kernel/net/tcp/tcp_tunables.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright 2016 Joyent, Inc.
  24  * Copyright 2013 Nexenta Systems, Inc.  All rights reserved.
  25  * Copyright (c) 2013 by Delphix. All rights reserved.
  26  */
  27 /* Copyright (c) 1990 Mentat Inc. */
  28
  29 #include <inet/ip.h>
  30 #include <inet/tcp_impl.h>
  31 #include <sys/multidata.h>
  32 #include <sys/sunddi.h>
  33
  34 /* Max size IP datagram is 64k - 1 */
  35 #define TCP_MSS_MAX_IPV4 (IP_MAXPACKET - (sizeof (ipha_t) + sizeof (tcpha_t)))
  36 #define TCP_MSS_MAX_IPV6 (IP_MAXPACKET - (sizeof (ip6_t) + sizeof (tcpha_t)))
  37
  38 /* Max of the above */
  39 #define TCP_MSS_MAX             TCP_MSS_MAX_IPV4
  40
  41 /*
  42  * returns the current list of listener limit configuration.
  43  */
  44 /* ARGSUSED */
  45 static int
  46 tcp_listener_conf_get(netstack_t *stack, mod_prop_info_t *pinfo,
  47     const char *ifname, void *val, uint_t psize, uint_t flags)
  48 {
  49         tcp_stack_t     *tcps = stack->netstack_tcp;
  50         tcp_listener_t  *tl;
  51         char            *pval = val;
  52         size_t          nbytes = 0, tbytes = 0;
  53         uint_t          size;
  54         int             err = 0;
  55
  56         bzero(pval, psize);
  57         size = psize;
  58
  59         if (flags & (MOD_PROP_DEFAULT|MOD_PROP_PERM|MOD_PROP_POSSIBLE))
  60                 return (0);
  61
  62         mutex_enter(&tcps->tcps_listener_conf_lock);
  63         for (tl = list_head(&tcps->tcps_listener_conf); tl != NULL;
  64             tl = list_next(&tcps->tcps_listener_conf, tl)) {
  65                 if (psize == size)
  66                         nbytes = snprintf(pval, size, "%d:%d",  tl->tl_port,
  67                             tl->tl_ratio);
  68                 else
  69                         nbytes = snprintf(pval, size, ",%d:%d",  tl->tl_port,
  70                             tl->tl_ratio);
  71                 size -= nbytes;
  72                 pval += nbytes;
  73                 tbytes += nbytes;
  74                 if (tbytes >= psize) {
  75                         /* Buffer overflow, stop copying information */
  76                         err = ENOBUFS;
  77                         break;
  78                 }
  79         }
  80
  81         mutex_exit(&tcps->tcps_listener_conf_lock);
  82         return (err);
  83 }
  84
  85 /*
  86  * add a new listener limit configuration.
  87  */
  88 /* ARGSUSED */
  89 static int
  90 tcp_listener_conf_add(netstack_t *stack, cred_t *cr, mod_prop_info_t *pinfo,
  91     const char *ifname, const void* pval, uint_t flags)
  92 {
  93         tcp_listener_t  *new_tl;
  94         tcp_listener_t  *tl;
  95         long            lport;
  96         long            ratio;
  97         char            *colon;
  98         tcp_stack_t     *tcps = stack->netstack_tcp;
  99
 100         if (flags & MOD_PROP_DEFAULT)
 101                 return (ENOTSUP);
 102
 103         if (ddi_strtol(pval, &colon, 10, &lport) != 0 || lport <= 0 ||
 104             lport > USHRT_MAX || *colon != ':') {
 105                 return (EINVAL);
 106         }
 107         if (ddi_strtol(colon + 1, NULL, 10, &ratio) != 0 || ratio <= 0)
 108                 return (EINVAL);
 109
 110         mutex_enter(&tcps->tcps_listener_conf_lock);
 111         for (tl = list_head(&tcps->tcps_listener_conf); tl != NULL;
 112             tl = list_next(&tcps->tcps_listener_conf, tl)) {
 113                 /* There is an existing entry, so update its ratio value. */
 114                 if (tl->tl_port == lport) {
 115                         tl->tl_ratio = ratio;
 116                         mutex_exit(&tcps->tcps_listener_conf_lock);
 117                         return (0);
 118                 }
 119         }
 120
 121         if ((new_tl = kmem_alloc(sizeof (tcp_listener_t), KM_NOSLEEP)) ==
 122             NULL) {
 123                 mutex_exit(&tcps->tcps_listener_conf_lock);
 124                 return (ENOMEM);
 125         }
 126
 127         new_tl->tl_port = lport;
 128         new_tl->tl_ratio = ratio;
 129         list_insert_tail(&tcps->tcps_listener_conf, new_tl);
 130         mutex_exit(&tcps->tcps_listener_conf_lock);
 131         return (0);
 132 }
 133
 134 /*
 135  * remove a listener limit configuration.
