xref: /illumos-gate/usr/src/uts/common/rpc/svc_rdma.c (revision b1d7ec75953cd517f5b7c3d9cb427ff8ec5d7d07)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
26 /* All Rights Reserved */
27 /*
28  * Portions of this source code were derived from Berkeley
29  * 4.3 BSD under license from the Regents of the University of
30  * California.
31  */
32 
33 /*
34  * Server side of RPC over RDMA in the kernel.
35  */
36 
37 #include <sys/param.h>
38 #include <sys/types.h>
39 #include <sys/user.h>
40 #include <sys/sysmacros.h>
41 #include <sys/proc.h>
42 #include <sys/file.h>
43 #include <sys/errno.h>
44 #include <sys/kmem.h>
45 #include <sys/debug.h>
46 #include <sys/systm.h>
47 #include <sys/cmn_err.h>
48 #include <sys/kstat.h>
49 #include <sys/vtrace.h>
50 #include <sys/debug.h>
51 
52 #include <rpc/types.h>
53 #include <rpc/xdr.h>
54 #include <rpc/auth.h>
55 #include <rpc/clnt.h>
56 #include <rpc/rpc_msg.h>
57 #include <rpc/svc.h>
58 #include <rpc/rpc_rdma.h>
59 #include <sys/ddi.h>
60 #include <sys/sunddi.h>
61 
62 #include <inet/common.h>
63 #include <inet/ip.h>
64 #include <inet/ip6.h>
65 
66 #include <nfs/nfs.h>
67 #include <sys/sdt.h>
68 
69 #define	SVC_RDMA_SUCCESS 0
70 #define	SVC_RDMA_FAIL -1
71 
72 #define	SVC_CREDIT_FACTOR (0.5)
73 
74 #define	MSG_IS_RPCSEC_GSS(msg)		\
75 	((msg)->rm_reply.rp_acpt.ar_verf.oa_flavor == RPCSEC_GSS)
76 
77 
78 uint32_t rdma_bufs_granted = RDMA_BUFS_GRANT;
79 
80 /*
81  * RDMA transport specific data associated with SVCMASTERXPRT
82  */
83 struct rdma_data {
84 	SVCMASTERXPRT 	*rd_xprt;	/* back ptr to SVCMASTERXPRT */
85 	struct rdma_svc_data rd_data;	/* rdma data */
86 	rdma_mod_t	*r_mod;		/* RDMA module containing ops ptr */
87 };
88 
89 /*
90  * Plugin connection specific data stashed away in clone SVCXPRT
91  */
92 struct clone_rdma_data {
93 	bool_t		cloned;		/* xprt cloned for thread processing */
94 	CONN		*conn;		/* RDMA connection */
95 	rdma_buf_t	rpcbuf;		/* RPC req/resp buffer */
96 	struct clist	*cl_reply;	/* reply chunk buffer info */
97 	struct clist	*cl_wlist;		/* write list clist */
98 };
99 
100 
101 #define	MAXADDRLEN	128	/* max length for address mask */
102 
103 /*
104  * Routines exported through ops vector.
105  */
106 static bool_t		svc_rdma_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
107 static bool_t		svc_rdma_ksend(SVCXPRT *, struct rpc_msg *);
108 static bool_t		svc_rdma_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
109 static bool_t		svc_rdma_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
110 void			svc_rdma_kdestroy(SVCMASTERXPRT *);
111 static int		svc_rdma_kdup(struct svc_req *, caddr_t, int,
112 				struct dupreq **, bool_t *);
113 static void		svc_rdma_kdupdone(struct dupreq *, caddr_t,
114 				void (*)(), int, int);
115 static int32_t		*svc_rdma_kgetres(SVCXPRT *, int);
116 static void		svc_rdma_kfreeres(SVCXPRT *);
117 static void		svc_rdma_kclone_destroy(SVCXPRT *);
118 static void		svc_rdma_kstart(SVCMASTERXPRT *);
119 void			svc_rdma_kstop(SVCMASTERXPRT *);
120 static void		svc_rdma_kclone_xprt(SVCXPRT *, SVCXPRT *);
121 
122 static int	svc_process_long_reply(SVCXPRT *, xdrproc_t,
123 			caddr_t, struct rpc_msg *, bool_t, int *,
124 			int *, int *, unsigned int *);
125 
126 static int	svc_compose_rpcmsg(SVCXPRT *, CONN *, xdrproc_t,
127 			caddr_t, rdma_buf_t *, XDR **, struct rpc_msg *,
128 			bool_t, uint_t *);
129 static bool_t rpcmsg_length(xdrproc_t,
130 		caddr_t,
131 		struct rpc_msg *, bool_t, int);
132 
133 /*
134  * Server transport operations vector.
135  */
136 struct svc_ops rdma_svc_ops = {
137 	svc_rdma_krecv,		/* Get requests */
138 	svc_rdma_kgetargs,	/* Deserialize arguments */
139 	svc_rdma_ksend,		/* Send reply */
140 	svc_rdma_kfreeargs,	/* Free argument data space */
141 	svc_rdma_kdestroy,	/* Destroy transport handle */
142 	svc_rdma_kdup,		/* Check entry in dup req cache */
143 	svc_rdma_kdupdone,	/* Mark entry in dup req cache as done */
144 	svc_rdma_kgetres,	/* Get pointer to response buffer */
145 	svc_rdma_kfreeres,	/* Destroy pre-serialized response header */
146 	svc_rdma_kclone_destroy,	/* Destroy a clone xprt */
147 	svc_rdma_kstart,	/* Tell `ready-to-receive' to rpcmod */
148 	svc_rdma_kclone_xprt	/* Transport specific clone xprt */
149 };
150 
151 /*
152  * Server statistics
153  * NOTE: This structure type is duplicated in the NFS fast path.
154  */
155 struct {
156 	kstat_named_t	rscalls;
157 	kstat_named_t	rsbadcalls;
158 	kstat_named_t	rsnullrecv;
159 	kstat_named_t	rsbadlen;
160 	kstat_named_t	rsxdrcall;
161 	kstat_named_t	rsdupchecks;
162 	kstat_named_t	rsdupreqs;
163 	kstat_named_t	rslongrpcs;
164 	kstat_named_t	rstotalreplies;
165 	kstat_named_t	rstotallongreplies;
166 	kstat_named_t	rstotalinlinereplies;
167 } rdmarsstat = {
168 	{ "calls",	KSTAT_DATA_UINT64 },
169 	{ "badcalls",	KSTAT_DATA_UINT64 },
170 	{ "nullrecv",	KSTAT_DATA_UINT64 },
171 	{ "badlen",	KSTAT_DATA_UINT64 },
172 	{ "xdrcall",	KSTAT_DATA_UINT64 },
173 	{ "dupchecks",	KSTAT_DATA_UINT64 },
174 	{ "dupreqs",	KSTAT_DATA_UINT64 },
175 	{ "longrpcs",	KSTAT_DATA_UINT64 },
176 	{ "totalreplies",	KSTAT_DATA_UINT64 },
177 	{ "totallongreplies",	KSTAT_DATA_UINT64 },
178 	{ "totalinlinereplies",	KSTAT_DATA_UINT64 },
179 };
180 
181 kstat_named_t *rdmarsstat_ptr = (kstat_named_t *)&rdmarsstat;
182 uint_t rdmarsstat_ndata = sizeof (rdmarsstat) / sizeof (kstat_named_t);
183 
184 #define	RSSTAT_INCR(x)	atomic_add_64(&rdmarsstat.x.value.ui64, 1)
185 /*
186  * Create a transport record.
