xref: /illumos-gate/usr/src/uts/common/rpc/svc_rdma.c (revision 47842382d52f28aa3173aa6b511781c322ccb6a2)
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 /* 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 	CONN		*conn;		/* RDMA connection */
94 	rdma_buf_t	rpcbuf;		/* RPC req/resp buffer */
95 	struct clist	*cl_reply;	/* reply chunk buffer info */
96 	struct clist	*cl_wlist;		/* write list clist */
97 };
98 
99 #define	MAXADDRLEN	128	/* max length for address mask */
100 
101 /*
102  * Routines exported through ops vector.
103  */
104 static bool_t		svc_rdma_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
105 static bool_t		svc_rdma_ksend(SVCXPRT *, struct rpc_msg *);
106 static bool_t		svc_rdma_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
107 static bool_t		svc_rdma_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
108 void			svc_rdma_kdestroy(SVCMASTERXPRT *);
109 static int		svc_rdma_kdup(struct svc_req *, caddr_t, int,
110 				struct dupreq **, bool_t *);
111 static void		svc_rdma_kdupdone(struct dupreq *, caddr_t,
112 				void (*)(), int, int);
113 static int32_t		*svc_rdma_kgetres(SVCXPRT *, int);
114 static void		svc_rdma_kfreeres(SVCXPRT *);
115 static void		svc_rdma_kclone_destroy(SVCXPRT *);
116 static void		svc_rdma_kstart(SVCMASTERXPRT *);
117 void			svc_rdma_kstop(SVCMASTERXPRT *);
118 
119 static int	svc_process_long_reply(SVCXPRT *, xdrproc_t,
120 			caddr_t, struct rpc_msg *, bool_t, int *,
121 			int *, int *, unsigned int *);
122 
123 static int	svc_compose_rpcmsg(SVCXPRT *, CONN *, xdrproc_t,
124 			caddr_t, rdma_buf_t *, XDR **, struct rpc_msg *,
125 			bool_t, uint_t *);
126 static bool_t rpcmsg_length(xdrproc_t,
127 		caddr_t,
128 		struct rpc_msg *, bool_t, int);
129 
130 /*
131  * Server transport operations vector.
132  */
133 struct svc_ops rdma_svc_ops = {
134 	svc_rdma_krecv,		/* Get requests */
135 	svc_rdma_kgetargs,	/* Deserialize arguments */
136 	svc_rdma_ksend,		/* Send reply */
137 	svc_rdma_kfreeargs,	/* Free argument data space */
138 	svc_rdma_kdestroy,	/* Destroy transport handle */
139 	svc_rdma_kdup,		/* Check entry in dup req cache */
140 	svc_rdma_kdupdone,	/* Mark entry in dup req cache as done */
141 	svc_rdma_kgetres,	/* Get pointer to response buffer */
142 	svc_rdma_kfreeres,	/* Destroy pre-serialized response header */
143 	svc_rdma_kclone_destroy,	/* Destroy a clone xprt */
144 	svc_rdma_kstart		/* Tell `ready-to-receive' to rpcmod */
145 };
146 
147 /*
148  * Server statistics
149  * NOTE: This structure type is duplicated in the NFS fast path.
150  */
151 struct {
152 	kstat_named_t	rscalls;
153 	kstat_named_t	rsbadcalls;
154 	kstat_named_t	rsnullrecv;
155 	kstat_named_t	rsbadlen;
156 	kstat_named_t	rsxdrcall;
157 	kstat_named_t	rsdupchecks;
158 	kstat_named_t	rsdupreqs;
159 	kstat_named_t	rslongrpcs;
160 	kstat_named_t	rstotalreplies;
161 	kstat_named_t	rstotallongreplies;
162 	kstat_named_t	rstotalinlinereplies;
163 } rdmarsstat = {
164 	{ "calls",	KSTAT_DATA_UINT64 },
165 	{ "badcalls",	KSTAT_DATA_UINT64 },
166 	{ "nullrecv",	KSTAT_DATA_UINT64 },
167 	{ "badlen",	KSTAT_DATA_UINT64 },
168 	{ "xdrcall",	KSTAT_DATA_UINT64 },
169 	{ "dupchecks",	KSTAT_DATA_UINT64 },
170 	{ "dupreqs",	KSTAT_DATA_UINT64 },
171 	{ "longrpcs",	KSTAT_DATA_UINT64 },
172 	{ "totalreplies",	KSTAT_DATA_UINT64 },
173 	{ "totallongreplies",	KSTAT_DATA_UINT64 },
174 	{ "totalinlinereplies",	KSTAT_DATA_UINT64 },
175 };
176 
177 kstat_named_t *rdmarsstat_ptr = (kstat_named_t *)&rdmarsstat;
178 uint_t rdmarsstat_ndata = sizeof (rdmarsstat) / sizeof (kstat_named_t);
179 
180 #define	RSSTAT_INCR(x)	atomic_add_64(&rdmarsstat.x.value.ui64, 1)
181 /*
182  * Create a transport record.
183  * The transport record, output buffer, and private data structure
184  * are allocated.  The output buffer is serialized into using xdrmem.
185  * There is one transport record per user process which implements a
186  * set of services.
187  */
188 /* ARGSUSED */
189 int
190 svc_rdma_kcreate(char *netid, SVC_CALLOUT_TABLE *sct, int id,
191     rdma_xprt_group_t *started_xprts)
192 {
193 	int error;
194 	SVCMASTERXPRT *xprt;
195 	struct rdma_data *rd;
196 	rdma_registry_t *rmod;
197 	rdma_xprt_record_t *xprt_rec;
198 	queue_t	*q;
199 	/*
200 	 * modload the RDMA plugins is not already done.
201 	 */
202 	if (!rdma_modloaded) {
203 		/*CONSTANTCONDITION*/
204 		ASSERT(sizeof (struct clone_rdma_data) <= SVC_P2LEN);
205 
206 		mutex_enter(&rdma_modload_lock);
207 		if (!rdma_modloaded) {
208 			error = rdma_modload();
209 		}
210 		mutex_exit(&rdma_modload_lock);
211 
212 		if (error)
213 			return (error);
214 	}
215 
216 	/*
217 	 * master_xprt_count is the count of master transport handles
218 	 * that were successfully created and are ready to recieve for
219 	 * RDMA based access.
