xref: /illumos-gate/usr/src/uts/common/rpc/svc_rdma.c (revision 4c28a617e3922d92a58e813a5b955eb526b9c386)
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 (c) 1983, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012 by Delphix. All rights reserved.
24  * Copyright 2013 Nexenta Systems, Inc.  All rights reserved.
25  */
26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
27 /* All Rights Reserved */
28 /*
29  * Portions of this source code were derived from Berkeley
30  * 4.3 BSD under license from the Regents of the University of
31  * California.
32  */
33 
34 /*
35  * Server side of RPC over RDMA in the kernel.
36  */
37 
38 #include <sys/param.h>
39 #include <sys/types.h>
40 #include <sys/user.h>
41 #include <sys/sysmacros.h>
42 #include <sys/proc.h>
43 #include <sys/file.h>
44 #include <sys/errno.h>
45 #include <sys/kmem.h>
46 #include <sys/debug.h>
47 #include <sys/systm.h>
48 #include <sys/cmn_err.h>
49 #include <sys/kstat.h>
50 #include <sys/vtrace.h>
51 #include <sys/debug.h>
52 
53 #include <rpc/types.h>
54 #include <rpc/xdr.h>
55 #include <rpc/auth.h>
56 #include <rpc/clnt.h>
57 #include <rpc/rpc_msg.h>
58 #include <rpc/svc.h>
59 #include <rpc/rpc_rdma.h>
60 #include <sys/ddi.h>
61 #include <sys/sunddi.h>
62 
63 #include <inet/common.h>
64 #include <inet/ip.h>
65 #include <inet/ip6.h>
66 
67 #include <nfs/nfs.h>
68 #include <sys/sdt.h>
69 
70 #define	SVC_RDMA_SUCCESS 0
71 #define	SVC_RDMA_FAIL -1
72 
73 #define	SVC_CREDIT_FACTOR (0.5)
74 
75 #define	MSG_IS_RPCSEC_GSS(msg)		\
76 	((msg)->rm_reply.rp_acpt.ar_verf.oa_flavor == RPCSEC_GSS)
77 
78 
79 uint32_t rdma_bufs_granted = RDMA_BUFS_GRANT;
80 
81 /*
82  * RDMA transport specific data associated with SVCMASTERXPRT
83  */
84 struct rdma_data {
85 	SVCMASTERXPRT 	*rd_xprt;	/* back ptr to SVCMASTERXPRT */
86 	struct rdma_svc_data rd_data;	/* rdma data */
87 	rdma_mod_t	*r_mod;		/* RDMA module containing ops ptr */
88 };
89 
90 /*
91  * Plugin connection specific data stashed away in clone SVCXPRT
92  */
93 struct clone_rdma_data {
94 	bool_t		cloned;		/* xprt cloned for thread processing */
95 	CONN		*conn;		/* RDMA connection */
96 	rdma_buf_t	rpcbuf;		/* RPC req/resp buffer */
97 	struct clist	*cl_reply;	/* reply chunk buffer info */
98 	struct clist	*cl_wlist;		/* write list clist */
99 };
100 
101 
102 #define	MAXADDRLEN	128	/* max length for address mask */
103 
104 /*
105  * Routines exported through ops vector.
106  */
107 static bool_t		svc_rdma_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
108 static bool_t		svc_rdma_ksend(SVCXPRT *, struct rpc_msg *);
109 static bool_t		svc_rdma_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
110 static bool_t		svc_rdma_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
111 void			svc_rdma_kdestroy(SVCMASTERXPRT *);
112 static int		svc_rdma_kdup(struct svc_req *, caddr_t, int,
113 				struct dupreq **, bool_t *);
114 static void		svc_rdma_kdupdone(struct dupreq *, caddr_t,
115 				void (*)(), int, int);
116 static int32_t		*svc_rdma_kgetres(SVCXPRT *, int);
117 static void		svc_rdma_kfreeres(SVCXPRT *);
118 static void		svc_rdma_kclone_destroy(SVCXPRT *);
119 static void		svc_rdma_kstart(SVCMASTERXPRT *);
120 void			svc_rdma_kstop(SVCMASTERXPRT *);
121 static void		svc_rdma_kclone_xprt(SVCXPRT *, SVCXPRT *);
122 static void		svc_rdma_ktattrs(SVCXPRT *, int, void **);
123 
124 static int	svc_process_long_reply(SVCXPRT *, xdrproc_t,
125 			caddr_t, struct rpc_msg *, bool_t, int *,
126 			int *, int *, unsigned int *);
127 
128 static int	svc_compose_rpcmsg(SVCXPRT *, CONN *, xdrproc_t,
129 			caddr_t, rdma_buf_t *, XDR **, struct rpc_msg *,
130 			bool_t, uint_t *);
131 static bool_t rpcmsg_length(xdrproc_t,
132 		caddr_t,
133 		struct rpc_msg *, bool_t, int);
134 
135 /*
136  * Server transport operations vector.
137  */
138 struct svc_ops rdma_svc_ops = {
139 	svc_rdma_krecv,		/* Get requests */
140 	svc_rdma_kgetargs,	/* Deserialize arguments */
141 	svc_rdma_ksend,		/* Send reply */
142 	svc_rdma_kfreeargs,	/* Free argument data space */
143 	svc_rdma_kdestroy,	/* Destroy transport handle */
144 	svc_rdma_kdup,		/* Check entry in dup req cache */
145 	svc_rdma_kdupdone,	/* Mark entry in dup req cache as done */
146 	svc_rdma_kgetres,	/* Get pointer to response buffer */
147 	svc_rdma_kfreeres,	/* Destroy pre-serialized response header */
148 	svc_rdma_kclone_destroy,	/* Destroy a clone xprt */
149 	svc_rdma_kstart,	/* Tell `ready-to-receive' to rpcmod */
150 	svc_rdma_kclone_xprt,	/* Transport specific clone xprt */
151 	svc_rdma_ktattrs	/* Get Transport Attributes */
152 };
153 
154 /*
155  * Server statistics
156  * NOTE: This structure type is duplicated in the NFS fast path.
157  */
158 struct {
159 	kstat_named_t	rscalls;
160 	kstat_named_t	rsbadcalls;
161 	kstat_named_t	rsnullrecv;
162 	kstat_named_t	rsbadlen;
163 	kstat_named_t	rsxdrcall;
164 	kstat_named_t	rsdupchecks;
165 	kstat_named_t	rsdupreqs;
166 	kstat_named_t	rslongrpcs;
167 	kstat_named_t	rstotalreplies;
168 	kstat_named_t	rstotallongreplies;
169 	kstat_named_t	rstotalinlinereplies;
170 } rdmarsstat = {
171 	{ "calls",	KSTAT_DATA_UINT64 },
172 	{ "badcalls",	KSTAT_DATA_UINT64 },
173 	{ "nullrecv",	KSTAT_DATA_UINT64 },
174 	{ "badlen",	KSTAT_DATA_UINT64 },
175 	{ "xdrcall",	KSTAT_DATA_UINT64 },
176 	{ "dupchecks",	KSTAT_DATA_UINT64 },
177 	{ "dupreqs",	KSTAT_DATA_UINT64 },
178 	{ "longrpcs",	KSTAT_DATA_UINT64 },
179 	{ "totalreplies",	KSTAT_DATA_UINT64 },
180 	{ "totallongreplies",	KSTAT_DATA_UINT64 },
181 	{ "totalinlinereplies",	KSTAT_DATA_UINT64 },
182 };
183 
184 kstat_named_t *rdmarsstat_ptr = (kstat_named_t *)&rdmarsstat;
185 uint_t rdmarsstat_ndata = sizeof (rdmarsstat) / sizeof (kstat_named_t);
186 
187 #define	RSSTAT_INCR(x)	atomic_inc_64(&rdmarsstat.x.value.ui64)
188 /*
189  * Create a transport record.
