2 * Copyright (c) 2016-2018 Netflix, Inc.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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.
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
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD$");
30 #include "opt_inet6.h"
32 #include "opt_tcpdebug.h"
35 * Some notes about usage.
37 * The tcp_hpts system is designed to provide a high precision timer
38 * system for tcp. Its main purpose is to provide a mechanism for
39 * pacing packets out onto the wire. It can be used in two ways
40 * by a given TCP stack (and those two methods can be used simultaneously).
42 * First, and probably the main thing its used by Rack and BBR, it can
43 * be used to call tcp_output() of a transport stack at some time in the future.
44 * The normal way this is done is that tcp_output() of the stack schedules
45 * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The
46 * slot is the time from now that the stack wants to be called but it
47 * must be converted to tcp_hpts's notion of slot. This is done with
48 * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical
49 * call from the tcp_output() routine might look like:
51 * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550));
53 * The above would schedule tcp_ouput() to be called in 550 useconds.
54 * Note that if using this mechanism the stack will want to add near
55 * its top a check to prevent unwanted calls (from user land or the
56 * arrival of incoming ack's). So it would add something like:
58 * if (inp->inp_in_hpts)
61 * to prevent output processing until the time alotted has gone by.
62 * Of course this is a bare bones example and the stack will probably
63 * have more consideration then just the above.
65 * Now the second function (actually two functions I guess :D)
66 * the tcp_hpts system provides is the ability to either abort
67 * a connection (later) or process input on a connection.
68 * Why would you want to do this? To keep processor locality
69 * and or not have to worry about untangling any recursive
70 * locks. The input function now is hooked to the new LRO
73 * In order to use the input redirection function the
74 * tcp stack must define an input function for
75 * tfb_do_queued_segments(). This function understands
76 * how to dequeue a array of packets that were input and
77 * knows how to call the correct processing routine.
79 * Locking in this is important as well so most likely the
80 * stack will need to define the tfb_do_segment_nounlock()
81 * splitting tfb_do_segment() into two parts. The main processing
82 * part that does not unlock the INP and returns a value of 1 or 0.
83 * It returns 0 if all is well and the lock was not released. It
84 * returns 1 if we had to destroy the TCB (a reset received etc).
85 * The remains of tfb_do_segment() then become just a simple call
86 * to the tfb_do_segment_nounlock() function and check the return
87 * code and possibly unlock.
89 * The stack must also set the flag on the INP that it supports this
90 * feature i.e. INP_SUPPORTS_MBUFQ. The LRO code recoginizes
91 * this flag as well and will queue packets when it is set.
92 * There are other flags as well INP_MBUF_QUEUE_READY and
93 * INP_DONT_SACK_QUEUE. The first flag tells the LRO code
94 * that we are in the pacer for output so there is no
95 * need to wake up the hpts system to get immediate
96 * input. The second tells the LRO code that its okay
97 * if a SACK arrives you can still defer input and let
98 * the current hpts timer run (this is usually set when
99 * a rack timer is up so we know SACK's are happening
100 * on the connection already and don't want to wakeup yet).
102 * There is a common functions within the rack_bbr_common code
103 * version i.e. ctf_do_queued_segments(). This function
104 * knows how to take the input queue of packets from
105 * tp->t_in_pkts and process them digging out
106 * all the arguments, calling any bpf tap and
107 * calling into tfb_do_segment_nounlock(). The common
108 * function (ctf_do_queued_segments()) requires that
109 * you have defined the tfb_do_segment_nounlock() as
112 * The second feature of the input side of hpts is the
113 * dropping of a connection. This is due to the way that
114 * locking may have occured on the INP_WLOCK. So if
115 * a stack wants to drop a connection it calls:
117 * tcp_set_inp_to_drop(tp, ETIMEDOUT)
119 * To schedule the tcp_hpts system to call
121 * tcp_drop(tp, drop_reason)
123 * at a future point. This is quite handy to prevent locking
124 * issues when dropping connections.
128 #include <sys/param.h>
130 #include <sys/interrupt.h>
131 #include <sys/module.h>
132 #include <sys/kernel.h>
133 #include <sys/hhook.h>
134 #include <sys/malloc.h>
135 #include <sys/mbuf.h>
136 #include <sys/proc.h> /* for proc0 declaration */
137 #include <sys/socket.h>
138 #include <sys/socketvar.h>
139 #include <sys/sysctl.h>
140 #include <sys/systm.h>
141 #include <sys/refcount.h>
142 #include <sys/sched.h>
143 #include <sys/queue.h>
145 #include <sys/counter.h>
146 #include <sys/time.h>
147 #include <sys/kthread.h>
148 #include <sys/kern_prefetch.h>
153 #include <net/route.h>
154 #include <net/vnet.h>
157 #include <net/netisr.h>
158 #include <net/rss_config.h>
161 #define TCPSTATES /* for logging */
163 #include <netinet/in.h>
164 #include <netinet/in_kdtrace.h>
165 #include <netinet/in_pcb.h>
166 #include <netinet/ip.h>
167 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
168 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
169 #include <netinet/ip_var.h>
170 #include <netinet/ip6.h>
171 #include <netinet6/in6_pcb.h>
172 #include <netinet6/ip6_var.h>
173 #include <netinet/tcp.h>
174 #include <netinet/tcp_fsm.h>
175 #include <netinet/tcp_seq.h>
176 #include <netinet/tcp_timer.h>
177 #include <netinet/tcp_var.h>
178 #include <netinet/tcpip.h>
179 #include <netinet/cc/cc.h>
180 #include <netinet/tcp_hpts.h>
181 #include <netinet/tcp_log_buf.h>
184 #include <netinet/tcp_debug.h>
185 #endif /* tcpdebug */
187 #include <netinet/tcp_offload.h>
191 MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts");
193 static int tcp_bind_threads = 1;
195 static int tcp_bind_threads = 2;
197 TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads);
199 static struct tcp_hptsi tcp_pace;
200 static int hpts_does_tp_logging = 0;
202 static void tcp_wakehpts(struct tcp_hpts_entry *p);
203 static void tcp_wakeinput(struct tcp_hpts_entry *p);
204 static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv);
205 static void tcp_hptsi(struct tcp_hpts_entry *hpts);
206 static void tcp_hpts_thread(void *ctx);
207 static void tcp_init_hptsi(void *st);
209 int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP;
210 static int32_t tcp_hpts_callout_skip_swi = 0;
212 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
213 "TCP Hpts controls");
215 #define timersub(tvp, uvp, vvp) \
217 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \
218 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \
219 if ((vvp)->tv_usec < 0) { \
221 (vvp)->tv_usec += 1000000; \
225 static int32_t tcp_hpts_precision = 120;
227 struct hpts_domain_info {
232 struct hpts_domain_info hpts_domains[MAXMEMDOM];
234 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW,
235 &tcp_hpts_precision, 120,
236 "Value for PRE() precision of callout");
238 counter_u64_t hpts_hopelessly_behind;
240 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, hopeless, CTLFLAG_RD,
241 &hpts_hopelessly_behind,
242 "Number of times hpts could not catch up and was behind hopelessly");
244 counter_u64_t hpts_loops;
