xref: /illumos-gate/usr/src/uts/common/rpc/svc_rdma.c (revision e4d060fb4c00d44cd578713eb9a921f594b733b8)
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 		xprt->xp_addrmask.maxlen =
268 		    xprt->xp_addrmask.len = sizeof (struct sockaddr_in);
269 		xprt->xp_addrmask.buf =
270 		    kmem_zalloc(xprt->xp_addrmask.len, KM_SLEEP);
271 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_addr.s_addr =
272 		    (uint32_t)~0;
273 		((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_family =
274 		    (ushort_t)~0;
275 
276 		/*
277 		 * Each of the plugins will have their own Service ID
278 		 * to listener specific mapping, like port number for VI
279 		 * and service name for IB.
280 		 */
281 		rd->rd_data.svcid = id;
282 		error = svc_xprt_register(xprt, id);
283 		if (error) {
284 			DTRACE_PROBE(krpc__e__svcrdma__xprt__reg);
285 			goto cleanup;
286 		}
287 
288 		SVC_START(xprt);
289 		if (!rd->rd_data.active) {
290 			svc_xprt_unregister(xprt);
291 			error = rd->rd_data.err_code;
292 			goto cleanup;
293 		}
294 
295 		/*
296 		 * This is set only when there is atleast one or more
297 		 * transports successfully created. We insert the pointer
298 		 * to the created RDMA master xprt into a separately maintained
299 		 * list. This way we can easily reference it later to cleanup,
300 		 * when NFS kRPC service pool is going away/unregistered.
301 		 */
302 		started_xprts->rtg_count ++;
303 		xprt_rec = kmem_alloc(sizeof (*xprt_rec), KM_SLEEP);
304 		xprt_rec->rtr_xprt_ptr = xprt;
305 		xprt_rec->rtr_next = started_xprts->rtg_listhead;
306 		started_xprts->rtg_listhead = xprt_rec;
307 		continue;
308 cleanup:
309 		SVC_DESTROY(xprt);
310 		if (error == RDMA_FAILED)
311 			error = EPROTONOSUPPORT;
312 	}
313 
314 	rw_exit(&rdma_lock);
315 
316 	/*
317 	 * Don't return any error even if a single plugin was started
318 	 * successfully.
319 	 */
320 	if (started_xprts->rtg_count == 0)
321 		return (error);
322 	return (0);
323 }
324 
325 /*
326  * Cleanup routine for freeing up memory allocated by
327  * svc_rdma_kcreate()
328  */
329 void
330 svc_rdma_kdestroy(SVCMASTERXPRT *xprt)
331 {
332 	struct rdma_data *rd = (struct rdma_data *)xprt->xp_p2;
333 
334 
335 	mutex_destroy(&xprt->xp_req_lock);
336 	mutex_destroy(&xprt->xp_thread_lock);
337 	kmem_free(rd, sizeof (*rd));
338 	kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
339 	kmem_free(xprt, sizeof (*xprt));
340 }
341 
342 
343 static void
344 svc_rdma_kstart(SVCMASTERXPRT *xprt)
345 {
346 	struct rdma_svc_data *svcdata;
347 	rdma_mod_t *rmod;
348 
349 	svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data;
350 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
351 
352 	/*
353 	 * Create a listener for  module at this port
354 	 */
355 
356 	if (rmod->rdma_count != 0)
357 		(*rmod->rdma_ops->rdma_svc_listen)(svcdata);
358 	else
359 		svcdata->err_code = RDMA_FAILED;
360 }
361 
362 void
363 svc_rdma_kstop(SVCMASTERXPRT *xprt)
364 {
365 	struct rdma_svc_data *svcdata;
366 	rdma_mod_t *rmod;
367 
368 	svcdata	= &((struct rdma_data *)xprt->xp_p2)->rd_data;
369 	rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod;
370 
371 	/*
372 	 * Call the stop listener routine for each plugin. If rdma_count is
373 	 * already zero set active to zero.