187  * The transport record, output buffer, and private data structure
188  * are allocated.  The output buffer is serialized into using xdrmem.
189  * There is one transport record per user process which implements a
190  * set of services.
191  */
192 /* ARGSUSED */
193 int
194 svc_rdma_kcreate(char *netid, SVC_CALLOUT_TABLE *sct, int id,
195     rdma_xprt_group_t *started_xprts)
196 {
197 	int error;
198 	SVCMASTERXPRT *xprt;
199 	struct rdma_data *rd;
200 	rdma_registry_t *rmod;
201 	rdma_xprt_record_t *xprt_rec;
202 	queue_t	*q;
203 	/*
204 	 * modload the RDMA plugins is not already done.
205 	 */
206 	if (!rdma_modloaded) {
207 		/*CONSTANTCONDITION*/
208 		ASSERT(sizeof (struct clone_rdma_data) <= SVC_P2LEN);
209 
210 		mutex_enter(&rdma_modload_lock);
211 		if (!rdma_modloaded) {
212 			error = rdma_modload();
213 		}
214 		mutex_exit(&rdma_modload_lock);
215 
216 		if (error)
217 			return (error);
218 	}
219 
220 	/*
221 	 * master_xprt_count is the count of master transport handles
222 	 * that were successfully created and are ready to recieve for
223 	 * RDMA based access.
224 	 */
225 	error = 0;
226 	xprt_rec = NULL;
227 	rw_enter(&rdma_lock, RW_READER);
228 	if (rdma_mod_head == NULL) {
229 		started_xprts->rtg_count = 0;
230 		rw_exit(&rdma_lock);
231 		if (rdma_dev_available)
232 			return (EPROTONOSUPPORT);
233 		else
234 			return (ENODEV);
235 	}
236 
237 	/*
238 	 * If we have reached here, then atleast one RDMA plugin has loaded.
239 	 * Create a master_xprt, make it start listenining on the device,
240 	 * if an error is generated, record it, we might need to shut
241 	 * the master_xprt.
242 	 * SVC_START() calls svc_rdma_kstart which calls plugin binding
243 	 * routines.
244 	 */
245 	for (rmod = rdma_mod_head; rmod != NULL; rmod = rmod->r_next) {
246 
247 		/*
248 		 * One SVCMASTERXPRT per RDMA plugin.
249 		 */
250 		xprt = kmem_zalloc(sizeof (*xprt), KM_SLEEP);
251 		xprt->xp_ops = &rdma_svc_ops;
252 		xprt->xp_sct = sct;
253 		xprt->xp_type = T_RDMA;
254 		mutex_init(&xprt->xp_req_lock, NULL, MUTEX_DEFAULT, NULL);
255 		mutex_init(&xprt->xp_thread_lock, NULL, MUTEX_DEFAULT, NULL);
256 		xprt->xp_req_head = (mblk_t *)0;
257 		xprt->xp_req_tail = (mblk_t *)0;
258 		xprt->xp_threads = 0;
259 		xprt->xp_detached_threads = 0;
260 
261 		rd = kmem_zalloc(sizeof (*rd), KM_SLEEP);
262 		xprt->xp_p2 = (caddr_t)rd;
263 		rd->rd_xprt = xprt;
264 		rd->r_mod = rmod->r_mod;
265 
266 		q = &rd->rd_data.q;
267 		xprt->xp_wq = q;
268 		q->q_ptr = &rd->rd_xprt;
269 		xprt->xp_netid = NULL;
270 
271 		xprt->xp_addrmask.maxlen =
272 		    xprt->xp_addrmask.len = sizeof (struct sockaddr_in);
273 		xprt->xp_addrmask.buf =
274 		    kmem_zalloc(xprt->xp_addrmask.len, KM_SLEEP);
275 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_addr.s_addr =
276 		    (uint32_t)~0;
277 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_family =
278 		    (ushort_t)~0;
279 
280 		/*
281 		 * Each of the plugins will have their own Service ID
282 		 * to listener specific mapping, like port number for VI
283 		 * and service name for IB.
284 		 */
285 		rd->rd_data.svcid = id;
286 		error = svc_xprt_register(xprt, id);
287 		if (error) {
288 			DTRACE_PROBE(krpc__e__svcrdma__xprt__reg);
289 			goto cleanup;
290 		}
291 
292 		SVC_START(xprt);
293 		if (!rd->rd_data.active) {
294 			svc_xprt_unregister(xprt);
295 			error = rd->rd_data.err_code;
296 			goto cleanup;
297 		}
298 
299 		/*
300 		 * This is set only when there is atleast one or more
301 		 * transports successfully created. We insert the pointer
302 		 * to the created RDMA master xprt into a separately maintained
303 		 * list. This way we can easily reference it later to cleanup,
304 		 * when NFS kRPC service pool is going away/unregistered.
305 		 */
306 		started_xprts->rtg_count ++;
307 		xprt_rec = kmem_alloc(sizeof (*xprt_rec), KM_SLEEP);
308 		xprt_rec->rtr_xprt_ptr = xprt;
309 		xprt_rec->rtr_next = started_xprts->rtg_listhead;
310 		started_xprts->rtg_listhead = xprt_rec;
311 		continue;
312 cleanup:
313 		SVC_DESTROY(xprt);
314 		if (error == RDMA_FAILED)
315 			error = EPROTONOSUPPORT;
316 	}
317 
318 	rw_exit(&rdma_lock);
319 
320 	/*
321 	 * Don't return any error even if a single plugin was started
322 	 * successfully.