246 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD,
247 &hpts_loops, "Number of times hpts had to loop to catch up");
250 counter_u64_t back_tosleep;
252 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD,
253 &back_tosleep, "Number of times hpts found no tcbs");
255 counter_u64_t combined_wheel_wrap;
257 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD,
258 &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
260 counter_u64_t wheel_wrap;
262 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, wheel_wrap, CTLFLAG_RD,
263 &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
265 static int32_t out_ts_percision = 0;
267 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW,
268 &out_ts_percision, 0,
269 "Do we use a percise timestamp for every output cts");
270 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW,
271 &hpts_does_tp_logging, 0,
272 "Do we add to any tp that has logging on pacer logs");
274 static int32_t max_pacer_loops = 10;
275 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW,
276 &max_pacer_loops, 10,
277 "What is the maximum number of times the pacer will loop trying to catch up");
279 #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2)
281 static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED;
285 sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS)
290 new = hpts_sleep_max;
291 error = sysctl_handle_int(oidp, &new, 0, req);
292 if (error == 0 && req->newptr) {
293 if ((new < (NUM_OF_HPTSI_SLOTS / 4)) ||
294 (new > HPTS_MAX_SLEEP_ALLOWED))
297 hpts_sleep_max = new;
302 SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep,
303 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
305 &sysctl_net_inet_tcp_hpts_max_sleep, "IU",
306 "Maximum time hpts will sleep");
308 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW,
309 &tcp_min_hptsi_time, 0,
310 "The minimum time the hpts must sleep before processing more slots");
312 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW,
313 &tcp_hpts_callout_skip_swi, 0,
314 "Do we have the callout call directly to the hpts?");
317 tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv,
318 int ticks_to_run, int idx)
320 union tcp_log_stackspecific log;
322 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
323 log.u_bbr.flex1 = hpts->p_nxt_slot;
324 log.u_bbr.flex2 = hpts->p_cur_slot;
325 log.u_bbr.flex3 = hpts->p_prev_slot;
326 log.u_bbr.flex4 = idx;
327 log.u_bbr.flex5 = hpts->p_curtick;
328 log.u_bbr.flex6 = hpts->p_on_queue_cnt;
329 log.u_bbr.use_lt_bw = 1;
330 log.u_bbr.inflight = ticks_to_run;
331 log.u_bbr.applimited = hpts->overidden_sleep;
332 log.u_bbr.delivered = hpts->saved_curtick;
333 log.u_bbr.timeStamp = tcp_tv_to_usectick(tv);
334 log.u_bbr.epoch = hpts->saved_curslot;
335 log.u_bbr.lt_epoch = hpts->saved_prev_slot;
336 log.u_bbr.pkts_out = hpts->p_delayed_by;
337 log.u_bbr.lost = hpts->p_hpts_sleep_time;
338 log.u_bbr.cur_del_rate = hpts->p_runningtick;
339 TCP_LOG_EVENTP(tp, NULL,
340 &tp->t_inpcb->inp_socket->so_rcv,
341 &tp->t_inpcb->inp_socket->so_snd,
347 hpts_timeout_swi(void *arg)
349 struct tcp_hpts_entry *hpts;
351 hpts = (struct tcp_hpts_entry *)arg;
352 swi_sched(hpts->ie_cookie, 0);
356 hpts_timeout_dir(void *arg)
358 tcp_hpts_thread(arg);
362 hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear)
365 if (mtx_owned(&hpts->p_mtx) == 0) {
366 /* We don't own the mutex? */
367 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
369 if (hpts->p_cpu != inp->inp_hpts_cpu) {
370 /* It is not the right cpu/mutex? */
371 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
373 if (inp->inp_in_hpts == 0) {
374 /* We are not on the hpts? */
375 panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp);
378 TAILQ_REMOVE(head, inp, inp_hpts);
379 hpts->p_on_queue_cnt--;
380 if (hpts->p_on_queue_cnt < 0) {
381 /* Count should not go negative .. */
383 panic("Hpts goes negative inp:%p hpts:%p",
386 hpts->p_on_queue_cnt = 0;
389 inp->inp_hpts_request = 0;
390 inp->inp_in_hpts = 0;
395 hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref)
398 if (mtx_owned(&hpts->p_mtx) == 0) {
399 /* We don't own the mutex? */
400 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
402 if (hpts->p_cpu != inp->inp_hpts_cpu) {
403 /* It is not the right cpu/mutex? */
404 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
406 if ((noref == 0) && (inp->inp_in_hpts == 1)) {
407 /* We are already on the hpts? */
408 panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp);
411 TAILQ_INSERT_TAIL(head, inp, inp_hpts);
412 inp->inp_in_hpts = 1;
413 hpts->p_on_queue_cnt++;
420 hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear)
423 if (mtx_owned(&hpts->p_mtx) == 0) {
424 /* We don't own the mutex? */
425 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
427 if (hpts->p_cpu != inp->inp_input_cpu) {
428 /* It is not the right cpu/mutex? */
429 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
431 if (inp->inp_in_input == 0) {
432 /* We are not on the input hpts? */
433 panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp);
436 TAILQ_REMOVE(&hpts->p_input, inp, inp_input);
437 hpts->p_on_inqueue_cnt--;
438 if (hpts->p_on_inqueue_cnt < 0) {
440 panic("Hpts in goes negative inp:%p hpts:%p",
443 hpts->p_on_inqueue_cnt = 0;
446 if (TAILQ_EMPTY(&hpts->p_input) &&
447 (hpts->p_on_inqueue_cnt != 0)) {
448 /* We should not be empty with a queue count */
449 panic("%s hpts:%p in_hpts input empty but cnt:%d",
450 __FUNCTION__, hpts, hpts->p_on_inqueue_cnt);
454 inp->inp_in_input = 0;
458 hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line)
461 if (mtx_owned(&hpts->p_mtx) == 0) {
462 /* We don't own the mutex? */
463 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
465 if (hpts->p_cpu != inp->inp_input_cpu) {
466 /* It is not the right cpu/mutex? */
467 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
469 if (inp->inp_in_input == 1) {
470 /* We are already on the input hpts? */
471 panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp);
474 TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input);
475 inp->inp_in_input = 1;
476 hpts->p_on_inqueue_cnt++;
481 tcp_wakehpts(struct tcp_hpts_entry *hpts)
483 HPTS_MTX_ASSERT(hpts);
484 if (hpts->p_hpts_wake_scheduled == 0) {
485 hpts->p_hpts_wake_scheduled = 1;
486 swi_sched(hpts->ie_cookie, 0);
491 tcp_wakeinput(struct tcp_hpts_entry *hpts)
493 HPTS_MTX_ASSERT(hpts);
494 if (hpts->p_hpts_wake_scheduled == 0) {
495 hpts->p_hpts_wake_scheduled = 1;
496 swi_sched(hpts->ie_cookie, 0);
500 struct tcp_hpts_entry *
501 tcp_cur_hpts(struct inpcb *inp)
504 struct tcp_hpts_entry *hpts;
506 hpts_num = inp->inp_hpts_cpu;
507 hpts = tcp_pace.rp_ent[hpts_num];
511 struct tcp_hpts_entry *
512 tcp_hpts_lock(struct inpcb *inp)
514 struct tcp_hpts_entry *hpts;
518 hpts_num = inp->inp_hpts_cpu;
519 hpts = tcp_pace.rp_ent[hpts_num];
521 if (mtx_owned(&hpts->p_mtx)) {
522 panic("Hpts:%p owns mtx prior-to lock line:%d",
526 mtx_lock(&hpts->p_mtx);
527 if (hpts_num != inp->inp_hpts_cpu) {
528 mtx_unlock(&hpts->p_mtx);
534 struct tcp_hpts_entry *
535 tcp_input_lock(struct inpcb *inp)
537 struct tcp_hpts_entry *hpts;
541 hpts_num = inp->inp_input_cpu;
542 hpts = tcp_pace.rp_ent[hpts_num];
544 if (mtx_owned(&hpts->p_mtx)) {
545 panic("Hpts:%p owns mtx prior-to lock line:%d",
549 mtx_lock(&hpts->p_mtx);
550 if (hpts_num != inp->inp_input_cpu) {
551 mtx_unlock(&hpts->p_mtx);
558 tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line)
562 if (inp->inp_flags2 & INP_FREED) {
564 * Need to play a special trick so that in_pcbrele_wlocked
565 * does not return 1 when it really should have returned 0.
