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"
31 #include "opt_tcpdebug.h"
33 * Some notes about usage.
35 * The tcp_hpts system is designed to provide a high precision timer
36 * system for tcp. Its main purpose is to provide a mechanism for
37 * pacing packets out onto the wire. It can be used in two ways
38 * by a given TCP stack (and those two methods can be used simultaneously).
40 * First, and probably the main thing its used by Rack and BBR, it can
41 * be used to call tcp_output() of a transport stack at some time in the future.
42 * The normal way this is done is that tcp_output() of the stack schedules
43 * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The
44 * slot is the time from now that the stack wants to be called but it
45 * must be converted to tcp_hpts's notion of slot. This is done with
46 * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical
47 * call from the tcp_output() routine might look like:
49 * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550));
51 * The above would schedule tcp_ouput() to be called in 550 useconds.
52 * Note that if using this mechanism the stack will want to add near
53 * its top a check to prevent unwanted calls (from user land or the
54 * arrival of incoming ack's). So it would add something like:
56 * if (inp->inp_in_hpts)
59 * to prevent output processing until the time alotted has gone by.
60 * Of course this is a bare bones example and the stack will probably
61 * have more consideration then just the above.
63 * Now the second function (actually two functions I guess :D)
64 * the tcp_hpts system provides is the ability to either abort
65 * a connection (later) or process input on a connection.
66 * Why would you want to do this? To keep processor locality
67 * and or not have to worry about untangling any recursive
68 * locks. The input function now is hooked to the new LRO
71 * In order to use the input redirection function the
72 * tcp stack must define an input function for
73 * tfb_do_queued_segments(). This function understands
74 * how to dequeue a array of packets that were input and
75 * knows how to call the correct processing routine.
77 * Locking in this is important as well so most likely the
78 * stack will need to define the tfb_do_segment_nounlock()
79 * splitting tfb_do_segment() into two parts. The main processing
80 * part that does not unlock the INP and returns a value of 1 or 0.
81 * It returns 0 if all is well and the lock was not released. It
82 * returns 1 if we had to destroy the TCB (a reset received etc).
83 * The remains of tfb_do_segment() then become just a simple call
84 * to the tfb_do_segment_nounlock() function and check the return
85 * code and possibly unlock.
87 * The stack must also set the flag on the INP that it supports this
88 * feature i.e. INP_SUPPORTS_MBUFQ. The LRO code recoginizes
89 * this flag as well and will queue packets when it is set.
90 * There are other flags as well INP_MBUF_QUEUE_READY and
91 * INP_DONT_SACK_QUEUE. The first flag tells the LRO code
92 * that we are in the pacer for output so there is no
93 * need to wake up the hpts system to get immediate
94 * input. The second tells the LRO code that its okay
95 * if a SACK arrives you can still defer input and let
96 * the current hpts timer run (this is usually set when
97 * a rack timer is up so we know SACK's are happening
98 * on the connection already and don't want to wakeup yet).
100 * There is a common functions within the rack_bbr_common code
101 * version i.e. ctf_do_queued_segments(). This function
102 * knows how to take the input queue of packets from
103 * tp->t_in_pkts and process them digging out
104 * all the arguments, calling any bpf tap and
105 * calling into tfb_do_segment_nounlock(). The common
106 * function (ctf_do_queued_segments()) requires that
107 * you have defined the tfb_do_segment_nounlock() as
110 * The second feature of the input side of hpts is the
111 * dropping of a connection. This is due to the way that
112 * locking may have occured on the INP_WLOCK. So if
113 * a stack wants to drop a connection it calls:
115 * tcp_set_inp_to_drop(tp, ETIMEDOUT)
117 * To schedule the tcp_hpts system to call
119 * tcp_drop(tp, drop_reason)
121 * at a future point. This is quite handy to prevent locking
122 * issues when dropping connections.
126 #include <sys/param.h>
128 #include <sys/interrupt.h>
129 #include <sys/module.h>
130 #include <sys/kernel.h>
131 #include <sys/hhook.h>
132 #include <sys/malloc.h>
133 #include <sys/mbuf.h>
134 #include <sys/proc.h> /* for proc0 declaration */
135 #include <sys/socket.h>
136 #include <sys/socketvar.h>
137 #include <sys/sysctl.h>
138 #include <sys/systm.h>
139 #include <sys/refcount.h>
140 #include <sys/sched.h>
141 #include <sys/queue.h>
143 #include <sys/counter.h>
144 #include <sys/time.h>
145 #include <sys/kthread.h>
146 #include <sys/kern_prefetch.h>
151 #include <net/route.h>
152 #include <net/vnet.h>
154 #define TCPSTATES /* for logging */
156 #include <netinet/in.h>
157 #include <netinet/in_kdtrace.h>
158 #include <netinet/in_pcb.h>
159 #include <netinet/ip.h>
160 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
161 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
162 #include <netinet/ip_var.h>
163 #include <netinet/ip6.h>
164 #include <netinet6/in6_pcb.h>
165 #include <netinet6/ip6_var.h>
166 #include <netinet/tcp.h>
167 #include <netinet/tcp_fsm.h>
168 #include <netinet/tcp_seq.h>
169 #include <netinet/tcp_timer.h>
170 #include <netinet/tcp_var.h>
171 #include <netinet/tcpip.h>
172 #include <netinet/cc/cc.h>
173 #include <netinet/tcp_hpts.h>
174 #include <netinet/tcp_log_buf.h>
177 #include <netinet/tcp_debug.h>
178 #endif /* tcpdebug */
180 #include <netinet/tcp_offload.h>
185 MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts");
