xref: /titanic_41/usr/src/uts/common/rpc/svc_cots.c (revision f6da83d4178694e7113b71d1e452f15b296f73d8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  *  Copyright (c) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
26 /*	  All Rights Reserved  	*/
27 
28 /*
29  * Portions of this source code were derived from Berkeley 4.3 BSD
30  * under license from the Regents of the University of California.
31  */
32 
33 /*
34  * svc_cots.c
35  * Server side for connection-oriented RPC in the kernel.
36  *
37  */
38 
39 #include <sys/param.h>
40 #include <sys/types.h>
41 #include <sys/sysmacros.h>
42 #include <sys/file.h>
43 #include <sys/stream.h>
44 #include <sys/strsubr.h>
45 #include <sys/strsun.h>
46 #include <sys/stropts.h>
47 #include <sys/tiuser.h>
48 #include <sys/timod.h>
49 #include <sys/tihdr.h>
50 #include <sys/fcntl.h>
51 #include <sys/errno.h>
52 #include <sys/kmem.h>
53 #include <sys/systm.h>
54 #include <sys/debug.h>
55 #include <sys/cmn_err.h>
56 #include <sys/kstat.h>
57 #include <sys/vtrace.h>
58 
59 #include <rpc/types.h>
60 #include <rpc/xdr.h>
61 #include <rpc/auth.h>
62 #include <rpc/rpc_msg.h>
63 #include <rpc/svc.h>
64 #include <inet/ip.h>
65 
66 #define	COTS_MAX_ALLOCSIZE	2048
67 #define	MSG_OFFSET		128	/* offset of call into the mblk */
68 #define	RM_HDR_SIZE		4	/* record mark header size */
69 
70 /*
71  * Routines exported through ops vector.
72  */
73 static bool_t		svc_cots_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
74 static bool_t		svc_cots_ksend(SVCXPRT *, struct rpc_msg *);
75 static bool_t		svc_cots_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
76 static bool_t		svc_cots_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
77 static void		svc_cots_kdestroy(SVCMASTERXPRT *);
78 static int		svc_cots_kdup(struct svc_req *, caddr_t, int,
79 				struct dupreq **, bool_t *);
80 static void		svc_cots_kdupdone(struct dupreq *, caddr_t,
81 				void (*)(), int, int);
82 static int32_t		*svc_cots_kgetres(SVCXPRT *, int);
83 static void		svc_cots_kfreeres(SVCXPRT *);
84 static void		svc_cots_kclone_destroy(SVCXPRT *);
85 static void		svc_cots_kstart(SVCMASTERXPRT *);
86 static void		svc_cots_ktattrs(SVCXPRT *, int, void **);
87 
88 /*
89  * Server transport operations vector.
90  */
91 struct svc_ops svc_cots_op = {
92 	svc_cots_krecv,		/* Get requests */
93 	svc_cots_kgetargs,	/* Deserialize arguments */
94 	svc_cots_ksend,		/* Send reply */
95 	svc_cots_kfreeargs,	/* Free argument data space */
96 	svc_cots_kdestroy,	/* Destroy transport handle */
97 	svc_cots_kdup,		/* Check entry in dup req cache */
98 	svc_cots_kdupdone,	/* Mark entry in dup req cache as done */
99 	svc_cots_kgetres,	/* Get pointer to response buffer */
100 	svc_cots_kfreeres,	/* Destroy pre-serialized response header */
101 	svc_cots_kclone_destroy, /* Destroy a clone xprt */
102 	svc_cots_kstart,	/* Tell `ready-to-receive' to rpcmod */
103 	NULL,			/* Transport specific clone xprt */
104 	svc_cots_ktattrs	/* Transport Attributes */
105 };
106 
107 /*
108  * Master transport private data.
109  * Kept in xprt->xp_p2.
110  */
111 struct cots_master_data {
112 	char	*cmd_src_addr;	/* client's address */
113 	int	cmd_xprt_started; /* flag for clone routine to call */
114 				/* rpcmod's start routine. */
115 	struct rpc_cots_server *cmd_stats;	/* stats for zone */
116 };
117 
118 /*
119  * Transport private data.
120  * Kept in clone_xprt->xp_p2buf.
121  */
122 typedef struct cots_data {
123 	mblk_t	*cd_mp;		/* pre-allocated reply message */
124 	mblk_t	*cd_req_mp;	/* request message */
125 } cots_data_t;
126 
127 /*
128  * Server statistics
129  * NOTE: This structure type is duplicated in the NFS fast path.
