Print this page
NEX-16052 NFS clients cannot reconnect to server after network loss (autofs)
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
Reviewed by: Joyce McIntosh <joyce.mcintosh@nexenta.com>
Reviewed by: Rick McNeal <rick.mcneal@nexenta.com>
Reviewed by: Gordon Ross <gordon.ross@nexenta.com>
NEX-4123 xdrmblk_getpos() is unreliable
Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com>
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
re #13613 rb4516 Tunables needs volatile keyword
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/rpc/clnt_cots.c
+++ new/usr/src/uts/common/rpc/clnt_cots.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
|
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 - * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
23 + * Copyright 2018 Nexenta Systems, Inc. All rights reserved.
24 24 * Copyright (c) 2016 by Delphix. All rights reserved.
25 25 */
26 26
27 27 /*
28 28 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
29 29 * Use is subject to license terms.
30 30 */
31 31
32 32 /*
33 33 * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T
34 34 * All Rights Reserved
35 35 */
36 36
37 37 /*
38 38 * Portions of this source code were derived from Berkeley 4.3 BSD
39 39 * under license from the Regents of the University of California.
40 40 */
41 41
42 42
43 43 /*
44 44 * Implements a kernel based, client side RPC over Connection Oriented
45 45 * Transports (COTS).
46 46 */
47 47
48 48 /*
49 49 * Much of this file has been re-written to let NFS work better over slow
50 50 * transports. A description follows.
51 51 *
52 52 * One of the annoying things about kRPC/COTS is that it will temporarily
53 53 * create more than one connection between a client and server. This
54 54 * happens because when a connection is made, the end-points entry in the
55 55 * linked list of connections (headed by cm_hd), is removed so that other
56 56 * threads don't mess with it. Went ahead and bit the bullet by keeping
57 57 * the endpoint on the connection list and introducing state bits,
58 58 * condition variables etc. to the connection entry data structure (struct
59 59 * cm_xprt).
60 60 *
61 61 * Here is a summary of the changes to cm-xprt:
62 62 *
63 63 * x_ctime is the timestamp of when the endpoint was last
64 64 * connected or disconnected. If an end-point is ever disconnected
65 65 * or re-connected, then any outstanding RPC request is presumed
66 66 * lost, telling clnt_cots_kcallit that it needs to re-send the
67 67 * request, not just wait for the original request's reply to
68 68 * arrive.
69 69 *
70 70 * x_thread flag which tells us if a thread is doing a connection attempt.
71 71 *
72 72 * x_waitdis flag which tells us we are waiting a disconnect ACK.
73 73 *
74 74 * x_needdis flag which tells us we need to send a T_DISCONN_REQ
75 75 * to kill the connection.
76 76 *
77 77 * x_needrel flag which tells us we need to send a T_ORDREL_REQ to
78 78 * gracefully close the connection.
79 79 *
80 80 * #defined bitmasks for the all the b_* bits so that more
81 81 * efficient (and at times less clumsy) masks can be used to
82 82 * manipulated state in cases where multiple bits have to
83 83 * set/cleared/checked in the same critical section.
84 84 *
85 85 * x_conn_cv and x_dis-_cv are new condition variables to let
86 86 * threads knows when the connection attempt is done, and to let
87 87 * the connecting thread know when the disconnect handshake is
88 88 * done.
89 89 *
90 90 * Added the CONN_HOLD() macro so that all reference holds have the same
91 91 * look and feel.
92 92 *
93 93 * In the private (cku_private) portion of the client handle,
94 94 *
95 95 * cku_flags replaces the cku_sent a boolean. cku_flags keeps
96 96 * track of whether a request as been sent, and whether the
97 97 * client's handles call record is on the dispatch list (so that
98 98 * the reply can be matched by XID to the right client handle).
99 99 * The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit()
100 100 * and still have the response find the right client handle so
101 101 * that the retry of CLNT_CALL() gets the result. Testing, found
102 102 * situations where if the timeout was increased, performance
103 103 * degraded. This was due to us hitting a window where the thread
104 104 * was back in rfscall() (probably printing server not responding)
105 105 * while the response came back but no place to put it.
106 106 *
107 107 * cku_ctime is just a cache of x_ctime. If they match,
108 108 * clnt_cots_kcallit() won't to send a retry (unless the maximum
109 109 * receive count limit as been reached). If the don't match, then
110 110 * we assume the request has been lost, and a retry of the request
111 111 * is needed.
112 112 *
113 113 * cku_recv_attempts counts the number of receive count attempts
114 114 * after one try is sent on the wire.
115 115 *
116 116 * Added the clnt_delay() routine so that interruptible and
117 117 * noninterruptible delays are possible.
118 118 *
119 119 * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to
120 120 * control how long the client delays before returned after getting
121 121 * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash
122 122 * a server that may be booting and not yet started nfsd.
123 123 *
124 124 * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable)
125 125 * Why don't we just wait forever (receive an infinite # of times)?
126 126 * Because the server may have rebooted. More insidious is that some
127 127 * servers (ours) will drop NFS/TCP requests in some cases. This is bad,
128 128 * but it is a reality.
129 129 *
130 130 * The case of a server doing orderly release really messes up the
131 131 * client's recovery, especially if the server's TCP implementation is
132 132 * buggy. It was found was that the kRPC/COTS client was breaking some
133 133 * TPI rules, such as not waiting for the acknowledgement of a
134 134 * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and
135 135 * T_DISCON_REQ in clnt_dispatch_notifyall()).
136 136 *
137 137 * One of things that we've seen is that a kRPC TCP endpoint goes into
138 138 * TIMEWAIT and a thus a reconnect takes a long time to satisfy because
139 139 * that the TIMEWAIT state takes a while to finish. If a server sends a
140 140 * T_ORDREL_IND, there is little point in an RPC client doing a
141 141 * T_ORDREL_REQ, because the RPC request isn't going to make it (the
142 142 * server is saying that it won't accept any more data). So kRPC was
143 143 * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the
144 144 * connection skips the TIMEWAIT state and goes straight to a bound state
145 145 * that kRPC can quickly switch to connected.
146 146 *
147 147 * Code that issues TPI request must use waitforack() to wait for the
148 148 * corresponding ack (assuming there is one) in any future modifications.
149 149 * This works around problems that may be introduced by breaking TPI rules
150 150 * (by submitting new calls before earlier requests have been acked) in the
151 151 * case of a signal or other early return. waitforack() depends on
152 152 * clnt_dispatch_notifyconn() to issue the wakeup when the ack
153 153 * arrives, so adding new TPI calls may require corresponding changes
154 154 * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on
155 155 * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure
156 156 * not to set it too low or TPI ACKS will be lost.
157 157 */
158 158
159 159 #include <sys/param.h>
160 160 #include <sys/types.h>
161 161 #include <sys/user.h>
162 162 #include <sys/systm.h>
163 163 #include <sys/sysmacros.h>
164 164 #include <sys/proc.h>
165 165 #include <sys/socket.h>
166 166 #include <sys/file.h>
167 167 #include <sys/stream.h>
168 168 #include <sys/strsubr.h>
169 169 #include <sys/stropts.h>
170 170 #include <sys/strsun.h>
171 171 #include <sys/timod.h>
172 172 #include <sys/tiuser.h>
173 173 #include <sys/tihdr.h>
174 174 #include <sys/t_kuser.h>
175 175 #include <sys/fcntl.h>
176 176 #include <sys/errno.h>
177 177 #include <sys/kmem.h>
178 178 #include <sys/debug.h>
179 179 #include <sys/systm.h>
180 180 #include <sys/kstat.h>
181 181 #include <sys/t_lock.h>
182 182 #include <sys/ddi.h>
183 183 #include <sys/cmn_err.h>
184 184 #include <sys/time.h>
185 185 #include <sys/isa_defs.h>
186 186 #include <sys/callb.h>
187 187 #include <sys/sunddi.h>
188 188 #include <sys/atomic.h>
189 189 #include <sys/sdt.h>
190 190
191 191 #include <netinet/in.h>
192 192 #include <netinet/tcp.h>
193 193
194 194 #include <rpc/types.h>
195 195 #include <rpc/xdr.h>
196 196 #include <rpc/auth.h>
197 197 #include <rpc/clnt.h>
198 198 #include <rpc/rpc_msg.h>
199 199
200 200 #define COTS_DEFAULT_ALLOCSIZE 2048
201 201
202 202 #define WIRE_HDR_SIZE 20 /* serialized call header, sans proc number */
203 203 #define MSG_OFFSET 128 /* offset of call into the mblk */
204 204
205 205 const char *kinet_ntop6(uchar_t *, char *, size_t);
206 206
207 207 static int clnt_cots_ksettimers(CLIENT *, struct rpc_timers *,
208 208 struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t);
209 209 static enum clnt_stat clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t,
210 210 caddr_t, xdrproc_t, caddr_t, struct timeval);
211 211 static void clnt_cots_kabort(CLIENT *);
212 212 static void clnt_cots_kerror(CLIENT *, struct rpc_err *);
213 213 static bool_t clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t);
214 214 static void clnt_cots_kdestroy(CLIENT *);
215 215 static bool_t clnt_cots_kcontrol(CLIENT *, int, char *);
216 216
217 217
218 218 /* List of transports managed by the connection manager. */
219 219 struct cm_xprt {
220 220 TIUSER *x_tiptr; /* transport handle */
221 221 queue_t *x_wq; /* send queue */
222 222 clock_t x_time; /* last time we handed this xprt out */
223 223 clock_t x_ctime; /* time we went to CONNECTED */
224 224 int x_tidu_size; /* TIDU size of this transport */
225 225 union {
226 226 struct {
227 227 unsigned int
228 228 #ifdef _BIT_FIELDS_HTOL
229 229 b_closing: 1, /* we've sent a ord rel on this conn */
230 230 b_dead: 1, /* transport is closed or disconn */
231 231 b_doomed: 1, /* too many conns, let this go idle */
232 232 b_connected: 1, /* this connection is connected */
233 233
234 234 b_ordrel: 1, /* do an orderly release? */
235 235 b_thread: 1, /* thread doing connect */
236 236 b_waitdis: 1, /* waiting for disconnect ACK */
237 237 b_needdis: 1, /* need T_DISCON_REQ */
238 238
239 239 b_needrel: 1, /* need T_ORDREL_REQ */
240 240 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */
241 241 /* disconnect during connect */
242 242
243 243 b_pad: 22;
244 244
245 245 #endif
246 246
247 247 #ifdef _BIT_FIELDS_LTOH
248 248 b_pad: 22,
249 249
250 250 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */
251 251 /* disconnect during connect */
252 252 b_needrel: 1, /* need T_ORDREL_REQ */
253 253
254 254 b_needdis: 1, /* need T_DISCON_REQ */
255 255 b_waitdis: 1, /* waiting for disconnect ACK */
256 256 b_thread: 1, /* thread doing connect */
257 257 b_ordrel: 1, /* do an orderly release? */
258 258
259 259 b_connected: 1, /* this connection is connected */
260 260 b_doomed: 1, /* too many conns, let this go idle */
261 261 b_dead: 1, /* transport is closed or disconn */
262 262 b_closing: 1; /* we've sent a ord rel on this conn */
263 263 #endif
264 264 } bit; unsigned int word;
265 265
266 266 #define x_closing x_state.bit.b_closing
267 267 #define x_dead x_state.bit.b_dead
268 268 #define x_doomed x_state.bit.b_doomed
269 269 #define x_connected x_state.bit.b_connected
270 270
271 271 #define x_ordrel x_state.bit.b_ordrel
272 272 #define x_thread x_state.bit.b_thread
273 273 #define x_waitdis x_state.bit.b_waitdis
274 274 #define x_needdis x_state.bit.b_needdis
275 275
276 276 #define x_needrel x_state.bit.b_needrel
277 277 #define x_early_disc x_state.bit.b_early_disc
278 278
279 279 #define x_state_flags x_state.word
280 280
281 281 #define X_CLOSING 0x80000000
282 282 #define X_DEAD 0x40000000
283 283 #define X_DOOMED 0x20000000
284 284 #define X_CONNECTED 0x10000000
285 285
286 286 #define X_ORDREL 0x08000000
287 287 #define X_THREAD 0x04000000
288 288 #define X_WAITDIS 0x02000000
289 289 #define X_NEEDDIS 0x01000000
290 290
291 291 #define X_NEEDREL 0x00800000
292 292 #define X_EARLYDISC 0x00400000
293 293
294 294 #define X_BADSTATES (X_CLOSING | X_DEAD | X_DOOMED)
295 295
296 296 } x_state;
297 297 int x_ref; /* number of users of this xprt */
298 298 int x_family; /* address family of transport */
299 299 dev_t x_rdev; /* device number of transport */
300 300 struct cm_xprt *x_next;
301 301
302 302 struct netbuf x_server; /* destination address */
303 303 struct netbuf x_src; /* src address (for retries) */
304 304 kmutex_t x_lock; /* lock on this entry */
305 305 kcondvar_t x_cv; /* to signal when can be closed */
306 306 kcondvar_t x_conn_cv; /* to signal when connection attempt */
307 307 /* is complete */
308 308 kstat_t *x_ksp;
309 309
310 310 kcondvar_t x_dis_cv; /* to signal when disconnect attempt */
311 311 /* is complete */
312 312 zoneid_t x_zoneid; /* zone this xprt belongs to */
313 313 };
314 314
315 315 typedef struct cm_kstat_xprt {
316 316 kstat_named_t x_wq;
317 317 kstat_named_t x_server;
318 318 kstat_named_t x_family;
319 319 kstat_named_t x_rdev;
320 320 kstat_named_t x_time;
321 321 kstat_named_t x_state;
322 322 kstat_named_t x_ref;
323 323 kstat_named_t x_port;
324 324 } cm_kstat_xprt_t;
325 325
326 326 static cm_kstat_xprt_t cm_kstat_template = {
327 327 { "write_queue", KSTAT_DATA_UINT32 },
328 328 { "server", KSTAT_DATA_STRING },
329 329 { "addr_family", KSTAT_DATA_UINT32 },
330 330 { "device", KSTAT_DATA_UINT32 },
331 331 { "time_stamp", KSTAT_DATA_UINT32 },
332 332 { "status", KSTAT_DATA_UINT32 },
333 333 { "ref_count", KSTAT_DATA_INT32 },
334 334 { "port", KSTAT_DATA_UINT32 },
335 335 };
336 336
337 337 /*
338 338 * The inverse of this is connmgr_release().
339 339 */
340 340 #define CONN_HOLD(Cm_entry) {\
341 341 mutex_enter(&(Cm_entry)->x_lock); \
342 342 (Cm_entry)->x_ref++; \
343 343 mutex_exit(&(Cm_entry)->x_lock); \
344 344 }
345 345
346 346
347 347 /*
348 348 * Private data per rpc handle. This structure is allocated by
349 349 * clnt_cots_kcreate, and freed by clnt_cots_kdestroy.
350 350 */
351 351 typedef struct cku_private_s {
352 352 CLIENT cku_client; /* client handle */
353 353 calllist_t cku_call; /* for dispatching calls */
354 354 struct rpc_err cku_err; /* error status */
355 355
356 356 struct netbuf cku_srcaddr; /* source address for retries */
357 357 int cku_addrfmly; /* for binding port */
358 358 struct netbuf cku_addr; /* remote address */
359 359 dev_t cku_device; /* device to use */
360 360 uint_t cku_flags;
361 361 #define CKU_ONQUEUE 0x1
362 362 #define CKU_SENT 0x2
363 363
364 364 bool_t cku_progress; /* for CLSET_PROGRESS */
365 365 uint32_t cku_xid; /* current XID */
366 366 clock_t cku_ctime; /* time stamp of when */
367 367 /* connection was created */
368 368 uint_t cku_recv_attempts;
369 369 XDR cku_outxdr; /* xdr routine for output */
370 370 XDR cku_inxdr; /* xdr routine for input */
371 371 char cku_rpchdr[WIRE_HDR_SIZE + 4];
372 372 /* pre-serialized rpc header */
373 373
374 374 uint_t cku_outbuflen; /* default output mblk length */
375 375 struct cred *cku_cred; /* credentials */
376 376 bool_t cku_nodelayonerr;
377 377 /* for CLSET_NODELAYONERR */
378 378 int cku_useresvport; /* Use reserved port */
379 379 struct rpc_cots_client *cku_stats; /* stats for zone */
380 380 } cku_private_t;
381 381
382 382 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *,
383 383 const struct timeval *, struct netbuf *, int, struct netbuf *,
384 384 struct rpc_err *, bool_t, bool_t, cred_t *);
385 385
386 386 static bool_t connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *,
387 387 int, calllist_t *, int *, bool_t reconnect,
388 388 const struct timeval *, bool_t, cred_t *);
389 389
390 390 static void *connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
391 391 t_uscalar_t length, uint_t align_size);
392 392 static bool_t connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr);
393 393 static bool_t connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
394 394 calllist_t *e, cred_t *cr);
395 395 static bool_t connmgr_setopt_int(queue_t *wq, int level, int name, int val,
396 396 calllist_t *e, cred_t *cr);
397 397 static bool_t connmgr_setopt(queue_t *, int, int, calllist_t *, cred_t *cr);
398 398 static void connmgr_sndrel(struct cm_xprt *);
399 399 static void connmgr_snddis(struct cm_xprt *);
400 400 static void connmgr_close(struct cm_xprt *);
401 401 static void connmgr_release(struct cm_xprt *);
402 402 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *,
403 403 cku_private_t *);
404 404
405 405 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *,
406 406 struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t,
407 407 bool_t, int, cred_t *);
408 408
409 409 static void connmgr_cancelconn(struct cm_xprt *);
410 410 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *,
411 411 bool_t);
412 412 static void connmgr_dis_and_wait(struct cm_xprt *);
413 413
414 414 static int clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t,
415 415 uint_t);
416 416
417 417 static int clnt_delay(clock_t, bool_t);
418 418
419 419 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t);
420 420
421 421 /*
422 422 * Operations vector for TCP/IP based RPC
423 423 */
424 424 static struct clnt_ops tcp_ops = {
425 425 clnt_cots_kcallit, /* do rpc call */
426 426 clnt_cots_kabort, /* abort call */
427 427 clnt_cots_kerror, /* return error status */
428 428 clnt_cots_kfreeres, /* free results */
429 429 clnt_cots_kdestroy, /* destroy rpc handle */
430 430 clnt_cots_kcontrol, /* the ioctl() of rpc */
431 431 clnt_cots_ksettimers, /* set retry timers */
432 432 };
433 433
434 434 static int rpc_kstat_instance = 0; /* keeps the current instance */
435 435 /* number for the next kstat_create */
436 436
437 437 static struct cm_xprt *cm_hd = NULL;
438 438 static kmutex_t connmgr_lock; /* for connection mngr's list of transports */
439 439
440 440 extern kmutex_t clnt_max_msg_lock;
441 441
442 442 static calllist_t *clnt_pending = NULL;
443 443 extern kmutex_t clnt_pending_lock;
444 444
445 445 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE;
446 446
447 447 static call_table_t *cots_call_ht;
448 448
449 449 static const struct rpc_cots_client {
450 450 kstat_named_t rccalls;
451 451 kstat_named_t rcbadcalls;
452 452 kstat_named_t rcbadxids;
453 453 kstat_named_t rctimeouts;
454 454 kstat_named_t rcnewcreds;
455 455 kstat_named_t rcbadverfs;
456 456 kstat_named_t rctimers;
457 457 kstat_named_t rccantconn;
458 458 kstat_named_t rcnomem;
459 459 kstat_named_t rcintrs;
460 460 } cots_rcstat_tmpl = {
461 461 { "calls", KSTAT_DATA_UINT64 },
462 462 { "badcalls", KSTAT_DATA_UINT64 },
463 463 { "badxids", KSTAT_DATA_UINT64 },
464 464 { "timeouts", KSTAT_DATA_UINT64 },
465 465 { "newcreds", KSTAT_DATA_UINT64 },
466 466 { "badverfs", KSTAT_DATA_UINT64 },
|
↓ open down ↓ |
433 lines elided |
↑ open up ↑ |
467 467 { "timers", KSTAT_DATA_UINT64 },
468 468 { "cantconn", KSTAT_DATA_UINT64 },
469 469 { "nomem", KSTAT_DATA_UINT64 },
470 470 { "interrupts", KSTAT_DATA_UINT64 }
471 471 };
472 472
473 473 #define COTSRCSTAT_INCR(p, x) \
474 474 atomic_inc_64(&(p)->x.value.ui64)
475 475
476 476 #define CLNT_MAX_CONNS 1 /* concurrent connections between clnt/srvr */
477 -int clnt_max_conns = CLNT_MAX_CONNS;
477 +volatile int clnt_max_conns = CLNT_MAX_CONNS;
478 478
479 479 #define CLNT_MIN_TIMEOUT 10 /* seconds to wait after we get a */
480 480 /* connection reset */
481 481 #define CLNT_MIN_CONNTIMEOUT 5 /* seconds to wait for a connection */
482 482
483 483
484 484 int clnt_cots_min_tout = CLNT_MIN_TIMEOUT;
485 485 int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT;
486 486
487 487 /*
488 488 * Limit the number of times we will attempt to receive a reply without
489 489 * re-sending a response.
