1 /*- 2 * Copyright (c) 2016-2018 Netflix Inc. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 */ 26 #include <sys/cdefs.h> 27 __FBSDID("$FreeBSD$"); 28 29 #include "opt_inet.h" 30 #include "opt_inet6.h" 31 #include "opt_ipsec.h" 32 #include "opt_tcpdebug.h" 33 /** 34 * Some notes about usage. 35 * 36 * The tcp_hpts system is designed to provide a high precision timer 37 * system for tcp. Its main purpose is to provide a mechanism for 38 * pacing packets out onto the wire. It can be used in two ways 39 * by a given TCP stack (and those two methods can be used simultaneously). 40 * 41 * First, and probably the main thing its used by Rack and BBR for, it can 42 * be used to call tcp_output() of a transport stack at some time in the future. 43 * The normal way this is done is that tcp_output() of the stack schedules 44 * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The 45 * slot is the time from now that the stack wants to be called but it 46 * must be converted to tcp_hpts's notion of slot. This is done with 47 * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical 48 * call from the tcp_output() routine might look like: 49 * 50 * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550)); 51 * 52 * The above would schedule tcp_ouput() to be called in 550 useconds. 53 * Note that if using this mechanism the stack will want to add near 54 * its top a check to prevent unwanted calls (from user land or the 55 * arrival of incoming ack's). So it would add something like: 56 * 57 * if (inp->inp_in_hpts) 58 * return; 59 * 60 * to prevent output processing until the time alotted has gone by. 61 * Of course this is a bare bones example and the stack will probably 62 * have more consideration then just the above. 63 * 64 * Now the tcp_hpts system will call tcp_output in one of two forms, 65 * it will first check to see if the stack as defined a 66 * tfb_tcp_output_wtime() function, if so that is the routine it 67 * will call, if that function is not defined then it will call the 68 * tfb_tcp_output() function. The only difference between these 69 * two calls is that the former passes the time in to the function 70 * so the function does not have to access the time (which tcp_hpts 71 * already has). What these functions do is of course totally up 72 * to the individual tcp stack. 73 * 74 * Now the second function (actually two functions I guess :D) 75 * the tcp_hpts system provides is the ability to either abort 76 * a connection (later) or process input on a connection. 77 * Why would you want to do this? To keep processor locality. 78 * 79 * So in order to use the input redirection function the 80 * stack changes its tcp_do_segment() routine to instead 81 * of process the data call the function: 82 * 83 * tcp_queue_pkt_to_input() 84 * 85 * You will note that the arguments to this function look 86 * a lot like tcp_do_segments's arguments. This function 87 * will assure that the tcp_hpts system will 88 * call the functions tfb_tcp_hpts_do_segment() from the 89 * correct CPU. Note that multiple calls can get pushed 90 * into the tcp_hpts system this will be indicated by 91 * the next to last argument to tfb_tcp_hpts_do_segment() 92 * (nxt_pkt). If nxt_pkt is a 1 then another packet is 93 * coming. If nxt_pkt is a 0 then this is the last call 94 * that the tcp_hpts system has available for the tcp stack. 95 * 96 * The other point of the input system is to be able to safely 97 * drop a tcp connection without worrying about the recursive 98 * locking that may be occuring on the INP_WLOCK. So if 99 * a stack wants to drop a connection it calls: 100 * 101 * tcp_set_inp_to_drop(tp, ETIMEDOUT) 102 * 103 * To schedule the tcp_hpts system to call 104 * 105 * tcp_drop(tp, drop_reason) 106 * 107 * at a future point. This is quite handy to prevent locking 108 * issues when dropping connections. 109 * 110 */ 111 112 #include <sys/param.h> 113 #include <sys/bus.h> 114 #include <sys/interrupt.h> 115 #include <sys/module.h> 116 #include <sys/kernel.h> 117 #include <sys/hhook.h> 118 #include <sys/malloc.h> 119 #include <sys/mbuf.h> 120 #include <sys/proc.h> /* for proc0 declaration */ 121 #include <sys/socket.h> 122 #include <sys/socketvar.h> 123 #include <sys/sysctl.h> 124 #include <sys/systm.h> 125 #include <sys/refcount.h> 126 #include <sys/sched.h> 127 #include <sys/queue.h> 128 #include <sys/smp.h> 129 #include <sys/counter.h> 130 #include <sys/time.h> 131 #include <sys/kthread.h> 132 #include <sys/kern_prefetch.h> 133 134 #include <vm/uma.h> 135 136 #include <net/route.h> 137 #include <net/vnet.h> 138 139 #define TCPSTATES /* for logging */ 140 141 #include <netinet/in.h> 142 #include <netinet/in_kdtrace.h> 143 #include <netinet/in_pcb.h> 144 #include <netinet/ip.h> 145 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 146 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 147 #include <netinet/ip_var.h> 148 #include <netinet/ip6.h> 149 #include <netinet6/in6_pcb.h> 150 #include <netinet6/ip6_var.h> 151 #define TCPOUTFLAGS 152 #include <netinet/tcp.h> 153 #include <netinet/tcp_fsm.h> 154 #include <netinet/tcp_seq.h> 155 #include <netinet/tcp_timer.h> 156 #include <netinet/tcp_var.h> 157 #include <netinet/tcpip.h> 158 #include <netinet/cc/cc.h> 159 #include <netinet/tcp_hpts.h> 160 161 #ifdef tcpdebug 162 #include <netinet/tcp_debug.h> 163 #endif /* tcpdebug */ 164 #ifdef tcp_offload 165 #include <netinet/tcp_offload.h> 166 #endif 167 168 #ifdef ipsec 169 #include <netipsec/ipsec.h> 170 #include <netipsec/ipsec6.h> 171 #endif /* ipsec */ 172 #include "opt_rss.h" 173 174 MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts"); 175 #ifdef RSS 176 static int tcp_bind_threads = 1; 177 #else 178 static int tcp_bind_threads = 0; 179 #endif 180 TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); 181 182 static uint32_t tcp_hpts_logging_size = DEFAULT_HPTS_LOG; 183 184 TUNABLE_INT("net.inet.tcp.hpts_logging_sz", &tcp_hpts_logging_size); 185 186 static struct tcp_hptsi tcp_pace; 187 188 static int 189 tcp_hptsi_lock_inpinfo(struct inpcb *inp, 190 struct tcpcb **tp); 191 static void tcp_wakehpts(struct tcp_hpts_entry *p); 192 static void tcp_wakeinput(struct tcp_hpts_entry *p); 193 static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv); 194 static void tcp_hptsi(struct tcp_hpts_entry *hpts, struct timeval *ctick); 195 static void tcp_hpts_thread(void *ctx); 196 static void tcp_init_hptsi(void *st); 197 198 int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP; 199 static int32_t tcp_hpts_callout_skip_swi = 0; 200 201 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW, 0, "TCP Hpts controls"); 202 203 #define timersub(tvp, uvp, vvp) \ 204 do { \ 205 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ 206 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ 207 if ((vvp)->tv_usec < 0) { \ 208 (vvp)->tv_sec--; \ 209 (vvp)->tv_usec += 1000000; \ 210 } \ 211 } while (0) 212 213 static int32_t logging_on = 0; 214 static int32_t hpts_sleep_max = (NUM_OF_HPTSI_SLOTS - 2); 215 static int32_t tcp_hpts_precision = 120; 216 217 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, 218 &tcp_hpts_precision, 120, 219 "Value for PRE() precision of callout"); 220 221 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW, 222 &logging_on, 0, 223 "Turn on logging if