include/sys/socketvar.h

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21
  22 /*
  23  * Copyright (c) 1996, 2010, Oracle and/or its affiliates. All rights reserved.
  24  */
  25
  26 /*      Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T     */
  27 /*        All Rights Reserved   */
  28
  29 /*
  30  * University Copyright- Copyright (c) 1982, 1986, 1988
  31  * The Regents of the University of California
  32  * All Rights Reserved
  33  *
  34  * University Acknowledgment- Portions of this document are derived from
  35  * software developed by the University of California, Berkeley, and its
  36  * contributors.
  37  */
  38 /*
  39  * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
  40  */
  41
  42 #ifndef _SYS_SOCKETVAR_H
  43 #define _SYS_SOCKETVAR_H
  44
  45 #include <sys/types.h>
  46 #include <sys/stream.h>
  47 #include <sys/t_lock.h>
  48 #include <sys/cred.h>
  49 #include <sys/vnode.h>
  50 #include <sys/file.h>
  51 #include <sys/param.h>
  52 #include <sys/zone.h>
  53 #include <sys/sdt.h>
  54 #include <sys/modctl.h>
  55 #include <sys/atomic.h>
  56 #include <sys/socket.h>
  57 #include <sys/ksocket.h>
  58 #include <sys/kstat.h>
  59 #include <sys/stdbool.h>
  60
  61 #ifdef _KERNEL
  62 #include <sys/vfs.h>
  63 #endif
  64
  65 #ifdef  __cplusplus
  66 extern "C" {
  67 #endif
  68
  69 /*
  70  * Internal representation of the address used to represent addresses
  71  * in the loopback transport for AF_UNIX. While the sockaddr_un is used
  72  * as the sockfs layer address for AF_UNIX the pathnames contained in
  73  * these addresses are not unique (due to relative pathnames) thus can not
  74  * be used in the transport.
  75  *
  76  * The transport level address consists of a magic number (used to separate the
  77  * name space for specific and implicit binds). For a specific bind
  78  * this is followed by a "vnode *" which ensures that all specific binds
  79  * have a unique transport level address. For implicit binds the latter
  80  * part of the address is a byte string (of the same length as a pointer)
  81  * that is assigned by the loopback transport.
  82  *
  83  * The uniqueness assumes that the loopback transport has a separate namespace
  84  * for sockets in order to avoid name conflicts with e.g. TLI use of the
  85  * same transport.
  86  */
  87 struct so_ux_addr {
  88         void    *soua_vp;       /* vnode pointer or assigned by tl */
  89         uint_t  soua_magic;     /* See below */
  90 };
  91
  92 #define SOU_MAGIC_EXPLICIT      0x75787670      /* "uxvp" */
  93 #define SOU_MAGIC_IMPLICIT      0x616e6f6e      /* "anon" */
  94
  95 struct sockaddr_ux {
  96         sa_family_t             sou_family;     /* AF_UNIX */
  97         struct so_ux_addr       sou_addr;
  98 };
  99
 100 #if defined(_KERNEL) || defined(_KMEMUSER)
 101
 102 #include <sys/socket_proto.h>
 103
 104 typedef struct sonodeops sonodeops_t;
 105 typedef struct sonode sonode_t;
 106
 107 struct sodirect_s;
 108
 109 /*
 110  * The sonode represents a socket. A sonode never exist in the file system
 111  * name space and can not be opened using open() - only the socket, socketpair
 112  * and accept calls create sonodes.
 113  *
 114  * The locking of sockfs uses the so_lock mutex plus the SOLOCKED and
 115  * SOREADLOCKED flags in so_flag. The mutex protects all the state in the
 116  * sonode. It is expected that the underlying transport protocol serializes
 117  * socket operations, so sockfs will not normally not single-thread
 118  * operations. However, certain sockets, including TPI based ones, can only
 119  * handle one control operation at a time. The SOLOCKED flag is used to
 120  * single-thread operations from sockfs users to prevent e.g. multiple bind()
 121  * calls to operate on the same sonode concurrently. The SOREADLOCKED flag is
 122  * used to ensure that only one thread sleeps in kstrgetmsg for a given
 123  * sonode. This is needed to ensure atomic operation for things like
 124  * MSG_WAITALL.
 125  *
 126  * The so_fallback_rwlock is used to ensure that for sockets that can
 127  * fall back to TPI, the fallback is not initiated until all pending
 128  * operations have completed.
 129  *
 130  * Note that so_lock is sometimes held across calls that might go to sleep
 131  * (kmem_alloc and soallocproto*). This implies that no other lock in
 132  * the system should be held when calling into sockfs; from the system call
 133  * side or from strrput (in case of TPI based sockets). If locks are held
 134  * while calling into sockfs the system might hang when running low on memory.
 135  */
 136 struct sonode {
 137         struct  vnode   *so_vnode;      /* vnode associated with this sonode */
 138
 139         sonodeops_t     *so_ops;        /* operations vector for this sonode */
 140         void            *so_priv;       /* sonode private data */
 141
 142         krwlock_t       so_fallback_rwlock;
 143         kmutex_t        so_lock;        /* protects sonode fields */
 144
 145         kcondvar_t      so_state_cv;    /* synchronize state changes */
 146         kcondvar_t      so_single_cv;   /* wait due to SOLOCKED */
 147         kcondvar_t      so_read_cv;     /* wait due to SOREADLOCKED */
 148
 149         /* These fields are protected by so_lock */
 150
 151         uint_t          so_state;       /* internal state flags SS_*, below */
 152         uint_t          so_mode;        /* characteristics on socket. SM_* */
 153         ushort_t        so_flag;        /* flags, see below */
 154         int             so_count;       /* count of opened references */
 155
 156         sock_connid_t   so_proto_connid; /* protocol generation number */
 157
 158         ushort_t        so_error;       /* error affecting connection */
 159
 160         struct sockparams *so_sockparams;       /* vnode or socket module */
 161         /* Needed to recreate the same socket for accept */
 162         short   so_family;
 163         short   so_type;
 164         short   so_protocol;
 165         /*
 166          * XXX: We removed socket versions, but SOV_STREAM, which was
 167          * documented as "Not a socket - just a stream", is special. We use
 168          * this flag as a direct replacement for now, but drop the SOV_* and
 169          * "version" terminology for clarity.