490 490 */
491 491 #define CLNT_MAXRECV_WITHOUT_RETRY 3
492 492 uint_t clnt_cots_maxrecv = CLNT_MAXRECV_WITHOUT_RETRY;
493 493
494 494 uint_t *clnt_max_msg_sizep;
495 495 void (*clnt_stop_idle)(queue_t *wq);
496 496
497 497 #define ptoh(p) (&((p)->cku_client))
498 498 #define htop(h) ((cku_private_t *)((h)->cl_private))
499 499
500 500 /*
501 501 * Times to retry
502 502 */
503 503 #define REFRESHES 2 /* authentication refreshes */
504 504
505 505 /*
506 506 * The following is used to determine the global default behavior for
507 507 * COTS when binding to a local port.
508 508 *
509 509 * If the value is set to 1 the default will be to select a reserved
510 510 * (aka privileged) port, if the value is zero the default will be to
511 511 * use non-reserved ports. Users of kRPC may override this by using
512 512 * CLNT_CONTROL() and CLSET_BINDRESVPORT.
513 513 */
514 514 int clnt_cots_do_bindresvport = 1;
515 515
516 516 static zone_key_t zone_cots_key;
517 517
518 518 /*
519 519 * Defaults TCP send and receive buffer size for RPC connections.
520 520 * These values can be tuned by /etc/system.
521 521 */
522 522 int rpc_send_bufsz = 1024*1024;
523 523 int rpc_recv_bufsz = 1024*1024;
524 524 /*
525 525 * To use system-wide default for TCP send and receive buffer size,
526 526 * use /etc/system to set rpc_default_tcp_bufsz to 1:
527 527 *
528 528 * set rpcmod:rpc_default_tcp_bufsz=1
529 529 */
530 530 int rpc_default_tcp_bufsz = 0;
531 531
532 532 /*
533 533 * We need to do this after all kernel threads in the zone have exited.
534 534 */
535 535 /* ARGSUSED */
536 536 static void
537 537 clnt_zone_destroy(zoneid_t zoneid, void *unused)
538 538 {
539 539 struct cm_xprt **cmp;
540 540 struct cm_xprt *cm_entry;
541 541 struct cm_xprt *freelist = NULL;
542 542
543 543 mutex_enter(&connmgr_lock);
544 544 cmp = &cm_hd;
545 545 while ((cm_entry = *cmp) != NULL) {
546 546 if (cm_entry->x_zoneid == zoneid) {
547 547 *cmp = cm_entry->x_next;
548 548 cm_entry->x_next = freelist;
549 549 freelist = cm_entry;
550 550 } else {
551 551 cmp = &cm_entry->x_next;
552 552 }
553 553 }
554 554 mutex_exit(&connmgr_lock);
555 555 while ((cm_entry = freelist) != NULL) {
556 556 freelist = cm_entry->x_next;
557 557 connmgr_close(cm_entry);
558 558 }
559 559 }
560 560
561 561 int
562 562 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
563 563 rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
564 564 {
565 565 CLIENT *h;
566 566 cku_private_t *p;
567 567 struct rpc_msg call_msg;
568 568 struct rpcstat *rpcstat;
569 569
570 570 RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
571 571
572 572 rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
573 573 ASSERT(rpcstat != NULL);
574 574
575 575 /* Allocate and intialize the client handle. */
576 576 p = kmem_zalloc(sizeof (*p), KM_SLEEP);
577 577
578 578 h = ptoh(p);
579 579
580 580 h->cl_private = (caddr_t)p;
581 581 h->cl_auth = authkern_create();
582 582 h->cl_ops = &tcp_ops;
583 583
584 584 cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
585 585 mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
586 586
587 587 /*
588 588 * If the current sanity check size in rpcmod is smaller
589 589 * than the size needed, then increase the sanity check.
590 590 */
591 591 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
592 592 max_msgsize > *clnt_max_msg_sizep) {
593 593 mutex_enter(&clnt_max_msg_lock);
594 594 if (max_msgsize > *clnt_max_msg_sizep)
595 595 *clnt_max_msg_sizep = max_msgsize;
596 596 mutex_exit(&clnt_max_msg_lock);
597 597 }
598 598
599 599 p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
600 600
601 601 /* Preserialize the call message header */
602 602
603 603 call_msg.rm_xid = 0;
604 604 call_msg.rm_direction = CALL;
605 605 call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
606 606 call_msg.rm_call.cb_prog = prog;
607 607 call_msg.rm_call.cb_vers = vers;
608 608
609 609 xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
610 610
611 611 if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
612 612 XDR_DESTROY(&p->cku_outxdr);
613 613 RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
614 614 "error\n");
615 615 auth_destroy(h->cl_auth);
616 616 kmem_free(p, sizeof (cku_private_t));
617 617 RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
618 618 return (EINVAL); /* XXX */
619 619 }
620 620 XDR_DESTROY(&p->cku_outxdr);
621 621
622 622 /*
623 623 * The zalloc initialized the fields below.
624 624 * p->cku_xid = 0;
625 625 * p->cku_flags = 0;
626 626 * p->cku_srcaddr.len = 0;
627 627 * p->cku_srcaddr.maxlen = 0;
628 628 */
629 629
630 630 p->cku_cred = cred;
631 631 p->cku_device = dev;
632 632 p->cku_addrfmly = family;
633 633 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
634 634 p->cku_addr.maxlen = addr->maxlen;
635 635 p->cku_addr.len = addr->len;
636 636 bcopy(addr->buf, p->cku_addr.buf, addr->len);
637 637 p->cku_stats = rpcstat->rpc_cots_client;
638 638 p->cku_useresvport = -1; /* value is has not been set */
639 639
640 640 *ncl = h;
641 641 return (0);
642 642 }
643 643
644 644 /*ARGSUSED*/
645 645 static void
646 646 clnt_cots_kabort(CLIENT *h)
647 647 {
648 648 }
649 649
650 650 /*
651 651 * Return error info on this handle.
652 652 */
653 653 static void
654 654 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
655 655 {
656 656 /* LINTED pointer alignment */
657 657 cku_private_t *p = htop(h);
658 658
659 659 *err = p->cku_err;
660 660 }
661 661
662 662 /*ARGSUSED*/
663 663 static bool_t
664 664 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
665 665 {
666 666 xdr_free(xdr_res, res_ptr);
667 667
668 668 return (TRUE);
669 669 }
670 670
671 671 static bool_t
672 672 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
673 673 {
674 674 cku_private_t *p = htop(h);
675 675
676 676 switch (cmd) {
677 677 case CLSET_PROGRESS:
678 678 p->cku_progress = TRUE;
679 679 return (TRUE);
680 680
681 681 case CLSET_XID:
682 682 if (arg == NULL)
683 683 return (FALSE);
684 684
685 685 p->cku_xid = *((uint32_t *)arg);
686 686 return (TRUE);
687 687
688 688 case CLGET_XID:
689 689 if (arg == NULL)
690 690 return (FALSE);
691 691
692 692 *((uint32_t *)arg) = p->cku_xid;
693 693 return (TRUE);
694 694
695 695 case CLSET_NODELAYONERR:
696 696 if (arg == NULL)
697 697 return (FALSE);
698 698
699 699 if (*((bool_t *)arg) == TRUE) {
700 700 p->cku_nodelayonerr = TRUE;
701 701 return (TRUE);
702 702 }
703 703 if (*((bool_t *)arg) == FALSE) {
704 704 p->cku_nodelayonerr = FALSE;
705 705 return (TRUE);
706 706 }
707 707 return (FALSE);
708 708
709 709 case CLGET_NODELAYONERR:
710 710 if (arg == NULL)
711 711 return (FALSE);
712 712
713 713 *((bool_t *)arg) = p->cku_nodelayonerr;
714 714 return (TRUE);
715 715
716 716 case CLSET_BINDRESVPORT:
717 717 if (arg == NULL)
718 718 return (FALSE);
719 719
720 720 if (*(int *)arg != 1 && *(int *)arg != 0)
721 721 return (FALSE);
722 722
723 723 p->cku_useresvport = *(int *)arg;
724 724
725 725 return (TRUE);
726 726
727 727 case CLGET_BINDRESVPORT:
728 728 if (arg == NULL)
729 729 return (FALSE);
730 730
731 731 *(int *)arg = p->cku_useresvport;
732 732
733 733 return (TRUE);
734 734
735 735 default:
736 736 return (FALSE);
737 737 }
738 738 }
739 739
740 740 /*
741 741 * Destroy rpc handle. Frees the space used for output buffer,
742 742 * private data, and handle structure.
743 743 */
744 744 static void
745 745 clnt_cots_kdestroy(CLIENT *h)
746 746 {
747 747 /* LINTED pointer alignment */
748 748 cku_private_t *p = htop(h);
749 749 calllist_t *call = &p->cku_call;
750 750
751 751 RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
752 752 RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
753 753
754 754 if (p->cku_flags & CKU_ONQUEUE) {
755 755 RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
756 756 "from dispatch list\n", p->cku_xid);
757 757 call_table_remove(call);
758 758 }
759 759
760 760 if (call->call_reply)
761 761 freemsg(call->call_reply);
762 762 cv_destroy(&call->call_cv);
763 763 mutex_destroy(&call->call_lock);
764 764
765 765 kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
766 766 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
767 767 kmem_free(p, sizeof (*p));
768 768 }
769 769
770 770 static int clnt_cots_pulls;
771 771 #define RM_HDR_SIZE 4 /* record mark header size */
772 772
773 773 /*
774 774 * Call remote procedure.
775 775 */
776 776 static enum clnt_stat
777 777 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
778 778 caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
779 779 {
780 780 /* LINTED pointer alignment */
781 781 cku_private_t *p = htop(h);
782 782 calllist_t *call = &p->cku_call;
783 783 XDR *xdrs;
784 784 struct rpc_msg reply_msg;
785 785 mblk_t *mp;
786 786 #ifdef RPCDEBUG
787 787 clock_t time_sent;
788 788 #endif
789 789 struct netbuf *retryaddr;
790 790 struct cm_xprt *cm_entry = NULL;
791 791 queue_t *wq;
792 792 int len, waitsecs, max_waitsecs;
793 793 int mpsize;
794 794 int refreshes = REFRESHES;
795 795 int interrupted;
796 796 int tidu_size;
797 797 enum clnt_stat status;
798 798 struct timeval cwait;
799 799 bool_t delay_first = FALSE;
800 800 clock_t ticks, now;
801 801
802 802 RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
803 803 COTSRCSTAT_INCR(p->cku_stats, rccalls);
804 804
805 805 RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
806 806 RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
807 807 /*
808 808 * Bug ID 1240234:
809 809 * Look out for zero length timeouts. We don't want to
810 810 * wait zero seconds for a connection to be established.
811 811 */
812 812 if (wait.tv_sec < clnt_cots_min_conntout) {
813 813 cwait.tv_sec = clnt_cots_min_conntout;
814 814 cwait.tv_usec = 0;
815 815 RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
816 816 wait.tv_sec);
817 817 RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
818 818 } else {
819 819 cwait = wait;
820 820 }
821 821
822 822 call_again:
823 823 if (cm_entry) {
824 824 connmgr_release(cm_entry);
825 825 cm_entry = NULL;
826 826 }
827 827
828 828 mp = NULL;
829 829
830 830 /*
831 831 * If the call is not a retry, allocate a new xid and cache it
832 832 * for future retries.
833 833 * Bug ID 1246045:
834 834 * Treat call as a retry for purposes of binding the source
835 835 * port only if we actually attempted to send anything on
836 836 * the previous call.
837 837 */
838 838 if (p->cku_xid == 0) {
839 839 p->cku_xid = alloc_xid();
840 840 call->call_zoneid = rpc_zoneid();
841 841
842 842 /*
843 843 * We need to ASSERT here that our xid != 0 because this
844 844 * determines whether or not our call record gets placed on
845 845 * the hash table or the linked list. By design, we mandate
846 846 * that RPC calls over cots must have xid's != 0, so we can
847 847 * ensure proper management of the hash table.
848 848 */
849 849 ASSERT(p->cku_xid != 0);
850 850
851 851 retryaddr = NULL;
852 852 p->cku_flags &= ~CKU_SENT;
853 853
854 854 if (p->cku_flags & CKU_ONQUEUE) {
855 855 RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
856 856 " one (%p)\n", (void *)call);
857 857 call_table_remove(call);
858 858 p->cku_flags &= ~CKU_ONQUEUE;
859 859 RPCLOG(64, "clnt_cots_kcallit: removing call from "
860 860 "dispatch list because xid was zero (now 0x%x)\n",
861 861 p->cku_xid);
862 862 }
863 863
864 864 if (call->call_reply != NULL) {
865 865 freemsg(call->call_reply);
866 866 call->call_reply = NULL;
867 867 }
868 868 } else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
869 869 retryaddr = NULL;
870 870
871 871 } else if (p->cku_flags & CKU_SENT) {
872 872 retryaddr = &p->cku_srcaddr;
873 873
874 874 } else {
875 875 /*
876 876 * Bug ID 1246045: Nothing was sent, so set retryaddr to
877 877 * NULL and let connmgr_get() bind to any source port it
878 878 * can get.
879 879 */
880 880 retryaddr = NULL;
881 881 }
882 882
883 883 RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
884 884 RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
885 885
886 886 p->cku_err.re_status = RPC_TIMEDOUT;
887 887 p->cku_err.re_errno = p->cku_err.re_terrno = 0;
888 888
889 889 cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
890 890
891 891 if (cm_entry == NULL) {
892 892 RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
893 893 clnt_sperrno(p->cku_err.re_status));
894 894
895 895 /*
896 896 * The reasons why we fail to create a connection are
897 897 * varied. In most cases we don't want the caller to
898 898 * immediately retry. This could have one or more
899 899 * bad effects. This includes flooding the net with
900 900 * connect requests to ports with no listener; a hard
901 901 * kernel loop due to all the "reserved" TCP ports being
902 902 * in use.
903 903 */
904 904 delay_first = TRUE;
905 905
906 906 /*
907 907 * Even if we end up returning EINTR, we still count a
908 908 * a "can't connect", because the connection manager
909 909 * might have been committed to waiting for or timing out on
910 910 * a connection.
911 911 */
912 912 COTSRCSTAT_INCR(p->cku_stats, rccantconn);
913 913 switch (p->cku_err.re_status) {
914 914 case RPC_INTR:
915 915 p->cku_err.re_errno = EINTR;
916 916
917 917 /*
918 918 * No need to delay because a UNIX signal(2)
919 919 * interrupted us. The caller likely won't
920 920 * retry the CLNT_CALL() and even if it does,
921 921 * we assume the caller knows what it is doing.
922 922 */
923 923 delay_first = FALSE;
924 924 break;
925 925
926 926 case RPC_TIMEDOUT:
927 927 p->cku_err.re_errno = ETIMEDOUT;
928 928
929 929 /*
930 930 * No need to delay because timed out already
931 931 * on the connection request and assume that the
932 932 * transport time out is longer than our minimum
933 933 * timeout, or least not too much smaller.
934 934 */
935 935 delay_first = FALSE;
936 936 break;
937 937
938 938 case RPC_SYSTEMERROR:
939 939 case RPC_TLIERROR:
940 940 /*
941 941 * We want to delay here because a transient
942 942 * system error has a better chance of going away
943 943 * if we delay a bit. If it's not transient, then
944 944 * we don't want end up in a hard kernel loop
945 945 * due to retries.
946 946 */
947 947 ASSERT(p->cku_err.re_errno != 0);
948 948 break;
949 949
950 950
951 951 case RPC_CANTCONNECT:
952 952 /*
953 953 * RPC_CANTCONNECT is set on T_ERROR_ACK which
954 954 * implies some error down in the TCP layer or
955 955 * below. If cku_nodelayonerror is set then we
956 956 * assume the caller knows not to try too hard.
957 957 */
958 958 RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
959 959 RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
960 960 RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
961 961 RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
962 962 if (p->cku_nodelayonerr == TRUE)
963 963 delay_first = FALSE;
964 964
965 965 p->cku_err.re_errno = EIO;
966 966
967 967 break;
968 968
969 969 case RPC_XPRTFAILED:
970 970 /*
971 971 * We want to delay here because we likely
972 972 * got a refused connection.