130  */
131 static const struct rpc_cots_server {
132 	kstat_named_t	rscalls;
133 	kstat_named_t	rsbadcalls;
134 	kstat_named_t	rsnullrecv;
135 	kstat_named_t	rsbadlen;
136 	kstat_named_t	rsxdrcall;
137 	kstat_named_t	rsdupchecks;
138 	kstat_named_t	rsdupreqs;
139 } cots_rsstat_tmpl = {
140 	{ "calls",	KSTAT_DATA_UINT64 },
141 	{ "badcalls",	KSTAT_DATA_UINT64 },
142 	{ "nullrecv",	KSTAT_DATA_UINT64 },
143 	{ "badlen",	KSTAT_DATA_UINT64 },
144 	{ "xdrcall",	KSTAT_DATA_UINT64 },
145 	{ "dupchecks",	KSTAT_DATA_UINT64 },
146 	{ "dupreqs",	KSTAT_DATA_UINT64 }
147 };
148 
149 #define	CLONE2STATS(clone_xprt)	\
150 	((struct cots_master_data *)(clone_xprt)->xp_master->xp_p2)->cmd_stats
151 #define	RSSTAT_INCR(s, x)	\
152 	atomic_add_64(&(s)->x.value.ui64, 1)
153 
154 /*
155  * Pointer to a transport specific `ready to receive' function in rpcmod
156  * (set from rpcmod).
157  */
158 void    (*mir_start)(queue_t *);
159 uint_t	*svc_max_msg_sizep;
160 
161 /*
162  * the address size of the underlying transport can sometimes be
163  * unknown (tinfo->ADDR_size == -1).  For this case, it is
164  * necessary to figure out what the size is so the correct amount
165  * of data is allocated.  This is an itterative process:
166  *	1. take a good guess (use T_MINADDRSIZE)
167  *	2. try it.
168  *	3. if it works then everything is ok
169  *	4. if the error is ENAMETOLONG, double the guess
170  *	5. go back to step 2.
171  */
172 #define	T_UNKNOWNADDRSIZE	(-1)
173 #define	T_MINADDRSIZE	32
174 
175 /*
176  * Create a transport record.
177  * The transport record, output buffer, and private data structure
178  * are allocated.  The output buffer is serialized into using xdrmem.
179  * There is one transport record per user process which implements a
180  * set of services.
181  */
182 static kmutex_t cots_kcreate_lock;
183 
184 int
185 svc_cots_kcreate(file_t *fp, uint_t max_msgsize, struct T_info_ack *tinfo,
186     SVCMASTERXPRT **nxprt)
187 {
188 	struct cots_master_data *cmd;
189 	int err, retval;
190 	SVCMASTERXPRT *xprt;
191 	struct rpcstat *rpcstat;
192 	struct T_addr_ack *ack_p;
193 	struct strioctl getaddr;
194 
195 	if (nxprt == NULL)
196 		return (EINVAL);
197 
198 	rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone);
199 	ASSERT(rpcstat != NULL);
200 
201 	xprt = kmem_zalloc(sizeof (SVCMASTERXPRT), KM_SLEEP);
202 
203 	cmd = kmem_zalloc(sizeof (*cmd) + sizeof (*ack_p)
204 	    + (2 * sizeof (sin6_t)), KM_SLEEP);
205 
206 	ack_p = (struct T_addr_ack *)&cmd[1];
207 
208 	if ((tinfo->TIDU_size > COTS_MAX_ALLOCSIZE) ||
209 	    (tinfo->TIDU_size <= 0))
210 		xprt->xp_msg_size = COTS_MAX_ALLOCSIZE;
211 	else {
212 		xprt->xp_msg_size = tinfo->TIDU_size -
213 		    (tinfo->TIDU_size % BYTES_PER_XDR_UNIT);
214 	}
215 
216 	xprt->xp_ops = &svc_cots_op;
217 	xprt->xp_p2 = (caddr_t)cmd;
218 	cmd->cmd_xprt_started = 0;
219 	cmd->cmd_stats = rpcstat->rpc_cots_server;
220 
221 	getaddr.ic_cmd = TI_GETINFO;
222 	getaddr.ic_timout = -1;
223 	getaddr.ic_len = sizeof (*ack_p) + (2 * sizeof (sin6_t));
224 	getaddr.ic_dp = (char *)ack_p;
225 	ack_p->PRIM_type = T_ADDR_REQ;
226 
227 	err = strioctl(fp->f_vnode, I_STR, (intptr_t)&getaddr,
228 	    0, K_TO_K, CRED(), &retval);
229 	if (err) {
230 		kmem_free(cmd, sizeof (*cmd) + sizeof (*ack_p) +
231 		    (2 * sizeof (sin6_t)));
232 		kmem_free(xprt, sizeof (SVCMASTERXPRT));
233 		return (err);
234 	}
235 
236 	xprt->xp_rtaddr.maxlen = ack_p->REMADDR_length;
237 	xprt->xp_rtaddr.len = ack_p->REMADDR_length;
238 	cmd->cmd_src_addr = xprt->xp_rtaddr.buf =
239 	    (char *)ack_p + ack_p->REMADDR_offset;
240 
241 	xprt->xp_lcladdr.maxlen = ack_p->LOCADDR_length;
242 	xprt->xp_lcladdr.len = ack_p->LOCADDR_length;
243 	xprt->xp_lcladdr.buf = (char *)ack_p + ack_p->LOCADDR_offset;
244 
245 	/*
246 	 * If the current sanity check size in rpcmod is smaller
247 	 * than the size needed for this xprt, then increase
248 	 * the sanity check.
