xref: /titanic_41/usr/src/uts/common/rpc/svc_cots.c (revision a31148363f598def767ac48c5d82e1572e44b935)
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 2010 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 /*
35  * svc_cots.c
36  * Server side for connection-oriented RPC in the kernel.
37  *
38  */
39 
40 #include <sys/param.h>
41 #include <sys/types.h>
42 #include <sys/sysmacros.h>
43 #include <sys/file.h>
44 #include <sys/stream.h>
45 #include <sys/strsubr.h>
46 #include <sys/strsun.h>
47 #include <sys/stropts.h>
48 #include <sys/tiuser.h>
49 #include <sys/timod.h>
50 #include <sys/tihdr.h>
51 #include <sys/fcntl.h>
52 #include <sys/errno.h>
53 #include <sys/kmem.h>
54 #include <sys/systm.h>
55 #include <sys/debug.h>
56 #include <sys/cmn_err.h>
57 #include <sys/kstat.h>
58 #include <sys/vtrace.h>
59 
60 #include <rpc/types.h>
61 #include <rpc/xdr.h>
62 #include <rpc/auth.h>
63 #include <rpc/rpc_msg.h>
64 #include <rpc/svc.h>
65 #include <inet/ip.h>
66 
67 #define	COTS_MAX_ALLOCSIZE	2048
68 #define	MSG_OFFSET		128	/* offset of call into the mblk */
69 #define	RM_HDR_SIZE		4	/* record mark header size */
70 
71 /*
72  * Routines exported through ops vector.
73  */
74 static bool_t		svc_cots_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *);
75 static bool_t		svc_cots_ksend(SVCXPRT *, struct rpc_msg *);
76 static bool_t		svc_cots_kgetargs(SVCXPRT *, xdrproc_t, caddr_t);
77 static bool_t		svc_cots_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t);
78 static void		svc_cots_kdestroy(SVCMASTERXPRT *);
79 static int		svc_cots_kdup(struct svc_req *, caddr_t, int,
80 				struct dupreq **, bool_t *);
81 static void		svc_cots_kdupdone(struct dupreq *, caddr_t,
82 				void (*)(), int, int);
83 static int32_t		*svc_cots_kgetres(SVCXPRT *, int);
84 static void		svc_cots_kfreeres(SVCXPRT *);
85 static void		svc_cots_kclone_destroy(SVCXPRT *);
86 static void		svc_cots_kstart(SVCMASTERXPRT *);
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 };
105 
106 /*
107  * Master transport private data.
108  * Kept in xprt->xp_p2.
109  */
110 struct cots_master_data {
111 	char	*cmd_src_addr;	/* client's address */
112 	int	cmd_xprt_started; /* flag for clone routine to call */
113 				/* rpcmod's start routine. */
114 	struct rpc_cots_server *cmd_stats;	/* stats for zone */
115 };
116 
117 /*
118  * Transport private data.
119  * Kept in clone_xprt->xp_p2buf.
120  */
121 typedef struct cots_data {
122 	mblk_t	*cd_mp;		/* pre-allocated reply message */
123 	mblk_t	*cd_req_mp;	/* request message */
124 } cots_data_t;
125 
126 /*
127  * Server statistics
128  * NOTE: This structure type is duplicated in the NFS fast path.
129  */
130 static const struct rpc_cots_server {
131 	kstat_named_t	rscalls;
132 	kstat_named_t	rsbadcalls;
133 	kstat_named_t	rsnullrecv;
134 	kstat_named_t	rsbadlen;
135 	kstat_named_t	rsxdrcall;
136 	kstat_named_t	rsdupchecks;
137 	kstat_named_t	rsdupreqs;
138 } cots_rsstat_tmpl = {
139 	{ "calls",	KSTAT_DATA_UINT64 },
140 	{ "badcalls",	KSTAT_DATA_UINT64 },
141 	{ "nullrecv",	KSTAT_DATA_UINT64 },
142 	{ "badlen",	KSTAT_DATA_UINT64 },
143 	{ "xdrcall",	KSTAT_DATA_UINT64 },
144 	{ "dupchecks",	KSTAT_DATA_UINT64 },
145 	{ "dupreqs",	KSTAT_DATA_UINT64 }
146 };
147 
148 #define	CLONE2STATS(clone_xprt)	\
149 	((struct cots_master_data *)(clone_xprt)->xp_master->xp_p2)->cmd_stats
150 #define	RSSTAT_INCR(s, x)	\
151 	atomic_add_64(&(s)->x.value.ui64, 1)
152 
153 /*
154  * Pointer to a transport specific `ready to receive' function in rpcmod
155  * (set from rpcmod).