187 static int tcp_bind_threads = 1;
189 static int tcp_bind_threads = 2;
191 TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads);
193 static struct tcp_hptsi tcp_pace;
194 static int hpts_does_tp_logging = 0;
196 static void tcp_wakehpts(struct tcp_hpts_entry *p);
197 static void tcp_wakeinput(struct tcp_hpts_entry *p);
198 static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv);
199 static void tcp_hptsi(struct tcp_hpts_entry *hpts);
200 static void tcp_hpts_thread(void *ctx);
201 static void tcp_init_hptsi(void *st);
203 int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP;
204 static int32_t tcp_hpts_callout_skip_swi = 0;
206 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
207 "TCP Hpts controls");
209 #define timersub(tvp, uvp, vvp) \
211 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \
212 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \
213 if ((vvp)->tv_usec < 0) { \
215 (vvp)->tv_usec += 1000000; \
219 static int32_t tcp_hpts_precision = 120;
221 struct hpts_domain_info {
226 struct hpts_domain_info hpts_domains[MAXMEMDOM];
228 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW,
229 &tcp_hpts_precision, 120,
230 "Value for PRE() precision of callout");
232 counter_u64_t hpts_hopelessly_behind;
234 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, hopeless, CTLFLAG_RD,
235 &hpts_hopelessly_behind,
236 "Number of times hpts could not catch up and was behind hopelessly");
238 counter_u64_t hpts_loops;
240 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD,
241 &hpts_loops, "Number of times hpts had to loop to catch up");
244 counter_u64_t back_tosleep;
246 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD,
247 &back_tosleep, "Number of times hpts found no tcbs");
249 counter_u64_t combined_wheel_wrap;
251 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD,
252 &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
254 counter_u64_t wheel_wrap;
256 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, wheel_wrap, CTLFLAG_RD,
257 &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
259 static int32_t out_ts_percision = 0;
261 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW,
262 &out_ts_percision, 0,
263 "Do we use a percise timestamp for every output cts");
264 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW,
265 &hpts_does_tp_logging, 0,
266 "Do we add to any tp that has logging on pacer logs");
268 static int32_t max_pacer_loops = 10;
269 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW,
270 &max_pacer_loops, 10,
271 "What is the maximum number of times the pacer will loop trying to catch up");
273 #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2)
275 static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED;
279 sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS)
284 new = hpts_sleep_max;
285 error = sysctl_handle_int(oidp, &new, 0, req);
286 if (error == 0 && req->newptr) {
287 if ((new < (NUM_OF_HPTSI_SLOTS / 4)) ||
288 (new > HPTS_MAX_SLEEP_ALLOWED))
291 hpts_sleep_max = new;
296 SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep,
297 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
299 &sysctl_net_inet_tcp_hpts_max_sleep, "IU",
300 "Maximum time hpts will sleep");
302 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW,
303 &tcp_min_hptsi_time, 0,
304 "The minimum time the hpts must sleep before processing more slots");
306 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW,
307 &tcp_hpts_callout_skip_swi, 0,
308 "Do we have the callout call directly to the hpts?");
311 tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv,
312 int ticks_to_run, int idx)
314 union tcp_log_stackspecific log;
316 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
317 log.u_bbr.flex1 = hpts->p_nxt_slot;
318 log.u_bbr.flex2 = hpts->p_cur_slot;
319 log.u_bbr.flex3 = hpts->p_prev_slot;
320 log.u_bbr.flex4 = idx;
321 log.u_bbr.flex5 = hpts->p_curtick;
322 log.u_bbr.flex6 = hpts->p_on_queue_cnt;
323 log.u_bbr.use_lt_bw = 1;
324 log.u_bbr.inflight = ticks_to_run;
325 log.u_bbr.applimited = hpts->overidden_sleep;
326 log.u_bbr.delivered = hpts->saved_curtick;
327 log.u_bbr.timeStamp = tcp_tv_to_usectick(tv);
328 log.u_bbr.epoch = hpts->saved_curslot;
329 log.u_bbr.lt_epoch = hpts->saved_prev_slot;
330 log.u_bbr.pkts_out = hpts->p_delayed_by;
331 log.u_bbr.lost = hpts->p_hpts_sleep_time;
332 log.u_bbr.cur_del_rate = hpts->p_runningtick;
333 TCP_LOG_EVENTP(tp, NULL,
334 &tp->t_inpcb->inp_socket->so_rcv,
335 &tp->t_inpcb->inp_socket->so_snd,
341 hpts_timeout_swi(void *arg)
343 struct tcp_hpts_entry *hpts;
345 hpts = (struct tcp_hpts_entry *)arg;
346 swi_sched(hpts->ie_cookie, 0);
350 hpts_timeout_dir(void *arg)
352 tcp_hpts_thread(arg);
356 hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear)
359 if (mtx_owned(&hpts->p_mtx) == 0) {
360 /* We don't own the mutex? */
361 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
363 if (hpts->p_cpu != inp->inp_hpts_cpu) {
364 /* It is not the right cpu/mutex? */
365 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
367 if (inp->inp_in_hpts == 0) {
368 /* We are not on the hpts? */
369 panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp);
372 TAILQ_REMOVE(head, inp, inp_hpts);
373 hpts->p_on_queue_cnt--;
374 if (hpts->p_on_queue_cnt < 0) {
375 /* Count should not go negative .. */
377 panic("Hpts goes negative inp:%p hpts:%p",
380 hpts->p_on_queue_cnt = 0;
383 inp->inp_hpts_request = 0;
384 inp->inp_in_hpts = 0;
389 hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref)
392 if (mtx_owned(&hpts->p_mtx) == 0) {
393 /* We don't own the mutex? */
394 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
396 if (hpts->p_cpu != inp->inp_hpts_cpu) {
397 /* It is not the right cpu/mutex? */
398 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
400 if ((noref == 0) && (inp->inp_in_hpts == 1)) {