249 	 */
250 	if (max_msgsize != 0 && svc_max_msg_sizep &&
251 	    max_msgsize > *svc_max_msg_sizep) {
252 
253 		/* This check needs a lock */
254 		mutex_enter(&cots_kcreate_lock);
255 		if (svc_max_msg_sizep && max_msgsize > *svc_max_msg_sizep)
256 			*svc_max_msg_sizep = max_msgsize;
257 		mutex_exit(&cots_kcreate_lock);
258 	}
259 
260 	*nxprt = xprt;
261 
262 	return (0);
263 }
264 
265 /*
266  * Destroy a master transport record.
267  * Frees the space allocated for a transport record.
268  */
269 static void
270 svc_cots_kdestroy(SVCMASTERXPRT *xprt)
271 {
272 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
273 
274 	ASSERT(cmd);
275 
276 	if (xprt->xp_netid)
277 		kmem_free(xprt->xp_netid, strlen(xprt->xp_netid) + 1);
278 	if (xprt->xp_addrmask.maxlen)
279 		kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
280 
281 	mutex_destroy(&xprt->xp_req_lock);
282 	mutex_destroy(&xprt->xp_thread_lock);
283 
284 	kmem_free(cmd, sizeof (*cmd) + sizeof (struct T_addr_ack) +
285 	    (2 * sizeof (sin6_t)));
286 
287 	kmem_free(xprt, sizeof (SVCMASTERXPRT));
288 }
289 
290 /*
291  * svc_tli_kcreate() calls this function at the end to tell
292  * rpcmod that the transport is ready to receive requests.
293  */
294 static void
295 svc_cots_kstart(SVCMASTERXPRT *xprt)
296 {
297 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
298 
299 	if (cmd->cmd_xprt_started == 0) {
300 		/*
301 		 * Acquire the xp_req_lock in order to use xp_wq
302 		 * safely (we don't want to qenable a queue that has
303 		 * already been closed).
304 		 */
305 		mutex_enter(&xprt->xp_req_lock);
306 		if (cmd->cmd_xprt_started == 0 &&
307 		    xprt->xp_wq != NULL) {
308 			(*mir_start)(xprt->xp_wq);
309 			cmd->cmd_xprt_started = 1;
310 		}
311 		mutex_exit(&xprt->xp_req_lock);
312 	}
313 }
314 
315 /*
316  * Transport-type specific part of svc_xprt_cleanup().
317  */
318 static void
319 svc_cots_kclone_destroy(SVCXPRT *clone_xprt)
320 {
321 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
322 
323 	if (cd->cd_req_mp) {
324 		freemsg(cd->cd_req_mp);
325 		cd->cd_req_mp = (mblk_t *)0;
326 	}
327 	ASSERT(cd->cd_mp == NULL);
328 }
329 
330 /*
331  * Transport Attributes.
332  */
333 static void
334 svc_cots_ktattrs(SVCXPRT *clone_xprt, int attrflag, void **tattr)
335 {
336 	*tattr = NULL;
337 
338 	switch (attrflag) {
339 	case SVC_TATTR_ADDRMASK:
340 		*tattr = (void *)&clone_xprt->xp_master->xp_addrmask;
341 	}
342 }
343 
344 /*
345  * Receive rpc requests.