220 	 */
221 	error = 0;
222 	xprt_rec = NULL;
223 	rw_enter(&rdma_lock, RW_READER);
224 	if (rdma_mod_head == NULL) {
225 		started_xprts->rtg_count = 0;
226 		rw_exit(&rdma_lock);
227 		if (rdma_dev_available)
228 			return (EPROTONOSUPPORT);
229 		else
230 			return (ENODEV);
231 	}
232 
233 	/*
234 	 * If we have reached here, then atleast one RDMA plugin has loaded.
235 	 * Create a master_xprt, make it start listenining on the device,
236 	 * if an error is generated, record it, we might need to shut
237 	 * the master_xprt.
238 	 * SVC_START() calls svc_rdma_kstart which calls plugin binding
239 	 * routines.
240 	 */
241 	for (rmod = rdma_mod_head; rmod != NULL; rmod = rmod->r_next) {
242 
243 		/*
244 		 * One SVCMASTERXPRT per RDMA plugin.
245 		 */
246 		xprt = kmem_zalloc(sizeof (*xprt), KM_SLEEP);
247 		xprt->xp_ops = &rdma_svc_ops;
248 		xprt->xp_sct = sct;
249 		xprt->xp_type = T_RDMA;
250 		mutex_init(&xprt->xp_req_lock, NULL, MUTEX_DEFAULT, NULL);
251 		mutex_init(&xprt->xp_thread_lock, NULL, MUTEX_DEFAULT, NULL);
252 		xprt->xp_req_head = (mblk_t *)0;
253 		xprt->xp_req_tail = (mblk_t *)0;
254 		xprt->xp_threads = 0;
255 		xprt->xp_detached_threads = 0;
256 
257 		rd = kmem_zalloc(sizeof (*rd), KM_SLEEP);
258 		xprt->xp_p2 = (caddr_t)rd;
259 		rd->rd_xprt = xprt;
260 		rd->r_mod = rmod->r_mod;
261 
262 		q = &rd->rd_data.q;
263 		xprt->xp_wq = q;
264 		q->q_ptr = &rd->rd_xprt;
265 		xprt->xp_netid = NULL;
266 
267 		if (netid != NULL) {
268 			xprt->xp_netid = kmem_alloc(strlen(netid) + 1,
269 			    KM_SLEEP);
270 			(void) strcpy(xprt->xp_netid, netid);
271 		}
272 
273 		xprt->xp_addrmask.maxlen =
274 		    xprt->xp_addrmask.len = sizeof (struct sockaddr_in);
275 		xprt->xp_addrmask.buf =
276 		    kmem_zalloc(xprt->xp_addrmask.len, KM_SLEEP);
277 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_addr.s_addr =
278 		    (uint32_t)~0;
279 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_family =
280 		    (ushort_t)~0;
281 
282 		/*
283 		 * Each of the plugins will have their own Service ID
284 		 * to listener specific mapping, like port number for VI
285 		 * and service name for IB.
286 		 */
287 		rd->rd_data.svcid = id;
288 		error = svc_xprt_register(xprt, id);
289 		if (error) {
290 			DTRACE_PROBE(krpc__e__svcrdma__xprt__reg);
291 			goto cleanup;
292 		}
293 
294 		SVC_START(xprt);
295 		if (!rd->rd_data.active) {
296 			svc_xprt_unregister(xprt);
297 			error = rd->rd_data.err_code;
298 			goto cleanup;
299 		}
300 
301 		/*
302 		 * This is set only when there is atleast one or more
303 		 * transports successfully created. We insert the pointer
304 		 * to the created RDMA master xprt into a separately maintained
305 		 * list. This way we can easily reference it later to cleanup,
306 		 * when NFS kRPC service pool is going away/unregistered.
307 		 */
308 		started_xprts->rtg_count ++;
309 		xprt_rec = kmem_alloc(sizeof (*xprt_rec), KM_SLEEP);
310 		xprt_rec->rtr_xprt_ptr = xprt;
311 		xprt_rec->rtr_next = started_xprts->rtg_listhead;
312 		started_xprts->rtg_listhead = xprt_rec;
313 		continue;
314 cleanup:
315 		SVC_DESTROY(xprt);
316 		if (error == RDMA_FAILED)
317 			error = EPROTONOSUPPORT;
318 	}
319 
320 	rw_exit(&rdma_lock);
321 
322 	/*
323 	 * Don't return any error even if a single plugin was started
324 	 * successfully.
325 	 */
326 	if (started_xprts->rtg_count == 0)
327 		return (error);
328 	return (0);
329 }
330 
331 /*
332  * Cleanup routine for freeing up memory allocated by
333  * svc_rdma_kcreate()
334  */
335 void
336 svc_rdma_kdestroy(SVCMASTERXPRT *xprt)
337 {
338 	struct rdma_data *rd = (struct rdma_data *)xprt->xp_p2;
339 
340 
341 	mutex_destroy(&xprt->xp_req_lock);
342 	mutex_destroy(&xprt->xp_thread_lock);
343 	kmem_free(xprt->xp_netid, strlen(xprt->xp_netid) + 1);
344 	kmem_free(rd, sizeof (*rd));
345 	kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
346 	kmem_free(xprt, sizeof (*xprt));
347 }
348 
349 
350 static void
351 svc_rdma_kstart(SVCMASTERXPRT *xprt)
352 {
353 	struct rdma_svc_data *svcdata;
354 	rdma_mod_t *rmod;
355 
356 	svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data;
357 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
358 
359 	/*
360 	 * Create a listener for  module at this port
361 	 */
362 
363 	if (rmod->rdma_count != 0)
364 		(*rmod->rdma_ops->rdma_svc_listen)(svcdata);
365 	else
366 		svcdata->err_code = RDMA_FAILED;
367 }
368 
369 void
370 svc_rdma_kstop(SVCMASTERXPRT *xprt)
371 {
372 	struct rdma_svc_data *svcdata;
373 	rdma_mod_t *rmod;
374 
375 	svcdata	= &((struct rdma_data *)xprt->xp_p2)->rd_data;
376 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
377 
378 	/*
379 	 * Call the stop listener routine for each plugin. If rdma_count is
380 	 * already zero set active to zero.