323 	 */
324 	if (started_xprts->rtg_count == 0)
325 		return (error);
326 	return (0);
327 }
328 
329 /*
330  * Cleanup routine for freeing up memory allocated by
331  * svc_rdma_kcreate()
332  */
333 void
334 svc_rdma_kdestroy(SVCMASTERXPRT *xprt)
335 {
336 	struct rdma_data *rd = (struct rdma_data *)xprt->xp_p2;
337 
338 
339 	mutex_destroy(&xprt->xp_req_lock);
340 	mutex_destroy(&xprt->xp_thread_lock);
341 	kmem_free(rd, sizeof (*rd));
342 	kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
343 	kmem_free(xprt, sizeof (*xprt));
344 }
345 
346 
347 static void
348 svc_rdma_kstart(SVCMASTERXPRT *xprt)
349 {
350 	struct rdma_svc_data *svcdata;
351 	rdma_mod_t *rmod;
352 
353 	svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data;
354 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
355 
356 	/*
357 	 * Create a listener for  module at this port
358 	 */
359 
360 	if (rmod->rdma_count != 0)
361 		(*rmod->rdma_ops->rdma_svc_listen)(svcdata);
362 	else
363 		svcdata->err_code = RDMA_FAILED;
364 }
365 
366 void
367 svc_rdma_kstop(SVCMASTERXPRT *xprt)
368 {
369 	struct rdma_svc_data *svcdata;
370 	rdma_mod_t *rmod;
371 
372 	svcdata	= &((struct rdma_data *)xprt->xp_p2)->rd_data;
373 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
374 
375 	/*
376 	 * Call the stop listener routine for each plugin. If rdma_count is
377 	 * already zero set active to zero.
378 	 */
379 	if (rmod->rdma_count != 0)
380 		(*rmod->rdma_ops->rdma_svc_stop)(svcdata);
381 	else
382 		svcdata->active = 0;
383 	if (svcdata->active)
384 		DTRACE_PROBE(krpc__e__svcrdma__kstop);
385 }
386 
387 /* ARGSUSED */
388 static void
389 svc_rdma_kclone_destroy(SVCXPRT *clone_xprt)
390 {
391 
392 	struct clone_rdma_data *cdrp;
393 	cdrp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
394 
395 	/*
396 	 * Only free buffers and release connection when cloned is set.
397 	 */
398 	if (cdrp->cloned != TRUE)
399 		return;
400 
401 	rdma_buf_free(cdrp->conn, &cdrp->rpcbuf);
402 	if (cdrp->cl_reply) {
403 		clist_free(cdrp->cl_reply);
404 		cdrp->cl_reply = NULL;
405 	}
406 	RDMA_REL_CONN(cdrp->conn);
407 
408 	cdrp->cloned = 0;
409 }
410 
411 /*
412  * Clone the xprt specific information.  It will be freed by
413  * SVC_CLONE_DESTROY.
414  */
415 static void
416 svc_rdma_kclone_xprt(SVCXPRT *src_xprt, SVCXPRT *dst_xprt)
417 {
418 	struct clone_rdma_data *srcp2;
419 	struct clone_rdma_data *dstp2;
420 
421 	srcp2 = (struct clone_rdma_data *)src_xprt->xp_p2buf;
422 	dstp2 = (struct clone_rdma_data *)dst_xprt->xp_p2buf;
423 
424 	if (srcp2->conn != NULL) {
425 		srcp2->cloned = TRUE;
426 		*dstp2 = *srcp2;
427 	}
428 }
429 
430 
431 static bool_t
432 svc_rdma_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
433 {
434 	XDR	*xdrs;
435 	CONN	*conn;
436 	rdma_recv_data_t	*rdp = (rdma_recv_data_t *)mp->b_rptr;
437 	struct clone_rdma_data *crdp;
438 	struct clist	*cl = NULL;
439 	struct clist	*wcl = NULL;
440 	struct clist	*cllong = NULL;
441 
442 	rdma_stat	status;
443 	uint32_t vers, op, pos, xid;
444 	uint32_t rdma_credit;
445 	uint32_t wcl_total_length = 0;
446 	bool_t	wwl = FALSE;
447 
448 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
449 	RSSTAT_INCR(rscalls);
450 	conn = rdp->conn;
451 
452 	status = rdma_svc_postrecv(conn);
453 	if (status != RDMA_SUCCESS) {
454 		DTRACE_PROBE(krpc__e__svcrdma__krecv__postrecv);
455 		goto badrpc_call;
456 	}
457 
458 	xdrs = &clone_xprt->xp_xdrin;
459 	xdrmem_create(xdrs, rdp->rpcmsg.addr, rdp->rpcmsg.len, XDR_DECODE);
460 	xid = *(uint32_t *)rdp->rpcmsg.addr;
461 	XDR_SETPOS(xdrs, sizeof (uint32_t));
462 
463 	if (! xdr_u_int(xdrs, &vers) ||
464 	    ! xdr_u_int(xdrs, &rdma_credit) ||
465 	    ! xdr_u_int(xdrs, &op)) {
466 		DTRACE_PROBE(krpc__e__svcrdma__krecv__uint);
467 		goto xdr_err;
468 	}
469 
470 	/* Checking if the status of the recv operation was normal */
471 	if (rdp->status != 0) {
472 		DTRACE_PROBE1(krpc__e__svcrdma__krecv__invalid__status,
473 		    int, rdp->status);
474 		goto badrpc_call;
475 	}
476 
477 	if (! xdr_do_clist(xdrs, &cl)) {
478 		DTRACE_PROBE(krpc__e__svcrdma__krecv__do__clist);
479 		goto xdr_err;
480 	}
481 
482 	if (!xdr_decode_wlist_svc(xdrs, &wcl, &wwl, &wcl_total_length, conn)) {
483 		DTRACE_PROBE(krpc__e__svcrdma__krecv__decode__wlist);
484 		if (cl)
485 			clist_free(cl);
486 		goto xdr_err;
487 	}
488 	crdp->cl_wlist = wcl;
489 
490 	crdp->cl_reply = NULL;
491 	(void) xdr_decode_reply_wchunk(xdrs, &crdp->cl_reply);
492 
493 	/*
494 	 * A chunk at 0 offset indicates that the RPC call message
495 	 * is in a chunk. Get the RPC call message chunk.
496 	 */
497 	if (cl != NULL && op == RDMA_NOMSG) {
498 
499 		/* Remove RPC call message chunk from chunklist */
500 		cllong = cl;
501 		cl = cl->c_next;
502 		cllong->c_next = NULL;
503 
504 
505 		/* Allocate and register memory for the RPC call msg chunk */
506 		cllong->rb_longbuf.type = RDMA_LONG_BUFFER;
507 		cllong->rb_longbuf.len = cllong->c_len > LONG_REPLY_LEN ?