374 	 */
375 	if (rmod->rdma_count != 0)
376 		(*rmod->rdma_ops->rdma_svc_stop)(svcdata);
377 	else
378 		svcdata->active = 0;
379 	if (svcdata->active)
380 		DTRACE_PROBE(krpc__e__svcrdma__kstop);
381 }
382 
383 /* ARGSUSED */
384 static void
385 svc_rdma_kclone_destroy(SVCXPRT *clone_xprt)
386 {
387 }
388 
389 static bool_t
390 svc_rdma_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
391 {
392 	XDR	*xdrs;
393 	CONN	*conn;
394 	rdma_recv_data_t	*rdp = (rdma_recv_data_t *)mp->b_rptr;
395 	struct clone_rdma_data *crdp;
396 	struct clist	*cl = NULL;
397 	struct clist	*wcl = NULL;
398 	struct clist	*cllong = NULL;
399 
400 	rdma_stat	status;
401 	uint32_t vers, op, pos, xid;
402 	uint32_t rdma_credit;
403 	uint32_t wcl_total_length = 0;
404 	bool_t	wwl = FALSE;
405 
406 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
407 	RSSTAT_INCR(rscalls);
408 	conn = rdp->conn;
409 
410 	status = rdma_svc_postrecv(conn);
411 	if (status != RDMA_SUCCESS) {
412 		DTRACE_PROBE(krpc__e__svcrdma__krecv__postrecv);
413 		goto badrpc_call;
414 	}
415 
416 	xdrs = &clone_xprt->xp_xdrin;
417 	xdrmem_create(xdrs, rdp->rpcmsg.addr, rdp->rpcmsg.len, XDR_DECODE);
418 	xid = *(uint32_t *)rdp->rpcmsg.addr;
419 	XDR_SETPOS(xdrs, sizeof (uint32_t));
420 
421 	if (! xdr_u_int(xdrs, &vers) ||
422 	    ! xdr_u_int(xdrs, &rdma_credit) ||
423 	    ! xdr_u_int(xdrs, &op)) {
424 		DTRACE_PROBE(krpc__e__svcrdma__krecv__uint);
425 		goto xdr_err;
426 	}
427 
428 	/* Checking if the status of the recv operation was normal */
429 	if (rdp->status != 0) {
430 		DTRACE_PROBE1(krpc__e__svcrdma__krecv__invalid__status,
431 		    int, rdp->status);
432 		goto badrpc_call;
433 	}
434 
435 	if (! xdr_do_clist(xdrs, &cl)) {
436 		DTRACE_PROBE(krpc__e__svcrdma__krecv__do__clist);
437 		goto xdr_err;
438 	}
439 
440 	if (!xdr_decode_wlist_svc(xdrs, &wcl, &wwl, &wcl_total_length, conn)) {
441 		DTRACE_PROBE(krpc__e__svcrdma__krecv__decode__wlist);
442 		if (cl)
443 			clist_free(cl);
444 		goto xdr_err;
445 	}
446 	crdp->cl_wlist = wcl;
447 
448 	crdp->cl_reply = NULL;
449 	(void) xdr_decode_reply_wchunk(xdrs, &crdp->cl_reply);
450 
451 	/*
452 	 * A chunk at 0 offset indicates that the RPC call message
453 	 * is in a chunk. Get the RPC call message chunk.
454 	 */
455 	if (cl != NULL && op == RDMA_NOMSG) {
456 
457 		/* Remove RPC call message chunk from chunklist */
458 		cllong = cl;
459 		cl = cl->c_next;
460 		cllong->c_next = NULL;
461 
462 
463 		/* Allocate and register memory for the RPC call msg chunk */
464 		cllong->rb_longbuf.type = RDMA_LONG_BUFFER;
465 		cllong->rb_longbuf.len = cllong->c_len > LONG_REPLY_LEN ?
466 		    cllong->c_len : LONG_REPLY_LEN;
467 
468 		if (rdma_buf_alloc(conn, &cllong->rb_longbuf)) {
469 			clist_free(cllong);
470 			goto cll_malloc_err;
471 		}
472 
473 		cllong->u.c_daddr3 = cllong->rb_longbuf.addr;
474 
475 		if (cllong->u.c_daddr == NULL) {
476 			DTRACE_PROBE(krpc__e__svcrdma__krecv__nomem);
477 			rdma_buf_free(conn, &cllong->rb_longbuf);
478 			clist_free(cllong);
479 			goto cll_malloc_err;
480 		}
481 
482 		status = clist_register(conn, cllong, CLIST_REG_DST);
483 		if (status) {
484 			DTRACE_PROBE(krpc__e__svcrdma__krecv__clist__reg);
485 			rdma_buf_free(conn, &cllong->rb_longbuf);
486 			clist_free(cllong);
487 			goto cll_malloc_err;
488 		}
489 
490 		/*
491 		 * Now read the RPC call message in
492 		 */
493 		status = RDMA_READ(conn, cllong, WAIT);
494 		if (status) {
495 			DTRACE_PROBE(krpc__e__svcrdma__krecv__read);
496 			(void) clist_deregister(conn, cllong);
497 			rdma_buf_free(conn, &cllong->rb_longbuf);
498 			clist_free(cllong);
499 			goto cll_malloc_err;
500 		}
501 
502 		status = clist_syncmem(conn, cllong, CLIST_REG_DST);
503 		(void) clist_deregister(conn, cllong);
504 
505 		xdrrdma_create(xdrs, (caddr_t)(uintptr_t)cllong->u.c_daddr3,
506 		    cllong->c_len, 0, cl, XDR_DECODE, conn);
507 
508 		crdp->rpcbuf = cllong->rb_longbuf;
509 		crdp->rpcbuf.len = cllong->c_len;
510 		clist_free(cllong);
511 		RDMA_BUF_FREE(conn, &rdp->rpcmsg);
512 	} else {
513 		pos = XDR_GETPOS(xdrs);
514 		xdrrdma_create(xdrs, rdp->rpcmsg.addr + pos,
515 		    rdp->rpcmsg.len - pos, 0, cl, XDR_DECODE, conn);
516 		crdp->rpcbuf = rdp->rpcmsg;
517 
518 		/* Use xdrrdmablk_ops to indicate there is a read chunk list */
519 		if (cl != NULL) {
520 			int32_t flg = XDR_RDMA_RLIST_REG;
521 
522 			XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
523 			xdrs->x_ops = &xdrrdmablk_ops;
524 		}
525 	}
526 
527 	if (crdp->cl_wlist) {
528 		int32_t flg = XDR_RDMA_WLIST_REG;
529 
530 		XDR_CONTROL(xdrs, XDR_RDMA_SET_WLIST, crdp->cl_wlist);
531 		XDR_CONTROL(xdrs, XDR_RDMA_SET_FLAGS, &flg);
532 	}
533 
534 	if (! xdr_callmsg(xdrs, msg)) {
535 		DTRACE_PROBE(krpc__e__svcrdma__krecv__callmsg);
536 		RSSTAT_INCR(rsxdrcall);
537 		goto callmsg_err;
538 	}
539 
540 	/*
541 	 * Point the remote transport address in the service_transport
542 	 * handle at the address in the request.
