xref: /illumos-gate/usr/src/uts/common/rpc/svc_cots.c (revision 3dbfc80346c4b24f1337e411111b9521c729cf9e)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  *  Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * Portions of this source code were derived from Berkeley 4.3 BSD
31  * under license from the Regents of the University of California.
32  */
33 
34 #pragma ident	"%Z%%M%	%I%	%E% SMI"
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 
90 /*
91  * Server transport operations vector.
92  */
93 struct svc_ops svc_cots_op = {
94 	svc_cots_krecv,		/* Get requests */
95 	svc_cots_kgetargs,	/* Deserialize arguments */
96 	svc_cots_ksend,		/* Send reply */
97 	svc_cots_kfreeargs,	/* Free argument data space */
98 	svc_cots_kdestroy,	/* Destroy transport handle */
99 	svc_cots_kdup,		/* Check entry in dup req cache */
100 	svc_cots_kdupdone,	/* Mark entry in dup req cache as done */
101 	svc_cots_kgetres,	/* Get pointer to response buffer */
102 	svc_cots_kfreeres,	/* Destroy pre-serialized response header */
103 	svc_cots_kclone_destroy, /* Destroy a clone xprt */
104 	svc_cots_kstart		/* Tell `ready-to-receive' to rpcmod */
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  * Receive rpc requests.
332  * Checks if the message is intact, and deserializes the call packet.
333  */
334 static bool_t
335 svc_cots_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
336 {
337 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
338 	XDR *xdrs = &clone_xprt->xp_xdrin;
339 	struct rpc_cots_server *stats = CLONE2STATS(clone_xprt);
340 
341 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KRECV_START,
342 	    "svc_cots_krecv_start:");
343 	RPCLOG(4, "svc_cots_krecv_start clone_xprt = %p:\n",
344 	    (void *)clone_xprt);
345 
346 	RSSTAT_INCR(stats, rscalls);
347 
348 	if (mp->b_datap->db_type != M_DATA) {
349 		RPCLOG(16, "svc_cots_krecv bad db_type %d\n",
350 		    mp->b_datap->db_type);
351 		goto bad;
352 	}
353 
354 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
355 
356 	TRACE_0(TR_FAC_KRPC, TR_XDR_CALLMSG_START,
357 	    "xdr_callmsg_start:");
358 	RPCLOG0(4, "xdr_callmsg_start:\n");
359 	if (!xdr_callmsg(xdrs, msg)) {
360 		TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
361 		    "xdr_callmsg_end:(%S)", "bad");
362 		RPCLOG0(1, "svc_cots_krecv xdr_callmsg failure\n");
363 		RSSTAT_INCR(stats, rsxdrcall);
364 		goto bad;
365 	}
366 	TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
367 	    "xdr_callmsg_end:(%S)", "good");
368 
369 	clone_xprt->xp_xid = msg->rm_xid;
370 	cd->cd_req_mp = mp;
371 
372 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
373 	    "svc_cots_krecv_end:(%S)", "good");
374 	RPCLOG0(4, "svc_cots_krecv_end:good\n");
375 	return (TRUE);
376 
377 bad:
378 	if (mp)
379 		freemsg(mp);
380 
381 	RSSTAT_INCR(stats, rsbadcalls);
382 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
383 	    "svc_cots_krecv_end:(%S)", "bad");
384 	return (FALSE);
385 }
386 
387 /*
388  * Send rpc reply.
389  */
390 static bool_t
391 svc_cots_ksend(SVCXPRT *clone_xprt, struct rpc_msg *msg)
392 {
393 	/* LINTED pointer alignment */
394 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
395 	XDR *xdrs = &(clone_xprt->xp_xdrout);
396 	int retval = FALSE;
397 	mblk_t *mp;
398 	xdrproc_t xdr_results;
399 	caddr_t xdr_location;
400 	bool_t has_args;
401 
402 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KSEND_START,
403 	    "svc_cots_ksend_start:");
404 
405 	/*
406 	 * If there is a result procedure specified in the reply message,
407 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
408 	 * We need to make sure it won't be processed twice, so we null
409 	 * it for xdr_replymsg here.