973 973 */
974 974 if (p->cku_err.re_errno == 0)
975 975 p->cku_err.re_errno = EIO;
976 976
977 977 RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
978 978 p->cku_err.re_errno);
979 979
980 980 break;
981 981
982 982 default:
983 983 /*
984 984 * We delay here because it is better to err
985 985 * on the side of caution. If we got here then
986 986 * status could have been RPC_SUCCESS, but we
987 987 * know that we did not get a connection, so
988 988 * force the rpc status to RPC_CANTCONNECT.
989 989 */
990 990 p->cku_err.re_status = RPC_CANTCONNECT;
991 991 p->cku_err.re_errno = EIO;
992 992 break;
993 993 }
994 994 if (delay_first == TRUE)
995 995 ticks = clnt_cots_min_tout * drv_usectohz(1000000);
996 996 goto cots_done;
997 997 }
998 998
999 999 /*
1000 1000 * If we've never sent any request on this connection (send count
1001 1001 * is zero, or the connection has been reset), cache the
1002 1002 * the connection's create time and send a request (possibly a retry)
1003 1003 */
1004 1004 if ((p->cku_flags & CKU_SENT) == 0 ||
1005 1005 p->cku_ctime != cm_entry->x_ctime) {
1006 1006 p->cku_ctime = cm_entry->x_ctime;
1007 1007
1008 1008 } else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
1009 1009 (call->call_reply != NULL ||
1010 1010 p->cku_recv_attempts < clnt_cots_maxrecv)) {
1011 1011
1012 1012 /*
1013 1013 * If we've sent a request and our call is on the dispatch
1014 1014 * queue and we haven't made too many receive attempts, then
1015 1015 * don't re-send, just receive.
1016 1016 */
1017 1017 p->cku_recv_attempts++;
1018 1018 goto read_again;
1019 1019 }
1020 1020
1021 1021 /*
1022 1022 * Now we create the RPC request in a STREAMS message. We have to do
1023 1023 * this after the call to connmgr_get so that we have the correct
1024 1024 * TIDU size for the transport.
1025 1025 */
1026 1026 tidu_size = cm_entry->x_tidu_size;
1027 1027 len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
1028 1028
1029 1029 while ((mp = allocb(len, BPRI_MED)) == NULL) {
1030 1030 if (strwaitbuf(len, BPRI_MED)) {
1031 1031 p->cku_err.re_status = RPC_SYSTEMERROR;
1032 1032 p->cku_err.re_errno = ENOSR;
1033 1033 COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1034 1034 goto cots_done;
1035 1035 }
1036 1036 }
1037 1037 xdrs = &p->cku_outxdr;
1038 1038 xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1039 1039 mpsize = MBLKSIZE(mp);
1040 1040 ASSERT(mpsize >= len);
1041 1041 ASSERT(mp->b_rptr == mp->b_datap->db_base);
1042 1042
1043 1043 /*
1044 1044 * If the size of mblk is not appreciably larger than what we
1045 1045 * asked, then resize the mblk to exactly len bytes. The reason for
1046 1046 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1047 1047 * (from TCP over ethernet), and the arguments to the RPC require
1048 1048 * 2800 bytes. Ideally we want the protocol to render two
1049 1049 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1050 1050 * mblk, and we allocate a second mblk for the remainder, the protocol
1051 1051 * module may generate 3 segments over the wire:
1052 1052 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1053 1053 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1054 1054 * the XDR encoding will generate two ~1400 byte mblks, and the
1055 1055 * protocol module is more likely to produce properly sized segments.
1056 1056 */
1057 1057 if ((mpsize >> 1) <= len)
1058 1058 mp->b_rptr += (mpsize - len);
1059 1059
1060 1060 /*
1061 1061 * Adjust b_rptr to reserve space for the non-data protocol headers
1062 1062 * any downstream modules might like to add, and for the
1063 1063 * record marking header.
1064 1064 */
1065 1065 mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1066 1066
1067 1067 if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1068 1068 /* Copy in the preserialized RPC header information. */
1069 1069 bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1070 1070
1071 1071 /* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1072 1072 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1073 1073 WIRE_HDR_SIZE));
1074 1074
1075 1075 ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1076 1076
1077 1077 /* Serialize the procedure number and the arguments. */
1078 1078 if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1079 1079 (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1080 1080 (!(*xdr_args)(xdrs, argsp))) {
1081 1081 XDR_DESTROY(xdrs);
1082 1082 p->cku_err.re_status = RPC_CANTENCODEARGS;
1083 1083 p->cku_err.re_errno = EIO;
1084 1084 goto cots_done;
1085 1085 }
1086 1086
1087 1087 (*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1088 1088 } else {
1089 1089 uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1090 1090 IXDR_PUT_U_INT32(uproc, procnum);
1091 1091
1092 1092 (*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1093 1093
1094 1094 /* Use XDR_SETPOS() to set the b_wptr. */
1095 1095 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1096 1096
1097 1097 /* Serialize the procedure number and the arguments. */
1098 1098 if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1099 1099 xdrs, xdr_args, argsp)) {
1100 1100 XDR_DESTROY(xdrs);
1101 1101 p->cku_err.re_status = RPC_CANTENCODEARGS;
1102 1102 p->cku_err.re_errno = EIO;
1103 1103 goto cots_done;
1104 1104 }
1105 1105 }
1106 1106
1107 1107 XDR_DESTROY(xdrs);
1108 1108
1109 1109 RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1110 1110 tidu_size);
1111 1111
1112 1112 wq = cm_entry->x_wq;
1113 1113 waitsecs = 0;
1114 1114
1115 1115 dispatch_again:
1116 1116 status = clnt_dispatch_send(wq, mp, call, p->cku_xid,
1117 1117 (p->cku_flags & CKU_ONQUEUE));
1118 1118
1119 1119 if ((status == RPC_CANTSEND) && (call->call_reason == ENOBUFS)) {
1120 1120 /*
1121 1121 * QFULL condition, allow some time for queue to drain
1122 1122 * and try again. Give up after waiting for all timeout
1123 1123 * specified for the call, or zone is going away.
1124 1124 */
1125 1125 max_waitsecs = wait.tv_sec ? wait.tv_sec : clnt_cots_min_tout;
1126 1126 if ((waitsecs++ < max_waitsecs) &&
1127 1127 !(zone_status_get(curproc->p_zone) >=
1128 1128 ZONE_IS_SHUTTING_DOWN)) {
1129 1129
1130 1130 /* wait 1 sec for queue to drain */
1131 1131 if (clnt_delay(drv_usectohz(1000000),
1132 1132 h->cl_nosignal) == EINTR) {
1133 1133 p->cku_err.re_errno = EINTR;
1134 1134 p->cku_err.re_status = RPC_INTR;
1135 1135
1136 1136 goto cots_done;
1137 1137 }
1138 1138
1139 1139 /* and try again */
1140 1140 goto dispatch_again;
1141 1141 }
1142 1142 p->cku_err.re_status = status;
1143 1143 p->cku_err.re_errno = call->call_reason;
1144 1144 DTRACE_PROBE(krpc__e__clntcots__kcallit__cantsend);
1145 1145
1146 1146 goto cots_done;
1147 1147 }
1148 1148
1149 1149 if (waitsecs) {
1150 1150 /* adjust timeout to account for time wait to send */
1151 1151 wait.tv_sec -= waitsecs;
1152 1152 if (wait.tv_sec < 0) {
1153 1153 /* pick up reply on next retry */
1154 1154 wait.tv_sec = 0;
1155 1155 }
1156 1156 DTRACE_PROBE2(clnt_cots__sendwait, CLIENT *, h,
1157 1157 int, waitsecs);
1158 1158 }
1159 1159
1160 1160 RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1161 1161 (uint_t)p->cku_xid);
1162 1162 p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1163 1163 p->cku_recv_attempts = 1;
1164 1164
1165 1165 #ifdef RPCDEBUG
1166 1166 time_sent = ddi_get_lbolt();
1167 1167 #endif
1168 1168
1169 1169 /*
1170 1170 * Wait for a reply or a timeout. If there is no error or timeout,
1171 1171 * (both indicated by call_status), call->call_reply will contain
1172 1172 * the RPC reply message.
1173 1173 */
1174 1174 read_again:
1175 1175 mutex_enter(&call->call_lock);
1176 1176 interrupted = 0;
1177 1177 if (call->call_status == RPC_TIMEDOUT) {
1178 1178 /*
1179 1179 * Indicate that the lwp is not to be stopped while waiting
1180 1180 * for this network traffic. This is to avoid deadlock while
1181 1181 * debugging a process via /proc and also to avoid recursive
1182 1182 * mutex_enter()s due to NFS page faults while stopping
1183 1183 * (NFS holds locks when it calls here).
1184 1184 */
1185 1185 clock_t cv_wait_ret;
1186 1186 clock_t timout;
1187 1187 clock_t oldlbolt;
1188 1188
1189 1189 klwp_t *lwp = ttolwp(curthread);
1190 1190
1191 1191 if (lwp != NULL)
1192 1192 lwp->lwp_nostop++;
1193 1193
1194 1194 oldlbolt = ddi_get_lbolt();
1195 1195 timout = wait.tv_sec * drv_usectohz(1000000) +
1196 1196 drv_usectohz(wait.tv_usec) + oldlbolt;
1197 1197 /*
1198 1198 * Iterate until the call_status is changed to something
1199 1199 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1200 1200 * something <=0 zero. The latter means that we timed
1201 1201 * out.
1202 1202 */
1203 1203 if (h->cl_nosignal)
1204 1204 while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1205 1205 &call->call_lock, timout)) > 0 &&
1206 1206 call->call_status == RPC_TIMEDOUT)
1207 1207 ;
1208 1208 else
1209 1209 while ((cv_wait_ret = cv_timedwait_sig(
1210 1210 &call->call_cv,
1211 1211 &call->call_lock, timout)) > 0 &&
1212 1212 call->call_status == RPC_TIMEDOUT)
1213 1213 ;
1214 1214
1215 1215 switch (cv_wait_ret) {
1216 1216 case 0:
1217 1217 /*
1218 1218 * If we got out of the above loop with
1219 1219 * cv_timedwait_sig() returning 0, then we were
1220 1220 * interrupted regardless what call_status is.
1221 1221 */
1222 1222 interrupted = 1;
1223 1223 break;
1224 1224 case -1:
1225 1225 /* cv_timedwait_sig() timed out */
1226 1226 break;
1227 1227 default:
1228 1228
1229 1229 /*
1230 1230 * We were cv_signaled(). If we didn't
1231 1231 * get a successful call_status and returned
1232 1232 * before time expired, delay up to clnt_cots_min_tout
1233 1233 * seconds so that the caller doesn't immediately
1234 1234 * try to call us again and thus force the
1235 1235 * same condition that got us here (such
1236 1236 * as a RPC_XPRTFAILED due to the server not
1237 1237 * listening on the end-point.
1238 1238 */
1239 1239 if (call->call_status != RPC_SUCCESS) {
1240 1240 clock_t curlbolt;
1241 1241 clock_t diff;
1242 1242
1243 1243 curlbolt = ddi_get_lbolt();
1244 1244 ticks = clnt_cots_min_tout *
1245 1245 drv_usectohz(1000000);
1246 1246 diff = curlbolt - oldlbolt;
1247 1247 if (diff < ticks) {
1248 1248 delay_first = TRUE;
1249 1249 if (diff > 0)
1250 1250 ticks -= diff;
1251 1251 }
1252 1252 }
1253 1253 break;
1254 1254 }
1255 1255
1256 1256 if (lwp != NULL)
1257 1257 lwp->lwp_nostop--;
1258 1258 }
1259 1259 /*
1260 1260 * Get the reply message, if any. This will be freed at the end
1261 1261 * whether or not an error occurred.
1262 1262 */
1263 1263 mp = call->call_reply;
1264 1264 call->call_reply = NULL;
1265 1265
1266 1266 /*
1267 1267 * call_err is the error info when the call is on dispatch queue.
1268 1268 * cku_err is the error info returned to the caller.
1269 1269 * Sync cku_err with call_err for local message processing.
1270 1270 */
1271 1271
1272 1272 status = call->call_status;
1273 1273 p->cku_err = call->call_err;
1274 1274 mutex_exit(&call->call_lock);
1275 1275
1276 1276 if (status != RPC_SUCCESS) {
1277 1277 switch (status) {
1278 1278 case RPC_TIMEDOUT:
1279 1279 now = ddi_get_lbolt();
1280 1280 if (interrupted) {
1281 1281 COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1282 1282 p->cku_err.re_status = RPC_INTR;
1283 1283 p->cku_err.re_errno = EINTR;
1284 1284 RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1285 1285 p->cku_xid);
1286 1286 RPCLOG(1, "signal interrupted at %ld", now);
1287 1287 RPCLOG(1, ", was sent at %ld\n", time_sent);
1288 1288 } else {
1289 1289 COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1290 1290 p->cku_err.re_errno = ETIMEDOUT;
1291 1291 RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1292 1292 now);
1293 1293 RPCLOG(1, ", was sent at %ld\n", time_sent);
1294 1294 }
1295 1295 break;
1296 1296
1297 1297 case RPC_XPRTFAILED:
1298 1298 if (p->cku_err.re_errno == 0)
1299 1299 p->cku_err.re_errno = EIO;
1300 1300
1301 1301 RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1302 1302 p->cku_err.re_errno);
1303 1303 break;
1304 1304
1305 1305 case RPC_SYSTEMERROR:
1306 1306 ASSERT(p->cku_err.re_errno);
1307 1307 RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1308 1308 p->cku_err.re_errno);
1309 1309 break;
1310 1310
1311 1311 default:
1312 1312 p->cku_err.re_status = RPC_SYSTEMERROR;
1313 1313 p->cku_err.re_errno = EIO;
1314 1314 RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1315 1315 clnt_sperrno(status));
1316 1316 break;
1317 1317 }
1318 1318 if (p->cku_err.re_status != RPC_TIMEDOUT) {
1319 1319
1320 1320 if (p->cku_flags & CKU_ONQUEUE) {
1321 1321 call_table_remove(call);
1322 1322 p->cku_flags &= ~CKU_ONQUEUE;
1323 1323 }
1324 1324
1325 1325 RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1326 1326 "taken off dispatch list\n", p->cku_xid);
1327 1327 if (call->call_reply) {
1328 1328 freemsg(call->call_reply);
1329 1329 call->call_reply = NULL;
1330 1330 }
1331 1331 } else if (wait.tv_sec != 0) {
1332 1332 /*
1333 1333 * We've sent the request over TCP and so we have
1334 1334 * every reason to believe it will get
1335 1335 * delivered. In which case returning a timeout is not
1336 1336 * appropriate.
1337 1337 */
1338 1338 if (p->cku_progress == TRUE &&
1339 1339 p->cku_recv_attempts < clnt_cots_maxrecv) {
1340 1340 p->cku_err.re_status = RPC_INPROGRESS;
1341 1341 }
1342 1342 }
1343 1343 goto cots_done;
1344 1344 }
1345 1345
1346 1346 xdrs = &p->cku_inxdr;
1347 1347 xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1348 1348
1349 1349 reply_msg.rm_direction = REPLY;
1350 1350 reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1351 1351 reply_msg.acpted_rply.ar_stat = SUCCESS;
1352 1352
1353 1353 reply_msg.acpted_rply.ar_verf = _null_auth;
1354 1354 /*
1355 1355 * xdr_results will be done in AUTH_UNWRAP.
1356 1356 */
1357 1357 reply_msg.acpted_rply.ar_results.where = NULL;
1358 1358 reply_msg.acpted_rply.ar_results.proc = xdr_void;
1359 1359
1360 1360 if (xdr_replymsg(xdrs, &reply_msg)) {
1361 1361 enum clnt_stat re_status;
1362 1362
1363 1363 _seterr_reply(&reply_msg, &p->cku_err);
1364 1364
1365 1365 re_status = p->cku_err.re_status;
1366 1366 if (re_status == RPC_SUCCESS) {
1367 1367 /*
1368 1368 * Reply is good, check auth.
1369 1369 */
1370 1370 if (!AUTH_VALIDATE(h->cl_auth,
1371 1371 &reply_msg.acpted_rply.ar_verf)) {
1372 1372 COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1373 1373 RPCLOG0(1, "clnt_cots_kcallit: validation "
1374 1374 "failure\n");
1375 1375 freemsg(mp);
1376 1376 (void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1377 1377 XDR_DESTROY(xdrs);
1378 1378 mutex_enter(&call->call_lock);
1379 1379 if (call->call_reply == NULL)
1380 1380 call->call_status = RPC_TIMEDOUT;
1381 1381 mutex_exit(&call->call_lock);
1382 1382 goto read_again;
1383 1383 } else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1384 1384 xdr_results, resultsp)) {
1385 1385 RPCLOG0(1, "clnt_cots_kcallit: validation "
1386 1386 "failure (unwrap)\n");
1387 1387 p->cku_err.re_status = RPC_CANTDECODERES;
1388 1388 p->cku_err.re_errno = EIO;
1389 1389 }
1390 1390 } else {
1391 1391 /* set errno in case we can't recover */
1392 1392 if (re_status != RPC_VERSMISMATCH &&
1393 1393 re_status != RPC_AUTHERROR &&
1394 1394 re_status != RPC_PROGVERSMISMATCH)
1395 1395 p->cku_err.re_errno = EIO;
1396 1396
1397 1397 if (re_status == RPC_AUTHERROR) {
1398 1398 /*
1399 1399 * Maybe our credential need to be refreshed
1400 1400 */
1401 1401 if (cm_entry) {
1402 1402 /*
1403 1403 * There is the potential that the
1404 1404 * cm_entry has/will be marked dead,
1405 1405 * so drop the connection altogether,
1406 1406 * force REFRESH to establish new
1407 1407 * connection.
1408 1408 */
1409 1409 connmgr_cancelconn(cm_entry);
1410 1410 cm_entry = NULL;
1411 1411 }
1412 1412
1413 1413 (void) xdr_rpc_free_verifier(xdrs,
1414 1414 &reply_msg);
1415 1415 XDR_DESTROY(xdrs);
1416 1416
1417 1417 if (p->cku_flags & CKU_ONQUEUE) {
1418 1418 call_table_remove(call);
1419 1419 p->cku_flags &= ~CKU_ONQUEUE;
1420 1420 }
1421 1421 RPCLOG(64,
1422 1422 "clnt_cots_kcallit: AUTH_ERROR, xid"
1423 1423 " 0x%x removed off dispatch list\n",
1424 1424 p->cku_xid);
1425 1425 if (call->call_reply) {
1426 1426 freemsg(call->call_reply);
1427 1427 call->call_reply = NULL;
1428 1428 }
1429 1429
1430 1430 if ((refreshes > 0) &&
1431 1431 AUTH_REFRESH(h->cl_auth, &reply_msg,
1432 1432 p->cku_cred)) {
1433 1433 refreshes--;
1434 1434 freemsg(mp);
1435 1435 mp = NULL;
1436 1436
1437 1437 COTSRCSTAT_INCR(p->cku_stats,
1438 1438 rcbadcalls);
1439 1439 COTSRCSTAT_INCR(p->cku_stats,
1440 1440 rcnewcreds);
1441 1441 goto call_again;
1442 1442 }
1443 1443
1444 1444 /*
1445 1445 * We have used the client handle to
1446 1446 * do an AUTH_REFRESH and the RPC status may
1447 1447 * be set to RPC_SUCCESS; Let's make sure to
1448 1448 * set it to RPC_AUTHERROR.