568 inp->inp_flags2 &= ~INP_FREED;
572 #ifndef INP_REF_DEBUG
573 if (in_pcbrele_wlocked(inp)) {
575 * This should not happen. We have the inpcb referred to by
576 * the main socket (why we are called) and the hpts. It
577 * should always return 0.
579 panic("inpcb:%p release ret 1",
583 if (__in_pcbrele_wlocked(inp, line)) {
585 * This should not happen. We have the inpcb referred to by
586 * the main socket (why we are called) and the hpts. It
587 * should always return 0.
589 panic("inpcb:%p release ret 1",
594 inp->inp_flags2 |= INP_FREED;
599 tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
601 if (inp->inp_in_hpts) {
602 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1);
603 tcp_remove_hpts_ref(inp, hpts, line);
608 tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
610 HPTS_MTX_ASSERT(hpts);
611 if (inp->inp_in_input) {
612 hpts_sane_input_remove(hpts, inp, 1);
613 tcp_remove_hpts_ref(inp, hpts, line);
618 * Called normally with the INP_LOCKED but it
619 * does not matter, the hpts lock is the key
620 * but the lock order allows us to hold the
621 * INP lock and then get the hpts lock.
623 * Valid values in the flags are
624 * HPTS_REMOVE_OUTPUT - remove from the output of the hpts.
625 * HPTS_REMOVE_INPUT - remove from the input of the hpts.
626 * Note that you can use one or both values together
627 * and get two actions.
630 __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line)
632 struct tcp_hpts_entry *hpts;
634 INP_WLOCK_ASSERT(inp);
635 if (flags & HPTS_REMOVE_OUTPUT) {
636 hpts = tcp_hpts_lock(inp);
637 tcp_hpts_remove_locked_output(hpts, inp, flags, line);
638 mtx_unlock(&hpts->p_mtx);
640 if (flags & HPTS_REMOVE_INPUT) {
641 hpts = tcp_input_lock(inp);
642 tcp_hpts_remove_locked_input(hpts, inp, flags, line);
643 mtx_unlock(&hpts->p_mtx);
648 hpts_tick(uint32_t wheel_tick, uint32_t plus)
651 * Given a slot on the wheel, what slot
652 * is that plus ticks out?
654 KASSERT(wheel_tick < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_tick));
655 return ((wheel_tick + plus) % NUM_OF_HPTSI_SLOTS);
659 tick_to_wheel(uint32_t cts_in_wticks)
662 * Given a timestamp in wheel ticks (10usec inc's)
663 * map it to our limited space wheel.
665 return (cts_in_wticks % NUM_OF_HPTSI_SLOTS);
669 hpts_ticks_diff(int prev_tick, int tick_now)
672 * Given two ticks that are someplace
673 * on our wheel. How far are they apart?
675 if (tick_now > prev_tick)
676 return (tick_now - prev_tick);
677 else if (tick_now == prev_tick)
679 * Special case, same means we can go all of our
680 * wheel less one slot.
682 return (NUM_OF_HPTSI_SLOTS - 1);
684 return ((NUM_OF_HPTSI_SLOTS - prev_tick) + tick_now);
688 * Given a tick on the wheel that is the current time
689 * mapped to the wheel (wheel_tick), what is the maximum
690 * distance forward that can be obtained without
691 * wrapping past either prev_tick or running_tick
692 * depending on the htps state? Also if passed
693 * a uint32_t *, fill it with the tick location.
695 * Note if you do not give this function the current
696 * time (that you think it is) mapped to the wheel
697 * then the results will not be what you expect and
698 * could lead to invalid inserts.
700 static inline int32_t
701 max_ticks_available(struct tcp_hpts_entry *hpts, uint32_t wheel_tick, uint32_t *target_tick)
703 uint32_t dis_to_travel, end_tick, pacer_to_now, avail_on_wheel;
705 if ((hpts->p_hpts_active == 1) &&
706 (hpts->p_wheel_complete == 0)) {
707 end_tick = hpts->p_runningtick;
708 /* Back up one tick */
710 end_tick = NUM_OF_HPTSI_SLOTS - 1;
714 *target_tick = end_tick;
717 * For the case where we are
718 * not active, or we have
719 * completed the pass over
720 * the wheel, we can use the
721 * prev tick and subtract one from it. This puts us
722 * as far out as possible on the wheel.
724 end_tick = hpts->p_prev_slot;
726 end_tick = NUM_OF_HPTSI_SLOTS - 1;
730 *target_tick = end_tick;
732 * Now we have close to the full wheel left minus the
733 * time it has been since the pacer went to sleep. Note
734 * that wheel_tick, passed in, should be the current time
735 * from the perspective of the caller, mapped to the wheel.
737 if (hpts->p_prev_slot != wheel_tick)
738 dis_to_travel = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
742 * dis_to_travel in this case is the space from when the
743 * pacer stopped (p_prev_slot) and where our wheel_tick
744 * is now. To know how many slots we can put it in we
745 * subtract from the wheel size. We would not want
746 * to place something after p_prev_slot or it will
749 return (NUM_OF_HPTSI_SLOTS - dis_to_travel);
752 * So how many slots are open between p_runningtick -> p_cur_slot
753 * that is what is currently un-available for insertion. Special
754 * case when we are at the last slot, this gets 1, so that
755 * the answer to how many slots are available is all but 1.
757 if (hpts->p_runningtick == hpts->p_cur_slot)
760 dis_to_travel = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
762 * How long has the pacer been running?