156  */
157 void    (*mir_start)(queue_t *);
158 uint_t	*svc_max_msg_sizep;
159 
160 /*
161  * the address size of the underlying transport can sometimes be
162  * unknown (tinfo->ADDR_size == -1).  For this case, it is
163  * necessary to figure out what the size is so the correct amount
164  * of data is allocated.  This is an itterative process:
165  *	1. take a good guess (use T_MINADDRSIZE)
166  *	2. try it.
167  *	3. if it works then everything is ok
168  *	4. if the error is ENAMETOLONG, double the guess
169  *	5. go back to step 2.
170  */
171 #define	T_UNKNOWNADDRSIZE	(-1)
172 #define	T_MINADDRSIZE	32
173 
174 /*
175  * Create a transport record.
176  * The transport record, output buffer, and private data structure
177  * are allocated.  The output buffer is serialized into using xdrmem.
178  * There is one transport record per user process which implements a
179  * set of services.
180  */
181 static kmutex_t cots_kcreate_lock;
182 
183 int
184 svc_cots_kcreate(file_t *fp, uint_t max_msgsize, struct T_info_ack *tinfo,
185     SVCMASTERXPRT **nxprt)
186 {
187 	struct cots_master_data *cmd;
188 	int err, retval;
189 	SVCMASTERXPRT *xprt;
190 	struct rpcstat *rpcstat;
191 	struct T_addr_ack *ack_p;
192 	struct strioctl getaddr;
193 
194 	if (nxprt == NULL)
195 		return (EINVAL);
196 
197 	rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone);
198 	ASSERT(rpcstat != NULL);
199 
200 	xprt = kmem_zalloc(sizeof (SVCMASTERXPRT), KM_SLEEP);
201 
202 	cmd = kmem_zalloc(sizeof (*cmd) + sizeof (*ack_p)
203 	    + (2 * sizeof (sin6_t)), KM_SLEEP);
204 
205 	ack_p = (struct T_addr_ack *)&cmd[1];
206 
207 	if ((tinfo->TIDU_size > COTS_MAX_ALLOCSIZE) ||
208 	    (tinfo->TIDU_size <= 0))
209 		xprt->xp_msg_size = COTS_MAX_ALLOCSIZE;
210 	else {
211 		xprt->xp_msg_size = tinfo->TIDU_size -
212 		    (tinfo->TIDU_size % BYTES_PER_XDR_UNIT);
213 	}
214 
215 	xprt->xp_ops = &svc_cots_op;
216 	xprt->xp_p2 = (caddr_t)cmd;
217 	cmd->cmd_xprt_started = 0;
218 	cmd->cmd_stats = rpcstat->rpc_cots_server;
219 
220 	getaddr.ic_cmd = TI_GETINFO;
221 	getaddr.ic_timout = -1;
222 	getaddr.ic_len = sizeof (*ack_p) + (2 * sizeof (sin6_t));
223 	getaddr.ic_dp = (char *)ack_p;
224 	ack_p->PRIM_type = T_ADDR_REQ;
225 
226 	err = strioctl(fp->f_vnode, I_STR, (intptr_t)&getaddr,
227 	    0, K_TO_K, CRED(), &retval);
228 	if (err) {
229 		kmem_free(cmd, sizeof (*cmd) + sizeof (*ack_p) +
230 		    (2 * sizeof (sin6_t)));
231 		kmem_free(xprt, sizeof (SVCMASTERXPRT));
232 		return (err);
233 	}
234 
235 	xprt->xp_rtaddr.maxlen = ack_p->REMADDR_length;
236 	xprt->xp_rtaddr.len = ack_p->REMADDR_length;
237 	cmd->cmd_src_addr = xprt->xp_rtaddr.buf =
238 	    (char *)ack_p + ack_p->REMADDR_offset;
239 
240 	xprt->xp_lcladdr.maxlen = ack_p->LOCADDR_length;
241 	xprt->xp_lcladdr.len = ack_p->LOCADDR_length;
242 	xprt->xp_lcladdr.buf = (char *)ack_p + ack_p->LOCADDR_offset;
243 
244 	/*
245 	 * If the current sanity check size in rpcmod is smaller
246 	 * than the size needed for this xprt, then increase
247 	 * the sanity check.
