xref: /illumos-gate/usr/src/uts/common/rpc/svc_cots.c (revision 3d75a287b67b9092b2820a0d7c3be218f843066b)
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 /*
23  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
24  *  Copyright (c) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright (c) 2012 by Delphix. All rights reserved.
26  */
27 
28 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
29 /*	  All Rights Reserved  	*/
30 
31 /*
32  * Portions of this source code were derived from Berkeley 4.3 BSD
33  * under license from the Regents of the University of California.
34  */
35 
36 /*
37  * svc_cots.c
38  * Server side for connection-oriented RPC in the kernel.
39  *
40  */
41 
42 #include <sys/param.h>
43 #include <sys/types.h>
44 #include <sys/sysmacros.h>
45 #include <sys/file.h>
46 #include <sys/stream.h>
47 #include <sys/strsubr.h>
48 #include <sys/strsun.h>
49 #include <sys/stropts.h>
50 #include <sys/tiuser.h>
51 #include <sys/timod.h>
52 #include <sys/tihdr.h>
53 #include <sys/fcntl.h>
54 #include <sys/errno.h>
55 #include <sys/kmem.h>
56 #include <sys/systm.h>
57 #include <sys/debug.h>
58 #include <sys/cmn_err.h>
59 #include <sys/kstat.h>
60 #include <sys/vtrace.h>
61 
62 #include <rpc/types.h>
63 #include <rpc/xdr.h>
64 #include <rpc/auth.h>
65 #include <rpc/rpc_msg.h>
66 #include <rpc/svc.h>
67 #include <inet/ip.h>
68 
69 #define	COTS_MAX_ALLOCSIZE	2048
70 #define	MSG_OFFSET		128	/* offset of call into the mblk */
71 #define	RM_HDR_SIZE		4	/* record mark header size */
72 
73 /*
74  * Routines exported through ops vector.
75  */
76 static bool_t		svc_cots_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
77 static bool_t		svc_cots_ksend(SVCXPRT *, struct rpc_msg *);
78 static bool_t		svc_cots_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
79 static bool_t		svc_cots_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
80 static void		svc_cots_kdestroy(SVCMASTERXPRT *);
81 static int		svc_cots_kdup(struct svc_req *, caddr_t, int,
82 				struct dupreq **, bool_t *);
83 static void		svc_cots_kdupdone(struct dupreq *, caddr_t,
84 				void (*)(), int, int);
85 static int32_t		*svc_cots_kgetres(SVCXPRT *, int);
86 static void		svc_cots_kfreeres(SVCXPRT *);
87 static void		svc_cots_kclone_destroy(SVCXPRT *);
88 static void		svc_cots_kstart(SVCMASTERXPRT *);
89 static void		svc_cots_ktattrs(SVCXPRT *, int, void **);
90 
91 /*
92  * Server transport operations vector.
93  */
94 struct svc_ops svc_cots_op = {
95 	svc_cots_krecv,		/* Get requests */
96 	svc_cots_kgetargs,	/* Deserialize arguments */
97 	svc_cots_ksend,		/* Send reply */
98 	svc_cots_kfreeargs,	/* Free argument data space */
99 	svc_cots_kdestroy,	/* Destroy transport handle */
100 	svc_cots_kdup,		/* Check entry in dup req cache */
101 	svc_cots_kdupdone,	/* Mark entry in dup req cache as done */
102 	svc_cots_kgetres,	/* Get pointer to response buffer */
103 	svc_cots_kfreeres,	/* Destroy pre-serialized response header */
104 	svc_cots_kclone_destroy, /* Destroy a clone xprt */
105 	svc_cots_kstart,	/* Tell `ready-to-receive' to rpcmod */
106 	NULL,			/* Transport specific clone xprt */
107 	svc_cots_ktattrs	/* Transport Attributes */
108 };
109 
110 /*
111  * Master transport private data.
112  * Kept in xprt->xp_p2.
113  */
114 struct cots_master_data {
115 	char	*cmd_src_addr;	/* client's address */
116 	int	cmd_xprt_started; /* flag for clone routine to call */
117 				/* rpcmod's start routine. */
118 	struct rpc_cots_server *cmd_stats;	/* stats for zone */
119 };
120 
121 /*
122  * Transport private data.
123  * Kept in clone_xprt->xp_p2buf.
124  */
125 typedef struct cots_data {
126 	mblk_t	*cd_mp;		/* pre-allocated reply message */
127 	mblk_t	*cd_req_mp;	/* request message */
128 } cots_data_t;
129 
130 /*
131  * Server statistics
132  * NOTE: This structure type is duplicated in the NFS fast path.