346  * Checks if the message is intact, and deserializes the call packet.
347  */
348 static bool_t
349 svc_cots_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
350 {
351 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
352 	XDR *xdrs = &clone_xprt->xp_xdrin;
353 	struct rpc_cots_server *stats = CLONE2STATS(clone_xprt);
354 
355 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KRECV_START,
356 	    "svc_cots_krecv_start:");
357 	RPCLOG(4, "svc_cots_krecv_start clone_xprt = %p:\n",
358 	    (void *)clone_xprt);
359 
360 	RSSTAT_INCR(stats, rscalls);
361 
362 	if (mp->b_datap->db_type != M_DATA) {
363 		RPCLOG(16, "svc_cots_krecv bad db_type %d\n",
364 		    mp->b_datap->db_type);
365 		goto bad;
366 	}
367 
368 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
369 
370 	TRACE_0(TR_FAC_KRPC, TR_XDR_CALLMSG_START,
371 	    "xdr_callmsg_start:");
372 	RPCLOG0(4, "xdr_callmsg_start:\n");
373 	if (!xdr_callmsg(xdrs, msg)) {
374 		TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
375 		    "xdr_callmsg_end:(%S)", "bad");
376 		RPCLOG0(1, "svc_cots_krecv xdr_callmsg failure\n");
377 		RSSTAT_INCR(stats, rsxdrcall);
378 		goto bad;
379 	}
380 	TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
381 	    "xdr_callmsg_end:(%S)", "good");
382 
383 	clone_xprt->xp_xid = msg->rm_xid;
384 	cd->cd_req_mp = mp;
385 
386 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
387 	    "svc_cots_krecv_end:(%S)", "good");
388 	RPCLOG0(4, "svc_cots_krecv_end:good\n");
389 	return (TRUE);
390 
391 bad:
392 	if (mp)
393 		freemsg(mp);
394 
395 	RSSTAT_INCR(stats, rsbadcalls);
396 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
397 	    "svc_cots_krecv_end:(%S)", "bad");
398 	return (FALSE);
399 }
400 
401 /*
402  * Send rpc reply.
403  */
404 static bool_t
405 svc_cots_ksend(SVCXPRT *clone_xprt, struct rpc_msg *msg)
406 {
407 	/* LINTED pointer alignment */
408 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
409 	XDR *xdrs = &(clone_xprt->xp_xdrout);
410 	int retval = FALSE;
411 	mblk_t *mp;
412 	xdrproc_t xdr_results;
413 	caddr_t xdr_location;
414 	bool_t has_args;
415 
416 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KSEND_START,
417 	    "svc_cots_ksend_start:");
418 
419 	/*
420 	 * If there is a result procedure specified in the reply message,
421 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
422 	 * We need to make sure it won't be processed twice, so we null
423 	 * it for xdr_replymsg here.
424 	 */
425 	has_args = FALSE;
426 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
427 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
428 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
429 			has_args = TRUE;
430 			xdr_location = msg->acpted_rply.ar_results.where;
431 			msg->acpted_rply.ar_results.proc = xdr_void;
432 			msg->acpted_rply.ar_results.where = NULL;
433 		}
434 	}
435 
436 	mp = cd->cd_mp;
437 	if (mp) {
438 		/*
439 		 * The program above pre-allocated an mblk and put
440 		 * the data in place.
441 		 */
442 		cd->cd_mp = (mblk_t *)NULL;
443 		if (!(xdr_replymsg_body(xdrs, msg) &&
444 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
445 		    xdr_results, xdr_location)))) {
446 			RPCLOG0(1, "svc_cots_ksend: "
447 			    "xdr_replymsg_body/SVCAUTH_WRAP failed\n");
448 			freemsg(mp);
449 			goto out;
450 		}
451 	} else {
452 		int	len;
453 		int	mpsize;
454 
455 		/*
456 		 * Leave space for protocol headers.
457 		 */
458 		len = MSG_OFFSET + clone_xprt->xp_msg_size;
459 
460 		/*
461 		 * Allocate an initial mblk for the response data.