764 if (hpts->p_cur_slot != wheel_tick) {
765 /* The pacer is a bit late */
766 pacer_to_now = hpts_ticks_diff(hpts->p_cur_slot, wheel_tick);
768 /* The pacer is right on time, now == pacers start time */
772 * To get the number left we can insert into we simply
773 * subract the distance the pacer has to run from how
774 * many slots there are.
776 avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel;
778 * Now how many of those we will eat due to the pacer's
779 * time (p_cur_slot) of start being behind the
780 * real time (wheel_tick)?
782 if (avail_on_wheel <= pacer_to_now) {
784 * Wheel wrap, we can't fit on the wheel, that
785 * is unusual the system must be way overloaded!
786 * Insert into the assured tick, and return special
789 counter_u64_add(combined_wheel_wrap, 1);
790 *target_tick = hpts->p_nxt_slot;
794 * We know how many slots are open
795 * on the wheel (the reverse of what
796 * is left to run. Take away the time
797 * the pacer started to now (wheel_tick)
798 * and that tells you how many slots are
799 * open that can be inserted into that won't
800 * be touched by the pacer until later.
802 return (avail_on_wheel - pacer_to_now);
807 tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref)
809 uint32_t need_wake = 0;
811 HPTS_MTX_ASSERT(hpts);
812 if (inp->inp_in_hpts == 0) {
813 /* Ok we need to set it on the hpts in the current slot */
814 inp->inp_hpts_request = 0;
815 if ((hpts->p_hpts_active == 0) ||
816 (hpts->p_wheel_complete)) {
818 * A sleeping hpts we want in next slot to run
819 * note that in this state p_prev_slot == p_cur_slot
821 inp->inp_hptsslot = hpts_tick(hpts->p_prev_slot, 1);
822 if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0))
824 } else if ((void *)inp == hpts->p_inp) {
826 * The hpts system is running and the caller
827 * was awoken by the hpts system.
828 * We can't allow you to go into the same slot we
829 * are in (we don't want a loop :-D).
831 inp->inp_hptsslot = hpts->p_nxt_slot;
833 inp->inp_hptsslot = hpts->p_runningtick;
834 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref);
837 * Activate the hpts if it is sleeping and its
840 hpts->p_direct_wake = 1;
848 __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line)
851 struct tcp_hpts_entry *hpts;
853 INP_WLOCK_ASSERT(inp);
854 hpts = tcp_hpts_lock(inp);
855 ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
856 mtx_unlock(&hpts->p_mtx);
862 check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line)
865 * Sanity checks for the pacer with invariants
868 if (inp_hptsslot >= NUM_OF_HPTSI_SLOTS)
869 panic("hpts:%p inp:%p slot:%d > max",
870 hpts, inp, inp_hptsslot);
871 if ((hpts->p_hpts_active) &&
872 (hpts->p_wheel_complete == 0)) {
874 * If the pacer is processing a arc
875 * of the wheel, we need to make
876 * sure we are not inserting within
879 int distance, yet_to_run;
881 distance = hpts_ticks_diff(hpts->p_runningtick, inp_hptsslot);
882 if (hpts->p_runningtick != hpts->p_cur_slot)
883 yet_to_run = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
885 yet_to_run = 0; /* processing last slot */
886 if (yet_to_run > distance) {
887 panic("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d",
888 hpts, inp, inp_hptsslot,
889 distance, yet_to_run,
890 hpts->p_runningtick, hpts->p_cur_slot);
897 tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, int32_t line,
898 struct hpts_diag *diag, struct timeval *tv)
900 uint32_t need_new_to = 0;
901 uint32_t wheel_cts, last_tick;
902 int32_t wheel_tick, maxticks;
903 int8_t need_wakeup = 0;
905 HPTS_MTX_ASSERT(hpts);
907 memset(diag, 0, sizeof(struct hpts_diag));
908 diag->p_hpts_active = hpts->p_hpts_active;
909 diag->p_prev_slot = hpts->p_prev_slot;
910 diag->p_runningtick = hpts->p_runningtick;
911 diag->p_nxt_slot = hpts->p_nxt_slot;
912 diag->p_cur_slot = hpts->p_cur_slot;
913 diag->p_curtick = hpts->p_curtick;
914 diag->p_lasttick = hpts->p_lasttick;
915 diag->slot_req = slot;
916 diag->p_on_min_sleep = hpts->p_on_min_sleep;
917 diag->hpts_sleep_time = hpts->p_hpts_sleep_time;
919 if (inp->inp_in_hpts == 0) {
922 tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
925 /* Get the current time relative to the wheel */
926 wheel_cts = tcp_tv_to_hptstick(tv);
927 /* Map it onto the wheel */
928 wheel_tick = tick_to_wheel(wheel_cts);
929 /* Now what's the max we can place it at? */
930 maxticks = max_ticks_available(hpts, wheel_tick, &last_tick);
932 diag->wheel_tick = wheel_tick;
933 diag->maxticks = maxticks;
934 diag->wheel_cts = wheel_cts;
937 /* The pacer is in a wheel wrap behind, yikes! */
940 * Reduce by 1 to prevent a forever loop in
941 * case something else is wrong. Note this
942 * probably does not hurt because the pacer
943 * if its true is so far behind we will be
944 * > 1second late calling anyway.
948 inp->inp_hptsslot = last_tick;
949 inp->inp_hpts_request = slot;
950 } else if (maxticks >= slot) {
951 /* It all fits on the wheel */
952 inp->inp_hpts_request = 0;
953 inp->inp_hptsslot = hpts_tick(wheel_tick, slot);
955 /* It does not fit */
956 inp->inp_hpts_request = slot - maxticks;
957 inp->inp_hptsslot = last_tick;
960 diag->slot_remaining = inp->inp_hpts_request;
961 diag->inp_hptsslot = inp->inp_hptsslot;
964 check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line);
966 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0);
967 if ((hpts->p_hpts_active == 0) &&
968 (inp->inp_hpts_request == 0) &&
969 (hpts->p_on_min_sleep == 0)) {
971 * The hpts is sleeping and not on a minimum
972 * sleep time, we need to figure out where
973 * it will wake up at and if we need to reschedule
976 uint32_t have_slept, yet_to_sleep;
978 /* Now do we need to restart the hpts's timer? */
979 have_slept = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
980 if (have_slept < hpts->p_hpts_sleep_time)
981 yet_to_sleep = hpts->p_hpts_sleep_time - have_slept;
983 /* We are over-due */
988 diag->have_slept = have_slept;
989 diag->yet_to_sleep = yet_to_sleep;
992 (yet_to_sleep > slot)) {
994 * We need to reschedule the hpts's time-out.
996 hpts->p_hpts_sleep_time = slot;
997 need_new_to = slot * HPTS_TICKS_PER_USEC;
1001 * Now how far is the hpts sleeping to? if active is 1, its
1002 * up and ticking we do nothing, otherwise we may need to
1003 * reschedule its callout if need_new_to is set from above.