543 	 */
544 	clone_xprt->xp_rtaddr.buf = conn->c_raddr.buf;
545 	clone_xprt->xp_rtaddr.len = conn->c_raddr.len;
546 	clone_xprt->xp_rtaddr.maxlen = conn->c_raddr.len;
547 
548 	clone_xprt->xp_lcladdr.buf = conn->c_laddr.buf;
549 	clone_xprt->xp_lcladdr.len = conn->c_laddr.len;
550 	clone_xprt->xp_lcladdr.maxlen = conn->c_laddr.len;
551 
552 	/*
553 	 * In case of RDMA, connection management is
554 	 * entirely done in rpcib module and netid in the
555 	 * SVCMASTERXPRT is NULL. Initialize the clone netid
556 	 * from the connection.
557 	 */
558 
559 	clone_xprt->xp_netid = conn->c_netid;
560 
561 	clone_xprt->xp_xid = xid;
562 	crdp->conn = conn;
563 
564 	freeb(mp);
565 
566 	return (TRUE);
567 
568 callmsg_err:
569 	rdma_buf_free(conn, &crdp->rpcbuf);
570 
571 cll_malloc_err:
572 	if (cl)
573 		clist_free(cl);
574 xdr_err:
575 	XDR_DESTROY(xdrs);
576 
577 badrpc_call:
578 	RDMA_BUF_FREE(conn, &rdp->rpcmsg);
579 	RDMA_REL_CONN(conn);
580 	freeb(mp);
581 	RSSTAT_INCR(rsbadcalls);
582 	return (FALSE);
583 }
584 
585 static int
586 svc_process_long_reply(SVCXPRT * clone_xprt,
587     xdrproc_t xdr_results, caddr_t xdr_location,
588     struct rpc_msg *msg, bool_t has_args, int *msglen,
589     int *freelen, int *numchunks, unsigned int *final_len)
590 {
591 	int status;
592 	XDR xdrslong;
593 	struct clist *wcl = NULL;
594 	int count = 0;
595 	int alloc_len;
596 	char  *memp;
597 	rdma_buf_t long_rpc = {0};
598 	struct clone_rdma_data *crdp;
599 
600 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
601 
602 	bzero(&xdrslong, sizeof (xdrslong));
603 
604 	/* Choose a size for the long rpc response */
605 	if (MSG_IS_RPCSEC_GSS(msg)) {
606 		alloc_len = RNDUP(MAX_AUTH_BYTES + *msglen);
607 	} else {
608 		alloc_len = RNDUP(*msglen);
609 	}
610 
611 	if (alloc_len <= 64 * 1024) {
612 		if (alloc_len > 32 * 1024) {
613 			alloc_len = 64 * 1024;
614 		} else {
615 			if (alloc_len > 16 * 1024) {
616 				alloc_len = 32 * 1024;
617 			} else {
618 				alloc_len = 16 * 1024;
619 			}
620 		}
621 	}
622 
623 	long_rpc.type = RDMA_LONG_BUFFER;
624 	long_rpc.len = alloc_len;
625 	if (rdma_buf_alloc(crdp->conn, &long_rpc)) {
626 		return (SVC_RDMA_FAIL);
627 	}
628 
629 	memp = long_rpc.addr;
630 	xdrmem_create(&xdrslong, memp, alloc_len, XDR_ENCODE);
631 
632 	msg->rm_xid = clone_xprt->xp_xid;
633 
634 	if (!(xdr_replymsg(&xdrslong, msg) &&
635 	    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, &xdrslong,
636 	    xdr_results, xdr_location)))) {
637 		rdma_buf_free(crdp->conn, &long_rpc);
638 		DTRACE_PROBE(krpc__e__svcrdma__longrep__authwrap);
639 		return (SVC_RDMA_FAIL);
640 	}
641 
642 	*final_len = XDR_GETPOS(&xdrslong);
643 
644 	DTRACE_PROBE1(krpc__i__replylen, uint_t, *final_len);
645 	*numchunks = 0;
646 	*freelen = 0;
647 
648 	wcl = crdp->cl_reply;
649 	wcl->rb_longbuf = long_rpc;
650 
651 	count = *final_len;
652 	while ((wcl != NULL) && (count > 0)) {
653 
654 		if (wcl->c_dmemhandle.mrc_rmr == 0)
655 			break;
656 
657 		DTRACE_PROBE2(krpc__i__write__chunks, uint32_t, count,
658 		    uint32_t, wcl->c_len);
659 
660 		if (wcl->c_len > count) {
661 			wcl->c_len = count;
662 		}
663 		wcl->w.c_saddr3 = (caddr_t)memp;
664 
665 		count -= wcl->c_len;
666 		*numchunks +=  1;
667 		memp += wcl->c_len;
668 		wcl = wcl->c_next;
669 	}
670 
671 	/*
672 	 * Make rest of the chunks 0-len
673 	 */
674 	while (wcl != NULL) {
675 		if (wcl->c_dmemhandle.mrc_rmr == 0)
676 			break;
677 		wcl->c_len = 0;
678 		wcl = wcl->c_next;