462 		 */
463 		while (!(mp = allocb(len, BPRI_LO))) {
464 			RPCLOG0(16, "svc_cots_ksend: allocb failed failed\n");
465 			if (strwaitbuf(len, BPRI_LO)) {
466 				TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
467 				    "svc_cots_ksend_end:(%S)", "strwaitbuf");
468 				RPCLOG0(1,
469 				    "svc_cots_ksend: strwaitbuf failed\n");
470 				goto out;
471 			}
472 		}
473 
474 		/*
475 		 * Initialize the XDR decode stream.  Additional mblks
476 		 * will be allocated if necessary.  They will be TIDU
477 		 * sized.
478 		 */
479 		xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
480 		mpsize = MBLKSIZE(mp);
481 		ASSERT(mpsize >= len);
482 		ASSERT(mp->b_rptr == mp->b_datap->db_base);
483 
484 		/*
485 		 * If the size of mblk is not appreciably larger than what we
486 		 * asked, then resize the mblk to exactly len bytes. Reason for
487 		 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
488 		 * (from TCP over ethernet), and the arguments to RPC require
489 		 * 2800 bytes. Ideally we want the protocol to render two
490 		 * ~1400 byte segments over the wire. If allocb() gives us a 2k
491 		 * mblk, and we allocate a second mblk for the rest, the
492 		 * protocol module may generate 3 segments over the wire:
493 		 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
494 		 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
495 		 * the XDR encoding will generate two ~1400 byte mblks, and the
496 		 * protocol module is more likely to produce properly sized
497 		 * segments.
498 		 */
499 		if ((mpsize >> 1) <= len) {
500 			mp->b_rptr += (mpsize - len);
501 		}
502 
503 		/*
504 		 * Adjust b_rptr to reserve space for the non-data protocol
505 		 * headers that any downstream modules might like to add, and
506 		 * for the record marking header.
507 		 */
508 		mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
509 
510 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
511 		ASSERT(mp->b_wptr == mp->b_rptr);
512 
513 		msg->rm_xid = clone_xprt->xp_xid;
514 
515 		TRACE_0(TR_FAC_KRPC, TR_XDR_REPLYMSG_START,
516 		    "xdr_replymsg_start:");
517 		if (!(xdr_replymsg(xdrs, msg) &&
518 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
519 		    xdr_results, xdr_location)))) {
520 			TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
521 			    "xdr_replymsg_end:(%S)", "bad");
522 			freemsg(mp);
523 			RPCLOG0(1, "svc_cots_ksend: xdr_replymsg/SVCAUTH_WRAP "
524 			    "failed\n");
525 			goto out;
526 		}
527 		TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
528 		    "xdr_replymsg_end:(%S)", "good");
529 	}
530 
531 	put(clone_xprt->xp_wq, mp);
532 	retval = TRUE;
533 
534 out:
535 	/*
536 	 * This is completely disgusting.  If public is set it is
537 	 * a pointer to a structure whose first field is the address
538 	 * of the function to free that structure and any related
539 	 * stuff.  (see rrokfree in nfs_xdr.c).
540 	 */
541 	if (xdrs->x_public) {
542 		/* LINTED pointer alignment */
543 		(**((int (**)())xdrs->x_public))(xdrs->x_public);
544 	}
545 
546 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
547 	    "svc_cots_ksend_end:(%S)", "done");
548 	return (retval);
549 }
550 
551 /*
552  * Deserialize arguments.
553  */
554 static bool_t
555 svc_cots_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
556     caddr_t args_ptr)
557 {
558 	return (SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
559 	    xdr_args, args_ptr));
560 }
561 
562 static bool_t
563 svc_cots_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
564     caddr_t args_ptr)
565 {
566 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
567 	mblk_t *mp;
568 	bool_t retval;
569 
570 	/*
571 	 * It is important to call the XDR routine before
572 	 * freeing the request mblk.  Structures in the
573 	 * XDR data may point into the mblk and require that
574 	 * the memory be intact during the free routine.
575 	 */
576 	if (args_ptr) {
577 		/* LINTED pointer alignment */
578 		XDR	*xdrs = &clone_xprt->xp_xdrin;
579 
580 		xdrs->x_op = XDR_FREE;
581 		retval = (*xdr_args)(xdrs, args_ptr);
582 	} else
583 		retval = TRUE;
584 
585 	if ((mp = cd->cd_req_mp) != NULL) {
586 		cd->cd_req_mp = (mblk_t *)0;
587 		freemsg(mp);
588 	}
589 
590 	return (retval);
591 }
592 
593 static int32_t *
594 svc_cots_kgetres(SVCXPRT *clone_xprt, int size)
595 {
596 	/* LINTED pointer alignment */
597 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
598 	XDR *xdrs = &clone_xprt->xp_xdrout;
599 	mblk_t *mp;
600 	int32_t *buf;
601 	struct rpc_msg rply;
602 	int len;
603 	int mpsize;
604 
605 	/*
606 	 * Leave space for protocol headers.