 170          */
 171         bool    so_is_stream;
 172
 173         /* Accept queue */
 174         kmutex_t        so_acceptq_lock;        /* protects accept queue */
 175         list_t          so_acceptq_list;        /* pending conns */
 176         list_t          so_acceptq_defer;       /* deferred conns */
 177         list_node_t     so_acceptq_node;        /* acceptq list node */
 178         unsigned int    so_acceptq_len;         /* # of conns (both lists) */
 179         unsigned int    so_backlog;             /* Listen backlog */
 180         kcondvar_t      so_acceptq_cv;          /* wait for new conn. */
 181         struct sonode   *so_listener;           /* parent socket */
 182
 183         /* Options */
 184         short   so_options;             /* From socket call, see socket.h */
 185         struct linger   so_linger;      /* SO_LINGER value */
 186 #define so_sndbuf       so_proto_props.sopp_txhiwat     /* SO_SNDBUF value */
 187 #define so_sndlowat     so_proto_props.sopp_txlowat     /* tx low water mark */
 188 #define so_rcvbuf       so_proto_props.sopp_rxhiwat     /* SO_RCVBUF value */
 189 #define so_rcvlowat     so_proto_props.sopp_rxlowat     /* rx low water mark */
 190 #define so_max_addr_len so_proto_props.sopp_maxaddrlen
 191 #define so_minpsz       so_proto_props.sopp_minpsz
 192 #define so_maxpsz       so_proto_props.sopp_maxpsz
 193
 194         clock_t so_sndtimeo;            /* send timeout */
 195         clock_t so_rcvtimeo;            /* recv timeout */
 196
 197         mblk_t  *so_oobmsg;             /* outofline oob data */
 198         ssize_t so_oobmark;             /* offset of the oob data */
 199
 200         pid_t   so_pgrp;                /* pgrp for signals */
 201
 202         cred_t          *so_peercred;   /* connected socket peer cred */
 203         pid_t           so_cpid;        /* connected socket peer cached pid */
 204         zoneid_t        so_zoneid;      /* opener's zoneid */
 205
 206         struct pollhead so_poll_list;   /* common pollhead */
 207         short           so_pollev;      /* events that should be generated */
 208
 209         /* Receive */
 210         unsigned int    so_rcv_queued;  /* # bytes on both rcv lists */
 211         mblk_t          *so_rcv_q_head; /* processing/copyout rcv queue */
 212         mblk_t          *so_rcv_q_last_head;
 213         mblk_t          *so_rcv_head;   /* protocol prequeue */
 214         mblk_t          *so_rcv_last_head;      /* last mblk in b_next chain */
 215         kcondvar_t      so_rcv_cv;      /* wait for data */
 216         uint_t          so_rcv_wanted;  /* # of bytes wanted by app */
 217         timeout_id_t    so_rcv_timer_tid;
 218
 219 #define so_rcv_thresh   so_proto_props.sopp_rcvthresh
 220 #define so_rcv_timer_interval so_proto_props.sopp_rcvtimer
 221
 222         kcondvar_t      so_snd_cv;      /* wait for snd buffers */
 223         uint32_t
 224                 so_snd_qfull: 1,        /* Transmit full */
 225                 so_rcv_wakeup: 1,
 226                 so_snd_wakeup: 1,
 227                 so_not_str: 1,  /* B_TRUE if not streams based socket */
 228                 so_pad_to_bit_31: 28;
 229
 230         /* Communication channel with protocol */
 231         sock_lower_handle_t     so_proto_handle;
 232         sock_downcalls_t        *so_downcalls;
 233
 234         struct sock_proto_props so_proto_props; /* protocol settings */
 235         boolean_t               so_flowctrld;   /* Flow controlled */
 236         uint_t                  so_copyflag;    /* Copy related flag */
 237         kcondvar_t              so_copy_cv;     /* Copy cond variable */
 238
 239         /* kernel sockets */
 240         ksocket_callbacks_t     so_ksock_callbacks;
 241         void                    *so_ksock_cb_arg;       /* callback argument */
 242         kcondvar_t              so_closing_cv;
 243
 244         /* != NULL for sodirect enabled socket */
 245         struct sodirect_s       *so_direct;
 246
 247         /* socket filters */
 248         uint_t                  so_filter_active;       /* # of active fil */
 249         uint_t                  so_filter_tx;           /* pending tx ops */
 250         struct sof_instance     *so_filter_top;         /* top of stack */
 251         struct sof_instance     *so_filter_bottom;      /* bottom of stack */
 252         clock_t                 so_filter_defertime;    /* time when deferred */
 253 };
 254
 255 #define SO_HAVE_DATA(so)                                                \
 256         /*                                                              \
 257          * For the (tid == 0) case we must check so_rcv_{q_,}head       \
 258          * rather than (so_rcv_queued > 0), since the latter does not   \
 259          * take into account mblks with only control/name information.  \
 260          */                                                             \
 261         ((so)->so_rcv_timer_tid == 0 && ((so)->so_rcv_head != NULL ||   \
 262         (so)->so_rcv_q_head != NULL)) ||                                \
 263         ((so)->so_state & SS_CANTRCVMORE)
 264
 265 /*
 266  * Events handled by the protocol (in case sd_poll is set)
 267  */
 268 #define SO_PROTO_POLLEV         (POLLIN|POLLRDNORM|POLLRDBAND)
 269
 270
 271 #endif /* _KERNEL || _KMEMUSER */
 272
 273 /* flags */
 274 #define SOMOD           0x0001          /* update socket modification time */
 275 #define SOACC           0x0002          /* update socket access time */
 276
 277 #define SOLOCKED        0x0010          /* use to serialize open/closes */
 278 #define SOREADLOCKED    0x0020          /* serialize kstrgetmsg calls */
 279 #define SOCLONE         0x0040          /* child of clone driver */
 280 #define SOASYNC_UNBIND  0x0080          /* wait for ACK of async unbind */
 281
 282 #define SOCK_IS_NONSTR(so)      ((so)->so_not_str)
 283
 284 /*
 285  * Socket state bits.