381 	 */
382 	if (rmod->rdma_count != 0)
383 		(*rmod->rdma_ops->rdma_svc_stop)(svcdata);
384 	else
385 		svcdata->active = 0;
386 	if (svcdata->active)
387 		DTRACE_PROBE(krpc__e__svcrdma__kstop);
388 }
389 
390 /* ARGSUSED */
391 static void
392 svc_rdma_kclone_destroy(SVCXPRT *clone_xprt)
393 {
394 }
395 
396 static bool_t
397 svc_rdma_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
398 {
399 	XDR	*xdrs;
400 	CONN	*conn;
401 
402 	rdma_recv_data_t	*rdp = (rdma_recv_data_t *)mp->b_rptr;
403 	struct clone_rdma_data *crdp;
404 	struct clist	*cl = NULL;
405 	struct clist	*wcl = NULL;
406 	struct clist	*cllong = NULL;
407 
408 	rdma_stat	status;
409 	uint32_t vers, op, pos, xid;
410 	uint32_t rdma_credit;
411 	uint32_t wcl_total_length = 0;
412 	bool_t	wwl = FALSE;
413 
414 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
415 	RSSTAT_INCR(rscalls);
416 	conn = rdp->conn;
417 
418 	status = rdma_svc_postrecv(conn);
419 	if (status != RDMA_SUCCESS) {
420 		DTRACE_PROBE(krpc__e__svcrdma__krecv__postrecv);
421 		goto badrpc_call;
422 	}
423 
424 	xdrs = &clone_xprt->xp_xdrin;
425 	xdrmem_create(xdrs, rdp->rpcmsg.addr, rdp->rpcmsg.len, XDR_DECODE);
426 	xid = *(uint32_t *)rdp->rpcmsg.addr;
427 	XDR_SETPOS(xdrs, sizeof (uint32_t));
428 
429 	if (! xdr_u_int(xdrs, &vers) ||
430 	    ! xdr_u_int(xdrs, &rdma_credit) ||
431 	    ! xdr_u_int(xdrs, &op)) {
432 		DTRACE_PROBE(krpc__e__svcrdma__krecv__uint);
433 		goto xdr_err;
434 	}
435 
436 	/* Checking if the status of the recv operation was normal */
437 	if (rdp->status != 0) {
438 		DTRACE_PROBE1(krpc__e__svcrdma__krecv__invalid__status,
439 		    int, rdp->status);
440 		goto badrpc_call;
441 	}
442 
443 	if (! xdr_do_clist(xdrs, &cl)) {
444 		DTRACE_PROBE(krpc__e__svcrdma__krecv__do__clist);
445 		goto xdr_err;
446 	}
447 
448 	if (!xdr_decode_wlist_svc(xdrs, &wcl, &wwl, &wcl_total_length, conn)) {
449 		DTRACE_PROBE(krpc__e__svcrdma__krecv__decode__wlist);
450 		if (cl)
451 			clist_free(cl);
452 		goto xdr_err;
453 	}
454 	crdp->cl_wlist = wcl;
455 
456 	crdp->cl_reply = NULL;
457 	(void) xdr_decode_reply_wchunk(xdrs, &crdp->cl_reply);
458 
459 	/*
460 	 * A chunk at 0 offset indicates that the RPC call message
461 	 * is in a chunk. Get the RPC call message chunk.
462 	 */
463 	if (cl != NULL && op == RDMA_NOMSG) {
464 
465 		/* Remove RPC call message chunk from chunklist */
466 		cllong = cl;
467 		cl = cl->c_next;
468 		cllong->c_next = NULL;
469 
470 
471 		/* Allocate and register memory for the RPC call msg chunk */
472 		cllong->rb_longbuf.type = RDMA_LONG_BUFFER;
473 		cllong->rb_longbuf.len = cllong->c_len > LONG_REPLY_LEN ?