607 	 */
608 	len = MSG_OFFSET + clone_xprt->xp_msg_size;
609 
610 	/*
611 	 * Allocate an initial mblk for the response data.
612 	 */
613 	while ((mp = allocb(len, BPRI_LO)) == NULL) {
614 		if (strwaitbuf(len, BPRI_LO))
615 			return (FALSE);
616 	}
617 
618 	/*
619 	 * Initialize the XDR decode stream.  Additional mblks
620 	 * will be allocated if necessary.  They will be TIDU
621 	 * sized.
622 	 */
623 	xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
624 	mpsize = MBLKSIZE(mp);
625 	ASSERT(mpsize >= len);
626 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
627 
628 	/*
629 	 * If the size of mblk is not appreciably larger than what we
630 	 * asked, then resize the mblk to exactly len bytes. Reason for
631 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
632 	 * (from TCP over ethernet), and the arguments to RPC require
633 	 * 2800 bytes. Ideally we want the protocol to render two
634 	 * ~1400 byte segments over the wire. If allocb() gives us a 2k
635 	 * mblk, and we allocate a second mblk for the rest, the
636 	 * protocol module may generate 3 segments over the wire:
637 	 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
638 	 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
639 	 * the XDR encoding will generate two ~1400 byte mblks, and the
640 	 * protocol module is more likely to produce properly sized
641 	 * segments.
642 	 */
643 	if ((mpsize >> 1) <= len) {
644 		mp->b_rptr += (mpsize - len);
645 	}
646 
647 	/*
648 	 * Adjust b_rptr to reserve space for the non-data protocol
649 	 * headers that any downstream modules might like to add, and
650 	 * for the record marking header.
651 	 */
652 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
653 
654 	XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
655 	ASSERT(mp->b_wptr == mp->b_rptr);
656 
657 	/*
658 	 * Assume a successful RPC since most of them are.
659 	 */
660 	rply.rm_xid = clone_xprt->xp_xid;
661 	rply.rm_direction = REPLY;
662 	rply.rm_reply.rp_stat = MSG_ACCEPTED;
663 	rply.acpted_rply.ar_verf = clone_xprt->xp_verf;
664 	rply.acpted_rply.ar_stat = SUCCESS;
665 
666 	if (!xdr_replymsg_hdr(xdrs, &rply)) {
667 		freeb(mp);
668 		return (NULL);
669 	}
670 
671 
672 	buf = XDR_INLINE(xdrs, size);
673 	if (buf == NULL) {
674 		ASSERT(cd->cd_mp == NULL);
675 		freemsg(mp);
676 	} else {
677 		cd->cd_mp = mp;
678 	}
679 	return (buf);
680 }
681 
682 static void
683 svc_cots_kfreeres(SVCXPRT *clone_xprt)
684 {
685 	cots_data_t *cd;
686 	mblk_t *mp;
687 
688 	cd = (cots_data_t *)clone_xprt->xp_p2buf;
689 	if ((mp = cd->cd_mp) != NULL) {
690 		cd->cd_mp = (mblk_t *)NULL;
691 		freemsg(mp);
692 	}
693 }
694 
695 /*
696  * the dup cacheing routines below provide a cache of non-failure
697  * transaction id's.  rpc service routines can use this to detect
698  * retransmissions and re-send a non-failure response.
699  */
700 
701 /*
702  * MAXDUPREQS is the number of cached items.  It should be adjusted
703  * to the service load so that there is likely to be a response entry
704  * when the first retransmission comes in.
705  */
706 #define	MAXDUPREQS	1024
707 
708 /*
709  * This should be appropriately scaled to MAXDUPREQS.
710  */
711 #define	DRHASHSZ	257
712 
713 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
714 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
715 #else
716 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
717 #endif
718 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
719 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
720 
721 static int	cotsndupreqs = 0;
722 int	cotsmaxdupreqs = MAXDUPREQS;
723 static kmutex_t cotsdupreq_lock;
724 static struct dupreq *cotsdrhashtbl[DRHASHSZ];
725 static int	cotsdrhashstat[DRHASHSZ];
726 
727 static void unhash(struct dupreq *);
728 
729 /*
730  * cotsdrmru points to the head of a circular linked list in lru order.