compiled in"); 224 225 counter_u64_t hpts_loops; 226 227 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD, 228 &hpts_loops, "Number of times hpts had to loop to catch up"); 229 230 counter_u64_t back_tosleep; 231 232 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD, 233 &back_tosleep, "Number of times hpts found no tcbs"); 234 235 static int32_t in_newts_every_tcb = 0; 236 237 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, in_tsperpcb, CTLFLAG_RW, 238 &in_newts_every_tcb, 0, 239 "Do we have a new cts every tcb we process for input"); 240 static int32_t in_ts_percision = 0; 241 242 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, in_tspercision, CTLFLAG_RW, 243 &in_ts_percision, 0, 244 "Do we use percise timestamp for clients on input"); 245 static int32_t out_newts_every_tcb = 0; 246 247 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tsperpcb, CTLFLAG_RW, 248 &out_newts_every_tcb, 0, 249 "Do we have a new cts every tcb we process for output"); 250 static int32_t out_ts_percision = 0; 251 252 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW, 253 &out_ts_percision, 0, 254 "Do we use a percise timestamp for every output cts"); 255 256 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, maxsleep, CTLFLAG_RW, 257 &hpts_sleep_max, 0, 258 "The maximum time the hpts will sleep <1 - 254>"); 259 260 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW, 261 &tcp_min_hptsi_time, 0, 262 "The minimum time the hpts must sleep before processing more slots"); 263 264 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW, 265 &tcp_hpts_callout_skip_swi, 0, 266 "Do we have the callout call directly to the hpts?"); 267 268 static void 269 __tcp_hpts_log_it(struct tcp_hpts_entry *hpts, struct inpcb *inp, int event, uint32_t slot, 270 uint32_t ticknow, int32_t line) 271 { 272 struct hpts_log *pl; 273 274 HPTS_MTX_ASSERT(hpts); 275 if (hpts->p_log == NULL) 276 return; 277 pl = &hpts->p_log[hpts->p_log_at]; 278 hpts->p_log_at++; 279 if (hpts->p_log_at >= hpts->p_logsize) { 280 hpts->p_log_at = 0; 281 hpts->p_log_wrapped = 1; 282 } 283 pl->inp = inp; 284 if (inp) { 285 pl->t_paceslot = inp->inp_hptsslot; 286 pl->t_hptsreq = inp->inp_hpts_request; 287 pl->p_onhpts = inp->inp_in_hpts; 288 pl->p_oninput = inp->inp_in_input; 289 } else { 290 pl->t_paceslot = 0; 291 pl->t_hptsreq = 0; 292 pl->p_onhpts = 0; 293 pl->p_oninput = 0; 294 } 295 pl->is_notempty = 1; 296 pl->event = event; 297 pl->line = line; 298 pl->cts = tcp_get_usecs(NULL); 299 pl->p_curtick = hpts->p_curtick; 300 pl->p_prevtick = hpts->p_prevtick; 301 pl->p_on_queue_cnt = hpts->p_on_queue_cnt; 302 pl->ticknow = ticknow; 303 pl->slot_req = slot; 304 pl->p_nxt_slot = hpts->p_nxt_slot; 305 pl->p_cur_slot = hpts->p_cur_slot; 306 pl->p_hpts_sleep_time = hpts->p_hpts_sleep_time; 307 pl->p_flags = (hpts->p_cpu & 0x7f); 308 pl->p_flags <<= 7; 309 pl->p_flags |= (hpts->p_num & 0x7f); 310 pl->p_flags <<= 2; 311 if (hpts->p_hpts_active) { 312 pl->p_flags |= HPTS_HPTS_ACTIVE; 313 } 314 } 315 316 #define tcp_hpts_log_it(a, b, c, d, e) __tcp_hpts_log_it(a, b, c, d, e, __LINE__) 317 318 static void 319 hpts_timeout_swi(void *arg) 320 { 321 struct tcp_hpts_entry *hpts; 322 323 hpts = (struct tcp_hpts_entry *)arg; 324 swi_sched(hpts->ie_cookie, 0); 325 } 326 327 static void 328 hpts_timeout_dir(void *arg) 329 { 330 tcp_hpts_thread(arg); 331 } 332 333 static inline void 334 hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear) 335 { 336 #ifdef INVARIANTS 337 if (mtx_owned(&hpts->p_mtx) == 0) { 338 /* We don't own the mutex? */ 339 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); 340 } 341 if (hpts->p_cpu != inp->inp_hpts_cpu) { 342 /* It is not the right cpu/mutex? */ 343 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); 344 } 345 if (inp->inp_in_hpts == 0) { 346 /* We are not on the hpts? */ 347 panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp); 348 } 349 if (TAILQ_EMPTY(head) && 350 (hpts->p_on_queue_cnt != 0)) { 351 /* We should not be empty with a queue count */ 352 panic("%s hpts:%p hpts bucket empty but cnt:%d", 353 __FUNCTION__, hpts, hpts->p_on_queue_cnt); 354 } 355 #endif 356 TAILQ_REMOVE(head, inp, inp_hpts); 357 hpts->p_on_queue_cnt--; 358 if (hpts->p_on_queue_cnt < 0) { 359 /* Count should not go negative .. */ 360 #ifdef INVARIANTS 361 panic("Hpts goes negative inp:%p hpts:%p", 362 inp, hpts); 363 #endif 364 hpts->p_on_queue_cnt = 0; 365 } 366 if (clear) { 367 inp->inp_hpts_request = 0; 368 inp->inp_in_hpts = 0; 369 } 370 } 371 372 static inline void 373 hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref) 374 { 375 #ifdef INVARIANTS 376 if (mtx_owned(&hpts->p_mtx) == 0) { 377 /* We don't own the mutex? */ 378 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); 379 } 380 if (hpts->p_cpu != inp->inp_hpts_cpu) { 381 /* It is not the right cpu/mutex? */ 382 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); 383 } 384 if ((noref == 0) && (inp->inp_in_hpts == 1)) { 385 /* We are already on the hpts? */ 386 panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp); 387 } 388 #endif 389 TAILQ_INSERT_TAIL(head, inp, inp_hpts); 390 inp->inp_in_hpts = 1; 391 hpts->p_on_queue_cnt++; 392 if (noref == 0) { 393 in_pcbref(inp); 394 } 395 } 396 397 static inline void 398 hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear) 399 { 400 #ifdef INVARIANTS 401 if (mtx_owned(&hpts->p_mtx) == 0) { 402 /* We don't own the mutex? */ 403 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); 404 } 405 if (hpts->p_cpu != inp->inp_input_cpu) { 406 /* It is not the right cpu/mutex? */ 407 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); 408 } 409 if (inp->inp_in_input == 0) { 410 /* We are not on the input hpts? */ 411 panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp); 412 } 413 #endif 414 TAILQ_REMOVE(&hpts->p_input, inp, inp_input); 415 hpts->p_on_inqueue_cnt--; 416 if (hpts->p_on_inqueue_cnt < 0) { 417 #ifdef INVARIANTS 418 panic("Hpts in goes negative inp:%p hpts:%p", 419 inp, hpts); 420 #endif 421 hpts->p_on_inqueue_cnt = 0; 422 } 423 #ifdef INVARIANTS 424 if (TAILQ_EMPTY(&hpts->p_input) && 425 (hpts->p_on_inqueue_cnt != 0)) { 426 /* We should not be empty with a queue count */ 427 panic("%s hpts:%p in_hpts input empty but cnt:%d", 428 __FUNCTION__, hpts, hpts->p_on_inqueue_cnt); 429 } 430 #endif 431 if (clear) 432 inp->inp_in_input = 0; 433 } 434 435 static inline void 436 hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line) 437 { 438 #ifdef INVARIANTS 439 if (mtx_owned(&hpts->p_mtx) == 0) { 440 /* We don't own the mutex? */ 441 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); 442 } 443 if (hpts->p_cpu != inp->inp_input_cpu) { 444 /* It is not the right cpu/mutex? */ 445 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); 446 } 447 if (inp->inp_in_input == 1) { 448 /* We are already on the input hpts? */ 449 panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp); 450 } 451 #endif 452 TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input); 453 inp->inp_in_input = 1; 454 hpts->p_on_inqueue_cnt++; 455 in_pcbref(inp); 456 } 457 458 static int 459 sysctl_tcp_hpts_log(SYSCTL_HANDLER_ARGS) 460 { 461 struct tcp_hpts_entry *hpts; 462 size_t sz; 463 int32_t logging_was, i; 464 int32_t error = 0; 465 466 /* 467 * HACK: Turn off logging