410 	 */
411 	has_args = FALSE;
412 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
413 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
414 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
415 			has_args = TRUE;
416 			xdr_location = msg->acpted_rply.ar_results.where;
417 			msg->acpted_rply.ar_results.proc = xdr_void;
418 			msg->acpted_rply.ar_results.where = NULL;
419 		}
420 	}
421 
422 	mp = cd->cd_mp;
423 	if (mp) {
424 		/*
425 		 * The program above pre-allocated an mblk and put
426 		 * the data in place.
427 		 */
428 		cd->cd_mp = (mblk_t *)NULL;
429 		if (!(xdr_replymsg_body(xdrs, msg) &&
430 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
431 		    xdr_results, xdr_location)))) {
432 			RPCLOG0(1, "svc_cots_ksend: "
433 			    "xdr_replymsg_body/SVCAUTH_WRAP failed\n");
434 			freemsg(mp);
435 			goto out;
436 		}
437 	} else {
438 		int	len;
439 		int	mpsize;
440 
441 		/*
442 		 * Leave space for protocol headers.
443 		 */
444 		len = MSG_OFFSET + clone_xprt->xp_msg_size;
445 
446 		/*
447 		 * Allocate an initial mblk for the response data.
448 		 */
449 		while (!(mp = allocb(len, BPRI_LO))) {
450 			RPCLOG0(16, "svc_cots_ksend: allocb failed failed\n");
451 			if (strwaitbuf(len, BPRI_LO)) {
452 				TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
453 				    "svc_cots_ksend_end:(%S)", "strwaitbuf");
454 				RPCLOG0(1,
455 				    "svc_cots_ksend: strwaitbuf failed\n");
456 				goto out;
457 			}
458 		}
459 
460 		/*
461 		 * Initialize the XDR decode stream.  Additional mblks
462 		 * will be allocated if necessary.  They will be TIDU
463 		 * sized.
464 		 */
465 		xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
466 		mpsize = MBLKSIZE(mp);
467 		ASSERT(mpsize >= len);
468 		ASSERT(mp->b_rptr == mp->b_datap->db_base);
469 
470 		/*
471 		 * If the size of mblk is not appreciably larger than what we
472 		 * asked, then resize the mblk to exactly len bytes. Reason for
473 		 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
474 		 * (from TCP over ethernet), and the arguments to RPC require
475 		 * 2800 bytes. Ideally we want the protocol to render two
476 		 * ~1400 byte segments over the wire. If allocb() gives us a 2k
477 		 * mblk, and we allocate a second mblk for the rest, the
478 		 * protocol module may generate 3 segments over the wire:
479 		 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
480 		 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
481 		 * the XDR encoding will generate two ~1400 byte mblks, and the
482 		 * protocol module is more likely to produce properly sized
483 		 * segments.
484 		 */
485 		if ((mpsize >> 1) <= len) {
486 			mp->b_rptr += (mpsize - len);
487 		}
488 
489 		/*
490 		 * Adjust b_rptr to reserve space for the non-data protocol
491 		 * headers that any downstream modules might like to add, and
492 		 * for the record marking header.