508 		    cllong->c_len : LONG_REPLY_LEN;
509 
510 		if (rdma_buf_alloc(conn, &cllong->rb_longbuf)) {
511 			clist_free(cllong);
512 			goto cll_malloc_err;
513 		}
514 
515 		cllong->u.c_daddr3 = cllong->rb_longbuf.addr;
516 
517 		if (cllong->u.c_daddr == NULL) {
518 			DTRACE_PROBE(krpc__e__svcrdma__krecv__nomem);
519 			rdma_buf_free(conn, &cllong->rb_longbuf);
520 			clist_free(cllong);
521 			goto cll_malloc_err;
522 		}
523 
524 		status = clist_register(conn, cllong, CLIST_REG_DST);
525 		if (status) {
526 			DTRACE_PROBE(krpc__e__svcrdma__krecv__clist__reg);
527 			rdma_buf_free(conn, &cllong->rb_longbuf);
528 			clist_free(cllong);
529 			goto cll_malloc_err;
530 		}
531 
532 		/*
533 		 * Now read the RPC call message in
534 		 */
535 		status = RDMA_READ(conn, cllong, WAIT);
536 		if (status) {
537 			DTRACE_PROBE(krpc__e__svcrdma__krecv__read);
538 			(void) clist_deregister(conn, cllong);
539 			rdma_buf_free(conn, &cllong->rb_longbuf);
540 			clist_free(cllong);
541 			goto cll_malloc_err;
542 		}
543 
544 		status = clist_syncmem(conn, cllong, CLIST_REG_DST);
545 		(void) clist_deregister(conn, cllong);
546 
547 		xdrrdma_create(xdrs, (caddr_t)(uintptr_t)cllong->u.c_daddr3,
548 		    cllong->c_len, 0, cl, XDR_DECODE, conn);
549 
550 		crdp->rpcbuf = cllong->rb_longbuf;
551 		crdp->rpcbuf.len = cllong->c_len;
552 		clist_free(cllong);
553 		RDMA_BUF_FREE(conn, &rdp->rpcmsg);
554 	} else {
555 		pos = XDR_GETPOS(xdrs);
556 		xdrrdma_create(xdrs, rdp->rpcmsg.addr + pos,
557 		    rdp->rpcmsg.len - pos, 0, cl, XDR_DECODE, conn);
558 		crdp->rpcbuf = rdp->rpcmsg;
559 
560 		/* Use xdrrdmablk_ops to indicate there is a read chunk list */
561 		if (cl != NULL) {
562 			int32_t flg = XDR_RDMA_RLIST_REG;
563 
564 			XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
565 			xdrs->x_ops = &xdrrdmablk_ops;
566 		}
567 	}
568 
569 	if (crdp->cl_wlist) {
570 		int32_t flg = XDR_RDMA_WLIST_REG;
571 
572 		XDR_CONTROL(xdrs, XDR_RDMA_SET_WLIST, crdp->cl_wlist);
573 		XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
574 	}
575 
576 	if (! xdr_callmsg(xdrs, msg)) {
577 		DTRACE_PROBE(krpc__e__svcrdma__krecv__callmsg);
578 		RSSTAT_INCR(rsxdrcall);
579 		goto callmsg_err;
580 	}
581 
582 	/*
583 	 * Point the remote transport address in the service_transport
584 	 * handle at the address in the request.
585 	 */
586 	clone_xprt->xp_rtaddr.buf = conn->c_raddr.buf;
587 	clone_xprt->xp_rtaddr.len = conn->c_raddr.len;
588 	clone_xprt->xp_rtaddr.maxlen = conn->c_raddr.len;
589 
590 	clone_xprt->xp_lcladdr.buf = conn->c_laddr.buf;
591 	clone_xprt->xp_lcladdr.len = conn->c_laddr.len;
592 	clone_xprt->xp_lcladdr.maxlen = conn->c_laddr.len;
593 
594 	/*
595 	 * In case of RDMA, connection management is
596 	 * entirely done in rpcib module and netid in the
597 	 * SVCMASTERXPRT is NULL. Initialize the clone netid
598 	 * from the connection.
599 	 */
600 
601 	clone_xprt->xp_netid = conn->c_netid;
602 
603 	clone_xprt->xp_xid = xid;
604 	crdp->conn = conn;
605 
606 	freeb(mp);
607 
608 	return (TRUE);
609 
610 callmsg_err:
611 	rdma_buf_free(conn, &crdp->rpcbuf);
612 
613 cll_malloc_err:
614 	if (cl)
615 		clist_free(cl);
616 xdr_err:
617 	XDR_DESTROY(xdrs);
618 
619 badrpc_call:
620 	RDMA_BUF_FREE(conn, &rdp->rpcmsg);
621 	RDMA_REL_CONN(conn);
622 	freeb(mp);
623 	RSSTAT_INCR(rsbadcalls);
624 	return (FALSE);
625 }
626 
627 static int
628 svc_process_long_reply(SVCXPRT * clone_xprt,
629     xdrproc_t xdr_results, caddr_t xdr_location,
630     struct rpc_msg *msg, bool_t has_args, int *msglen,
631     int *freelen, int *numchunks, unsigned int *final_len)
632 {
633 	int status;
634 	XDR xdrslong;
635 	struct clist *wcl = NULL;
636 	int count = 0;
637 	int alloc_len;
638 	char  *memp;
639 	rdma_buf_t long_rpc = {0};
640 	struct clone_rdma_data *crdp;
641 
642 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
643 
644 	bzero(&xdrslong, sizeof (xdrslong));
645 
646 	/* Choose a size for the long rpc response */
647 	if (MSG_IS_RPCSEC_GSS(msg)) {
648 		alloc_len = RNDUP(MAX_AUTH_BYTES + *msglen);
649 	} else {
650 		alloc_len = RNDUP(*msglen);
651 	}
652 
653 	if (alloc_len <= 64 * 1024) {
654 		if (alloc_len > 32 * 1024) {
655 			alloc_len = 64 * 1024;
656 		} else {
657 			if (alloc_len > 16 * 1024) {
658 				alloc_len = 32 * 1024;
659 			} else {
660 				alloc_len = 16 * 1024;
661 			}
662 		}
663 	}
664 
665 	long_rpc.type = RDMA_LONG_BUFFER;
666 	long_rpc.len = alloc_len;
667 	if (rdma_buf_alloc(crdp->conn, &long_rpc)) {
668 		return (SVC_RDMA_FAIL);
669 	}
670 
671 	memp = long_rpc.addr;
672 	xdrmem_create(&xdrslong, memp, alloc_len, XDR_ENCODE);
673 
674 	msg->rm_xid = clone_xprt->xp_xid;
675 
676 	if (!(xdr_replymsg(&xdrslong, msg) &&
677 	    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, &xdrslong,
678 	    xdr_results, xdr_location)))) {
679 		rdma_buf_free(crdp->conn, &long_rpc);
680 		DTRACE_PROBE(krpc__e__svcrdma__longrep__authwrap);
681 		return (SVC_RDMA_FAIL);
682 	}
683 
684 	*final_len = XDR_GETPOS(&xdrslong);
685 
686 	DTRACE_PROBE1(krpc__i__replylen, uint_t, *final_len);
687 	*numchunks = 0;
688 	*freelen = 0;
689 
690 	wcl = crdp->cl_reply;
691 	wcl->rb_longbuf = long_rpc;
692 
693 	count = *final_len;
694 	while ((wcl != NULL) && (count > 0)) {
695 
696 		if (wcl->c_dmemhandle.mrc_rmr == 0)
697 			break;
698 
699 		DTRACE_PROBE2(krpc__i__write__chunks, uint32_t, count,
700 		    uint32_t, wcl->c_len);
701 
702 		if (wcl->c_len > count) {
703 			wcl->c_len = count;
704 		}