so no locks are required this really needs 468 * a memory barrier :) 469 */ 470 logging_was = logging_on; 471 logging_on = 0; 472 if (!req->oldptr) { 473 /* How much? */ 474 sz = 0; 475 for (i = 0; i < tcp_pace.rp_num_hptss; i++) { 476 hpts = tcp_pace.rp_ent[i]; 477 if (hpts->p_log == NULL) 478 continue; 479 sz += (sizeof(struct hpts_log) * hpts->p_logsize); 480 } 481 error = SYSCTL_OUT(req, 0, sz); 482 } else { 483 for (i = 0; i < tcp_pace.rp_num_hptss; i++) { 484 hpts = tcp_pace.rp_ent[i]; 485 if (hpts->p_log == NULL) 486 continue; 487 if (hpts->p_log_wrapped) 488 sz = (sizeof(struct hpts_log) * hpts->p_logsize); 489 else 490 sz = (sizeof(struct hpts_log) * hpts->p_log_at); 491 error = SYSCTL_OUT(req, hpts->p_log, sz); 492 } 493 } 494 logging_on = logging_was; 495 return error; 496 } 497 498 SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, log, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 499 0, 0, sysctl_tcp_hpts_log, "A", "tcp hptsi log"); 500 501 502 /* 503 * Try to get the INP_INFO lock. 504 * 505 * This function always succeeds in getting the lock. It will clear 506 * *tpp and return (1) if something critical changed while the inpcb 507 * was unlocked. Otherwise, it will leave *tpp unchanged and return (0). 508 * 509 * This function relies on the fact that the hpts always holds a 510 * reference on the inpcb while the segment is on the hptsi wheel and 511 * in the input queue. 512 * 513 */ 514 static int 515 tcp_hptsi_lock_inpinfo(struct inpcb *inp, struct tcpcb **tpp) 516 { 517 struct tcp_function_block *tfb; 518 struct tcpcb *tp; 519 void *ptr; 520 521 /* Try the easy way. */ 522 if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) 523 return (0); 524 525 /* 526 * OK, let's try the hard way. We'll save the function pointer block 527 * to make sure that doesn't change while we aren't holding the 528 * lock. 529 */ 530 tp = *tpp; 531 tfb = tp->t_fb; 532 ptr = tp->t_fb_ptr; 533 INP_WUNLOCK(inp); 534 INP_INFO_RLOCK(&V_tcbinfo); 535 INP_WLOCK(inp); 536 /* If the session went away, return an error. */ 537 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || 538 (inp->inp_flags2 & INP_FREED)) { 539 *tpp = NULL; 540 return (1); 541 } 542 /* 543 * If the function block or stack-specific data block changed, 544 * report an error. 545 */ 546 tp = intotcpcb(inp); 547 if ((tp->t_fb != tfb) && (tp->t_fb_ptr != ptr)) { 548 *tpp = NULL; 549 return (1); 550 } 551 return (0); 552 } 553 554 555 static void 556 tcp_wakehpts(struct tcp_hpts_entry *hpts) 557 { 558 HPTS_MTX_ASSERT(hpts); 559 swi_sched(hpts->ie_cookie, 0); 560 if (hpts->p_hpts_active == 2) { 561 /* Rare sleeping on a ENOBUF */ 562 wakeup_one(hpts); 563 } 564 } 565 566 static void 567 tcp_wakeinput(struct tcp_hpts_entry *hpts) 568 { 569 HPTS_MTX_ASSERT(hpts); 570 swi_sched(hpts->ie_cookie, 0); 571 if (hpts->p_hpts_active == 2) { 572 /* Rare sleeping on a ENOBUF */ 573 wakeup_one(hpts); 574 } 575 } 576 577 struct tcp_hpts_entry * 578 tcp_cur_hpts(struct inpcb *inp) 579 { 580 int32_t hpts_num; 581 struct tcp_hpts_entry *hpts; 582 583 hpts_num = inp->inp_hpts_cpu; 584 hpts = tcp_pace.rp_ent[hpts_num]; 585 return (hpts); 586 } 587 588 struct tcp_hpts_entry * 589 tcp_hpts_lock(struct inpcb *inp) 590 { 591 struct tcp_hpts_entry *hpts; 592 int32_t hpts_num; 593 594 again: 595 hpts_num = inp->inp_hpts_cpu; 596 hpts = tcp_pace.rp_ent[hpts_num]; 597 #ifdef INVARIANTS 598 if (mtx_owned(&hpts->p_mtx)) { 599 panic("Hpts:%p owns mtx prior-to lock line:%d", 600 hpts, __LINE__); 601 } 602 #endif 603 mtx_lock(&hpts->p_mtx); 604 if (hpts_num != inp->inp_hpts_cpu) { 605 mtx_unlock(&hpts->p_mtx); 606 goto again; 607 } 608 return (hpts); 609 } 610 611 struct tcp_hpts_entry * 612 tcp_input_lock(struct inpcb *inp) 613 { 614 struct tcp_hpts_entry *hpts; 615 int32_t hpts_num; 616 617 again: 618 hpts_num = inp->inp_input_cpu; 619 hpts = tcp_pace.rp_ent[hpts_num]; 620 #ifdef INVARIANTS 621 if (mtx_owned(&hpts->p_mtx)) { 622 panic("Hpts:%p owns mtx prior-to lock line:%d", 623 hpts, __LINE__); 624 } 625 #endif 626 mtx_lock(&hpts->p_mtx); 627 if (hpts_num != inp->inp_input_cpu) { 628 mtx_unlock(&hpts->p_mtx); 629 goto again; 630 } 631 return (hpts); 632 } 633 634 static void 635 tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line) 636 { 637 int32_t add_freed; 638 639 if (inp->inp_flags2 & INP_FREED) { 640 /* 641 * Need to play a special trick so that in_pcbrele_wlocked 642 * does not return 1 when it really should have returned 0. 643 */ 644 add_freed = 1; 645 inp->inp_flags2 &= ~INP_FREED; 646 } else { 647 add_freed = 0; 648 } 649 #ifndef INP_REF_DEBUG 650 if (in_pcbrele_wlocked(inp)) { 651 /* 652 * This should not happen. We have the inpcb referred to by 653 * the main socket (why we are called) and the hpts. It 654 * should always return 0. 655 */ 656 panic("inpcb:%p release ret 1", 657 inp); 658 } 659 #else 660 if (__in_pcbrele_wlocked(inp, line)) { 661 /* 662 * This should not happen. We have the inpcb referred to by 663 * the main socket (why we are called) and the hpts. It 664 * should always return 0. 665 */ 666 panic("inpcb:%p release ret 1", 667 inp); 668 } 669 #endif 670 if (add_freed) { 671 inp->inp_flags2 |= INP_FREED; 672 } 673 } 674 675 static void 676 tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) 677 { 678 if (inp->inp_in_hpts) { 679 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1); 680 tcp_remove_hpts_ref(inp, hpts, line); 681 } 682 } 683 684 static void 685 tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) 686 { 687 HPTS_MTX_ASSERT(hpts); 688 if (inp->inp_in_input) { 689 hpts_sane_input_remove(hpts, inp, 1); 690 tcp_remove_hpts_ref(inp, hpts, line); 691 } 692 } 693 694 /* 695 * Called normally with the INP_LOCKED but it 696 * does not matter, the hpts lock is the key 697 * but the lock order allows us to hold the 698 * INP lock and then get the hpts lock. 699 * 700 * Valid values in the flags are 701 * HPTS_REMOVE_OUTPUT - remove from the output of the hpts. 702 * HPTS_REMOVE_INPUT - remove from the input of the hpts. 703 * Note that you can or both values together and get two 704 * actions. 705 */ 706 void 707 __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line) 708 { 709 struct tcp_hpts_entry *hpts; 710 711 INP_WLOCK_ASSERT(inp); 712 if (flags & HPTS_REMOVE_OUTPUT) { 713 hpts = tcp_hpts_lock(inp); 714 tcp_hpts_remove_locked_output(hpts, inp, flags, line); 715 mtx_unlock(&hpts->p_mtx); 716 } 717 if (flags & HPTS_REMOVE_INPUT) { 718 hpts = tcp_input_lock(inp); 719 tcp_hpts_remove_locked_input(hpts, inp, flags, line); 720 mtx_unlock(&hpts->p_mtx); 721 } 722 } 723 724 static inline int 725 hpts_tick(struct tcp_hpts_entry *hpts, int32_t plus) 726 { 727 return ((hpts->p_prevtick + plus) % NUM_OF_HPTSI_SLOTS); 728 } 729 730 static int 731 tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref) 732 { 733 int32_t need_wake = 0; 734 uint32_t ticknow = 0; 735 736 HPTS_MTX_ASSERT(hpts); 737 if (inp->inp_in_hpts == 0) { 738 /* Ok we need to set it on the hpts in the current slot */ 739 if (hpts->p_hpts_active == 0) { 740 /* A sleeping hpts we want in next slot to run */ 741 if (logging_on) { 742 tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_SLEEPER, 0, 743 hpts_tick(hpts, 1)); 744 } 745 inp->inp_hptsslot = hpts_tick(hpts, 1); 746 inp->inp_hpts_request = 0; 747 if (logging_on) { 748 tcp_hpts_log_it(hpts, inp, HPTSLOG_SLEEP_BEFORE, 1, ticknow); 749 } 750 need_wake = 1; 751 } else if ((void *)inp == hpts->p_inp) { 752 /* 753 * We can't allow you to go into the same slot we 754 * are in. We must put you out. 755 */ 756 inp->inp_hptsslot = hpts->p_nxt_slot; 757 } else 758 inp->inp_hptsslot = hpts->p_cur_slot; 759 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref); 760 inp->inp_hpts_request = 0; 761 if (logging_on) { 762 tcp_hpts_log_it(hpts, inp, HPTSLOG_IMMEDIATE, 0, 0); 763 } 764 if (need_wake) { 765 /* 766 * Activate the hpts if it is sleeping and its 767 * timeout is not 1. 