190  * The transport record, output buffer, and private data structure
191  * are allocated.  The output buffer is serialized into using xdrmem.
192  * There is one transport record per user process which implements a
193  * set of services.
194  */
195 /* ARGSUSED */
196 int
197 svc_rdma_kcreate(char *netid, SVC_CALLOUT_TABLE *sct, int id,
198     rdma_xprt_group_t *started_xprts)
199 {
200 	int error;
201 	SVCMASTERXPRT *xprt;
202 	struct rdma_data *rd;
203 	rdma_registry_t *rmod;
204 	rdma_xprt_record_t *xprt_rec;
205 	queue_t	*q;
206 	/*
207 	 * modload the RDMA plugins is not already done.
208 	 */
209 	if (!rdma_modloaded) {
210 		/*CONSTANTCONDITION*/
211 		ASSERT(sizeof (struct clone_rdma_data) <= SVC_P2LEN);
212 
213 		mutex_enter(&rdma_modload_lock);
214 		if (!rdma_modloaded) {
215 			error = rdma_modload();
216 		}
217 		mutex_exit(&rdma_modload_lock);
218 
219 		if (error)
220 			return (error);
221 	}
222 
223 	/*
224 	 * master_xprt_count is the count of master transport handles
225 	 * that were successfully created and are ready to recieve for
226 	 * RDMA based access.
227 	 */
228 	error = 0;
229 	xprt_rec = NULL;
230 	rw_enter(&rdma_lock, RW_READER);
231 	if (rdma_mod_head == NULL) {
232 		started_xprts->rtg_count = 0;
233 		rw_exit(&rdma_lock);
234 		if (rdma_dev_available)
235 			return (EPROTONOSUPPORT);
236 		else
237 			return (ENODEV);
238 	}
239 
240 	/*
241 	 * If we have reached here, then atleast one RDMA plugin has loaded.
242 	 * Create a master_xprt, make it start listenining on the device,
243 	 * if an error is generated, record it, we might need to shut
244 	 * the master_xprt.
245 	 * SVC_START() calls svc_rdma_kstart which calls plugin binding
246 	 * routines.
247 	 */
248 	for (rmod = rdma_mod_head; rmod != NULL; rmod = rmod->r_next) {
249 
250 		/*
251 		 * One SVCMASTERXPRT per RDMA plugin.
252 		 */
253 		xprt = kmem_zalloc(sizeof (*xprt), KM_SLEEP);
254 		xprt->xp_ops = &rdma_svc_ops;
255 		xprt->xp_sct = sct;
256 		xprt->xp_type = T_RDMA;
257 		mutex_init(&xprt->xp_req_lock, NULL, MUTEX_DEFAULT, NULL);
258 		mutex_init(&xprt->xp_thread_lock, NULL, MUTEX_DEFAULT, NULL);
259 		xprt->xp_req_head = (mblk_t *)0;
260 		xprt->xp_req_tail = (mblk_t *)0;
261 		xprt->xp_full = FALSE;
262 		xprt->xp_enable = FALSE;
263 		xprt->xp_reqs = 0;
264 		xprt->xp_size = 0;
265 		xprt->xp_threads = 0;
266 		xprt->xp_detached_threads = 0;
267 
268 		rd = kmem_zalloc(sizeof (*rd), KM_SLEEP);
269 		xprt->xp_p2 = (caddr_t)rd;
270 		rd->rd_xprt = xprt;
271 		rd->r_mod = rmod->r_mod;
272 
273 		q = &rd->rd_data.q;
274 		xprt->xp_wq = q;
275 		q->q_ptr = &rd->rd_xprt;
276 		xprt->xp_netid = NULL;
277 
278 		/*
279 		 * Each of the plugins will have their own Service ID
280 		 * to listener specific mapping, like port number for VI
281 		 * and service name for IB.
282 		 */
283 		rd->rd_data.svcid = id;
284 		error = svc_xprt_register(xprt, id);
285 		if (error) {
286 			DTRACE_PROBE(krpc__e__svcrdma__xprt__reg);
287 			goto cleanup;
288 		}
289 
290 		SVC_START(xprt);
291 		if (!rd->rd_data.active) {
292 			svc_xprt_unregister(xprt);
293 			error = rd->rd_data.err_code;
294 			goto cleanup;
295 		}
296 
297 		/*
298 		 * This is set only when there is atleast one or more
299 		 * transports successfully created. We insert the pointer
300 		 * to the created RDMA master xprt into a separately maintained
301 		 * list. This way we can easily reference it later to cleanup,
302 		 * when NFS kRPC service pool is going away/unregistered.
303 		 */
304 		started_xprts->rtg_count ++;
305 		xprt_rec = kmem_alloc(sizeof (*xprt_rec), KM_SLEEP);
306 		xprt_rec->rtr_xprt_ptr = xprt;
307 		xprt_rec->rtr_next = started_xprts->rtg_listhead;
308 		started_xprts->rtg_listhead = xprt_rec;
309 		continue;
310 cleanup:
311 		SVC_DESTROY(xprt);
312 		if (error == RDMA_FAILED)
313 			error = EPROTONOSUPPORT;
314 	}
315 
316 	rw_exit(&rdma_lock);
317 
318 	/*
319 	 * Don't return any error even if a single plugin was started
320 	 * successfully.
321 	 */
322 	if (started_xprts->rtg_count == 0)
323 		return (error);
324 	return (0);
325 }
326 
327 /*
328  * Cleanup routine for freeing up memory allocated by
329  * svc_rdma_kcreate()
330  */
331 void
332 svc_rdma_kdestroy(SVCMASTERXPRT *xprt)
333 {
334 	struct rdma_data *rd = (struct rdma_data *)xprt->xp_p2;
335 
336 
337 	mutex_destroy(&xprt->xp_req_lock);
338 	mutex_destroy(&xprt->xp_thread_lock);
339 	kmem_free(rd, sizeof (*rd));
340 	kmem_free(xprt, sizeof (*xprt));
341 }
342 
343 
344 static void
345 svc_rdma_kstart(SVCMASTERXPRT *xprt)
346 {
347 	struct rdma_svc_data *svcdata;
348 	rdma_mod_t *rmod;
349 
350 	svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data;
351 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
352 
353 	/*
354 	 * Create a listener for  module at this port
355 	 */
356 
357 	if (rmod->rdma_count != 0)
358 		(*rmod->rdma_ops->rdma_svc_listen)(svcdata);
359 	else
360 		svcdata->err_code = RDMA_FAILED;
361 }
362 
363 void
364 svc_rdma_kstop(SVCMASTERXPRT *xprt)
365 {
366 	struct rdma_svc_data *svcdata;
367 	rdma_mod_t *rmod;
368 
369 	svcdata	= &((struct rdma_data *)xprt->xp_p2)->rd_data;
370 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
371 
372 	/*
373 	 * Call the stop listener routine for each plugin. If rdma_count is
374 	 * already zero set active to zero.