493 		 */
494 		mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
495 
496 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
497 		ASSERT(mp->b_wptr == mp->b_rptr);
498 
499 		msg->rm_xid = clone_xprt->xp_xid;
500 
501 		TRACE_0(TR_FAC_KRPC, TR_XDR_REPLYMSG_START,
502 		    "xdr_replymsg_start:");
503 		if (!(xdr_replymsg(xdrs, msg) &&
504 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
505 		    xdr_results, xdr_location)))) {
506 			TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
507 			    "xdr_replymsg_end:(%S)", "bad");
508 			freemsg(mp);
509 			RPCLOG0(1, "svc_cots_ksend: xdr_replymsg/SVCAUTH_WRAP "
510 			    "failed\n");
511 			goto out;
512 		}
513 		TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
514 		    "xdr_replymsg_end:(%S)", "good");
515 	}
516 
517 	put(clone_xprt->xp_wq, mp);
518 	retval = TRUE;
519 
520 out:
521 	/*
522 	 * This is completely disgusting.  If public is set it is
523 	 * a pointer to a structure whose first field is the address
524 	 * of the function to free that structure and any related
525 	 * stuff.  (see rrokfree in nfs_xdr.c).
526 	 */
527 	if (xdrs->x_public) {
528 		/* LINTED pointer alignment */
529 		(**((int (**)())xdrs->x_public))(xdrs->x_public);
530 	}
531 
532 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
533 	    "svc_cots_ksend_end:(%S)", "done");
534 	return (retval);
535 }
536 
537 /*
538  * Deserialize arguments.
539  */
540 static bool_t
541 svc_cots_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
542     caddr_t args_ptr)
543 {
544 	return (SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
545 	    xdr_args, args_ptr));
546 }
547 
548 static bool_t
549 svc_cots_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
550     caddr_t args_ptr)
551 {
552 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
553 	mblk_t *mp;
554 	bool_t retval;
555 
556 	/*
557 	 * It is important to call the XDR routine before
558 	 * freeing the request mblk.  Structures in the
559 	 * XDR data may point into the mblk and require that
560 	 * the memory be intact during the free routine.
561 	 */
562 	if (args_ptr) {
563 		/* LINTED pointer alignment */
564 		XDR	*xdrs = &clone_xprt->xp_xdrin;
565 
566 		xdrs->x_op = XDR_FREE;
567 		retval = (*xdr_args)(xdrs, args_ptr);
568 	} else
569 		retval = TRUE;
570 
571 	if ((mp = cd->cd_req_mp) != NULL) {
572 		cd->cd_req_mp = (mblk_t *)0;
573 		freemsg(mp);
574 	}
575 
576 	return (retval);
577 }
578 
579 static int32_t *
580 svc_cots_kgetres(SVCXPRT *clone_xprt, int size)
581 {
582 	/* LINTED pointer alignment */
583 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
584 	XDR *xdrs = &clone_xprt->xp_xdrout;
585 	mblk_t *mp;
586 	int32_t *buf;
587 	struct rpc_msg rply;
588 	int len;
589 	int mpsize;
590 
591 	/*
592 	 * Leave space for protocol headers.
593 	 */
594 	len = MSG_OFFSET + clone_xprt->xp_msg_size;
595 
596 	/*
597 	 * Allocate an initial mblk for the response data.
598 	 */
599 	while ((mp = allocb(len, BPRI_LO)) == NULL) {
600 		if (strwaitbuf(len, BPRI_LO))
601 			return (FALSE);
602 	}
603 
604 	/*
605 	 * Initialize the XDR decode stream.  Additional mblks
606 	 * will be allocated if necessary.  They will be TIDU
607 	 * sized.
608 	 */
609 	xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
610 	mpsize = MBLKSIZE(mp);
611 	ASSERT(mpsize >= len);
612 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
613 
614 	/*
615 	 * If the size of mblk is not appreciably larger than what we
616 	 * asked, then resize the mblk to exactly len bytes. Reason for
617 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
618 	 * (from TCP over ethernet), and the arguments to RPC require
619 	 * 2800 bytes. Ideally we want the protocol to render two
620 	 * ~1400 byte segments over the wire. If allocb() gives us a 2k
621 	 * mblk, and we allocate a second mblk for the rest, the
622 	 * protocol module may generate 3 segments over the wire:
623 	 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
624 	 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
625 	 * the XDR encoding will generate two ~1400 byte mblks, and the
626 	 * protocol module is more likely to produce properly sized
627 	 * segments.