 286  */
 287 #define SS_ISCONNECTED          0x00000001 /* socket connected to a peer */
 288 #define SS_ISCONNECTING         0x00000002 /* in process, connecting to peer */
 289 #define SS_ISDISCONNECTING      0x00000004 /* in process of disconnecting */
 290 #define SS_CANTSENDMORE         0x00000008 /* can't send more data to peer */
 291
 292 #define SS_CANTRCVMORE          0x00000010 /* can't receive more data */
 293 #define SS_ISBOUND              0x00000020 /* socket is bound */
 294 #define SS_NDELAY               0x00000040 /* FNDELAY non-blocking */
 295 #define SS_NONBLOCK             0x00000080 /* O_NONBLOCK non-blocking */
 296
 297 #define SS_ASYNC                0x00000100 /* async i/o notify */
 298 #define SS_ACCEPTCONN           0x00000200 /* listen done */
 299 /*      unused                  0x00000400 */   /* was SS_HASCONNIND */
 300 #define SS_SAVEDEOR             0x00000800 /* Saved MSG_EOR rcv side state */
 301
 302 #define SS_RCVATMARK            0x00001000 /* at mark on input */
 303 #define SS_OOBPEND              0x00002000 /* OOB pending or present - poll */
 304 #define SS_HAVEOOBDATA          0x00004000 /* OOB data present */
 305 #define SS_HADOOBDATA           0x00008000 /* OOB data consumed */
 306 #define SS_CLOSING              0x00010000 /* in process of closing */
 307
 308 #define SS_FIL_DEFER            0x00020000 /* filter deferred notification */
 309 #define SS_FILOP_OK             0x00040000 /* socket can attach filters */
 310 #define SS_FIL_RCV_FLOWCTRL     0x00080000 /* filter asserted rcv flow ctrl */
 311 #define SS_FIL_SND_FLOWCTRL     0x00100000 /* filter asserted snd flow ctrl */
 312 #define SS_FIL_STOP             0x00200000 /* no more filter actions */
 313
 314 #define SS_SODIRECT             0x00400000 /* transport supports sodirect */
 315
 316 #define SS_SENTLASTREADSIG      0x01000000 /* last rx signal has been sent */
 317 #define SS_SENTLASTWRITESIG     0x02000000 /* last tx signal has been sent */
 318
 319 #define SS_FALLBACK_DRAIN       0x20000000 /* data was/is being drained */
 320 #define SS_FALLBACK_PENDING     0x40000000 /* fallback is pending */
 321 #define SS_FALLBACK_COMP        0x80000000 /* fallback has completed */
 322
 323
 324 /* Set of states when the socket can't be rebound */
 325 #define SS_CANTREBIND   (SS_ISCONNECTED|SS_ISCONNECTING|SS_ISDISCONNECTING|\
 326                             SS_CANTSENDMORE|SS_CANTRCVMORE|SS_ACCEPTCONN)
 327
 328 /*
 329  * Sockets that can fall back to TPI must ensure that fall back is not
 330  * initiated while a thread is using a socket.
 331  */
 332 #define SO_BLOCK_FALLBACK(so, fn)                               \
 333         ASSERT(MUTEX_NOT_HELD(&(so)->so_lock));                 \
 334         rw_enter(&(so)->so_fallback_rwlock, RW_READER);         \
 335         if ((so)->so_state & (SS_FALLBACK_COMP|SS_FILOP_OK)) {  \
 336                 if ((so)->so_state & SS_FALLBACK_COMP) {        \
 337                         rw_exit(&(so)->so_fallback_rwlock);     \
 338                         return (fn);                            \
 339                 } else {                                        \
 340                         mutex_enter(&(so)->so_lock);            \
 341                         (so)->so_state &= ~SS_FILOP_OK;         \
 342                         mutex_exit(&(so)->so_lock);             \
 343                 }                                               \
 344         }
 345
 346 #define SO_UNBLOCK_FALLBACK(so) {                       \
 347         rw_exit(&(so)->so_fallback_rwlock);             \
 348 }
 349
 350 #define SO_SND_FLOWCTRLD(so)    \
 351         ((so)->so_snd_qfull || (so)->so_state & SS_FIL_SND_FLOWCTRL)
 352
 353 /* Poll events */
 354 #define SO_POLLEV_IN            0x1     /* POLLIN wakeup needed */
 355 #define SO_POLLEV_ALWAYS        0x2     /* wakeups */
 356
 357 /*
 358  * Characteristics of sockets. Not changed after the socket is created.
 359  */
 360 #define SM_PRIV                 0x001   /* privileged for broadcast, raw... */
 361 #define SM_ATOMIC               0x002   /* atomic data transmission */
 362 #define SM_ADDR                 0x004   /* addresses given with messages */
 363 #define SM_CONNREQUIRED         0x008   /* connection required by protocol */
 364
 365 #define SM_FDPASSING            0x010   /* passes file descriptors */
 366 #define SM_EXDATA               0x020   /* Can handle T_EXDATA_REQ */
 367 #define SM_OPTDATA              0x040   /* Can handle T_OPTDATA_REQ */
 368 #define SM_BYTESTREAM           0x080   /* Byte stream - can use M_DATA */
 369
 370 #define SM_ACCEPTOR_ID          0x100   /* so_acceptor_id is valid */
 371
 372 #define SM_KERNEL               0x200   /* kernel socket */
 373
 374 /* The modes below are only for non-streams sockets */
 375 #define SM_ACCEPTSUPP           0x400   /* can handle accept() */
 376 #define SM_SENDFILESUPP         0x800   /* Private: proto supp sendfile  */
 377
 378 #if defined(_KERNEL) || defined(_KMEMUSER)
 379
 380 /*
 381  * sonode create and destroy functions.