705 		wcl->w.c_saddr3 = (caddr_t)memp;
706 
707 		count -= wcl->c_len;
708 		*numchunks +=  1;
709 		memp += wcl->c_len;
710 		wcl = wcl->c_next;
711 	}
712 
713 	/*
714 	 * Make rest of the chunks 0-len
715 	 */
716 	while (wcl != NULL) {
717 		if (wcl->c_dmemhandle.mrc_rmr == 0)
718 			break;
719 		wcl->c_len = 0;
720 		wcl = wcl->c_next;
721 	}
722 
723 	wcl = crdp->cl_reply;
724 
725 	/*
726 	 * MUST fail if there are still more data
727 	 */
728 	if (count > 0) {
729 		rdma_buf_free(crdp->conn, &long_rpc);
730 		DTRACE_PROBE(krpc__e__svcrdma__longrep__dlen__clist);
731 		return (SVC_RDMA_FAIL);
732 	}
733 
734 	if (clist_register(crdp->conn, wcl, CLIST_REG_SOURCE) != RDMA_SUCCESS) {
735 		rdma_buf_free(crdp->conn, &long_rpc);
736 		DTRACE_PROBE(krpc__e__svcrdma__longrep__clistreg);
737 		return (SVC_RDMA_FAIL);
738 	}
739 
740 	status = clist_syncmem(crdp->conn, wcl, CLIST_REG_SOURCE);
741 
742 	if (status) {
743 		(void) clist_deregister(crdp->conn, wcl);
744 		rdma_buf_free(crdp->conn, &long_rpc);
745 		DTRACE_PROBE(krpc__e__svcrdma__longrep__syncmem);
746 		return (SVC_RDMA_FAIL);
747 	}
748 
749 	status = RDMA_WRITE(crdp->conn, wcl, WAIT);
750 
751 	(void) clist_deregister(crdp->conn, wcl);
752 	rdma_buf_free(crdp->conn, &wcl->rb_longbuf);
753 
754 	if (status != RDMA_SUCCESS) {
755 		DTRACE_PROBE(krpc__e__svcrdma__longrep__write);
756 		return (SVC_RDMA_FAIL);
757 	}
758 
759 	return (SVC_RDMA_SUCCESS);
760 }
761 
762 
763 static int
764 svc_compose_rpcmsg(SVCXPRT * clone_xprt, CONN * conn, xdrproc_t xdr_results,
765     caddr_t xdr_location, rdma_buf_t *rpcreply, XDR ** xdrs,
766     struct rpc_msg *msg, bool_t has_args, uint_t *len)
767 {
768 	/*
769 	 * Get a pre-allocated buffer for rpc reply
770 	 */
771 	rpcreply->type = SEND_BUFFER;
772 	if (rdma_buf_alloc(conn, rpcreply)) {
773 		DTRACE_PROBE(krpc__e__svcrdma__rpcmsg__reply__nofreebufs);
774 		return (SVC_RDMA_FAIL);
775 	}
776 
777 	xdrrdma_create(*xdrs, rpcreply->addr, rpcreply->len,
778 	    0, NULL, XDR_ENCODE, conn);
779 
780 	msg->rm_xid = clone_xprt->xp_xid;
781 
782 	if (has_args) {
783 		if (!(xdr_replymsg(*xdrs, msg) &&
784 		    (!has_args ||
785 		    SVCAUTH_WRAP(&clone_xprt->xp_auth, *xdrs,
786 		    xdr_results, xdr_location)))) {
787 			rdma_buf_free(conn, rpcreply);
788 			DTRACE_PROBE(
789 			    krpc__e__svcrdma__rpcmsg__reply__authwrap1);
790 			return (SVC_RDMA_FAIL);
791 		}
792 	} else {
793 		if (!xdr_replymsg(*xdrs, msg)) {
794 			rdma_buf_free(conn, rpcreply);
795 			DTRACE_PROBE(
796 			    krpc__e__svcrdma__rpcmsg__reply__authwrap2);
797 			return (SVC_RDMA_FAIL);
798 		}
799 	}
800 
801 	*len = XDR_GETPOS(*xdrs);
802 
803 	return (SVC_RDMA_SUCCESS);
804 }
805 
806 /*
807  * Send rpc reply.
808  */
809 static bool_t
810 svc_rdma_ksend(SVCXPRT * clone_xprt, struct rpc_msg *msg)
811 {
812 	XDR *xdrs_rpc = &(clone_xprt->xp_xdrout);
813 	XDR xdrs_rhdr;
814 	CONN *conn = NULL;
815 	rdma_buf_t rbuf_resp = {0}, rbuf_rpc_resp = {0};
816 
817 	struct clone_rdma_data *crdp;
818 	struct clist *cl_read = NULL;
819 	struct clist *cl_send = NULL;
820 	struct clist *cl_write = NULL;
821 	xdrproc_t xdr_results;		/* results XDR encoding function */
822 	caddr_t xdr_location;		/* response results pointer */
823 
824 	int retval = FALSE;
825 	int status, msglen, num_wreply_segments = 0;
826 	uint32_t rdma_credit = 0;
827 	int freelen = 0;
828 	bool_t has_args;
829 	uint_t  final_resp_len, rdma_response_op, vers;
830 
831 	bzero(&xdrs_rhdr, sizeof (XDR));
832 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
833 	conn = crdp->conn;
834 
835 	/*
836 	 * If there is a result procedure specified in the reply message,
837 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
838 	 * We need to make sure it won't be processed twice, so we null
839 	 * it for xdr_replymsg here.
840 	 */
841 	has_args = FALSE;
842 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
843 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
844 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
845 			has_args = TRUE;
846 			xdr_location = msg->acpted_rply.ar_results.where;
847 			msg->acpted_rply.ar_results.proc = xdr_void;
848 			msg->acpted_rply.ar_results.where = NULL;
849 		}
850 	}
851 
852 	/*
853 	 * Given the limit on the inline response size (RPC_MSG_SZ),
854 	 * there is a need to make a guess as to the overall size of
855 	 * the response.  If the resultant size is beyond the inline
856 	 * size, then the server needs to use the "reply chunk list"
857 	 * provided by the client (if the client provided one).  An
858 	 * example of this type of response would be a READDIR
859 	 * response (e.g. a small directory read would fit in RPC_MSG_SZ
860 	 * and that is the preference but it may not fit)
861 	 *
862 	 * Combine the encoded size and the size of the true results
863 	 * and then make the decision about where to encode and send results.
864 	 *
865 	 * One important note, this calculation is ignoring the size
866 	 * of the encoding of the authentication overhead.  The reason
867 	 * for this is rooted in the complexities of access to the
868 	 * encoded size of RPCSEC_GSS related authentiation,
869 	 * integrity, and privacy.
870 	 *
871 	 * If it turns out that the encoded authentication bumps the
872 	 * response over the RPC_MSG_SZ limit, then it may need to
873 	 * attempt to encode for the reply chunk list.
874 	 */
875 
876 	/*
877 	 * Calculating the "sizeof" the RPC response header and the
878 	 * encoded results.