375 	 */
376 	if (rmod->rdma_count != 0)
377 		(*rmod->rdma_ops->rdma_svc_stop)(svcdata);
378 	else
379 		svcdata->active = 0;
380 	if (svcdata->active)
381 		DTRACE_PROBE(krpc__e__svcrdma__kstop);
382 }
383 
384 /* ARGSUSED */
385 static void
386 svc_rdma_kclone_destroy(SVCXPRT *clone_xprt)
387 {
388 
389 	struct clone_rdma_data *cdrp;
390 	cdrp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
391 
392 	/*
393 	 * Only free buffers and release connection when cloned is set.
394 	 */
395 	if (cdrp->cloned != TRUE)
396 		return;
397 
398 	rdma_buf_free(cdrp->conn, &cdrp->rpcbuf);
399 	if (cdrp->cl_reply) {
400 		clist_free(cdrp->cl_reply);
401 		cdrp->cl_reply = NULL;
402 	}
403 	RDMA_REL_CONN(cdrp->conn);
404 
405 	cdrp->cloned = 0;
406 }
407 
408 /*
409  * Clone the xprt specific information.  It will be freed by
410  * SVC_CLONE_DESTROY.
411  */
412 static void
413 svc_rdma_kclone_xprt(SVCXPRT *src_xprt, SVCXPRT *dst_xprt)
414 {
415 	struct clone_rdma_data *srcp2;
416 	struct clone_rdma_data *dstp2;
417 
418 	srcp2 = (struct clone_rdma_data *)src_xprt->xp_p2buf;
419 	dstp2 = (struct clone_rdma_data *)dst_xprt->xp_p2buf;
420 
421 	if (srcp2->conn != NULL) {
422 		srcp2->cloned = TRUE;
423 		*dstp2 = *srcp2;
424 	}
425 }
426 
427 static void
428 svc_rdma_ktattrs(SVCXPRT *clone_xprt, int attrflag, void **tattr)
429 {
430 	CONN	*conn;
431 	*tattr = NULL;
432 
433 	switch (attrflag) {
434 	case SVC_TATTR_ADDRMASK:
435 		conn = ((struct clone_rdma_data *)clone_xprt->xp_p2buf)->conn;
436 		ASSERT(conn != NULL);
437 		if (conn)
438 			*tattr = (void *)&conn->c_addrmask;
439 	}
440 }
441 
442 static bool_t
443 svc_rdma_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
444 {
445 	XDR	*xdrs;
446 	CONN	*conn;
447 	rdma_recv_data_t	*rdp = (rdma_recv_data_t *)mp->b_rptr;
448 	struct clone_rdma_data *crdp;
449 	struct clist	*cl = NULL;
450 	struct clist	*wcl = NULL;
451 	struct clist	*cllong = NULL;
452 
453 	rdma_stat	status;
454 	uint32_t vers, op, pos, xid;
455 	uint32_t rdma_credit;
456 	uint32_t wcl_total_length = 0;
457 	bool_t	wwl = FALSE;
458 
459 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
460 	RSSTAT_INCR(rscalls);
461 	conn = rdp->conn;
462 
463 	status = rdma_svc_postrecv(conn);
464 	if (status != RDMA_SUCCESS) {
465 		DTRACE_PROBE(krpc__e__svcrdma__krecv__postrecv);
466 		goto badrpc_call;
467 	}
468 
469 	xdrs = &clone_xprt->xp_xdrin;
470 	xdrmem_create(xdrs, rdp->rpcmsg.addr, rdp->rpcmsg.len, XDR_DECODE);
471 	xid = *(uint32_t *)rdp->rpcmsg.addr;
472 	XDR_SETPOS(xdrs, sizeof (uint32_t));
473 
474 	if (! xdr_u_int(xdrs, &vers) ||
475 	    ! xdr_u_int(xdrs, &rdma_credit) ||
476 	    ! xdr_u_int(xdrs, &op)) {
477 		DTRACE_PROBE(krpc__e__svcrdma__krecv__uint);
478 		goto xdr_err;
479 	}
480 
481 	/* Checking if the status of the recv operation was normal */
482 	if (rdp->status != 0) {
483 		DTRACE_PROBE1(krpc__e__svcrdma__krecv__invalid__status,
484 		    int, rdp->status);
485 		goto badrpc_call;
486 	}
487 
488 	if (! xdr_do_clist(xdrs, &cl)) {
489 		DTRACE_PROBE(krpc__e__svcrdma__krecv__do__clist);
490 		goto xdr_err;
491 	}
492 
493 	if (!xdr_decode_wlist_svc(xdrs, &wcl, &wwl, &wcl_total_length, conn)) {
494 		DTRACE_PROBE(krpc__e__svcrdma__krecv__decode__wlist);
495 		if (cl)
496 			clist_free(cl);
497 		goto xdr_err;
498 	}
499 	crdp->cl_wlist = wcl;
500 
501 	crdp->cl_reply = NULL;
502 	(void) xdr_decode_reply_wchunk(xdrs, &crdp->cl_reply);
503 
504 	/*
505 	 * A chunk at 0 offset indicates that the RPC call message
506 	 * is in a chunk. Get the RPC call message chunk.
507 	 */
508 	if (cl != NULL && op == RDMA_NOMSG) {
509 
510 		/* Remove RPC call message chunk from chunklist */
511 		cllong = cl;
512 		cl = cl->c_next;
513 		cllong->c_next = NULL;
514 
515 
516 		/* Allocate and register memory for the RPC call msg chunk */
517 		cllong->rb_longbuf.type = RDMA_LONG_BUFFER;
518 		cllong->rb_longbuf.len = cllong->c_len > LONG_REPLY_LEN ?