 382  */
 383 typedef struct sonode *(*so_create_func_t)(struct sockparams *,
 384     int, int, int, int, int *, cred_t *);
 385 typedef void (*so_destroy_func_t)(struct sonode *);
 386
 387 /* STREAM device information */
 388 typedef struct sdev_info {
 389         char    *sd_devpath;
 390         int     sd_devpathlen; /* Is 0 if sp_devpath is a static string */
 391         vnode_t *sd_vnode;
 392 } sdev_info_t;
 393
 394 #define SOCKMOD_VERSION_1       1
 395 #define SOCKMOD_VERSION         2
 396
 397 /* name of the TPI pseudo socket module */
 398 #define SOTPI_SMOD_NAME         "socktpi"
 399
 400 typedef struct __smod_priv_s {
 401         so_create_func_t        smodp_sock_create_func;
 402         so_destroy_func_t       smodp_sock_destroy_func;
 403         so_proto_fallback_func_t smodp_proto_fallback_func;
 404         const char              *smodp_fallback_devpath_v4;
 405         const char              *smodp_fallback_devpath_v6;
 406 } __smod_priv_t;
 407
 408 /*
 409  * Socket module register information
 410  */
 411 typedef struct smod_reg_s {
 412         int             smod_version;
 413         char            *smod_name;
 414         size_t          smod_uc_version;
 415         size_t          smod_dc_version;
 416         so_proto_create_func_t  smod_proto_create_func;
 417
 418         /* __smod_priv_data must be NULL */
 419         __smod_priv_t   *__smod_priv;
 420 } smod_reg_t;
 421
 422 /*
 423  * Socket module information
 424  */
 425 typedef struct smod_info {
 426         int             smod_version;
 427         char            *smod_name;
 428         uint_t          smod_refcnt;            /* # of entries */
 429         size_t          smod_uc_version;        /* upcall version */
 430         size_t          smod_dc_version;        /* down call version */
 431         so_proto_create_func_t  smod_proto_create_func;
 432         so_proto_fallback_func_t smod_proto_fallback_func;
 433         const char              *smod_fallback_devpath_v4;
 434         const char              *smod_fallback_devpath_v6;
 435         so_create_func_t        smod_sock_create_func;
 436         so_destroy_func_t       smod_sock_destroy_func;
 437         list_node_t     smod_node;
 438 } smod_info_t;
 439
 440 typedef struct sockparams_stats {
 441         kstat_named_t   sps_nfallback;  /* # of fallbacks to TPI */
 442         kstat_named_t   sps_nactive;    /* # of active sockets */
 443         kstat_named_t   sps_ncreate;    /* total # of created sockets */
 444 } sockparams_stats_t;
 445
 446 /*
 447  * sockparams
 448  *
 449  * Used for mapping family/type/protocol to a socket module or STREAMS device
 450  */
 451 struct sockparams {
 452         /*
 453          * The family, type, protocol, sdev_info and smod_name are
 454          * set when the entry is created, and they will never change
 455          * thereafter.
 456          */
 457         int             sp_family;
 458         int             sp_type;
 459         int             sp_protocol;
 460
 461         sdev_info_t     sp_sdev_info;   /* STREAM device */
 462         char            *sp_smod_name;  /* socket module name */
 463
 464         kmutex_t        sp_lock;        /* lock for refcnt and smod_info */
 465         uint64_t        sp_refcnt;      /* entry reference count */
 466         smod_info_t     *sp_smod_info;  /* socket module */
 467
 468         sockparams_stats_t sp_stats;
 469         kstat_t         *sp_kstat;
 470
 471         /*
 472          * The entries below are only modified while holding
 473          * sockconf_lock as a writer.
 474          */
 475         int             sp_flags;       /* see below */
 476         list_node_t     sp_node;
 477
 478         list_t          sp_auto_filters; /* list of automatic filters */
 479         list_t          sp_prog_filters; /* list of programmatic filters */
 480 };
 481
 482 struct sof_entry;
 483
 484 typedef struct sp_filter {
 485         struct sof_entry *spf_filter;
 486         list_node_t     spf_node;
 487 } sp_filter_t;
 488
 489
 490 /*
 491  * sockparams flags
 492  */
 493 #define SOCKPARAMS_EPHEMERAL    0x1     /* temp. entry, not on global list */
 494
 495 extern void sockparams_init(void);
 496 extern struct sockparams *sockparams_hold_ephemeral_bydev(int, int, int,
 497     const char *, int, int *);
 498 extern struct sockparams *sockparams_hold_ephemeral_bymod(int, int, int,
 499     const char *, int, int *);
 500 extern void sockparams_ephemeral_drop_last_ref(struct sockparams *);
 501
 502 extern struct sockparams *sockparams_create(int, int, int, char *, char *, int,
 503     int, int, int *);
 504 extern void     sockparams_destroy(struct sockparams *);
 505 extern int      sockparams_add(struct sockparams *);
 506 extern int      sockparams_delete(int, int, int);
 507 extern int      sockparams_new_filter(struct sof_entry *);
 508 extern void     sockparams_filter_cleanup(struct sof_entry *);
 509 extern int      sockparams_copyout_socktable(uintptr_t);
 510
 511 extern void smod_init(void);
 512 extern void smod_add(smod_info_t *);
 513 extern int smod_register(const smod_reg_t *);
 514 extern int smod_unregister(const char *);
 515 extern smod_info_t *smod_lookup_byname(const char *);
 516
 517 #define SOCKPARAMS_HAS_DEVICE(sp)                                       \
 518         ((sp)->sp_sdev_info.sd_devpath != NULL)
 519
 520 /* Increase the smod_info_t reference count */
 521 #define SMOD_INC_REF(smodp) {                                           \
 522         ASSERT((smodp) != NULL);                                        \
 523         DTRACE_PROBE1(smodinfo__inc__ref, struct smod_info *, (smodp)); \
 524         atomic_inc_uint(&(smodp)->smod_refcnt);                         \
 525 }
 526
 527 /*
 528  * Decreace the socket module entry reference count.
 529  * When no one mapping to the entry, we try to unload the module from the
 530  * kernel. If the module can't unload, just leave the module entry with
 531  * a zero refcnt.