248 	 */
249 	if (max_msgsize != 0 && svc_max_msg_sizep &&
250 	    max_msgsize > *svc_max_msg_sizep) {
251 
252 		/* This check needs a lock */
253 		mutex_enter(&cots_kcreate_lock);
254 		if (svc_max_msg_sizep && max_msgsize > *svc_max_msg_sizep)
255 			*svc_max_msg_sizep = max_msgsize;
256 		mutex_exit(&cots_kcreate_lock);
257 	}
258 
259 	*nxprt = xprt;
260 
261 	return (0);
262 }
263 
264 /*
265  * Destroy a master transport record.
266  * Frees the space allocated for a transport record.
267  */
268 static void
269 svc_cots_kdestroy(SVCMASTERXPRT *xprt)
270 {
271 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
272 
273 	ASSERT(cmd);
274 
275 	if (xprt->xp_netid)
276 		kmem_free(xprt->xp_netid, strlen(xprt->xp_netid) + 1);
277 	if (xprt->xp_addrmask.maxlen)
278 		kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen);
279 
280 	mutex_destroy(&xprt->xp_req_lock);
281 	mutex_destroy(&xprt->xp_thread_lock);
282 
283 	kmem_free(cmd, sizeof (*cmd) + sizeof (struct T_addr_ack) +
284 	    (2 * sizeof (sin6_t)));
285 
286 	kmem_free(xprt, sizeof (SVCMASTERXPRT));
287 }
288 
289 /*
290  * svc_tli_kcreate() calls this function at the end to tell
291  * rpcmod that the transport is ready to receive requests.
292  */
293 static void
294 svc_cots_kstart(SVCMASTERXPRT *xprt)
295 {
296 	struct cots_master_data *cmd = (struct cots_master_data *)xprt->xp_p2;
297 
298 	if (cmd->cmd_xprt_started == 0) {
299 		/*
300 		 * Acquire the xp_req_lock in order to use xp_wq
301 		 * safely (we don't want to qenable a queue that has
302 		 * already been closed).
303 		 */
304 		mutex_enter(&xprt->xp_req_lock);
305 		if (cmd->cmd_xprt_started == 0 &&
306 		    xprt->xp_wq != NULL) {
307 			(*mir_start)(xprt->xp_wq);
308 			cmd->cmd_xprt_started = 1;
309 		}
310 		mutex_exit(&xprt->xp_req_lock);
311 	}
312 }
313 
314 /*
315  * Transport-type specific part of svc_xprt_cleanup().
316  */
317 static void
318 svc_cots_kclone_destroy(SVCXPRT *clone_xprt)
319 {
320 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
321 
322 	if (cd->cd_req_mp) {
323 		freemsg(cd->cd_req_mp);
324 		cd->cd_req_mp = (mblk_t *)0;
325 	}
326 	ASSERT(cd->cd_mp == NULL);
327 }
328 
329 /*
330  * Receive rpc requests.
331  * Checks if the message is intact, and deserializes the call packet.
332  */
333 static bool_t
334 svc_cots_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg)
335 {
336 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
337 	XDR *xdrs = &clone_xprt->xp_xdrin;
338 	struct rpc_cots_server *stats = CLONE2STATS(clone_xprt);
339 
340 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KRECV_START,
341 	    "svc_cots_krecv_start:");
342 	RPCLOG(4, "svc_cots_krecv_start clone_xprt = %p:\n",
343 	    (void *)clone_xprt);
344 
345 	RSSTAT_INCR(stats, rscalls);
346 
347 	if (mp->b_datap->db_type != M_DATA) {
348 		RPCLOG(16, "svc_cots_krecv bad db_type %d\n",
349 		    mp->b_datap->db_type);
350 		goto bad;
351 	}
352 
353 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
354 
355 	TRACE_0(TR_FAC_KRPC, TR_XDR_CALLMSG_START,
356 	    "xdr_callmsg_start:");
357 	RPCLOG0(4, "xdr_callmsg_start:\n");
358 	if (!xdr_callmsg(xdrs, msg)) {
359 		TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
360 		    "xdr_callmsg_end:(%S)", "bad");
361 		RPCLOG0(1, "svc_cots_krecv xdr_callmsg failure\n");
362 		RSSTAT_INCR(stats, rsxdrcall);
363 		goto bad;
364 	}
365 	TRACE_1(TR_FAC_KRPC, TR_XDR_CALLMSG_END,
366 	    "xdr_callmsg_end:(%S)", "good");
367 
368 	clone_xprt->xp_xid = msg->rm_xid;
369 	cd->cd_req_mp = mp;
370 
371 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
372 	    "svc_cots_krecv_end:(%S)", "good");
373 	RPCLOG0(4, "svc_cots_krecv_end:good\n");
374 	return (TRUE);
375 
376 bad:
377 	if (mp)
378 		freemsg(mp);
379 
380 	RSSTAT_INCR(stats, rsbadcalls);
381 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KRECV_END,
382 	    "svc_cots_krecv_end:(%S)", "bad");
383 	return (FALSE);
384 }
385 
386 /*
387  * Send rpc reply.