679 	}
680 
681 	wcl = crdp->cl_reply;
682 
683 	/*
684 	 * MUST fail if there are still more data
685 	 */
686 	if (count > 0) {
687 		rdma_buf_free(crdp->conn, &long_rpc);
688 		DTRACE_PROBE(krpc__e__svcrdma__longrep__dlen__clist);
689 		return (SVC_RDMA_FAIL);
690 	}
691 
692 	if (clist_register(crdp->conn, wcl, CLIST_REG_SOURCE) != RDMA_SUCCESS) {
693 		rdma_buf_free(crdp->conn, &long_rpc);
694 		DTRACE_PROBE(krpc__e__svcrdma__longrep__clistreg);
695 		return (SVC_RDMA_FAIL);
696 	}
697 
698 	status = clist_syncmem(crdp->conn, wcl, CLIST_REG_SOURCE);
699 
700 	if (status) {
701 		(void) clist_deregister(crdp->conn, wcl);
702 		rdma_buf_free(crdp->conn, &long_rpc);
703 		DTRACE_PROBE(krpc__e__svcrdma__longrep__syncmem);
704 		return (SVC_RDMA_FAIL);
705 	}
706 
707 	status = RDMA_WRITE(crdp->conn, wcl, WAIT);
708 
709 	(void) clist_deregister(crdp->conn, wcl);
710 	rdma_buf_free(crdp->conn, &wcl->rb_longbuf);
711 
712 	if (status != RDMA_SUCCESS) {
713 		DTRACE_PROBE(krpc__e__svcrdma__longrep__write);
714 		return (SVC_RDMA_FAIL);
715 	}
716 
717 	return (SVC_RDMA_SUCCESS);
718 }
719 
720 
721 static int
722 svc_compose_rpcmsg(SVCXPRT * clone_xprt, CONN * conn, xdrproc_t xdr_results,
723     caddr_t xdr_location, rdma_buf_t *rpcreply, XDR ** xdrs,
724     struct rpc_msg *msg, bool_t has_args, uint_t *len)
725 {
726 	/*
727 	 * Get a pre-allocated buffer for rpc reply
728 	 */
729 	rpcreply->type = SEND_BUFFER;
730 	if (rdma_buf_alloc(conn, rpcreply)) {
731 		DTRACE_PROBE(krpc__e__svcrdma__rpcmsg__reply__nofreebufs);
732 		return (SVC_RDMA_FAIL);
733 	}
734 
735 	xdrrdma_create(*xdrs, rpcreply->addr, rpcreply->len,
736 	    0, NULL, XDR_ENCODE, conn);
737 
738 	msg->rm_xid = clone_xprt->xp_xid;
739 
740 	if (has_args) {
741 		if (!(xdr_replymsg(*xdrs, msg) &&
742 		    (!has_args ||
743 		    SVCAUTH_WRAP(&clone_xprt->xp_auth, *xdrs,
744 		    xdr_results, xdr_location)))) {
745 			rdma_buf_free(conn, rpcreply);
746 			DTRACE_PROBE(
747 			    krpc__e__svcrdma__rpcmsg__reply__authwrap1);
748 			return (SVC_RDMA_FAIL);
749 		}
750 	} else {
751 		if (!xdr_replymsg(*xdrs, msg)) {
752 			rdma_buf_free(conn, rpcreply);
753 			DTRACE_PROBE(
754 			    krpc__e__svcrdma__rpcmsg__reply__authwrap2);
755 			return (SVC_RDMA_FAIL);
756 		}
757 	}
758 
759 	*len = XDR_GETPOS(*xdrs);
760 
761 	return (SVC_RDMA_SUCCESS);
762 }
763 
764 /*
765  * Send rpc reply.
766  */
767 static bool_t
768 svc_rdma_ksend(SVCXPRT * clone_xprt, struct rpc_msg *msg)
769 {
770 	XDR *xdrs_rpc = &(clone_xprt->xp_xdrout);
771 	XDR xdrs_rhdr;
772 	CONN *conn = NULL;
773 	rdma_buf_t rbuf_resp = {0}, rbuf_rpc_resp = {0};
774 
775 	struct clone_rdma_data *crdp;
776 	struct clist *cl_read = NULL;
777 	struct clist *cl_send = NULL;
778 	struct clist *cl_write = NULL;
779 	xdrproc_t xdr_results;		/* results XDR encoding function */
780 	caddr_t xdr_location;		/* response results pointer */
781 
782 	int retval = FALSE;
783 	int status, msglen, num_wreply_segments = 0;
784 	uint32_t rdma_credit = 0;
785 	int freelen = 0;
786 	bool_t has_args;
787 	uint_t  final_resp_len, rdma_response_op, vers;
788 
789 	bzero(&xdrs_rhdr, sizeof (XDR));
790 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
791 	conn = crdp->conn;
792 
793 	/*
794 	 * If there is a result procedure specified in the reply message,
795 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
796 	 * We need to make sure it won't be processed twice, so we null
797 	 * it for xdr_replymsg here.