 136  */
 137 /* ARGSUSED */
 138 static int
 139 tcp_listener_conf_del(netstack_t *stack, cred_t *cr, mod_prop_info_t *pinfo,
 140     const char *ifname, const void* pval, uint_t flags)
 141 {
 142         tcp_listener_t  *tl;
 143         long            lport;
 144         tcp_stack_t     *tcps = stack->netstack_tcp;
 145
 146         if (flags & MOD_PROP_DEFAULT)
 147                 return (ENOTSUP);
 148
 149         if (ddi_strtol(pval, NULL, 10, &lport) != 0 || lport <= 0 ||
 150             lport > USHRT_MAX) {
 151                 return (EINVAL);
 152         }
 153         mutex_enter(&tcps->tcps_listener_conf_lock);
 154         for (tl = list_head(&tcps->tcps_listener_conf); tl != NULL;
 155             tl = list_next(&tcps->tcps_listener_conf, tl)) {
 156                 if (tl->tl_port == lport) {
 157                         list_remove(&tcps->tcps_listener_conf, tl);
 158                         mutex_exit(&tcps->tcps_listener_conf_lock);
 159                         kmem_free(tl, sizeof (tcp_listener_t));
 160                         return (0);
 161                 }
 162         }
 163         mutex_exit(&tcps->tcps_listener_conf_lock);
 164         return (ESRCH);
 165 }
 166
 167 static int
 168 tcp_set_buf_prop(netstack_t *stack, cred_t *cr, mod_prop_info_t *pinfo,
 169     const char *ifname, const void *pval, uint_t flags)
 170 {
 171         return (mod_set_buf_prop(stack->netstack_tcp->tcps_propinfo_tbl, stack,
 172             cr, pinfo, ifname, pval, flags));
 173 }
 174
 175 static int
 176 tcp_get_buf_prop(netstack_t *stack, mod_prop_info_t *pinfo, const char *ifname,
 177     void *val, uint_t psize, uint_t flags)
 178 {
 179         return (mod_get_buf_prop(stack->netstack_tcp->tcps_propinfo_tbl, stack,
 180             pinfo, ifname, val, psize, flags));
 181 }
 182
 183 /*
 184  * Special checkers for smallest/largest anonymous port so they don't
 185  * ever happen to be (largest < smallest).