133  */
134 static const struct rpc_cots_server {
135 	kstat_named_t	rscalls;
136 	kstat_named_t	rsbadcalls;
137 	kstat_named_t	rsnullrecv;
138 	kstat_named_t	rsbadlen;
139 	kstat_named_t	rsxdrcall;
140 	kstat_named_t	rsdupchecks;
141 	kstat_named_t	rsdupreqs;
142 } cots_rsstat_tmpl = {
143 	{ "calls",	KSTAT_DATA_UINT64 },
144 	{ "badcalls",	KSTAT_DATA_UINT64 },
145 	{ "nullrecv",	KSTAT_DATA_UINT64 },
146 	{ "badlen",	KSTAT_DATA_UINT64 },
147 	{ "xdrcall",	KSTAT_DATA_UINT64 },
148 	{ "dupchecks",	KSTAT_DATA_UINT64 },
149 	{ "dupreqs",	KSTAT_DATA_UINT64 }
150 };
151 
152 #define	CLONE2STATS(clone_xprt)	\
153 	((struct cots_master_data *)(clone_xprt)->xp_master->xp_p2)->cmd_stats
154 #define	RSSTAT_INCR(s, x)	\
155 	atomic_inc_64(&(s)->x.value.ui64)
156 
157 /*
158  * Pointer to a transport specific `ready to receive' function in rpcmod
159  * (set from rpcmod).
160  */
161 void    (*mir_start)(queue_t *);
162 uint_t	*svc_max_msg_sizep;
163 
164 /*
165  * the address size of the underlying transport can sometimes be
166  * unknown (tinfo->ADDR_size == -1).  For this case, it is
167  * necessary to figure out what the size is so the correct amount
168  * of data is allocated.  This is an itterative process:
169  *	1. take a good guess (use T_MINADDRSIZE)
170  *	2. try it.
171  *	3. if it works then everything is ok
172  *	4. if the error is ENAMETOLONG, double the guess
173  *	5. go back to step 2.
174  */
175 #define	T_UNKNOWNADDRSIZE	(-1)
176 #define	T_MINADDRSIZE	32
177 
178 /*
179  * Create a transport record.
180  * The transport record, output buffer, and private data structure
181  * are allocated.  The output buffer is serialized into using xdrmem.
182  * There is one transport record per user process which implements a
183  * set of services.
184  */
185 static kmutex_t cots_kcreate_lock;
186 
187 int
188 svc_cots_kcreate(file_t *fp, uint_t max_msgsize, struct T_info_ack *tinfo,
189     SVCMASTERXPRT **nxprt)
190 {
191 	struct cots_master_data *cmd;
192 	int err, retval;
193 	SVCMASTERXPRT *xprt;
194 	struct rpcstat *rpcstat;
195 	struct T_addr_ack *ack_p;
196 	struct strioctl getaddr;
197 
198 	if (nxprt == NULL)
199 		return (EINVAL);
200 
201 	rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone);
202 	ASSERT(rpcstat != NULL);
203 
204 	xprt = kmem_zalloc(sizeof (SVCMASTERXPRT), KM_SLEEP);
205 
206 	cmd = kmem_zalloc(sizeof (*cmd) + sizeof (*ack_p)
207 	    + (2 * sizeof (sin6_t)), KM_SLEEP);
208 
209 	ack_p = (struct T_addr_ack *)&cmd[1];
210 
211 	if ((tinfo->TIDU_size > COTS_MAX_ALLOCSIZE) ||
212 	    (tinfo->TIDU_size <= 0))
213 		xprt->xp_msg_size = COTS_MAX_ALLOCSIZE;
214 	else {
215 		xprt->xp_msg_size = tinfo->TIDU_size -
216 		    (tinfo->TIDU_size % BYTES_PER_XDR_UNIT);
217 	}
218 
219 	xprt->xp_ops = &svc_cots_op;
220 	xprt->xp_p2 = (caddr_t)cmd;
221 	cmd->cmd_xprt_started = 0;
222 	cmd->cmd_stats = rpcstat->rpc_cots_server;
223 
224 	getaddr.ic_cmd = TI_GETINFO;
225 	getaddr.ic_timout = -1;
226 	getaddr.ic_len = sizeof (*ack_p) + (2 * sizeof (sin6_t));
227 	getaddr.ic_dp = (char *)ack_p;
228 	ack_p->PRIM_type = T_ADDR_REQ;
229 
230 	err = strioctl(fp->f_vnode, I_STR, (intptr_t)&getaddr,
231 	    0, K_TO_K, CRED(), &retval);
232 	if (err) {
233 		kmem_free(cmd, sizeof (*cmd) + sizeof (*ack_p) +
234 		    (2 * sizeof (sin6_t)));
235 		kmem_free(xprt, sizeof (SVCMASTERXPRT));
236 		return (err);
237 	}
238 
239 	xprt->xp_rtaddr.maxlen = ack_p->REMADDR_length;
240 	xprt->xp_rtaddr.len = ack_p->REMADDR_length;
241 	cmd->cmd_src_addr = xprt->xp_rtaddr.buf =
242 	    (char *)ack_p + ack_p->REMADDR_offset;
243 
244 	xprt->xp_lcladdr.maxlen = ack_p->LOCADDR_length;
245 	xprt->xp_lcladdr.len = ack_p->LOCADDR_length;
246 	xprt->xp_lcladdr.buf = (char *)ack_p + ack_p->LOCADDR_offset;
247 
248 	/*
249 	 * If the current sanity check size in rpcmod is smaller
250 	 * than the size needed for this xprt, then increase
251 	 * the sanity check.
