xref: /freebsd/sys/fs/nfsserver/nfs_nfsdcache.c (revision 298cf604ccf133b101c6fad42d1a078a1fac58ca)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 /*
38  * Here is the basic algorithm:
39  * First, some design criteria I used:
40  * - I think a false hit is more serious than a false miss
41  * - A false hit for an RPC that has Op(s) that order via seqid# must be
42  *   avoided at all cost
43  * - A valid hit will probably happen a long time after the original reply
44  *   and the TCP socket that the original request was received on will no
45  *   longer be active
46  *   (The long time delay implies to me that LRU is not appropriate.)
47  * - The mechanism will satisfy the requirements of ordering Ops with seqid#s
48  *   in them as well as minimizing the risk of redoing retried non-idempotent
49  *   Ops.
50  * Because it is biased towards avoiding false hits, multiple entries with
51  * the same xid are to be expected, especially for the case of the entry
52  * in the cache being related to a seqid# sequenced Op.
53  *
54  * The basic algorithm I'm about to code up:
55  * - Null RPCs bypass the cache and are just done
56  * For TCP
57  * 	- key on <xid, NFS version> (as noted above, there can be several
58  * 				     entries with the same key)
59  * 	When a request arrives:
60  * 		For all that match key
61  * 		- if RPC# != OR request_size !=
62  * 			- not a match with this one
63  * 		- if NFSv4 and received on same TCP socket OR
64  *			received on a TCP connection created before the
65  *			entry was cached
66  * 			- not a match with this one
67  * 			(V2,3 clients might retry on same TCP socket)
68  * 		- calculate checksum on first N bytes of NFS XDR
69  * 		- if checksum !=
70  * 			- not a match for this one
71  * 		If any of the remaining ones that match has a
72  * 			seqid_refcnt > 0
73  * 			- not a match (go do RPC, using new cache entry)
74  * 		If one match left
75  * 			- a hit (reply from cache)
76  * 		else
77  * 			- miss (go do RPC, using new cache entry)
78  *
79  * 	During processing of NFSv4 request:
80  * 		- set a flag when a non-idempotent Op is processed
81  * 		- when an Op that uses a seqid# (Open,...) is processed
82  * 			- if same seqid# as referenced entry in cache
83  * 				- free new cache entry
84  * 				- reply from referenced cache entry
85  * 			  else if next seqid# in order
86  * 				- free referenced cache entry
87  * 				- increment seqid_refcnt on new cache entry
88  * 				- set pointer from Openowner/Lockowner to
89  * 					new cache entry (aka reference it)
90  * 			  else if first seqid# in sequence
91  * 				- increment seqid_refcnt on new cache entry
92  * 				- set pointer from Openowner/Lockowner to
93  * 					new cache entry (aka reference it)
94  *
95  * 	At end of RPC processing:
96  * 		- if seqid_refcnt > 0 OR flagged non-idempotent on new
97  * 			cache entry
98  * 			- save reply in cache entry
99  * 			- calculate checksum on first N bytes of NFS XDR
100  * 				request
101  * 			- note op and length of XDR request (in bytes)
102  * 			- timestamp it
103  * 		  else
104  * 			- free new cache entry
105  * 		- Send reply (noting info for socket activity check, below)
106  *
107  * 	For cache entries saved above:
108  * 		- if saved since seqid_refcnt was > 0
109  * 			- free when seqid_refcnt decrements to 0
110  * 			  (when next one in sequence is processed above, or
111  * 			   when Openowner/Lockowner is discarded)
112  * 		  else { non-idempotent Op(s) }
113  * 			- free when
114  * 				- some further activity observed on same
115  * 					socket
116  * 				  (I'm not yet sure how I'm going to do
117  * 				   this. Maybe look at the TCP connection
118  * 				   to see if the send_tcp_sequence# is well
119  * 				   past sent reply OR K additional RPCs
120  * 				   replied on same socket OR?)
121  * 			  OR
122  * 				- when very old (hours, days, weeks?)
123  *
124  * For UDP (v2, 3 only), pretty much the old way:
125  * - key on <xid, NFS version, RPC#, Client host ip#>
126  *   (at most one entry for each key)
127  *
128  * When a Request arrives:
129  * - if a match with entry via key
130  * 	- if RPC marked In_progress
131  * 		- discard request (don't send reply)
132  * 	  else
133  * 		- reply from cache
134  * 		- timestamp cache entry
135  *   else
136  * 	- add entry to cache, marked In_progress
137  * 	- do RPC
138  * 	- when RPC done
139  * 		- if RPC# non-idempotent
140  * 			- mark entry Done (not In_progress)
141  * 			- save reply
142  * 			- timestamp cache entry
143  * 		  else
144  * 			- free cache entry
145  * 		- send reply
146  *
147  * Later, entries with saved replies are free'd a short time (few minutes)
148  * after reply sent (timestamp).