768 */ 769 if (logging_on) { 770 tcp_hpts_log_it(hpts, inp, HPTSLOG_WAKEUP_HPTS, 0, ticknow); 771 } 772 hpts->p_direct_wake = 1; 773 tcp_wakehpts(hpts); 774 } 775 } 776 return (need_wake); 777 } 778 779 int 780 __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line) 781 { 782 int32_t ret; 783 struct tcp_hpts_entry *hpts; 784 785 INP_WLOCK_ASSERT(inp); 786 hpts = tcp_hpts_lock(inp); 787 ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0); 788 mtx_unlock(&hpts->p_mtx); 789 return (ret); 790 } 791 792 static void 793 tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, uint32_t cts, int32_t line, 794 struct hpts_diag *diag, int32_t noref) 795 { 796 int32_t need_new_to = 0; 797 int32_t need_wakeup = 0; 798 uint32_t largest_slot; 799 uint32_t ticknow = 0; 800 uint32_t slot_calc; 801 802 HPTS_MTX_ASSERT(hpts); 803 if (diag) { 804 memset(diag, 0, sizeof(struct hpts_diag)); 805 diag->p_hpts_active = hpts->p_hpts_active; 806 diag->p_nxt_slot = hpts->p_nxt_slot; 807 diag->p_cur_slot = hpts->p_cur_slot; 808 diag->slot_req = slot; 809 } 810 if ((inp->inp_in_hpts == 0) || noref) { 811 inp->inp_hpts_request = slot; 812 if (slot == 0) { 813 /* Immediate */ 814 tcp_queue_to_hpts_immediate_locked(inp, hpts, line, noref); 815 return; 816 } 817 if (hpts->p_hpts_active) { 818 /* 819 * Its slot - 1 since nxt_slot is the next tick that 820 * will go off since the hpts is awake 821 */ 822 if (logging_on) { 823 tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_NORMAL, slot, 0); 824 } 825 /* 826 * We want to make sure that we don't place a inp in 827 * the range of p_cur_slot <-> p_nxt_slot. If we 828 * take from p_nxt_slot to the end, plus p_cur_slot 829 * and then take away 2, we will know how many is 830 * the max slots we can use. 831 */ 832 if (hpts->p_nxt_slot > hpts->p_cur_slot) { 833 /* 834 * Non-wrap case nxt_slot <-> cur_slot we 835 * don't want to land in. So the diff gives 836 * us what is taken away from the number of 837 * slots. 838 */ 839 largest_slot = NUM_OF_HPTSI_SLOTS - (hpts->p_nxt_slot - hpts->p_cur_slot); 840 } else if (hpts->p_nxt_slot == hpts->p_cur_slot) { 841 largest_slot = NUM_OF_HPTSI_SLOTS - 2; 842 } else { 843 /* 844 * Wrap case so the diff gives us the number 845 * of slots that we can land in. 846 */ 847 largest_slot = hpts->p_cur_slot - hpts->p_nxt_slot; 848 } 849 /* 850 * We take away two so we never have a problem (20 851 * usec's) out of 1024000 usecs 852 */ 853 largest_slot -= 2; 854 if (inp->inp_hpts_request > largest_slot) { 855 /* 856 * Restrict max jump of slots and remember 857 * leftover 858 */ 859 slot = largest_slot; 860 inp->inp_hpts_request -= largest_slot; 861 } else { 862 /* This one will run when we hit it */ 863 inp->inp_hpts_request = 0; 864 } 865 if (hpts->p_nxt_slot == hpts->p_cur_slot) 866 slot_calc = (hpts->p_nxt_slot + slot) % NUM_OF_HPTSI_SLOTS; 867 else 868 slot_calc = (hpts->p_nxt_slot + slot - 1) % NUM_OF_HPTSI_SLOTS; 869 if (slot_calc == hpts->p_cur_slot) { 870 #ifdef INVARIANTS 871 /* TSNH */ 872 panic("Hpts:%p impossible slot calculation slot_calc:%u slot:%u largest:%u\n", 873 hpts, slot_calc, slot, largest_slot); 874 #endif 875 if (slot_calc) 876 slot_calc--; 877 else 878 slot_calc = NUM_OF_HPTSI_SLOTS - 1; 879 } 880 inp->inp_hptsslot = slot_calc; 881 if (diag) { 882 diag->inp_hptsslot = inp->inp_hptsslot; 883 } 884 } else { 885 /* 886 * The hpts is sleeping, we need to figure out where 887 * it will wake up at and if we need to reschedule 888 * its time-out. 889 */ 890 uint32_t have_slept, yet_to_sleep; 891 uint32_t slot_now; 892 struct timeval tv; 893 894 ticknow = tcp_gethptstick(&tv); 895 slot_now = ticknow % NUM_OF_HPTSI_SLOTS; 896 /* 897 * The user wants to be inserted at (slot_now + 898 * slot) % NUM_OF_HPTSI_SLOTS, so lets set that up. 899 */ 900 largest_slot = NUM_OF_HPTSI_SLOTS - 2; 901 if (inp->inp_hpts_request > largest_slot) { 902 /* Adjust the residual in inp_hpts_request */ 903 slot = largest_slot; 904 inp->inp_hpts_request -= largest_slot; 905 } else { 906 /* No residual it all fits */ 907 inp->inp_hpts_request = 0; 908 } 909 inp->inp_hptsslot = (slot_now + slot) % NUM_OF_HPTSI_SLOTS; 910 if (diag) { 911 diag->slot_now = slot_now; 912 diag->inp_hptsslot = inp->inp_hptsslot; 913 diag->p_on_min_sleep = hpts->p_on_min_sleep; 914 } 915 if (logging_on) { 916 tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_SLEEPER, slot, ticknow); 917 } 918 /* Now do we need to restart the hpts's timer? */ 919 if (TSTMP_GT(ticknow, hpts->p_curtick)) 920 have_slept = ticknow - hpts->p_curtick; 921 else 922 have_slept = 0; 923 if (have_slept < hpts->p_hpts_sleep_time) { 924 /* This should be what happens */ 925 yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; 926 } else { 927 /* We are over-due */ 928 yet_to_sleep = 0; 929 need_wakeup = 1; 930 } 931 if (diag) { 932 diag->have_slept = have_slept; 933 diag->yet_to_sleep = yet_to_sleep; 934 diag->hpts_sleep_time = hpts->p_hpts_sleep_time; 935 } 936 if ((hpts->p_on_min_sleep == 0) && (yet_to_sleep > slot)) { 937 /* 938 * We need to reschedule the hptss time-out. 939 */ 940 hpts->p_hpts_sleep_time = slot; 941 need_new_to = slot * HPTS_TICKS_PER_USEC; 942 } 943 } 944 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref); 945 if (logging_on) { 946 tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERTED, slot, ticknow); 947 } 948 /* 949 * Now how far is the hpts sleeping to? if active is 1, its 950 * up and ticking we do nothing, otherwise we may need to 951 * reschedule its callout if need_new_to is set from above. 952 */ 953 if (need_wakeup) { 954 if (logging_on) { 955 tcp_hpts_log_it(hpts, inp, HPTSLOG_RESCHEDULE, 1, 0); 956 } 957 hpts->p_direct_wake = 1; 958 tcp_wakehpts(hpts); 959 if (diag) { 960 diag->need_new_to = 0; 961 diag->co_ret = 0xffff0000; 962 } 963 } else if (need_new_to) { 964 int32_t co_ret; 965 struct timeval tv; 966 sbintime_t sb; 967 968 tv.tv_sec = 0; 969 tv.tv_usec = 0; 970 while (need_new_to > HPTS_USEC_IN_SEC) { 971 tv.tv_sec++; 972 need_new_to -= HPTS_USEC_IN_SEC; 973 } 974 tv.tv_usec = need_new_to; 975 sb = tvtosbt(tv); 976 if (tcp_hpts_callout_skip_swi == 0) { 977 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, 978 hpts_timeout_swi, hpts, hpts->p_cpu, 979 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); 980 } else { 981 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, 982 hpts_timeout_dir, hpts, 983 hpts->p_cpu, 984 C_PREL(tcp_hpts_precision)); 985 } 986 if (diag) { 987 diag->need_new_to = need_new_to; 988 diag->co_ret = co_ret; 989 } 990 } 991 } else { 992 #ifdef INVARIANTS 993 panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp); 994 #endif 995 } 996 } 997 998 uint32_t 999 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag){ 1000 struct tcp_hpts_entry *hpts; 1001 uint32_t slot_on, cts; 1002 struct timeval tv; 1003 1004 /* 1005 * We now return the next-slot the hpts will be on, beyond its 1006 * current run (if up) or where it was when it stopped if it is 1007 * sleeping. 