252 	 */
253 	if (max_msgsize != 0 && svc_max_msg_sizep &&
254 	    max_msgsize > *svc_max_msg_sizep) {
255 
256 		/* This check needs a lock */
257 		mutex_enter(&cots_kcreate_lock);
258 		if (svc_max_msg_sizep && max_msgsize > *svc_max_msg_sizep)
259 			*svc_max_msg_sizep = max_msgsize;
260 		mutex_exit(&cots_kcreate_lock);
261 	}
262 
263 	*nxprt = xprt;
264 
265 	return (0);
266 }
267 
268 /*
269  * Destroy a master transport record.
270  * Frees the space allocated for a transport record.
271  */
272 static void
273 svc_cots_kdestroy(SVCMASTERXPRT *xprt)
274 {
275 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
276 
277 	ASSERT(cmd);
278 
279 	if (xprt->xp_netid)
280 		kmem_free(xprt->xp_netid, strlen(xprt->xp_netid) + 1);
281 	if (xprt->xp_addrmask.maxlen)
282 		kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
283 
284 	mutex_destroy(&xprt->xp_req_lock);
285 	mutex_destroy(&xprt->xp_thread_lock);
286 
287 	kmem_free(cmd, sizeof (*cmd) + sizeof (struct T_addr_ack) +
288 	    (2 * sizeof (sin6_t)));
289 
290 	kmem_free(xprt, sizeof (SVCMASTERXPRT));
291 }
292 
293 /*
294  * svc_tli_kcreate() calls this function at the end to tell
295  * rpcmod that the transport is ready to receive requests.
296  */
297 static void
298 svc_cots_kstart(SVCMASTERXPRT *xprt)
299 {
300 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
301 
302 	if (cmd->cmd_xprt_started == 0) {
303 		/*
304 		 * Acquire the xp_req_lock in order to use xp_wq
305 		 * safely (we don't want to qenable a queue that has
306 		 * already been closed).
307 		 */
308 		mutex_enter(&xprt->xp_req_lock);
309 		if (cmd->cmd_xprt_started == 0 &&
310 		    xprt->xp_wq != NULL) {
311 			(*mir_start)(xprt->xp_wq);
312 			cmd->cmd_xprt_started = 1;
313 		}
314 		mutex_exit(&xprt->xp_req_lock);
315 	}
316 }
317 
318 /*
319  * Transport-type specific part of svc_xprt_cleanup().
320  */
321 static void
322 svc_cots_kclone_destroy(SVCXPRT *clone_xprt)
323 {
324 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
325 
326 	if (cd->cd_req_mp) {
327 		freemsg(cd->cd_req_mp);
328 		cd->cd_req_mp = (mblk_t *)0;
329 	}
330 	ASSERT(cd->cd_mp == NULL);
331 }
332 
333 /*
334  * Transport Attributes.
335  */
336 static void
337 svc_cots_ktattrs(SVCXPRT *clone_xprt, int attrflag, void **tattr)
338 {
339 	*tattr = NULL;
340 
341 	switch (attrflag) {
342 	case SVC_TATTR_ADDRMASK:
343 		*tattr = (void *)&clone_xprt->xp_master->xp_addrmask;
344 	}
345 }
346 
347 /*
348  * Receive rpc requests.