401 /* We are already on the hpts? */
402 panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp);
405 TAILQ_INSERT_TAIL(head, inp, inp_hpts);
406 inp->inp_in_hpts = 1;
407 hpts->p_on_queue_cnt++;
414 hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear)
417 if (mtx_owned(&hpts->p_mtx) == 0) {
418 /* We don't own the mutex? */
419 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
421 if (hpts->p_cpu != inp->inp_input_cpu) {
422 /* It is not the right cpu/mutex? */
423 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
425 if (inp->inp_in_input == 0) {
426 /* We are not on the input hpts? */
427 panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp);
430 TAILQ_REMOVE(&hpts->p_input, inp, inp_input);
431 hpts->p_on_inqueue_cnt--;
432 if (hpts->p_on_inqueue_cnt < 0) {
434 panic("Hpts in goes negative inp:%p hpts:%p",
437 hpts->p_on_inqueue_cnt = 0;
440 if (TAILQ_EMPTY(&hpts->p_input) &&
441 (hpts->p_on_inqueue_cnt != 0)) {
442 /* We should not be empty with a queue count */
443 panic("%s hpts:%p in_hpts input empty but cnt:%d",
444 __FUNCTION__, hpts, hpts->p_on_inqueue_cnt);
448 inp->inp_in_input = 0;
452 hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line)
455 if (mtx_owned(&hpts->p_mtx) == 0) {
456 /* We don't own the mutex? */
457 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
459 if (hpts->p_cpu != inp->inp_input_cpu) {
460 /* It is not the right cpu/mutex? */
461 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
463 if (inp->inp_in_input == 1) {
464 /* We are already on the input hpts? */
465 panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp);
468 TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input);
469 inp->inp_in_input = 1;
470 hpts->p_on_inqueue_cnt++;
475 tcp_wakehpts(struct tcp_hpts_entry *hpts)
477 HPTS_MTX_ASSERT(hpts);
478 if (hpts->p_hpts_wake_scheduled == 0) {
479 hpts->p_hpts_wake_scheduled = 1;
480 swi_sched(hpts->ie_cookie, 0);
485 tcp_wakeinput(struct tcp_hpts_entry *hpts)
487 HPTS_MTX_ASSERT(hpts);
488 if (hpts->p_hpts_wake_scheduled == 0) {
489 hpts->p_hpts_wake_scheduled = 1;
490 swi_sched(hpts->ie_cookie, 0);
494 struct tcp_hpts_entry *
495 tcp_cur_hpts(struct inpcb *inp)
498 struct tcp_hpts_entry *hpts;
500 hpts_num = inp->inp_hpts_cpu;
501 hpts = tcp_pace.rp_ent[hpts_num];
505 struct tcp_hpts_entry *
506 tcp_hpts_lock(struct inpcb *inp)
508 struct tcp_hpts_entry *hpts;
512 hpts_num = inp->inp_hpts_cpu;
513 hpts = tcp_pace.rp_ent[hpts_num];
515 if (mtx_owned(&hpts->p_mtx)) {
516 panic("Hpts:%p owns mtx prior-to lock line:%d",
520 mtx_lock(&hpts->p_mtx);
521 if (hpts_num != inp->inp_hpts_cpu) {
522 mtx_unlock(&hpts->p_mtx);
528 struct tcp_hpts_entry *
529 tcp_input_lock(struct inpcb *inp)
531 struct tcp_hpts_entry *hpts;
535 hpts_num = inp->inp_input_cpu;
536 hpts = tcp_pace.rp_ent[hpts_num];
538 if (mtx_owned(&hpts->p_mtx)) {
539 panic("Hpts:%p owns mtx prior-to lock line:%d",
543 mtx_lock(&hpts->p_mtx);
544 if (hpts_num != inp->inp_input_cpu) {
545 mtx_unlock(&hpts->p_mtx);
552 tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line)
556 if (inp->inp_flags2 & INP_FREED) {
558 * Need to play a special trick so that in_pcbrele_wlocked
559 * does not return 1 when it really should have returned 0.
562 inp->inp_flags2 &= ~INP_FREED;
566 #ifndef INP_REF_DEBUG
567 if (in_pcbrele_wlocked(inp)) {
569 * This should not happen. We have the inpcb referred to by
570 * the main socket (why we are called) and the hpts. It
571 * should always return 0.
573 panic("inpcb:%p release ret 1",
577 if (__in_pcbrele_wlocked(inp, line)) {
579 * This should not happen. We have the inpcb referred to by
580 * the main socket (why we are called) and the hpts. It
581 * should always return 0.
583 panic("inpcb:%p release ret 1",
588 inp->inp_flags2 |= INP_FREED;
593 tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
595 if (inp->inp_in_hpts) {
596 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1);
597 tcp_remove_hpts_ref(inp, hpts, line);
602 tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
604 HPTS_MTX_ASSERT(hpts);
605 if (inp->inp_in_input) {
606 hpts_sane_input_remove(hpts, inp, 1);
607 tcp_remove_hpts_ref(inp, hpts, line);
612 * Called normally with the INP_LOCKED but it
613 * does not matter, the hpts lock is the key
614 * but the lock order allows us to hold the
615 * INP lock and then get the hpts lock.
617 * Valid values in the flags are
618 * HPTS_REMOVE_OUTPUT - remove from the output of the hpts.
619 * HPTS_REMOVE_INPUT - remove from the input of the hpts.
620 * Note that you can use one or both values together
621 * and get two actions.
624 __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line)
626 struct tcp_hpts_entry *hpts;
628 INP_WLOCK_ASSERT(inp);
629 if (flags & HPTS_REMOVE_OUTPUT) {
630 hpts = tcp_hpts_lock(inp);
631 tcp_hpts_remove_locked_output(hpts, inp, flags, line);
632 mtx_unlock(&hpts->p_mtx);
634 if (flags & HPTS_REMOVE_INPUT) {
635 hpts = tcp_input_lock(inp);
636 tcp_hpts_remove_locked_input(hpts, inp, flags, line);
637 mtx_unlock(&hpts->p_mtx);
642 hpts_tick(uint32_t wheel_tick, uint32_t plus)
645 * Given a slot on the wheel, what slot
646 * is that plus ticks out?
648 KASSERT(wheel_tick < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_tick));
649 return ((wheel_tick + plus) % NUM_OF_HPTSI_SLOTS);
653 tick_to_wheel(uint32_t cts_in_wticks)
656 * Given a timestamp in wheel ticks (10usec inc's)
657 * map it to our limited space wheel.
659 return (cts_in_wticks % NUM_OF_HPTSI_SLOTS);
663 hpts_ticks_diff(int prev_tick, int tick_now)
666 * Given two ticks that are someplace
667 * on our wheel. How far are they apart?
669 if (tick_now > prev_tick)
670 return (tick_now - prev_tick);
671 else if (tick_now == prev_tick)
673 * Special case, same means we can go all of our
674 * wheel less one slot.