 532  */
 533 #define SMOD_DEC_REF(smodp, modname) {                                  \
 534         ASSERT((smodp) != NULL);                                        \
 535         ASSERT((smodp)->smod_refcnt != 0);                              \
 536         atomic_dec_uint(&(smodp)->smod_refcnt);                         \
 537         /*                                                              \
 538          * No need to atomically check the return value because the     \
 539          * socket module framework will verify that no one is using     \
 540          * the module before unloading. Worst thing that can happen     \
 541          * here is multiple calls to mod_remove_by_name(), which is OK. \
 542          */                                                             \
 543         if ((smodp)->smod_refcnt == 0)                                  \
 544                 (void) mod_remove_by_name(modname);                     \
 545 }
 546
 547 /* Increase the reference count */
 548 #define SOCKPARAMS_INC_REF(sp) {                                        \
 549         ASSERT((sp) != NULL);                                           \
 550         DTRACE_PROBE1(sockparams__inc__ref, struct sockparams *, (sp)); \
 551         mutex_enter(&(sp)->sp_lock);                                    \
 552         (sp)->sp_refcnt++;                                              \
 553         ASSERT((sp)->sp_refcnt != 0);                                   \
 554         mutex_exit(&(sp)->sp_lock);                                     \
 555 }
 556
 557 /*
 558  * Decrease the reference count.
 559  *
 560  * If the sockparams is ephemeral, then the thread dropping the last ref
 561  * count will destroy the entry.
 562  */
 563 #define SOCKPARAMS_DEC_REF(sp) {                                        \
 564         ASSERT((sp) != NULL);                                           \
 565         DTRACE_PROBE1(sockparams__dec__ref, struct sockparams *, (sp)); \
 566         mutex_enter(&(sp)->sp_lock);                                    \
 567         ASSERT((sp)->sp_refcnt > 0);                                    \
 568         if ((sp)->sp_refcnt == 1) {                                     \
 569                 if ((sp)->sp_flags & SOCKPARAMS_EPHEMERAL) {            \
 570                         mutex_exit(&(sp)->sp_lock);                     \
 571                         sockparams_ephemeral_drop_last_ref((sp));       \
 572                 } else {                                                \
 573                         (sp)->sp_refcnt--;                              \
 574                         if ((sp)->sp_smod_info != NULL) {               \
 575                                 SMOD_DEC_REF((sp)->sp_smod_info,        \
 576                                     (sp)->sp_smod_name);                \
 577                         }                                               \
 578                         (sp)->sp_smod_info = NULL;                      \
 579                         mutex_exit(&(sp)->sp_lock);                     \
 580                 }                                                       \
 581         } else {                                                        \
 582                 (sp)->sp_refcnt--;                                      \
 583                 mutex_exit(&(sp)->sp_lock);                             \
 584         }                                                               \
 585 }
 586
 587 /*
 588  * Used to traverse the list of AF_UNIX sockets to construct the kstat
 589  * for netstat(1m).
 590  */
 591 struct socklist {
 592         kmutex_t        sl_lock;
 593         struct sonode   *sl_list;
 594 };
 595
 596 extern struct socklist socklist;
 597 /*
 598  * ss_full_waits is the number of times the reader thread
 599  * waits when the queue is full and ss_empty_waits is the number
 600  * of times the consumer thread waits when the queue is empty.
 601  * No locks for these as they are just indicators of whether
 602  * disk or network or both is slow or fast.
 603  */
 604 struct sendfile_stats {
 605         uint32_t ss_file_cached;
 606         uint32_t ss_file_not_cached;
 607         uint32_t ss_full_waits;
 608         uint32_t ss_empty_waits;
 609         uint32_t ss_file_segmap;
 610 };
 611
 612 /*
 613  * A single sendfile request is represented by snf_req.
 614  */
 615 typedef struct snf_req {
 616         struct snf_req  *sr_next;
 617         mblk_t          *sr_mp_head;
 618         mblk_t          *sr_mp_tail;
 619         kmutex_t        sr_lock;
 620         kcondvar_t      sr_cv;
 621         uint_t          sr_qlen;
 622         int             sr_hiwat;
 623         int             sr_lowat;
 624         int             sr_operation;
 625         struct vnode    *sr_vp;
 626         file_t          *sr_fp;
 627         ssize_t         sr_maxpsz;
 628         uoff_t          sr_file_off;
 629         uoff_t          sr_file_size;
 630 #define SR_READ_DONE    0x80000000
 631         int             sr_read_error;
 632         int             sr_write_error;
 633 } snf_req_t;
 634
 635 /* A queue of sendfile requests */
 636 struct sendfile_queue {
 637         snf_req_t       *snfq_req_head;
 638         snf_req_t       *snfq_req_tail;
 639         kmutex_t        snfq_lock;
 640         kcondvar_t      snfq_cv;
 641         int             snfq_svc_threads;       /* # of service threads */
 642         int             snfq_idle_cnt;          /* # of idling threads */
 643         int             snfq_max_threads;
 644         int             snfq_req_cnt;           /* Number of requests */
 645 };
 646
 647 #define READ_OP                 1