349  * Checks if the message is intact, and deserializes the call packet.
350  */
351 static bool_t
352 svc_cots_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
353 {
354 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
355 	XDR *xdrs = &clone_xprt->xp_xdrin;
356 	struct rpc_cots_server *stats = CLONE2STATS(clone_xprt);
357 
358 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KRECV_START,
359 	    "svc_cots_krecv_start:");
360 	RPCLOG(4, "svc_cots_krecv_start clone_xprt = %p:\n",
361 	    (void *)clone_xprt);
362 
363 	RSSTAT_INCR(stats, rscalls);
364 
365 	if (mp->b_datap->db_type != M_DATA) {
366 		RPCLOG(16, "svc_cots_krecv bad db_type %d\n",
367 		    mp->b_datap->db_type);
368 		goto bad;
369 	}
370 
371 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
372 
373 	TRACE_0(TR_FAC_KRPC, TR_XDR_CALLMSG_START,
374 	    "xdr_callmsg_start:");
375 	RPCLOG0(4, "xdr_callmsg_start:\n");
376 	if (!xdr_callmsg(xdrs, msg)) {
377 		XDR_DESTROY(xdrs);
378 		TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
379 		    "xdr_callmsg_end:(%S)", "bad");
380 		RPCLOG0(1, "svc_cots_krecv xdr_callmsg failure\n");
381 		RSSTAT_INCR(stats, rsxdrcall);
382 		goto bad;
383 	}
384 	TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
385 	    "xdr_callmsg_end:(%S)", "good");
386 
387 	clone_xprt->xp_xid = msg->rm_xid;
388 	cd->cd_req_mp = mp;
389 
390 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
391 	    "svc_cots_krecv_end:(%S)", "good");
392 	RPCLOG0(4, "svc_cots_krecv_end:good\n");
393 	return (TRUE);
394 
395 bad:
396 	if (mp)
397 		freemsg(mp);
398 
399 	RSSTAT_INCR(stats, rsbadcalls);
400 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
401 	    "svc_cots_krecv_end:(%S)", "bad");
402 	return (FALSE);
403 }
404 
405 /*
406  * Send rpc reply.
407  */
408 static bool_t
409 svc_cots_ksend(SVCXPRT *clone_xprt, struct rpc_msg *msg)
410 {
411 	/* LINTED pointer alignment */
412 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
413 	XDR *xdrs = &(clone_xprt->xp_xdrout);
414 	int retval = FALSE;
415 	mblk_t *mp;
416 	xdrproc_t xdr_results;
417 	caddr_t xdr_location;
418 	bool_t has_args;
419 
420 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KSEND_START,
421 	    "svc_cots_ksend_start:");
422 
423 	/*
424 	 * If there is a result procedure specified in the reply message,
425 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
426 	 * We need to make sure it won't be processed twice, so we null
427 	 * it for xdr_replymsg here.
428 	 */
429 	has_args = FALSE;
430 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
431 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
432 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
433 			has_args = TRUE;
434 			xdr_location = msg->acpted_rply.ar_results.where;
435 			msg->acpted_rply.ar_results.proc = xdr_void;
436 			msg->acpted_rply.ar_results.where = NULL;
437 		}
438 	}
439 
440 	mp = cd->cd_mp;
441 	if (mp) {
442 		/*
443 		 * The program above pre-allocated an mblk and put
444 		 * the data in place.
445 		 */
446 		cd->cd_mp = (mblk_t *)NULL;
447 		if (!(xdr_replymsg_body(xdrs, msg) &&
448 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
449 		    xdr_results, xdr_location)))) {
450 			XDR_DESTROY(xdrs);
451 			RPCLOG0(1, "svc_cots_ksend: "
452 			    "xdr_replymsg_body/SVCAUTH_WRAP failed\n");
453 			freemsg(mp);
454 			goto out;
455 		}
456 	} else {
457 		int	len;
458 		int	mpsize;
459 
460 		/*
461 		 * Leave space for protocol headers.
462 		 */
463 		len = MSG_OFFSET + clone_xprt->xp_msg_size;
464 
465 		/*
466 		 * Allocate an initial mblk for the response data.
467 		 */
468 		while (!(mp = allocb(len, BPRI_LO))) {
469 			RPCLOG0(16, "svc_cots_ksend: allocb failed failed\n");
470 			if (strwaitbuf(len, BPRI_LO)) {
471 				TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
472 				    "svc_cots_ksend_end:(%S)", "strwaitbuf");
473 				RPCLOG0(1,
474 				    "svc_cots_ksend: strwaitbuf failed\n");
475 				goto out;
476 			}
477 		}
478 
479 		/*
480 		 * Initialize the XDR encode stream.  Additional mblks
481 		 * will be allocated if necessary.  They will be TIDU
482 		 * sized.