1006 hpts->p_direct_wake = 1;
1009 diag->need_new_to = 0;
1010 diag->co_ret = 0xffff0000;
1012 } else if (need_new_to) {
1019 while (need_new_to > HPTS_USEC_IN_SEC) {
1021 need_new_to -= HPTS_USEC_IN_SEC;
1023 tv.tv_usec = need_new_to;
1025 if (tcp_hpts_callout_skip_swi == 0) {
1026 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1027 hpts_timeout_swi, hpts, hpts->p_cpu,
1028 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1030 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1031 hpts_timeout_dir, hpts,
1033 C_PREL(tcp_hpts_precision));
1036 diag->need_new_to = need_new_to;
1037 diag->co_ret = co_ret;
1042 panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp);
1048 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag)
1050 struct tcp_hpts_entry *hpts;
1055 * We now return the next-slot the hpts will be on, beyond its
1056 * current run (if up) or where it was when it stopped if it is
1059 INP_WLOCK_ASSERT(inp);
1060 hpts = tcp_hpts_lock(inp);
1062 tcp_hpts_insert_locked(hpts, inp, slot, line, diag, &tv);
1063 slot_on = hpts->p_nxt_slot;
1064 mtx_unlock(&hpts->p_mtx);
1069 __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){
1070 return (tcp_hpts_insert_diag(inp, slot, line, NULL));
1073 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line)
1077 HPTS_MTX_ASSERT(hpts);
1078 if (inp->inp_in_input == 0) {
1079 /* Ok we need to set it on the hpts in the current slot */
1080 hpts_sane_input_insert(hpts, inp, line);
1082 if (hpts->p_hpts_active == 0) {
1084 * Activate the hpts if it is sleeping.
1087 hpts->p_direct_wake = 1;
1088 tcp_wakeinput(hpts);
1090 } else if (hpts->p_hpts_active == 0) {
1092 hpts->p_direct_wake = 1;
1093 tcp_wakeinput(hpts);
1099 __tcp_queue_to_input(struct inpcb *inp, int line)
1101 struct tcp_hpts_entry *hpts;
1104 hpts = tcp_input_lock(inp);
1105 ret = __tcp_queue_to_input_locked(inp, hpts, line);
1106 mtx_unlock(&hpts->p_mtx);
1111 __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line)
1113 struct tcp_hpts_entry *hpts;
1116 tp = intotcpcb(inp);
1117 hpts = tcp_input_lock(tp->t_inpcb);
1118 if (inp->inp_in_input == 0) {
1119 /* Ok we need to set it on the hpts in the current slot */
1120 hpts_sane_input_insert(hpts, inp, line);
1121 if (hpts->p_hpts_active == 0) {
1123 * Activate the hpts if it is sleeping.
1125 hpts->p_direct_wake = 1;
1126 tcp_wakeinput(hpts);
1128 } else if (hpts->p_hpts_active == 0) {
1129 hpts->p_direct_wake = 1;
1130 tcp_wakeinput(hpts);
1132 inp->inp_hpts_drop_reas = reason;
1133 mtx_unlock(&hpts->p_mtx);
1137 hpts_random_cpu(struct inpcb *inp){
1139 * No flow type set distribute the load randomly.
1145 * If one has been set use it i.e. we want both in and out on the
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);
1153 /* Nothing set use a random number */
1155 cpuid = (ran & 0xffff) % mp_ncpus;
1160 hpts_cpuid(struct inpcb *inp)
1163 #if !defined(RSS) && defined(NUMA)
1164 struct hpts_domain_info *di;
1168 * If one has been set use it i.e. we want both in and out on the
1171 if (inp->inp_input_cpu_set) {
1172 return (inp->inp_input_cpu);
1173 } else if (inp->inp_hpts_cpu_set) {
1174 return (inp->inp_hpts_cpu);
1176 /* If one is set the other must be the same */
1178 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
1179 if (cpuid == NETISR_CPUID_NONE)
1180 return (hpts_random_cpu(inp));
1185 * We don't have a flowid -> cpuid mapping, so cheat and just map
1186 * unknown cpuids to curcpu. Not the best, but apparently better
1187 * than defaulting to swi 0.
1190 if (inp->inp_flowtype == M_HASHTYPE_NONE)
1191 return (hpts_random_cpu(inp));
1193 * Hash to a thread based on the flowid. If we are using numa,
1194 * then restrict the hash to the numa domain where the inp lives.
1197 if (tcp_bind_threads == 2 && inp->inp_numa_domain != M_NODOM) {
1198 di = &hpts_domains[inp->inp_numa_domain];
1199 cpuid = di->cpu[inp->inp_flowid % di->count];
1202 cpuid = inp->inp_flowid % mp_ncpus;
1209 tcp_drop_in_pkts(struct tcpcb *tp)
1218 tp->t_in_pkt = NULL;
1228 * Do NOT try to optimize the processing of inp's
1229 * by first pulling off all the inp's into a temporary
1230 * list (e.g. TAILQ_CONCAT). If you do that the subtle
1231 * interactions of switching CPU's will kill because of
1232 * problems in the linked list manipulation. Basically
1233 * you would switch cpu's with the hpts mutex locked
1234 * but then while you were processing one of the inp's
1235 * some other one that you switch will get a new
1236 * packet on the different CPU. It will insert it
1237 * on the new hpts's input list. Creating a temporary
1238 * link in the inp will not fix it either, since
1239 * the other hpts will be doing the same thing and
1240 * you will both end up using the temporary link.
1242 * You will die in an ASSERT for tailq corruption if you
1243 * run INVARIANTS or you will die horribly without
1244 * INVARIANTS in some unknown way with a corrupt linked
1248 tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv)
1252 uint16_t drop_reason;
1254 uint32_t did_prefetch = 0;
1257 HPTS_MTX_ASSERT(hpts);
1260 while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) {
1261 HPTS_MTX_ASSERT(hpts);
1262 hpts_sane_input_remove(hpts, inp, 0);
1263 if (inp->inp_input_cpu_set == 0) {
1269 drop_reason = inp->inp_hpts_drop_reas;
1270 inp->inp_in_input = 0;
1271 mtx_unlock(&hpts->p_mtx);
1274 CURVNET_SET(inp->inp_vnet);
1276 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1277 (inp->inp_flags2 & INP_FREED)) {
1280 if (in_pcbrele_wlocked(inp) == 0) {
1286 mtx_lock(&hpts->p_mtx);
1289 tp = intotcpcb(inp);
1290 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1294 /* This tcb is being destroyed for drop_reason */
1295 tcp_drop_in_pkts(tp);
1296 tp = tcp_drop(tp, drop_reason);
1300 if (in_pcbrele_wlocked(inp) == 0)
1305 mtx_lock(&hpts->p_mtx);
1310 * Setup so the next time we will move to the right
1311 * CPU. This should be a rare event. It will
1312 * sometimes happens when we are the client side
1313 * (usually not the server). Somehow tcp_output()
1314 * gets called before the tcp_do_segment() sets the
1315 * intial state. This means the r_cpu and r_hpts_cpu
1316 * is 0. We get on the hpts, and then tcp_input()
1317 * gets called setting up the r_cpu to the correct
1318 * value. The hpts goes off and sees the mis-match.