519 		    cllong->c_len : LONG_REPLY_LEN;
520 
521 		if (rdma_buf_alloc(conn, &cllong->rb_longbuf)) {
522 			clist_free(cllong);
523 			goto cll_malloc_err;
524 		}
525 
526 		cllong->u.c_daddr3 = cllong->rb_longbuf.addr;
527 
528 		if (cllong->u.c_daddr == NULL) {
529 			DTRACE_PROBE(krpc__e__svcrdma__krecv__nomem);
530 			rdma_buf_free(conn, &cllong->rb_longbuf);
531 			clist_free(cllong);
532 			goto cll_malloc_err;
533 		}
534 
535 		status = clist_register(conn, cllong, CLIST_REG_DST);
536 		if (status) {
537 			DTRACE_PROBE(krpc__e__svcrdma__krecv__clist__reg);
538 			rdma_buf_free(conn, &cllong->rb_longbuf);
539 			clist_free(cllong);
540 			goto cll_malloc_err;
541 		}
542 
543 		/*
544 		 * Now read the RPC call message in
545 		 */
546 		status = RDMA_READ(conn, cllong, WAIT);
547 		if (status) {
548 			DTRACE_PROBE(krpc__e__svcrdma__krecv__read);
549 			(void) clist_deregister(conn, cllong);
550 			rdma_buf_free(conn, &cllong->rb_longbuf);
551 			clist_free(cllong);
552 			goto cll_malloc_err;
553 		}
554 
555 		status = clist_syncmem(conn, cllong, CLIST_REG_DST);
556 		(void) clist_deregister(conn, cllong);
557 
558 		xdrrdma_create(xdrs, (caddr_t)(uintptr_t)cllong->u.c_daddr3,
559 		    cllong->c_len, 0, cl, XDR_DECODE, conn);
560 
561 		crdp->rpcbuf = cllong->rb_longbuf;
562 		crdp->rpcbuf.len = cllong->c_len;
563 		clist_free(cllong);
564 		RDMA_BUF_FREE(conn, &rdp->rpcmsg);
565 	} else {
566 		pos = XDR_GETPOS(xdrs);
567 		xdrrdma_create(xdrs, rdp->rpcmsg.addr + pos,
568 		    rdp->rpcmsg.len - pos, 0, cl, XDR_DECODE, conn);
569 		crdp->rpcbuf = rdp->rpcmsg;
570 
571 		/* Use xdrrdmablk_ops to indicate there is a read chunk list */
572 		if (cl != NULL) {
573 			int32_t flg = XDR_RDMA_RLIST_REG;
574 
575 			XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
576 			xdrs->x_ops = &xdrrdmablk_ops;
577 		}
578 	}
579 
580 	if (crdp->cl_wlist) {
581 		int32_t flg = XDR_RDMA_WLIST_REG;
582 
583 		XDR_CONTROL(xdrs, XDR_RDMA_SET_WLIST, crdp->cl_wlist);
584 		XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
585 	}
586 
587 	if (! xdr_callmsg(xdrs, msg)) {
588 		DTRACE_PROBE(krpc__e__svcrdma__krecv__callmsg);
589 		RSSTAT_INCR(rsxdrcall);
590 		goto callmsg_err;
591 	}
592 
593 	/*
594 	 * Point the remote transport address in the service_transport
595 	 * handle at the address in the request.
596 	 */
597 	clone_xprt->xp_rtaddr.buf = conn->c_raddr.buf;
598 	clone_xprt->xp_rtaddr.len = conn->c_raddr.len;
599 	clone_xprt->xp_rtaddr.maxlen = conn->c_raddr.len;
600 
601 	clone_xprt->xp_lcladdr.buf = conn->c_laddr.buf;
602 	clone_xprt->xp_lcladdr.len = conn->c_laddr.len;
603 	clone_xprt->xp_lcladdr.maxlen = conn->c_laddr.len;
604 
605 	/*
606 	 * In case of RDMA, connection management is
607 	 * entirely done in rpcib module and netid in the
608 	 * SVCMASTERXPRT is NULL. Initialize the clone netid
609 	 * from the connection.
610 	 */
611 
612 	clone_xprt->xp_netid = conn->c_netid;
613 
614 	clone_xprt->xp_xid = xid;
615 	crdp->conn = conn;
616 
617 	freeb(mp);
618 
619 	return (TRUE);
620 
621 callmsg_err:
622 	rdma_buf_free(conn, &crdp->rpcbuf);
623 
624 cll_malloc_err:
625 	if (cl)
626 		clist_free(cl);
627 xdr_err:
628 	XDR_DESTROY(xdrs);
629 
630 badrpc_call:
631 	RDMA_BUF_FREE(conn, &rdp->rpcmsg);
632 	RDMA_REL_CONN(conn);
633 	freeb(mp);
634 	RSSTAT_INCR(rsbadcalls);
635 	return (FALSE);
636 }
637 
638 static int
639 svc_process_long_reply(SVCXPRT * clone_xprt,
640     xdrproc_t xdr_results, caddr_t xdr_location,
641     struct rpc_msg *msg, bool_t has_args, int *msglen,
642     int *freelen, int *numchunks, unsigned int *final_len)
643 {
644 	int status;
645 	XDR xdrslong;
646 	struct clist *wcl = NULL;
647 	int count = 0;
648 	int alloc_len;
649 	char  *memp;
650 	rdma_buf_t long_rpc = {0};
651 	struct clone_rdma_data *crdp;
652 
653 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
654 
655 	bzero(&xdrslong, sizeof (xdrslong));
656 
657 	/* Choose a size for the long rpc response */
658 	if (MSG_IS_RPCSEC_GSS(msg)) {
659 		alloc_len = RNDUP(MAX_AUTH_BYTES + *msglen);
660 	} else {
661 		alloc_len = RNDUP(*msglen);
662 	}
663 
664 	if (alloc_len <= 64 * 1024) {
665 		if (alloc_len > 32 * 1024) {
666 			alloc_len = 64 * 1024;
667 		} else {
668 			if (alloc_len > 16 * 1024) {
669 				alloc_len = 32 * 1024;
670 			} else {
671 				alloc_len = 16 * 1024;
672 			}
673 		}
674 	}
675 
676 	long_rpc.type = RDMA_LONG_BUFFER;
677 	long_rpc.len = alloc_len;
678 	if (rdma_buf_alloc(crdp->conn, &long_rpc)) {
679 		return (SVC_RDMA_FAIL);
680 	}
681 
682 	memp = long_rpc.addr;
683 	xdrmem_create(&xdrslong, memp, alloc_len, XDR_ENCODE);
684 
685 	msg->rm_xid = clone_xprt->xp_xid;
686 
687 	if (!(xdr_replymsg(&xdrslong, msg) &&
688 	    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, &xdrslong,
689 	    xdr_results, xdr_location)))) {
690 		rdma_buf_free(crdp->conn, &long_rpc);
691 		DTRACE_PROBE(krpc__e__svcrdma__longrep__authwrap);
692 		return (SVC_RDMA_FAIL);
693 	}
694 
695 	*final_len = XDR_GETPOS(&xdrslong);
696 
697 	DTRACE_PROBE1(krpc__i__replylen, uint_t, *final_len);
698 	*numchunks = 0;
699 	*freelen = 0;
700 
701 	wcl = crdp->cl_reply;
702 	wcl->rb_longbuf = long_rpc;
703 
704 	count = *final_len;
705 	while ((wcl != NULL) && (count > 0)) {
706 
707 		if (wcl->c_dmemhandle.mrc_rmr == 0)
708 			break;
709 