483 		 */
484 		xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
485 		mpsize = MBLKSIZE(mp);
486 		ASSERT(mpsize >= len);
487 		ASSERT(mp->b_rptr == mp->b_datap->db_base);
488 
489 		/*
490 		 * If the size of mblk is not appreciably larger than what we
491 		 * asked, then resize the mblk to exactly len bytes. Reason for
492 		 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
493 		 * (from TCP over ethernet), and the arguments to RPC require
494 		 * 2800 bytes. Ideally we want the protocol to render two
495 		 * ~1400 byte segments over the wire. If allocb() gives us a 2k
496 		 * mblk, and we allocate a second mblk for the rest, the
497 		 * protocol module may generate 3 segments over the wire:
498 		 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
499 		 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
500 		 * the XDR encoding will generate two ~1400 byte mblks, and the
501 		 * protocol module is more likely to produce properly sized
502 		 * segments.
503 		 */
504 		if ((mpsize >> 1) <= len) {
505 			mp->b_rptr += (mpsize - len);
506 		}
507 
508 		/*
509 		 * Adjust b_rptr to reserve space for the non-data protocol
510 		 * headers that any downstream modules might like to add, and
511 		 * for the record marking header.
512 		 */
513 		mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
514 
515 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
516 		ASSERT(mp->b_wptr == mp->b_rptr);
517 
518 		msg->rm_xid = clone_xprt->xp_xid;
519 
520 		TRACE_0(TR_FAC_KRPC, TR_XDR_REPLYMSG_START,
521 		    "xdr_replymsg_start:");
522 		if (!(xdr_replymsg(xdrs, msg) &&
523 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
524 		    xdr_results, xdr_location)))) {
525 			XDR_DESTROY(xdrs);
526 			TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
527 			    "xdr_replymsg_end:(%S)", "bad");
528 			freemsg(mp);
529 			RPCLOG0(1, "svc_cots_ksend: xdr_replymsg/SVCAUTH_WRAP "
530 			    "failed\n");
531 			goto out;
532 		}
533 		TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
534 		    "xdr_replymsg_end:(%S)", "good");
535 	}
536 
537 	XDR_DESTROY(xdrs);
538 
539 	put(clone_xprt->xp_wq, mp);
540 	retval = TRUE;
541 
542 out:
543 	/*
544 	 * This is completely disgusting.  If public is set it is
545 	 * a pointer to a structure whose first field is the address
546 	 * of the function to free that structure and any related
547 	 * stuff.  (see rrokfree in nfs_xdr.c).
548 	 */
549 	if (xdrs->x_public) {
550 		/* LINTED pointer alignment */
551 		(**((int (**)())xdrs->x_public))(xdrs->x_public);
552 	}
553 
554 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
555 	    "svc_cots_ksend_end:(%S)", "done");
556 	return (retval);
557 }
558 
559 /*
560  * Deserialize arguments.
561  */
562 static bool_t
563 svc_cots_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
564     caddr_t args_ptr)
565 {
566 	return (SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
567 	    xdr_args, args_ptr));
568 }
569 
570 static bool_t
571 svc_cots_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
572     caddr_t args_ptr)
573 {
574 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
575 	/* LINTED pointer alignment */
576 	XDR *xdrs = &clone_xprt->xp_xdrin;
577 	mblk_t *mp;
578 	bool_t retval;
579 
580 	/*
581 	 * It is important to call the XDR routine before
582 	 * freeing the request mblk.  Structures in the
583 	 * XDR data may point into the mblk and require that
584 	 * the memory be intact during the free routine.
585 	 */
586 	if (args_ptr) {
587 		xdrs->x_op = XDR_FREE;
588 		retval = (*xdr_args)(xdrs, args_ptr);
589 	} else
590 		retval = TRUE;
591 
592 	XDR_DESTROY(xdrs);
593 
594 	if ((mp = cd->cd_req_mp) != NULL) {
595 		cd->cd_req_mp = (mblk_t *)0;
596 		freemsg(mp);
597 	}
598 
599 	return (retval);
600 }
601 
602 static int32_t *
603 svc_cots_kgetres(SVCXPRT *clone_xprt, int size)
604 {
605 	/* LINTED pointer alignment */
606 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
607 	XDR *xdrs = &clone_xprt->xp_xdrout;
608 	mblk_t *mp;
609 	int32_t *buf;
610 	struct rpc_msg rply;
611 	int len;
612 	int mpsize;
613 
614 	/*
615 	 * Leave space for protocol headers.