1008 */ 1009 INP_WLOCK_ASSERT(inp); 1010 hpts = tcp_hpts_lock(inp); 1011 if (in_ts_percision) 1012 microuptime(&tv); 1013 else 1014 getmicrouptime(&tv); 1015 cts = tcp_tv_to_usectick(&tv); 1016 tcp_hpts_insert_locked(hpts, inp, slot, cts, line, diag, 0); 1017 slot_on = hpts->p_nxt_slot; 1018 mtx_unlock(&hpts->p_mtx); 1019 return (slot_on); 1020 } 1021 1022 uint32_t 1023 __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){ 1024 return (tcp_hpts_insert_diag(inp, slot, line, NULL)); 1025 } 1026 1027 int 1028 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line) 1029 { 1030 int32_t retval = 0; 1031 1032 HPTS_MTX_ASSERT(hpts); 1033 if (inp->inp_in_input == 0) { 1034 /* Ok we need to set it on the hpts in the current slot */ 1035 hpts_sane_input_insert(hpts, inp, line); 1036 retval = 1; 1037 if (hpts->p_hpts_active == 0) { 1038 /* 1039 * Activate the hpts if it is sleeping. 1040 */ 1041 if (logging_on) { 1042 tcp_hpts_log_it(hpts, inp, HPTSLOG_WAKEUP_INPUT, 0, 0); 1043 } 1044 retval = 2; 1045 hpts->p_direct_wake = 1; 1046 tcp_wakeinput(hpts); 1047 } 1048 } else if (hpts->p_hpts_active == 0) { 1049 retval = 4; 1050 hpts->p_direct_wake = 1; 1051 tcp_wakeinput(hpts); 1052 } 1053 return (retval); 1054 } 1055 1056 void 1057 tcp_queue_pkt_to_input(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 1058 int32_t tlen, int32_t drop_hdrlen, uint8_t iptos, uint8_t ti_locked) 1059 { 1060 /* Setup packet for input first */ 1061 INP_WLOCK_ASSERT(tp->t_inpcb); 1062 m->m_pkthdr.pace_thoff = (uint16_t) ((caddr_t)th - mtod(m, caddr_t)); 1063 m->m_pkthdr.pace_tlen = (uint16_t) tlen; 1064 m->m_pkthdr.pace_drphdrlen = drop_hdrlen; 1065 m->m_pkthdr.pace_tos = iptos; 1066 m->m_pkthdr.pace_lock = (uint8_t) ti_locked; 1067 if (tp->t_in_pkt == NULL) { 1068 tp->t_in_pkt = m; 1069 tp->t_tail_pkt = m; 1070 } else { 1071 tp->t_tail_pkt->m_nextpkt = m; 1072 tp->t_tail_pkt = m; 1073 } 1074 } 1075 1076 1077 int32_t 1078 __tcp_queue_to_input(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 1079 int32_t tlen, int32_t drop_hdrlen, uint8_t iptos, uint8_t ti_locked, int32_t line){ 1080 struct tcp_hpts_entry *hpts; 1081 int32_t ret; 1082 1083 tcp_queue_pkt_to_input(tp, m, th, tlen, drop_hdrlen, iptos, ti_locked); 1084 hpts = tcp_input_lock(tp->t_inpcb); 1085 ret = __tcp_queue_to_input_locked(tp->t_inpcb, hpts, line); 1086 mtx_unlock(&hpts->p_mtx); 1087 return (ret); 1088 } 1089 1090 void 1091 __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line) 1092 { 1093 struct tcp_hpts_entry *hpts; 1094 struct tcpcb *tp; 1095 1096 tp = intotcpcb(inp); 1097 hpts = tcp_input_lock(tp->t_inpcb); 1098 if (inp->inp_in_input == 0) { 1099 /* Ok we need to set it on the hpts in the current slot */ 1100 hpts_sane_input_insert(hpts, inp, line); 1101 if (hpts->p_hpts_active == 0) { 1102 /* 1103 * Activate the hpts if it is sleeping. 1104 */ 1105 hpts->p_direct_wake = 1; 1106 tcp_wakeinput(hpts); 1107 } 1108 } else if (hpts->p_hpts_active == 0) { 1109 hpts->p_direct_wake = 1; 1110 tcp_wakeinput(hpts); 1111 } 1112 inp->inp_hpts_drop_reas = reason; 1113 mtx_unlock(&hpts->p_mtx); 1114 } 1115 1116 static uint16_t 1117 hpts_random_cpu(struct inpcb *inp){ 1118 /* 1119 * No flow type set distribute the load randomly. 1120 */ 1121 uint16_t cpuid; 1122 uint32_t ran; 1123 1124 /* 1125 * If one has been set use it i.e. we want both in and out on the 1126 * same hpts. 1127 */ 1128 if (inp->inp_input_cpu_set) { 1129 return (inp->inp_input_cpu); 1130 } else if (inp->inp_hpts_cpu_set) { 1131 return (inp->inp_hpts_cpu); 1132 } 1133 /* Nothing set use a random number */ 1134 ran = arc4random(); 1135 cpuid = (ran & 0xffff) % mp_ncpus; 1136 return (cpuid); 1137 } 1138 1139 static uint16_t 1140 hpts_cpuid(struct inpcb *inp){ 1141 uint16_t cpuid; 1142 1143 1144 /* 1145 * If one has been set use it i.e. we want both in and out on the 1146 * same hpts. 1147 */ 1148 if (inp->inp_input_cpu_set) { 1149 return (inp->inp_input_cpu); 1150 } else if (inp->inp_hpts_cpu_set) { 1151 return (inp->inp_hpts_cpu); 1152 } 1153 /* If one is set the other must be the same */ 1154 #ifdef RSS 1155 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); 1156 if (cpuid == NETISR_CPUID_NONE) 1157 return (hpts_random_cpu(inp)); 1158 else 1159 return (cpuid); 1160 #else 1161 /* 1162 * We don't have a flowid -> cpuid mapping, so cheat and just map 1163 * unknown cpuids to curcpu. Not the best, but apparently better 1164 * than defaulting to swi 0. 1165 */ 1166 if (inp->inp_flowtype != M_HASHTYPE_NONE) { 1167 cpuid = inp->inp_flowid % mp_ncpus; 1168 return (cpuid); 1169 } 1170 cpuid = hpts_random_cpu(inp); 1171 return (cpuid); 1172 #endif 1173 } 1174 1175 /* 1176 * Do NOT try to optimize the processing of inp's 1177 * by first pulling off all the inp's into a temporary 1178 * list (e.g. TAILQ_CONCAT). If you do that the subtle 1179 * interactions of switching CPU's will kill because of 1180 * problems in the linked list manipulation. Basically 1181 * you would switch cpu's with the hpts mutex locked 1182 * but then while you were processing one of the inp's 1183 * some other one that you switch will get a new 1184 * packet on the different CPU. It will insert it 1185 * on the new hptss input list. Creating a temporary 1186 * link in the inp will not fix it either, since 1187 * the other hpts will be doing the same thing and 1188 * you will both end up using the temporary link. 1189 * 1190 * You will die in an ASSERT for tailq corruption if you 1191 * run INVARIANTS or you will die horribly without 1192 * INVARIANTS in some unknown way with a corrupt linked 1193 * list. 1194 */ 1195 static void 1196 tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv) 1197 { 1198 struct mbuf *m, *n; 1199 struct tcpcb *tp; 1200 struct inpcb *inp; 1201 uint16_t drop_reason; 1202 int16_t set_cpu; 1203 uint32_t did_prefetch = 0; 1204 int32_t ti_locked = TI_UNLOCKED; 1205 1206 HPTS_MTX_ASSERT(hpts); 1207 while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) { 1208 HPTS_MTX_ASSERT(hpts); 1209 hpts_sane_input_remove(hpts, inp, 0); 1210 if (inp->inp_input_cpu_set == 0) { 1211 set_cpu = 1; 1212 } else { 1213 set_cpu = 0; 1214 } 1215 hpts->p_inp = inp; 1216 drop_reason = inp->inp_hpts_drop_reas; 1217 inp->inp_in_input = 0; 1218 tp = intotcpcb(inp); 1219 mtx_unlock(&hpts->p_mtx); 1220 CURVNET_SET(tp->t_vnet); 1221 if (drop_reason) { 1222 INP_INFO_RLOCK(&V_tcbinfo); 1223 ti_locked = TI_RLOCKED; 1224 } else { 1225 ti_locked = TI_UNLOCKED; 1226 } 1227 INP_WLOCK(inp); 1228 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || 1229 (inp->inp_flags2 & INP_FREED)) { 1230 out: 1231 hpts->p_inp = NULL; 1232 if (ti_locked == TI_RLOCKED) { 1233 INP_INFO_RUNLOCK(&V_tcbinfo); 1234 } 1235 if (in_pcbrele_wlocked(inp) == 0) { 1236 INP_WUNLOCK(inp); 1237 } 1238 ti_locked = TI_UNLOCKED; 1239 CURVNET_RESTORE(); 1240 mtx_lock(&hpts->p_mtx); 1241 continue; 1242 } 1243 if ((tp == NULL) || (tp->t_inpcb == NULL)) { 1244 goto out; 1245 } 1246 if (drop_reason) { 1247 /* This tcb is being destroyed for drop_reason */ 1248 m = tp->t_in_pkt; 1249 if (m) 1250 n = m->m_nextpkt; 1251 else 1252 n = NULL; 1253 tp->t_in_pkt = NULL; 1254 while (m) { 1255 m_freem(m); 1256 m = n; 1257 if (m) 1258 n = m->m_nextpkt; 1259 } 1260 tp = tcp_drop(tp, drop_reason); 1261 INP_INFO_RUNLOCK(&V_tcbinfo); 1262 if (tp == NULL) { 1263 INP_WLOCK(inp); 1264 } 1265 if (in_pcbrele_wlocked(inp) == 0) 1266 INP_WUNLOCK(inp); 1267 CURVNET_RESTORE(); 1268 mtx_lock(&hpts->p_mtx); 1269 continue; 1270 } 1271 if (set_cpu) { 1272 /* 1273 * Setup so the next time we will move to the right 1274 * CPU. This should be a rare event. It will 1275 * sometimes happens when we are the client side 1276 * (usually not the server). Somehow tcp_output() 1277 * gets called before the tcp_do_segment() sets the 1278 * intial state. This means the r_cpu and r_hpts_cpu 1279 * is 0. We get on the hpts, and then tcp_input() 1280 * gets called setting up the r_cpu to the correct 1281 * value. The hpts goes off and sees the mis-match. 