388  */
389 static bool_t
390 svc_cots_ksend(SVCXPRT *clone_xprt, struct rpc_msg *msg)
391 {
392 	/* LINTED pointer alignment */
393 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
394 	XDR *xdrs = &(clone_xprt->xp_xdrout);
395 	int retval = FALSE;
396 	mblk_t *mp;
397 	xdrproc_t xdr_results;
398 	caddr_t xdr_location;
399 	bool_t has_args;
400 
401 	TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KSEND_START,
402 	    "svc_cots_ksend_start:");
403 
404 	/*
405 	 * If there is a result procedure specified in the reply message,
406 	 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP.
407 	 * We need to make sure it won't be processed twice, so we null
408 	 * it for xdr_replymsg here.
409 	 */
410 	has_args = FALSE;
411 	if (msg->rm_reply.rp_stat == MSG_ACCEPTED &&
412 	    msg->rm_reply.rp_acpt.ar_stat == SUCCESS) {
413 		if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) {
414 			has_args = TRUE;
415 			xdr_location = msg->acpted_rply.ar_results.where;
416 			msg->acpted_rply.ar_results.proc = xdr_void;
417 			msg->acpted_rply.ar_results.where = NULL;
418 		}
419 	}
420 
421 	mp = cd->cd_mp;
422 	if (mp) {
423 		/*
424 		 * The program above pre-allocated an mblk and put
425 		 * the data in place.
426 		 */
427 		cd->cd_mp = (mblk_t *)NULL;
428 		if (!(xdr_replymsg_body(xdrs, msg) &&
429 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
430 		    xdr_results, xdr_location)))) {
431 			RPCLOG0(1, "svc_cots_ksend: "
432 			    "xdr_replymsg_body/SVCAUTH_WRAP failed\n");
433 			freemsg(mp);
434 			goto out;
435 		}
436 	} else {
437 		int	len;
438 		int	mpsize;
439 
440 		/*
441 		 * Leave space for protocol headers.
442 		 */
443 		len = MSG_OFFSET + clone_xprt->xp_msg_size;
444 
445 		/*
446 		 * Allocate an initial mblk for the response data.
447 		 */
448 		while (!(mp = allocb(len, BPRI_LO))) {
449 			RPCLOG0(16, "svc_cots_ksend: allocb failed failed\n");
450 			if (strwaitbuf(len, BPRI_LO)) {
451 				TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
452 				    "svc_cots_ksend_end:(%S)", "strwaitbuf");
453 				RPCLOG0(1,
454 				    "svc_cots_ksend: strwaitbuf failed\n");
455 				goto out;
456 			}
457 		}
458 
459 		/*
460 		 * Initialize the XDR decode stream.  Additional mblks
461 		 * will be allocated if necessary.  They will be TIDU
462 		 * sized.
463 		 */
464 		xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
465 		mpsize = MBLKSIZE(mp);
466 		ASSERT(mpsize >= len);
467 		ASSERT(mp->b_rptr == mp->b_datap->db_base);
468 
469 		/*
470 		 * If the size of mblk is not appreciably larger than what we
471 		 * asked, then resize the mblk to exactly len bytes. Reason for
472 		 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
473 		 * (from TCP over ethernet), and the arguments to RPC require
474 		 * 2800 bytes. Ideally we want the protocol to render two
475 		 * ~1400 byte segments over the wire. If allocb() gives us a 2k
476 		 * mblk, and we allocate a second mblk for the rest, the
477 		 * protocol module may generate 3 segments over the wire:
478 		 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
479 		 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
480 		 * the XDR encoding will generate two ~1400 byte mblks, and the
481 		 * protocol module is more likely to produce properly sized
482 		 * segments.