731  * cotsdrmru->dr_next == drlru
732  */
733 struct dupreq *cotsdrmru;
734 
735 /*
736  * PSARC 2003/523 Contract Private Interface
737  * svc_cots_kdup
738  * Changes must be reviewed by Solaris File Sharing
739  * Changes must be communicated to contract-2003-523@sun.com
740  *
741  * svc_cots_kdup searches the request cache and returns 0 if the
742  * request is not found in the cache.  If it is found, then it
743  * returns the state of the request (in progress or done) and
744  * the status or attributes that were part of the original reply.
745  *
746  * If DUP_DONE (there is a duplicate) svc_cots_kdup copies over the
747  * value of the response. In that case, also return in *dupcachedp
748  * whether the response free routine is cached in the dupreq - in which case
749  * the caller should not be freeing it, because it will be done later
750  * in the svc_cots_kdup code when the dupreq is reused.
751  */
752 static int
753 svc_cots_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
754 	bool_t *dupcachedp)
755 {
756 	struct rpc_cots_server *stats = CLONE2STATS(req->rq_xprt);
757 	struct dupreq *dr;
758 	uint32_t xid;
759 	uint32_t drhash;
760 	int status;
761 
762 	xid = REQTOXID(req);
763 	mutex_enter(&cotsdupreq_lock);
764 	RSSTAT_INCR(stats, rsdupchecks);
765 	/*
766 	 * Check to see whether an entry already exists in the cache.
767 	 */
768 	dr = cotsdrhashtbl[XIDHASH(xid)];
769 	while (dr != NULL) {
770 		if (dr->dr_xid == xid &&
771 		    dr->dr_proc == req->rq_proc &&
772 		    dr->dr_prog == req->rq_prog &&
773 		    dr->dr_vers == req->rq_vers &&
774 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
775 		    bcmp((caddr_t)dr->dr_addr.buf,
776 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
777 		    dr->dr_addr.len) == 0) {
778 			status = dr->dr_status;
779 			if (status == DUP_DONE) {
780 				bcopy(dr->dr_resp.buf, res, size);
781 				if (dupcachedp != NULL)
782 					*dupcachedp = (dr->dr_resfree != NULL);
783 				TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KDUP_DONE,
784 				    "svc_cots_kdup: DUP_DONE");
785 			} else {
786 				dr->dr_status = DUP_INPROGRESS;
787 				*drpp = dr;
788 				TRACE_0(TR_FAC_KRPC,
789 				    TR_SVC_COTS_KDUP_INPROGRESS,
790 				    "svc_cots_kdup: DUP_INPROGRESS");
791 			}
792 			RSSTAT_INCR(stats, rsdupreqs);
793 			mutex_exit(&cotsdupreq_lock);
794 			return (status);
795 		}
796 		dr = dr->dr_chain;
797 	}
798 
799 	/*
800 	 * There wasn't an entry, either allocate a new one or recycle
801 	 * an old one.
802 	 */
803 	if (cotsndupreqs < cotsmaxdupreqs) {
804 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
805 		if (dr == NULL) {
806 			mutex_exit(&cotsdupreq_lock);
807 			return (DUP_ERROR);
808 		}
809 		dr->dr_resp.buf = NULL;
810 		dr->dr_resp.maxlen = 0;
811 		dr->dr_addr.buf = NULL;
812 		dr->dr_addr.maxlen = 0;
813 		if (cotsdrmru) {
814 			dr->dr_next = cotsdrmru->dr_next;
815 			cotsdrmru->dr_next = dr;
816 		} else {
817 			dr->dr_next = dr;
818 		}
819 		cotsndupreqs++;
820 	} else {
821 		dr = cotsdrmru->dr_next;
822 		while (dr->dr_status == DUP_INPROGRESS) {
823 			dr = dr->dr_next;
824 			if (dr == cotsdrmru->dr_next) {
825 				cmn_err(CE_WARN, "svc_cots_kdup no slots free");
826 				mutex_exit(&cotsdupreq_lock);
827 				return (DUP_ERROR);
828 			}
829 		}
830 		unhash(dr);
831 		if (dr->dr_resfree) {
832 			(*dr->dr_resfree)(dr->dr_resp.buf);
833 		}
834 	}
835 	dr->dr_resfree = NULL;
836 	cotsdrmru = dr;
837 
838 	dr->dr_xid = REQTOXID(req);
839 	dr->dr_prog = req->rq_prog;
840 	dr->dr_vers = req->rq_vers;