474 		    cllong->c_len : LONG_REPLY_LEN;
475 
476 		if (rdma_buf_alloc(conn, &cllong->rb_longbuf)) {
477 			clist_free(cllong);
478 			goto cll_malloc_err;
479 		}
480 
481 		cllong->u.c_daddr3 = cllong->rb_longbuf.addr;
482 
483 		if (cllong->u.c_daddr == NULL) {
484 			DTRACE_PROBE(krpc__e__svcrdma__krecv__nomem);
485 			rdma_buf_free(conn, &cllong->rb_longbuf);
486 			clist_free(cllong);
487 			goto cll_malloc_err;
488 		}
489 
490 		status = clist_register(conn, cllong, CLIST_REG_DST);
491 		if (status) {
492 			DTRACE_PROBE(krpc__e__svcrdma__krecv__clist__reg);
493 			rdma_buf_free(conn, &cllong->rb_longbuf);
494 			clist_free(cllong);
495 			goto cll_malloc_err;
496 		}
497 
498 		/*
499 		 * Now read the RPC call message in
500 		 */
501 		status = RDMA_READ(conn, cllong, WAIT);
502 		if (status) {
503 			DTRACE_PROBE(krpc__e__svcrdma__krecv__read);
504 			(void) clist_deregister(conn, cllong, CLIST_REG_DST);
505 			rdma_buf_free(conn, &cllong->rb_longbuf);
506 			clist_free(cllong);
507 			goto cll_malloc_err;
508 		}
509 
510 		status = clist_syncmem(conn, cllong, CLIST_REG_DST);
511 		(void) clist_deregister(conn, cllong, CLIST_REG_DST);
512 
513 		xdrrdma_create(xdrs, (caddr_t)(uintptr_t)cllong->u.c_daddr3,
514 		    cllong->c_len, 0, cl, XDR_DECODE, conn);
515 
516 		crdp->rpcbuf = cllong->rb_longbuf;
517 		crdp->rpcbuf.len = cllong->c_len;
518 		clist_free(cllong);
519 		RDMA_BUF_FREE(conn, &rdp->rpcmsg);
520 	} else {
521 		pos = XDR_GETPOS(xdrs);
522 		xdrrdma_create(xdrs, rdp->rpcmsg.addr + pos,
523 		    rdp->rpcmsg.len - pos, 0, cl, XDR_DECODE, conn);
524 		crdp->rpcbuf = rdp->rpcmsg;
525 
526 		/* Use xdrrdmablk_ops to indicate there is a read chunk list */
527 		if (cl != NULL) {
528 			int32_t flg = XDR_RDMA_RLIST_REG;
529 
530 			XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
531 			xdrs->x_ops = &xdrrdmablk_ops;
532 		}
533 	}
534 
535 	if (crdp->cl_wlist) {
536 		int32_t flg = XDR_RDMA_WLIST_REG;
537 
538 		XDR_CONTROL(xdrs, XDR_RDMA_SET_WLIST, crdp->cl_wlist);
539 		XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
540 	}
541 
542 	if (! xdr_callmsg(xdrs, msg)) {
543 		DTRACE_PROBE(krpc__e__svcrdma__krecv__callmsg);
544 		RSSTAT_INCR(rsxdrcall);
545 		goto callmsg_err;
546 	}
547 
548 	/*
549 	 * Point the remote transport address in the service_transport
550 	 * handle at the address in the request.
551 	 */
552 	clone_xprt->xp_rtaddr.buf = conn->c_raddr.buf;
553 	clone_xprt->xp_rtaddr.len = conn->c_raddr.len;
554 	clone_xprt->xp_rtaddr.maxlen = conn->c_raddr.len;
555 	clone_xprt->xp_xid = xid;
556 	crdp->conn = conn;
557 
558 	freeb(mp);
559 
560 	return (TRUE);
561 
562 callmsg_err:
563 	rdma_buf_free(conn, &crdp->rpcbuf);
564 
565 cll_malloc_err:
566 	if (cl)
567 		clist_free(cl);
568 xdr_err:
569 	XDR_DESTROY(xdrs);
570 
571 badrpc_call:
572 	RDMA_BUF_FREE(conn, &rdp->rpcmsg);
573 	RDMA_REL_CONN(conn);
574 	freeb(mp);
575 	RSSTAT_INCR(rsbadcalls);
576 	return (FALSE);
577 }
578 
579 static int
580 svc_process_long_reply(SVCXPRT * clone_xprt,
581     xdrproc_t xdr_results, caddr_t xdr_location,
582     struct rpc_msg *msg, bool_t has_args, int *msglen,
583     int *freelen, int *numchunks, unsigned int *final_len)
584 {
585 	int status;
586 	XDR xdrslong;
587 	struct clist *wcl = NULL;
588 	int count = 0;
589 	int alloc_len;
590 	char  *memp;
591 	rdma_buf_t long_rpc = {0};
592 	struct clone_rdma_data *crdp;
593 
594 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
595 
596 	bzero(&xdrslong, sizeof (xdrslong));
597 
598 	/* Choose a size for the long rpc response */
599 	if (MSG_IS_RPCSEC_GSS(msg)) {
600 		alloc_len = RNDUP(MAX_AUTH_BYTES + *msglen);
601 	} else {
602 		alloc_len = RNDUP(*msglen);
603 	}
604 
605 	if (alloc_len <= 64 * 1024) {
606 		if (alloc_len > 32 * 1024) {
607 			alloc_len = 64 * 1024;
608 		} else {
609 			if (alloc_len > 16 * 1024) {
610 				alloc_len = 32 * 1024;
611 			} else {
612 				alloc_len = 16 * 1024;
613 			}
614 		}
615 	}
616 
617 	long_rpc.type = RDMA_LONG_BUFFER;
618 	long_rpc.len = alloc_len;
619 	if (rdma_buf_alloc(crdp->conn, &long_rpc)) {
620 		return (SVC_RDMA_FAIL);
621 	}
622 
623 	memp = long_rpc.addr;
624 	xdrmem_create(&xdrslong, memp, alloc_len, XDR_ENCODE);
625 
626 	msg->rm_xid = clone_xprt->xp_xid;
627 
628 	if (!(xdr_replymsg(&xdrslong, msg) &&
629 	    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, &xdrslong,
630 	    xdr_results, xdr_location)))) {
631 		rdma_buf_free(crdp->conn, &long_rpc);
632 		DTRACE_PROBE(krpc__e__svcrdma__longrep__authwrap);
633 		return (SVC_RDMA_FAIL);
634 	}
635 
636 	*final_len = XDR_GETPOS(&xdrslong);
637 
638 	*numchunks = 0;
639 	*freelen = 0;
640 
641 	wcl = crdp->cl_reply;
642 	wcl->rb_longbuf = long_rpc;
643 
644 	count = *final_len;
645 	while (wcl != NULL) {
646 		if (wcl->c_dmemhandle.mrc_rmr == 0)
647 			break;
648 
649 		if (wcl->c_len > count) {
650 			wcl->c_len = count;
651 		}
652 		wcl->w.c_saddr3 = (caddr_t)memp;
653 
654 		count -= wcl->c_len;