628 	 */
629 	if ((mpsize >> 1) <= len) {
630 		mp->b_rptr += (mpsize - len);
631 	}
632 
633 	/*
634 	 * Adjust b_rptr to reserve space for the non-data protocol
635 	 * headers that any downstream modules might like to add, and
636 	 * for the record marking header.
637 	 */
638 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
639 
640 	XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
641 	ASSERT(mp->b_wptr == mp->b_rptr);
642 
643 	/*
644 	 * Assume a successful RPC since most of them are.
645 	 */
646 	rply.rm_xid = clone_xprt->xp_xid;
647 	rply.rm_direction = REPLY;
648 	rply.rm_reply.rp_stat = MSG_ACCEPTED;
649 	rply.acpted_rply.ar_verf = clone_xprt->xp_verf;
650 	rply.acpted_rply.ar_stat = SUCCESS;
651 
652 	if (!xdr_replymsg_hdr(xdrs, &rply)) {
653 		freeb(mp);
654 		return (NULL);
655 	}
656 
657 
658 	buf = XDR_INLINE(xdrs, size);
659 	if (buf == NULL) {
660 		ASSERT(cd->cd_mp == NULL);
661 		freemsg(mp);
662 	} else {
663 		cd->cd_mp = mp;
664 	}
665 	return (buf);
666 }
667 
668 static void
669 svc_cots_kfreeres(SVCXPRT *clone_xprt)
670 {
671 	cots_data_t *cd;
672 	mblk_t *mp;
673 
674 	cd = (cots_data_t *)clone_xprt->xp_p2buf;
675 	if ((mp = cd->cd_mp) != NULL) {
676 		cd->cd_mp = (mblk_t *)NULL;
677 		freemsg(mp);
678 	}
679 }
680 
681 /*
682  * the dup cacheing routines below provide a cache of non-failure
683  * transaction id's.  rpc service routines can use this to detect
684  * retransmissions and re-send a non-failure response.
685  */
686 
687 /*
688  * MAXDUPREQS is the number of cached items.  It should be adjusted
689  * to the service load so that there is likely to be a response entry
690  * when the first retransmission comes in.
691  */
692 #define	MAXDUPREQS	1024
693 
694 /*
695  * This should be appropriately scaled to MAXDUPREQS.
696  */
697 #define	DRHASHSZ	257
698 
699 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
700 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
701 #else
702 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
703 #endif
704 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
705 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
706 
707 static int	cotsndupreqs = 0;
708 static int	cotsmaxdupreqs = MAXDUPREQS;
709 static kmutex_t cotsdupreq_lock;
710 static struct dupreq *cotsdrhashtbl[DRHASHSZ];
711 static int	cotsdrhashstat[DRHASHSZ];
712 
713 static void unhash(struct dupreq *);
714 
715 /*
716  * cotsdrmru points to the head of a circular linked list in lru order.
717  * cotsdrmru->dr_next == drlru
718  */
719 struct dupreq *cotsdrmru;
720 
721 /*
722  * PSARC 2003/523 Contract Private Interface
723  * svc_cots_kdup
724  * Changes must be reviewed by Solaris File Sharing
725  * Changes must be communicated to contract-2003-523@sun.com
726  *
727  * svc_cots_kdup searches the request cache and returns 0 if the
728  * request is not found in the cache.  If it is found, then it
729  * returns the state of the request (in progress or done) and
730  * the status or attributes that were part of the original reply.
731  *
732  * If DUP_DONE (there is a duplicate) svc_cots_kdup copies over the
733  * value of the response. In that case, also return in *dupcachedp
734  * whether the response free routine is cached in the dupreq - in which case
735  * the caller should not be freeing it, because it will be done later
736  * in the svc_cots_kdup code when the dupreq is reused.