483 		 */
484 		if ((mpsize >> 1) <= len) {
485 			mp->b_rptr += (mpsize - len);
486 		}
487 
488 		/*
489 		 * Adjust b_rptr to reserve space for the non-data protocol
490 		 * headers that any downstream modules might like to add, and
491 		 * for the record marking header.
492 		 */
493 		mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
494 
495 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
496 		ASSERT(mp->b_wptr == mp->b_rptr);
497 
498 		msg->rm_xid = clone_xprt->xp_xid;
499 
500 		TRACE_0(TR_FAC_KRPC, TR_XDR_REPLYMSG_START,
501 		    "xdr_replymsg_start:");
502 		if (!(xdr_replymsg(xdrs, msg) &&
503 		    (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs,
504 		    xdr_results, xdr_location)))) {
505 			TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
506 			    "xdr_replymsg_end:(%S)", "bad");
507 			freemsg(mp);
508 			RPCLOG0(1, "svc_cots_ksend: xdr_replymsg/SVCAUTH_WRAP "
509 			    "failed\n");
510 			goto out;
511 		}
512 		TRACE_1(TR_FAC_KRPC, TR_XDR_REPLYMSG_END,
513 		    "xdr_replymsg_end:(%S)", "good");
514 	}
515 
516 	put(clone_xprt->xp_wq, mp);
517 	retval = TRUE;
518 
519 out:
520 	/*
521 	 * This is completely disgusting.  If public is set it is
522 	 * a pointer to a structure whose first field is the address
523 	 * of the function to free that structure and any related
524 	 * stuff.  (see rrokfree in nfs_xdr.c).
525 	 */
526 	if (xdrs->x_public) {
527 		/* LINTED pointer alignment */
528 		(**((int (**)())xdrs->x_public))(xdrs->x_public);
529 	}
530 
531 	TRACE_1(TR_FAC_KRPC, TR_SVC_COTS_KSEND_END,
532 	    "svc_cots_ksend_end:(%S)", "done");
533 	return (retval);
534 }
535 
536 /*
537  * Deserialize arguments.
538  */
539 static bool_t
540 svc_cots_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
541     caddr_t args_ptr)
542 {
543 	return (SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin,
544 	    xdr_args, args_ptr));
545 }
546 
547 static bool_t
548 svc_cots_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args,
549     caddr_t args_ptr)
550 {
551 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
552 	mblk_t *mp;
553 	bool_t retval;
554 
555 	/*
556 	 * It is important to call the XDR routine before
557 	 * freeing the request mblk.  Structures in the
558 	 * XDR data may point into the mblk and require that
559 	 * the memory be intact during the free routine.
560 	 */
561 	if (args_ptr) {
562 		/* LINTED pointer alignment */
563 		XDR	*xdrs = &clone_xprt->xp_xdrin;
564 
565 		xdrs->x_op = XDR_FREE;
566 		retval = (*xdr_args)(xdrs, args_ptr);
567 	} else
568 		retval = TRUE;
569 
570 	if ((mp = cd->cd_req_mp) != NULL) {
571 		cd->cd_req_mp = (mblk_t *)0;
572 		freemsg(mp);
573 	}
574 
575 	return (retval);
576 }
577 
578 static int32_t *
579 svc_cots_kgetres(SVCXPRT *clone_xprt, int size)
580 {
581 	/* LINTED pointer alignment */
582 	cots_data_t *cd = (cots_data_t *)clone_xprt->xp_p2buf;
583 	XDR *xdrs = &clone_xprt->xp_xdrout;
584 	mblk_t *mp;
585 	int32_t *buf;
586 	struct rpc_msg rply;
587 	int len;
588 	int mpsize;
589 
590 	/*
591 	 * Leave space for protocol headers.
592 	 */
593 	len = MSG_OFFSET + clone_xprt->xp_msg_size;
594 
595 	/*
596 	 * Allocate an initial mblk for the response data.
597 	 */
598 	while ((mp = allocb(len, BPRI_LO)) == NULL) {
599 		if (strwaitbuf(len, BPRI_LO))
600 			return (FALSE);
601 	}
602 
603 	/*
604 	 * Initialize the XDR decode stream.  Additional mblks
605 	 * will be allocated if necessary.  They will be TIDU
606 	 * sized.