 186  */
 187 /* ARGSUSED */
 188 static int
 189 tcp_smallest_anon_set(netstack_t *stack, cred_t *cr, mod_prop_info_t *pinfo,
 190     const char *ifname, const void *pval, uint_t flags)
 191 {
 192         unsigned long new_value;
 193         tcp_stack_t *tcps = stack->netstack_tcp;
 194         int err;
 195
 196         if ((err = mod_uint32_value(pval, pinfo, flags, &new_value)) != 0)
 197                 return (err);
 198         /* mod_uint32_value() + pinfo guarantees we're in TCP port range. */
 199         if ((uint32_t)new_value > tcps->tcps_largest_anon_port)
 200                 return (ERANGE);
 201         pinfo->prop_cur_uval = (uint32_t)new_value;
 202         return (0);
 203 }
 204
 205 /* ARGSUSED */
 206 static int
 207 tcp_largest_anon_set(netstack_t *stack, cred_t *cr, mod_prop_info_t *pinfo,
 208     const char *ifname, const void *pval, uint_t flags)
 209 {
 210         unsigned long new_value;
 211         tcp_stack_t *tcps = stack->netstack_tcp;
 212         int err;
 213
 214         if ((err = mod_uint32_value(pval, pinfo, flags, &new_value)) != 0)
 215                 return (err);
 216         /* mod_uint32_value() + pinfo guarantees we're in TCP port range. */
 217         if ((uint32_t)new_value < tcps->tcps_smallest_anon_port)
 218                 return (ERANGE);
 219         pinfo->prop_cur_uval = (uint32_t)new_value;
 220         return (0);
 221 }
 222
 223 /*
 224  * All of these are alterable, within the min/max values given, at run time.
 225  *
 226  * Note: All those tunables which do not start with "_" are Committed and
 227  * therefore are public. See PSARC 2010/080.
 228  */
 229 mod_prop_info_t tcp_propinfo_tbl[] = {
 230         /* tunable - 0 */
 231         { "_time_wait_interval", MOD_PROTO_TCP,
 232             mod_set_uint32, mod_get_uint32,
 233             {1*SECONDS, TCP_TIME_WAIT_MAX, 1*MINUTES}, {1*MINUTES} },
 234
 235         { "_conn_req_max_q", MOD_PROTO_TCP,
 236             mod_set_uint32, mod_get_uint32,
 237             {1, UINT32_MAX, 128}, {128} },
 238
 239         { "_conn_req_max_q0", MOD_PROTO_TCP,
 240             mod_set_uint32, mod_get_uint32,
 241             {0, UINT32_MAX, 1024}, {1024} },
 242
 243         { "_conn_req_min", MOD_PROTO_TCP,
 244             mod_set_uint32, mod_get_uint32,
 245             {1, 1024, 1}, {1} },
 246
 247         { "_conn_grace_period", MOD_PROTO_TCP,
 248             mod_set_uint32, mod_get_uint32,
 249             {0*MS, 20*SECONDS, 0*MS}, {0*MS} },
 250
 251         { "_cwnd_max", MOD_PROTO_TCP,
 252             mod_set_uint32, mod_get_uint32,
 253             {128, ULP_MAX_BUF, 1024*1024}, {1024*1024} },
 254
 255         { "_debug", MOD_PROTO_TCP,
 256             mod_set_uint32, mod_get_uint32,
 257             {0, 10, 0}, {0} },
 258
 259         { "smallest_nonpriv_port", MOD_PROTO_TCP,
 260             mod_set_uint32, mod_get_uint32,
 261             {1024, (32*1024), 1024}, {1024} },
 262
 263         { "_ip_abort_cinterval", MOD_PROTO_TCP,
 264             mod_set_uint32, mod_get_uint32,
 265             {1*SECONDS, UINT32_MAX, 3*MINUTES}, {3*MINUTES} },