676 return (NUM_OF_HPTSI_SLOTS - 1);
678 return ((NUM_OF_HPTSI_SLOTS - prev_tick) + tick_now);
682 * Given a tick on the wheel that is the current time
683 * mapped to the wheel (wheel_tick), what is the maximum
684 * distance forward that can be obtained without
685 * wrapping past either prev_tick or running_tick
686 * depending on the htps state? Also if passed
687 * a uint32_t *, fill it with the tick location.
689 * Note if you do not give this function the current
690 * time (that you think it is) mapped to the wheel
691 * then the results will not be what you expect and
692 * could lead to invalid inserts.
694 static inline int32_t
695 max_ticks_available(struct tcp_hpts_entry *hpts, uint32_t wheel_tick, uint32_t *target_tick)
697 uint32_t dis_to_travel, end_tick, pacer_to_now, avail_on_wheel;
699 if ((hpts->p_hpts_active == 1) &&
700 (hpts->p_wheel_complete == 0)) {
701 end_tick = hpts->p_runningtick;
702 /* Back up one tick */
704 end_tick = NUM_OF_HPTSI_SLOTS - 1;
708 *target_tick = end_tick;
711 * For the case where we are
712 * not active, or we have
713 * completed the pass over
714 * the wheel, we can use the
715 * prev tick and subtract one from it. This puts us
716 * as far out as possible on the wheel.
718 end_tick = hpts->p_prev_slot;
720 end_tick = NUM_OF_HPTSI_SLOTS - 1;
724 *target_tick = end_tick;
726 * Now we have close to the full wheel left minus the
727 * time it has been since the pacer went to sleep. Note
728 * that wheel_tick, passed in, should be the current time
729 * from the perspective of the caller, mapped to the wheel.
731 if (hpts->p_prev_slot != wheel_tick)
732 dis_to_travel = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
736 * dis_to_travel in this case is the space from when the
737 * pacer stopped (p_prev_slot) and where our wheel_tick
738 * is now. To know how many slots we can put it in we
739 * subtract from the wheel size. We would not want
740 * to place something after p_prev_slot or it will
743 return (NUM_OF_HPTSI_SLOTS - dis_to_travel);
746 * So how many slots are open between p_runningtick -> p_cur_slot
747 * that is what is currently un-available for insertion. Special
748 * case when we are at the last slot, this gets 1, so that
749 * the answer to how many slots are available is all but 1.
751 if (hpts->p_runningtick == hpts->p_cur_slot)
754 dis_to_travel = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
756 * How long has the pacer been running?
758 if (hpts->p_cur_slot != wheel_tick) {
759 /* The pacer is a bit late */
760 pacer_to_now = hpts_ticks_diff(hpts->p_cur_slot, wheel_tick);
762 /* The pacer is right on time, now == pacers start time */
766 * To get the number left we can insert into we simply
767 * subract the distance the pacer has to run from how
768 * many slots there are.
770 avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel;
772 * Now how many of those we will eat due to the pacer's
773 * time (p_cur_slot) of start being behind the
774 * real time (wheel_tick)?
776 if (avail_on_wheel <= pacer_to_now) {
778 * Wheel wrap, we can't fit on the wheel, that
779 * is unusual the system must be way overloaded!
780 * Insert into the assured tick, and return special
783 counter_u64_add(combined_wheel_wrap, 1);
784 *target_tick = hpts->p_nxt_slot;
788 * We know how many slots are open
789 * on the wheel (the reverse of what
790 * is left to run. Take away the time
791 * the pacer started to now (wheel_tick)
792 * and that tells you how many slots are
793 * open that can be inserted into that won't
794 * be touched by the pacer until later.
796 return (avail_on_wheel - pacer_to_now);
801 tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref)
803 uint32_t need_wake = 0;
805 HPTS_MTX_ASSERT(hpts);
806 if (inp->inp_in_hpts == 0) {
807 /* Ok we need to set it on the hpts in the current slot */
808 inp->inp_hpts_request = 0;
809 if ((hpts->p_hpts_active == 0) ||
810 (hpts->p_wheel_complete)) {
812 * A sleeping hpts we want in next slot to run
813 * note that in this state p_prev_slot == p_cur_slot
815 inp->inp_hptsslot = hpts_tick(hpts->p_prev_slot, 1);
816 if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0))
818 } else if ((void *)inp == hpts->p_inp) {
820 * The hpts system is running and the caller
821 * was awoken by the hpts system.
822 * We can't allow you to go into the same slot we
823 * are in (we don't want a loop :-D).
825 inp->inp_hptsslot = hpts->p_nxt_slot;
827 inp->inp_hptsslot = hpts->p_runningtick;
828 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref);
831 * Activate the hpts if it is sleeping and its
834 hpts->p_direct_wake = 1;
842 __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line)
845 struct tcp_hpts_entry *hpts;
847 INP_WLOCK_ASSERT(inp);
848 hpts = tcp_hpts_lock(inp);
849 ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
850 mtx_unlock(&hpts->p_mtx);
856 check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line)
859 * Sanity checks for the pacer with invariants
862 if (inp_hptsslot >= NUM_OF_HPTSI_SLOTS)
863 panic("hpts:%p inp:%p slot:%d > max",
864 hpts, inp, inp_hptsslot);
865 if ((hpts->p_hpts_active) &&
866 (hpts->p_wheel_complete == 0)) {
868 * If the pacer is processing a arc
869 * of the wheel, we need to make
870 * sure we are not inserting within
873 int distance, yet_to_run;
875 distance = hpts_ticks_diff(hpts->p_runningtick, inp_hptsslot);
876 if (hpts->p_runningtick != hpts->p_cur_slot)
877 yet_to_run = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
879 yet_to_run = 0; /* processing last slot */
880 if (yet_to_run > distance) {
881 panic("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d",
882 hpts, inp, inp_hptsslot,
883 distance, yet_to_run,
884 hpts->p_runningtick, hpts->p_cur_slot);
891 tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, int32_t line,
892 struct hpts_diag *diag, struct timeval *tv)
894 uint32_t need_new_to = 0;
895 uint32_t wheel_cts, last_tick;
896 int32_t wheel_tick, maxticks;
897 int8_t need_wakeup = 0;
899 HPTS_MTX_ASSERT(hpts);
901 memset(diag, 0, sizeof(struct hpts_diag));
902 diag->p_hpts_active = hpts->p_hpts_active;
903 diag->p_prev_slot = hpts->p_prev_slot;
904 diag->p_runningtick = hpts->p_runningtick;
905 diag->p_nxt_slot = hpts->p_nxt_slot;
906 diag->p_cur_slot = hpts->p_cur_slot;
907 diag->p_curtick = hpts->p_curtick;
908 diag->p_lasttick = hpts->p_lasttick;
909 diag->slot_req = slot;
910 diag->p_on_min_sleep = hpts->p_on_min_sleep;
911 diag->hpts_sleep_time = hpts->p_hpts_sleep_time;
913 if (inp->inp_in_hpts == 0) {
916 tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
919 /* Get the current time relative to the wheel */
920 wheel_cts = tcp_tv_to_hptstick(tv);
921 /* Map it onto the wheel */
922 wheel_tick = tick_to_wheel(wheel_cts);
923 /* Now what's the max we can place it at? */
924 maxticks = max_ticks_available(hpts, wheel_tick, &last_tick);
926 diag->wheel_tick = wheel_tick;
927 diag->maxticks = maxticks;
928 diag->wheel_cts = wheel_cts;
931 /* The pacer is in a wheel wrap behind, yikes! */
934 * Reduce by 1 to prevent a forever loop in
935 * case something else is wrong. Note this
936 * probably does not hurt because the pacer
937 * if its true is so far behind we will be
938 * > 1second late calling anyway.