607 	 */
608 	xdrmblk_init(xdrs, mp, XDR_ENCODE, clone_xprt->xp_msg_size);
609 	mpsize = MBLKSIZE(mp);
610 	ASSERT(mpsize >= len);
611 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
612 
613 	/*
614 	 * If the size of mblk is not appreciably larger than what we
615 	 * asked, then resize the mblk to exactly len bytes. Reason for
616 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
617 	 * (from TCP over ethernet), and the arguments to RPC require
618 	 * 2800 bytes. Ideally we want the protocol to render two
619 	 * ~1400 byte segments over the wire. If allocb() gives us a 2k
620 	 * mblk, and we allocate a second mblk for the rest, the
621 	 * protocol module may generate 3 segments over the wire:
622 	 * 1460 bytes for the first, 448 (2048 - 1600) for the 2nd, and
623 	 * 892 for the 3rd. If we "waste" 448 bytes in the first mblk,
624 	 * the XDR encoding will generate two ~1400 byte mblks, and the
625 	 * protocol module is more likely to produce properly sized
626 	 * segments.
627 	 */
628 	if ((mpsize >> 1) <= len) {
629 		mp->b_rptr += (mpsize - len);
630 	}
631 
632 	/*
633 	 * Adjust b_rptr to reserve space for the non-data protocol
634 	 * headers that any downstream modules might like to add, and
635 	 * for the record marking header.
636 	 */
637 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
638 
639 	XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
640 	ASSERT(mp->b_wptr == mp->b_rptr);
641 
642 	/*
643 	 * Assume a successful RPC since most of them are.
644 	 */
645 	rply.rm_xid = clone_xprt->xp_xid;
646 	rply.rm_direction = REPLY;
647 	rply.rm_reply.rp_stat = MSG_ACCEPTED;
648 	rply.acpted_rply.ar_verf = clone_xprt->xp_verf;
649 	rply.acpted_rply.ar_stat = SUCCESS;
650 
651 	if (!xdr_replymsg_hdr(xdrs, &rply)) {
652 		freeb(mp);
653 		return (NULL);
654 	}
655 
656 
657 	buf = XDR_INLINE(xdrs, size);
658 	if (buf == NULL) {
659 		ASSERT(cd->cd_mp == NULL);
660 		freemsg(mp);
661 	} else {
662 		cd->cd_mp = mp;
663 	}
664 	return (buf);
665 }
666 
667 static void
668 svc_cots_kfreeres(SVCXPRT *clone_xprt)
669 {
670 	cots_data_t *cd;
671 	mblk_t *mp;
672 
673 	cd = (cots_data_t *)clone_xprt->xp_p2buf;
674 	if ((mp = cd->cd_mp) != NULL) {
675 		cd->cd_mp = (mblk_t *)NULL;
676 		freemsg(mp);
677 	}
678 }
679 
680 /*
681  * the dup cacheing routines below provide a cache of non-failure
682  * transaction id's.  rpc service routines can use this to detect
683  * retransmissions and re-send a non-failure response.
684  */
685 
686 /*
687  * MAXDUPREQS is the number of cached items.  It should be adjusted
688  * to the service load so that there is likely to be a response entry
689  * when the first retransmission comes in.
690  */
691 #define	MAXDUPREQS	1024
692 
693 /*
694  * This should be appropriately scaled to MAXDUPREQS.
695  */
696 #define	DRHASHSZ	257
697 
698 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0)
699 #define	XIDHASH(xid)	((xid) & (DRHASHSZ - 1))
700 #else
701 #define	XIDHASH(xid)	((xid) % DRHASHSZ)
702 #endif
703 #define	DRHASH(dr)	XIDHASH((dr)->dr_xid)
704 #define	REQTOXID(req)	((req)->rq_xprt->xp_xid)
705 
706 static int	cotsndupreqs = 0;
707 int	cotsmaxdupreqs = MAXDUPREQS;
708 static kmutex_t cotsdupreq_lock;
709 static struct dupreq *cotsdrhashtbl[DRHASHSZ];
710 static int	cotsdrhashstat[DRHASHSZ];
711 
712 static void unhash(struct dupreq *);
713 
714 /*
715  * cotsdrmru points to the head of a circular linked list in lru order.