1282 * We simply correct it here and the CPU will switch 1283 * to the new hpts nextime the tcb gets added to the 1284 * the hpts (not this time) :-) 1285 */ 1286 tcp_set_hpts(inp); 1287 } 1288 m = tp->t_in_pkt; 1289 n = NULL; 1290 if (m != NULL && 1291 (m->m_pkthdr.pace_lock == TI_RLOCKED || 1292 tp->t_state != TCPS_ESTABLISHED)) { 1293 ti_locked = TI_RLOCKED; 1294 if (tcp_hptsi_lock_inpinfo(inp, &tp)) { 1295 CURVNET_RESTORE(); 1296 goto out; 1297 } 1298 m = tp->t_in_pkt; 1299 } 1300 if (in_newts_every_tcb) { 1301 if (in_ts_percision) 1302 microuptime(tv); 1303 else 1304 getmicrouptime(tv); 1305 } 1306 if (tp->t_fb_ptr != NULL) { 1307 kern_prefetch(tp->t_fb_ptr, &did_prefetch); 1308 did_prefetch = 1; 1309 } 1310 /* Any input work to do, if so do it first */ 1311 if ((m != NULL) && (m == tp->t_in_pkt)) { 1312 struct tcphdr *th; 1313 int32_t tlen, drop_hdrlen, nxt_pkt; 1314 uint8_t iptos; 1315 1316 n = m->m_nextpkt; 1317 tp->t_in_pkt = tp->t_tail_pkt = NULL; 1318 while (m) { 1319 th = (struct tcphdr *)(mtod(m, caddr_t)+m->m_pkthdr.pace_thoff); 1320 tlen = m->m_pkthdr.pace_tlen; 1321 drop_hdrlen = m->m_pkthdr.pace_drphdrlen; 1322 iptos = m->m_pkthdr.pace_tos; 1323 m->m_nextpkt = NULL; 1324 if (n) 1325 nxt_pkt = 1; 1326 else 1327 nxt_pkt = 0; 1328 inp->inp_input_calls = 1; 1329 if (tp->t_fb->tfb_tcp_hpts_do_segment) { 1330 /* Use the hpts specific do_segment */ 1331 (*tp->t_fb->tfb_tcp_hpts_do_segment) (m, th, inp->inp_socket, 1332 tp, drop_hdrlen, 1333 tlen, iptos, ti_locked, nxt_pkt, tv); 1334 } else { 1335 /* Use the default do_segment */ 1336 (*tp->t_fb->tfb_tcp_do_segment) (m, th, inp->inp_socket, 1337 tp, drop_hdrlen, 1338 tlen, iptos, ti_locked); 1339 } 1340 /* 1341 * Do segment returns unlocked we need the 1342 * lock again but we also need some kasserts 1343 * here. 1344 */ 1345 INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); 1346 INP_UNLOCK_ASSERT(inp); 1347 m = n; 1348 if (m) 1349 n = m->m_nextpkt; 1350 if (m != NULL && 1351 m->m_pkthdr.pace_lock == TI_RLOCKED) { 1352 INP_INFO_RLOCK(&V_tcbinfo); 1353 ti_locked = TI_RLOCKED; 1354 } else 1355 ti_locked = TI_UNLOCKED; 1356 INP_WLOCK(inp); 1357 /* 1358 * Since we have an opening here we must 1359 * re-check if the tcb went away while we 1360 * were getting the lock(s). 1361 */ 1362 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || 1363 (inp->inp_flags2 & INP_FREED)) { 1364 out_free: 1365 while (m) { 1366 m_freem(m); 1367 m = n; 1368 if (m) 1369 n = m->m_nextpkt; 1370 } 1371 goto out; 1372 } 1373 /* 1374 * Now that we hold the INP lock, check if 1375 * we need to upgrade our lock. 1376 */ 1377 if (ti_locked == TI_UNLOCKED && 1378 (tp->t_state != TCPS_ESTABLISHED)) { 1379 ti_locked = TI_RLOCKED; 1380 if (tcp_hptsi_lock_inpinfo(inp, &tp)) 1381 goto out_free; 1382 } 1383 } /** end while(m) */ 1384 } /** end if ((m != NULL) && (m == tp->t_in_pkt)) */ 1385 if (in_pcbrele_wlocked(inp) == 0) 1386 INP_WUNLOCK(inp); 1387 if (ti_locked == TI_RLOCKED) 1388 INP_INFO_RUNLOCK(&V_tcbinfo); 1389 INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); 1390 INP_UNLOCK_ASSERT(inp); 1391 ti_locked = TI_UNLOCKED; 1392 mtx_lock(&hpts->p_mtx); 1393 hpts->p_inp = NULL; 1394 CURVNET_RESTORE(); 1395 } 1396 } 1397 1398 static int 1399 tcp_hpts_est_run(struct tcp_hpts_entry *hpts) 1400 { 1401 int32_t ticks_to_run; 1402 1403 if (hpts->p_prevtick && (SEQ_GT(hpts->p_curtick, hpts->p_prevtick))) { 1404 ticks_to_run = hpts->p_curtick - hpts->p_prevtick; 1405 if (ticks_to_run >= (NUM_OF_HPTSI_SLOTS - 1)) { 1406 ticks_to_run = NUM_OF_HPTSI_SLOTS - 2; 1407 } 1408 } else { 1409 if (hpts->p_prevtick == hpts->p_curtick) { 1410 /* This happens when we get woken up right away */ 1411 return (-1); 1412 } 1413 ticks_to_run = 1; 1414 } 1415 /* Set in where we will be when we catch up */ 1416 hpts->p_nxt_slot = (hpts->p_cur_slot + ticks_to_run) % NUM_OF_HPTSI_SLOTS; 1417 if (hpts->p_nxt_slot == hpts->p_cur_slot) { 1418 panic("Impossible math -- hpts:%p p_nxt_slot:%d p_cur_slot:%d ticks_to_run:%d", 1419 hpts, hpts->p_nxt_slot, hpts->p_cur_slot, ticks_to_run); 1420 } 1421 return (ticks_to_run); 1422 } 1423 1424 static void 1425 tcp_hptsi(struct tcp_hpts_entry *hpts, struct timeval *ctick) 1426 { 1427 struct tcpcb *tp; 1428 struct inpcb *inp = NULL, *ninp; 1429 struct timeval tv; 1430 int32_t ticks_to_run, i, error, tick_now, interum_tick; 1431 int32_t paced_cnt = 0; 1432 int32_t did_prefetch = 0; 1433 int32_t prefetch_ninp = 0; 1434 int32_t prefetch_tp = 0; 1435 uint32_t cts; 1436 int16_t set_cpu; 1437 1438 HPTS_MTX_ASSERT(hpts); 1439 hpts->p_curtick = tcp_tv_to_hptstick(ctick); 1440 cts = tcp_tv_to_usectick(ctick); 1441 memcpy(&tv, ctick, sizeof(struct timeval)); 1442 hpts->p_cur_slot = hpts_tick(hpts, 1); 1443 1444 /* Figure out if we had missed ticks */ 1445 again: 1446 HPTS_MTX_ASSERT(hpts); 1447 ticks_to_run = tcp_hpts_est_run(hpts); 1448 if (!TAILQ_EMPTY(&hpts->p_input)) { 1449 tcp_input_data(hpts, &tv); 1450 } 1451 #ifdef INVARIANTS 1452 if (TAILQ_EMPTY(&hpts->p_input) && 1453 (hpts->p_on_inqueue_cnt != 0)) { 1454 panic("tp:%p in_hpts input empty but cnt:%d", 1455 hpts, hpts->p_on_inqueue_cnt); 1456 } 1457 #endif 1458 HPTS_MTX_ASSERT(hpts); 1459 /* Reset the ticks to run and time if we need too */ 1460 interum_tick = tcp_gethptstick(&tv); 1461 if (interum_tick != hpts->p_curtick) { 1462 /* Save off the new time we execute to */ 1463 *ctick = tv; 1464 hpts->p_curtick = interum_tick; 1465 cts = tcp_tv_to_usectick(&tv); 1466 hpts->p_cur_slot = hpts_tick(hpts, 1); 1467 ticks_to_run = tcp_hpts_est_run(hpts); 1468 } 1469 if (ticks_to_run == -1) { 1470 goto no_run; 1471 } 1472 if (logging_on) { 1473 tcp_hpts_log_it(hpts, inp, HPTSLOG_SETTORUN, ticks_to_run, 0); 1474 } 1475 if (hpts->p_on_queue_cnt == 0) { 1476 goto no_one; 1477 } 1478 HPTS_MTX_ASSERT(hpts); 1479 for (i = 0; i < ticks_to_run; i++) { 1480 /* 1481 * Calculate our delay, if there are no extra ticks there 1482 * was not any 1483 */ 1484 hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC; 1485 HPTS_MTX_ASSERT(hpts); 1486 while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_cur_slot])) != NULL) { 1487 /* For debugging */ 1488 if (logging_on) { 1489 tcp_hpts_log_it(hpts, inp, HPTSLOG_HPTSI, ticks_to_run, i); 1490 } 1491 hpts->p_inp = inp; 1492 paced_cnt++; 1493 if (hpts->p_cur_slot != inp->inp_hptsslot) { 1494 panic("Hpts:%p inp:%p slot mis-aligned %u vs %u", 1495 hpts, inp, hpts->p_cur_slot, inp->inp_hptsslot); 1496 } 1497 /* Now pull it */ 1498 if (inp->inp_hpts_cpu_set == 0) { 1499 set_cpu = 1; 1500 } else { 1501 set_cpu = 0; 1502 } 1503 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_cur_slot], 0); 1504 if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_cur_slot])) != NULL) { 1505 /* We prefetch the next inp if possible */ 1506 kern_prefetch(ninp, &prefetch_ninp); 1507 prefetch_ninp = 1; 1508 } 1509 if (inp->inp_hpts_request) { 1510 /* 1511 * This guy is deferred out further in time 1512 * then our wheel had on it. Push him back 1513 * on the wheel. 