942 inp->inp_hptsslot = last_tick;
943 inp->inp_hpts_request = slot;
944 } else if (maxticks >= slot) {
945 /* It all fits on the wheel */
946 inp->inp_hpts_request = 0;
947 inp->inp_hptsslot = hpts_tick(wheel_tick, slot);
949 /* It does not fit */
950 inp->inp_hpts_request = slot - maxticks;
951 inp->inp_hptsslot = last_tick;
954 diag->slot_remaining = inp->inp_hpts_request;
955 diag->inp_hptsslot = inp->inp_hptsslot;
958 check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line);
960 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0);
961 if ((hpts->p_hpts_active == 0) &&
962 (inp->inp_hpts_request == 0) &&
963 (hpts->p_on_min_sleep == 0)) {
965 * The hpts is sleeping and not on a minimum
966 * sleep time, we need to figure out where
967 * it will wake up at and if we need to reschedule
970 uint32_t have_slept, yet_to_sleep;
972 /* Now do we need to restart the hpts's timer? */
973 have_slept = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
974 if (have_slept < hpts->p_hpts_sleep_time)
975 yet_to_sleep = hpts->p_hpts_sleep_time - have_slept;
977 /* We are over-due */
982 diag->have_slept = have_slept;
983 diag->yet_to_sleep = yet_to_sleep;
986 (yet_to_sleep > slot)) {
988 * We need to reschedule the hpts's time-out.
990 hpts->p_hpts_sleep_time = slot;
991 need_new_to = slot * HPTS_TICKS_PER_USEC;
995 * Now how far is the hpts sleeping to? if active is 1, its
996 * up and ticking we do nothing, otherwise we may need to
997 * reschedule its callout if need_new_to is set from above.
1000 hpts->p_direct_wake = 1;
1003 diag->need_new_to = 0;
1004 diag->co_ret = 0xffff0000;
1006 } else if (need_new_to) {
1013 while (need_new_to > HPTS_USEC_IN_SEC) {
1015 need_new_to -= HPTS_USEC_IN_SEC;
1017 tv.tv_usec = need_new_to;
1019 if (tcp_hpts_callout_skip_swi == 0) {
1020 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1021 hpts_timeout_swi, hpts, hpts->p_cpu,
1022 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1024 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1025 hpts_timeout_dir, hpts,
1027 C_PREL(tcp_hpts_precision));
1030 diag->need_new_to = need_new_to;
1031 diag->co_ret = co_ret;
1036 panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp);
1042 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag)
1044 struct tcp_hpts_entry *hpts;
1049 * We now return the next-slot the hpts will be on, beyond its
1050 * current run (if up) or where it was when it stopped if it is
1053 INP_WLOCK_ASSERT(inp);
1054 hpts = tcp_hpts_lock(inp);
1056 tcp_hpts_insert_locked(hpts, inp, slot, line, diag, &tv);
1057 slot_on = hpts->p_nxt_slot;
1058 mtx_unlock(&hpts->p_mtx);
1063 __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){
1064 return (tcp_hpts_insert_diag(inp, slot, line, NULL));
1067 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line)
1071 HPTS_MTX_ASSERT(hpts);
1072 if (inp->inp_in_input == 0) {
1073 /* Ok we need to set it on the hpts in the current slot */
1074 hpts_sane_input_insert(hpts, inp, line);
1076 if (hpts->p_hpts_active == 0) {
1078 * Activate the hpts if it is sleeping.
1081 hpts->p_direct_wake = 1;
1082 tcp_wakeinput(hpts);
1084 } else if (hpts->p_hpts_active == 0) {
1086 hpts->p_direct_wake = 1;
1087 tcp_wakeinput(hpts);
1093 __tcp_queue_to_input(struct inpcb *inp, int line)
1095 struct tcp_hpts_entry *hpts;
1098 hpts = tcp_input_lock(inp);
1099 ret = __tcp_queue_to_input_locked(inp, hpts, line);
1100 mtx_unlock(&hpts->p_mtx);
1105 __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line)
1107 struct tcp_hpts_entry *hpts;
1110 tp = intotcpcb(inp);
1111 hpts = tcp_input_lock(tp->t_inpcb);
1112 if (inp->inp_in_input == 0) {
1113 /* Ok we need to set it on the hpts in the current slot */
1114 hpts_sane_input_insert(hpts, inp, line);
1115 if (hpts->p_hpts_active == 0) {
1117 * Activate the hpts if it is sleeping.
1119 hpts->p_direct_wake = 1;
1120 tcp_wakeinput(hpts);
1122 } else if (hpts->p_hpts_active == 0) {
1123 hpts->p_direct_wake = 1;
1124 tcp_wakeinput(hpts);
1126 inp->inp_hpts_drop_reas = reason;
1127 mtx_unlock(&hpts->p_mtx);
1131 hpts_random_cpu(struct inpcb *inp){
1133 * No flow type set distribute the load randomly.