798 	 */
799 	has_args = FALSE;
800 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
801 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
802 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
803 			has_args = TRUE;
804 			xdr_location = msg->acpted_rply.ar_results.where;
805 			msg->acpted_rply.ar_results.proc = xdr_void;
806 			msg->acpted_rply.ar_results.where = NULL;
807 		}
808 	}
809 
810 	/*
811 	 * Given the limit on the inline response size (RPC_MSG_SZ),
812 	 * there is a need to make a guess as to the overall size of
813 	 * the response.  If the resultant size is beyond the inline
814 	 * size, then the server needs to use the "reply chunk list"
815 	 * provided by the client (if the client provided one).  An
816 	 * example of this type of response would be a READDIR
817 	 * response (e.g. a small directory read would fit in RPC_MSG_SZ
818 	 * and that is the preference but it may not fit)
819 	 *
820 	 * Combine the encoded size and the size of the true results
821 	 * and then make the decision about where to encode and send results.
822 	 *
823 	 * One important note, this calculation is ignoring the size
824 	 * of the encoding of the authentication overhead.  The reason
825 	 * for this is rooted in the complexities of access to the
826 	 * encoded size of RPCSEC_GSS related authentiation,
827 	 * integrity, and privacy.
828 	 *
829 	 * If it turns out that the encoded authentication bumps the
830 	 * response over the RPC_MSG_SZ limit, then it may need to
831 	 * attempt to encode for the reply chunk list.
832 	 */
833 
834 	/*
835 	 * Calculating the "sizeof" the RPC response header and the
836 	 * encoded results.
837 	 */
838 	msglen = xdr_sizeof(xdr_replymsg, msg);
839 
840 	if (msglen > 0) {
841 		RSSTAT_INCR(rstotalreplies);
842 	}
843 	if (has_args)
844 		msglen += xdrrdma_sizeof(xdr_results, xdr_location,
845 		    rdma_minchunk, NULL, NULL);
846 
847 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__msglen, int, msglen);
848 
849 	status = SVC_RDMA_SUCCESS;
850 
851 	if (msglen < RPC_MSG_SZ) {
852 		/*
853 		 * Looks like the response will fit in the inline
854 		 * response; let's try
855 		 */
856 		RSSTAT_INCR(rstotalinlinereplies);
857 
858 		rdma_response_op = RDMA_MSG;
859 
860 		status = svc_compose_rpcmsg(clone_xprt, conn, xdr_results,
861 		    xdr_location, &rbuf_rpc_resp, &xdrs_rpc, msg,
862 		    has_args, &final_resp_len);
863 
864 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_status,
865 		    int, status);
866 		DTRACE_PROBE1(krpc__i__srdma__ksend__compose_len,
867 		    int, final_resp_len);
868 
869 		if (status == SVC_RDMA_SUCCESS && crdp->cl_reply) {
870 			clist_free(crdp->cl_reply);
871 			crdp->cl_reply = NULL;
872 		}
873 	}
874 
875 	/*
876 	 * If the encode failed (size?) or the message really is
877 	 * larger than what is allowed, try the response chunk list.
878 	 */
879 	if (status != SVC_RDMA_SUCCESS || msglen >= RPC_MSG_SZ) {
880 		/*
881 		 * attempting to use a reply chunk list when there
882 		 * isn't one won't get very far...
883 		 */
884 		if (crdp->cl_reply == NULL) {
885 			DTRACE_PROBE(krpc__e__svcrdma__ksend__noreplycl);
886 			goto out;
887 		}
888 
889 		RSSTAT_INCR(rstotallongreplies);
890 
891 		msglen = xdr_sizeof(xdr_replymsg, msg);
892 		msglen += xdrrdma_sizeof(xdr_results, xdr_location, 0,
893 		    NULL, NULL);
894 
895 		status = svc_process_long_reply(clone_xprt, xdr_results,
896 		    xdr_location, msg, has_args, &msglen, &freelen,
897 		    &num_wreply_segments, &final_resp_len);
898 
899 		DTRACE_PROBE1(krpc__i__svcrdma__ksend__longreplen,
900 		    int, final_resp_len);
901 
902 		if (status != SVC_RDMA_SUCCESS) {
903 			DTRACE_PROBE(krpc__e__svcrdma__ksend__compose__failed);
904 			goto out;
905 		}
906 
907 		rdma_response_op = RDMA_NOMSG;
908 	}
909 
910 	DTRACE_PROBE1(krpc__i__svcrdma__ksend__rdmamsg__len,
911 	    int, final_resp_len);
912 
913 	rbuf_resp.type = SEND_BUFFER;
914 	if (rdma_buf_alloc(conn, &rbuf_resp)) {
915 		rdma_buf_free(conn, &rbuf_rpc_resp);
916 		DTRACE_PROBE(krpc__e__svcrdma__ksend__nofreebufs);
917 		goto out;
918 	}
919 
920 	rdma_credit = rdma_bufs_granted;
921 
922 	vers = RPCRDMA_VERS;
923 	xdrmem_create(&xdrs_rhdr, rbuf_resp.addr, rbuf_resp.len, XDR_ENCODE);
924 	(*(uint32_t *)rbuf_resp.addr) = msg->rm_xid;