1319 * We simply correct it here and the CPU will switch
1320 * to the new hpts nextime the tcb gets added to the
1321 * the hpts (not this time) :-)
1325 if (tp->t_fb_ptr != NULL) {
1326 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1329 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1330 if (inp->inp_in_input)
1331 tcp_hpts_remove(inp, HPTS_REMOVE_INPUT);
1332 dropped = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1334 /* Re-acquire the wlock so we can release the reference */
1337 } else if (tp->t_in_pkt) {
1339 * We reach here only if we had a
1340 * stack that supported INP_SUPPORTS_MBUFQ
1341 * and then somehow switched to a stack that
1342 * does not. The packets are basically stranded
1343 * and would hang with the connection until
1344 * cleanup without this code. Its not the
1345 * best way but I know of no other way to
1346 * handle it since the stack needs functions
1347 * it does not have to handle queued packets.
1349 tcp_drop_in_pkts(tp);
1351 if (in_pcbrele_wlocked(inp) == 0)
1353 INP_UNLOCK_ASSERT(inp);
1357 mtx_lock(&hpts->p_mtx);
1363 tcp_hptsi(struct tcp_hpts_entry *hpts)
1366 struct inpcb *inp = NULL, *ninp;
1368 int32_t ticks_to_run, i, error;
1369 int32_t paced_cnt = 0;
1370 int32_t loop_cnt = 0;
1371 int32_t did_prefetch = 0;
1372 int32_t prefetch_ninp = 0;
1373 int32_t prefetch_tp = 0;
1374 int32_t wrap_loop_cnt = 0;
1377 HPTS_MTX_ASSERT(hpts);
1380 /* record previous info for any logging */
1381 hpts->saved_lasttick = hpts->p_lasttick;
1382 hpts->saved_curtick = hpts->p_curtick;
1383 hpts->saved_curslot = hpts->p_cur_slot;
1384 hpts->saved_prev_slot = hpts->p_prev_slot;
1386 hpts->p_lasttick = hpts->p_curtick;
1387 hpts->p_curtick = tcp_gethptstick(&tv);
1388 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1389 if ((hpts->p_on_queue_cnt == 0) ||
1390 (hpts->p_lasttick == hpts->p_curtick)) {
1392 * No time has yet passed,
1395 hpts->p_prev_slot = hpts->p_cur_slot;
1396 hpts->p_lasttick = hpts->p_curtick;
1400 hpts->p_wheel_complete = 0;
1401 HPTS_MTX_ASSERT(hpts);
1402 ticks_to_run = hpts_ticks_diff(hpts->p_prev_slot, hpts->p_cur_slot);
1403 if (((hpts->p_curtick - hpts->p_lasttick) > ticks_to_run) &&
1404 (hpts->p_on_queue_cnt != 0)) {
1406 * Wheel wrap is occuring, basically we
1407 * are behind and the distance between
1408 * run's has spread so much it has exceeded
1409 * the time on the wheel (1.024 seconds). This
1410 * is ugly and should NOT be happening. We
1411 * need to run the entire wheel. We last processed
1412 * p_prev_slot, so that needs to be the last slot
1413 * we run. The next slot after that should be our
1414 * reserved first slot for new, and then starts
1415 * the running postion. Now the problem is the
1416 * reserved "not to yet" place does not exist
1417 * and there may be inp's in there that need
1418 * running. We can merge those into the
1419 * first slot at the head.
1422 hpts->p_nxt_slot = hpts_tick(hpts->p_prev_slot, 1);
1423 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 2);
1425 * Adjust p_cur_slot to be where we are starting from
1426 * hopefully we will catch up (fat chance if something
1427 * is broken this bad :( )
1429 hpts->p_cur_slot = hpts->p_prev_slot;
1431 * The next slot has guys to run too, and that would
1432 * be where we would normally start, lets move them into
1433 * the next slot (p_prev_slot + 2) so that we will
1434 * run them, the extra 10usecs of late (by being
1435 * put behind) does not really matter in this situation.
1439 * To prevent a panic we need to update the inpslot to the
1440 * new location. This is safe since it takes both the
1441 * INP lock and the pacer mutex to change the inp_hptsslot.
1443 TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts) {
1444 inp->inp_hptsslot = hpts->p_runningtick;
1447 TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningtick],
1448 &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts);
1449 ticks_to_run = NUM_OF_HPTSI_SLOTS - 1;
1450 counter_u64_add(wheel_wrap, 1);
1453 * Nxt slot is always one after p_runningtick though
1454 * its not used usually unless we are doing wheel wrap.
1456 hpts->p_nxt_slot = hpts->p_prev_slot;
1457 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 1);
1460 if (TAILQ_EMPTY(&hpts->p_input) &&
1461 (hpts->p_on_inqueue_cnt != 0)) {
1462 panic("tp:%p in_hpts input empty but cnt:%d",
1463 hpts, hpts->p_on_inqueue_cnt);
1466 HPTS_MTX_ASSERT(hpts);
1467 if (hpts->p_on_queue_cnt == 0) {
1470 HPTS_MTX_ASSERT(hpts);
1471 for (i = 0; i < ticks_to_run; i++) {
1473 * Calculate our delay, if there are no extra ticks there
1474 * was not any (i.e. if ticks_to_run == 1, no delay).
1476 hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC;
1477 HPTS_MTX_ASSERT(hpts);
1478 while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1483 if (hpts->p_runningtick != inp->inp_hptsslot) {
1484 panic("Hpts:%p inp:%p slot mis-aligned %u vs %u",
1485 hpts, inp, hpts->p_runningtick, inp->inp_hptsslot);
1489 if (inp->inp_hpts_cpu_set == 0) {
1494 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningtick], 0);
1495 if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1496 /* We prefetch the next inp if possible */
1497 kern_prefetch(ninp, &prefetch_ninp);
1500 if (inp->inp_hpts_request) {
1502 * This guy is deferred out further in time
1503 * then our wheel had available on it.
1504 * Push him back on the wheel or run it
1507 uint32_t maxticks, last_tick, remaining_slots;
1509 remaining_slots = ticks_to_run - (i + 1);
1510 if (inp->inp_hpts_request > remaining_slots) {
1512 * How far out can we go?