1514 */ 1515 int32_t remaining_slots; 1516 1517 remaining_slots = ticks_to_run - (i + 1); 1518 if (inp->inp_hpts_request > remaining_slots) { 1519 /* 1520 * Keep INVARIANTS happy by clearing 1521 * the flag 1522 */ 1523 tcp_hpts_insert_locked(hpts, inp, inp->inp_hpts_request, cts, __LINE__, NULL, 1); 1524 hpts->p_inp = NULL; 1525 continue; 1526 } 1527 inp->inp_hpts_request = 0; 1528 } 1529 /* 1530 * We clear the hpts flag here after dealing with 1531 * remaining slots. This way anyone looking with the 1532 * TCB lock will see its on the hpts until just 1533 * before we unlock. 1534 */ 1535 inp->inp_in_hpts = 0; 1536 mtx_unlock(&hpts->p_mtx); 1537 INP_WLOCK(inp); 1538 if (in_pcbrele_wlocked(inp)) { 1539 mtx_lock(&hpts->p_mtx); 1540 if (logging_on) 1541 tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 1); 1542 hpts->p_inp = NULL; 1543 continue; 1544 } 1545 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 1546 out_now: 1547 #ifdef INVARIANTS 1548 if (mtx_owned(&hpts->p_mtx)) { 1549 panic("Hpts:%p owns mtx prior-to lock line:%d", 1550 hpts, __LINE__); 1551 } 1552 #endif 1553 INP_WUNLOCK(inp); 1554 mtx_lock(&hpts->p_mtx); 1555 if (logging_on) 1556 tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 3); 1557 hpts->p_inp = NULL; 1558 continue; 1559 } 1560 tp = intotcpcb(inp); 1561 if ((tp == NULL) || (tp->t_inpcb == NULL)) { 1562 goto out_now; 1563 } 1564 if (set_cpu) { 1565 /* 1566 * Setup so the next time we will move to 1567 * the right CPU. This should be a rare 1568 * event. It will sometimes happens when we 1569 * are the client side (usually not the 1570 * server). Somehow tcp_output() gets called 1571 * before the tcp_do_segment() sets the 1572 * intial state. This means the r_cpu and 1573 * r_hpts_cpu is 0. We get on the hpts, and 1574 * then tcp_input() gets called setting up 1575 * the r_cpu to the correct value. The hpts 1576 * goes off and sees the mis-match. We 1577 * simply correct it here and the CPU will 1578 * switch to the new hpts nextime the tcb 1579 * gets added to the the hpts (not this one) 1580 * :-) 1581 */ 1582 tcp_set_hpts(inp); 1583 } 1584 if (out_newts_every_tcb) { 1585 struct timeval sv; 1586 1587 if (out_ts_percision) 1588 microuptime(&sv); 1589 else 1590 getmicrouptime(&sv); 1591 cts = tcp_tv_to_usectick(&sv); 1592 } 1593 CURVNET_SET(tp->t_vnet); 1594 /* 1595 * There is a hole here, we get the refcnt on the 1596 * inp so it will still be preserved but to make 1597 * sure we can get the INP we need to hold the p_mtx 1598 * above while we pull out the tp/inp, as long as 1599 * fini gets the lock first we are assured of having 1600 * a sane INP we can lock and test. 1601 */ 1602 #ifdef INVARIANTS 1603 if (mtx_owned(&hpts->p_mtx)) { 1604 panic("Hpts:%p owns mtx before tcp-output:%d", 1605 hpts, __LINE__); 1606 } 1607 #endif 1608 if (tp->t_fb_ptr != NULL) { 1609 kern_prefetch(tp->t_fb_ptr, &did_prefetch); 1610 did_prefetch = 1; 1611 } 1612 inp->inp_hpts_calls = 1; 1613 if (tp->t_fb->tfb_tcp_output_wtime != NULL) { 1614 error = (*tp->t_fb->tfb_tcp_output_wtime) (tp, &tv); 1615 } else { 1616 error = tp->t_fb->tfb_tcp_output(tp); 1617 } 1618 if (ninp && ninp->inp_ppcb) { 1619 /* 1620 * If we have a nxt inp, see if we can 1621 * prefetch its ppcb. Note this may seem 1622 * "risky" since we have no locks (other 1623 * than the previous inp) and there no 1624 * assurance that ninp was not pulled while 1625 * we were processing inp and freed. If this 1626 * occured it could mean that either: 1627 * 1628 * a) Its NULL (which is fine we won't go 1629 * here) <or> b) Its valid (which is cool we 1630 * will prefetch it) <or> c) The inp got 1631 * freed back to the slab which was 1632 * reallocated. Then the piece of memory was 1633 * re-used and something else (not an 1634 * address) is in inp_ppcb. If that occurs 1635 * we don't crash, but take a TLB shootdown 1636 * performance hit (same as if it was NULL 1637 * and we tried to pre-fetch it). 1638 * 1639 * Considering that the likelyhood of <c> is 1640 * quite rare we will take a risk on doing 1641 * this. If performance drops after testing 1642 * we can always take this out. NB: the 1643 * kern_prefetch on amd64 actually has 1644 * protection against a bad address now via 1645 * the DMAP_() tests. This will prevent the 1646 * TLB hit, and instead if <c> occurs just 1647 * cause us to load cache with a useless 1648 * address (to us). 1649 */ 1650 kern_prefetch(ninp->inp_ppcb, &prefetch_tp); 1651 prefetch_tp = 1; 1652 } 1653 INP_WUNLOCK(inp); 1654 INP_UNLOCK_ASSERT(inp); 1655 CURVNET_RESTORE(); 1656 #ifdef INVARIANTS 1657 if (mtx_owned(&hpts->p_mtx)) { 1658 panic("Hpts:%p owns mtx prior-to lock line:%d", 1659 hpts, __LINE__); 1660 } 1661 #endif 1662 mtx_lock(&hpts->p_mtx); 1663 if (logging_on) 1664 tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 4); 1665 hpts->p_inp = NULL; 1666 } 1667 HPTS_MTX_ASSERT(hpts); 1668 hpts->p_inp = NULL; 1669 hpts->p_cur_slot++; 1670 if (hpts->p_cur_slot >= NUM_OF_HPTSI_SLOTS) { 1671 hpts->p_cur_slot = 0; 1672 } 1673 } 1674 no_one: 1675 HPTS_MTX_ASSERT(hpts); 1676 hpts->p_prevtick = hpts->p_curtick; 1677 hpts->p_delayed_by = 0; 1678 /* 1679 * Check to see if we took an excess amount of time and need to run 1680 * more ticks (if we did not hit eno-bufs). 1681 */ 1682 /* Re-run any input that may be there */ 1683 (void)tcp_gethptstick(&tv); 1684 if (!TAILQ_EMPTY(&hpts->p_input)) { 1685 tcp_input_data(hpts, &tv); 1686 } 1687 #ifdef INVARIANTS 1688 if (TAILQ_EMPTY(&hpts->p_input) && 1689 (hpts->p_on_inqueue_cnt != 0)) { 1690 panic("tp:%p in_hpts input empty but cnt:%d", 1691 hpts, hpts->p_on_inqueue_cnt); 1692 } 1693 #endif 1694 tick_now = tcp_gethptstick(&tv); 1695 if (SEQ_GT(tick_now, hpts->p_prevtick)) { 1696 struct timeval res; 1697 1698 /* Did we really spend a full tick or more in here? */ 1699 timersub(&tv, ctick, &res); 1700 if (res.tv_sec || (res.tv_usec >= HPTS_TICKS_PER_USEC)) { 1701 counter_u64_add(hpts_loops, 1); 1702 if (logging_on) { 1703 tcp_hpts_log_it(hpts, inp, HPTSLOG_TOLONG, (uint32_t) res.tv_usec, tick_now); 1704 } 1705 *ctick = res; 1706 hpts->p_curtick = tick_now; 1707 goto again; 1708 } 1709 } 1710 no_run: 1711 { 1712 uint32_t t = 0, i, fnd = 0; 1713 1714 if (hpts->p_on_queue_cnt) { 1715 1716 1717 /* 1718 * Find next slot that is occupied and use that to 1719 * be the sleep time. 1720 */ 1721 for (i = 1, t = hpts->p_nxt_slot; i < NUM_OF_HPTSI_SLOTS; i++) { 1722 if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) { 1723 fnd = 1; 1724 break; 1725 } 1726 t = (t + 1) % NUM_OF_HPTSI_SLOTS; 1727 } 1728 if (fnd) { 1729 hpts->p_hpts_sleep_time = i; 1730 } else { 1731 counter_u64_add(back_tosleep, 