616 	 */
617 	len = MSG_OFFSET + clone_xprt->xp_msg_size;
618 
619 	/*
620 	 * Allocate an initial mblk for the response data.
621 	 */
622 	while ((mp = allocb(len, BPRI_LO)) == NULL) {
623 		if (strwaitbuf(len, BPRI_LO))
624 			return (NULL);
625 	}
626 
627 	/*
628 	 * Initialize the XDR encode stream.  Additional mblks
629 	 * will be allocated if necessary.  They will be TIDU
630 	 * sized.
631 	 */
632 	xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
633 	mpsize = MBLKSIZE(mp);
634 	ASSERT(mpsize >= len);
635 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
636 
637 	/*
638 	 * If the size of mblk is not appreciably larger than what we
639 	 * asked, then resize the mblk to exactly len bytes. Reason for
640 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
641 	 * (from TCP over ethernet), and the arguments to RPC require
642 	 * 2800 bytes. Ideally we want the protocol to render two
643 	 * ~1400 byte segments over the wire. If allocb() gives us a 2k
644 	 * mblk, and we allocate a second mblk for the rest, the
645 	 * protocol module may generate 3 segments over the wire:
646 	 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
647 	 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
648 	 * the XDR encoding will generate two ~1400 byte mblks, and the
649 	 * protocol module is more likely to produce properly sized
650 	 * segments.
651 	 */
652 	if ((mpsize >> 1) <= len) {
653 		mp->b_rptr += (mpsize - len);
654 	}
655 
656 	/*
657 	 * Adjust b_rptr to reserve space for the non-data protocol
658 	 * headers that any downstream modules might like to add, and
659 	 * for the record marking header.
660 	 */
661 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
662 
663 	XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
664 	ASSERT(mp->b_wptr == mp->b_rptr);
665 
666 	/*
667 	 * Assume a successful RPC since most of them are.
668 	 */
669 	rply.rm_xid = clone_xprt->xp_xid;
670 	rply.rm_direction = REPLY;
671 	rply.rm_reply.rp_stat = MSG_ACCEPTED;
672 	rply.acpted_rply.ar_verf = clone_xprt->xp_verf;
673 	rply.acpted_rply.ar_stat = SUCCESS;
674 
675 	if (!xdr_replymsg_hdr(xdrs, &rply)) {
676 		XDR_DESTROY(xdrs);
677 		freeb(mp);
678 		return (NULL);
679 	}
680 
681 	buf = XDR_INLINE(xdrs, size);
682 	if (buf == NULL) {
683 		XDR_DESTROY(xdrs);
684 		ASSERT(cd->cd_mp == NULL);
685 		freemsg(mp);
686 	} else {
687 		cd->cd_mp = mp;
688 	}
689 	return (buf);
690 }
691 
692 static void
693 svc_cots_kfreeres(SVCXPRT *clone_xprt)
694 {
695 	cots_data_t *cd;
696 	mblk_t *mp;
697 
698 	cd = (cots_data_t *)clone_xprt->xp_p2buf;
699 	if ((mp = cd->cd_mp) != NULL) {
700 		XDR_DESTROY(&clone_xprt->xp_xdrout);
701 		cd->cd_mp = (mblk_t *)NULL;
702 		freemsg(mp);
703 	}
704 }
705 
706 /*
707  * the dup cacheing routines below provide a cache of non-failure
708  * transaction id's.  rpc service routines can use this to detect
709  * retransmissions and re-send a non-failure response.
710  */
711 
712 /*
713  * MAXDUPREQS is the number of cached items.  It should be adjusted
714  * to the service load so that there is likely to be a response entry
715  * when the first retransmission comes in.
716  */
717 #define	MAXDUPREQS	8192
718 
719 /*
720  * This should be appropriately scaled to MAXDUPREQS.  To produce as less as
721  * possible collisions it is suggested to set this to a prime.
722  */
723 #define	DRHASHSZ	2053
724 
725 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
726 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
727 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
728 
729 static int	cotsndupreqs = 0;
730 int	cotsmaxdupreqs = MAXDUPREQS;
731 static kmutex_t cotsdupreq_lock;
732 static struct dupreq *cotsdrhashtbl[DRHASHSZ];
733 static int	cotsdrhashstat[DRHASHSZ];
734 
735 static void unhash(struct dupreq *);
736 
737 /*
738  * cotsdrmru points to the head of a circular linked list in lru order.