1139 * If one has been set use it i.e. we want both in and out on the
1142 if (inp->inp_input_cpu_set) {
1143 return (inp->inp_input_cpu);
1144 } else if (inp->inp_hpts_cpu_set) {
1145 return (inp->inp_hpts_cpu);
1147 /* Nothing set use a random number */
1149 cpuid = (ran & 0xffff) % mp_ncpus;
1154 hpts_cpuid(struct inpcb *inp){
1157 struct hpts_domain_info *di;
1161 * If one has been set use it i.e. we want both in and out on the
1164 if (inp->inp_input_cpu_set) {
1165 return (inp->inp_input_cpu);
1166 } else if (inp->inp_hpts_cpu_set) {
1167 return (inp->inp_hpts_cpu);
1169 /* If one is set the other must be the same */
1171 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
1172 if (cpuid == NETISR_CPUID_NONE)
1173 return (hpts_random_cpu(inp));
1178 * We don't have a flowid -> cpuid mapping, so cheat and just map
1179 * unknown cpuids to curcpu. Not the best, but apparently better
1180 * than defaulting to swi 0.
1183 if (inp->inp_flowtype == M_HASHTYPE_NONE)
1184 return (hpts_random_cpu(inp));
1186 * Hash to a thread based on the flowid. If we are using numa,
1187 * then restrict the hash to the numa domain where the inp lives.
1190 if (tcp_bind_threads == 2 && inp->inp_numa_domain != M_NODOM) {
1191 di = &hpts_domains[inp->inp_numa_domain];
1192 cpuid = di->cpu[inp->inp_flowid % di->count];
1195 cpuid = inp->inp_flowid % mp_ncpus;
1202 tcp_drop_in_pkts(struct tcpcb *tp)
1211 tp->t_in_pkt = NULL;
1221 * Do NOT try to optimize the processing of inp's
1222 * by first pulling off all the inp's into a temporary
1223 * list (e.g. TAILQ_CONCAT). If you do that the subtle
1224 * interactions of switching CPU's will kill because of
1225 * problems in the linked list manipulation. Basically
1226 * you would switch cpu's with the hpts mutex locked
1227 * but then while you were processing one of the inp's
1228 * some other one that you switch will get a new
1229 * packet on the different CPU. It will insert it
1230 * on the new hpts's input list. Creating a temporary
1231 * link in the inp will not fix it either, since
1232 * the other hpts will be doing the same thing and
1233 * you will both end up using the temporary link.
1235 * You will die in an ASSERT for tailq corruption if you
1236 * run INVARIANTS or you will die horribly without
1237 * INVARIANTS in some unknown way with a corrupt linked
1241 tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv)
1245 uint16_t drop_reason;
1247 uint32_t did_prefetch = 0;
1250 HPTS_MTX_ASSERT(hpts);
1253 while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) {
1254 HPTS_MTX_ASSERT(hpts);
1255 hpts_sane_input_remove(hpts, inp, 0);
1256 if (inp->inp_input_cpu_set == 0) {
1262 drop_reason = inp->inp_hpts_drop_reas;
1263 inp->inp_in_input = 0;
1264 mtx_unlock(&hpts->p_mtx);
1267 CURVNET_SET(inp->inp_vnet);
1269 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1270 (inp->inp_flags2 & INP_FREED)) {
1273 if (in_pcbrele_wlocked(inp) == 0) {
1279 mtx_lock(&hpts->p_mtx);
1282 tp = intotcpcb(inp);
1283 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1287 /* This tcb is being destroyed for drop_reason */
1288 tcp_drop_in_pkts(tp);
1289 tp = tcp_drop(tp, drop_reason);
1293 if (in_pcbrele_wlocked(inp) == 0)
1298 mtx_lock(&hpts->p_mtx);
1303 * Setup so the next time we will move to the right
1304 * CPU. This should be a rare event. It will
1305 * sometimes happens when we are the client side
1306 * (usually not the server). Somehow tcp_output()
1307 * gets called before the tcp_do_segment() sets the
1308 * intial state. This means the r_cpu and r_hpts_cpu
1309 * is 0. We get on the hpts, and then tcp_input()
1310 * gets called setting up the r_cpu to the correct
1311 * value. The hpts goes off and sees the mis-match.
1312 * We simply correct it here and the CPU will switch
1313 * to the new hpts nextime the tcb gets added to the
1314 * the hpts (not this time) :-)
1318 if (tp->t_fb_ptr != NULL) {
1319 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1322 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1323 if (inp->inp_in_input)
1324 tcp_hpts_remove(inp, HPTS_REMOVE_INPUT);
1325 dropped = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1327 /* Re-acquire the wlock so we can release the reference */
1330 } else if (tp->t_in_pkt) {
1332 * We reach here only if we had a
1333 * stack that supported INP_SUPPORTS_MBUFQ
1334 * and then somehow switched to a stack that
1335 * does not. The packets are basically stranded
1336 * and would hang with the connection until
1337 * cleanup without this code. Its not the
1338 * best way but I know of no other way to
1339 * handle it since the stack needs functions
1340 * it does not have to handle queued packets.