 266
 267         { "_ip_abort_linterval", MOD_PROTO_TCP,
 268             mod_set_uint32, mod_get_uint32,
 269             {1*SECONDS, UINT32_MAX, 3*MINUTES}, {3*MINUTES} },
 270
 271         /* tunable - 10 */
 272         { "_ip_abort_interval", MOD_PROTO_TCP,
 273             mod_set_uint32, mod_get_uint32,
 274             {500*MS, UINT32_MAX, 5*MINUTES}, {5*MINUTES} },
 275
 276         { "_ip_notify_cinterval", MOD_PROTO_TCP,
 277             mod_set_uint32, mod_get_uint32,
 278             {1*SECONDS, UINT32_MAX, 10*SECONDS},
 279             {10*SECONDS} },
 280
 281         { "_ip_notify_interval", MOD_PROTO_TCP,
 282             mod_set_uint32, mod_get_uint32,
 283             {500*MS, UINT32_MAX, 10*SECONDS}, {10*SECONDS} },
 284
 285         { "_ipv4_ttl", MOD_PROTO_TCP,
 286             mod_set_uint32, mod_get_uint32,
 287             {1, 255, 64}, {64} },
 288
 289         { "_keepalive_interval", MOD_PROTO_TCP,
 290             mod_set_uint32, mod_get_uint32,
 291             {1*SECONDS, 10*DAYS, 2*HOURS}, {2*HOURS} },
 292
 293         { "_maxpsz_multiplier", MOD_PROTO_TCP,
 294             mod_set_uint32, mod_get_uint32,
 295             {0, 100, 10}, {10} },
 296
 297         { "_mss_def_ipv4", MOD_PROTO_TCP,
 298             mod_set_uint32, mod_get_uint32,
 299             {1, TCP_MSS_MAX_IPV4, 536}, {536} },
 300
 301         { "_mss_max_ipv4", MOD_PROTO_TCP,
 302             mod_set_uint32, mod_get_uint32,
 303             {1, TCP_MSS_MAX_IPV4, TCP_MSS_MAX_IPV4},
 304             {TCP_MSS_MAX_IPV4} },
 305
 306         { "_mss_min", MOD_PROTO_TCP,
 307             mod_set_uint32, mod_get_uint32,
 308             {1, TCP_MSS_MAX, 108}, {108} },
 309
 310         { "_naglim_def", MOD_PROTO_TCP,
 311             mod_set_uint32, mod_get_uint32,
 312             {1, (64*1024)-1, (4*1024)-1}, {(4*1024)-1} },
 313
 314         /* tunable - 20 */
 315         { "_rexmit_interval_initial", MOD_PROTO_TCP,
 316             mod_set_uint32, mod_get_uint32,
 317             {1*MS, 20*SECONDS, 1*SECONDS}, {1*SECONDS} },
 318
 319         { "_rexmit_interval_max", MOD_PROTO_TCP,
 320             mod_set_uint32, mod_get_uint32,
 321             {1*MS, 2*HOURS, 60*SECONDS}, {60*SECONDS} },
 322
 323         { "_rexmit_interval_min", MOD_PROTO_TCP,
 324             mod_set_uint32, mod_get_uint32,
 325             {1*MS, 2*HOURS, 400*MS}, {400*MS} },
 326
 327         { "_deferred_ack_interval", MOD_PROTO_TCP,
 328             mod_set_uint32, mod_get_uint32,
 329             {1*MS, 1*MINUTES, 100*MS}, {100*MS} },
 330
 331         { "_snd_lowat_fraction", MOD_PROTO_TCP,
 332             mod_set_uint32, mod_get_uint32,
 333             {0, 16, 10}, {10} },
 334
 335         { "dupack_fast_retrans", MOD_PROTO_TCP,
 336             mod_set_uint32, mod_get_uint32,
 337             {1, 10000, 3}, {3} },
 338
 339         { "_ignore_path_mtu", MOD_PROTO_TCP,
 340             mod_set_boolean, mod_get_boolean,
 341             {B_FALSE}, {B_FALSE} },
 342
 343         { "smallest_anon_port", MOD_PROTO_TCP,
 344             tcp_smallest_anon_set, mod_get_uint32,