655 		*numchunks +=  1;
656 		if (count == 0)
657 			break;
658 		memp += wcl->c_len;
659 		wcl = wcl->c_next;
660 	}
661 
662 	wcl = crdp->cl_reply;
663 
664 	/*
665 	 * MUST fail if there are still more data
666 	 */
667 	if (count > 0) {
668 		rdma_buf_free(crdp->conn, &long_rpc);
669 		DTRACE_PROBE(krpc__e__svcrdma__longrep__dlen__clist);
670 		return (SVC_RDMA_FAIL);
671 	}
672 
673 	if (clist_register(crdp->conn, wcl, CLIST_REG_SOURCE) != RDMA_SUCCESS) {
674 		rdma_buf_free(crdp->conn, &long_rpc);
675 		DTRACE_PROBE(krpc__e__svcrdma__longrep__clistreg);
676 		return (SVC_RDMA_FAIL);
677 	}
678 
679 	status = clist_syncmem(crdp->conn, wcl, CLIST_REG_SOURCE);
680 
681 	if (status) {
682 		(void) clist_deregister(crdp->conn, wcl, CLIST_REG_SOURCE);
683 		rdma_buf_free(crdp->conn, &long_rpc);
684 		DTRACE_PROBE(krpc__e__svcrdma__longrep__syncmem);
685 		return (SVC_RDMA_FAIL);
686 	}
687 
688 	status = RDMA_WRITE(crdp->conn, wcl, WAIT);
689 
690 	(void) clist_deregister(crdp->conn, wcl, CLIST_REG_SOURCE);
691 	rdma_buf_free(crdp->conn, &wcl->rb_longbuf);
692 
693 	if (status != RDMA_SUCCESS) {
694 		DTRACE_PROBE(krpc__e__svcrdma__longrep__write);
695 		return (SVC_RDMA_FAIL);
696 	}
697 
698 	return (SVC_RDMA_SUCCESS);
699 }
700 
701 
702 static int
703 svc_compose_rpcmsg(SVCXPRT * clone_xprt, CONN * conn, xdrproc_t xdr_results,
704     caddr_t xdr_location, rdma_buf_t *rpcreply, XDR ** xdrs,
705     struct rpc_msg *msg, bool_t has_args, uint_t *len)
706 {
707 	/*
708 	 * Get a pre-allocated buffer for rpc reply
709 	 */
710 	rpcreply->type = SEND_BUFFER;
711 	if (rdma_buf_alloc(conn, rpcreply)) {
712 		DTRACE_PROBE(krpc__e__svcrdma__rpcmsg__reply__nofreebufs);
713 		return (SVC_RDMA_FAIL);
714 	}
715 
716 	xdrrdma_create(*xdrs, rpcreply->addr, rpcreply->len,
717 	    0, NULL, XDR_ENCODE, conn);
718 
719 	msg->rm_xid = clone_xprt->xp_xid;
720 
721 	if (has_args) {
722 		if (!(xdr_replymsg(*xdrs, msg) &&
723 		    (!has_args ||
724 		    SVCAUTH_WRAP(&clone_xprt->xp_auth, *xdrs,
725 		    xdr_results, xdr_location)))) {
726 			rdma_buf_free(conn, rpcreply);
727 			DTRACE_PROBE(
728 			    krpc__e__svcrdma__rpcmsg__reply__authwrap1);
729 			return (SVC_RDMA_FAIL);
730 		}
731 	} else {
732 		if (!xdr_replymsg(*xdrs, msg)) {
733 			rdma_buf_free(conn, rpcreply);
734 			DTRACE_PROBE(
735 			    krpc__e__svcrdma__rpcmsg__reply__authwrap2);
736 			return (SVC_RDMA_FAIL);
737 		}
738 	}
739 
740 	*len = XDR_GETPOS(*xdrs);
741 
742 	return (SVC_RDMA_SUCCESS);
743 }
744 
745 /*
746  * Send rpc reply.
747  */
748 static bool_t
749 svc_rdma_ksend(SVCXPRT * clone_xprt, struct rpc_msg *msg)
750 {
751 	XDR *xdrs_rpc = &(clone_xprt->xp_xdrout);
752 	XDR xdrs_rhdr;
753 	CONN *conn = NULL;
754 	rdma_buf_t rbuf_resp = {0}, rbuf_rpc_resp = {0};
755 
756 	struct clone_rdma_data *crdp;
757 	struct clist *cl_read = NULL;
758 	struct clist *cl_send = NULL;
759 	struct clist *cl_write = NULL;
760 	xdrproc_t xdr_results;		/* results XDR encoding function */
761 	caddr_t xdr_location;		/* response results pointer */
762 
763 	int retval = FALSE;
764 	int status, msglen, num_wreply_segments = 0;
765 	uint32_t rdma_credit = 0;
766 	int freelen = 0;
767 	bool_t has_args;
768 	uint_t  final_resp_len, rdma_response_op, vers;
769 
770 	bzero(&xdrs_rhdr, sizeof (XDR));
771 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
772 	conn = crdp->conn;
773 
774 	/*
775 	 * If there is a result procedure specified in the reply message,
776 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
777 	 * We need to make sure it won't be processed twice, so we null
778 	 * it for xdr_replymsg here.
779 	 */
780 	has_args = FALSE;
781 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
782 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
783 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
784 			has_args = TRUE;
785 			xdr_location = msg->acpted_rply.ar_results.where;
786 			msg->acpted_rply.ar_results.proc = xdr_void;
787 			msg->acpted_rply.ar_results.where = NULL;
788 		}
789 	}
790 
791 	/*
792 	 * Given the limit on the inline response size (RPC_MSG_SZ),
793 	 * there is a need to make a guess as to the overall size of
794 	 * the response.  If the resultant size is beyond the inline
795 	 * size, then the server needs to use the "reply chunk list"
796 	 * provided by the client (if the client provided one).  An
797 	 * example of this type of response would be a READDIR
798 	 * response (e.g. a small directory read would fit in RPC_MSG_SZ
799 	 * and that is the preference but it may not fit)
800 	 *
801 	 * Combine the encoded size and the size of the true results
802 	 * and then make the decision about where to encode and send results.
803 	 *
804 	 * One important note, this calculation is ignoring the size
805 	 * of the encoding of the authentication overhead.  The reason
806 	 * for this is rooted in the complexities of access to the
807 	 * encoded size of RPCSEC_GSS related authentiation,
808 	 * integrity, and privacy.