 648 #define SNFQ_TIMEOUT            (60 * 5 * hz)   /* 5 minutes */
 649
 650 /* Socket network operations switch */
 651 struct sonodeops {
 652         int     (*sop_init)(struct sonode *, struct sonode *, cred_t *,
 653                     int);
 654         int     (*sop_accept)(struct sonode *, int, cred_t *, struct sonode **);
 655         int     (*sop_bind)(struct sonode *, struct sockaddr *, socklen_t,
 656                     int, cred_t *);
 657         int     (*sop_listen)(struct sonode *, int, cred_t *);
 658         int     (*sop_connect)(struct sonode *, struct sockaddr *,
 659                     socklen_t, int, int, cred_t *);
 660         int     (*sop_recvmsg)(struct sonode *, struct msghdr *,
 661                     struct uio *, cred_t *);
 662         int     (*sop_sendmsg)(struct sonode *, struct msghdr *,
 663                     struct uio *, cred_t *);
 664         int     (*sop_sendmblk)(struct sonode *, struct msghdr *, int,
 665                     cred_t *, mblk_t **);
 666         int     (*sop_getpeername)(struct sonode *, struct sockaddr *,
 667                     socklen_t *, boolean_t, cred_t *);
 668         int     (*sop_getsockname)(struct sonode *, struct sockaddr *,
 669                     socklen_t *, cred_t *);
 670         int     (*sop_shutdown)(struct sonode *, int, cred_t *);
 671         int     (*sop_getsockopt)(struct sonode *, int, int, void *,
 672                     socklen_t *, int, cred_t *);
 673         int     (*sop_setsockopt)(struct sonode *, int, int, const void *,
 674                     socklen_t, cred_t *);
 675         int     (*sop_ioctl)(struct sonode *, int, intptr_t, int,
 676                     cred_t *, int32_t *);
 677         int     (*sop_poll)(struct sonode *, short, int, short *,
 678                     struct pollhead **);
 679         int     (*sop_close)(struct sonode *, int, cred_t *);
 680 };
 681
 682 #define SOP_INIT(so, flag, cr, flags)   \
 683         ((so)->so_ops->sop_init((so), (flag), (cr), (flags)))
 684 #define SOP_ACCEPT(so, fflag, cr, nsop) \
 685         ((so)->so_ops->sop_accept((so), (fflag), (cr), (nsop)))
 686 #define SOP_BIND(so, name, namelen, flags, cr)  \
 687         ((so)->so_ops->sop_bind((so), (name), (namelen), (flags), (cr)))
 688 #define SOP_LISTEN(so, backlog, cr)     \
 689         ((so)->so_ops->sop_listen((so), (backlog), (cr)))
 690 #define SOP_CONNECT(so, name, namelen, fflag, flags, cr)        \
 691         ((so)->so_ops->sop_connect((so), (name), (namelen), (fflag), (flags), \
 692         (cr)))
 693 #define SOP_RECVMSG(so, msg, uiop, cr)  \
 694         ((so)->so_ops->sop_recvmsg((so), (msg), (uiop), (cr)))
 695 #define SOP_SENDMSG(so, msg, uiop, cr)  \
 696         ((so)->so_ops->sop_sendmsg((so), (msg), (uiop), (cr)))
 697 #define SOP_SENDMBLK(so, msg, size, cr, mpp)    \
 698         ((so)->so_ops->sop_sendmblk((so), (msg), (size), (cr), (mpp)))
 699 #define SOP_GETPEERNAME(so, addr, addrlen, accept, cr)  \
 700         ((so)->so_ops->sop_getpeername((so), (addr), (addrlen), (accept), (cr)))
 701 #define SOP_GETSOCKNAME(so, addr, addrlen, cr)  \
 702         ((so)->so_ops->sop_getsockname((so), (addr), (addrlen), (cr)))
 703 #define SOP_SHUTDOWN(so, how, cr)       \
 704         ((so)->so_ops->sop_shutdown((so), (how), (cr)))
 705 #define SOP_GETSOCKOPT(so, level, optionname, optval, optlenp, flags, cr) \
 706         ((so)->so_ops->sop_getsockopt((so), (level), (optionname),      \
 707             (optval), (optlenp), (flags), (cr)))
 708 #define SOP_SETSOCKOPT(so, level, optionname, optval, optlen, cr)       \
 709         ((so)->so_ops->sop_setsockopt((so), (level), (optionname),      \
 710             (optval), (optlen), (cr)))
 711 #define SOP_IOCTL(so, cmd, arg, mode, cr, rvalp)        \
 712         ((so)->so_ops->sop_ioctl((so), (cmd), (arg), (mode), (cr), (rvalp)))
 713 #define SOP_POLL(so, events, anyyet, reventsp, phpp) \
 714         ((so)->so_ops->sop_poll((so), (events), (anyyet), (reventsp), (phpp)))
 715 #define SOP_CLOSE(so, flag, cr) \
 716         ((so)->so_ops->sop_close((so), (flag), (cr)))
 717
 718 #endif /* defined(_KERNEL) || defined(_KMEMUSER) */
 719
 720 #ifdef _KERNEL
 721
 722 #define ISALIGNED_cmsghdr(addr) \
 723                 (((uintptr_t)(addr) & (_CMSG_HDR_ALIGNMENT - 1)) == 0)
 724
 725 #define ROUNDUP_cmsglen(len) \
 726         (((len) + _CMSG_HDR_ALIGNMENT - 1) & ~(_CMSG_HDR_ALIGNMENT - 1))
 727
 728 #define IS_NON_STREAM_SOCK(vp) \
 729         ((vp)->v_type == VSOCK && (vp)->v_stream == NULL)
 730 /*
 731  * Macros that operate on struct cmsghdr.
 732  * Used in parsing msg_control.
 733  * The CMSG_VALID macro does not assume that the last option buffer is padded.
 734  */
 735 #define CMSG_NEXT(cmsg)                                         \
 736         (struct cmsghdr *)((uintptr_t)(cmsg) +                  \
 737             ROUNDUP_cmsglen((cmsg)->cmsg_len))
 738 #define CMSG_CONTENT(cmsg)      (&((cmsg)[1]))
 739 #define CMSG_CONTENTLEN(cmsg)   ((cmsg)->cmsg_len - sizeof (struct cmsghdr))
 740 #define CMSG_VALID(cmsg, start, end)                                    \
 741         (ISALIGNED_cmsghdr(cmsg) &&                                     \
 742         ((uintptr_t)(cmsg) >= (uintptr_t)(start)) &&                    \
 743         ((uintptr_t)(cmsg) < (uintptr_t)(end)) &&                       \
 744         ((ssize_t)(cmsg)->cmsg_len >= sizeof (struct cmsghdr)) &&       \
 745         ((uintptr_t)(cmsg) + (cmsg)->cmsg_len <= (uintptr_t)(end)))
 746
 747 /*
 748  * Maximum size of any argument that is copied in (addresses, options,
 749  * access rights). MUST be at least MAXPATHLEN + 3.
 750  * BSD and SunOS 4.X limited this to MLEN or MCLBYTES.