1342 tcp_drop_in_pkts(tp);
1344 if (in_pcbrele_wlocked(inp) == 0)
1346 INP_UNLOCK_ASSERT(inp);
1350 mtx_lock(&hpts->p_mtx);
1356 tcp_hptsi(struct tcp_hpts_entry *hpts)
1359 struct inpcb *inp = NULL, *ninp;
1361 int32_t ticks_to_run, i, error;
1362 int32_t paced_cnt = 0;
1363 int32_t loop_cnt = 0;
1364 int32_t did_prefetch = 0;
1365 int32_t prefetch_ninp = 0;
1366 int32_t prefetch_tp = 0;
1367 int32_t wrap_loop_cnt = 0;
1370 HPTS_MTX_ASSERT(hpts);
1373 /* record previous info for any logging */
1374 hpts->saved_lasttick = hpts->p_lasttick;
1375 hpts->saved_curtick = hpts->p_curtick;
1376 hpts->saved_curslot = hpts->p_cur_slot;
1377 hpts->saved_prev_slot = hpts->p_prev_slot;
1379 hpts->p_lasttick = hpts->p_curtick;
1380 hpts->p_curtick = tcp_gethptstick(&tv);
1381 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1382 if ((hpts->p_on_queue_cnt == 0) ||
1383 (hpts->p_lasttick == hpts->p_curtick)) {
1385 * No time has yet passed,
1388 hpts->p_prev_slot = hpts->p_cur_slot;
1389 hpts->p_lasttick = hpts->p_curtick;
1393 hpts->p_wheel_complete = 0;
1394 HPTS_MTX_ASSERT(hpts);
1395 ticks_to_run = hpts_ticks_diff(hpts->p_prev_slot, hpts->p_cur_slot);
1396 if (((hpts->p_curtick - hpts->p_lasttick) > ticks_to_run) &&
1397 (hpts->p_on_queue_cnt != 0)) {
1399 * Wheel wrap is occuring, basically we
1400 * are behind and the distance between
1401 * run's has spread so much it has exceeded
1402 * the time on the wheel (1.024 seconds). This
1403 * is ugly and should NOT be happening. We
1404 * need to run the entire wheel. We last processed
1405 * p_prev_slot, so that needs to be the last slot
1406 * we run. The next slot after that should be our
1407 * reserved first slot for new, and then starts
1408 * the running postion. Now the problem is the
1409 * reserved "not to yet" place does not exist
1410 * and there may be inp's in there that need
1411 * running. We can merge those into the
1412 * first slot at the head.
1415 hpts->p_nxt_slot = hpts_tick(hpts->p_prev_slot, 1);
1416 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 2);
1418 * Adjust p_cur_slot to be where we are starting from
1419 * hopefully we will catch up (fat chance if something
1420 * is broken this bad :( )
1422 hpts->p_cur_slot = hpts->p_prev_slot;
1424 * The next slot has guys to run too, and that would
1425 * be where we would normally start, lets move them into
1426 * the next slot (p_prev_slot + 2) so that we will
1427 * run them, the extra 10usecs of late (by being
1428 * put behind) does not really matter in this situation.
1432 * To prevent a panic we need to update the inpslot to the
1433 * new location. This is safe since it takes both the
1434 * INP lock and the pacer mutex to change the inp_hptsslot.
1436 TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts) {
1437 inp->inp_hptsslot = hpts->p_runningtick;
1440 TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningtick],
1441 &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts);
1442 ticks_to_run = NUM_OF_HPTSI_SLOTS - 1;
1443 counter_u64_add(wheel_wrap, 1);
1446 * Nxt slot is always one after p_runningtick though
1447 * its not used usually unless we are doing wheel wrap.
1449 hpts->p_nxt_slot = hpts->p_prev_slot;
1450 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 1);
1453 if (TAILQ_EMPTY(&hpts->p_input) &&
1454 (hpts->p_on_inqueue_cnt != 0)) {
1455 panic("tp:%p in_hpts input empty but cnt:%d",
1456 hpts, hpts->p_on_inqueue_cnt);
1459 HPTS_MTX_ASSERT(hpts);
1460 if (hpts->p_on_queue_cnt == 0) {
1463 HPTS_MTX_ASSERT(hpts);
1464 for (i = 0; i < ticks_to_run; i++) {
1466 * Calculate our delay, if there are no extra ticks there
1467 * was not any (i.e. if ticks_to_run == 1, no delay).
1469 hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC;
1470 HPTS_MTX_ASSERT(hpts);
1471 while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1476 if (hpts->p_runningtick != inp->inp_hptsslot) {
1477 panic("Hpts:%p inp:%p slot mis-aligned %u vs %u",
1478 hpts, inp, hpts->p_runningtick, inp->inp_hptsslot);
1482 if (inp->inp_hpts_cpu_set == 0) {
1487 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningtick], 0);
1488 if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1489 /* We prefetch the next inp if possible */
1490 kern_prefetch(ninp, &prefetch_ninp);
1493 if (inp->inp_hpts_request) {
1495 * This guy is deferred out further in time
1496 * then our wheel had available on it.
1497 * Push him back on the wheel or run it
1500 uint32_t maxticks, last_tick, remaining_slots;
1502 remaining_slots = ticks_to_run - (i + 1);
1503 if (inp->inp_hpts_request > remaining_slots) {
1505 * How far out can we go?
1507 maxticks = max_ticks_available(hpts, hpts->p_cur_slot, &last_tick);
1508 if (maxticks >= inp->inp_hpts_request) {
1509 /* we can place it finally to be processed */
1510 inp->inp_hptsslot = hpts_tick(hpts->p_runningtick, inp->inp_hpts_request);
1511 inp->inp_hpts_request = 0;
1513 /* Work off some more time */
1514 inp->inp_hptsslot = last_tick;
1515 inp->inp_hpts_request-= maxticks;
1517 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], __LINE__, 1);
1521 inp->inp_hpts_request = 0;
1522 /* Fall through we will so do it now */
1525 * We clear the hpts flag here after dealing with
1526 * remaining slots. This way anyone looking with the
1527 * TCB lock will see its on the hpts until just
1530 inp->inp_in_hpts = 0;
1531 mtx_unlock(&hpts->p_mtx);
1533 if (in_pcbrele_wlocked(inp)) {
1534 mtx_lock(&hpts->p_mtx);
1538 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1539 (inp->inp_flags2 & INP_FREED)) {
1542 if (mtx_owned(&hpts->p_mtx)) {
1543 panic("Hpts:%p owns mtx prior-to lock line:%d",
1548 mtx_lock(&hpts->p_mtx);
1552 tp = intotcpcb(inp);
1553 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1558 * Setup so the next time we will move to
1559 * the right CPU. This should be a rare
1560 * event. It will sometimes happens when we
1561 * are the client side (usually not the
1562 * server). Somehow tcp_output() gets called
1563 * before the tcp_do_segment() sets the
1564 * intial state. This means the r_cpu and
1565 * r_hpts_cpu is 0. We get on the hpts, and
1566 * then tcp_input() gets called setting up
1567 * the r_cpu to the correct value. The hpts
1568 * goes off and sees the mis-match. We
1569 * simply correct it here and the CPU will
1570 * switch to the new hpts nextime the tcb
1571 * gets added to the the hpts (not this one)
1577 CURVNET_SET(inp->inp_vnet);
1579 /* Lets do any logging that we might want to */
1580 if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1581 tcp_hpts_log(hpts, tp, &tv, ticks_to_run, i);
1584 * There is a hole here, we get the refcnt on the
1585 * inp so it will still be preserved but to make
1586 * sure we can get the INP we need to hold the p_mtx
1587 * above while we pull out the tp/inp, as long as
1588 * fini gets the lock first we are assured of having
1589 * a sane INP we can lock and test.