809 	 *
810 	 * If it turns out that the encoded authentication bumps the
811 	 * response over the RPC_MSG_SZ limit, then it may need to
812 	 * attempt to encode for the reply chunk list.
813 	 */
814 
815 	/*
816 	 * Calculating the "sizeof" the RPC response header and the
817 	 * encoded results.
818 	 */
819 	msglen = xdr_sizeof(xdr_replymsg, msg);
820 
821 	if (msglen > 0) {
822 		RSSTAT_INCR(rstotalreplies);
823 	}
824 	if (has_args)
825 		msglen += xdrrdma_sizeof(xdr_results, xdr_location,
826 		    rdma_minchunk, NULL, NULL);
827 
828 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__msglen, int, msglen);
829 
830 	status = SVC_RDMA_SUCCESS;
831 
832 	if (msglen < RPC_MSG_SZ) {
833 		/*
834 		 * Looks like the response will fit in the inline
835 		 * response; let's try
836 		 */
837 		RSSTAT_INCR(rstotalinlinereplies);
838 
839 		rdma_response_op = RDMA_MSG;
840 
841 		status = svc_compose_rpcmsg(clone_xprt, conn, xdr_results,
842 		    xdr_location, &rbuf_rpc_resp, &xdrs_rpc, msg,
843 		    has_args, &final_resp_len);
844 
845 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_status,
846 		    int, status);
847 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_len,
848 		    int, final_resp_len);
849 
850 		if (status == SVC_RDMA_SUCCESS && crdp->cl_reply) {
851 			clist_free(crdp->cl_reply);
852 			crdp->cl_reply = NULL;
853 		}
854 	}
855 
856 	/*
857 	 * If the encode failed (size?) or the message really is
858 	 * larger than what is allowed, try the response chunk list.
859 	 */
860 	if (status != SVC_RDMA_SUCCESS || msglen >= RPC_MSG_SZ) {
861 		/*
862 		 * attempting to use a reply chunk list when there
863 		 * isn't one won't get very far...
864 		 */
865 		if (crdp->cl_reply == NULL) {
866 			DTRACE_PROBE(krpc__e__svcrdma__ksend__noreplycl);
867 			goto out;
868 		}
869 
870 		RSSTAT_INCR(rstotallongreplies);
871 
872 		msglen = xdr_sizeof(xdr_replymsg, msg);
873 		msglen += xdrrdma_sizeof(xdr_results, xdr_location, 0,
874 		    NULL, NULL);
875 
876 		status = svc_process_long_reply(clone_xprt, xdr_results,
877 		    xdr_location, msg, has_args, &msglen, &freelen,
878 		    &num_wreply_segments, &final_resp_len);
879 
880 		DTRACE_PROBE1(krpc__i__svcrdma__ksend__longreplen,
881 		    int, final_resp_len);
882 
883 		if (status != SVC_RDMA_SUCCESS) {
884 			DTRACE_PROBE(krpc__e__svcrdma__ksend__compose__failed);
885 			goto out;
886 		}
887 
888 		rdma_response_op = RDMA_NOMSG;
889 	}
890 
891 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__rdmamsg__len,
892 	    int, final_resp_len);
893 
894 	rbuf_resp.type = SEND_BUFFER;
895 	if (rdma_buf_alloc(conn, &rbuf_resp)) {
896 		rdma_buf_free(conn, &rbuf_rpc_resp);
897 		DTRACE_PROBE(krpc__e__svcrdma__ksend__nofreebufs);
898 		goto out;
899 	}
900 
901 	rdma_credit = rdma_bufs_granted;
902 
903 	vers = RPCRDMA_VERS;
904 	xdrmem_create(&xdrs_rhdr, rbuf_resp.addr, rbuf_resp.len, XDR_ENCODE);
905 	(*(uint32_t *)rbuf_resp.addr) = msg->rm_xid;
906 	/* Skip xid and set the xdr position accordingly. */
907 	XDR_SETPOS(&xdrs_rhdr, sizeof (uint32_t));
908 	if (!xdr_u_int(&xdrs_rhdr, &vers) ||
909 	    !xdr_u_int(&xdrs_rhdr, &rdma_credit) ||
910 	    !xdr_u_int(&xdrs_rhdr, &rdma_response_op)) {
911 		rdma_buf_free(conn, &rbuf_rpc_resp);
912 		rdma_buf_free(conn, &rbuf_resp);
913 		DTRACE_PROBE(krpc__e__svcrdma__ksend__uint);
914 		goto out;
915 	}
916 
917 	/*
918 	 * Now XDR the read chunk list, actually always NULL
919 	 */
920 	(void) xdr_encode_rlist_svc(&xdrs_rhdr, cl_read);
921 
922 	/*
923 	 * encode write list -- we already drove RDMA_WRITEs
924 	 */
925 	cl_write = crdp->cl_wlist;
926 	if (!xdr_encode_wlist(&xdrs_rhdr, cl_write)) {
927 		DTRACE_PROBE(krpc__e__svcrdma__ksend__enc__wlist);
928 		rdma_buf_free(conn, &rbuf_rpc_resp);
929 		rdma_buf_free(conn, &rbuf_resp);
930 		goto out;
931 	}
932 
933 	/*
934 	 * XDR encode the RDMA_REPLY write chunk
935 	 */
936 	if (!xdr_encode_reply_wchunk(&xdrs_rhdr, crdp->cl_reply,
937 	    num_wreply_segments)) {
938 		rdma_buf_free(conn, &rbuf_rpc_resp);
939 		rdma_buf_free(conn, &rbuf_resp);
940 		goto out;
941 	}
942 
943 	clist_add(&cl_send, 0, XDR_GETPOS(&xdrs_rhdr), &rbuf_resp.handle,
944 	    rbuf_resp.addr, NULL, NULL);
945 
946 	if (rdma_response_op == RDMA_MSG) {
947 		clist_add(&cl_send, 0, final_resp_len, &rbuf_rpc_resp.handle,
948 		    rbuf_rpc_resp.addr, NULL, NULL);
949 	}
950 
951 	status = RDMA_SEND(conn, cl_send, msg->rm_xid);
952 
953 	if (status == RDMA_SUCCESS) {
954 		retval = TRUE;
955 	}
956 
957 out:
958 	/*
959 	 * Free up sendlist chunks
960 	 */
961 	if (cl_send != NULL)
962 		clist_free(cl_send);
963 
964 	/*
965 	 * Destroy private data for xdr rdma
966 	 */
967 	if (clone_xprt->xp_xdrout.x_ops != NULL) {
968 		XDR_DESTROY(&(clone_xprt->xp_xdrout));
969 	}
970 
971 	if (crdp->cl_reply) {
972 		clist_free(crdp->cl_reply);
973 		crdp->cl_reply = NULL;
974 	}
975 
976 	/*
977 	 * This is completely disgusting.  If public is set it is
978 	 * a pointer to a structure whose first field is the address
979 	 * of the function to free that structure and any related
980 	 * stuff.  (see rrokfree in nfs_xdr.c).