710 		DTRACE_PROBE2(krpc__i__write__chunks, uint32_t, count,
711 		    uint32_t, wcl->c_len);
712 
713 		if (wcl->c_len > count) {
714 			wcl->c_len = count;
715 		}
716 		wcl->w.c_saddr3 = (caddr_t)memp;
717 
718 		count -= wcl->c_len;
719 		*numchunks +=  1;
720 		memp += wcl->c_len;
721 		wcl = wcl->c_next;
722 	}
723 
724 	/*
725 	 * Make rest of the chunks 0-len
726 	 */
727 	while (wcl != NULL) {
728 		if (wcl->c_dmemhandle.mrc_rmr == 0)
729 			break;
730 		wcl->c_len = 0;
731 		wcl = wcl->c_next;
732 	}
733 
734 	wcl = crdp->cl_reply;
735 
736 	/*
737 	 * MUST fail if there are still more data
738 	 */
739 	if (count > 0) {
740 		rdma_buf_free(crdp->conn, &long_rpc);
741 		DTRACE_PROBE(krpc__e__svcrdma__longrep__dlen__clist);
742 		return (SVC_RDMA_FAIL);
743 	}
744 
745 	if (clist_register(crdp->conn, wcl, CLIST_REG_SOURCE) != RDMA_SUCCESS) {
746 		rdma_buf_free(crdp->conn, &long_rpc);
747 		DTRACE_PROBE(krpc__e__svcrdma__longrep__clistreg);
748 		return (SVC_RDMA_FAIL);
749 	}
750 
751 	status = clist_syncmem(crdp->conn, wcl, CLIST_REG_SOURCE);
752 
753 	if (status) {
754 		(void) clist_deregister(crdp->conn, wcl);
755 		rdma_buf_free(crdp->conn, &long_rpc);
756 		DTRACE_PROBE(krpc__e__svcrdma__longrep__syncmem);
757 		return (SVC_RDMA_FAIL);
758 	}
759 
760 	status = RDMA_WRITE(crdp->conn, wcl, WAIT);
761 
762 	(void) clist_deregister(crdp->conn, wcl);
763 	rdma_buf_free(crdp->conn, &wcl->rb_longbuf);
764 
765 	if (status != RDMA_SUCCESS) {
766 		DTRACE_PROBE(krpc__e__svcrdma__longrep__write);
767 		return (SVC_RDMA_FAIL);
768 	}
769 
770 	return (SVC_RDMA_SUCCESS);
771 }
772 
773 
774 static int
775 svc_compose_rpcmsg(SVCXPRT * clone_xprt, CONN * conn, xdrproc_t xdr_results,
776     caddr_t xdr_location, rdma_buf_t *rpcreply, XDR ** xdrs,
777     struct rpc_msg *msg, bool_t has_args, uint_t *len)
778 {
779 	/*
780 	 * Get a pre-allocated buffer for rpc reply
781 	 */
782 	rpcreply->type = SEND_BUFFER;
783 	if (rdma_buf_alloc(conn, rpcreply)) {
784 		DTRACE_PROBE(krpc__e__svcrdma__rpcmsg__reply__nofreebufs);
785 		return (SVC_RDMA_FAIL);
786 	}
787 
788 	xdrrdma_create(*xdrs, rpcreply->addr, rpcreply->len,
789 	    0, NULL, XDR_ENCODE, conn);
790 
791 	msg->rm_xid = clone_xprt->xp_xid;
792 
793 	if (has_args) {
794 		if (!(xdr_replymsg(*xdrs, msg) &&
795 		    (!has_args ||
796 		    SVCAUTH_WRAP(&clone_xprt->xp_auth, *xdrs,
797 		    xdr_results, xdr_location)))) {
798 			rdma_buf_free(conn, rpcreply);
799 			DTRACE_PROBE(
800 			    krpc__e__svcrdma__rpcmsg__reply__authwrap1);
801 			return (SVC_RDMA_FAIL);
802 		}
803 	} else {
804 		if (!xdr_replymsg(*xdrs, msg)) {
805 			rdma_buf_free(conn, rpcreply);
806 			DTRACE_PROBE(
807 			    krpc__e__svcrdma__rpcmsg__reply__authwrap2);
808 			return (SVC_RDMA_FAIL);
809 		}
810 	}
811 
812 	*len = XDR_GETPOS(*xdrs);
813 
814 	return (SVC_RDMA_SUCCESS);
815 }
816 
817 /*
818  * Send rpc reply.
819  */
820 static bool_t
821 svc_rdma_ksend(SVCXPRT * clone_xprt, struct rpc_msg *msg)
822 {
823 	XDR *xdrs_rpc = &(clone_xprt->xp_xdrout);
824 	XDR xdrs_rhdr;
825 	CONN *conn = NULL;
826 	rdma_buf_t rbuf_resp = {0}, rbuf_rpc_resp = {0};
827 
828 	struct clone_rdma_data *crdp;
829 	struct clist *cl_read = NULL;
830 	struct clist *cl_send = NULL;
831 	struct clist *cl_write = NULL;
832 	xdrproc_t xdr_results;		/* results XDR encoding function */
833 	caddr_t xdr_location;		/* response results pointer */
834 
835 	int retval = FALSE;
836 	int status, msglen, num_wreply_segments = 0;
837 	uint32_t rdma_credit = 0;
838 	int freelen = 0;
839 	bool_t has_args;
840 	uint_t  final_resp_len, rdma_response_op, vers;
841 
842 	bzero(&xdrs_rhdr, sizeof (XDR));
843 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
844 	conn = crdp->conn;
845 
846 	/*
847 	 * If there is a result procedure specified in the reply message,
848 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
849 	 * We need to make sure it won't be processed twice, so we null
850 	 * it for xdr_replymsg here.
851 	 */
852 	has_args = FALSE;
853 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
854 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
855 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
856 			has_args = TRUE;
857 			xdr_location = msg->acpted_rply.ar_results.where;
858 			msg->acpted_rply.ar_results.proc = xdr_void;
859 			msg->acpted_rply.ar_results.where = NULL;
860 		}
861 	}
862 
863 	/*
864 	 * Given the limit on the inline response size (RPC_MSG_SZ),
865 	 * there is a need to make a guess as to the overall size of
866 	 * the response.  If the resultant size is beyond the inline
867 	 * size, then the server needs to use the "reply chunk list"
868 	 * provided by the client (if the client provided one).  An
869 	 * example of this type of response would be a READDIR
870 	 * response (e.g. a small directory read would fit in RPC_MSG_SZ
871 	 * and that is the preference but it may not fit)
872 	 *
873 	 * Combine the encoded size and the size of the true results
874 	 * and then make the decision about where to encode and send results.
875 	 *
876 	 * One important note, this calculation is ignoring the size
877 	 * of the encoding of the authentication overhead.  The reason
878 	 * for this is rooted in the complexities of access to the
879 	 * encoded size of RPCSEC_GSS related authentiation,
880 	 * integrity, and privacy.