1449 1449 */
1450 1450 p->cku_err.re_status = RPC_AUTHERROR;
1451 1451
1452 1452 /*
1453 1453 * Map recoverable and unrecoverable
1454 1454 * authentication errors to appropriate errno
1455 1455 */
1456 1456 switch (p->cku_err.re_why) {
1457 1457 case AUTH_TOOWEAK:
1458 1458 /*
1459 1459 * This could be a failure where the
1460 1460 * server requires use of a reserved
1461 1461 * port, check and optionally set the
1462 1462 * client handle useresvport trying
1463 1463 * one more time. Next go round we
1464 1464 * fall out with the tooweak error.
1465 1465 */
1466 1466 if (p->cku_useresvport != 1) {
1467 1467 p->cku_useresvport = 1;
1468 1468 p->cku_xid = 0;
1469 1469 freemsg(mp);
1470 1470 mp = NULL;
1471 1471 goto call_again;
1472 1472 }
1473 1473 /* FALLTHRU */
1474 1474 case AUTH_BADCRED:
1475 1475 case AUTH_BADVERF:
1476 1476 case AUTH_INVALIDRESP:
1477 1477 case AUTH_FAILED:
1478 1478 case RPCSEC_GSS_NOCRED:
1479 1479 case RPCSEC_GSS_FAILED:
1480 1480 p->cku_err.re_errno = EACCES;
1481 1481 break;
1482 1482 case AUTH_REJECTEDCRED:
1483 1483 case AUTH_REJECTEDVERF:
1484 1484 default: p->cku_err.re_errno = EIO;
1485 1485 break;
1486 1486 }
1487 1487 RPCLOG(1, "clnt_cots_kcallit : authentication"
1488 1488 " failed with RPC_AUTHERROR of type %d\n",
1489 1489 (int)p->cku_err.re_why);
1490 1490 goto cots_done;
1491 1491 }
1492 1492 }
1493 1493 } else {
1494 1494 /* reply didn't decode properly. */
1495 1495 p->cku_err.re_status = RPC_CANTDECODERES;
1496 1496 p->cku_err.re_errno = EIO;
1497 1497 RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1498 1498 }
1499 1499
1500 1500 (void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1501 1501 XDR_DESTROY(xdrs);
1502 1502
1503 1503 if (p->cku_flags & CKU_ONQUEUE) {
1504 1504 call_table_remove(call);
1505 1505 p->cku_flags &= ~CKU_ONQUEUE;
1506 1506 }
1507 1507
1508 1508 RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1509 1509 p->cku_xid);
1510 1510 RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1511 1511 cots_done:
1512 1512 if (cm_entry)
1513 1513 connmgr_release(cm_entry);
1514 1514
1515 1515 if (mp != NULL)
1516 1516 freemsg(mp);
1517 1517 if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1518 1518 freemsg(call->call_reply);
1519 1519 call->call_reply = NULL;
1520 1520 }
1521 1521 if (p->cku_err.re_status != RPC_SUCCESS) {
1522 1522 RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1523 1523 COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1524 1524 }
1525 1525
1526 1526 /*
1527 1527 * No point in delaying if the zone is going away.
1528 1528 */
1529 1529 if (delay_first == TRUE &&
1530 1530 !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1531 1531 if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1532 1532 p->cku_err.re_errno = EINTR;
1533 1533 p->cku_err.re_status = RPC_INTR;
1534 1534 }
1535 1535 }
1536 1536 return (p->cku_err.re_status);
1537 1537 }
1538 1538
1539 1539 /*
1540 1540 * Kinit routine for cots. This sets up the correct operations in
1541 1541 * the client handle, as the handle may have previously been a clts
1542 1542 * handle, and clears the xid field so there is no way a new call
1543 1543 * could be mistaken for a retry. It also sets in the handle the
1544 1544 * information that is passed at create/kinit time but needed at
1545 1545 * call time, as cots creates the transport at call time - device,
1546 1546 * address of the server, protocol family.
1547 1547 */
1548 1548 void
1549 1549 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1550 1550 int max_msgsize, cred_t *cred)
1551 1551 {
1552 1552 /* LINTED pointer alignment */
1553 1553 cku_private_t *p = htop(h);
1554 1554 calllist_t *call = &p->cku_call;
1555 1555
1556 1556 h->cl_ops = &tcp_ops;
1557 1557 if (p->cku_flags & CKU_ONQUEUE) {
1558 1558 call_table_remove(call);
1559 1559 p->cku_flags &= ~CKU_ONQUEUE;
1560 1560 RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1561 1561 " dispatch list\n", p->cku_xid);
1562 1562 }
1563 1563
1564 1564 if (call->call_reply != NULL) {
1565 1565 freemsg(call->call_reply);
1566 1566 call->call_reply = NULL;
1567 1567 }
1568 1568
1569 1569 call->call_bucket = NULL;
1570 1570 call->call_hash = 0;
1571 1571
1572 1572 /*
1573 1573 * We don't clear cku_flags here, because clnt_cots_kcallit()
1574 1574 * takes care of handling the cku_flags reset.
1575 1575 */
1576 1576 p->cku_xid = 0;
1577 1577 p->cku_device = dev;
1578 1578 p->cku_addrfmly = family;
1579 1579 p->cku_cred = cred;
1580 1580
1581 1581 if (p->cku_addr.maxlen < addr->len) {
1582 1582 if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL)
1583 1583 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1584 1584 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1585 1585 p->cku_addr.maxlen = addr->maxlen;
1586 1586 }
1587 1587
1588 1588 p->cku_addr.len = addr->len;
1589 1589 bcopy(addr->buf, p->cku_addr.buf, addr->len);
1590 1590
1591 1591 /*
1592 1592 * If the current sanity check size in rpcmod is smaller
1593 1593 * than the size needed, then increase the sanity check.
1594 1594 */
1595 1595 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1596 1596 max_msgsize > *clnt_max_msg_sizep) {
1597 1597 mutex_enter(&clnt_max_msg_lock);
1598 1598 if (max_msgsize > *clnt_max_msg_sizep)
1599 1599 *clnt_max_msg_sizep = max_msgsize;
1600 1600 mutex_exit(&clnt_max_msg_lock);
1601 1601 }
1602 1602 }
1603 1603
1604 1604 /*
1605 1605 * ksettimers is a no-op for cots, with the exception of setting the xid.
1606 1606 */
1607 1607 /* ARGSUSED */
1608 1608 static int
1609 1609 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1610 1610 int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg, uint32_t xid)
1611 1611 {
1612 1612 /* LINTED pointer alignment */
1613 1613 cku_private_t *p = htop(h);
1614 1614
1615 1615 if (xid)
1616 1616 p->cku_xid = xid;
1617 1617 COTSRCSTAT_INCR(p->cku_stats, rctimers);
1618 1618 return (0);
1619 1619 }
1620 1620
1621 1621 extern void rpc_poptimod(struct vnode *);
1622 1622 extern int kstr_push(struct vnode *, char *);
1623 1623
1624 1624 int
1625 1625 conn_kstat_update(kstat_t *ksp, int rw)
1626 1626 {
1627 1627 struct cm_xprt *cm_entry;
1628 1628 struct cm_kstat_xprt *cm_ksp_data;
1629 1629 uchar_t *b;
1630 1630 char *fbuf;
1631 1631
1632 1632 if (rw == KSTAT_WRITE)
1633 1633 return (EACCES);
1634 1634 if (ksp == NULL || ksp->ks_private == NULL)
1635 1635 return (EIO);
1636 1636 cm_entry = (struct cm_xprt *)ksp->ks_private;
1637 1637 cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1638 1638
1639 1639 cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1640 1640 cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1641 1641 cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1642 1642 cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1643 1643 cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1644 1644 cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1645 1645
1646 1646 if (cm_entry->x_server.buf) {
1647 1647 fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1648 1648 if (cm_entry->x_family == AF_INET &&
1649 1649 cm_entry->x_server.len ==
1650 1650 sizeof (struct sockaddr_in)) {
1651 1651 struct sockaddr_in *sa;
1652 1652 sa = (struct sockaddr_in *)
1653 1653 cm_entry->x_server.buf;
1654 1654 b = (uchar_t *)&sa->sin_addr;
1655 1655 (void) sprintf(fbuf,
1656 1656 "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1657 1657 b[2] & 0xFF, b[3] & 0xFF);
1658 1658 cm_ksp_data->x_port.value.ui32 = ntohs(sa->sin_port);
1659 1659 } else if (cm_entry->x_family == AF_INET6 &&
1660 1660 cm_entry->x_server.len >=
1661 1661 sizeof (struct sockaddr_in6)) {
1662 1662 /* extract server IP address & port */
1663 1663 struct sockaddr_in6 *sin6;
1664 1664 sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1665 1665 (void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1666 1666 INET6_ADDRSTRLEN);
1667 1667 cm_ksp_data->x_port.value.ui32 = ntohs(sin6->sin6_port);
1668 1668 } else {
1669 1669 struct sockaddr_in *sa;
1670 1670
1671 1671 sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1672 1672 b = (uchar_t *)&sa->sin_addr;
1673 1673 (void) sprintf(fbuf,
1674 1674 "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1675 1675 b[2] & 0xFF, b[3] & 0xFF);
1676 1676 }
1677 1677 KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1678 1678 strlen(fbuf) + 1;
1679 1679 }
1680 1680
1681 1681 return (0);
1682 1682 }
1683 1683
1684 1684
1685 1685 /*
1686 1686 * We want a version of delay which is interruptible by a UNIX signal
1687 1687 * Return EINTR if an interrupt occured.
1688 1688 */
1689 1689 static int
1690 1690 clnt_delay(clock_t ticks, bool_t nosignal)
1691 1691 {
1692 1692 if (nosignal == TRUE) {
1693 1693 delay(ticks);
1694 1694 return (0);
1695 1695 }
1696 1696 return (delay_sig(ticks));
1697 1697 }
1698 1698
1699 1699 /*
1700 1700 * Wait for a connection until a timeout, or until we are
1701 1701 * signalled that there has been a connection state change.
1702 1702 */
1703 1703 static enum clnt_stat
1704 1704 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1705 1705 bool_t nosignal)
1706 1706 {
1707 1707 bool_t interrupted;
1708 1708 clock_t timout, cv_stat;
1709 1709 enum clnt_stat clstat;
1710 1710 unsigned int old_state;
1711 1711
1712 1712 ASSERT(MUTEX_HELD(&connmgr_lock));
1713 1713 /*
1714 1714 * We wait for the transport connection to be made, or an
1715 1715 * indication that it could not be made.
1716 1716 */
1717 1717 clstat = RPC_TIMEDOUT;
1718 1718 interrupted = FALSE;
1719 1719
1720 1720 old_state = cm_entry->x_state_flags;
1721 1721 /*
1722 1722 * Now loop until cv_timedwait{_sig} returns because of
1723 1723 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1724 1724 * cv_signalled for various other reasons too. So loop
1725 1725 * until there is a state change on the connection.
1726 1726 */
1727 1727
1728 1728 timout = waitp->tv_sec * drv_usectohz(1000000) +
1729 1729 drv_usectohz(waitp->tv_usec) + ddi_get_lbolt();
1730 1730
1731 1731 if (nosignal) {
1732 1732 while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1733 1733 &connmgr_lock, timout)) > 0 &&
1734 1734 cm_entry->x_state_flags == old_state)
1735 1735 ;
1736 1736 } else {
1737 1737 while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1738 1738 &connmgr_lock, timout)) > 0 &&
1739 1739 cm_entry->x_state_flags == old_state)
1740 1740 ;
1741 1741
1742 1742 if (cv_stat == 0) /* got intr signal? */
1743 1743 interrupted = TRUE;
1744 1744 }
1745 1745
1746 1746 if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1747 1747 X_CONNECTED) {
1748 1748 clstat = RPC_SUCCESS;
1749 1749 } else {
1750 1750 if (interrupted == TRUE)
1751 1751 clstat = RPC_INTR;
1752 1752 RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1753 1753 clnt_sperrno(clstat));
1754 1754 }
1755 1755
1756 1756 return (clstat);
1757 1757 }
1758 1758
1759 1759 /*
1760 1760 * Primary interface for how RPC grabs a connection.
1761 1761 */
1762 1762 static struct cm_xprt *
1763 1763 connmgr_wrapget(
1764 1764 struct netbuf *retryaddr,
1765 1765 const struct timeval *waitp,
1766 1766 cku_private_t *p)
1767 1767 {
1768 1768 struct cm_xprt *cm_entry;
1769 1769
1770 1770 cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1771 1771 &p->cku_srcaddr, &p->cku_err, p->cku_device,
1772 1772 p->cku_client.cl_nosignal, p->cku_useresvport, p->cku_cred);
1773 1773
1774 1774 if (cm_entry == NULL) {
1775 1775 /*
1776 1776 * Re-map the call status to RPC_INTR if the err code is
1777 1777 * EINTR. This can happen if calls status is RPC_TLIERROR.
1778 1778 * However, don't re-map if signalling has been turned off.
1779 1779 * XXX Really need to create a separate thread whenever
1780 1780 * there isn't an existing connection.
1781 1781 */
1782 1782 if (p->cku_err.re_errno == EINTR) {
1783 1783 if (p->cku_client.cl_nosignal == TRUE)
1784 1784 p->cku_err.re_errno = EIO;
1785 1785 else
1786 1786 p->cku_err.re_status = RPC_INTR;
1787 1787 }
1788 1788 }
1789 1789
1790 1790 return (cm_entry);
1791 1791 }
1792 1792
1793 1793 /*
1794 1794 * Obtains a transport to the server specified in addr. If a suitable transport
1795 1795 * does not already exist in the list of cached transports, a new connection
1796 1796 * is created, connected, and added to the list. The connection is for sending
1797 1797 * only - the reply message may come back on another transport connection.
1798 1798 *
1799 1799 * To implement round-robin load balancing with multiple client connections,
1800 1800 * the last entry on the list is always selected. Once the entry is selected
1801 1801 * it's re-inserted to the head of the list.
1802 1802 */
1803 1803 static struct cm_xprt *
1804 1804 connmgr_get(
1805 1805 struct netbuf *retryaddr,
1806 1806 const struct timeval *waitp, /* changed to a ptr to converse stack */
1807 1807 struct netbuf *destaddr,
1808 1808 int addrfmly,
1809 1809 struct netbuf *srcaddr,
1810 1810 struct rpc_err *rpcerr,
1811 1811 dev_t device,
1812 1812 bool_t nosignal,
1813 1813 int useresvport,
1814 1814 cred_t *cr)
1815 1815 {
1816 1816 struct cm_xprt *cm_entry;
1817 1817 struct cm_xprt *lru_entry;
1818 1818 struct cm_xprt **cmp, **prev;
1819 1819 queue_t *wq;
1820 1820 TIUSER *tiptr;
1821 1821 int i;
1822 1822 int retval;
1823 1823 int tidu_size;
1824 1824 bool_t connected;
1825 1825 zoneid_t zoneid = rpc_zoneid();
1826 1826
1827 1827 /*
1828 1828 * If the call is not a retry, look for a transport entry that
1829 1829 * goes to the server of interest.
1830 1830 */
1831 1831 mutex_enter(&connmgr_lock);
1832 1832
1833 1833 if (retryaddr == NULL) {
1834 1834 use_new_conn:
1835 1835 i = 0;
1836 1836 cm_entry = lru_entry = NULL;
1837 1837
1838 1838 prev = cmp = &cm_hd;
1839 1839 while ((cm_entry = *cmp) != NULL) {
1840 1840 ASSERT(cm_entry != cm_entry->x_next);
1841 1841 /*
1842 1842 * Garbage collect conections that are marked
1843 1843 * for needs disconnect.
1844 1844 */
1845 1845 if (cm_entry->x_needdis) {
1846 1846 CONN_HOLD(cm_entry);
1847 1847 connmgr_dis_and_wait(cm_entry);
1848 1848 connmgr_release(cm_entry);
1849 1849 /*
1850 1850 * connmgr_lock could have been
1851 1851 * dropped for the disconnect
1852 1852 * processing so start over.
1853 1853 */
1854 1854 goto use_new_conn;
1855 1855 }
1856 1856
1857 1857 /*
1858 1858 * Garbage collect the dead connections that have
1859 1859 * no threads working on them.
1860 1860 */
1861 1861 if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1862 1862 X_DEAD) {
1863 1863 mutex_enter(&cm_entry->x_lock);
1864 1864 if (cm_entry->x_ref != 0) {
1865 1865 /*
1866 1866 * Currently in use.
1867 1867 * Cleanup later.
1868 1868 */
1869 1869 cmp = &cm_entry->x_next;
1870 1870 mutex_exit(&cm_entry->x_lock);
1871 1871 continue;
1872 1872 }
1873 1873 mutex_exit(&cm_entry->x_lock);
1874 1874 *cmp = cm_entry->x_next;
1875 1875 mutex_exit(&connmgr_lock);
1876 1876 connmgr_close(cm_entry);
1877 1877 mutex_enter(&connmgr_lock);
1878 1878 goto use_new_conn;
1879 1879 }
1880 1880
1881 1881
1882 1882 if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1883 1883 cm_entry->x_zoneid == zoneid &&
1884 1884 cm_entry->x_rdev == device &&
1885 1885 destaddr->len == cm_entry->x_server.len &&
1886 1886 bcmp(destaddr->buf, cm_entry->x_server.buf,
1887 1887 destaddr->len) == 0) {
1888 1888 /*
1889 1889 * If the matching entry isn't connected,
1890 1890 * attempt to reconnect it.
1891 1891 */
1892 1892 if (cm_entry->x_connected == FALSE) {
1893 1893 /*
1894 1894 * We don't go through trying
1895 1895 * to find the least recently
1896 1896 * used connected because
1897 1897 * connmgr_reconnect() briefly
1898 1898 * dropped the connmgr_lock,
1899 1899 * allowing a window for our
1900 1900 * accounting to be messed up.
1901 1901 * In any case, a re-connected
1902 1902 * connection is as good as
1903 1903 * a LRU connection.