925 	/* Skip xid and set the xdr position accordingly. */
926 	XDR_SETPOS(&xdrs_rhdr, sizeof (uint32_t));
927 	if (!xdr_u_int(&xdrs_rhdr, &vers) ||
928 	    !xdr_u_int(&xdrs_rhdr, &rdma_credit) ||
929 	    !xdr_u_int(&xdrs_rhdr, &rdma_response_op)) {
930 		rdma_buf_free(conn, &rbuf_rpc_resp);
931 		rdma_buf_free(conn, &rbuf_resp);
932 		DTRACE_PROBE(krpc__e__svcrdma__ksend__uint);
933 		goto out;
934 	}
935 
936 	/*
937 	 * Now XDR the read chunk list, actually always NULL
938 	 */
939 	(void) xdr_encode_rlist_svc(&xdrs_rhdr, cl_read);
940 
941 	/*
942 	 * encode write list -- we already drove RDMA_WRITEs
943 	 */
944 	cl_write = crdp->cl_wlist;
945 	if (!xdr_encode_wlist(&xdrs_rhdr, cl_write)) {
946 		DTRACE_PROBE(krpc__e__svcrdma__ksend__enc__wlist);
947 		rdma_buf_free(conn, &rbuf_rpc_resp);
948 		rdma_buf_free(conn, &rbuf_resp);
949 		goto out;
950 	}
951 
952 	/*
953 	 * XDR encode the RDMA_REPLY write chunk
954 	 */
955 	if (!xdr_encode_reply_wchunk(&xdrs_rhdr, crdp->cl_reply,
956 	    num_wreply_segments)) {
957 		rdma_buf_free(conn, &rbuf_rpc_resp);
958 		rdma_buf_free(conn, &rbuf_resp);
959 		goto out;
960 	}
961 
962 	clist_add(&cl_send, 0, XDR_GETPOS(&xdrs_rhdr), &rbuf_resp.handle,
963 	    rbuf_resp.addr, NULL, NULL);
964 
965 	if (rdma_response_op == RDMA_MSG) {
966 		clist_add(&cl_send, 0, final_resp_len, &rbuf_rpc_resp.handle,
967 		    rbuf_rpc_resp.addr, NULL, NULL);
968 	}
969 
970 	status = RDMA_SEND(conn, cl_send, msg->rm_xid);
971 
972 	if (status == RDMA_SUCCESS) {
973 		retval = TRUE;
974 	}
975 
976 out:
977 	/*
978 	 * Free up sendlist chunks
979 	 */
980 	if (cl_send != NULL)
981 		clist_free(cl_send);
982 
983 	/*
984 	 * Destroy private data for xdr rdma
985 	 */
986 	if (clone_xprt->xp_xdrout.x_ops != NULL) {
987 		XDR_DESTROY(&(clone_xprt->xp_xdrout));
988 	}
989 
990 	if (crdp->cl_reply) {
991 		clist_free(crdp->cl_reply);
992 		crdp->cl_reply = NULL;
993 	}
994 
995 	/*
996 	 * This is completely disgusting.  If public is set it is
997 	 * a pointer to a structure whose first field is the address
998 	 * of the function to free that structure and any related
999 	 * stuff.  (see rrokfree in nfs_xdr.c).
1000 	 */
1001 	if (xdrs_rpc->x_public) {
1002 		/* LINTED pointer alignment */
1003 		(**((int (**)()) xdrs_rpc->x_public)) (xdrs_rpc->x_public);
1004 	}
1005 
1006 	if (xdrs_rhdr.x_ops != NULL) {
1007 		XDR_DESTROY(&xdrs_rhdr);
1008 	}
1009 
1010 	return (retval);
1011 }
1012 
1013 /*
1014  * Deserialize arguments.
1015  */
1016 static bool_t
1017 svc_rdma_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, caddr_t args_ptr)
1018 {
1019 	if ((SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
1020 	    xdr_args, args_ptr)) != TRUE)
1021 		return (FALSE);
1022 	return (TRUE);
1023 }
1024 
1025 static bool_t
1026 svc_rdma_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
1027     caddr_t args_ptr)
1028 {
1029 	struct clone_rdma_data *crdp;
1030 	bool_t retval;
1031 
1032 	crdp = (struct clone_rdma_data *)clone_xprt->xp_p2buf;
1033 
1034 	/*
1035 	 * Free the args if needed then XDR_DESTROY
1036 	 */
1037 	if (args_ptr) {
1038 		XDR	*xdrs = &clone_xprt->xp_xdrin;
1039 
1040 		xdrs->x_op = XDR_FREE;
1041 		retval = (*xdr_args)(xdrs, args_ptr);
1042 	}
1043 
1044 	XDR_DESTROY(&(clone_xprt->xp_xdrin));
1045 	rdma_buf_free(crdp->conn, &crdp->rpcbuf);
1046 	if (crdp->cl_reply) {
1047 		clist_free(crdp->cl_reply);
1048 		crdp->cl_reply = NULL;
1049 	}
1050 	RDMA_REL_CONN(crdp->conn);
1051 
1052 	return (retval);
1053 }
1054 
1055 /* ARGSUSED */
1056 static int32_t *
1057 svc_rdma_kgetres(SVCXPRT *clone_xprt, int size)
1058 {
1059 	return (NULL);
1060 }
1061 
1062 /* ARGSUSED */
1063 static void
1064 svc_rdma_kfreeres(SVCXPRT *clone_xprt)
1065 {
1066 }
1067 
1068 /*
1069  * the dup cacheing routines below provide a cache of non-failure
1070  * transaction id's.  rpc service routines can use this to detect
1071  * retransmissions and re-send a non-failure response.
1072  */
1073 
1074 /*
1075  * MAXDUPREQS is the number of cached items.  It should be adjusted
1076  * to the service load so that there is likely to be a response entry
1077  * when the first retransmission comes in.
1078  */
1079 #define	MAXDUPREQS	1024
1080 
1081 /*
1082  * This should be appropriately scaled to MAXDUPREQS.