149  * Reference: Chet Juszczak, "Improving the Performance and Correctness
150  *		of an NFS Server", in Proc. Winter 1989 USENIX Conference,
151  *		pages 53-63. San Diego, February 1989.
152  *	 for the UDP case.
153  * nfsrc_floodlevel is set to the allowable upper limit for saved replies
154  *	for TCP. For V3, a reply won't be saved when the flood level is
155  *	hit. For V4, the non-idempotent Op will return NFSERR_RESOURCE in
156  *	that case. This level should be set high enough that this almost
157  *	never happens.
158  */
159 #ifndef APPLEKEXT
160 #include <fs/nfs/nfsport.h>
161 
162 extern struct nfsstats newnfsstats;
163 NFSCACHEMUTEX;
164 int nfsrc_floodlevel = NFSRVCACHE_FLOODLEVEL, nfsrc_tcpsavedreplies = 0;
165 #endif	/* !APPLEKEXT */
166 
167 static int nfsrc_tcpnonidempotent = 1;
168 static int nfsrc_udphighwater = NFSRVCACHE_UDPHIGHWATER, nfsrc_udpcachesize = 0;
169 static TAILQ_HEAD(, nfsrvcache) nfsrvudplru;
170 static struct nfsrvhashhead nfsrvhashtbl[NFSRVCACHE_HASHSIZE],
171     nfsrvudphashtbl[NFSRVCACHE_HASHSIZE];
172 /*
173  * and the reverse mapping from generic to Version 2 procedure numbers
174  */
175 static int newnfsv2_procid[NFS_V3NPROCS] = {
176 	NFSV2PROC_NULL,
177 	NFSV2PROC_GETATTR,
178 	NFSV2PROC_SETATTR,
179 	NFSV2PROC_LOOKUP,
180 	NFSV2PROC_NOOP,
181 	NFSV2PROC_READLINK,
182 	NFSV2PROC_READ,
183 	NFSV2PROC_WRITE,
184 	NFSV2PROC_CREATE,
185 	NFSV2PROC_MKDIR,
186 	NFSV2PROC_SYMLINK,
187 	NFSV2PROC_CREATE,
188 	NFSV2PROC_REMOVE,
189 	NFSV2PROC_RMDIR,
190 	NFSV2PROC_RENAME,
191 	NFSV2PROC_LINK,
192 	NFSV2PROC_READDIR,
193 	NFSV2PROC_NOOP,
194 	NFSV2PROC_STATFS,
195 	NFSV2PROC_NOOP,
196 	NFSV2PROC_NOOP,
197 	NFSV2PROC_NOOP,
198 };
199 
200 #define	NFSRCUDPHASH(xid) \
201 	(&nfsrvudphashtbl[((xid) + ((xid) >> 24)) % NFSRVCACHE_HASHSIZE])
202 #define	NFSRCHASH(xid) \
203 	(&nfsrvhashtbl[((xid) + ((xid) >> 24)) % NFSRVCACHE_HASHSIZE])
204 #define	TRUE	1
205 #define	FALSE	0
206 #define	NFSRVCACHE_CHECKLEN	100
207 
208 /* True iff the rpc reply is an nfs status ONLY! */
209 static int nfsv2_repstat[NFS_V3NPROCS] = {
210 	FALSE,
211 	FALSE,
212 	FALSE,
213 	FALSE,
214 	FALSE,
215 	FALSE,
216 	FALSE,
217 	FALSE,
218 	FALSE,
219 	FALSE,
220 	TRUE,
221 	TRUE,
222 	TRUE,
223 	TRUE,
224 	FALSE,
225 	TRUE,
226 	FALSE,
227 	FALSE,
228 	FALSE,
229 	FALSE,
230 	FALSE,
231 	FALSE,
232 };
233 
234 /*
235  * Will NFS want to work over IPv6 someday?