1904 1904 */
1905 1905 return (connmgr_wrapconnect(cm_entry,
1906 1906 waitp, destaddr, addrfmly, srcaddr,
1907 1907 rpcerr, TRUE, nosignal, cr));
1908 1908 }
1909 1909 i++;
1910 1910
1911 1911 /* keep track of the last entry */
1912 1912 lru_entry = cm_entry;
1913 1913 prev = cmp;
1914 1914 }
1915 1915 cmp = &cm_entry->x_next;
1916 1916 }
1917 1917
1918 1918 if (i > clnt_max_conns) {
1919 1919 RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1920 1920 " %p\n", (void *)lru_entry->x_tiptr);
1921 1921 lru_entry->x_doomed = TRUE;
1922 1922 goto use_new_conn;
1923 1923 }
1924 1924
1925 1925 /*
1926 1926 * If we are at the maximum number of connections to
1927 1927 * the server, hand back the least recently used one.
1928 1928 */
1929 1929 if (i == clnt_max_conns) {
1930 1930 /*
1931 1931 * Copy into the handle the source address of
1932 1932 * the connection, which we will use in case of
1933 1933 * a later retry.
1934 1934 */
1935 1935 if (srcaddr->len != lru_entry->x_src.len) {
1936 1936 if (srcaddr->len > 0)
1937 1937 kmem_free(srcaddr->buf,
1938 1938 srcaddr->maxlen);
1939 1939 srcaddr->buf = kmem_zalloc(
1940 1940 lru_entry->x_src.len, KM_SLEEP);
1941 1941 srcaddr->maxlen = srcaddr->len =
1942 1942 lru_entry->x_src.len;
1943 1943 }
1944 1944 bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len);
1945 1945 RPCLOG(2, "connmgr_get: call going out on %p\n",
1946 1946 (void *)lru_entry);
1947 1947 lru_entry->x_time = ddi_get_lbolt();
1948 1948 CONN_HOLD(lru_entry);
1949 1949
1950 1950 if ((i > 1) && (prev != &cm_hd)) {
1951 1951 /*
1952 1952 * remove and re-insert entry at head of list.
1953 1953 */
1954 1954 *prev = lru_entry->x_next;
1955 1955 lru_entry->x_next = cm_hd;
1956 1956 cm_hd = lru_entry;
1957 1957 }
1958 1958
1959 1959 mutex_exit(&connmgr_lock);
1960 1960 return (lru_entry);
1961 1961 }
1962 1962
1963 1963 } else {
1964 1964 /*
1965 1965 * This is the retry case (retryaddr != NULL). Retries must
1966 1966 * be sent on the same source port as the original call.
1967 1967 */
1968 1968
1969 1969 /*
1970 1970 * Walk the list looking for a connection with a source address
1971 1971 * that matches the retry address.
1972 1972 */
1973 1973 start_retry_loop:
1974 1974 cmp = &cm_hd;
1975 1975 while ((cm_entry = *cmp) != NULL) {
1976 1976 ASSERT(cm_entry != cm_entry->x_next);
1977 1977
1978 1978 /*
1979 1979 * determine if this connection matches the passed
1980 1980 * in retry address. If it does not match, advance
1981 1981 * to the next element on the list.
1982 1982 */
1983 1983 if (zoneid != cm_entry->x_zoneid ||
1984 1984 device != cm_entry->x_rdev ||
1985 1985 retryaddr->len != cm_entry->x_src.len ||
1986 1986 bcmp(retryaddr->buf, cm_entry->x_src.buf,
1987 1987 retryaddr->len) != 0) {
1988 1988 cmp = &cm_entry->x_next;
1989 1989 continue;
1990 1990 }
1991 1991 /*
1992 1992 * Garbage collect conections that are marked
1993 1993 * for needs disconnect.
1994 1994 */
1995 1995 if (cm_entry->x_needdis) {
1996 1996 CONN_HOLD(cm_entry);
1997 1997 connmgr_dis_and_wait(cm_entry);
1998 1998 connmgr_release(cm_entry);
1999 1999 /*
2000 2000 * connmgr_lock could have been
2001 2001 * dropped for the disconnect
2002 2002 * processing so start over.
2003 2003 */
2004 2004 goto start_retry_loop;
2005 2005 }
2006 2006 /*
2007 2007 * Garbage collect the dead connections that have
2008 2008 * no threads working on them.
2009 2009 */
2010 2010 if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
2011 2011 X_DEAD) {
2012 2012 mutex_enter(&cm_entry->x_lock);
2013 2013 if (cm_entry->x_ref != 0) {
2014 2014 /*
2015 2015 * Currently in use.
2016 2016 * Cleanup later.
2017 2017 */
2018 2018 cmp = &cm_entry->x_next;
2019 2019 mutex_exit(&cm_entry->x_lock);
2020 2020 continue;
2021 2021 }
2022 2022 mutex_exit(&cm_entry->x_lock);
2023 2023 *cmp = cm_entry->x_next;
2024 2024 mutex_exit(&connmgr_lock);
2025 2025 connmgr_close(cm_entry);
2026 2026 mutex_enter(&connmgr_lock);
2027 2027 goto start_retry_loop;
2028 2028 }
2029 2029
2030 2030 /*
2031 2031 * Sanity check: if the connection with our source
2032 2032 * port is going to some other server, something went
2033 2033 * wrong, as we never delete connections (i.e. release
2034 2034 * ports) unless they have been idle. In this case,
2035 2035 * it is probably better to send the call out using
2036 2036 * a new source address than to fail it altogether,
2037 2037 * since that port may never be released.
2038 2038 */
2039 2039 if (destaddr->len != cm_entry->x_server.len ||
2040 2040 bcmp(destaddr->buf, cm_entry->x_server.buf,
2041 2041 destaddr->len) != 0) {
2042 2042 RPCLOG(1, "connmgr_get: tiptr %p"
2043 2043 " is going to a different server"
2044 2044 " with the port that belongs"
2045 2045 " to us!\n", (void *)cm_entry->x_tiptr);
2046 2046 retryaddr = NULL;
2047 2047 goto use_new_conn;
2048 2048 }
2049 2049
2050 2050 /*
2051 2051 * If the connection of interest is not connected and we
2052 2052 * can't reconnect it, then the server is probably
2053 2053 * still down. Return NULL to the caller and let it
2054 2054 * retry later if it wants to. We have a delay so the
2055 2055 * machine doesn't go into a tight retry loop. If the
2056 2056 * entry was already connected, or the reconnected was
2057 2057 * successful, return this entry.
2058 2058 */
2059 2059 if (cm_entry->x_connected == FALSE) {
2060 2060 return (connmgr_wrapconnect(cm_entry,
2061 2061 waitp, destaddr, addrfmly, NULL,
2062 2062 rpcerr, TRUE, nosignal, cr));
2063 2063 } else {
2064 2064 CONN_HOLD(cm_entry);
2065 2065
2066 2066 cm_entry->x_time = ddi_get_lbolt();
2067 2067 mutex_exit(&connmgr_lock);
2068 2068 RPCLOG(2, "connmgr_get: found old "
2069 2069 "transport %p for retry\n",
2070 2070 (void *)cm_entry);
2071 2071 return (cm_entry);
2072 2072 }
2073 2073 }
2074 2074
2075 2075 /*
2076 2076 * We cannot find an entry in the list for this retry.
2077 2077 * Either the entry has been removed temporarily to be
2078 2078 * reconnected by another thread, or the original call
2079 2079 * got a port but never got connected,
2080 2080 * and hence the transport never got put in the
2081 2081 * list. Fall through to the "create new connection" code -
2082 2082 * the former case will fail there trying to rebind the port,
2083 2083 * and the later case (and any other pathological cases) will
2084 2084 * rebind and reconnect and not hang the client machine.
2085 2085 */
2086 2086 RPCLOG0(8, "connmgr_get: no entry in list for retry\n");
2087 2087 }
2088 2088 /*
2089 2089 * Set up a transport entry in the connection manager's list.
2090 2090 */
2091 2091 cm_entry = (struct cm_xprt *)
2092 2092 kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP);
2093 2093
2094 2094 cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2095 2095 bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len);
2096 2096 cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len;
2097 2097
2098 2098 cm_entry->x_state_flags = X_THREAD;
2099 2099 cm_entry->x_ref = 1;
2100 2100 cm_entry->x_family = addrfmly;
2101 2101 cm_entry->x_rdev = device;
2102 2102 cm_entry->x_zoneid = zoneid;
2103 2103 mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL);
2104 2104 cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL);
2105 2105 cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL);
2106 2106 cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL);
2107 2107
2108 2108 /*
2109 2109 * Note that we add this partially initialized entry to the
2110 2110 * connection list. This is so that we don't have connections to
2111 2111 * the same server.
2112 2112 *
2113 2113 * Note that x_src is not initialized at this point. This is because
2114 2114 * retryaddr might be NULL in which case x_src is whatever
2115 2115 * t_kbind/bindresvport gives us. If another thread wants a
2116 2116 * connection to the same server, seemingly we have an issue, but we
2117 2117 * don't. If the other thread comes in with retryaddr == NULL, then it
2118 2118 * will never look at x_src, and it will end up waiting in
2119 2119 * connmgr_cwait() for the first thread to finish the connection
2120 2120 * attempt. If the other thread comes in with retryaddr != NULL, then
2121 2121 * that means there was a request sent on a connection, in which case
2122 2122 * the the connection should already exist. Thus the first thread
2123 2123 * never gets here ... it finds the connection it its server in the
2124 2124 * connection list.
2125 2125 *
2126 2126 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd
2127 2127 * thread will skip us because x_src.len == 0.
2128 2128 */
2129 2129 cm_entry->x_next = cm_hd;
2130 2130 cm_hd = cm_entry;
2131 2131 mutex_exit(&connmgr_lock);
2132 2132
2133 2133 /*
2134 2134 * Either we didn't find an entry to the server of interest, or we
2135 2135 * don't have the maximum number of connections to that server -
2136 2136 * create a new connection.
2137 2137 */
2138 2138 RPCLOG0(8, "connmgr_get: creating new connection\n");
2139 2139 rpcerr->re_status = RPC_TLIERROR;
2140 2140
2141 2141 i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, zone_kcred());
2142 2142 if (i) {
2143 2143 RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i);
2144 2144 rpcerr->re_errno = i;
2145 2145 connmgr_cancelconn(cm_entry);
2146 2146 return (NULL);
2147 2147 }
2148 2148 rpc_poptimod(tiptr->fp->f_vnode);
2149 2149
2150 2150 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0,
2151 2151 K_TO_K, kcred, &retval)) {
2152 2152 RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i);
2153 2153 (void) t_kclose(tiptr, 1);
2154 2154 rpcerr->re_errno = i;
2155 2155 connmgr_cancelconn(cm_entry);
2156 2156 return (NULL);
2157 2157 }
2158 2158
2159 2159 if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K,
2160 2160 kcred, &retval)) {
2161 2161 RPCLOG(1, "connmgr_get: can't set client status with cots "
2162 2162 "module, %d\n", i);
2163 2163 (void) t_kclose(tiptr, 1);
2164 2164 rpcerr->re_errno = i;
2165 2165 connmgr_cancelconn(cm_entry);
2166 2166 return (NULL);
2167 2167 }
2168 2168
2169 2169 mutex_enter(&connmgr_lock);
2170 2170
2171 2171 wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next;
2172 2172 cm_entry->x_wq = wq;
2173 2173
2174 2174 mutex_exit(&connmgr_lock);
2175 2175
2176 2176 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0,
2177 2177 K_TO_K, kcred, &retval)) {
2178 2178 RPCLOG(1, "connmgr_get: can't push timod, %d\n", i);
2179 2179 (void) t_kclose(tiptr, 1);
2180 2180 rpcerr->re_errno = i;
2181 2181 connmgr_cancelconn(cm_entry);
2182 2182 return (NULL);
2183 2183 }
2184 2184
2185 2185 /*
2186 2186 * If the caller has not specified reserved port usage then
2187 2187 * take the system default.
2188 2188 */
2189 2189 if (useresvport == -1)
2190 2190 useresvport = clnt_cots_do_bindresvport;
2191 2191
2192 2192 if ((useresvport || retryaddr != NULL) &&
2193 2193 (addrfmly == AF_INET || addrfmly == AF_INET6)) {
2194 2194 bool_t alloc_src = FALSE;
2195 2195
2196 2196 if (srcaddr->len != destaddr->len) {
2197 2197 kmem_free(srcaddr->buf, srcaddr->maxlen);
2198 2198 srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2199 2199 srcaddr->maxlen = destaddr->len;
2200 2200 srcaddr->len = destaddr->len;
2201 2201 alloc_src = TRUE;
2202 2202 }
2203 2203
2204 2204 if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) {
2205 2205 (void) t_kclose(tiptr, 1);
2206 2206 RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: "
2207 2207 "%p\n", (void *)retryaddr);
2208 2208
2209 2209 /*
2210 2210 * 1225408: If we allocated a source address, then it
2211 2211 * is either garbage or all zeroes. In that case
2212 2212 * we need to clear srcaddr.
2213 2213 */
2214 2214 if (alloc_src == TRUE) {
2215 2215 kmem_free(srcaddr->buf, srcaddr->maxlen);
2216 2216 srcaddr->maxlen = srcaddr->len = 0;
2217 2217 srcaddr->buf = NULL;
2218 2218 }
2219 2219 rpcerr->re_errno = i;
2220 2220 connmgr_cancelconn(cm_entry);
2221 2221 return (NULL);
2222 2222 }
2223 2223 } else {
2224 2224 if ((i = t_kbind(tiptr, NULL, NULL)) != 0) {
2225 2225 RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i);
2226 2226 (void) t_kclose(tiptr, 1);
2227 2227 rpcerr->re_errno = i;
2228 2228 connmgr_cancelconn(cm_entry);
2229 2229 return (NULL);
2230 2230 }
2231 2231 }
2232 2232
2233 2233 {
2234 2234 /*
2235 2235 * Keep the kernel stack lean. Don't move this call
2236 2236 * declaration to the top of this function because a
2237 2237 * call is declared in connmgr_wrapconnect()
2238 2238 */
2239 2239 calllist_t call;
2240 2240
2241 2241 bzero(&call, sizeof (call));
2242 2242 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2243 2243
2244 2244 /*
2245 2245 * This is a bound end-point so don't close it's stream.
2246 2246 */
2247 2247 connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly,
2248 2248 &call, &tidu_size, FALSE, waitp, nosignal, cr);
2249 2249 *rpcerr = call.call_err;
2250 2250 cv_destroy(&call.call_cv);
2251 2251
2252 2252 }
2253 2253
2254 2254 mutex_enter(&connmgr_lock);
2255 2255
2256 2256 /*
2257 2257 * Set up a transport entry in the connection manager's list.
2258 2258 */
2259 2259 cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP);
2260 2260 bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2261 2261 cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2262 2262
2263 2263 cm_entry->x_tiptr = tiptr;
2264 2264 cm_entry->x_time = ddi_get_lbolt();
2265 2265
2266 2266 if (tiptr->tp_info.servtype == T_COTS_ORD)
2267 2267 cm_entry->x_ordrel = TRUE;
2268 2268 else
2269 2269 cm_entry->x_ordrel = FALSE;
2270 2270
2271 2271 cm_entry->x_tidu_size = tidu_size;
2272 2272
|
↓ open down ↓ |
1785 lines elided |
↑ open up ↑ |
2273 2273 if (cm_entry->x_early_disc) {
2274 2274 /*
2275 2275 * We need to check if a disconnect request has come
2276 2276 * while we are connected, if so, then we need to
2277 2277 * set rpcerr->re_status appropriately before returning
2278 2278 * NULL to caller.
2279 2279 */
2280 2280 if (rpcerr->re_status == RPC_SUCCESS)
2281 2281 rpcerr->re_status = RPC_XPRTFAILED;
2282 2282 cm_entry->x_connected = FALSE;
2283 + cm_entry->x_dead = TRUE;
2283 2284 } else
2284 2285 cm_entry->x_connected = connected;
2285 2286
2286 2287 /*
2287 2288 * There could be a discrepancy here such that
2288 2289 * x_early_disc is TRUE yet connected is TRUE as well
2289 2290 * and the connection is actually connected. In that case
2290 2291 * lets be conservative and declare the connection as not
2291 2292 * connected.
2292 2293 */
2293 2294 cm_entry->x_early_disc = FALSE;
2294 2295 cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2295 2296 cm_entry->x_ctime = ddi_get_lbolt();
2296 2297
2297 2298 /*
2298 2299 * Notify any threads waiting that the connection attempt is done.
2299 2300 */
2300 2301 cm_entry->x_thread = FALSE;
2301 2302 cv_broadcast(&cm_entry->x_conn_cv);
2302 2303
2303 2304 if (cm_entry->x_connected == FALSE) {
2304 2305 mutex_exit(&connmgr_lock);
2305 2306 connmgr_release(cm_entry);
2306 2307 return (NULL);
2307 2308 }
2308 2309
2309 2310 mutex_exit(&connmgr_lock);
2310 2311
2311 2312 return (cm_entry);
2312 2313 }
2313 2314
2314 2315 /*
2315 2316 * Keep the cm_xprt entry on the connecton list when making a connection. This
2316 2317 * is to prevent multiple connections to a slow server from appearing.
2317 2318 * We use the bit field x_thread to tell if a thread is doing a connection
2318 2319 * which keeps other interested threads from messing with connection.
2319 2320 * Those other threads just wait if x_thread is set.
2320 2321 *
2321 2322 * If x_thread is not set, then we do the actual work of connecting via
2322 2323 * connmgr_connect().
2323 2324 *
2324 2325 * mutex convention: called with connmgr_lock held, returns with it released.
2325 2326 */
2326 2327 static struct cm_xprt *
2327 2328 connmgr_wrapconnect(
2328 2329 struct cm_xprt *cm_entry,
2329 2330 const struct timeval *waitp,
2330 2331 struct netbuf *destaddr,
2331 2332 int addrfmly,
2332 2333 struct netbuf *srcaddr,
2333 2334 struct rpc_err *rpcerr,
2334 2335 bool_t reconnect,
2335 2336 bool_t nosignal,
2336 2337 cred_t *cr)
2337 2338 {
2338 2339 ASSERT(MUTEX_HELD(&connmgr_lock));
2339 2340 /*
2340 2341 * Hold this entry as we are about to drop connmgr_lock.
2341 2342 */
2342 2343 CONN_HOLD(cm_entry);
2343 2344
2344 2345 /*
2345 2346 * If there is a thread already making a connection for us, then
2346 2347 * wait for it to complete the connection.
2347 2348 */
2348 2349 if (cm_entry->x_thread == TRUE) {
2349 2350 rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2350 2351
2351 2352 if (rpcerr->re_status != RPC_SUCCESS) {
2352 2353 mutex_exit(&connmgr_lock);
2353 2354 connmgr_release(cm_entry);
2354 2355 return (NULL);
2355 2356 }
2356 2357 } else {
2357 2358 bool_t connected;
2358 2359 calllist_t call;
2359 2360
2360 2361 cm_entry->x_thread = TRUE;
2361 2362
2362 2363 while (cm_entry->x_needrel == TRUE) {
2363 2364 cm_entry->x_needrel = FALSE;
2364 2365
2365 2366 connmgr_sndrel(cm_entry);
2366 2367 delay(drv_usectohz(1000000));
2367 2368
2368 2369 mutex_enter(&connmgr_lock);
2369 2370 }
2370 2371
2371 2372 /*
2372 2373 * If we need to send a T_DISCON_REQ, send one.