881 	 *
882 	 * If it turns out that the encoded authentication bumps the
883 	 * response over the RPC_MSG_SZ limit, then it may need to
884 	 * attempt to encode for the reply chunk list.
885 	 */
886 
887 	/*
888 	 * Calculating the "sizeof" the RPC response header and the
889 	 * encoded results.
890 	 */
891 	msglen = xdr_sizeof(xdr_replymsg, msg);
892 
893 	if (msglen > 0) {
894 		RSSTAT_INCR(rstotalreplies);
895 	}
896 	if (has_args)
897 		msglen += xdrrdma_sizeof(xdr_results, xdr_location,
898 		    rdma_minchunk, NULL, NULL);
899 
900 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__msglen, int, msglen);
901 
902 	status = SVC_RDMA_SUCCESS;
903 
904 	if (msglen < RPC_MSG_SZ) {
905 		/*
906 		 * Looks like the response will fit in the inline
907 		 * response; let's try
908 		 */
909 		RSSTAT_INCR(rstotalinlinereplies);
910 
911 		rdma_response_op = RDMA_MSG;
912 
913 		status = svc_compose_rpcmsg(clone_xprt, conn, xdr_results,
914 		    xdr_location, &rbuf_rpc_resp, &xdrs_rpc, msg,
915 		    has_args, &final_resp_len);
916 
917 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_status,
918 		    int, status);
919 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_len,
920 		    int, final_resp_len);
921 
922 		if (status == SVC_RDMA_SUCCESS && crdp->cl_reply) {
923 			clist_free(crdp->cl_reply);
924 			crdp->cl_reply = NULL;
925 		}
926 	}
927 
928 	/*
929 	 * If the encode failed (size?) or the message really is
930 	 * larger than what is allowed, try the response chunk list.
931 	 */
932 	if (status != SVC_RDMA_SUCCESS || msglen >= RPC_MSG_SZ) {
933 		/*
934 		 * attempting to use a reply chunk list when there
935 		 * isn't one won't get very far...
936 		 */
937 		if (crdp->cl_reply == NULL) {
938 			DTRACE_PROBE(krpc__e__svcrdma__ksend__noreplycl);
939 			goto out;
940 		}
941 
942 		RSSTAT_INCR(rstotallongreplies);
943 
944 		msglen = xdr_sizeof(xdr_replymsg, msg);
945 		msglen += xdrrdma_sizeof(xdr_results, xdr_location, 0,
946 		    NULL, NULL);
947 
948 		status = svc_process_long_reply(clone_xprt, xdr_results,
949 		    xdr_location, msg, has_args, &msglen, &freelen,
950 		    &num_wreply_segments, &final_resp_len);
951 
952 		DTRACE_PROBE1(krpc__i__svcrdma__ksend__longreplen,
953 		    int, final_resp_len);
954 
955 		if (status != SVC_RDMA_SUCCESS) {
956 			DTRACE_PROBE(krpc__e__svcrdma__ksend__compose__failed);
957 			goto out;
958 		}
959 
960 		rdma_response_op = RDMA_NOMSG;
961 	}
962 
963 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__rdmamsg__len,
964 	    int, final_resp_len);
965 
966 	rbuf_resp.type = SEND_BUFFER;
967 	if (rdma_buf_alloc(conn, &rbuf_resp)) {
968 		rdma_buf_free(conn, &rbuf_rpc_resp);
969 		DTRACE_PROBE(krpc__e__svcrdma__ksend__nofreebufs);
970 		goto out;
971 	}
972 
973 	rdma_credit = rdma_bufs_granted;
974 
975 	vers = RPCRDMA_VERS;
976 	xdrmem_create(&xdrs_rhdr, rbuf_resp.addr, rbuf_resp.len, XDR_ENCODE);
977 	(*(uint32_t *)rbuf_resp.addr) = msg->rm_xid;
978 	/* Skip xid and set the xdr position accordingly. */
979 	XDR_SETPOS(&xdrs_rhdr, sizeof (uint32_t));
980 	if (!xdr_u_int(&xdrs_rhdr, &vers) ||
981 	    !xdr_u_int(&xdrs_rhdr, &rdma_credit) ||
982 	    !xdr_u_int(&xdrs_rhdr, &rdma_response_op)) {
983 		rdma_buf_free(conn, &rbuf_rpc_resp);
984 		rdma_buf_free(conn, &rbuf_resp);
985 		DTRACE_PROBE(krpc__e__svcrdma__ksend__uint);
986 		goto out;
987 	}
988 
989 	/*
990 	 * Now XDR the read chunk list, actually always NULL
991 	 */
992 	(void) xdr_encode_rlist_svc(&xdrs_rhdr, cl_read);
993 
994 	/*
995 	 * encode write list -- we already drove RDMA_WRITEs
996 	 */
997 	cl_write = crdp->cl_wlist;
998 	if (!xdr_encode_wlist(&xdrs_rhdr, cl_write)) {
999 		DTRACE_PROBE(krpc__e__svcrdma__ksend__enc__wlist);
1000 		rdma_buf_free(conn, &rbuf_rpc_resp);
1001 		rdma_buf_free(conn, &rbuf_resp);
1002 		goto out;
1003 	}
1004 
1005 	/*
1006 	 * XDR encode the RDMA_REPLY write chunk
1007 	 */
1008 	if (!xdr_encode_reply_wchunk(&xdrs_rhdr, crdp->cl_reply,
1009 	    num_wreply_segments)) {
1010 		rdma_buf_free(conn, &rbuf_rpc_resp);
1011 		rdma_buf_free(conn, &rbuf_resp);
1012 		goto out;
1013 	}
1014 
1015 	clist_add(&cl_send, 0, XDR_GETPOS(&xdrs_rhdr), &rbuf_resp.handle,
1016 	    rbuf_resp.addr, NULL, NULL);
1017 
1018 	if (rdma_response_op == RDMA_MSG) {
1019 		clist_add(&cl_send, 0, final_resp_len, &rbuf_rpc_resp.handle,
1020 		    rbuf_rpc_resp.addr, NULL, NULL);
1021 	}
1022 
1023 	status = RDMA_SEND(conn, cl_send, msg->rm_xid);
1024 
1025 	if (status == RDMA_SUCCESS) {
1026 		retval = TRUE;
1027 	}
1028 
1029 out:
1030 	/*
1031 	 * Free up sendlist chunks
1032 	 */
1033 	if (cl_send != NULL)
1034 		clist_free(cl_send);
1035 
1036 	/*
1037 	 * Destroy private data for xdr rdma
1038 	 */
1039 	if (clone_xprt->xp_xdrout.x_ops != NULL) {
1040 		XDR_DESTROY(&(clone_xprt->xp_xdrout));
1041 	}
1042 
1043 	if (crdp->cl_reply) {
1044 		clist_free(crdp->cl_reply);
1045 		crdp->cl_reply = NULL;
1046 	}
1047 
1048 	/*
1049 	 * This is completely disgusting.  If public is set it is
1050 	 * a pointer to a structure whose first field is the address
1051 	 * of the function to free that structure and any related
1052 	 * stuff.  (see rrokfree in nfs_xdr.c).