716  * cotsdrmru->dr_next == drlru
717  */
718 struct dupreq *cotsdrmru;
719 
720 /*
721  * PSARC 2003/523 Contract Private Interface
722  * svc_cots_kdup
723  * Changes must be reviewed by Solaris File Sharing
724  * Changes must be communicated to contract-2003-523@sun.com
725  *
726  * svc_cots_kdup searches the request cache and returns 0 if the
727  * request is not found in the cache.  If it is found, then it
728  * returns the state of the request (in progress or done) and
729  * the status or attributes that were part of the original reply.
730  *
731  * If DUP_DONE (there is a duplicate) svc_cots_kdup copies over the
732  * value of the response. In that case, also return in *dupcachedp
733  * whether the response free routine is cached in the dupreq - in which case
734  * the caller should not be freeing it, because it will be done later
735  * in the svc_cots_kdup code when the dupreq is reused.
736  */
737 static int
738 svc_cots_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp,
739 	bool_t *dupcachedp)
740 {
741 	struct rpc_cots_server *stats = CLONE2STATS(req->rq_xprt);
742 	struct dupreq *dr;
743 	uint32_t xid;
744 	uint32_t drhash;
745 	int status;
746 
747 	xid = REQTOXID(req);
748 	mutex_enter(&cotsdupreq_lock);
749 	RSSTAT_INCR(stats, rsdupchecks);
750 	/*
751 	 * Check to see whether an entry already exists in the cache.
752 	 */
753 	dr = cotsdrhashtbl[XIDHASH(xid)];
754 	while (dr != NULL) {
755 		if (dr->dr_xid == xid &&
756 		    dr->dr_proc == req->rq_proc &&
757 		    dr->dr_prog == req->rq_prog &&
758 		    dr->dr_vers == req->rq_vers &&
759 		    dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len &&
760 		    bcmp((caddr_t)dr->dr_addr.buf,
761 		    (caddr_t)req->rq_xprt->xp_rtaddr.buf,
762 		    dr->dr_addr.len) == 0) {
763 			status = dr->dr_status;
764 			if (status == DUP_DONE) {
765 				bcopy(dr->dr_resp.buf, res, size);
766 				if (dupcachedp != NULL)
767 					*dupcachedp = (dr->dr_resfree != NULL);
768 				TRACE_0(TR_FAC_KRPC, TR_SVC_COTS_KDUP_DONE,
769 				    "svc_cots_kdup: DUP_DONE");
770 			} else {
771 				dr->dr_status = DUP_INPROGRESS;
772 				*drpp = dr;
773 				TRACE_0(TR_FAC_KRPC,
774 				    TR_SVC_COTS_KDUP_INPROGRESS,
775 				    "svc_cots_kdup: DUP_INPROGRESS");
776 			}
777 			RSSTAT_INCR(stats, rsdupreqs);
778 			mutex_exit(&cotsdupreq_lock);
779 			return (status);
780 		}
781 		dr = dr->dr_chain;
782 	}
783 
784 	/*
785 	 * There wasn't an entry, either allocate a new one or recycle
786 	 * an old one.
787 	 */
788 	if (cotsndupreqs < cotsmaxdupreqs) {
789 		dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP);
790 		if (dr == NULL) {
791 			mutex_exit(&cotsdupreq_lock);
792 			return (DUP_ERROR);
793 		}
794 		dr->dr_resp.buf = NULL;
795 		dr->dr_resp.maxlen = 0;
796 		dr->dr_addr.buf = NULL;
797 		dr->dr_addr.maxlen = 0;
798 		if (cotsdrmru) {
799 			dr->dr_next = cotsdrmru->dr_next;
800 			cotsdrmru->dr_next = dr;
801 		} else {
802 			dr->dr_next = dr;
803 		}
804 		cotsndupreqs++;
805 	} else {
806 		dr = cotsdrmru->dr_next;
807 		while (dr->dr_status == DUP_INPROGRESS) {
808 			dr = dr->dr_next;
809 			if (dr == cotsdrmru->dr_next) {
810 				cmn_err(CE_WARN, "svc_cots_kdup no slots free");
811 				mutex_exit(&cotsdupreq_lock);
812 				return (DUP_ERROR);
813 			}
814 		}
815 		unhash(dr);
816 		if (dr->dr_resfree) {
817 			(*dr->dr_resfree)(dr->dr_resp.buf);
818 		}
819 	}
820 	dr->dr_resfree = NULL;
821 	cotsdrmru = dr;
822 
823 	dr->dr_xid = REQTOXID(req);
824 	dr->dr_prog = req->rq_prog;
825 	dr->dr_vers = req->rq_vers;
826 	dr->dr_proc = req->rq_proc;