879 	 */
880 	msglen = xdr_sizeof(xdr_replymsg, msg);
881 
882 	if (msglen > 0) {
883 		RSSTAT_INCR(rstotalreplies);
884 	}
885 	if (has_args)
886 		msglen += xdrrdma_sizeof(xdr_results, xdr_location,
887 		    rdma_minchunk, NULL, NULL);
888 
889 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__msglen, int, msglen);
890 
891 	status = SVC_RDMA_SUCCESS;
892 
893 	if (msglen < RPC_MSG_SZ) {
894 		/*
895 		 * Looks like the response will fit in the inline
896 		 * response; let's try
897 		 */
898 		RSSTAT_INCR(rstotalinlinereplies);
899 
900 		rdma_response_op = RDMA_MSG;
901 
902 		status = svc_compose_rpcmsg(clone_xprt, conn, xdr_results,
903 		    xdr_location, &rbuf_rpc_resp, &xdrs_rpc, msg,
904 		    has_args, &final_resp_len);
905 
906 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_status,
907 		    int, status);
908 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_len,
909 		    int, final_resp_len);
910 
911 		if (status == SVC_RDMA_SUCCESS && crdp->cl_reply) {
912 			clist_free(crdp->cl_reply);
913 			crdp->cl_reply = NULL;
914 		}
915 	}
916 
917 	/*
918 	 * If the encode failed (size?) or the message really is
919 	 * larger than what is allowed, try the response chunk list.
920 	 */
921 	if (status != SVC_RDMA_SUCCESS || msglen >= RPC_MSG_SZ) {
922 		/*
923 		 * attempting to use a reply chunk list when there
924 		 * isn't one won't get very far...
925 		 */
926 		if (crdp->cl_reply == NULL) {
927 			DTRACE_PROBE(krpc__e__svcrdma__ksend__noreplycl);
928 			goto out;
929 		}
930 
931 		RSSTAT_INCR(rstotallongreplies);
932 
933 		msglen = xdr_sizeof(xdr_replymsg, msg);
934 		msglen += xdrrdma_sizeof(xdr_results, xdr_location, 0,
935 		    NULL, NULL);
936 
937 		status = svc_process_long_reply(clone_xprt, xdr_results,
938 		    xdr_location, msg, has_args, &msglen, &freelen,
939 		    &num_wreply_segments, &final_resp_len);
940 
941 		DTRACE_PROBE1(krpc__i__svcrdma__ksend__longreplen,
942 		    int, final_resp_len);
943 
944 		if (status != SVC_RDMA_SUCCESS) {
945 			DTRACE_PROBE(krpc__e__svcrdma__ksend__compose__failed);
946 			goto out;
947 		}
948 
949 		rdma_response_op = RDMA_NOMSG;
950 	}
951 
952 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__rdmamsg__len,
953 	    int, final_resp_len);
954 
955 	rbuf_resp.type = SEND_BUFFER;
956 	if (rdma_buf_alloc(conn, &rbuf_resp)) {
957 		rdma_buf_free(conn, &rbuf_rpc_resp);
958 		DTRACE_PROBE(krpc__e__svcrdma__ksend__nofreebufs);
959 		goto out;
960 	}
961 
962 	rdma_credit = rdma_bufs_granted;
963 
964 	vers = RPCRDMA_VERS;
965 	xdrmem_create(&xdrs_rhdr, rbuf_resp.addr, rbuf_resp.len, XDR_ENCODE);
966 	(*(uint32_t *)rbuf_resp.addr) = msg->rm_xid;
967 	/* Skip xid and set the xdr position accordingly. */
968 	XDR_SETPOS(&xdrs_rhdr, sizeof (uint32_t));
969 	if (!xdr_u_int(&xdrs_rhdr, &vers) ||
970 	    !xdr_u_int(&xdrs_rhdr, &rdma_credit) ||
971 	    !xdr_u_int(&xdrs_rhdr, &rdma_response_op)) {
972 		rdma_buf_free(conn, &rbuf_rpc_resp);
973 		rdma_buf_free(conn, &rbuf_resp);
974 		DTRACE_PROBE(krpc__e__svcrdma__ksend__uint);
975 		goto out;
976 	}
977 
978 	/*
979 	 * Now XDR the read chunk list, actually always NULL
980 	 */
981 	(void) xdr_encode_rlist_svc(&xdrs_rhdr, cl_read);
982 
983 	/*
984 	 * encode write list -- we already drove RDMA_WRITEs
985 	 */
986 	cl_write = crdp->cl_wlist;
987 	if (!xdr_encode_wlist(&xdrs_rhdr, cl_write)) {
988 		DTRACE_PROBE(krpc__e__svcrdma__ksend__enc__wlist);
989 		rdma_buf_free(conn, &rbuf_rpc_resp);
990 		rdma_buf_free(conn, &rbuf_resp);
991 		goto out;
992 	}
993 
994 	/*
995 	 * XDR encode the RDMA_REPLY write chunk
996 	 */
997 	if (!xdr_encode_reply_wchunk(&xdrs_rhdr, crdp->cl_reply,
998 	    num_wreply_segments)) {
999 		rdma_buf_free(conn, &rbuf_rpc_resp);
1000 		rdma_buf_free(conn, &rbuf_resp);
1001 		goto out;
1002 	}
1003 
1004 	clist_add(&cl_send, 0, XDR_GETPOS(&xdrs_rhdr), &rbuf_resp.handle,
1005 	    rbuf_resp.addr, NULL, NULL);
1006 
1007 	if (rdma_response_op == RDMA_MSG) {
1008 		clist_add(&cl_send, 0, final_resp_len, &rbuf_rpc_resp.handle,
1009 		    rbuf_rpc_resp.addr, NULL, NULL);
1010 	}
1011 
1012 	status = RDMA_SEND(conn, cl_send, msg->rm_xid);
1013 
1014 	if (status == RDMA_SUCCESS) {
1015 		retval = TRUE;
1016 	}
1017 
1018 out:
1019 	/*
1020 	 * Free up sendlist chunks
1021 	 */
1022 	if (cl_send != NULL)
1023 		clist_free(cl_send);
1024 
1025 	/*
1026 	 * Destroy private data for xdr rdma
1027 	 */
1028 	if (clone_xprt->xp_xdrout.x_ops != NULL) {
1029 		XDR_DESTROY(&(clone_xprt->xp_xdrout));
1030 	}
1031 
1032 	if (crdp->cl_reply) {
1033 		clist_free(crdp->cl_reply);
1034 		crdp->cl_reply = NULL;
1035 	}
1036 
1037 	/*
1038 	 * This is completely disgusting.  If public is set it is
1039 	 * a pointer to a structure whose first field is the address
1040 	 * of the function to free that structure and any related
1041 	 * stuff.  (see rrokfree in nfs_xdr.c).