236  */
237 #define	NETFAMILY(rp) \
238 		(((rp)->rc_flag & RC_INETIPV6) ? AF_INET6 : AF_INET)
239 
240 /* local functions */
241 static int nfsrc_getudp(struct nfsrv_descript *nd, struct nfsrvcache *newrp);
242 static int nfsrc_gettcp(struct nfsrv_descript *nd, struct nfsrvcache *newrp);
243 static void nfsrc_lock(struct nfsrvcache *rp);
244 static void nfsrc_unlock(struct nfsrvcache *rp);
245 static void nfsrc_wanted(struct nfsrvcache *rp);
246 static void nfsrc_freecache(struct nfsrvcache *rp);
247 static void nfsrc_trimcache(u_int64_t, struct socket *);
248 static int nfsrc_activesocket(struct nfsrvcache *rp, u_int64_t,
249     struct socket *);
250 static int nfsrc_getlenandcksum(mbuf_t m1, u_int16_t *cksum);
251 static void nfsrc_marksametcpconn(u_int64_t);
252 
253 /*
254  * Initialize the server request cache list
255  */
256 APPLESTATIC void
257 nfsrvd_initcache(void)
258 {
259 	int i;
260 	static int inited = 0;
261 
262 	if (inited)
263 		return;
264 	inited = 1;
265 	for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
266 		LIST_INIT(&nfsrvudphashtbl[i]);
267 		LIST_INIT(&nfsrvhashtbl[i]);
268 	}
269 	TAILQ_INIT(&nfsrvudplru);
270 	nfsrc_tcpsavedreplies = 0;
271 	nfsrc_udpcachesize = 0;
272 	newnfsstats.srvcache_tcppeak = 0;
273 	newnfsstats.srvcache_size = 0;
274 }
275 
276 /*
277  * Get a cache entry for this request. Basically just malloc a new one
278  * and then call nfsrc_getudp() or nfsrc_gettcp() to do the rest.
279  * Call nfsrc_trimcache() to clean up the cache before returning.
280  */
281 APPLESTATIC int
282 nfsrvd_getcache(struct nfsrv_descript *nd, struct socket *so)
283 {
284 	struct nfsrvcache *newrp;
285 	int ret;
286 
287 	if (nd->nd_procnum == NFSPROC_NULL)
288 		panic("nfsd cache null");
289 	MALLOC(newrp, struct nfsrvcache *, sizeof (struct nfsrvcache),
290 	    M_NFSRVCACHE, M_WAITOK);
291 	NFSBZERO((caddr_t)newrp, sizeof (struct nfsrvcache));
292 	if (nd->nd_flag & ND_NFSV4)
293 		newrp->rc_flag = RC_NFSV4;
294 	else if (nd->nd_flag & ND_NFSV3)
295 		newrp->rc_flag = RC_NFSV3;
296 	else
297 		newrp->rc_flag = RC_NFSV2;
298 	newrp->rc_xid = nd->nd_retxid;
299 	newrp->rc_proc = nd->nd_procnum;
300 	newrp->rc_sockref = nd->nd_sockref;
301 	newrp->rc_cachetime = nd->nd_tcpconntime;
302 	if (nd->nd_flag & ND_SAMETCPCONN)
303 		newrp->rc_flag |= RC_SAMETCPCONN;
304 	if (nd->nd_nam2 != NULL) {
305 		newrp->rc_flag |= RC_UDP;
306 		ret = nfsrc_getudp(nd, newrp);
307 	} else {
308 		ret = nfsrc_gettcp(nd, newrp);
309 	}
310 	nfsrc_trimcache(nd->nd_sockref, so);
311 	NFSEXITCODE2(0, nd);
312 	return (ret);
313 }
314 
315 /*
316  * For UDP (v2, v3):
317  * - key on <xid, NFS version, RPC#, Client host ip#>
318  *   (at most one entry for each key)
319  */
320 static int
321 nfsrc_getudp(struct nfsrv_descript *nd, struct nfsrvcache *newrp)
322 {
323 	struct nfsrvcache *rp;
324 	struct sockaddr_in *saddr;
325 	struct sockaddr_in6 *saddr6;
326 	struct nfsrvhashhead *hp;
327 	int ret = 0;
328 
329 	hp = NFSRCUDPHASH(newrp->rc_xid);
330 loop:
331 	NFSLOCKCACHE();
332 	LIST_FOREACH(rp, hp, rc_hash) {
333 	    if (newrp->rc_xid == rp->rc_xid &&
334 		newrp->rc_proc == rp->rc_proc &&
335 		(newrp->rc_flag & rp->rc_flag & RC_NFSVERS) &&
336 		nfsaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
337 			if ((rp->rc_flag & RC_LOCKED) != 0) {
338 				rp->rc_flag |= RC_WANTED;
339 				(void)mtx_sleep(rp, NFSCACHEMUTEXPTR,
340 				    (PZERO - 1) | PDROP, "nfsrc", 10 * hz);
341 				goto loop;
342 			}
343 			if (rp->rc_flag == 0)
344 				panic("nfs udp cache0");
345 			rp->rc_flag |= RC_LOCKED;
346 			TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
347 			TAILQ_INSERT_TAIL(&nfsrvudplru, rp, rc_lru);
348 			if (rp->rc_flag & RC_INPROG) {
349 				newnfsstats.srvcache_inproghits++;
350 				NFSUNLOCKCACHE();
351 				ret = RC_DROPIT;
352 			} else if (rp->rc_flag & RC_REPSTATUS) {
353 				/*
354 				 * V2 only.