2373 2374 */
2374 2375 connmgr_dis_and_wait(cm_entry);
2375 2376
2376 2377 mutex_exit(&connmgr_lock);
2377 2378
2378 2379 bzero(&call, sizeof (call));
2379 2380 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2380 2381
2381 2382 connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2382 2383 destaddr, addrfmly, &call, &cm_entry->x_tidu_size,
2383 2384 reconnect, waitp, nosignal, cr);
2384 2385
2385 2386 *rpcerr = call.call_err;
2386 2387 cv_destroy(&call.call_cv);
2387 2388
2388 2389 mutex_enter(&connmgr_lock);
2389 2390
2390 2391
|
↓ open down ↓ |
98 lines elided |
↑ open up ↑ |
2391 2392 if (cm_entry->x_early_disc) {
2392 2393 /*
2393 2394 * We need to check if a disconnect request has come
2394 2395 * while we are connected, if so, then we need to
2395 2396 * set rpcerr->re_status appropriately before returning
2396 2397 * NULL to caller.
2397 2398 */
2398 2399 if (rpcerr->re_status == RPC_SUCCESS)
2399 2400 rpcerr->re_status = RPC_XPRTFAILED;
2400 2401 cm_entry->x_connected = FALSE;
2402 + cm_entry->x_dead = TRUE;
2401 2403 } else
2402 2404 cm_entry->x_connected = connected;
2403 2405
2404 2406 /*
2405 2407 * There could be a discrepancy here such that
2406 2408 * x_early_disc is TRUE yet connected is TRUE as well
2407 2409 * and the connection is actually connected. In that case
2408 2410 * lets be conservative and declare the connection as not
2409 2411 * connected.
2410 2412 */
2411 2413
2412 2414 cm_entry->x_early_disc = FALSE;
2413 2415 cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2414 2416
2415 2417
2416 2418 /*
2417 2419 * connmgr_connect() may have given up before the connection
2418 2420 * actually timed out. So ensure that before the next
2419 2421 * connection attempt we do a disconnect.
2420 2422 */
2421 2423 cm_entry->x_ctime = ddi_get_lbolt();
2422 2424 cm_entry->x_thread = FALSE;
2423 2425
2424 2426 cv_broadcast(&cm_entry->x_conn_cv);
2425 2427
2426 2428 if (cm_entry->x_connected == FALSE) {
2427 2429 mutex_exit(&connmgr_lock);
2428 2430 connmgr_release(cm_entry);
2429 2431 return (NULL);
2430 2432 }
2431 2433 }
2432 2434
2433 2435 if (srcaddr != NULL) {
2434 2436 /*
2435 2437 * Copy into the handle the
2436 2438 * source address of the
2437 2439 * connection, which we will use
2438 2440 * in case of a later retry.
2439 2441 */
2440 2442 if (srcaddr->len != cm_entry->x_src.len) {
2441 2443 if (srcaddr->maxlen > 0)
2442 2444 kmem_free(srcaddr->buf, srcaddr->maxlen);
2443 2445 srcaddr->buf = kmem_zalloc(cm_entry->x_src.len,
2444 2446 KM_SLEEP);
2445 2447 srcaddr->maxlen = srcaddr->len =
2446 2448 cm_entry->x_src.len;
2447 2449 }
2448 2450 bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2449 2451 }
2450 2452 cm_entry->x_time = ddi_get_lbolt();
2451 2453 mutex_exit(&connmgr_lock);
2452 2454 return (cm_entry);
2453 2455 }
2454 2456
2455 2457 /*
2456 2458 * If we need to send a T_DISCON_REQ, send one.
2457 2459 */
2458 2460 static void
2459 2461 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2460 2462 {
2461 2463 ASSERT(MUTEX_HELD(&connmgr_lock));
2462 2464 for (;;) {
2463 2465 while (cm_entry->x_needdis == TRUE) {
2464 2466 RPCLOG(8, "connmgr_dis_and_wait: need "
2465 2467 "T_DISCON_REQ for connection 0x%p\n",
2466 2468 (void *)cm_entry);
2467 2469 cm_entry->x_needdis = FALSE;
2468 2470 cm_entry->x_waitdis = TRUE;
2469 2471
2470 2472 connmgr_snddis(cm_entry);
2471 2473
2472 2474 mutex_enter(&connmgr_lock);
2473 2475 }
2474 2476
2475 2477 if (cm_entry->x_waitdis == TRUE) {
2476 2478 clock_t timout;
2477 2479
2478 2480 RPCLOG(8, "connmgr_dis_and_wait waiting for "
2479 2481 "T_DISCON_REQ's ACK for connection %p\n",
2480 2482 (void *)cm_entry);
2481 2483
2482 2484 timout = clnt_cots_min_conntout * drv_usectohz(1000000);
2483 2485
2484 2486 /*
2485 2487 * The TPI spec says that the T_DISCON_REQ
2486 2488 * will get acknowledged, but in practice
2487 2489 * the ACK may never get sent. So don't
2488 2490 * block forever.
2489 2491 */
2490 2492 (void) cv_reltimedwait(&cm_entry->x_dis_cv,
2491 2493 &connmgr_lock, timout, TR_CLOCK_TICK);
2492 2494 }
2493 2495 /*
2494 2496 * If we got the ACK, break. If we didn't,
2495 2497 * then send another T_DISCON_REQ.
2496 2498 */
2497 2499 if (cm_entry->x_waitdis == FALSE) {
2498 2500 break;
2499 2501 } else {
2500 2502 RPCLOG(8, "connmgr_dis_and_wait: did"
2501 2503 "not get T_DISCON_REQ's ACK for "
2502 2504 "connection %p\n", (void *)cm_entry);
2503 2505 cm_entry->x_needdis = TRUE;
2504 2506 }
2505 2507 }
2506 2508 }
2507 2509
2508 2510 static void
2509 2511 connmgr_cancelconn(struct cm_xprt *cm_entry)
2510 2512 {
2511 2513 /*
2512 2514 * Mark the connection table entry as dead; the next thread that
2513 2515 * goes through connmgr_release() will notice this and deal with it.
2514 2516 */
2515 2517 mutex_enter(&connmgr_lock);
2516 2518 cm_entry->x_dead = TRUE;
2517 2519
2518 2520 /*
2519 2521 * Notify any threads waiting for the connection that it isn't
2520 2522 * going to happen.
2521 2523 */
2522 2524 cm_entry->x_thread = FALSE;
2523 2525 cv_broadcast(&cm_entry->x_conn_cv);
2524 2526 mutex_exit(&connmgr_lock);
2525 2527
2526 2528 connmgr_release(cm_entry);
2527 2529 }
2528 2530
2529 2531 static void
2530 2532 connmgr_close(struct cm_xprt *cm_entry)
2531 2533 {
2532 2534 mutex_enter(&cm_entry->x_lock);
2533 2535 while (cm_entry->x_ref != 0) {
2534 2536 /*
2535 2537 * Must be a noninterruptible wait.
2536 2538 */
2537 2539 cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2538 2540 }
2539 2541
2540 2542 if (cm_entry->x_tiptr != NULL)
2541 2543 (void) t_kclose(cm_entry->x_tiptr, 1);
2542 2544
2543 2545 mutex_exit(&cm_entry->x_lock);
2544 2546 if (cm_entry->x_ksp != NULL) {
2545 2547 mutex_enter(&connmgr_lock);
2546 2548 cm_entry->x_ksp->ks_private = NULL;
2547 2549 mutex_exit(&connmgr_lock);
2548 2550
2549 2551 /*
2550 2552 * Must free the buffer we allocated for the
2551 2553 * server address in the update function
2552 2554 */
2553 2555 if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2554 2556 x_server.value.str.addr.ptr != NULL)
2555 2557 kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2556 2558 ks_data))->x_server.value.str.addr.ptr,
2557 2559 INET6_ADDRSTRLEN);
2558 2560 kmem_free(cm_entry->x_ksp->ks_data,
2559 2561 cm_entry->x_ksp->ks_data_size);
2560 2562 kstat_delete(cm_entry->x_ksp);
2561 2563 }
2562 2564
2563 2565 mutex_destroy(&cm_entry->x_lock);
2564 2566 cv_destroy(&cm_entry->x_cv);
2565 2567 cv_destroy(&cm_entry->x_conn_cv);
2566 2568 cv_destroy(&cm_entry->x_dis_cv);
2567 2569
2568 2570 if (cm_entry->x_server.buf != NULL)
2569 2571 kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2570 2572 if (cm_entry->x_src.buf != NULL)
2571 2573 kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2572 2574 kmem_free(cm_entry, sizeof (struct cm_xprt));
2573 2575 }
2574 2576
2575 2577 /*
2576 2578 * Called by KRPC after sending the call message to release the connection
2577 2579 * it was using.
2578 2580 */
2579 2581 static void
2580 2582 connmgr_release(struct cm_xprt *cm_entry)
2581 2583 {
2582 2584 mutex_enter(&cm_entry->x_lock);
2583 2585 cm_entry->x_ref--;
2584 2586 if (cm_entry->x_ref == 0)
2585 2587 cv_signal(&cm_entry->x_cv);
2586 2588 mutex_exit(&cm_entry->x_lock);
2587 2589 }
2588 2590
2589 2591 /*
2590 2592 * Set TCP receive and xmit buffer size for RPC connections.
2591 2593 */
2592 2594 static bool_t
2593 2595 connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr)
2594 2596 {
2595 2597 int ok = FALSE;
2596 2598 int val;
2597 2599
2598 2600 if (rpc_default_tcp_bufsz)
2599 2601 return (FALSE);
2600 2602
2601 2603 /*
2602 2604 * Only set new buffer size if it's larger than the system
2603 2605 * default buffer size. If smaller buffer size is needed
2604 2606 * then use /etc/system to set rpc_default_tcp_bufsz to 1.
2605 2607 */
2606 2608 ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_RCVBUF, &val, e, cr);
2607 2609 if ((ok == TRUE) && (val < rpc_send_bufsz)) {
2608 2610 ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_RCVBUF,
2609 2611 rpc_send_bufsz, e, cr);
2610 2612 DTRACE_PROBE2(krpc__i__connmgr_rcvbufsz,
2611 2613 int, ok, calllist_t *, e);
2612 2614 }
2613 2615
2614 2616 ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_SNDBUF, &val, e, cr);
2615 2617 if ((ok == TRUE) && (val < rpc_recv_bufsz)) {
2616 2618 ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_SNDBUF,
2617 2619 rpc_recv_bufsz, e, cr);
2618 2620 DTRACE_PROBE2(krpc__i__connmgr_sndbufsz,
2619 2621 int, ok, calllist_t *, e);
2620 2622 }
2621 2623 return (TRUE);
2622 2624 }
2623 2625
2624 2626 /*
2625 2627 * Given an open stream, connect to the remote. Returns true if connected,
2626 2628 * false otherwise.
2627 2629 */
2628 2630 static bool_t
2629 2631 connmgr_connect(
2630 2632 struct cm_xprt *cm_entry,
2631 2633 queue_t *wq,
2632 2634 struct netbuf *addr,
2633 2635 int addrfmly,
2634 2636 calllist_t *e,
2635 2637 int *tidu_ptr,
2636 2638 bool_t reconnect,
2637 2639 const struct timeval *waitp,
2638 2640 bool_t nosignal,
2639 2641 cred_t *cr)
2640 2642 {
2641 2643 mblk_t *mp;
2642 2644 struct T_conn_req *tcr;
2643 2645 struct T_info_ack *tinfo;
2644 2646 int interrupted, error;
2645 2647 int tidu_size, kstat_instance;
2646 2648
2647 2649 /* if it's a reconnect, flush any lingering data messages */
2648 2650 if (reconnect)
2649 2651 (void) putctl1(wq, M_FLUSH, FLUSHRW);
2650 2652
2651 2653 /*
2652 2654 * Note: if the receiver uses SCM_UCRED/getpeerucred the pid will
2653 2655 * appear as -1.
2654 2656 */
2655 2657 mp = allocb_cred(sizeof (*tcr) + addr->len, cr, NOPID);
2656 2658 if (mp == NULL) {
2657 2659 /*
2658 2660 * This is unfortunate, but we need to look up the stats for
2659 2661 * this zone to increment the "memory allocation failed"
2660 2662 * counter. curproc->p_zone is safe since we're initiating a
2661 2663 * connection and not in some strange streams context.
2662 2664 */
2663 2665 struct rpcstat *rpcstat;
2664 2666
2665 2667 rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
2666 2668 ASSERT(rpcstat != NULL);
2667 2669
2668 2670 RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2669 2671 "sending conn request\n");
2670 2672 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2671 2673 e->call_status = RPC_SYSTEMERROR;
2672 2674 e->call_reason = ENOSR;
2673 2675 return (FALSE);
2674 2676 }
2675 2677
2676 2678 /* Set TCP buffer size for RPC connections if needed */
2677 2679 if (addrfmly == AF_INET || addrfmly == AF_INET6)
2678 2680 (void) connmgr_setbufsz(e, wq, cr);
2679 2681
2680 2682 mp->b_datap->db_type = M_PROTO;
2681 2683 tcr = (struct T_conn_req *)mp->b_rptr;
2682 2684 bzero(tcr, sizeof (*tcr));
2683 2685 tcr->PRIM_type = T_CONN_REQ;
2684 2686 tcr->DEST_length = addr->len;
2685 2687 tcr->DEST_offset = sizeof (struct T_conn_req);
2686 2688 mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2687 2689
2688 2690 bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2689 2691 mp->b_wptr += tcr->DEST_length;
2690 2692
2691 2693 RPCLOG(8, "connmgr_connect: sending conn request on queue "
2692 2694 "%p", (void *)wq);
2693 2695 RPCLOG(8, " call %p\n", (void *)wq);
2694 2696 /*
2695 2697 * We use the entry in the handle that is normally used for
2696 2698 * waiting for RPC replies to wait for the connection accept.
2697 2699 */
2698 2700 if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2699 2701 DTRACE_PROBE(krpc__e__connmgr__connect__cantsend);
2700 2702 freemsg(mp);
2701 2703 return (FALSE);
2702 2704 }
2703 2705
2704 2706 mutex_enter(&clnt_pending_lock);
2705 2707
2706 2708 /*
2707 2709 * We wait for the transport connection to be made, or an
2708 2710 * indication that it could not be made.
2709 2711 */
2710 2712 interrupted = 0;
2711 2713
2712 2714 /*
2713 2715 * waitforack should have been called with T_OK_ACK, but the
2714 2716 * present implementation needs to be passed T_INFO_ACK to
2715 2717 * work correctly.
2716 2718 */
2717 2719 error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2718 2720 if (error == EINTR)
2719 2721 interrupted = 1;
2720 2722 if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2721 2723 /*
2722 2724 * No time to lose; we essentially have been signaled to
2723 2725 * quit.
2724 2726 */
2725 2727 interrupted = 1;
2726 2728 }
2727 2729 #ifdef RPCDEBUG
2728 2730 if (error == ETIME)
2729 2731 RPCLOG0(8, "connmgr_connect: giving up "
2730 2732 "on connection attempt; "
2731 2733 "clnt_dispatch notifyconn "
2732 2734 "diagnostic 'no one waiting for "
2733 2735 "connection' should not be "
2734 2736 "unexpected\n");
2735 2737 #endif
2736 2738 if (e->call_prev)
2737 2739 e->call_prev->call_next = e->call_next;
2738 2740 else
2739 2741 clnt_pending = e->call_next;
2740 2742 if (e->call_next)
2741 2743 e->call_next->call_prev = e->call_prev;
2742 2744 mutex_exit(&clnt_pending_lock);
2743 2745
2744 2746 if (e->call_status != RPC_SUCCESS || error != 0) {
2745 2747 if (interrupted)
2746 2748 e->call_status = RPC_INTR;
2747 2749 else if (error == ETIME)
2748 2750 e->call_status = RPC_TIMEDOUT;
2749 2751 else if (error == EPROTO) {
2750 2752 e->call_status = RPC_SYSTEMERROR;
2751 2753 e->call_reason = EPROTO;
2752 2754 }
2753 2755
2754 2756 RPCLOG(8, "connmgr_connect: can't connect, status: "
2755 2757 "%s\n", clnt_sperrno(e->call_status));
2756 2758
2757 2759 if (e->call_reply) {
2758 2760 freemsg(e->call_reply);
2759 2761 e->call_reply = NULL;
2760 2762 }
2761 2763
2762 2764 return (FALSE);
2763 2765 }
2764 2766 /*
2765 2767 * The result of the "connection accept" is a T_info_ack
2766 2768 * in the call_reply field.
2767 2769 */
2768 2770 ASSERT(e->call_reply != NULL);
2769 2771 mp = e->call_reply;
2770 2772 e->call_reply = NULL;
2771 2773 tinfo = (struct T_info_ack *)mp->b_rptr;
2772 2774
2773 2775 tidu_size = tinfo->TIDU_size;
2774 2776 tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2775 2777 if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2776 2778 tidu_size = COTS_DEFAULT_ALLOCSIZE;
2777 2779 *tidu_ptr = tidu_size;
2778 2780
2779 2781 freemsg(mp);
2780 2782
2781 2783 /*
2782 2784 * Set up the pertinent options. NODELAY is so the transport doesn't
2783 2785 * buffer up RPC messages on either end. This may not be valid for
2784 2786 * all transports. Failure to set this option is not cause to
2785 2787 * bail out so we return success anyway. Note that lack of NODELAY
2786 2788 * or some other way to flush the message on both ends will cause
2787 2789 * lots of retries and terrible performance.
2788 2790 */
2789 2791 if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2790 2792 (void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e, cr);
2791 2793 if (e->call_status == RPC_XPRTFAILED)
2792 2794 return (FALSE);
2793 2795 }
2794 2796
2795 2797 /*
2796 2798 * Since we have a connection, we now need to figure out if
2797 2799 * we need to create a kstat. If x_ksp is not NULL then we
2798 2800 * are reusing a connection and so we do not need to create
2799 2801 * another kstat -- lets just return.
2800 2802 */
2801 2803 if (cm_entry->x_ksp != NULL)
2802 2804 return (TRUE);
2803 2805
2804 2806 /*
2805 2807 * We need to increment rpc_kstat_instance atomically to prevent
2806 2808 * two kstats being created with the same instance.