981 	 */
982 	if (xdrs_rpc->x_public) {
983 		/* LINTED pointer alignment */
984 		(**((int (**)()) xdrs_rpc->x_public)) (xdrs_rpc->x_public);
985 	}
986 
987 	if (xdrs_rhdr.x_ops != NULL) {
988 		XDR_DESTROY(&xdrs_rhdr);
989 	}
990 
991 	return (retval);
992 }
993 
994 /*
995  * Deserialize arguments.
996  */
997 static bool_t
998 svc_rdma_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, caddr_t args_ptr)
999 {
1000 	if ((SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
1001 	    xdr_args, args_ptr)) != TRUE)
1002 		return (FALSE);
1003 	return (TRUE);
1004 }
1005 
1006 static bool_t
1007 svc_rdma_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
1008     caddr_t args_ptr)
1009 {
1010 	struct clone_rdma_data *crdp;
1011 	bool_t retval;
1012 
1013 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
1014 
1015 	/*
1016 	 * Free the args if needed then XDR_DESTROY
1017 	 */
1018 	if (args_ptr) {
1019 		XDR	*xdrs = &clone_xprt->xp_xdrin;
1020 
1021 		xdrs->x_op = XDR_FREE;
1022 		retval = (*xdr_args)(xdrs, args_ptr);
1023 	}
1024 
1025 	XDR_DESTROY(&(clone_xprt->xp_xdrin));
1026 	rdma_buf_free(crdp->conn, &crdp->rpcbuf);
1027 	if (crdp->cl_reply) {
1028 		clist_free(crdp->cl_reply);
1029 		crdp->cl_reply = NULL;
1030 	}
1031 	RDMA_REL_CONN(crdp->conn);
1032 
1033 	return (retval);
1034 }
1035 
1036 /* ARGSUSED */
1037 static int32_t *
1038 svc_rdma_kgetres(SVCXPRT *clone_xprt, int size)
1039 {
1040 	return (NULL);
1041 }
1042 
1043 /* ARGSUSED */
1044 static void
1045 svc_rdma_kfreeres(SVCXPRT *clone_xprt)
1046 {
1047 }
1048 
1049 /*
1050  * the dup cacheing routines below provide a cache of non-failure
1051  * transaction id's.  rpc service routines can use this to detect
1052  * retransmissions and re-send a non-failure response.
1053  */
1054 
1055 /*
1056  * MAXDUPREQS is the number of cached items.  It should be adjusted
1057  * to the service load so that there is likely to be a response entry
1058  * when the first retransmission comes in.
1059  */
1060 #define	MAXDUPREQS	1024
1061 
1062 /*
1063  * This should be appropriately scaled to MAXDUPREQS.
1064  */
1065 #define	DRHASHSZ	257
1066 
1067 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
1068 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
1069 #else
1070 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
1071 #endif
1072 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
1073 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
1074 
1075 static int	rdmandupreqs = 0;
1076 int	rdmamaxdupreqs = MAXDUPREQS;
1077 static kmutex_t rdmadupreq_lock;
1078 static struct dupreq *rdmadrhashtbl[DRHASHSZ];
1079 static int	rdmadrhashstat[DRHASHSZ];
1080 
1081 static void unhash(struct dupreq *);
1082 
1083 /*
1084  * rdmadrmru points to the head of a circular linked list in lru order.
1085  * rdmadrmru->dr_next == drlru
1086  */
1087 struct dupreq *rdmadrmru;
1088 
1089 /*
1090  * svc_rdma_kdup searches the request cache and returns 0 if the
1091  * request is not found in the cache.  If it is found, then it
1092  * returns the state of the request (in progress or done) and
1093  * the status or attributes that were part of the original reply.
1094  */
1095 static int
1096 svc_rdma_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
1097 	bool_t *dupcachedp)
1098 {
1099 	struct dupreq *dr;
1100 	uint32_t xid;
1101 	uint32_t drhash;
1102 	int status;
1103 
1104 	xid = REQTOXID(req);
1105 	mutex_enter(&rdmadupreq_lock);
1106 	RSSTAT_INCR(rsdupchecks);
1107 	/*
1108 	 * Check to see whether an entry already exists in the cache.
1109 	 */
1110 	dr = rdmadrhashtbl[XIDHASH(xid)];
1111 	while (dr != NULL) {
1112 		if (dr->dr_xid == xid &&
1113 		    dr->dr_proc == req->rq_proc &&
1114 		    dr->dr_prog == req->rq_prog &&
1115 		    dr->dr_vers == req->rq_vers &&
1116 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
1117 		    bcmp((caddr_t)dr->dr_addr.buf,
1118 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
1119 		    dr->dr_addr.len) == 0) {
1120 			status = dr->dr_status;
1121 			if (status == DUP_DONE) {
1122 				bcopy(dr->dr_resp.buf, res, size);
1123 				if (dupcachedp != NULL)
1124 					*dupcachedp = (dr->dr_resfree != NULL);
1125 			} else {
1126 				dr->dr_status = DUP_INPROGRESS;
1127 				*drpp = dr;
1128 			}
1129 			RSSTAT_INCR(rsdupreqs);
1130 			mutex_exit(&rdmadupreq_lock);
1131 			return (status);
1132 		}
1133 		dr = dr->dr_chain;
1134 	}
1135 
1136 	/*
1137 	 * There wasn't an entry, either allocate a new one or recycle
1138 	 * an old one.