355 				 */
356 				newnfsstats.srvcache_nonidemdonehits++;
357 				NFSUNLOCKCACHE();
358 				nfsrvd_rephead(nd);
359 				*(nd->nd_errp) = rp->rc_status;
360 				ret = RC_REPLY;
361 				rp->rc_timestamp = NFSD_MONOSEC +
362 					NFSRVCACHE_UDPTIMEOUT;
363 			} else if (rp->rc_flag & RC_REPMBUF) {
364 				newnfsstats.srvcache_nonidemdonehits++;
365 				NFSUNLOCKCACHE();
366 				nd->nd_mreq = m_copym(rp->rc_reply, 0,
367 					M_COPYALL, M_WAITOK);
368 				ret = RC_REPLY;
369 				rp->rc_timestamp = NFSD_MONOSEC +
370 					NFSRVCACHE_UDPTIMEOUT;
371 			} else {
372 				panic("nfs udp cache1");
373 			}
374 			nfsrc_unlock(rp);
375 			free((caddr_t)newrp, M_NFSRVCACHE);
376 			goto out;
377 		}
378 	}
379 	newnfsstats.srvcache_misses++;
380 	newnfsstats.srvcache_size++;
381 	nfsrc_udpcachesize++;
382 
383 	newrp->rc_flag |= RC_INPROG;
384 	saddr = NFSSOCKADDR(nd->nd_nam, struct sockaddr_in *);
385 	if (saddr->sin_family == AF_INET)
386 		newrp->rc_inet = saddr->sin_addr.s_addr;
387 	else if (saddr->sin_family == AF_INET6) {
388 		saddr6 = (struct sockaddr_in6 *)saddr;
389 		NFSBCOPY((caddr_t)&saddr6->sin6_addr, (caddr_t)&newrp->rc_inet6,
390 		    sizeof (struct in6_addr));
391 		newrp->rc_flag |= RC_INETIPV6;
392 	}
393 	LIST_INSERT_HEAD(hp, newrp, rc_hash);
394 	TAILQ_INSERT_TAIL(&nfsrvudplru, newrp, rc_lru);
395 	NFSUNLOCKCACHE();
396 	nd->nd_rp = newrp;
397 	ret = RC_DOIT;
398 
399 out:
400 	NFSEXITCODE2(0, nd);
401 	return (ret);
402 }
403 
404 /*
405  * Update a request cache entry after the rpc has been done
406  */
407 APPLESTATIC struct nfsrvcache *
408 nfsrvd_updatecache(struct nfsrv_descript *nd, struct socket *so)
409 {
410 	struct nfsrvcache *rp;
411 	struct nfsrvcache *retrp = NULL;
412 	mbuf_t m;
413 
414 	rp = nd->nd_rp;
415 	if (!rp)
416 		panic("nfsrvd_updatecache null rp");
417 	nd->nd_rp = NULL;
418 	NFSLOCKCACHE();
419 	nfsrc_lock(rp);
420 	if (!(rp->rc_flag & RC_INPROG))
421 		panic("nfsrvd_updatecache not inprog");
422 	rp->rc_flag &= ~RC_INPROG;
423 	if (rp->rc_flag & RC_UDP) {
424 		TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
425 		TAILQ_INSERT_TAIL(&nfsrvudplru, rp, rc_lru);
426 	}
427 
428 	/*
429 	 * Reply from cache is a special case returned by nfsrv_checkseqid().