737  */
738 static int
739 svc_cots_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
740 	bool_t *dupcachedp)
741 {
742 	struct rpc_cots_server *stats = CLONE2STATS(req->rq_xprt);
743 	struct dupreq *dr;
744 	uint32_t xid;
745 	uint32_t drhash;
746 	int status;
747 
748 	xid = REQTOXID(req);
749 	mutex_enter(&cotsdupreq_lock);
750 	RSSTAT_INCR(stats, rsdupchecks);
751 	/*
752 	 * Check to see whether an entry already exists in the cache.
753 	 */
754 	dr = cotsdrhashtbl[XIDHASH(xid)];
755 	while (dr != NULL) {
756 		if (dr->dr_xid == xid &&
757 		    dr->dr_proc == req->rq_proc &&
758 		    dr->dr_prog == req->rq_prog &&
759 		    dr->dr_vers == req->rq_vers &&
760 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
761 		    bcmp((caddr_t)dr->dr_addr.buf,
762 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
763 		    dr->dr_addr.len) == 0) {
764 			status = dr->dr_status;
765 			if (status == DUP_DONE) {
766 				bcopy(dr->dr_resp.buf, res, size);
767 				if (dupcachedp != NULL)
768 					*dupcachedp = (dr->dr_resfree != NULL);
769 				TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KDUP_DONE,
770 				    "svc_cots_kdup: DUP_DONE");
771 			} else {
772 				dr->dr_status = DUP_INPROGRESS;
773 				*drpp = dr;
774 				TRACE_0(TR_FAC_KRPC,
775 				    TR_SVC_COTS_KDUP_INPROGRESS,
776 				    "svc_cots_kdup: DUP_INPROGRESS");
777 			}
778 			RSSTAT_INCR(stats, rsdupreqs);
779 			mutex_exit(&cotsdupreq_lock);
780 			return (status);
781 		}
782 		dr = dr->dr_chain;
783 	}
784 
785 	/*
786 	 * There wasn't an entry, either allocate a new one or recycle
787 	 * an old one.
788 	 */
789 	if (cotsndupreqs < cotsmaxdupreqs) {
790 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
791 		if (dr == NULL) {
792 			mutex_exit(&cotsdupreq_lock);
793 			return (DUP_ERROR);
794 		}
795 		dr->dr_resp.buf = NULL;
796 		dr->dr_resp.maxlen = 0;
797 		dr->dr_addr.buf = NULL;
798 		dr->dr_addr.maxlen = 0;
799 		if (cotsdrmru) {
800 			dr->dr_next = cotsdrmru->dr_next;
801 			cotsdrmru->dr_next = dr;
802 		} else {
803 			dr->dr_next = dr;
804 		}
805 		cotsndupreqs++;
806 	} else {
807 		dr = cotsdrmru->dr_next;
808 		while (dr->dr_status == DUP_INPROGRESS) {
809 			dr = dr->dr_next;
810 			if (dr == cotsdrmru->dr_next) {
811 				cmn_err(CE_WARN, "svc_cots_kdup no slots free");
812 				mutex_exit(&cotsdupreq_lock);
813 				return (DUP_ERROR);
814 			}
815 		}
816 		unhash(dr);
817 		if (dr->dr_resfree) {
818 			(*dr->dr_resfree)(dr->dr_resp.buf);
819 		}
820 	}
821 	dr->dr_resfree = NULL;
822 	cotsdrmru = dr;
823 
824 	dr->dr_xid = REQTOXID(req);
825 	dr->dr_prog = req->rq_prog;
826 	dr->dr_vers = req->rq_vers;
827 	dr->dr_proc = req->rq_proc;
828 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
829 		if (dr->dr_addr.buf != NULL)
830 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
831 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
832 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
833 		if (dr->dr_addr.buf == NULL) {
834 			dr->dr_addr.maxlen = 0;
835 			dr->dr_status = DUP_DROP;
836 			mutex_exit(&cotsdupreq_lock);
837 			return (DUP_ERROR);