1053 	 */
1054 	if (xdrs_rpc->x_public) {
1055 		/* LINTED pointer alignment */
1056 		(**((int (**)()) xdrs_rpc->x_public)) (xdrs_rpc->x_public);
1057 	}
1058 
1059 	if (xdrs_rhdr.x_ops != NULL) {
1060 		XDR_DESTROY(&xdrs_rhdr);
1061 	}
1062 
1063 	return (retval);
1064 }
1065 
1066 /*
1067  * Deserialize arguments.
1068  */
1069 static bool_t
1070 svc_rdma_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, caddr_t args_ptr)
1071 {
1072 	if ((SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
1073 	    xdr_args, args_ptr)) != TRUE)
1074 		return (FALSE);
1075 	return (TRUE);
1076 }
1077 
1078 static bool_t
1079 svc_rdma_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
1080     caddr_t args_ptr)
1081 {
1082 	struct clone_rdma_data *crdp;
1083 	bool_t retval;
1084 
1085 	/*
1086 	 * If the cloned bit is true, then this transport specific
1087 	 * rmda data has been duplicated into another cloned xprt. Do
1088 	 * not free, or release the connection, it is still in use.  The
1089 	 * buffers will be freed and the connection released later by
1090 	 * SVC_CLONE_DESTROY().
1091 	 */
1092 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
1093 	if (crdp->cloned == TRUE) {
1094 		crdp->cloned = 0;
1095 		return (TRUE);
1096 	}
1097 
1098 	/*
1099 	 * Free the args if needed then XDR_DESTROY
1100 	 */
1101 	if (args_ptr) {
1102 		XDR	*xdrs = &clone_xprt->xp_xdrin;
1103 
1104 		xdrs->x_op = XDR_FREE;
1105 		retval = (*xdr_args)(xdrs, args_ptr);
1106 	}
1107 
1108 	XDR_DESTROY(&(clone_xprt->xp_xdrin));
1109 	rdma_buf_free(crdp->conn, &crdp->rpcbuf);
1110 	if (crdp->cl_reply) {
1111 		clist_free(crdp->cl_reply);
1112 		crdp->cl_reply = NULL;
1113 	}
1114 	RDMA_REL_CONN(crdp->conn);
1115 
1116 	return (retval);
1117 }
1118 
1119 /* ARGSUSED */
1120 static int32_t *
1121 svc_rdma_kgetres(SVCXPRT *clone_xprt, int size)
1122 {
1123 	return (NULL);
1124 }
1125 
1126 /* ARGSUSED */
1127 static void
1128 svc_rdma_kfreeres(SVCXPRT *clone_xprt)
1129 {
1130 }
1131 
1132 /*
1133  * the dup cacheing routines below provide a cache of non-failure
1134  * transaction id's.  rpc service routines can use this to detect
1135  * retransmissions and re-send a non-failure response.
1136  */
1137 
1138 /*
1139  * MAXDUPREQS is the number of cached items.  It should be adjusted
1140  * to the service load so that there is likely to be a response entry
1141  * when the first retransmission comes in.
1142  */
1143 #define	MAXDUPREQS	8192
1144 
1145 /*
1146  * This should be appropriately scaled to MAXDUPREQS.  To produce as less as
1147  * possible collisions it is suggested to set this to a prime.
1148  */
1149 #define	DRHASHSZ	2053
1150 
1151 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
1152 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
1153 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
1154 
1155 static int	rdmandupreqs = 0;
1156 int	rdmamaxdupreqs = MAXDUPREQS;
1157 static kmutex_t rdmadupreq_lock;
1158 static struct dupreq *rdmadrhashtbl[DRHASHSZ];
1159 static int	rdmadrhashstat[DRHASHSZ];
1160 
1161 static void unhash(struct dupreq *);
1162 
1163 /*
1164  * rdmadrmru points to the head of a circular linked list in lru order.
1165  * rdmadrmru->dr_next == drlru
1166  */
1167 struct dupreq *rdmadrmru;
1168 
1169 /*
1170  * svc_rdma_kdup searches the request cache and returns 0 if the
1171  * request is not found in the cache.  If it is found, then it
1172  * returns the state of the request (in progress or done) and
1173  * the status or attributes that were part of the original reply.
1174  */
1175 static int
1176 svc_rdma_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
1177     bool_t *dupcachedp)
1178 {
1179 	struct dupreq *dr;
1180 	uint32_t xid;
1181 	uint32_t drhash;
1182 	int status;
1183 
1184 	xid = REQTOXID(req);
1185 	mutex_enter(&rdmadupreq_lock);
1186 	RSSTAT_INCR(rsdupchecks);
1187 	/*
1188 	 * Check to see whether an entry already exists in the cache.
1189 	 */
1190 	dr = rdmadrhashtbl[XIDHASH(xid)];
1191 	while (dr != NULL) {
1192 		if (dr->dr_xid == xid &&
1193 		    dr->dr_proc == req->rq_proc &&
1194 		    dr->dr_prog == req->rq_prog &&
1195 		    dr->dr_vers == req->rq_vers &&
1196 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
1197 		    bcmp((caddr_t)dr->dr_addr.buf,
1198 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
1199 		    dr->dr_addr.len) == 0) {
1200 			status = dr->dr_status;
1201 			if (status == DUP_DONE) {
1202 				bcopy(dr->dr_resp.buf, res, size);
1203 				if (dupcachedp != NULL)
1204 					*dupcachedp = (dr->dr_resfree != NULL);
1205 			} else {
1206 				dr->dr_status = DUP_INPROGRESS;
1207 				*drpp = dr;
1208 			}
1209 			RSSTAT_INCR(rsdupreqs);
1210 			mutex_exit(&rdmadupreq_lock);
1211 			return (status);
1212 		}
1213 		dr = dr->dr_chain;
1214 	}
1215 
1216 	/*
1217 	 * There wasn't an entry, either allocate a new one or recycle
1218 	 * an old one.