739  * cotsdrmru->dr_next == drlru
740  */
741 struct dupreq *cotsdrmru;
742 
743 /*
744  * PSARC 2003/523 Contract Private Interface
745  * svc_cots_kdup
746  * Changes must be reviewed by Solaris File Sharing
747  * Changes must be communicated to contract-2003-523@sun.com
748  *
749  * svc_cots_kdup searches the request cache and returns 0 if the
750  * request is not found in the cache.  If it is found, then it
751  * returns the state of the request (in progress or done) and
752  * the status or attributes that were part of the original reply.
753  *
754  * If DUP_DONE (there is a duplicate) svc_cots_kdup copies over the
755  * value of the response. In that case, also return in *dupcachedp
756  * whether the response free routine is cached in the dupreq - in which case
757  * the caller should not be freeing it, because it will be done later
758  * in the svc_cots_kdup code when the dupreq is reused.
759  */
760 static int
761 svc_cots_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
762     bool_t *dupcachedp)
763 {
764 	struct rpc_cots_server *stats = CLONE2STATS(req->rq_xprt);
765 	struct dupreq *dr;
766 	uint32_t xid;
767 	uint32_t drhash;
768 	int status;
769 
770 	xid = REQTOXID(req);
771 	mutex_enter(&cotsdupreq_lock);
772 	RSSTAT_INCR(stats, rsdupchecks);
773 	/*
774 	 * Check to see whether an entry already exists in the cache.
775 	 */
776 	dr = cotsdrhashtbl[XIDHASH(xid)];
777 	while (dr != NULL) {
778 		if (dr->dr_xid == xid &&
779 		    dr->dr_proc == req->rq_proc &&
780 		    dr->dr_prog == req->rq_prog &&
781 		    dr->dr_vers == req->rq_vers &&
782 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
783 		    bcmp((caddr_t)dr->dr_addr.buf,
784 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
785 		    dr->dr_addr.len) == 0) {
786 			status = dr->dr_status;
787 			if (status == DUP_DONE) {
788 				bcopy(dr->dr_resp.buf, res, size);
789 				if (dupcachedp != NULL)
790 					*dupcachedp = (dr->dr_resfree != NULL);
791 				TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KDUP_DONE,
792 				    "svc_cots_kdup: DUP_DONE");
793 			} else {
794 				dr->dr_status = DUP_INPROGRESS;
795 				*drpp = dr;
796 				TRACE_0(TR_FAC_KRPC,
797 				    TR_SVC_COTS_KDUP_INPROGRESS,
798 				    "svc_cots_kdup: DUP_INPROGRESS");
799 			}
800 			RSSTAT_INCR(stats, rsdupreqs);
801 			mutex_exit(&cotsdupreq_lock);
802 			return (status);
803 		}
804 		dr = dr->dr_chain;
805 	}
806 
807 	/*
808 	 * There wasn't an entry, either allocate a new one or recycle
809 	 * an old one.