430 	 */
431 	if (nd->nd_repstat == NFSERR_REPLYFROMCACHE) {
432 		newnfsstats.srvcache_nonidemdonehits++;
433 		NFSUNLOCKCACHE();
434 		nd->nd_repstat = 0;
435 		if (nd->nd_mreq)
436 			mbuf_freem(nd->nd_mreq);
437 		if (!(rp->rc_flag & RC_REPMBUF))
438 			panic("reply from cache");
439 		nd->nd_mreq = m_copym(rp->rc_reply, 0,
440 		    M_COPYALL, M_WAITOK);
441 		rp->rc_timestamp = NFSD_MONOSEC + NFSRVCACHE_TCPTIMEOUT;
442 		nfsrc_unlock(rp);
443 		goto out;
444 	}
445 
446 	/*
447 	 * If rc_refcnt > 0, save it
448 	 * For UDP, save it if ND_SAVEREPLY is set
449 	 * For TCP, save it if ND_SAVEREPLY and nfsrc_tcpnonidempotent is set
450 	 */
451 	if (nd->nd_repstat != NFSERR_DONTREPLY &&
452 	    (rp->rc_refcnt > 0 ||
453 	     ((nd->nd_flag & ND_SAVEREPLY) && (rp->rc_flag & RC_UDP)) ||
454 	     ((nd->nd_flag & ND_SAVEREPLY) && !(rp->rc_flag & RC_UDP) &&
455 	      nfsrc_tcpsavedreplies <= nfsrc_floodlevel &&
456 	      nfsrc_tcpnonidempotent))) {
457 		if (rp->rc_refcnt > 0) {
458 			if (!(rp->rc_flag & RC_NFSV4))
459 				panic("update_cache refcnt");
460 			rp->rc_flag |= RC_REFCNT;
461 		}
462 		if ((nd->nd_flag & ND_NFSV2) &&
463 		    nfsv2_repstat[newnfsv2_procid[nd->nd_procnum]]) {
464 			rp->rc_status = nd->nd_repstat;
465 			rp->rc_flag |= RC_REPSTATUS;
466 			NFSUNLOCKCACHE();
467 		} else {
468 			if (!(rp->rc_flag & RC_UDP)) {
469 			    nfsrc_tcpsavedreplies++;
470 			    if (nfsrc_tcpsavedreplies >
471 				newnfsstats.srvcache_tcppeak)
472 				newnfsstats.srvcache_tcppeak =
473 				    nfsrc_tcpsavedreplies;
474 			}
475 			NFSUNLOCKCACHE();
476 			m = m_copym(nd->nd_mreq, 0, M_COPYALL, M_WAITOK);
477 			NFSLOCKCACHE();
478 			rp->rc_reply = m;
479 			rp->rc_flag |= RC_REPMBUF;
480 			NFSUNLOCKCACHE();
481 		}
482 		if (rp->rc_flag & RC_UDP) {
483 			rp->rc_timestamp = NFSD_MONOSEC +
484 			    NFSRVCACHE_UDPTIMEOUT;
485 			nfsrc_unlock(rp);
486 		} else {
487 			rp->rc_timestamp = NFSD_MONOSEC +
488 			    NFSRVCACHE_TCPTIMEOUT;
489 			if (rp->rc_refcnt > 0)
490 				nfsrc_unlock(rp);
491 			else
492 				retrp = rp;
493 		}
494 	} else {
495 		nfsrc_freecache(rp);
496 		NFSUNLOCKCACHE();
497 	}
498 
499 out:
500 	nfsrc_trimcache(nd->nd_sockref, so);
501 	NFSEXITCODE2(0, nd);
502 	return (retrp);
503 }
504 
505 /*
506  * Invalidate and, if possible, free an in prog cache entry.
507  * Must not sleep.
508  */
509 APPLESTATIC void
510 nfsrvd_delcache(struct nfsrvcache *rp)
511 {
512 
513 	if (!(rp->rc_flag & RC_INPROG))
514 		panic("nfsrvd_delcache not in prog");
515 	NFSLOCKCACHE();
516 	rp->rc_flag &= ~RC_INPROG;
517 	if (rp->rc_refcnt == 0 && !(rp->rc_flag & RC_LOCKED))
518 		nfsrc_freecache(rp);
519 	NFSUNLOCKCACHE();
520 }
521 
522 /*
523  * Called after nfsrvd_updatecache() once the reply is sent, to update
524  * the entry for nfsrc_activesocket() and unlock it. The argument is
525  * the pointer returned by nfsrvd_updatecache().