1042 	 */
1043 	if (xdrs_rpc->x_public) {
1044 		/* LINTED pointer alignment */
1045 		(**((int (**)()) xdrs_rpc->x_public)) (xdrs_rpc->x_public);
1046 	}
1047 
1048 	if (xdrs_rhdr.x_ops != NULL) {
1049 		XDR_DESTROY(&xdrs_rhdr);
1050 	}
1051 
1052 	return (retval);
1053 }
1054 
1055 /*
1056  * Deserialize arguments.
1057  */
1058 static bool_t
1059 svc_rdma_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, caddr_t args_ptr)
1060 {
1061 	if ((SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
1062 	    xdr_args, args_ptr)) != TRUE)
1063 		return (FALSE);
1064 	return (TRUE);
1065 }
1066 
1067 static bool_t
1068 svc_rdma_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
1069     caddr_t args_ptr)
1070 {
1071 	struct clone_rdma_data *crdp;
1072 	bool_t retval;
1073 
1074 	/*
1075 	 * If the cloned bit is true, then this transport specific
1076 	 * rmda data has been duplicated into another cloned xprt. Do
1077 	 * not free, or release the connection, it is still in use.  The
1078 	 * buffers will be freed and the connection released later by
1079 	 * SVC_CLONE_DESTROY().
1080 	 */
1081 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
1082 	if (crdp->cloned == TRUE) {
1083 		crdp->cloned = 0;
1084 		return (TRUE);
1085 	}
1086 
1087 	/*
1088 	 * Free the args if needed then XDR_DESTROY
1089 	 */
1090 	if (args_ptr) {
1091 		XDR	*xdrs = &clone_xprt->xp_xdrin;
1092 
1093 		xdrs->x_op = XDR_FREE;
1094 		retval = (*xdr_args)(xdrs, args_ptr);
1095 	}
1096 
1097 	XDR_DESTROY(&(clone_xprt->xp_xdrin));
1098 	rdma_buf_free(crdp->conn, &crdp->rpcbuf);
1099 	if (crdp->cl_reply) {
1100 		clist_free(crdp->cl_reply);
1101 		crdp->cl_reply = NULL;
1102 	}
1103 	RDMA_REL_CONN(crdp->conn);
1104 
1105 	return (retval);
1106 }
1107 
1108 /* ARGSUSED */
1109 static int32_t *
1110 svc_rdma_kgetres(SVCXPRT *clone_xprt, int size)
1111 {
1112 	return (NULL);
1113 }
1114 
1115 /* ARGSUSED */
1116 static void
1117 svc_rdma_kfreeres(SVCXPRT *clone_xprt)
1118 {
1119 }
1120 
1121 /*
1122  * the dup cacheing routines below provide a cache of non-failure
1123  * transaction id's.  rpc service routines can use this to detect
1124  * retransmissions and re-send a non-failure response.
1125  */
1126 
1127 /*
1128  * MAXDUPREQS is the number of cached items.  It should be adjusted
1129  * to the service load so that there is likely to be a response entry
1130  * when the first retransmission comes in.
1131  */
1132 #define	MAXDUPREQS	1024
1133 
1134 /*
1135  * This should be appropriately scaled to MAXDUPREQS.
1136  */
1137 #define	DRHASHSZ	257
1138 
1139 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
1140 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
1141 #else
1142 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
1143 #endif
1144 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
1145 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
1146 
1147 static int	rdmandupreqs = 0;
1148 int	rdmamaxdupreqs = MAXDUPREQS;
1149 static kmutex_t rdmadupreq_lock;
1150 static struct dupreq *rdmadrhashtbl[DRHASHSZ];
1151 static int	rdmadrhashstat[DRHASHSZ];
1152 
1153 static void unhash(struct dupreq *);
1154 
1155 /*
1156  * rdmadrmru points to the head of a circular linked list in lru order.
1157  * rdmadrmru->dr_next == drlru
1158  */
1159 struct dupreq *rdmadrmru;
1160 
1161 /*
1162  * svc_rdma_kdup searches the request cache and returns 0 if the
1163  * request is not found in the cache.  If it is found, then it
1164  * returns the state of the request (in progress or done) and
1165  * the status or attributes that were part of the original reply.
1166  */
1167 static int
1168 svc_rdma_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
1169 	bool_t *dupcachedp)
1170 {
1171 	struct dupreq *dr;
1172 	uint32_t xid;
1173 	uint32_t drhash;
1174 	int status;
1175 
1176 	xid = REQTOXID(req);
1177 	mutex_enter(&rdmadupreq_lock);
1178 	RSSTAT_INCR(rsdupchecks);
1179 	/*
1180 	 * Check to see whether an entry already exists in the cache.
1181 	 */
1182 	dr = rdmadrhashtbl[XIDHASH(xid)];
1183 	while (dr != NULL) {
1184 		if (dr->dr_xid == xid &&
1185 		    dr->dr_proc == req->rq_proc &&
1186 		    dr->dr_prog == req->rq_prog &&
1187 		    dr->dr_vers == req->rq_vers &&
1188 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
1189 		    bcmp((caddr_t)dr->dr_addr.buf,
1190 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
1191 		    dr->dr_addr.len) == 0) {
1192 			status = dr->dr_status;
1193 			if (status == DUP_DONE) {
1194 				bcopy(dr->dr_resp.buf, res, size);
1195 				if (dupcachedp != NULL)
1196 					*dupcachedp = (dr->dr_resfree != NULL);
1197 			} else {
1198 				dr->dr_status = DUP_INPROGRESS;
1199 				*drpp = dr;
1200 			}
1201 			RSSTAT_INCR(rsdupreqs);
1202 			mutex_exit(&rdmadupreq_lock);
1203 			return (status);
1204 		}
1205 		dr = dr->dr_chain;
1206 	}
1207 
1208 	/*
1209 	 * There wasn't an entry, either allocate a new one or recycle
1210 	 * an old one.