1139 	 */
1140 	if (rdmandupreqs < rdmamaxdupreqs) {
1141 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
1142 		if (dr == NULL) {
1143 			mutex_exit(&rdmadupreq_lock);
1144 			return (DUP_ERROR);
1145 		}
1146 		dr->dr_resp.buf = NULL;
1147 		dr->dr_resp.maxlen = 0;
1148 		dr->dr_addr.buf = NULL;
1149 		dr->dr_addr.maxlen = 0;
1150 		if (rdmadrmru) {
1151 			dr->dr_next = rdmadrmru->dr_next;
1152 			rdmadrmru->dr_next = dr;
1153 		} else {
1154 			dr->dr_next = dr;
1155 		}
1156 		rdmandupreqs++;
1157 	} else {
1158 		dr = rdmadrmru->dr_next;
1159 		while (dr->dr_status == DUP_INPROGRESS) {
1160 			dr = dr->dr_next;
1161 			if (dr == rdmadrmru->dr_next) {
1162 				mutex_exit(&rdmadupreq_lock);
1163 				return (DUP_ERROR);
1164 			}
1165 		}
1166 		unhash(dr);
1167 		if (dr->dr_resfree) {
1168 			(*dr->dr_resfree)(dr->dr_resp.buf);
1169 		}
1170 	}
1171 	dr->dr_resfree = NULL;
1172 	rdmadrmru = dr;
1173 
1174 	dr->dr_xid = REQTOXID(req);
1175 	dr->dr_prog = req->rq_prog;
1176 	dr->dr_vers = req->rq_vers;
1177 	dr->dr_proc = req->rq_proc;
1178 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
1179 		if (dr->dr_addr.buf != NULL)
1180 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
1181 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
1182 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
1183 		if (dr->dr_addr.buf == NULL) {
1184 			dr->dr_addr.maxlen = 0;
1185 			dr->dr_status = DUP_DROP;
1186 			mutex_exit(&rdmadupreq_lock);
1187 			return (DUP_ERROR);
1188 		}
1189 	}
1190 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
1191 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
1192 	if (dr->dr_resp.maxlen < size) {
1193 		if (dr->dr_resp.buf != NULL)
1194 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
1195 		dr->dr_resp.maxlen = (unsigned int)size;
1196 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
1197 		if (dr->dr_resp.buf == NULL) {
1198 			dr->dr_resp.maxlen = 0;
1199 			dr->dr_status = DUP_DROP;
1200 			mutex_exit(&rdmadupreq_lock);
1201 			return (DUP_ERROR);
1202 		}
1203 	}
1204 	dr->dr_status = DUP_INPROGRESS;
1205 
1206 	drhash = (uint32_t)DRHASH(dr);
1207 	dr->dr_chain = rdmadrhashtbl[drhash];
1208 	rdmadrhashtbl[drhash] = dr;
1209 	rdmadrhashstat[drhash]++;
1210 	mutex_exit(&rdmadupreq_lock);
1211 	*drpp = dr;
1212 	return (DUP_NEW);
1213 }
1214 
1215 /*
1216  * svc_rdma_kdupdone marks the request done (DUP_DONE or DUP_DROP)
1217  * and stores the response.
1218  */
1219 static void
1220 svc_rdma_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
1221 	int size, int status)
1222 {
1223 	ASSERT(dr->dr_resfree == NULL);
1224 	if (status == DUP_DONE) {
1225 		bcopy(res, dr->dr_resp.buf, size);
1226 		dr->dr_resfree = dis_resfree;
1227 	}
1228 	dr->dr_status = status;
1229 }
1230 
1231 /*
1232  * This routine expects that the mutex, rdmadupreq_lock, is already held.
1233  */
1234 static void
1235 unhash(struct dupreq *dr)
1236 {
1237 	struct dupreq *drt;
1238 	struct dupreq *drtprev = NULL;
1239 	uint32_t drhash;
1240 
1241 	ASSERT(MUTEX_HELD(&rdmadupreq_lock));
1242 
1243 	drhash = (uint32_t)DRHASH(dr);
1244 	drt = rdmadrhashtbl[drhash];
1245 	while (drt != NULL) {
1246 		if (drt == dr) {
1247 			rdmadrhashstat[drhash]--;
1248 			if (drtprev == NULL) {
1249 				rdmadrhashtbl[drhash] = drt->dr_chain;
1250 			} else {
1251 				drtprev->dr_chain = drt->dr_chain;
1252 			}
1253 			return;
1254 		}
1255 		drtprev = drt;
1256 		drt = drt->dr_chain;
1257 	}
1258 }
1259 
1260 bool_t
1261 rdma_get_wchunk(struct svc_req *req, iovec_t *iov, struct clist *wlist)
1262 {
1263 	struct clist	*clist;
1264 	uint32_t	tlen;
1265 
1266 	if (req->rq_xprt->xp_type != T_RDMA) {
1267 		return (FALSE);
1268 	}
1269 
1270 	tlen = 0;
1271 	clist = wlist;
1272 	while (clist) {
1273 		tlen += clist->c_len;
1274 		clist = clist->c_next;
1275 	}
1276 
1277 	/*
1278 	 * set iov to addr+len of first segment of first wchunk of
1279 	 * wlist sent by client.  krecv() already malloc'd a buffer
1280 	 * large enough, but registration is deferred until we write
1281 	 * the buffer back to (NFS) client using RDMA_WRITE.
1282 	 */
1283 	iov->iov_base = (caddr_t)(uintptr_t)wlist->w.c_saddr;
1284 	iov->iov_len = tlen;
1285 
1286 	return (TRUE);
1287 }
1288