526  */
527 APPLESTATIC void
528 nfsrvd_sentcache(struct nfsrvcache *rp, struct socket *so, int err)
529 {
530 	tcp_seq tmp_seq;
531 
532 	if (!(rp->rc_flag & RC_LOCKED))
533 		panic("nfsrvd_sentcache not locked");
534 	if (!err) {
535 		if ((so->so_proto->pr_domain->dom_family != AF_INET &&
536 		     so->so_proto->pr_domain->dom_family != AF_INET6) ||
537 		     so->so_proto->pr_protocol != IPPROTO_TCP)
538 			panic("nfs sent cache");
539 		if (nfsrv_getsockseqnum(so, &tmp_seq)) {
540 			NFSLOCKCACHE();
541 			rp->rc_tcpseq = tmp_seq;
542 			rp->rc_flag |= RC_TCPSEQ;
543 			NFSUNLOCKCACHE();
544 		}
545 	}
546 	nfsrc_unlock(rp);
547 }
548 
549 /*
550  * Get a cache entry for TCP
551  * - key on <xid, nfs version>
552  *   (allow multiple entries for a given key)
553  */
554 static int
555 nfsrc_gettcp(struct nfsrv_descript *nd, struct nfsrvcache *newrp)
556 {
557 	struct nfsrvcache *rp, *nextrp;
558 	int i;
559 	struct nfsrvcache *hitrp;
560 	struct nfsrvhashhead *hp, nfsrc_templist;
561 	int hit, ret = 0;
562 
563 	hp = NFSRCHASH(newrp->rc_xid);
564 	newrp->rc_reqlen = nfsrc_getlenandcksum(nd->nd_mrep, &newrp->rc_cksum);
565 tryagain:
566 	NFSLOCKCACHE();
567 	hit = 1;
568 	LIST_INIT(&nfsrc_templist);
569 	/*
570 	 * Get all the matches and put them on the temp list.
571 	 */
572 	rp = LIST_FIRST(hp);
573 	while (rp != LIST_END(hp)) {
574 		nextrp = LIST_NEXT(rp, rc_hash);
575 		if (newrp->rc_xid == rp->rc_xid &&
576 		    (!(rp->rc_flag & RC_INPROG) ||
577 		     ((newrp->rc_flag & RC_SAMETCPCONN) &&
578 		      newrp->rc_sockref == rp->rc_sockref)) &&
579 		    (newrp->rc_flag & rp->rc_flag & RC_NFSVERS) &&
580 		    newrp->rc_proc == rp->rc_proc &&
581 		    ((newrp->rc_flag & RC_NFSV4) &&
582 		     newrp->rc_sockref != rp->rc_sockref &&
583 		     newrp->rc_cachetime >= rp->rc_cachetime)
584 		    && newrp->rc_reqlen == rp->rc_reqlen &&
585 		    newrp->rc_cksum == rp->rc_cksum) {
586 			LIST_REMOVE(rp, rc_hash);
587 			LIST_INSERT_HEAD(&nfsrc_templist, rp, rc_hash);
588 		}
589 		rp = nextrp;
590 	}
591 
592 	/*
593 	 * Now, use nfsrc_templist to decide if there is a match.
594 	 */
595 	i = 0;
596 	LIST_FOREACH(rp, &nfsrc_templist, rc_hash) {
597 		i++;
598 		if (rp->rc_refcnt > 0) {
599 			hit = 0;
600 			break;
601 		}
602 	}
603 	/*
604 	 * Can be a hit only if one entry left.
605 	 * Note possible hit entry and put nfsrc_templist back on hash
606 	 * list.
607 	 */
608 	if (i != 1)
609 		hit = 0;
610 	hitrp = rp = LIST_FIRST(&nfsrc_templist);
611 	while (rp != LIST_END(&nfsrc_templist)) {
612 		nextrp = LIST_NEXT(rp, rc_hash);
613 		LIST_REMOVE(rp, rc_hash);
614 		LIST_INSERT_HEAD(hp, rp, rc_hash);
615 		rp = nextrp;
616 	}
617 	if (LIST_FIRST(&nfsrc_templist) != LIST_END(&nfsrc_templist))
618 		panic("nfs gettcp cache templist");
619 
620 	if (hit) {
621 		rp = hitrp;
622 		if ((rp->rc_flag & RC_LOCKED) != 0) {
623 			rp->rc_flag |= RC_WANTED;
624 			(void)mtx_sleep(rp, NFSCACHEMUTEXPTR,
625 			    (PZERO - 1) | PDROP, "nfsrc", 10 * hz);
626 			goto tryagain;
627 		}
628 		if (rp->rc_flag == 0)
629 			panic("nfs tcp cache0");
630 		rp->rc_flag |= RC_LOCKED;
631 		if (rp->rc_flag & RC_INPROG) {
632 			newnfsstats.srvcache_inproghits++;
633 			NFSUNLOCKCACHE();
634 			if (newrp->rc_sockref == rp->rc_sockref)
635 				nfsrc_marksametcpconn(rp->rc_sockref);
636 			ret = RC_DROPIT;
637 		} else if (rp->rc_flag & RC_REPSTATUS) {
638 			/*
639 			 * V2 only.