1514 maxticks = max_ticks_available(hpts, hpts->p_cur_slot, &last_tick);
1515 if (maxticks >= inp->inp_hpts_request) {
1516 /* we can place it finally to be processed */
1517 inp->inp_hptsslot = hpts_tick(hpts->p_runningtick, inp->inp_hpts_request);
1518 inp->inp_hpts_request = 0;
1520 /* Work off some more time */
1521 inp->inp_hptsslot = last_tick;
1522 inp->inp_hpts_request-= maxticks;
1524 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], __LINE__, 1);
1528 inp->inp_hpts_request = 0;
1529 /* Fall through we will so do it now */
1532 * We clear the hpts flag here after dealing with
1533 * remaining slots. This way anyone looking with the
1534 * TCB lock will see its on the hpts until just
1537 inp->inp_in_hpts = 0;
1538 mtx_unlock(&hpts->p_mtx);
1540 if (in_pcbrele_wlocked(inp)) {
1541 mtx_lock(&hpts->p_mtx);
1545 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1546 (inp->inp_flags2 & INP_FREED)) {
1549 if (mtx_owned(&hpts->p_mtx)) {
1550 panic("Hpts:%p owns mtx prior-to lock line:%d",
1555 mtx_lock(&hpts->p_mtx);
1559 tp = intotcpcb(inp);
1560 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1565 * Setup so the next time we will move to
1566 * the right CPU. This should be a rare
1567 * event. It will sometimes happens when we
1568 * are the client side (usually not the
1569 * server). Somehow tcp_output() gets called
1570 * before the tcp_do_segment() sets the
1571 * intial state. This means the r_cpu and
1572 * r_hpts_cpu is 0. We get on the hpts, and
1573 * then tcp_input() gets called setting up
1574 * the r_cpu to the correct value. The hpts
1575 * goes off and sees the mis-match. We
1576 * simply correct it here and the CPU will
1577 * switch to the new hpts nextime the tcb
1578 * gets added to the the hpts (not this one)
1584 CURVNET_SET(inp->inp_vnet);
1586 /* Lets do any logging that we might want to */
1587 if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1588 tcp_hpts_log(hpts, tp, &tv, ticks_to_run, i);
1591 * There is a hole here, we get the refcnt on the
1592 * inp so it will still be preserved but to make
1593 * sure we can get the INP we need to hold the p_mtx
1594 * above while we pull out the tp/inp, as long as
1595 * fini gets the lock first we are assured of having
1596 * a sane INP we can lock and test.
1599 if (mtx_owned(&hpts->p_mtx)) {
1600 panic("Hpts:%p owns mtx before tcp-output:%d",
1604 if (tp->t_fb_ptr != NULL) {
1605 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1608 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1609 error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1611 /* The input killed the connection */
1615 inp->inp_hpts_calls = 1;
1616 error = tp->t_fb->tfb_tcp_output(tp);
1617 inp->inp_hpts_calls = 0;
1618 if (ninp && ninp->inp_ppcb) {
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:
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).
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
1650 kern_prefetch(ninp->inp_ppcb, &prefetch_tp);
1658 INP_UNLOCK_ASSERT(inp);
1660 if (mtx_owned(&hpts->p_mtx)) {
1661 panic("Hpts:%p owns mtx prior-to lock line:%d",
1665 mtx_lock(&hpts->p_mtx);
1668 HPTS_MTX_ASSERT(hpts);
1670 hpts->p_runningtick++;
1671 if (hpts->p_runningtick >= NUM_OF_HPTSI_SLOTS) {
1672 hpts->p_runningtick = 0;
1676 HPTS_MTX_ASSERT(hpts);
1677 hpts->p_delayed_by = 0;
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).
1683 if (TAILQ_EMPTY(&hpts->p_input) &&
1684 (hpts->p_on_inqueue_cnt != 0)) {
1685 panic("tp:%p in_hpts input empty but cnt:%d",
1686 hpts, hpts->p_on_inqueue_cnt);
1689 hpts->p_prev_slot = hpts->p_cur_slot;
1690 hpts->p_lasttick = hpts->p_curtick;
1691 if (loop_cnt > max_pacer_loops) {
1693 * Something is serious slow we have
1694 * looped through processing the wheel
1695 * and by the time we cleared the
1696 * needs to run max_pacer_loops time
1697 * we still needed to run. That means
1698 * the system is hopelessly behind and
1699 * can never catch up :(
1701 * We will just lie to this thread
1702 * and let it thing p_curtick is
1703 * correct. When it next awakens
1704 * it will find itself further behind.
1706 counter_u64_add(hpts_hopelessly_behind, 1);
1709 hpts->p_curtick = tcp_gethptstick(&tv);
1710 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1711 if ((wrap_loop_cnt < 2) &&
1712 (hpts->p_lasttick != hpts->p_curtick)) {
1713 counter_u64_add(hpts_loops, 1);
1719 * Set flag to tell that we are done for
1720 * any slot input that happens during
1723 hpts->p_wheel_complete = 1;
1725 * Run any input that may be there not covered
1728 if (!TAILQ_EMPTY(&hpts->p_input)) {
1729 tcp_input_data(hpts, &tv);
1731 * Now did we spend too long running
1732 * input and need to run more ticks?
1734 KASSERT(hpts->p_prev_slot == hpts->p_cur_slot,
1735 ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts,
1736 hpts->p_prev_slot, hpts->p_cur_slot));
1737 KASSERT(hpts->p_lasttick == hpts->p_curtick,
1738 ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts,
1739 hpts->p_lasttick, hpts->p_curtick));
1740 hpts->p_curtick = tcp_gethptstick(&tv);
1741 if (hpts->p_lasttick != hpts->p_curtick) {
1742 counter_u64_add(hpts_loops, 1);
1743 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1748 uint32_t t = 0, i, fnd = 0;
1750 if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) {
1752 * Find next slot that is occupied and use that to
1753 * be the sleep time.