810 	 */
811 	if (cotsndupreqs < cotsmaxdupreqs) {
812 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
813 		if (dr == NULL) {
814 			mutex_exit(&cotsdupreq_lock);
815 			return (DUP_ERROR);
816 		}
817 		dr->dr_resp.buf = NULL;
818 		dr->dr_resp.maxlen = 0;
819 		dr->dr_addr.buf = NULL;
820 		dr->dr_addr.maxlen = 0;
821 		if (cotsdrmru) {
822 			dr->dr_next = cotsdrmru->dr_next;
823 			cotsdrmru->dr_next = dr;
824 		} else {
825 			dr->dr_next = dr;
826 		}
827 		cotsndupreqs++;
828 	} else {
829 		dr = cotsdrmru->dr_next;
830 		while (dr->dr_status == DUP_INPROGRESS) {
831 			dr = dr->dr_next;
832 			if (dr == cotsdrmru->dr_next) {
833 				cmn_err(CE_WARN, "svc_cots_kdup no slots free");
834 				mutex_exit(&cotsdupreq_lock);
835 				return (DUP_ERROR);
836 			}
837 		}
838 		unhash(dr);
839 		if (dr->dr_resfree) {
840 			(*dr->dr_resfree)(dr->dr_resp.buf);
841 		}
842 	}
843 	dr->dr_resfree = NULL;
844 	cotsdrmru = dr;
845 
846 	dr->dr_xid = REQTOXID(req);
847 	dr->dr_prog = req->rq_prog;
848 	dr->dr_vers = req->rq_vers;
849 	dr->dr_proc = req->rq_proc;
850 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
851 		if (dr->dr_addr.buf != NULL)
852 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
853 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
854 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
855 		if (dr->dr_addr.buf == NULL) {
856 			dr->dr_addr.maxlen = 0;
857 			dr->dr_status = DUP_DROP;
858 			mutex_exit(&cotsdupreq_lock);
859 			return (DUP_ERROR);
860 		}
861 	}
862 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
863 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
864 	if (dr->dr_resp.maxlen < size) {
865 		if (dr->dr_resp.buf != NULL)
866 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
867 		dr->dr_resp.maxlen = (unsigned int)size;
868 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
869 		if (dr->dr_resp.buf == NULL) {
870 			dr->dr_resp.maxlen = 0;
871 			dr->dr_status = DUP_DROP;
872 			mutex_exit(&cotsdupreq_lock);
873 			return (DUP_ERROR);
874 		}
875 	}
876 	dr->dr_status = DUP_INPROGRESS;
877 
878 	drhash = (uint32_t)DRHASH(dr);
879 	dr->dr_chain = cotsdrhashtbl[drhash];
880 	cotsdrhashtbl[drhash] = dr;
881 	cotsdrhashstat[drhash]++;
882 	mutex_exit(&cotsdupreq_lock);
883 	*drpp = dr;
884 	return (DUP_NEW);
885 }
886 
887 /*
888  * PSARC 2003/523 Contract Private Interface
889  * svc_cots_kdupdone
890  * Changes must be reviewed by Solaris File Sharing
891  * Changes must be communicated to contract-2003-523@sun.com
892  *
893  * svc_cots_kdupdone marks the request done (DUP_DONE or DUP_DROP)
894  * and stores the response.
895  */
896 static void
897 svc_cots_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
898     int size, int status)
899 {
900 	ASSERT(dr->dr_resfree == NULL);
901 	if (status == DUP_DONE) {
902 		bcopy(res, dr->dr_resp.buf, size);
903 		dr->dr_resfree = dis_resfree;
904 	}
905 	dr->dr_status = status;
906 }
907 
908 /*
909  * This routine expects that the mutex, cotsdupreq_lock, is already held.
910  */
911 static void
912 unhash(struct dupreq *dr)
913 {
914 	struct dupreq *drt;
915 	struct dupreq *drtprev = NULL;
916 	uint32_t drhash;
917 
918 	ASSERT(MUTEX_HELD(&cotsdupreq_lock));
919 
920 	drhash = (uint32_t)DRHASH(dr);
921 	drt = cotsdrhashtbl[drhash];
922 	while (drt != NULL) {
923 		if (drt == dr) {
924 			cotsdrhashstat[drhash]--;
925 			if (drtprev == NULL) {
926 				cotsdrhashtbl[drhash] = drt->dr_chain;
927 			} else {
928 				drtprev->dr_chain = drt->dr_chain;
929 			}
930 			return;
931 		}
932 		drtprev = drt;
933 		drt = drt->dr_chain;
934 	}
935 }
936 
937 void
938 svc_cots_stats_init(zoneid_t zoneid, struct rpc_cots_server **statsp)
939 {
940 	*statsp = (struct rpc_cots_server *)rpcstat_zone_init_common(zoneid,
941 	    "unix", "rpc_cots_server", (const kstat_named_t *)&cots_rsstat_tmpl,
942 	    sizeof (cots_rsstat_tmpl));
943 }
944 
945 void
946 svc_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_server **statsp)
947 {
948 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_server");
949 	kmem_free(*statsp, sizeof (cots_rsstat_tmpl));
950 }
951 
952 void
953 svc_cots_init(void)
954 {
955 	/*
956 	 * Check to make sure that the cots private data will fit into
957 	 * the stack buffer allocated by svc_run.  The ASSERT is a safety
958 	 * net if the cots_data_t structure ever changes.
959 	 */
960 	/*CONSTANTCONDITION*/
961 	ASSERT(sizeof (cots_data_t) <= SVC_P2LEN);
962 
963 	mutex_init(&cots_kcreate_lock, NULL, MUTEX_DEFAULT, NULL);
964 	mutex_init(&cotsdupreq_lock, NULL, MUTEX_DEFAULT, NULL);
965 }
966