640 			 */
641 			newnfsstats.srvcache_nonidemdonehits++;
642 			NFSUNLOCKCACHE();
643 			if (newrp->rc_sockref == rp->rc_sockref)
644 				nfsrc_marksametcpconn(rp->rc_sockref);
645 			ret = RC_REPLY;
646 			nfsrvd_rephead(nd);
647 			*(nd->nd_errp) = rp->rc_status;
648 			rp->rc_timestamp = NFSD_MONOSEC +
649 				NFSRVCACHE_TCPTIMEOUT;
650 		} else if (rp->rc_flag & RC_REPMBUF) {
651 			newnfsstats.srvcache_nonidemdonehits++;
652 			NFSUNLOCKCACHE();
653 			if (newrp->rc_sockref == rp->rc_sockref)
654 				nfsrc_marksametcpconn(rp->rc_sockref);
655 			ret = RC_REPLY;
656 			nd->nd_mreq = m_copym(rp->rc_reply, 0,
657 				M_COPYALL, M_WAITOK);
658 			rp->rc_timestamp = NFSD_MONOSEC +
659 				NFSRVCACHE_TCPTIMEOUT;
660 		} else {
661 			panic("nfs tcp cache1");
662 		}
663 		nfsrc_unlock(rp);
664 		free((caddr_t)newrp, M_NFSRVCACHE);
665 		goto out;
666 	}
667 	newnfsstats.srvcache_misses++;
668 	newnfsstats.srvcache_size++;
669 
670 	/*
671 	 * For TCP, multiple entries for a key are allowed, so don't
672 	 * chain it into the hash table until done.
673 	 */
674 	newrp->rc_cachetime = NFSD_MONOSEC;
675 	newrp->rc_flag |= RC_INPROG;
676 	LIST_INSERT_HEAD(hp, newrp, rc_hash);
677 	NFSUNLOCKCACHE();
678 	nd->nd_rp = newrp;
679 	ret = RC_DOIT;
680 
681 out:
682 	NFSEXITCODE2(0, nd);
683 	return (ret);
684 }
685 
686 /*
687  * Lock a cache entry.
688  * Also puts a mutex lock on the cache list.
689  */
690 static void
691 nfsrc_lock(struct nfsrvcache *rp)
692 {
693 	NFSCACHELOCKREQUIRED();
694 	while ((rp->rc_flag & RC_LOCKED) != 0) {
695 		rp->rc_flag |= RC_WANTED;
696 		(void)mtx_sleep(rp, NFSCACHEMUTEXPTR, PZERO - 1,
697 		    "nfsrc", 0);
698 	}
699 	rp->rc_flag |= RC_LOCKED;
700 }
701 
702 /*
703  * Unlock a cache entry.
704  */
705 static void
706 nfsrc_unlock(struct nfsrvcache *rp)
707 {
708 
709 	NFSLOCKCACHE();
710 	rp->rc_flag &= ~RC_LOCKED;
711 	nfsrc_wanted(rp);
712 	NFSUNLOCKCACHE();
713 }
714 
715 /*
716  * Wakeup anyone wanting entry.
717  */
718 static void
719 nfsrc_wanted(struct nfsrvcache *rp)
720 {
721 	if (rp->rc_flag & RC_WANTED) {
722 		rp->rc_flag &= ~RC_WANTED;
723 		wakeup((caddr_t)rp);
724 	}
725 }
726 
727 /*
728  * Free up the entry.
729  * Must not sleep.
730  */
731 static void
732 nfsrc_freecache(struct nfsrvcache *rp)
733 {
734 
735 	NFSCACHELOCKREQUIRED();
736 	LIST_REMOVE(rp, rc_hash);
737 	if (rp->rc_flag & RC_UDP) {
738 		TAILQ_REMOVE(&nfsrvudplru, rp, rc_lru);
739 		nfsrc_udpcachesize--;
740 	}
741 	nfsrc_wanted(rp);
742 	if (rp->rc_flag & RC_REPMBUF) {
743 		mbuf_freem(rp->rc_reply);
744 		if (!(rp->rc_flag & RC_UDP))
745 			nfsrc_tcpsavedreplies--;
746 	}
747 	FREE((caddr_t)rp, M_NFSRVCACHE);
748 	newnfsstats.srvcache_size--;
749 }
750 
751 /*
752  * Clean out the cache. Called when nfsserver module is unloaded.
753  */
754 APPLESTATIC void
755 nfsrvd_cleancache(void)
756 {
757 	struct nfsrvcache *rp, *nextrp;
758 	int i;
759 
760 	NFSLOCKCACHE();
761 	for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
762 		LIST_FOREACH_SAFE(rp, &nfsrvhashtbl[i], rc_hash, nextrp) {
763 			nfsrc_freecache(rp);
764 		}
765 	}
766 	for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
767 		LIST_FOREACH_SAFE(rp, &nfsrvudphashtbl[i], rc_hash, nextrp) {
768 			nfsrc_freecache(rp);
769 		}
770 	}
771 	newnfsstats.srvcache_size = 0;
772 	nfsrc_tcpsavedreplies = 0;
773 	NFSUNLOCKCACHE();
774 }
775 
776 /*
777  * The basic rule is to get rid of entries that are expired.