 751  */
 752 #define SO_MAXARGSIZE   8192
 753
 754 /*
 755  * Convert between vnode and sonode
 756  */
 757 #define VTOSO(vp)       ((struct sonode *)((vp)->v_data))
 758 #define SOTOV(sp)       ((sp)->so_vnode)
 759
 760 /*
 761  * Internal flags for sobind()
 762  */
 763 #define _SOBIND_REBIND          0x01    /* Bind to existing local address */
 764 #define _SOBIND_UNSPEC          0x02    /* Bind to unspecified address */
 765 #define _SOBIND_LOCK_HELD       0x04    /* so_excl_lock held by caller */
 766 #define _SOBIND_NOXLATE         0x08    /* No addr translation for AF_UNIX */
 767 #define _SOBIND_LISTEN          0x40    /* Make into SS_ACCEPTCONN */
 768 #define _SOBIND_SOCKETPAIR      0x80    /* Internal flag for so_socketpair() */
 769                                         /* to enable listen with backlog = 1 */
 770
 771 /*
 772  * Internal flags for sounbind()
 773  */
 774 #define _SOUNBIND_REBIND        0x01    /* Don't clear fields - will rebind */
 775
 776 /*
 777  * Internal flags for soconnect()
 778  */
 779 #define _SOCONNECT_NOXLATE      0x01    /* No addr translation for AF_UNIX */
 780 #define _SOCONNECT_DID_BIND     0x02    /* Unbind when connect fails */
 781
 782 /*
 783  * Internal flags for sodisconnect()
 784  */
 785 #define _SODISCONNECT_LOCK_HELD 0x01    /* so_excl_lock held by caller */
 786
 787 /*
 788  * Internal flags for soallocproto*()
 789  */
 790 #define _ALLOC_NOSLEEP          0       /* Don't sleep for memory */
 791 #define _ALLOC_INTR             1       /* Sleep until interrupt */
 792 #define _ALLOC_SLEEP            2       /* Sleep forever */
 793
 794 /*
 795  * Internal structure for handling AF_UNIX file descriptor passing
 796  */
 797 struct fdbuf {
 798         int             fd_size;        /* In bytes, for kmem_free */
 799         int             fd_numfd;       /* Number of elements below */
 800         char            *fd_ebuf;       /* Extra buffer to free  */
 801         int             fd_ebuflen;
 802         frtn_t          fd_frtn;
 803         struct file     *fd_fds[1];     /* One or more */
 804 };
 805 #define FDBUF_HDRSIZE   (sizeof (struct fdbuf) - sizeof (struct file *))
 806
 807 /*
 808  * Variable that can be patched to set what version of socket socket()
 809  * will create.
 810  */
 811 extern int so_default_version;
 812
 813 #ifdef DEBUG
 814 /* Turn on extra testing capabilities */
 815 #define SOCK_TEST
 816 #endif /* DEBUG */
 817
 818 #ifdef DEBUG
 819 char    *pr_state(uint_t, uint_t);
 820 char    *pr_addr(int, struct sockaddr *, t_uscalar_t);
 821 int     so_verify_oobstate(struct sonode *);
 822 #endif /* DEBUG */
 823
 824 /*
 825  * DEBUG macros
 826  */
 827 #if defined(DEBUG)
 828 #define SOCK_DEBUG
 829
 830 extern int sockdebug;
 831 extern int sockprinterr;
 832
 833 #define eprint(args)    printf args
 834 #define eprintso(so, args) \
 835 { if (sockprinterr && ((so)->so_options & SO_DEBUG)) printf args; }
 836 #define eprintline(error)                                       \
 837 {                                                               \
 838         if (error != EINTR && (sockprinterr || sockdebug > 0))  \
 839                 printf("socket error %d: line %d file %s\n",    \
 840                         (error), __LINE__, __FILE__);           \
 841 }
 842
 843 #define eprintsoline(so, error)                                 \
 844 { if (sockprinterr && ((so)->so_options & SO_DEBUG))            \
 845         printf("socket(%p) error %d: line %d file %s\n",        \
 846                 (void *)(so), (error), __LINE__, __FILE__);     \
 847 }
 848 #define dprint(level, args)     { if (sockdebug > (level)) printf args; }
 849 #define dprintso(so, level, args) \
 850 { if (sockdebug > (level) && ((so)->so_options & SO_DEBUG)) printf args; }
 851
 852 #else /* define(DEBUG) */
 853
 854 #define eprint(args)            {}
 855 #define eprintso(so, args)      {}
 856 #define eprintline(error)       {}
 857 #define eprintsoline(so, error) {}
 858 #define dprint(level, args)     {}
 859 #define dprintso(so, level, args) {}
 860
 861 #endif /* defined(DEBUG) */
 862
 863 extern const struct vnodeops            socket_vnodeops;
 864
 865 extern dev_t                            sockdev;
 866
 867 extern krwlock_t                        sockconf_lock;
 868
 869 /*
 870  * sockfs functions
 871  */
 872 extern int      sock_getmsg(vnode_t *, struct strbuf *, struct strbuf *,
 873                         uchar_t *, int *, int, rval_t *);
 874 extern int      sock_putmsg(vnode_t *, struct strbuf *, struct strbuf *,
 875                         uchar_t, int, int);
 876 extern int      sogetvp(char *, vnode_t **, int);
 877 extern int      sockinit(int, char *);
 878 extern int      solookup(int, int, int, struct sockparams **);
 879 extern void     so_lock_single(struct sonode *);
 880 extern void     so_unlock_single(struct sonode *, int);
 881 extern int      so_lock_read(struct sonode *, int);
 882 extern int      so_lock_read_intr(struct sonode *, int);
 883 extern void     so_unlock_read(struct sonode *);
 884 extern void     *sogetoff(mblk_t *, t_uscalar_t, t_uscalar_t, uint_t);
 885 extern void     so_getopt_srcaddr(void *, t_uscalar_t,
 886                         void **, t_uscalar_t *);
 887 extern int      so_getopt_unix_close(void *, t_uscalar_t);
 888 extern void     fdbuf_free(struct fdbuf *);
 889 extern mblk_t   *fdbuf_allocmsg(int, struct fdbuf *);
 890 extern int      fdbuf_create(void *, int, struct fdbuf **);
 891 extern void     so_closefds(void *, t_uscalar_t, int);
 892 extern int      so_getfdopt(void *, t_uscalar_t, void **, int *);
 893 t_uscalar_t     so_optlen(void *, t_uscalar_t);
 894 extern void     so_cmsg2opt(void *, t_uscalar_t, mblk_t *);
 895 extern t_uscalar_t
 896                 so_cmsglen(mblk_t *, void *, t_uscalar_t);
 897 extern int      so_opt2cmsg(mblk_t *, void *, t_uscalar_t,
 898                         void *, t_uscalar_t);
 899 extern void     soisconnecting(struct sonode *);
 900 extern void     soisconnected(struct sonode *);
 901 extern void     soisdisconnected(struct sonode *, int);
 902 extern void     socantsendmore(struct sonode *);
 903 extern void     socantrcvmore(struct sonode *);
 904 extern void     soseterror(struct sonode *, int);
 905 extern int      sogeterr(struct sonode *, boolean_t);
 906 extern int      sowaitconnected(struct sonode *, int, int);
 907
 908 extern ssize_t  soreadfile(file_t *, uchar_t *, uoff_t, int *, size_t);
 909 extern void     *sock_kstat_init(zoneid_t);
 910 extern void     sock_kstat_fini(zoneid_t, void *);
 911 extern struct sonode *getsonode(int, int *, file_t **);
 912 /*
 913  * Function wrappers (mostly around the sonode switch) for
 914  * backward compatibility.