1211 	 */
1212 	if (rdmandupreqs < rdmamaxdupreqs) {
1213 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
1214 		if (dr == NULL) {
1215 			mutex_exit(&rdmadupreq_lock);
1216 			return (DUP_ERROR);
1217 		}
1218 		dr->dr_resp.buf = NULL;
1219 		dr->dr_resp.maxlen = 0;
1220 		dr->dr_addr.buf = NULL;
1221 		dr->dr_addr.maxlen = 0;
1222 		if (rdmadrmru) {
1223 			dr->dr_next = rdmadrmru->dr_next;
1224 			rdmadrmru->dr_next = dr;
1225 		} else {
1226 			dr->dr_next = dr;
1227 		}
1228 		rdmandupreqs++;
1229 	} else {
1230 		dr = rdmadrmru->dr_next;
1231 		while (dr->dr_status == DUP_INPROGRESS) {
1232 			dr = dr->dr_next;
1233 			if (dr == rdmadrmru->dr_next) {
1234 				mutex_exit(&rdmadupreq_lock);
1235 				return (DUP_ERROR);
1236 			}
1237 		}
1238 		unhash(dr);
1239 		if (dr->dr_resfree) {
1240 			(*dr->dr_resfree)(dr->dr_resp.buf);
1241 		}
1242 	}
1243 	dr->dr_resfree = NULL;
1244 	rdmadrmru = dr;
1245 
1246 	dr->dr_xid = REQTOXID(req);
1247 	dr->dr_prog = req->rq_prog;
1248 	dr->dr_vers = req->rq_vers;
1249 	dr->dr_proc = req->rq_proc;
1250 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
1251 		if (dr->dr_addr.buf != NULL)
1252 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
1253 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
1254 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
1255 		if (dr->dr_addr.buf == NULL) {
1256 			dr->dr_addr.maxlen = 0;
1257 			dr->dr_status = DUP_DROP;
1258 			mutex_exit(&rdmadupreq_lock);
1259 			return (DUP_ERROR);
1260 		}
1261 	}
1262 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
1263 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
1264 	if (dr->dr_resp.maxlen < size) {
1265 		if (dr->dr_resp.buf != NULL)
1266 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
1267 		dr->dr_resp.maxlen = (unsigned int)size;
1268 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
1269 		if (dr->dr_resp.buf == NULL) {
1270 			dr->dr_resp.maxlen = 0;
1271 			dr->dr_status = DUP_DROP;
1272 			mutex_exit(&rdmadupreq_lock);
1273 			return (DUP_ERROR);
1274 		}
1275 	}
1276 	dr->dr_status = DUP_INPROGRESS;
1277 
1278 	drhash = (uint32_t)DRHASH(dr);
1279 	dr->dr_chain = rdmadrhashtbl[drhash];
1280 	rdmadrhashtbl[drhash] = dr;
1281 	rdmadrhashstat[drhash]++;
1282 	mutex_exit(&rdmadupreq_lock);
1283 	*drpp = dr;
1284 	return (DUP_NEW);
1285 }
1286 
1287 /*
1288  * svc_rdma_kdupdone marks the request done (DUP_DONE or DUP_DROP)
1289  * and stores the response.
1290  */
1291 static void
1292 svc_rdma_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
1293 	int size, int status)
1294 {
1295 	ASSERT(dr->dr_resfree == NULL);
1296 	if (status == DUP_DONE) {
1297 		bcopy(res, dr->dr_resp.buf, size);
1298 		dr->dr_resfree = dis_resfree;
1299 	}
1300 	dr->dr_status = status;
1301 }
1302 
1303 /*
1304  * This routine expects that the mutex, rdmadupreq_lock, is already held.
1305  */
1306 static void
1307 unhash(struct dupreq *dr)
1308 {
1309 	struct dupreq *drt;
1310 	struct dupreq *drtprev = NULL;
1311 	uint32_t drhash;
1312 
1313 	ASSERT(MUTEX_HELD(&rdmadupreq_lock));
1314 
1315 	drhash = (uint32_t)DRHASH(dr);
1316 	drt = rdmadrhashtbl[drhash];
1317 	while (drt != NULL) {
1318 		if (drt == dr) {
1319 			rdmadrhashstat[drhash]--;
1320 			if (drtprev == NULL) {
1321 				rdmadrhashtbl[drhash] = drt->dr_chain;
1322 			} else {
1323 				drtprev->dr_chain = drt->dr_chain;
1324 			}
1325 			return;
1326 		}
1327 		drtprev = drt;
1328 		drt = drt->dr_chain;
1329 	}
1330 }
1331 
1332 bool_t
1333 rdma_get_wchunk(struct svc_req *req, iovec_t *iov, struct clist *wlist)
1334 {
1335 	struct clist	*clist;
1336 	uint32_t	tlen;
1337 
1338 	if (req->rq_xprt->xp_type != T_RDMA) {
1339 		return (FALSE);
1340 	}
1341 
1342 	tlen = 0;
1343 	clist = wlist;
1344 	while (clist) {
1345 		tlen += clist->c_len;
1346 		clist = clist->c_next;
1347 	}
1348 
1349 	/*
1350 	 * set iov to addr+len of first segment of first wchunk of
1351 	 * wlist sent by client.  krecv() already malloc'd a buffer
1352 	 * large enough, but registration is deferred until we write
1353 	 * the buffer back to (NFS) client using RDMA_WRITE.
1354 	 */
1355 	iov->iov_base = (caddr_t)(uintptr_t)wlist->w.c_saddr;
1356 	iov->iov_len = tlen;
1357 
1358 	return (TRUE);
1359 }
1360 
1361 /*
1362  * routine to setup the read chunk lists
1363  */
1364 
1365 int
1366 rdma_setup_read_chunks(struct clist *wcl, uint32_t count, int *wcl_len)
1367 {
1368 	int		data_len, avail_len;
1369 	uint_t		round_len;
1370 
1371 	data_len = avail_len = 0;
1372 
1373 	while (wcl != NULL && count > 0) {
1374 		if (wcl->c_dmemhandle.mrc_rmr == 0)
1375 			break;
1376 
1377 		if (wcl->c_len < count) {
1378 			data_len += wcl->c_len;
1379 			avail_len = 0;
1380 		} else {
1381 			data_len += count;
1382 			avail_len = wcl->c_len - count;
1383 			wcl->c_len = count;
1384 		}
1385 		count -= wcl->c_len;
1386 
1387 		if (count == 0)
1388 			break;
1389 
1390 		wcl = wcl->c_next;
1391 	}
1392 
1393 	/*
1394 	 * MUST fail if there are still more data
1395 	 */
1396 	if (count > 0) {
1397 		DTRACE_PROBE2(krpc__e__rdma_setup_read_chunks_clist_len,
1398 		    int, data_len, int, count);
1399 		return (FALSE);
1400 	}
1401 
1402 	/*
1403 	 * Round up the last chunk to 4-byte boundary
1404 	 */
1405 	*wcl_len = roundup(data_len, BYTES_PER_XDR_UNIT);
1406 	round_len = *wcl_len - data_len;
1407 
1408 	if (round_len) {
1409 
1410 		/*
1411 		 * If there is space in the current chunk,
1412 		 * add the roundup to the chunk.
1413 		 */
1414 		if (avail_len >= round_len) {
1415 			wcl->c_len += round_len;
1416 		} else  {
1417 			/*
1418 			 * try the next one.
1419 			 */
1420 			wcl = wcl->c_next;
1421 			if ((wcl == NULL) || (wcl->c_len < round_len)) {
1422 				DTRACE_PROBE1(
1423 				    krpc__e__rdma_setup_read_chunks_rndup,
1424 				    int, round_len);
1425 				return (FALSE);
1426 			}
1427 			wcl->c_len = round_len;
1428 		}
1429 	}
1430 
1431 	wcl = wcl->c_next;
1432 
1433 	/*
1434 	 * Make rest of the chunks 0-len
1435 	 */
1436 
1437 	clist_zero_len(wcl);
1438 
1439 	return (TRUE);
1440 }
1441