778  */
779 static void
780 nfsrc_trimcache(u_int64_t sockref, struct socket *so)
781 {
782 	struct nfsrvcache *rp, *nextrp;
783 	int i;
784 
785 	NFSLOCKCACHE();
786 	TAILQ_FOREACH_SAFE(rp, &nfsrvudplru, rc_lru, nextrp) {
787 		if (!(rp->rc_flag & (RC_INPROG|RC_LOCKED|RC_WANTED))
788 		     && rp->rc_refcnt == 0
789 		     && ((rp->rc_flag & RC_REFCNT) ||
790 			 NFSD_MONOSEC > rp->rc_timestamp ||
791 			 nfsrc_udpcachesize > nfsrc_udphighwater))
792 			nfsrc_freecache(rp);
793 	}
794 	for (i = 0; i < NFSRVCACHE_HASHSIZE; i++) {
795 		LIST_FOREACH_SAFE(rp, &nfsrvhashtbl[i], rc_hash, nextrp) {
796 			if (!(rp->rc_flag & (RC_INPROG|RC_LOCKED|RC_WANTED))
797 			     && rp->rc_refcnt == 0
798 			     && ((rp->rc_flag & RC_REFCNT) ||
799 				 NFSD_MONOSEC > rp->rc_timestamp ||
800 				 nfsrc_activesocket(rp, sockref, so)))
801 				nfsrc_freecache(rp);
802 		}
803 	}
804 	NFSUNLOCKCACHE();
805 }
806 
807 /*
808  * Add a seqid# reference to the cache entry.
809  */
810 APPLESTATIC void
811 nfsrvd_refcache(struct nfsrvcache *rp)
812 {
813 
814 	NFSLOCKCACHE();
815 	if (rp->rc_refcnt < 0)
816 		panic("nfs cache refcnt");
817 	rp->rc_refcnt++;
818 	NFSUNLOCKCACHE();
819 }
820 
821 /*
822  * Dereference a seqid# cache entry.
823  */
824 APPLESTATIC void
825 nfsrvd_derefcache(struct nfsrvcache *rp)
826 {
827 
828 	NFSLOCKCACHE();
829 	if (rp->rc_refcnt <= 0)
830 		panic("nfs cache derefcnt");
831 	rp->rc_refcnt--;
832 	if (rp->rc_refcnt == 0 && !(rp->rc_flag & (RC_LOCKED | RC_INPROG)))
833 		nfsrc_freecache(rp);
834 	NFSUNLOCKCACHE();
835 }
836 
837 /*
838  * Check to see if the socket is active.
839  * Return 1 if the reply has been received/acknowledged by the client,
840  * 0 otherwise.
841  * XXX - Uses tcp internals.
842  */
843 static int
844 nfsrc_activesocket(struct nfsrvcache *rp, u_int64_t cur_sockref,
845     struct socket *cur_so)
846 {
847 	int ret = 0;
848 
849 	if (!(rp->rc_flag & RC_TCPSEQ))
850 		return (ret);
851 	/*
852 	 * If the sockref is the same, it is the same TCP connection.
853 	 */
854 	if (cur_sockref == rp->rc_sockref)
855 		ret = nfsrv_checksockseqnum(cur_so, rp->rc_tcpseq);
856 	return (ret);
857 }
858 
859 /*
860  * Calculate the length of the mbuf list and a checksum on the first up to
861  * NFSRVCACHE_CHECKLEN bytes.
862  */
863 static int
864 nfsrc_getlenandcksum(mbuf_t m1, u_int16_t *cksum)
865 {
866 	int len = 0, cklen;
867 	mbuf_t m;
868 
869 	m = m1;
870 	while (m) {
871 		len += mbuf_len(m);
872 		m = mbuf_next(m);
873 	}
874 	cklen = (len > NFSRVCACHE_CHECKLEN) ? NFSRVCACHE_CHECKLEN : len;
875 	*cksum = in_cksum(m1, cklen);
876 	return (len);
877 }
878 
879 /*
880  * Mark a TCP connection that is seeing retries. Should never happen for
881  * NFSv4.
882  */
883 static void
884 nfsrc_marksametcpconn(u_int64_t sockref)
885 {
886 }
887 
888