838 		}
839 	}
840 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
841 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
842 	if (dr->dr_resp.maxlen < size) {
843 		if (dr->dr_resp.buf != NULL)
844 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
845 		dr->dr_resp.maxlen = (unsigned int)size;
846 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
847 		if (dr->dr_resp.buf == NULL) {
848 			dr->dr_resp.maxlen = 0;
849 			dr->dr_status = DUP_DROP;
850 			mutex_exit(&cotsdupreq_lock);
851 			return (DUP_ERROR);
852 		}
853 	}
854 	dr->dr_status = DUP_INPROGRESS;
855 
856 	drhash = (uint32_t)DRHASH(dr);
857 	dr->dr_chain = cotsdrhashtbl[drhash];
858 	cotsdrhashtbl[drhash] = dr;
859 	cotsdrhashstat[drhash]++;
860 	mutex_exit(&cotsdupreq_lock);
861 	*drpp = dr;
862 	return (DUP_NEW);
863 }
864 
865 /*
866  * PSARC 2003/523 Contract Private Interface
867  * svc_cots_kdupdone
868  * Changes must be reviewed by Solaris File Sharing
869  * Changes must be communicated to contract-2003-523@sun.com
870  *
871  * svc_cots_kdupdone marks the request done (DUP_DONE or DUP_DROP)
872  * and stores the response.
873  */
874 static void
875 svc_cots_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
876 	int size, int status)
877 {
878 	ASSERT(dr->dr_resfree == NULL);
879 	if (status == DUP_DONE) {
880 		bcopy(res, dr->dr_resp.buf, size);
881 		dr->dr_resfree = dis_resfree;
882 	}
883 	dr->dr_status = status;
884 }
885 
886 /*
887  * This routine expects that the mutex, cotsdupreq_lock, is already held.
888  */
889 static void
890 unhash(struct dupreq *dr)
891 {
892 	struct dupreq *drt;
893 	struct dupreq *drtprev = NULL;
894 	uint32_t drhash;
895 
896 	ASSERT(MUTEX_HELD(&cotsdupreq_lock));
897 
898 	drhash = (uint32_t)DRHASH(dr);
899 	drt = cotsdrhashtbl[drhash];
900 	while (drt != NULL) {
901 		if (drt == dr) {
902 			cotsdrhashstat[drhash]--;
903 			if (drtprev == NULL) {
904 				cotsdrhashtbl[drhash] = drt->dr_chain;
905 			} else {
906 				drtprev->dr_chain = drt->dr_chain;
907 			}
908 			return;
909 		}
910 		drtprev = drt;
911 		drt = drt->dr_chain;
912 	}
913 }
914 
915 void
916 svc_cots_stats_init(zoneid_t zoneid, struct rpc_cots_server **statsp)
917 {
918 	*statsp = (struct rpc_cots_server *)rpcstat_zone_init_common(zoneid,
919 	    "unix", "rpc_cots_server", (const kstat_named_t *)&cots_rsstat_tmpl,
920 	    sizeof (cots_rsstat_tmpl));
921 }
922 
923 void
924 svc_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_server **statsp)
925 {
926 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_server");
927 	kmem_free(*statsp, sizeof (cots_rsstat_tmpl));
928 }
929 
930 void
931 svc_cots_init(void)
932 {
933 	/*
934 	 * Check to make sure that the cots private data will fit into
935 	 * the stack buffer allocated by svc_run.  The ASSERT is a safety
936 	 * net if the cots_data_t structure ever changes.
937 	 */
938 	/*CONSTANTCONDITION*/
939 	ASSERT(sizeof (cots_data_t) <= SVC_P2LEN);
940 
941 	mutex_init(&cots_kcreate_lock, NULL, MUTEX_DEFAULT, NULL);
942 	mutex_init(&cotsdupreq_lock, NULL, MUTEX_DEFAULT, NULL);
943 }
944