827 	if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) {
828 		if (dr->dr_addr.buf != NULL)
829 			kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen);
830 		dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len;
831 		dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP);
832 		if (dr->dr_addr.buf == NULL) {
833 			dr->dr_addr.maxlen = 0;
834 			dr->dr_status = DUP_DROP;
835 			mutex_exit(&cotsdupreq_lock);
836 			return (DUP_ERROR);
837 		}
838 	}
839 	dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len;
840 	bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len);
841 	if (dr->dr_resp.maxlen < size) {
842 		if (dr->dr_resp.buf != NULL)
843 			kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen);
844 		dr->dr_resp.maxlen = (unsigned int)size;
845 		dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP);
846 		if (dr->dr_resp.buf == NULL) {
847 			dr->dr_resp.maxlen = 0;
848 			dr->dr_status = DUP_DROP;
849 			mutex_exit(&cotsdupreq_lock);
850 			return (DUP_ERROR);
851 		}
852 	}
853 	dr->dr_status = DUP_INPROGRESS;
854 
855 	drhash = (uint32_t)DRHASH(dr);
856 	dr->dr_chain = cotsdrhashtbl[drhash];
857 	cotsdrhashtbl[drhash] = dr;
858 	cotsdrhashstat[drhash]++;
859 	mutex_exit(&cotsdupreq_lock);
860 	*drpp = dr;
861 	return (DUP_NEW);
862 }
863 
864 /*
865  * PSARC 2003/523 Contract Private Interface
866  * svc_cots_kdupdone
867  * Changes must be reviewed by Solaris File Sharing
868  * Changes must be communicated to contract-2003-523@sun.com
869  *
870  * svc_cots_kdupdone marks the request done (DUP_DONE or DUP_DROP)
871  * and stores the response.
872  */
873 static void
874 svc_cots_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(),
875 	int size, int status)
876 {
877 	ASSERT(dr->dr_resfree == NULL);
878 	if (status == DUP_DONE) {
879 		bcopy(res, dr->dr_resp.buf, size);
880 		dr->dr_resfree = dis_resfree;
881 	}
882 	dr->dr_status = status;
883 }
884 
885 /*
886  * This routine expects that the mutex, cotsdupreq_lock, is already held.
887  */
888 static void
889 unhash(struct dupreq *dr)
890 {
891 	struct dupreq *drt;
892 	struct dupreq *drtprev = NULL;
893 	uint32_t drhash;
894 
895 	ASSERT(MUTEX_HELD(&cotsdupreq_lock));
896 
897 	drhash = (uint32_t)DRHASH(dr);
898 	drt = cotsdrhashtbl[drhash];
899 	while (drt != NULL) {
900 		if (drt == dr) {
901 			cotsdrhashstat[drhash]--;
902 			if (drtprev == NULL) {
903 				cotsdrhashtbl[drhash] = drt->dr_chain;
904 			} else {
905 				drtprev->dr_chain = drt->dr_chain;
906 			}
907 			return;
908 		}
909 		drtprev = drt;
910 		drt = drt->dr_chain;
911 	}
912 }
913 
914 void
915 svc_cots_stats_init(zoneid_t zoneid, struct rpc_cots_server **statsp)
916 {
917 	*statsp = (struct rpc_cots_server *)rpcstat_zone_init_common(zoneid,
918 	    "unix", "rpc_cots_server", (const kstat_named_t *)&cots_rsstat_tmpl,
919 	    sizeof (cots_rsstat_tmpl));
920 }
921 
922 void
923 svc_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_server **statsp)
924 {
925 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_server");
926 	kmem_free(*statsp, sizeof (cots_rsstat_tmpl));
927 }
928 
929 void
930 svc_cots_init(void)
931 {
932 	/*
933 	 * Check to make sure that the cots private data will fit into
934 	 * the stack buffer allocated by svc_run.  The ASSERT is a safety
935 	 * net if the cots_data_t structure ever changes.
936 	 */
937 	/*CONSTANTCONDITION*/
938 	ASSERT(sizeof (cots_data_t) <= SVC_P2LEN);
939 
940 	mutex_init(&cots_kcreate_lock, NULL, MUTEX_DEFAULT, NULL);
941 	mutex_init(&cotsdupreq_lock, NULL, MUTEX_DEFAULT, NULL);
942 }
943