 345             {1024, ULP_MAX_PORT, 32*1024}, {32*1024} },
 346
 347         { "largest_anon_port", MOD_PROTO_TCP,
 348             tcp_largest_anon_set, mod_get_uint32,
 349             {1024, ULP_MAX_PORT, ULP_MAX_PORT},
 350             {ULP_MAX_PORT} },
 351
 352         { "send_buf", MOD_PROTO_TCP,
 353             tcp_set_buf_prop, tcp_get_buf_prop,
 354             {TCP_XMIT_LOWATER, ULP_MAX_BUF, TCP_XMIT_HIWATER},
 355             {TCP_XMIT_HIWATER} },
 356
 357         /* tunable - 30 */
 358         { "_xmit_lowat", MOD_PROTO_TCP,
 359             mod_set_uint32, mod_get_uint32,
 360             {TCP_XMIT_LOWATER, ULP_MAX_BUF, TCP_XMIT_LOWATER},
 361             {TCP_XMIT_LOWATER} },
 362
 363         { "recv_buf", MOD_PROTO_TCP,
 364             tcp_set_buf_prop, tcp_get_buf_prop,
 365             {TCP_RECV_LOWATER, ULP_MAX_BUF, TCP_RECV_HIWATER},
 366             {TCP_RECV_HIWATER} },
 367
 368         { "_recv_hiwat_minmss", MOD_PROTO_TCP,
 369             mod_set_uint32, mod_get_uint32,
 370             {1, 65536, 4}, {4} },
 371
 372         { "_fin_wait_2_flush_interval", MOD_PROTO_TCP,
 373             mod_set_uint32, mod_get_uint32,
 374             {1*SECONDS, 2*HOURS, 60*SECONDS},
 375             {60*SECONDS} },
 376
 377         { "max_buf", MOD_PROTO_TCP,
 378             mod_set_uint32, mod_get_uint32,
 379             {8192, ULP_MAX_BUF, 1024*1024}, {1024*1024} },
 380
 381         { "_rtt_updates", MOD_PROTO_TCP,
 382             mod_set_uint32, mod_get_uint32,
 383             {0, 65536, 20}, {20} },
 384
 385         { "_wscale_always", MOD_PROTO_TCP,
 386             mod_set_boolean, mod_get_boolean,
 387             {B_TRUE}, {B_TRUE} },
 388
 389         { "_tstamp_always", MOD_PROTO_TCP,
 390             mod_set_boolean, mod_get_boolean,
 391             {B_FALSE}, {B_FALSE} },
 392
 393         { "_tstamp_if_wscale", MOD_PROTO_TCP,
 394             mod_set_boolean, mod_get_boolean,
 395             {B_TRUE}, {B_TRUE} },
 396
 397         /* tunable - 40 */
 398         { "_rexmit_interval_extra", MOD_PROTO_TCP,
 399             mod_set_uint32, mod_get_uint32,
 400             {0*MS, 2*HOURS, 0*MS}, {0*MS} },
 401
 402         { "_deferred_acks_max", MOD_PROTO_TCP,
 403             mod_set_uint32, mod_get_uint32,
 404             {0, 16, 2}, {2} },
 405
 406         { "_slow_start_after_idle", MOD_PROTO_TCP,
 407             mod_set_uint32, mod_get_uint32,
 408             {0, 16384, 0}, {0} },
 409
 410         { "_slow_start_initial", MOD_PROTO_TCP,
 411             mod_set_uint32, mod_get_uint32,
 412             {0, 16, 0}, {0} },
 413
 414         { "sack", MOD_PROTO_TCP,
 415             mod_set_uint32, mod_get_uint32,
 416             {0, 2, 2}, {2} },
 417
 418         { "_ipv6_hoplimit", MOD_PROTO_TCP,
 419             mod_set_uint32, mod_get_uint32,
 420             {0, IPV6_MAX_HOPS, IPV6_DEFAULT_HOPS},
 421             {IPV6_DEFAULT_HOPS} },
 422
 423         { "_mss_def_ipv6", MOD_PROTO_TCP,
 424             mod_set_uint32, mod_get_uint32,
 425             {1, TCP_MSS_MAX_IPV6, 1220}, {1220} },
 426