2807 2809 */
2808 2810 kstat_instance = atomic_inc_32_nv((uint32_t *)&rpc_kstat_instance);
2809 2811
2810 2812 if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2811 2813 "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2812 2814 (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2813 2815 KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2814 2816 return (TRUE);
2815 2817 }
2816 2818
2817 2819 cm_entry->x_ksp->ks_lock = &connmgr_lock;
2818 2820 cm_entry->x_ksp->ks_private = cm_entry;
2819 2821 cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2820 2822 + sizeof (cm_kstat_template));
2821 2823 cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2822 2824 KM_SLEEP);
2823 2825 bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2824 2826 cm_entry->x_ksp->ks_data_size);
2825 2827 ((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2826 2828 x_server.value.str.addr.ptr =
2827 2829 kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2828 2830
2829 2831 cm_entry->x_ksp->ks_update = conn_kstat_update;
2830 2832 kstat_install(cm_entry->x_ksp);
2831 2833 return (TRUE);
2832 2834 }
2833 2835
2834 2836 /*
2835 2837 * Verify that the specified offset falls within the mblk and
2836 2838 * that the resulting pointer is aligned.
2837 2839 * Returns NULL if not.
2838 2840 *
2839 2841 * code from fs/sockfs/socksubr.c
2840 2842 */
2841 2843 static void *
2842 2844 connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
2843 2845 t_uscalar_t length, uint_t align_size)
2844 2846 {
2845 2847 uintptr_t ptr1, ptr2;
2846 2848
2847 2849 ASSERT(mp && mp->b_wptr >= mp->b_rptr);
2848 2850 ptr1 = (uintptr_t)mp->b_rptr + offset;
2849 2851 ptr2 = (uintptr_t)ptr1 + length;
2850 2852 if (ptr1 < (uintptr_t)mp->b_rptr || ptr2 > (uintptr_t)mp->b_wptr) {
2851 2853 return (NULL);
2852 2854 }
2853 2855 if ((ptr1 & (align_size - 1)) != 0) {
2854 2856 return (NULL);
2855 2857 }
2856 2858 return ((void *)ptr1);
2857 2859 }
2858 2860
2859 2861 static bool_t
2860 2862 connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
2861 2863 calllist_t *e, cred_t *cr)
2862 2864 {
2863 2865 mblk_t *mp;
2864 2866 struct opthdr *opt, *opt_res;
2865 2867 struct T_optmgmt_req *tor;
2866 2868 struct T_optmgmt_ack *opt_ack;
2867 2869 struct timeval waitp;
2868 2870 int error;
2869 2871
2870 2872 mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2871 2873 sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2872 2874 if (mp == NULL)
2873 2875 return (FALSE);
2874 2876
2875 2877 mp->b_datap->db_type = M_PROTO;
2876 2878 tor = (struct T_optmgmt_req *)(mp->b_rptr);
2877 2879 tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2878 2880 tor->MGMT_flags = T_CURRENT;
2879 2881 tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2880 2882 tor->OPT_offset = sizeof (struct T_optmgmt_req);
2881 2883
2882 2884 opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2883 2885 opt->level = level;
2884 2886 opt->name = name;
2885 2887 opt->len = sizeof (int);
2886 2888 mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2887 2889 sizeof (int);
2888 2890
2889 2891 /*
2890 2892 * We will use this connection regardless
2891 2893 * of whether or not the option is readable.
2892 2894 */
2893 2895 if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2894 2896 DTRACE_PROBE(krpc__e__connmgr__getopt__cantsend);
2895 2897 freemsg(mp);
2896 2898 return (FALSE);
2897 2899 }
2898 2900
2899 2901 mutex_enter(&clnt_pending_lock);
2900 2902
2901 2903 waitp.tv_sec = clnt_cots_min_conntout;
2902 2904 waitp.tv_usec = 0;
2903 2905 error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2904 2906
2905 2907 if (e->call_prev)
2906 2908 e->call_prev->call_next = e->call_next;
2907 2909 else
2908 2910 clnt_pending = e->call_next;
2909 2911 if (e->call_next)
2910 2912 e->call_next->call_prev = e->call_prev;
2911 2913 mutex_exit(&clnt_pending_lock);
2912 2914
2913 2915 /* get reply message */
2914 2916 mp = e->call_reply;
2915 2917 e->call_reply = NULL;
2916 2918
2917 2919 if ((!mp) || (e->call_status != RPC_SUCCESS) || (error != 0)) {
2918 2920
2919 2921 DTRACE_PROBE4(krpc__e__connmgr_getopt, int, name,
2920 2922 int, e->call_status, int, error, mblk_t *, mp);
2921 2923
2922 2924 if (mp)
2923 2925 freemsg(mp);
2924 2926 return (FALSE);
2925 2927 }
2926 2928
2927 2929 opt_ack = (struct T_optmgmt_ack *)mp->b_rptr;
2928 2930 opt_res = (struct opthdr *)connmgr_opt_getoff(mp, opt_ack->OPT_offset,
2929 2931 opt_ack->OPT_length, __TPI_ALIGN_SIZE);
2930 2932
2931 2933 if (!opt_res) {
2932 2934 DTRACE_PROBE4(krpc__e__connmgr_optres, mblk_t *, mp, int, name,
2933 2935 int, opt_ack->OPT_offset, int, opt_ack->OPT_length);
2934 2936 freemsg(mp);
2935 2937 return (FALSE);
2936 2938 }
2937 2939 *val = *(int *)&opt_res[1];
2938 2940
2939 2941 DTRACE_PROBE2(connmgr_getopt__ok, int, name, int, *val);
2940 2942
2941 2943 freemsg(mp);
2942 2944 return (TRUE);
2943 2945 }
2944 2946
2945 2947 /*
2946 2948 * Called by connmgr_connect to set an option on the new stream.
2947 2949 */
2948 2950 static bool_t
2949 2951 connmgr_setopt_int(queue_t *wq, int level, int name, int val,
2950 2952 calllist_t *e, cred_t *cr)
2951 2953 {
2952 2954 mblk_t *mp;
2953 2955 struct opthdr *opt;
2954 2956 struct T_optmgmt_req *tor;
2955 2957 struct timeval waitp;
2956 2958 int error;
2957 2959
2958 2960 mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2959 2961 sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2960 2962 if (mp == NULL) {
2961 2963 RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2962 2964 "request\n");
2963 2965 return (FALSE);
2964 2966 }
2965 2967
2966 2968 mp->b_datap->db_type = M_PROTO;
2967 2969 tor = (struct T_optmgmt_req *)(mp->b_rptr);
2968 2970 tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2969 2971 tor->MGMT_flags = T_NEGOTIATE;
2970 2972 tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2971 2973 tor->OPT_offset = sizeof (struct T_optmgmt_req);
2972 2974
2973 2975 opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2974 2976 opt->level = level;
2975 2977 opt->name = name;
2976 2978 opt->len = sizeof (int);
2977 2979 *(int *)((char *)opt + sizeof (*opt)) = val;
2978 2980 mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2979 2981 sizeof (int);
2980 2982
2981 2983 /*
2982 2984 * We will use this connection regardless
2983 2985 * of whether or not the option is settable.
2984 2986 */
2985 2987 if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2986 2988 DTRACE_PROBE(krpc__e__connmgr__setopt__cantsend);
2987 2989 freemsg(mp);
2988 2990 return (FALSE);
2989 2991 }
2990 2992
2991 2993 mutex_enter(&clnt_pending_lock);
2992 2994
2993 2995 waitp.tv_sec = clnt_cots_min_conntout;
2994 2996 waitp.tv_usec = 0;
2995 2997 error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2996 2998
2997 2999 if (e->call_prev)
2998 3000 e->call_prev->call_next = e->call_next;
2999 3001 else
3000 3002 clnt_pending = e->call_next;
3001 3003 if (e->call_next)
3002 3004 e->call_next->call_prev = e->call_prev;
3003 3005 mutex_exit(&clnt_pending_lock);
3004 3006
3005 3007 if (e->call_reply != NULL) {
3006 3008 freemsg(e->call_reply);
3007 3009 e->call_reply = NULL;
3008 3010 }
3009 3011
3010 3012 if (e->call_status != RPC_SUCCESS || error != 0) {
3011 3013 RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
3012 3014 return (FALSE);
3013 3015 }
3014 3016 RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
3015 3017 return (TRUE);
3016 3018 }
3017 3019
3018 3020 static bool_t
3019 3021 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e, cred_t *cr)
3020 3022 {
3021 3023 return (connmgr_setopt_int(wq, level, name, 1, e, cr));
3022 3024 }
3023 3025
3024 3026 #ifdef DEBUG
3025 3027
3026 3028 /*
3027 3029 * This is a knob to let us force code coverage in allocation failure
3028 3030 * case.
3029 3031 */
3030 3032 static int connmgr_failsnd;
3031 3033 #define CONN_SND_ALLOC(Size, Pri) \
3032 3034 ((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
3033 3035
3034 3036 #else
3035 3037
3036 3038 #define CONN_SND_ALLOC(Size, Pri) allocb(Size, Pri)
3037 3039
3038 3040 #endif
3039 3041
3040 3042 /*
3041 3043 * Sends an orderly release on the specified queue.
3042 3044 * Entered with connmgr_lock. Exited without connmgr_lock
3043 3045 */
3044 3046 static void
3045 3047 connmgr_sndrel(struct cm_xprt *cm_entry)
3046 3048 {
3047 3049 struct T_ordrel_req *torr;
3048 3050 mblk_t *mp;
3049 3051 queue_t *q = cm_entry->x_wq;
3050 3052 ASSERT(MUTEX_HELD(&connmgr_lock));
3051 3053 mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
3052 3054 if (mp == NULL) {
3053 3055 cm_entry->x_needrel = TRUE;
3054 3056 mutex_exit(&connmgr_lock);
3055 3057 RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
3056 3058 "to queue %p\n", (void *)q);
3057 3059 return;
3058 3060 }
3059 3061 mutex_exit(&connmgr_lock);
3060 3062
3061 3063 mp->b_datap->db_type = M_PROTO;
3062 3064 torr = (struct T_ordrel_req *)(mp->b_rptr);
3063 3065 torr->PRIM_type = T_ORDREL_REQ;
3064 3066 mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
3065 3067
3066 3068 RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
3067 3069 put(q, mp);
3068 3070 }
3069 3071
3070 3072 /*
3071 3073 * Sends an disconnect on the specified queue.
3072 3074 * Entered with connmgr_lock. Exited without connmgr_lock
3073 3075 */
3074 3076 static void
3075 3077 connmgr_snddis(struct cm_xprt *cm_entry)
3076 3078 {
3077 3079 struct T_discon_req *tdis;
3078 3080 mblk_t *mp;
3079 3081 queue_t *q = cm_entry->x_wq;
3080 3082
3081 3083 ASSERT(MUTEX_HELD(&connmgr_lock));
3082 3084 mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
3083 3085 if (mp == NULL) {
3084 3086 cm_entry->x_needdis = TRUE;
3085 3087 mutex_exit(&connmgr_lock);
3086 3088 RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
3087 3089 "to queue %p\n", (void *)q);
3088 3090 return;
3089 3091 }
3090 3092 mutex_exit(&connmgr_lock);
3091 3093
3092 3094 mp->b_datap->db_type = M_PROTO;
3093 3095 tdis = (struct T_discon_req *)mp->b_rptr;
3094 3096 tdis->PRIM_type = T_DISCON_REQ;
3095 3097 mp->b_wptr = mp->b_rptr + sizeof (*tdis);
3096 3098
3097 3099 RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
3098 3100 put(q, mp);
3099 3101 }
3100 3102
3101 3103 /*
3102 3104 * Sets up the entry for receiving replies, and calls rpcmod's write put proc
3103 3105 * (through put) to send the call.
3104 3106 */
3105 3107 static int
3106 3108 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
3107 3109 uint_t queue_flag)
3108 3110 {
3109 3111 ASSERT(e != NULL);
3110 3112
3111 3113 e->call_status = RPC_TIMEDOUT; /* optimistic, eh? */
3112 3114 e->call_reason = 0;
3113 3115 e->call_wq = q;
3114 3116 e->call_xid = xid;
3115 3117 e->call_notified = FALSE;
3116 3118
3117 3119 if (!canput(q)) {
3118 3120 e->call_status = RPC_CANTSEND;
3119 3121 e->call_reason = ENOBUFS;
3120 3122 return (RPC_CANTSEND);
3121 3123 }
3122 3124
3123 3125 /*
3124 3126 * If queue_flag is set then the calllist_t is already on the hash
3125 3127 * queue. In this case just send the message and return.
3126 3128 */
3127 3129 if (queue_flag) {
3128 3130 put(q, mp);
3129 3131 return (RPC_SUCCESS);
3130 3132
3131 3133 }
3132 3134
3133 3135 /*
3134 3136 * Set up calls for RPC requests (with XID != 0) on the hash
3135 3137 * queue for fast lookups and place other calls (i.e.
3136 3138 * connection management) on the linked list.
3137 3139 */
3138 3140 if (xid != 0) {
3139 3141 RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
3140 3142 "dispatch list\n", xid);
3141 3143 e->call_hash = call_hash(xid, clnt_cots_hash_size);
3142 3144 e->call_bucket = &cots_call_ht[e->call_hash];
3143 3145 call_table_enter(e);
3144 3146 } else {
3145 3147 mutex_enter(&clnt_pending_lock);
3146 3148 if (clnt_pending)
3147 3149 clnt_pending->call_prev = e;
3148 3150 e->call_next = clnt_pending;
3149 3151 e->call_prev = NULL;
3150 3152 clnt_pending = e;
3151 3153 mutex_exit(&clnt_pending_lock);
3152 3154 }
3153 3155
3154 3156 put(q, mp);
3155 3157 return (RPC_SUCCESS);
3156 3158 }
3157 3159
3158 3160 /*
3159 3161 * Called by rpcmod to notify a client with a clnt_pending call that its reply
3160 3162 * has arrived. If we can't find a client waiting for this reply, we log
3161 3163 * the error and return.
3162 3164 */
3163 3165 bool_t
3164 3166 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
3165 3167 {
3166 3168 calllist_t *e = NULL;
3167 3169 call_table_t *chtp;
3168 3170 uint32_t xid;
3169 3171 uint_t hash;
3170 3172
3171 3173 if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
3172 3174 (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
3173 3175 xid = *((uint32_t *)mp->b_rptr);
3174 3176 else {
3175 3177 int i = 0;
3176 3178 unsigned char *p = (unsigned char *)&xid;
3177 3179 unsigned char *rptr;
3178 3180 mblk_t *tmp = mp;
3179 3181
3180 3182 /*
3181 3183 * Copy the xid, byte-by-byte into xid.
3182 3184 */
3183 3185 while (tmp) {
3184 3186 rptr = tmp->b_rptr;
3185 3187 while (rptr < tmp->b_wptr) {
3186 3188 *p++ = *rptr++;
3187 3189 if (++i >= sizeof (xid))
3188 3190 goto done_xid_copy;
3189 3191 }
3190 3192 tmp = tmp->b_cont;
3191 3193 }
3192 3194
3193 3195 /*
3194 3196 * If we got here, we ran out of mblk space before the
3195 3197 * xid could be copied.
3196 3198 */
3197 3199 ASSERT(tmp == NULL && i < sizeof (xid));
3198 3200
3199 3201 RPCLOG0(1,
3200 3202 "clnt_dispatch_notify: message less than size of xid\n");
3201 3203 return (FALSE);
3202 3204
3203 3205 }
3204 3206 done_xid_copy:
3205 3207
3206 3208 hash = call_hash(xid, clnt_cots_hash_size);
3207 3209 chtp = &cots_call_ht[hash];
3208 3210 /* call_table_find returns with the hash bucket locked */
3209 3211 call_table_find(chtp, xid, e);
3210 3212
3211 3213 if (e != NULL) {
3212 3214 /*
3213 3215 * Found thread waiting for this reply
3214 3216 */
3215 3217 mutex_enter(&e->call_lock);
3216 3218
3217 3219 /*
3218 3220 * verify that the reply is coming in on
3219 3221 * the same zone that it was sent from.
3220 3222 */
3221 3223 if (e->call_zoneid != zoneid) {
3222 3224 mutex_exit(&e->call_lock);
3223 3225 mutex_exit(&chtp->ct_lock);
3224 3226 RPCLOG0(1, "clnt_dispatch_notify: incorrect zoneid\n");
3225 3227 return (FALSE);
3226 3228 }
3227 3229
3228 3230 if (e->call_reply)
3229 3231 /*
3230 3232 * This can happen under the following scenario:
3231 3233 * clnt_cots_kcallit() times out on the response,
3232 3234 * rfscall() repeats the CLNT_CALL() with
3233 3235 * the same xid, clnt_cots_kcallit() sends the retry,
3234 3236 * thereby putting the clnt handle on the pending list,
3235 3237 * the first response arrives, signalling the thread
3236 3238 * in clnt_cots_kcallit(). Before that thread is
3237 3239 * dispatched, the second response arrives as well,
3238 3240 * and clnt_dispatch_notify still finds the handle on
3239 3241 * the pending list, with call_reply set. So free the
3240 3242 * old reply now.
3241 3243 *
3242 3244 * It is also possible for a response intended for
3243 3245 * an RPC call with a different xid to reside here.
3244 3246 * This can happen if the thread that owned this
3245 3247 * client handle prior to the current owner bailed
3246 3248 * out and left its call record on the dispatch
3247 3249 * queue. A window exists where the response can
3248 3250 * arrive before the current owner dispatches its
3249 3251 * RPC call.
3250 3252 *
3251 3253 * In any case, this is the very last point where we
3252 3254 * can safely check the call_reply field before
3253 3255 * placing the new response there.
3254 3256 */
3255 3257 freemsg(e->call_reply);
3256 3258 e->call_reply = mp;
3257 3259 e->call_status = RPC_SUCCESS;
3258 3260 e->call_notified = TRUE;
3259 3261 cv_signal(&e->call_cv);
3260 3262 mutex_exit(&e->call_lock);
3261 3263 mutex_exit(&chtp->ct_lock);
3262 3264 return (TRUE);
3263 3265 } else {
3264 3266 zone_t *zone;
3265 3267 struct rpcstat *rpcstat;
3266 3268
3267 3269 mutex_exit(&chtp->ct_lock);
3268 3270 RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
3269 3271 xid);
3270 3272 /*
3271 3273 * This is unfortunate, but we need to lookup the zone so we
3272 3274 * can increment its "rcbadxids" counter.
3273 3275 */
3274 3276 zone = zone_find_by_id(zoneid);
3275 3277 if (zone == NULL) {
3276 3278 /*
3277 3279 * The zone went away...