841 	dr->dr_proc = req->rq_proc;
842 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
843 		if (dr->dr_addr.buf != NULL)
844 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
845 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
846 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
847 		if (dr->dr_addr.buf == NULL) {
848 			dr->dr_addr.maxlen = 0;
849 			dr->dr_status = DUP_DROP;
850 			mutex_exit(&cotsdupreq_lock);
851 			return (DUP_ERROR);
852 		}
853 	}
854 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
855 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
856 	if (dr->dr_resp.maxlen < size) {
857 		if (dr->dr_resp.buf != NULL)
858 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
859 		dr->dr_resp.maxlen = (unsigned int)size;
860 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
861 		if (dr->dr_resp.buf == NULL) {
862 			dr->dr_resp.maxlen = 0;
863 			dr->dr_status = DUP_DROP;
864 			mutex_exit(&cotsdupreq_lock);
865 			return (DUP_ERROR);
866 		}
867 	}
868 	dr->dr_status = DUP_INPROGRESS;
869 
870 	drhash = (uint32_t)DRHASH(dr);
871 	dr->dr_chain = cotsdrhashtbl[drhash];
872 	cotsdrhashtbl[drhash] = dr;
873 	cotsdrhashstat[drhash]++;
874 	mutex_exit(&cotsdupreq_lock);
875 	*drpp = dr;
876 	return (DUP_NEW);
877 }
878 
879 /*
880  * PSARC 2003/523 Contract Private Interface
881  * svc_cots_kdupdone
882  * Changes must be reviewed by Solaris File Sharing
883  * Changes must be communicated to contract-2003-523@sun.com
884  *
885  * svc_cots_kdupdone marks the request done (DUP_DONE or DUP_DROP)
886  * and stores the response.
887  */
888 static void
889 svc_cots_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
890 	int size, int status)
891 {
892 	ASSERT(dr->dr_resfree == NULL);
893 	if (status == DUP_DONE) {
894 		bcopy(res, dr->dr_resp.buf, size);
895 		dr->dr_resfree = dis_resfree;
896 	}
897 	dr->dr_status = status;
898 }
899 
900 /*
901  * This routine expects that the mutex, cotsdupreq_lock, is already held.
902  */
903 static void
904 unhash(struct dupreq *dr)
905 {
906 	struct dupreq *drt;
907 	struct dupreq *drtprev = NULL;
908 	uint32_t drhash;
909 
910 	ASSERT(MUTEX_HELD(&cotsdupreq_lock));
911 
912 	drhash = (uint32_t)DRHASH(dr);
913 	drt = cotsdrhashtbl[drhash];
914 	while (drt != NULL) {
915 		if (drt == dr) {
916 			cotsdrhashstat[drhash]--;
917 			if (drtprev == NULL) {
918 				cotsdrhashtbl[drhash] = drt->dr_chain;
919 			} else {
920 				drtprev->dr_chain = drt->dr_chain;
921 			}
922 			return;
923 		}
924 		drtprev = drt;
925 		drt = drt->dr_chain;
926 	}
927 }
928 
929 void
930 svc_cots_stats_init(zoneid_t zoneid, struct rpc_cots_server **statsp)
931 {
932 	*statsp = (struct rpc_cots_server *)rpcstat_zone_init_common(zoneid,
933 	    "unix", "rpc_cots_server", (const kstat_named_t *)&cots_rsstat_tmpl,
934 	    sizeof (cots_rsstat_tmpl));
935 }
936 
937 void
938 svc_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_server **statsp)
939 {
940 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_server");
941 	kmem_free(*statsp, sizeof (cots_rsstat_tmpl));
942 }
943 
944 void
945 svc_cots_init(void)
946 {
947 	/*
948 	 * Check to make sure that the cots private data will fit into
949 	 * the stack buffer allocated by svc_run.  The ASSERT is a safety
950 	 * net if the cots_data_t structure ever changes.
951 	 */
952 	/*CONSTANTCONDITION*/
953 	ASSERT(sizeof (cots_data_t) <= SVC_P2LEN);
954 
955 	mutex_init(&cots_kcreate_lock, NULL, MUTEX_DEFAULT, NULL);
956 	mutex_init(&cotsdupreq_lock, NULL, MUTEX_DEFAULT, NULL);
957 }
958