1083  */
1084 #define	DRHASHSZ	257
1085 
1086 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
1087 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
1088 #else
1089 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
1090 #endif
1091 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
1092 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
1093 
1094 static int	rdmandupreqs = 0;
1095 int	rdmamaxdupreqs = MAXDUPREQS;
1096 static kmutex_t rdmadupreq_lock;
1097 static struct dupreq *rdmadrhashtbl[DRHASHSZ];
1098 static int	rdmadrhashstat[DRHASHSZ];
1099 
1100 static void unhash(struct dupreq *);
1101 
1102 /*
1103  * rdmadrmru points to the head of a circular linked list in lru order.
1104  * rdmadrmru->dr_next == drlru
1105  */
1106 struct dupreq *rdmadrmru;
1107 
1108 /*
1109  * svc_rdma_kdup searches the request cache and returns 0 if the
1110  * request is not found in the cache.  If it is found, then it
1111  * returns the state of the request (in progress or done) and
1112  * the status or attributes that were part of the original reply.
1113  */
1114 static int
1115 svc_rdma_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
1116 	bool_t *dupcachedp)
1117 {
1118 	struct dupreq *dr;
1119 	uint32_t xid;
1120 	uint32_t drhash;
1121 	int status;
1122 
1123 	xid = REQTOXID(req);
1124 	mutex_enter(&rdmadupreq_lock);
1125 	RSSTAT_INCR(rsdupchecks);
1126 	/*
1127 	 * Check to see whether an entry already exists in the cache.
1128 	 */
1129 	dr = rdmadrhashtbl[XIDHASH(xid)];
1130 	while (dr != NULL) {
1131 		if (dr->dr_xid == xid &&
1132 		    dr->dr_proc == req->rq_proc &&
1133 		    dr->dr_prog == req->rq_prog &&
1134 		    dr->dr_vers == req->rq_vers &&
1135 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
1136 		    bcmp((caddr_t)dr->dr_addr.buf,
1137 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
1138 		    dr->dr_addr.len) == 0) {
1139 			status = dr->dr_status;
1140 			if (status == DUP_DONE) {
1141 				bcopy(dr->dr_resp.buf, res, size);
1142 				if (dupcachedp != NULL)
1143 					*dupcachedp = (dr->dr_resfree != NULL);
1144 			} else {
1145 				dr->dr_status = DUP_INPROGRESS;
1146 				*drpp = dr;
1147 			}
1148 			RSSTAT_INCR(rsdupreqs);
1149 			mutex_exit(&rdmadupreq_lock);
1150 			return (status);
1151 		}
1152 		dr = dr->dr_chain;
1153 	}
1154 
1155 	/*
1156 	 * There wasn't an entry, either allocate a new one or recycle
1157 	 * an old one.
1158 	 */
1159 	if (rdmandupreqs < rdmamaxdupreqs) {
1160 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
1161 		if (dr == NULL) {
1162 			mutex_exit(&rdmadupreq_lock);
1163 			return (DUP_ERROR);
1164 		}
1165 		dr->dr_resp.buf = NULL;
1166 		dr->dr_resp.maxlen = 0;
1167 		dr->dr_addr.buf = NULL;
1168 		dr->dr_addr.maxlen = 0;
1169 		if (rdmadrmru) {
1170 			dr->dr_next = rdmadrmru->dr_next;
1171 			rdmadrmru->dr_next = dr;
1172 		} else {
1173 			dr->dr_next = dr;
1174 		}
1175 		rdmandupreqs++;
1176 	} else {
1177 		dr = rdmadrmru->dr_next;
1178 		while (dr->dr_status == DUP_INPROGRESS) {
1179 			dr = dr->dr_next;
1180 			if (dr == rdmadrmru->dr_next) {
1181 				mutex_exit(&rdmadupreq_lock);
1182 				return (DUP_ERROR);
1183 			}
1184 		}
1185 		unhash(dr);
1186 		if (dr->dr_resfree) {
1187 			(*dr->dr_resfree)(dr->dr_resp.buf);
1188 		}
1189 	}
1190 	dr->dr_resfree = NULL;
1191 	rdmadrmru = dr;
1192 
1193 	dr->dr_xid = REQTOXID(req);
1194 	dr->dr_prog = req->rq_prog;
1195 	dr->dr_vers = req->rq_vers;
1196 	dr->dr_proc = req->rq_proc;
1197 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
1198 		if (dr->dr_addr.buf != NULL)
1199 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
1200 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
1201 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
1202 		if (dr->dr_addr.buf == NULL) {
1203 			dr->dr_addr.maxlen = 0;
1204 			dr->dr_status = DUP_DROP;
1205 			mutex_exit(&rdmadupreq_lock);
1206 			return (DUP_ERROR);
1207 		}
1208 	}
1209 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
1210 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
1211 	if (dr->dr_resp.maxlen < size) {
1212 		if (dr->dr_resp.buf != NULL)
1213 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
1214 		dr->dr_resp.maxlen = (unsigned int)size;
1215 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
1216 		if (dr->dr_resp.buf == NULL) {
1217 			dr->dr_resp.maxlen = 0;
1218 			dr->dr_status = DUP_DROP;
1219 			mutex_exit(&rdmadupreq_lock);
1220 			return (DUP_ERROR);
1221 		}
1222 	}
1223 	dr->dr_status = DUP_INPROGRESS;
1224 
1225 	drhash = (uint32_t)DRHASH(dr);
1226 	dr->dr_chain = rdmadrhashtbl[drhash];
1227 	rdmadrhashtbl[drhash] = dr;
1228 	rdmadrhashstat[drhash]++;
1229 	mutex_exit(&rdmadupreq_lock);
1230 	*drpp = dr;
1231 	return (DUP_NEW);
1232 }
1233 
1234 /*
1235  * svc_rdma_kdupdone marks the request done (DUP_DONE or DUP_DROP)
1236  * and stores the response.