1219 	 */
1220 	if (rdmandupreqs < rdmamaxdupreqs) {
1221 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
1222 		if (dr == NULL) {
1223 			mutex_exit(&rdmadupreq_lock);
1224 			return (DUP_ERROR);
1225 		}
1226 		dr->dr_resp.buf = NULL;
1227 		dr->dr_resp.maxlen = 0;
1228 		dr->dr_addr.buf = NULL;
1229 		dr->dr_addr.maxlen = 0;
1230 		if (rdmadrmru) {
1231 			dr->dr_next = rdmadrmru->dr_next;
1232 			rdmadrmru->dr_next = dr;
1233 		} else {
1234 			dr->dr_next = dr;
1235 		}
1236 		rdmandupreqs++;
1237 	} else {
1238 		dr = rdmadrmru->dr_next;
1239 		while (dr->dr_status == DUP_INPROGRESS) {
1240 			dr = dr->dr_next;
1241 			if (dr == rdmadrmru->dr_next) {
1242 				mutex_exit(&rdmadupreq_lock);
1243 				return (DUP_ERROR);
1244 			}
1245 		}
1246 		unhash(dr);
1247 		if (dr->dr_resfree) {
1248 			(*dr->dr_resfree)(dr->dr_resp.buf);
1249 		}
1250 	}
1251 	dr->dr_resfree = NULL;
1252 	rdmadrmru = dr;
1253 
1254 	dr->dr_xid = REQTOXID(req);
1255 	dr->dr_prog = req->rq_prog;
1256 	dr->dr_vers = req->rq_vers;
1257 	dr->dr_proc = req->rq_proc;
1258 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
1259 		if (dr->dr_addr.buf != NULL)
1260 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
1261 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
1262 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
1263 		if (dr->dr_addr.buf == NULL) {
1264 			dr->dr_addr.maxlen = 0;
1265 			dr->dr_status = DUP_DROP;
1266 			mutex_exit(&rdmadupreq_lock);
1267 			return (DUP_ERROR);
1268 		}
1269 	}
1270 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
1271 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
1272 	if (dr->dr_resp.maxlen < size) {
1273 		if (dr->dr_resp.buf != NULL)
1274 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
1275 		dr->dr_resp.maxlen = (unsigned int)size;
1276 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
1277 		if (dr->dr_resp.buf == NULL) {
1278 			dr->dr_resp.maxlen = 0;
1279 			dr->dr_status = DUP_DROP;
1280 			mutex_exit(&rdmadupreq_lock);
1281 			return (DUP_ERROR);
1282 		}
1283 	}
1284 	dr->dr_status = DUP_INPROGRESS;
1285 
1286 	drhash = (uint32_t)DRHASH(dr);
1287 	dr->dr_chain = rdmadrhashtbl[drhash];
1288 	rdmadrhashtbl[drhash] = dr;
1289 	rdmadrhashstat[drhash]++;
1290 	mutex_exit(&rdmadupreq_lock);
1291 	*drpp = dr;
1292 	return (DUP_NEW);
1293 }
1294 
1295 /*
1296  * svc_rdma_kdupdone marks the request done (DUP_DONE or DUP_DROP)
1297  * and stores the response.
1298  */
1299 static void
1300 svc_rdma_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
1301     int size, int status)
1302 {
1303 	ASSERT(dr->dr_resfree == NULL);
1304 	if (status == DUP_DONE) {
1305 		bcopy(res, dr->dr_resp.buf, size);
1306 		dr->dr_resfree = dis_resfree;
1307 	}
1308 	dr->dr_status = status;
1309 }
1310 
1311 /*
1312  * This routine expects that the mutex, rdmadupreq_lock, is already held.
1313  */
1314 static void
1315 unhash(struct dupreq *dr)
1316 {
1317 	struct dupreq *drt;
1318 	struct dupreq *drtprev = NULL;
1319 	uint32_t drhash;
1320 
1321 	ASSERT(MUTEX_HELD(&rdmadupreq_lock));
1322 
1323 	drhash = (uint32_t)DRHASH(dr);
1324 	drt = rdmadrhashtbl[drhash];
1325 	while (drt != NULL) {
1326 		if (drt == dr) {
1327 			rdmadrhashstat[drhash]--;
1328 			if (drtprev == NULL) {
1329 				rdmadrhashtbl[drhash] = drt->dr_chain;
1330 			} else {
1331 				drtprev->dr_chain = drt->dr_chain;
1332 			}
1333 			return;
1334 		}
1335 		drtprev = drt;
1336 		drt = drt->dr_chain;
1337 	}
1338 }
1339 
1340 bool_t
1341 rdma_get_wchunk(struct svc_req *req, iovec_t *iov, struct clist *wlist)
1342 {
1343 	struct clist	*clist;
1344 	uint32_t	tlen;
1345 
1346 	if (req->rq_xprt->xp_type != T_RDMA) {
1347 		return (FALSE);
1348 	}
1349 
1350 	tlen = 0;
1351 	clist = wlist;
1352 	while (clist) {
1353 		tlen += clist->c_len;
1354 		clist = clist->c_next;
1355 	}
1356 
1357 	/*
1358 	 * set iov to addr+len of first segment of first wchunk of
1359 	 * wlist sent by client.  krecv() already malloc'd a buffer
1360 	 * large enough, but registration is deferred until we write
1361 	 * the buffer back to (NFS) client using RDMA_WRITE.
1362 	 */
1363 	iov->iov_base = (caddr_t)(uintptr_t)wlist->w.c_saddr;
1364 	iov->iov_len = tlen;
1365 
1366 	return (TRUE);
1367 }
1368 
1369 /*
1370  * routine to setup the read chunk lists
1371  */
1372 
1373 int
1374 rdma_setup_read_chunks(struct clist *wcl, uint32_t count, int *wcl_len)
1375 {
1376 	int		data_len, avail_len;
1377 	uint_t		round_len;
1378 
1379 	data_len = avail_len = 0;
1380 
1381 	while (wcl != NULL && count > 0) {
1382 		if (wcl->c_dmemhandle.mrc_rmr == 0)
1383 			break;
1384 
1385 		if (wcl->c_len < count) {
1386 			data_len += wcl->c_len;
1387 			avail_len = 0;
1388 		} else {
1389 			data_len += count;
1390 			avail_len = wcl->c_len - count;
1391 			wcl->c_len = count;
1392 		}
1393 		count -= wcl->c_len;
1394 
1395 		if (count == 0)
1396 			break;
1397 
1398 		wcl = wcl->c_next;
1399 	}
1400 
1401 	/*
1402 	 * MUST fail if there are still more data
1403 	 */
1404 	if (count > 0) {
1405 		DTRACE_PROBE2(krpc__e__rdma_setup_read_chunks_clist_len,
1406 		    int, data_len, int, count);
1407 		return (FALSE);
1408 	}
1409 
1410 	/*
1411 	 * Round up the last chunk to 4-byte boundary
1412 	 */
1413 	*wcl_len = roundup(data_len, BYTES_PER_XDR_UNIT);
1414 	round_len = *wcl_len - data_len;
1415 
1416 	if (round_len) {
1417 
1418 		/*
1419 		 * If there is space in the current chunk,
1420 		 * add the roundup to the chunk.
1421 		 */
1422 		if (avail_len >= round_len) {
1423 			wcl->c_len += round_len;
1424 		} else  {
1425 			/*
1426 			 * try the next one.
1427 			 */
1428 			wcl = wcl->c_next;
1429 			if ((wcl == NULL) || (wcl->c_len < round_len)) {
1430 				DTRACE_PROBE1(
1431 				    krpc__e__rdma_setup_read_chunks_rndup,
1432 				    int, round_len);
1433 				return (FALSE);
1434 			}
1435 			wcl->c_len = round_len;
1436 		}
1437 	}
1438 
1439 	wcl = wcl->c_next;
1440 
1441 	/*
1442 	 * Make rest of the chunks 0-len
1443 	 */
1444 
1445 	clist_zero_len(wcl);
1446 
1447 	return (TRUE);
1448 }
1449