1); 1732 #ifdef INVARIANTS 1733 panic("Hpts:%p cnt:%d but non found", hpts, hpts->p_on_queue_cnt); 1734 #endif 1735 hpts->p_on_queue_cnt = 0; 1736 goto non_found; 1737 } 1738 t++; 1739 } else { 1740 /* No one on the wheel sleep for all but 2 slots */ 1741 non_found: 1742 if (hpts_sleep_max == 0) 1743 hpts_sleep_max = 1; 1744 hpts->p_hpts_sleep_time = min((NUM_OF_HPTSI_SLOTS - 2), hpts_sleep_max); 1745 t = 0; 1746 } 1747 if (logging_on) { 1748 tcp_hpts_log_it(hpts, inp, HPTSLOG_SLEEPSET, t, (hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC)); 1749 } 1750 } 1751 } 1752 1753 void 1754 __tcp_set_hpts(struct inpcb *inp, int32_t line) 1755 { 1756 struct tcp_hpts_entry *hpts; 1757 1758 INP_WLOCK_ASSERT(inp); 1759 hpts = tcp_hpts_lock(inp); 1760 if ((inp->inp_in_hpts == 0) && 1761 (inp->inp_hpts_cpu_set == 0)) { 1762 inp->inp_hpts_cpu = hpts_cpuid(inp); 1763 inp->inp_hpts_cpu_set = 1; 1764 } 1765 mtx_unlock(&hpts->p_mtx); 1766 hpts = tcp_input_lock(inp); 1767 if ((inp->inp_input_cpu_set == 0) && 1768 (inp->inp_in_input == 0)) { 1769 inp->inp_input_cpu = hpts_cpuid(inp); 1770 inp->inp_input_cpu_set = 1; 1771 } 1772 mtx_unlock(&hpts->p_mtx); 1773 } 1774 1775 uint16_t 1776 tcp_hpts_delayedby(struct inpcb *inp){ 1777 return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by); 1778 } 1779 1780 static void 1781 tcp_hpts_thread(void *ctx) 1782 { 1783 struct tcp_hpts_entry *hpts; 1784 struct timeval tv; 1785 sbintime_t sb; 1786 1787 hpts = (struct tcp_hpts_entry *)ctx; 1788 mtx_lock(&hpts->p_mtx); 1789 if (hpts->p_direct_wake) { 1790 /* Signaled by input */ 1791 if (logging_on) 1792 tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 1, 1); 1793 callout_stop(&hpts->co); 1794 } else { 1795 /* Timed out */ 1796 if (callout_pending(&hpts->co) || 1797 !callout_active(&hpts->co)) { 1798 if (logging_on) 1799 tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 2, 2); 1800 mtx_unlock(&hpts->p_mtx); 1801 return; 1802 } 1803 callout_deactivate(&hpts->co); 1804 if (logging_on) 1805 tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 3, 3); 1806 } 1807 hpts->p_hpts_active = 1; 1808 (void)tcp_gethptstick(&tv); 1809 tcp_hptsi(hpts, &tv); 1810 HPTS_MTX_ASSERT(hpts); 1811 tv.tv_sec = 0; 1812 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; 1813 if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) { 1814 tv.tv_usec = tcp_min_hptsi_time; 1815 hpts->p_on_min_sleep = 1; 1816 } else { 1817 /* Clear the min sleep flag */ 1818 hpts->p_on_min_sleep = 0; 1819 } 1820 hpts->p_hpts_active = 0; 1821 sb = tvtosbt(tv); 1822 if (tcp_hpts_callout_skip_swi == 0) { 1823 callout_reset_sbt_on(&hpts->co, sb, 0, 1824 hpts_timeout_swi, hpts, hpts->p_cpu, 1825 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); 1826 } else { 1827 callout_reset_sbt_on(&hpts->co, sb, 0, 1828 hpts_timeout_dir, hpts, 1829 hpts->p_cpu, 1830 C_PREL(tcp_hpts_precision)); 1831 } 1832 hpts->p_direct_wake = 0; 1833 mtx_unlock(&hpts->p_mtx); 1834 } 1835 1836 #undef timersub 1837 1838 static void 1839 tcp_init_hptsi(void *st) 1840 { 1841 int32_t i, j, error, bound = 0, created = 0; 1842 size_t sz, asz; 1843 struct timeval tv; 1844 sbintime_t sb; 1845 struct tcp_hpts_entry *hpts; 1846 char unit[16]; 1847 uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1848 1849 tcp_pace.rp_proc = NULL; 1850 tcp_pace.rp_num_hptss = ncpus; 1851 hpts_loops = counter_u64_alloc(M_WAITOK); 1852 back_tosleep = counter_u64_alloc(M_WAITOK); 1853 1854 sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *)); 1855 tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); 1856 asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS; 1857 for (i = 0; i < tcp_pace.rp_num_hptss; i++) { 1858 tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry), 1859 M_TCPHPTS, M_WAITOK | M_ZERO); 1860 tcp_pace.rp_ent[i]->p_hptss = malloc(asz, 1861 M_TCPHPTS, M_WAITOK); 1862 hpts = tcp_pace.rp_ent[i]; 1863 /* 1864 * Init all the hpts structures that are not specifically 1865 * zero'd by the allocations. Also lets attach them to the 1866 * appropriate sysctl block as well. 1867 */ 1868 mtx_init(&hpts->p_mtx, "tcp_hpts_lck", 1869 "hpts", MTX_DEF | MTX_DUPOK); 1870 TAILQ_INIT(&hpts->p_input); 1871 for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) { 1872 TAILQ_INIT(&hpts->p_hptss[j]); 1873 } 1874 sysctl_ctx_init(&hpts->hpts_ctx); 1875 sprintf(unit, "%d", i); 1876 hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx, 1877 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts), 1878 OID_AUTO, 1879 unit, 1880 CTLFLAG_RW, 0, 1881 ""); 1882 SYSCTL_ADD_INT(&hpts->hpts_ctx, 1883 SYSCTL_CHILDREN(hpts->hpts_root), 1884 OID_AUTO, "in_qcnt", CTLFLAG_RD, 1885 &hpts->p_on_inqueue_cnt, 0, 1886 "Count TCB's awaiting input processing"); 1887 SYSCTL_ADD_INT(&hpts->hpts_ctx, 1888 SYSCTL_CHILDREN(hpts->hpts_root), 1889 OID_AUTO, "out_qcnt", CTLFLAG_RD, 1890 &hpts->p_on_queue_cnt, 0, 1891 "Count TCB's awaiting output processing"); 1892 SYSCTL_ADD_UINT(&hpts->hpts_ctx, 1893 SYSCTL_CHILDREN(hpts->hpts_root), 1894 OID_AUTO, "active", CTLFLAG_RD, 1895 &hpts->p_hpts_active, 0, 1896 "Is the hpts active"); 1897 SYSCTL_ADD_UINT(&hpts->hpts_ctx, 1898 SYSCTL_CHILDREN(hpts->hpts_root), 1899 OID_AUTO, "curslot", CTLFLAG_RD, 1900 &hpts->p_cur_slot, 0, 1901 "What the current slot is if active"); 1902 SYSCTL_ADD_UINT(&hpts->hpts_ctx, 1903 SYSCTL_CHILDREN(hpts->hpts_root), 1904 OID_AUTO, "curtick", CTLFLAG_RD, 1905 &hpts->p_curtick, 0, 1906 "What the current tick on if active"); 1907 SYSCTL_ADD_UINT(&hpts->hpts_ctx, 1908 SYSCTL_CHILDREN(hpts->hpts_root), 1909 OID_AUTO, "logsize", CTLFLAG_RD, 1910 &hpts->p_logsize, 0, 1911 "Hpts logging buffer size"); 1912 hpts->p_hpts_sleep_time = NUM_OF_HPTSI_SLOTS - 2; 1913 hpts->p_num = i; 1914 hpts->p_prevtick = hpts->p_curtick = tcp_gethptstick(&tv); 1915 hpts->p_prevtick -= 1; 1916 hpts->p_prevtick %= NUM_OF_HPTSI_SLOTS; 1917 hpts->p_cpu = 0xffff; 1918 hpts->p_nxt_slot = 1; 1919 hpts->p_logsize = tcp_hpts_logging_size; 1920 if (hpts->p_logsize) { 1921 sz = (sizeof(struct hpts_log) * hpts->p_logsize); 1922 hpts->p_log = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); 1923 } 1924 callout_init(&hpts->co, 1); 1925 } 1926 /* 1927 * Now lets start ithreads to handle the hptss. 1928 */ 1929 CPU_FOREACH(i) { 1930 hpts = tcp_pace.rp_ent[i]; 1931 hpts->p_cpu = i; 1932 error = swi_add(&hpts->ie, "hpts", 1933 tcp_hpts_thread, (void *)hpts, 1934 SWI_NET, INTR_MPSAFE, &hpts->ie_cookie); 1935 if (error) { 1936 panic("Can't add hpts:%p i:%d err:%d", 1937 hpts, i, error); 1938 } 1939 created++; 1940 if (tcp_bind_threads) { 1941 if (intr_event_bind(hpts->ie, i) == 0) 1942 bound++; 1943 } 1944 tv.tv_sec = 0; 1945 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; 1946 sb = tvtosbt(tv); 1947 if (tcp_hpts_callout_skip_swi == 0) { 1948 callout_reset_sbt_on(&hpts->co, sb, 0, 1949 hpts_timeout_swi, hpts, hpts->p_cpu, 1950 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); 1951 } else { 1952 callout_reset_sbt_on(&hpts->co, sb, 0, 1953 hpts_timeout_dir, hpts, 1954 hpts->p_cpu, 1955 C_PREL(tcp_hpts_precision)); 1956 } 1957 } 1958 printf("TCP Hpts created %d swi interrupt thread and bound %d\n", 1959 created, bound); 1960 return; 1961 } 1962 1963 SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL); 1964 MODULE_VERSION(tcphpts, 1); 1965