1237  */
1238 static void
1239 svc_rdma_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
1240 	int size, int status)
1241 {
1242 	ASSERT(dr->dr_resfree == NULL);
1243 	if (status == DUP_DONE) {
1244 		bcopy(res, dr->dr_resp.buf, size);
1245 		dr->dr_resfree = dis_resfree;
1246 	}
1247 	dr->dr_status = status;
1248 }
1249 
1250 /*
1251  * This routine expects that the mutex, rdmadupreq_lock, is already held.
1252  */
1253 static void
1254 unhash(struct dupreq *dr)
1255 {
1256 	struct dupreq *drt;
1257 	struct dupreq *drtprev = NULL;
1258 	uint32_t drhash;
1259 
1260 	ASSERT(MUTEX_HELD(&rdmadupreq_lock));
1261 
1262 	drhash = (uint32_t)DRHASH(dr);
1263 	drt = rdmadrhashtbl[drhash];
1264 	while (drt != NULL) {
1265 		if (drt == dr) {
1266 			rdmadrhashstat[drhash]--;
1267 			if (drtprev == NULL) {
1268 				rdmadrhashtbl[drhash] = drt->dr_chain;
1269 			} else {
1270 				drtprev->dr_chain = drt->dr_chain;
1271 			}
1272 			return;
1273 		}
1274 		drtprev = drt;
1275 		drt = drt->dr_chain;
1276 	}
1277 }
1278 
1279 bool_t
1280 rdma_get_wchunk(struct svc_req *req, iovec_t *iov, struct clist *wlist)
1281 {
1282 	struct clist	*clist;
1283 	uint32_t	tlen;
1284 
1285 	if (req->rq_xprt->xp_type != T_RDMA) {
1286 		return (FALSE);
1287 	}
1288 
1289 	tlen = 0;
1290 	clist = wlist;
1291 	while (clist) {
1292 		tlen += clist->c_len;
1293 		clist = clist->c_next;
1294 	}
1295 
1296 	/*
1297 	 * set iov to addr+len of first segment of first wchunk of
1298 	 * wlist sent by client.  krecv() already malloc'd a buffer
1299 	 * large enough, but registration is deferred until we write
1300 	 * the buffer back to (NFS) client using RDMA_WRITE.
1301 	 */
1302 	iov->iov_base = (caddr_t)(uintptr_t)wlist->w.c_saddr;
1303 	iov->iov_len = tlen;
1304 
1305 	return (TRUE);
1306 }
1307 
1308 /*
1309  * routine to setup the read chunk lists
1310  */
1311 
1312 int
1313 rdma_setup_read_chunks(struct clist *wcl, uint32_t count, int *wcl_len)
1314 {
1315 	int		data_len, avail_len;
1316 	uint_t		round_len;
1317 
1318 	data_len = avail_len = 0;
1319 
1320 	while (wcl != NULL && count > 0) {
1321 		if (wcl->c_dmemhandle.mrc_rmr == 0)
1322 			break;
1323 
1324 		if (wcl->c_len < count) {
1325 			data_len += wcl->c_len;
1326 			avail_len = 0;
1327 		} else {
1328 			data_len += count;
1329 			avail_len = wcl->c_len - count;
1330 			wcl->c_len = count;
1331 		}
1332 		count -= wcl->c_len;
1333 
1334 		if (count == 0)
1335 			break;
1336 
1337 		wcl = wcl->c_next;
1338 	}
1339 
1340 	/*
1341 	 * MUST fail if there are still more data
1342 	 */
1343 	if (count > 0) {
1344 		DTRACE_PROBE2(krpc__e__rdma_setup_read_chunks_clist_len,
1345 		    int, data_len, int, count);
1346 		return (FALSE);
1347 	}
1348 
1349 	/*
1350 	 * Round up the last chunk to 4-byte boundary
1351 	 */
1352 	*wcl_len = roundup(data_len, BYTES_PER_XDR_UNIT);
1353 	round_len = *wcl_len - data_len;
1354 
1355 	if (round_len) {
1356 
1357 		/*
1358 		 * If there is space in the current chunk,
1359 		 * add the roundup to the chunk.
1360 		 */
1361 		if (avail_len >= round_len) {
1362 			wcl->c_len += round_len;
1363 		} else  {
1364 			/*
1365 			 * try the next one.
1366 			 */
1367 			wcl = wcl->c_next;
1368 			if ((wcl == NULL) || (wcl->c_len < round_len)) {
1369 				DTRACE_PROBE1(
1370 				    krpc__e__rdma_setup_read_chunks_rndup,
1371 				    int, round_len);
1372 				return (FALSE);
1373 			}
1374 			wcl->c_len = round_len;
1375 		}
1376 	}
1377 
1378 	wcl = wcl->c_next;
1379 
1380 	/*
1381 	 * Make rest of the chunks 0-len
1382 	 */
1383 
1384 	clist_zero_len(wcl);
1385 
1386 	return (TRUE);
1387 }
1388