3278 3280 */
3279 3281 return (FALSE);
3280 3282 }
3281 3283 rpcstat = zone_getspecific(rpcstat_zone_key, zone);
3282 3284 if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
3283 3285 /*
3284 3286 * Not interested
3285 3287 */
3286 3288 zone_rele(zone);
3287 3289 return (FALSE);
3288 3290 }
3289 3291 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
3290 3292 zone_rele(zone);
3291 3293 }
3292 3294 return (FALSE);
3293 3295 }
3294 3296
3295 3297 /*
3296 3298 * Called by rpcmod when a non-data indication arrives. The ones in which we
3297 3299 * are interested are connection indications and options acks. We dispatch
3298 3300 * based on the queue the indication came in on. If we are not interested in
3299 3301 * what came in, we return false to rpcmod, who will then pass it upstream.
3300 3302 */
3301 3303 bool_t
3302 3304 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
3303 3305 {
3304 3306 calllist_t *e;
3305 3307 int type;
3306 3308
3307 3309 ASSERT((q->q_flag & QREADR) == 0);
3308 3310
3309 3311 type = ((union T_primitives *)mp->b_rptr)->type;
3310 3312 RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
3311 3313 rpc_tpiprim2name(type));
3312 3314 mutex_enter(&clnt_pending_lock);
3313 3315 for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
3314 3316 if (e == NULL) {
3315 3317 mutex_exit(&clnt_pending_lock);
3316 3318 RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
3317 3319 "for connection on queue 0x%p\n", (void *)q);
3318 3320 return (FALSE);
3319 3321 }
3320 3322 if (e->call_wq == q)
3321 3323 break;
3322 3324 }
3323 3325
3324 3326 switch (type) {
3325 3327 case T_CONN_CON:
3326 3328 /*
3327 3329 * The transport is now connected, send a T_INFO_REQ to get
3328 3330 * the tidu size.
3329 3331 */
3330 3332 mutex_exit(&clnt_pending_lock);
3331 3333 ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
3332 3334 sizeof (struct T_info_req));
3333 3335 mp->b_rptr = mp->b_datap->db_base;
3334 3336 ((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
3335 3337 mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
3336 3338 mp->b_datap->db_type = M_PCPROTO;
3337 3339 put(q, mp);
3338 3340 return (TRUE);
3339 3341 case T_INFO_ACK:
3340 3342 case T_OPTMGMT_ACK:
3341 3343 e->call_status = RPC_SUCCESS;
3342 3344 e->call_reply = mp;
3343 3345 e->call_notified = TRUE;
3344 3346 cv_signal(&e->call_cv);
3345 3347 break;
3346 3348 case T_ERROR_ACK:
3347 3349 e->call_status = RPC_CANTCONNECT;
3348 3350 e->call_reply = mp;
3349 3351 e->call_notified = TRUE;
3350 3352 cv_signal(&e->call_cv);
3351 3353 break;
3352 3354 case T_OK_ACK:
3353 3355 /*
3354 3356 * Great, but we are really waiting for a T_CONN_CON
3355 3357 */
3356 3358 freemsg(mp);
3357 3359 break;
3358 3360 default:
3359 3361 mutex_exit(&clnt_pending_lock);
3360 3362 RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3361 3363 return (FALSE);
3362 3364 }
3363 3365
3364 3366 mutex_exit(&clnt_pending_lock);
3365 3367 return (TRUE);
3366 3368 }
3367 3369
3368 3370 /*
3369 3371 * Called by rpcmod when the transport is (or should be) going away. Informs
3370 3372 * all callers waiting for replies and marks the entry in the connection
3371 3373 * manager's list as unconnected, and either closing (close handshake in
3372 3374 * progress) or dead.
3373 3375 */
3374 3376 void
3375 3377 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3376 3378 {
3377 3379 calllist_t *e;
3378 3380 call_table_t *ctp;
3379 3381 struct cm_xprt *cm_entry;
3380 3382 int have_connmgr_lock;
3381 3383 int i;
3382 3384
3383 3385 ASSERT((q->q_flag & QREADR) == 0);
3384 3386
3385 3387 RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3386 3388 RPCLOG(1, " received a notifcation prim type [%s]",
3387 3389 rpc_tpiprim2name(msg_type));
3388 3390 RPCLOG(1, " and reason %d\n", reason);
3389 3391
3390 3392 /*
3391 3393 * Find the transport entry in the connection manager's list, close
3392 3394 * the transport and delete the entry. In the case where rpcmod's
3393 3395 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3394 3396 * should gracefully close the connection.
3395 3397 */
3396 3398 have_connmgr_lock = 1;
3397 3399 mutex_enter(&connmgr_lock);
3398 3400 for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3399 3401 ASSERT(cm_entry != cm_entry->x_next);
3400 3402 if (cm_entry->x_wq == q) {
3401 3403 ASSERT(MUTEX_HELD(&connmgr_lock));
3402 3404 ASSERT(have_connmgr_lock == 1);
3403 3405 switch (msg_type) {
3404 3406 case T_ORDREL_REQ:
3405 3407
3406 3408 if (cm_entry->x_dead) {
3407 3409 RPCLOG(1, "idle timeout on dead "
3408 3410 "connection: %p\n",
3409 3411 (void *)cm_entry);
3410 3412 if (clnt_stop_idle != NULL)
3411 3413 (*clnt_stop_idle)(q);
3412 3414 break;
3413 3415 }
3414 3416
3415 3417 /*
3416 3418 * Only mark the connection as dead if it is
3417 3419 * connected and idle.
3418 3420 * An unconnected connection has probably
3419 3421 * gone idle because the server is down,
3420 3422 * and when it comes back up there will be
3421 3423 * retries that need to use that connection.
3422 3424 */
3423 3425 if (cm_entry->x_connected ||
3424 3426 cm_entry->x_doomed) {
3425 3427 if (cm_entry->x_ordrel) {
3426 3428 if (cm_entry->x_closing ==
3427 3429 TRUE) {
3428 3430 /*
3429 3431 * The connection is
3430 3432 * obviously wedged due
3431 3433 * to a bug or problem
3432 3434 * with the transport.
3433 3435 * Mark it as dead.
3434 3436 * Otherwise we can
3435 3437 * leak connections.
3436 3438 */
3437 3439 cm_entry->x_dead = TRUE;
3438 3440 mutex_exit(
3439 3441 &connmgr_lock);
3440 3442 have_connmgr_lock = 0;
3441 3443 if (clnt_stop_idle !=
3442 3444 NULL)
3443 3445 (*clnt_stop_idle)(q);
3444 3446 break;
3445 3447 }
3446 3448 cm_entry->x_closing = TRUE;
3447 3449 connmgr_sndrel(cm_entry);
3448 3450 have_connmgr_lock = 0;
3449 3451 } else {
3450 3452 cm_entry->x_dead = TRUE;
3451 3453 mutex_exit(&connmgr_lock);
3452 3454 have_connmgr_lock = 0;
3453 3455 if (clnt_stop_idle != NULL)
3454 3456 (*clnt_stop_idle)(q);
3455 3457 }
3456 3458 } else {
3457 3459 /*
3458 3460 * We don't mark the connection
3459 3461 * as dead, but we turn off the
3460 3462 * idle timer.
3461 3463 */
3462 3464 mutex_exit(&connmgr_lock);
3463 3465 have_connmgr_lock = 0;
3464 3466 if (clnt_stop_idle != NULL)
3465 3467 (*clnt_stop_idle)(q);
3466 3468 RPCLOG(1, "clnt_dispatch_notifyall:"
3467 3469 " ignoring timeout from rpcmod"
3468 3470 " (q %p) because we are not "
3469 3471 " connected\n", (void *)q);
3470 3472 }
3471 3473 break;
3472 3474 case T_ORDREL_IND:
3473 3475 /*
3474 3476 * If this entry is marked closing, then we are
3475 3477 * completing a close handshake, and the
3476 3478 * connection is dead. Otherwise, the server is
3477 3479 * trying to close. Since the server will not
3478 3480 * be sending any more RPC replies, we abort
3479 3481 * the connection, including flushing
3480 3482 * any RPC requests that are in-transit.
3481 3483 * In either case, mark the entry as dead so
3482 3484 * that it can be closed by the connection
3483 3485 * manager's garbage collector.
3484 3486 */
3485 3487 cm_entry->x_dead = TRUE;
3486 3488 if (cm_entry->x_closing) {
3487 3489 mutex_exit(&connmgr_lock);
3488 3490 have_connmgr_lock = 0;
3489 3491 if (clnt_stop_idle != NULL)
3490 3492 (*clnt_stop_idle)(q);
3491 3493 } else {
3492 3494 /*
3493 3495 * if we're getting a disconnect
3494 3496 * before we've finished our
3495 3497 * connect attempt, mark it for
3496 3498 * later processing
3497 3499 */
3498 3500 if (cm_entry->x_thread)
3499 3501 cm_entry->x_early_disc = TRUE;
3500 3502 else
3501 3503 cm_entry->x_connected = FALSE;
3502 3504 cm_entry->x_waitdis = TRUE;
3503 3505 connmgr_snddis(cm_entry);
3504 3506 have_connmgr_lock = 0;
3505 3507 }
3506 3508 break;
3507 3509
3508 3510 case T_ERROR_ACK:
3509 3511 case T_OK_ACK:
3510 3512 cm_entry->x_waitdis = FALSE;
3511 3513 cv_signal(&cm_entry->x_dis_cv);
3512 3514 mutex_exit(&connmgr_lock);
3513 3515 return;
3514 3516
3515 3517 case T_DISCON_REQ:
3516 3518 if (cm_entry->x_thread)
3517 3519 cm_entry->x_early_disc = TRUE;
3518 3520 else
3519 3521 cm_entry->x_connected = FALSE;
3520 3522 cm_entry->x_waitdis = TRUE;
3521 3523
3522 3524 connmgr_snddis(cm_entry);
3523 3525 have_connmgr_lock = 0;
3524 3526 break;
3525 3527
3526 3528 case T_DISCON_IND:
3527 3529 default:
3528 3530 /*
3529 3531 * if we're getting a disconnect before
3530 3532 * we've finished our connect attempt,
3531 3533 * mark it for later processing
3532 3534 */
3533 3535 if (cm_entry->x_closing) {
3534 3536 cm_entry->x_dead = TRUE;
3535 3537 mutex_exit(&connmgr_lock);
3536 3538 have_connmgr_lock = 0;
3537 3539 if (clnt_stop_idle != NULL)
3538 3540 (*clnt_stop_idle)(q);
3539 3541 } else {
3540 3542 if (cm_entry->x_thread) {
3541 3543 cm_entry->x_early_disc = TRUE;
3542 3544 } else {
3543 3545 cm_entry->x_dead = TRUE;
3544 3546 cm_entry->x_connected = FALSE;
3545 3547 }
3546 3548 }
3547 3549 break;
3548 3550 }
3549 3551 break;
3550 3552 }
3551 3553 }
3552 3554
3553 3555 if (have_connmgr_lock)
3554 3556 mutex_exit(&connmgr_lock);
3555 3557
3556 3558 if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3557 3559 RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3558 3560 "connmgr entry for discon ack\n", (void *)q);
3559 3561 return;
3560 3562 }
3561 3563
3562 3564 /*
3563 3565 * Then kick all the clnt_pending calls out of their wait. There
3564 3566 * should be no clnt_pending calls in the case of rpcmod's idle
3565 3567 * timer firing.
3566 3568 */
3567 3569 for (i = 0; i < clnt_cots_hash_size; i++) {
3568 3570 ctp = &cots_call_ht[i];
3569 3571 mutex_enter(&ctp->ct_lock);
3570 3572 for (e = ctp->ct_call_next;
3571 3573 e != (calllist_t *)ctp;
3572 3574 e = e->call_next) {
3573 3575 if (e->call_wq == q && e->call_notified == FALSE) {
3574 3576 RPCLOG(1,
3575 3577 "clnt_dispatch_notifyall for queue %p ",
3576 3578 (void *)q);
3577 3579 RPCLOG(1, "aborting clnt_pending call %p\n",
3578 3580 (void *)e);
3579 3581
3580 3582 if (msg_type == T_DISCON_IND)
3581 3583 e->call_reason = reason;
3582 3584 e->call_notified = TRUE;
3583 3585 e->call_status = RPC_XPRTFAILED;
3584 3586 cv_signal(&e->call_cv);
3585 3587 }
3586 3588 }
3587 3589 mutex_exit(&ctp->ct_lock);
3588 3590 }
3589 3591
3590 3592 mutex_enter(&clnt_pending_lock);
3591 3593 for (e = clnt_pending; e; e = e->call_next) {
3592 3594 /*
3593 3595 * Only signal those RPC handles that haven't been
3594 3596 * signalled yet. Otherwise we can get a bogus call_reason.
3595 3597 * This can happen if thread A is making a call over a
3596 3598 * connection. If the server is killed, it will cause
3597 3599 * reset, and reason will default to EIO as a result of
3598 3600 * a T_ORDREL_IND. Thread B then attempts to recreate
3599 3601 * the connection but gets a T_DISCON_IND. If we set the
3600 3602 * call_reason code for all threads, then if thread A
3601 3603 * hasn't been dispatched yet, it will get the wrong
3602 3604 * reason. The bogus call_reason can make it harder to
3603 3605 * discriminate between calls that fail because the
3604 3606 * connection attempt failed versus those where the call
3605 3607 * may have been executed on the server.
3606 3608 */
3607 3609 if (e->call_wq == q && e->call_notified == FALSE) {
3608 3610 RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3609 3611 (void *)q);
3610 3612 RPCLOG(1, " aborting clnt_pending call %p\n",
3611 3613 (void *)e);
3612 3614
3613 3615 if (msg_type == T_DISCON_IND)
3614 3616 e->call_reason = reason;
3615 3617 e->call_notified = TRUE;
3616 3618 /*
3617 3619 * Let the caller timeout, else it will retry
3618 3620 * immediately.
3619 3621 */
3620 3622 e->call_status = RPC_XPRTFAILED;
3621 3623
3622 3624 /*
3623 3625 * We used to just signal those threads
3624 3626 * waiting for a connection, (call_xid = 0).
3625 3627 * That meant that threads waiting for a response
3626 3628 * waited till their timeout expired. This
3627 3629 * could be a long time if they've specified a
3628 3630 * maximum timeout. (2^31 - 1). So we
3629 3631 * Signal all threads now.
3630 3632 */
3631 3633 cv_signal(&e->call_cv);
3632 3634 }
3633 3635 }
3634 3636 mutex_exit(&clnt_pending_lock);
3635 3637 }
3636 3638
3637 3639
3638 3640 /*ARGSUSED*/
3639 3641 /*
3640 3642 * after resuming a system that's been suspended for longer than the
3641 3643 * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3642 3644 * generates "NFS server X not responding" and "NFS server X ok" messages;
3643 3645 * here we reset inet connections to cause a re-connect and avoid those
3644 3646 * NFS messages. see 4045054
3645 3647 */
3646 3648 boolean_t
3647 3649 connmgr_cpr_reset(void *arg, int code)
3648 3650 {
3649 3651 struct cm_xprt *cxp;
3650 3652
3651 3653 if (code == CB_CODE_CPR_CHKPT)
3652 3654 return (B_TRUE);
3653 3655
3654 3656 if (mutex_tryenter(&connmgr_lock) == 0)
3655 3657 return (B_FALSE);
3656 3658 for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3657 3659 if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3658 3660 cxp->x_connected == TRUE) {
3659 3661 if (cxp->x_thread)
3660 3662 cxp->x_early_disc = TRUE;
3661 3663 else
3662 3664 cxp->x_connected = FALSE;
3663 3665 cxp->x_needdis = TRUE;
3664 3666 }
3665 3667 }
3666 3668 mutex_exit(&connmgr_lock);
3667 3669 return (B_TRUE);
3668 3670 }
3669 3671
3670 3672 void
3671 3673 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3672 3674 {
3673 3675
3674 3676 *statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3675 3677 "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3676 3678 sizeof (cots_rcstat_tmpl));
3677 3679 }
3678 3680
3679 3681 void
3680 3682 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3681 3683 {
3682 3684 rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3683 3685 kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3684 3686 }
3685 3687
3686 3688 void
3687 3689 clnt_cots_init(void)
3688 3690 {
3689 3691 mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3690 3692 mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3691 3693
3692 3694 if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3693 3695 clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3694 3696
3695 3697 cots_call_ht = call_table_init(clnt_cots_hash_size);
3696 3698 zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3697 3699 }
3698 3700
3699 3701 void
3700 3702 clnt_cots_fini(void)
3701 3703 {
3702 3704 (void) zone_key_delete(zone_cots_key);
3703 3705 }
3704 3706
3705 3707 /*
3706 3708 * Wait for TPI ack, returns success only if expected ack is received
3707 3709 * within timeout period.
3708 3710 */
3709 3711
3710 3712 static int
3711 3713 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3712 3714 bool_t nosignal)
3713 3715 {
3714 3716 union T_primitives *tpr;
3715 3717 clock_t timout;
3716 3718 int cv_stat = 1;
3717 3719
3718 3720 ASSERT(MUTEX_HELD(&clnt_pending_lock));
3719 3721 while (e->call_reply == NULL) {
3720 3722 if (waitp != NULL) {
3721 3723 timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3722 3724 drv_usectohz(waitp->tv_usec);
3723 3725 if (nosignal)
3724 3726 cv_stat = cv_reltimedwait(&e->call_cv,
3725 3727 &clnt_pending_lock, timout, TR_CLOCK_TICK);
3726 3728 else
3727 3729 cv_stat = cv_reltimedwait_sig(&e->call_cv,
3728 3730 &clnt_pending_lock, timout, TR_CLOCK_TICK);
3729 3731 } else {
3730 3732 if (nosignal)
3731 3733 cv_wait(&e->call_cv, &clnt_pending_lock);
3732 3734 else
3733 3735 cv_stat = cv_wait_sig(&e->call_cv,
3734 3736 &clnt_pending_lock);
3735 3737 }
3736 3738 if (cv_stat == -1)
3737 3739 return (ETIME);
3738 3740 if (cv_stat == 0)
3739 3741 return (EINTR);
3740 3742 /*
3741 3743 * if we received an error from the server and we know a reply
3742 3744 * is not going to be sent, do not wait for the full timeout,
3743 3745 * return now.
3744 3746 */
3745 3747 if (e->call_status == RPC_XPRTFAILED)
3746 3748 return (e->call_reason);
3747 3749 }
3748 3750 tpr = (union T_primitives *)e->call_reply->b_rptr;
3749 3751 if (tpr->type == ack_prim)
3750 3752 return (0); /* Success */
3751 3753
3752 3754 if (tpr->type == T_ERROR_ACK) {
3753 3755 if (tpr->error_ack.TLI_error == TSYSERR)
3754 3756 return (tpr->error_ack.UNIX_error);
3755 3757 else
3756 3758 return (t_tlitosyserr(tpr->error_ack.TLI_error));
3757 3759 }
3758 3760
3759 3761 return (EPROTO); /* unknown or unexpected primitive */
3760 3762 }
|
↓ open down ↓ |
1350 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX