * stat back.
518 	 */
519 
520 	if (ret)
521 		stat =  rdma_wait.svc_stat;
522 	else
523 		stat = RDMA_INTR;
524 
525 	mutex_exit(&rdma_wait.svc_lock);
526 
527 	return (stat);
528 }
529