 427         { "_mss_max_ipv6", MOD_PROTO_TCP,
 428             mod_set_uint32, mod_get_uint32,
 429             {1, TCP_MSS_MAX_IPV6, TCP_MSS_MAX_IPV6},
 430             {TCP_MSS_MAX_IPV6} },
 431
 432         { "_rev_src_routes", MOD_PROTO_TCP,
 433             mod_set_boolean, mod_get_boolean,
 434             {B_FALSE}, {B_FALSE} },
 435
 436         { "_local_dack_interval", MOD_PROTO_TCP,
 437             mod_set_uint32, mod_get_uint32,
 438             {10*MS, 500*MS, 50*MS}, {50*MS} },
 439
 440         /* tunable - 50 */
 441         { "_local_dacks_max", MOD_PROTO_TCP,
 442             mod_set_uint32, mod_get_uint32,
 443             {0, 16, 8}, {8} },
 444
 445         { "ecn", MOD_PROTO_TCP,
 446             mod_set_uint32, mod_get_uint32,
 447             {0, 2, 1}, {1} },
 448
 449         { "_rst_sent_rate_enabled", MOD_PROTO_TCP,
 450             mod_set_boolean, mod_get_boolean,
 451             {B_TRUE}, {B_TRUE} },
 452
 453         { "_rst_sent_rate", MOD_PROTO_TCP,
 454             mod_set_uint32, mod_get_uint32,
 455             {0, UINT32_MAX, 40}, {40} },
 456
 457         { "_push_timer_interval", MOD_PROTO_TCP,
 458             mod_set_uint32, mod_get_uint32,
 459             {0, 100*MS, 50*MS}, {50*MS} },
 460
 461         { "_use_smss_as_mss_opt", MOD_PROTO_TCP,
 462             mod_set_boolean, mod_get_boolean,
 463             {B_FALSE}, {B_FALSE} },
 464
 465         { "_keepalive_abort_interval", MOD_PROTO_TCP,
 466             mod_set_uint32, mod_get_uint32,
 467             {0, UINT32_MAX, 8*MINUTES}, {8*MINUTES} },
 468
 469         /*
 470          * tcp_wroff_xtra is the extra space in front of TCP/IP header for link
 471          * layer header.  It has to be a multiple of 8.
 472          */
 473         { "_wroff_xtra", MOD_PROTO_TCP,
 474             mod_set_aligned, mod_get_uint32,
 475             {0, 256, 32}, {32} },
 476
 477         { "_dev_flow_ctl", MOD_PROTO_TCP,
 478             mod_set_boolean, mod_get_boolean,
 479             {B_FALSE}, {B_FALSE} },
 480
 481         { "_reass_timeout", MOD_PROTO_TCP,
 482             mod_set_uint32, mod_get_uint32,
 483             {0, UINT32_MAX, 100*SECONDS}, {100*SECONDS} },
 484
 485         /* tunable - 60 */
 486         { "extra_priv_ports", MOD_PROTO_TCP,
 487             mod_set_extra_privports, mod_get_extra_privports,
 488             {1, ULP_MAX_PORT, 0}, {0} },
 489
 490         { "_listener_limit_conf", MOD_PROTO_TCP,
 491             NULL, tcp_listener_conf_get, {0}, {0} },
 492
 493         { "_listener_limit_conf_add", MOD_PROTO_TCP,
 494             tcp_listener_conf_add, NULL, {0}, {0} },
 495
 496         { "_listener_limit_conf_del", MOD_PROTO_TCP,
 497             tcp_listener_conf_del, NULL, {0}, {0} },
 498
 499         { "_iss_incr", MOD_PROTO_TCP,
 500             mod_set_uint32, mod_get_uint32,
 501             {1, ISS_INCR, ISS_INCR},
 502             {ISS_INCR} },
 503
 504         { "?", MOD_PROTO_TCP, NULL, mod_get_allprop, {0}, {0} },
 505
 506         { NULL, 0, NULL, NULL, {0}, {0} }
 507 };
 508
 509 int tcp_propinfo_count = A_CNT(tcp_propinfo_tbl);