1755 for (i = 0, t = hpts_tick(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) {
1756 if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) {
1760 t = (t + 1) % NUM_OF_HPTSI_SLOTS;
1763 hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max);
1766 panic("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt);
1768 counter_u64_add(back_tosleep, 1);
1769 hpts->p_on_queue_cnt = 0;
1772 } else if (wrap_loop_cnt >= 2) {
1773 /* Special case handling */
1774 hpts->p_hpts_sleep_time = tcp_min_hptsi_time;
1776 /* No one on the wheel sleep for all but 400 slots or sleep max */
1778 hpts->p_hpts_sleep_time = hpts_sleep_max;
1784 __tcp_set_hpts(struct inpcb *inp, int32_t line)
1786 struct tcp_hpts_entry *hpts;
1788 INP_WLOCK_ASSERT(inp);
1789 hpts = tcp_hpts_lock(inp);
1790 if ((inp->inp_in_hpts == 0) &&
1791 (inp->inp_hpts_cpu_set == 0)) {
1792 inp->inp_hpts_cpu = hpts_cpuid(inp);
1793 inp->inp_hpts_cpu_set = 1;
1795 mtx_unlock(&hpts->p_mtx);
1796 hpts = tcp_input_lock(inp);
1797 if ((inp->inp_input_cpu_set == 0) &&
1798 (inp->inp_in_input == 0)) {
1799 inp->inp_input_cpu = hpts_cpuid(inp);
1800 inp->inp_input_cpu_set = 1;
1802 mtx_unlock(&hpts->p_mtx);
1806 tcp_hpts_delayedby(struct inpcb *inp){
1807 return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by);
1811 tcp_hpts_thread(void *ctx)
1813 struct tcp_hpts_entry *hpts;
1814 struct epoch_tracker et;
1818 hpts = (struct tcp_hpts_entry *)ctx;
1819 mtx_lock(&hpts->p_mtx);
1820 if (hpts->p_direct_wake) {
1821 /* Signaled by input */
1822 callout_stop(&hpts->co);
1825 if (callout_pending(&hpts->co) ||
1826 !callout_active(&hpts->co)) {
1827 mtx_unlock(&hpts->p_mtx);
1830 callout_deactivate(&hpts->co);
1832 hpts->p_hpts_wake_scheduled = 0;
1833 hpts->p_hpts_active = 1;
1834 NET_EPOCH_ENTER(et);
1837 HPTS_MTX_ASSERT(hpts);
1839 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1840 if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) {
1841 hpts->overidden_sleep = tv.tv_usec;
1842 tv.tv_usec = tcp_min_hptsi_time;
1843 hpts->p_on_min_sleep = 1;
1845 /* Clear the min sleep flag */
1846 hpts->overidden_sleep = 0;
1847 hpts->p_on_min_sleep = 0;
1849 hpts->p_hpts_active = 0;
1851 if (tcp_hpts_callout_skip_swi == 0) {
1852 callout_reset_sbt_on(&hpts->co, sb, 0,
1853 hpts_timeout_swi, hpts, hpts->p_cpu,
1854 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1856 callout_reset_sbt_on(&hpts->co, sb, 0,
1857 hpts_timeout_dir, hpts,
1859 C_PREL(tcp_hpts_precision));
1861 hpts->p_direct_wake = 0;
1862 mtx_unlock(&hpts->p_mtx);
1868 tcp_init_hptsi(void *st)
1870 int32_t i, j, error, bound = 0, created = 0;
1874 struct tcp_hpts_entry *hpts;
1878 uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1881 tcp_pace.rp_proc = NULL;
1882 tcp_pace.rp_num_hptss = ncpus;
1883 hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK);
1884 hpts_loops = counter_u64_alloc(M_WAITOK);
1885 back_tosleep = counter_u64_alloc(M_WAITOK);
1886 combined_wheel_wrap = counter_u64_alloc(M_WAITOK);
1887 wheel_wrap = counter_u64_alloc(M_WAITOK);
1888 sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *));
1889 tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO);
1890 asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS;
1891 for (i = 0; i < tcp_pace.rp_num_hptss; i++) {
1892 tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry),
1893 M_TCPHPTS, M_WAITOK | M_ZERO);
1894 tcp_pace.rp_ent[i]->p_hptss = malloc(asz,
1895 M_TCPHPTS, M_WAITOK);
1896 hpts = tcp_pace.rp_ent[i];
1898 * Init all the hpts structures that are not specifically
1899 * zero'd by the allocations. Also lets attach them to the
1900 * appropriate sysctl block as well.
1902 mtx_init(&hpts->p_mtx, "tcp_hpts_lck",
1903 "hpts", MTX_DEF | MTX_DUPOK);
1904 TAILQ_INIT(&hpts->p_input);
1905 for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) {
1906 TAILQ_INIT(&hpts->p_hptss[j]);
1908 sysctl_ctx_init(&hpts->hpts_ctx);
1909 sprintf(unit, "%d", i);
1910 hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx,
1911 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts),
1914 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1916 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1917 SYSCTL_CHILDREN(hpts->hpts_root),
1918 OID_AUTO, "in_qcnt", CTLFLAG_RD,
1919 &hpts->p_on_inqueue_cnt, 0,
1920 "Count TCB's awaiting input processing");
1921 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1922 SYSCTL_CHILDREN(hpts->hpts_root),
1923 OID_AUTO, "out_qcnt", CTLFLAG_RD,
1924 &hpts->p_on_queue_cnt, 0,
1925 "Count TCB's awaiting output processing");
1926 SYSCTL_ADD_U16(&hpts->hpts_ctx,
1927 SYSCTL_CHILDREN(hpts->hpts_root),
1928 OID_AUTO, "active", CTLFLAG_RD,
1929 &hpts->p_hpts_active, 0,
1930 "Is the hpts active");
1931 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1932 SYSCTL_CHILDREN(hpts->hpts_root),
1933 OID_AUTO, "curslot", CTLFLAG_RD,
1934 &hpts->p_cur_slot, 0,
1935 "What the current running pacers goal");
1936 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1937 SYSCTL_CHILDREN(hpts->hpts_root),
1938 OID_AUTO, "runtick", CTLFLAG_RD,
1939 &hpts->p_runningtick, 0,
1940 "What the running pacers current slot is");
1941 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1942 SYSCTL_CHILDREN(hpts->hpts_root),
1943 OID_AUTO, "curtick", CTLFLAG_RD,
1944 &hpts->p_curtick, 0,
1945 "What the running pacers last tick mapped to the wheel was");
1946 hpts->p_hpts_sleep_time = hpts_sleep_max;
1948 hpts->p_curtick = tcp_gethptstick(&tv);
1949 hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1950 hpts->p_cpu = 0xffff;
1951 hpts->p_nxt_slot = hpts_tick(hpts->p_cur_slot, 1);
1952 callout_init(&hpts->co, 1);
1955 /* Don't try to bind to NUMA domains if we don't have any */
1956 if (vm_ndomains == 1 && tcp_bind_threads == 2)
1957 tcp_bind_threads = 0;
1960 * Now lets start ithreads to handle the hptss.
1963 hpts = tcp_pace.rp_ent[i];
1965 error = swi_add(&hpts->ie, "hpts",
1966 tcp_hpts_thread, (void *)hpts,
1967 SWI_NET, INTR_MPSAFE, &hpts->ie_cookie);
1969 panic("Can't add hpts:%p i:%d err:%d",
1973 if (tcp_bind_threads == 1) {
1974 if (intr_event_bind(hpts->ie, i) == 0)
1976 } else if (tcp_bind_threads == 2) {
1978 domain = pc->pc_domain;
1979 CPU_COPY(&cpuset_domain[domain], &cs);
1980 if (intr_event_bind_ithread_cpuset(hpts->ie, &cs)
1983 count = hpts_domains[domain].count;
1984 hpts_domains[domain].cpu[count] = i;
1985 hpts_domains[domain].count++;
1989 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1991 if (tcp_hpts_callout_skip_swi == 0) {
1992 callout_reset_sbt_on(&hpts->co, sb, 0,
1993 hpts_timeout_swi, hpts, hpts->p_cpu,
1994 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1996 callout_reset_sbt_on(&hpts->co, sb, 0,
1997 hpts_timeout_dir, hpts,
1999 C_PREL(tcp_hpts_precision));
2003 * If we somehow have an empty domain, fall back to choosing
2004 * among all htps threads.
2006 for (i = 0; i < vm_ndomains; i++) {
2007 if (hpts_domains[i].count == 0) {
2008 tcp_bind_threads = 0;
2013 printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n",
2015 tcp_bind_threads == 2 ? "NUMA domains" : "cpus");
2018 SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL);
2019 MODULE_VERSION(tcphpts, 1);