 915  */
 916 extern int      soaccept(struct sonode *, int, struct sonode **);
 917 extern int      sobind(struct sonode *, struct sockaddr *, socklen_t,
 918                     int, int);
 919 extern int      solisten(struct sonode *, int);
 920 extern int      soconnect(struct sonode *, struct sockaddr *, socklen_t,
 921                     int, int);
 922 extern int      sorecvmsg(struct sonode *, struct msghdr *, struct uio *);
 923 extern int      sosendmsg(struct sonode *, struct msghdr *, struct uio *);
 924 extern int      soshutdown(struct sonode *, int);
 925 extern int      sogetsockopt(struct sonode *, int, int, void *, socklen_t *,
 926                     int);
 927 extern int      sosetsockopt(struct sonode *, int, int, const void *,
 928                     t_uscalar_t);
 929
 930 extern struct sonode    *socreate(struct sockparams *, int, int, int, int *);
 931
 932 extern int      so_copyin(const void *, void *, size_t, int);
 933 extern int      so_copyout(const void *, void *, size_t, int);
 934
 935 #endif
 936
 937 /*
 938  * Internal structure for obtaining sonode information from the socklist.
 939  * These types match those corresponding in the sonode structure.
 940  * This is not a published interface, and may change at any time.
 941  */
 942 struct sockinfo {
 943         uint_t          si_size;                /* real length of this struct */
 944         short           si_family;
 945         short           si_type;
 946         ushort_t        si_flag;
 947         uint_t          si_state;
 948         uint_t          si_ux_laddr_sou_magic;
 949         uint_t          si_ux_faddr_sou_magic;
 950         t_scalar_t      si_serv_type;
 951         t_uscalar_t     si_laddr_soa_len;
 952         t_uscalar_t     si_faddr_soa_len;
 953         uint16_t        si_laddr_family;
 954         uint16_t        si_faddr_family;
 955         char            si_laddr_sun_path[MAXPATHLEN + 1]; /* NULL terminated */
 956         char            si_faddr_sun_path[MAXPATHLEN + 1];
 957         boolean_t       si_faddr_noxlate;
 958         zoneid_t        si_szoneid;
 959 };
 960
 961 /*
 962  * Subcodes for sockconf() system call
 963  */
 964 #define SOCKCONFIG_ADD_SOCK             0
 965 #define SOCKCONFIG_REMOVE_SOCK          1
 966 #define SOCKCONFIG_ADD_FILTER           2
 967 #define SOCKCONFIG_REMOVE_FILTER        3
 968 #define SOCKCONFIG_GET_SOCKTABLE        4
 969
 970 /*
 971  * Data structures for configuring socket filters.
 972  */
 973
 974 /*
 975  * Placement hint for automatic filters
 976  */
 977 typedef enum {
 978         SOF_HINT_NONE,
 979         SOF_HINT_TOP,
 980         SOF_HINT_BOTTOM,
 981         SOF_HINT_BEFORE,
 982         SOF_HINT_AFTER
 983 } sof_hint_t;
 984
 985 /*
 986  * Socket tuple. Used by sockconfig_filter_props to list socket
 987  * types of interest.
 988  */
 989 typedef struct sof_socktuple {
 990         int     sofst_family;
 991         int     sofst_type;
 992         int     sofst_protocol;
 993 } sof_socktuple_t;
 994
 995 /*
 996  * Socket filter properties used by sockconfig() system call.
 997  */
 998 struct sockconfig_filter_props {
 999         char            *sfp_modname;
1000         boolean_t       sfp_autoattach;
1001         sof_hint_t      sfp_hint;
1002         char            *sfp_hintarg;
1003         uint_t          sfp_socktuple_cnt;
1004         sof_socktuple_t *sfp_socktuple;
1005 };
1006
1007 /*
1008  * Data structures for the in-kernel socket configuration table.
1009  */
1010 typedef struct sockconfig_socktable_entry {
1011         int             se_family;
1012         int             se_type;
1013         int             se_protocol;
1014         int             se_refcnt;
1015         int             se_flags;
1016         char            se_modname[MODMAXNAMELEN];
1017         char            se_strdev[MAXPATHLEN];
1018 } sockconfig_socktable_entry_t;
1019
1020 typedef struct sockconfig_socktable {
1021         uint_t          num_of_entries;
1022         sockconfig_socktable_entry_t *st_entries;
1023 } sockconfig_socktable_t;
1024
1025 #ifdef  _SYSCALL32
1026
1027 typedef struct sof_socktuple32 {
1028         int32_t sofst_family;
1029         int32_t sofst_type;
1030         int32_t sofst_protocol;
1031 } sof_socktuple32_t;
1032
1033 struct sockconfig_filter_props32 {
1034         caddr32_t       sfp_modname;
1035         boolean_t       sfp_autoattach;
1036         sof_hint_t      sfp_hint;
1037         caddr32_t       sfp_hintarg;
1038         uint32_t        sfp_socktuple_cnt;
1039         caddr32_t       sfp_socktuple;
1040 };
1041
1042 typedef struct sockconfig_socktable32 {
1043         uint_t          num_of_entries;
1044         caddr32_t       st_entries;
1045 } sockconfig_socktable32_t;
1046
1047 #endif  /* _SYSCALL32 */
1048
1049 #define SOCKMOD_PATH    "socketmod"     /* dir where sockmods are stored */
1050
1051 #ifdef  __cplusplus
1052 }
1053 #endif
1054
1055 #endif  /* _SYS_SOCKETVAR_H */