1592 if (mtx_owned(&hpts->p_mtx)) {
1593 panic("Hpts:%p owns mtx before tcp-output:%d",
1597 if (tp->t_fb_ptr != NULL) {
1598 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1601 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1602 error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1604 /* The input killed the connection */
1608 inp->inp_hpts_calls = 1;
1609 error = tp->t_fb->tfb_tcp_output(tp);
1610 inp->inp_hpts_calls = 0;
1611 if (ninp && ninp->inp_ppcb) {
1613 * If we have a nxt inp, see if we can
1614 * prefetch its ppcb. Note this may seem
1615 * "risky" since we have no locks (other
1616 * than the previous inp) and there no
1617 * assurance that ninp was not pulled while
1618 * we were processing inp and freed. If this
1619 * occured it could mean that either:
1621 * a) Its NULL (which is fine we won't go
1622 * here) <or> b) Its valid (which is cool we
1623 * will prefetch it) <or> c) The inp got
1624 * freed back to the slab which was
1625 * reallocated. Then the piece of memory was
1626 * re-used and something else (not an
1627 * address) is in inp_ppcb. If that occurs
1628 * we don't crash, but take a TLB shootdown
1629 * performance hit (same as if it was NULL
1630 * and we tried to pre-fetch it).
1632 * Considering that the likelyhood of <c> is
1633 * quite rare we will take a risk on doing
1634 * this. If performance drops after testing
1635 * we can always take this out. NB: the
1636 * kern_prefetch on amd64 actually has
1637 * protection against a bad address now via
1638 * the DMAP_() tests. This will prevent the
1639 * TLB hit, and instead if <c> occurs just
1640 * cause us to load cache with a useless
1643 kern_prefetch(ninp->inp_ppcb, &prefetch_tp);
1651 INP_UNLOCK_ASSERT(inp);
1653 if (mtx_owned(&hpts->p_mtx)) {
1654 panic("Hpts:%p owns mtx prior-to lock line:%d",
1658 mtx_lock(&hpts->p_mtx);
1661 HPTS_MTX_ASSERT(hpts);
1663 hpts->p_runningtick++;
1664 if (hpts->p_runningtick >= NUM_OF_HPTSI_SLOTS) {
1665 hpts->p_runningtick = 0;
1669 HPTS_MTX_ASSERT(hpts);
1670 hpts->p_delayed_by = 0;
1672 * Check to see if we took an excess amount of time and need to run
1673 * more ticks (if we did not hit eno-bufs).
1676 if (TAILQ_EMPTY(&hpts->p_input) &&
1677 (hpts->p_on_inqueue_cnt != 0)) {
1678 panic("tp:%p in_hpts input empty but cnt:%d",
1679 hpts, hpts->p_on_inqueue_cnt);
1682 hpts->p_prev_slot = hpts->p_cur_slot;
1683 hpts->p_lasttick = hpts->p_curtick;
1684 if (loop_cnt > max_pacer_loops) {
1686 * Something is serious slow we have
1687 * looped through processing the wheel
1688 * and by the time we cleared the
1689 * needs to run max_pacer_loops time
1690 * we still needed to run. That means
1691 * the system is hopelessly behind and
1692 * can never catch up :(
1694 * We will just lie to this thread
1695 * and let it thing p_curtick is
1696 * correct. When it next awakens
1697 * it will find itself further behind.
1699 counter_u64_add(hpts_hopelessly_behind, 1);
1702 hpts->p_curtick = tcp_gethptstick(&tv);
1703 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1704 if ((wrap_loop_cnt < 2) &&
1705 (hpts->p_lasttick != hpts->p_curtick)) {
1706 counter_u64_add(hpts_loops, 1);
1712 * Set flag to tell that we are done for
1713 * any slot input that happens during
1716 hpts->p_wheel_complete = 1;
1718 * Run any input that may be there not covered
1721 if (!TAILQ_EMPTY(&hpts->p_input)) {
1722 tcp_input_data(hpts, &tv);
1724 * Now did we spend too long running
1725 * input and need to run more ticks?
1727 KASSERT(hpts->p_prev_slot == hpts->p_cur_slot,
1728 ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts,
1729 hpts->p_prev_slot, hpts->p_cur_slot));
1730 KASSERT(hpts->p_lasttick == hpts->p_curtick,
1731 ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts,
1732 hpts->p_lasttick, hpts->p_curtick));
1733 hpts->p_curtick = tcp_gethptstick(&tv);
1734 if (hpts->p_lasttick != hpts->p_curtick) {
1735 counter_u64_add(hpts_loops, 1);
1736 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1741 uint32_t t = 0, i, fnd = 0;
1743 if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) {
1745 * Find next slot that is occupied and use that to
1746 * be the sleep time.
1748 for (i = 0, t = hpts_tick(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) {
1749 if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) {
1753 t = (t + 1) % NUM_OF_HPTSI_SLOTS;
1756 hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max);
1759 panic("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt);
1761 counter_u64_add(back_tosleep, 1);
1762 hpts->p_on_queue_cnt = 0;
1765 } else if (wrap_loop_cnt >= 2) {
1766 /* Special case handling */
1767 hpts->p_hpts_sleep_time = tcp_min_hptsi_time;
1769 /* No one on the wheel sleep for all but 400 slots or sleep max */
1771 hpts->p_hpts_sleep_time = hpts_sleep_max;
1777 __tcp_set_hpts(struct inpcb *inp, int32_t line)
1779 struct tcp_hpts_entry *hpts;
1781 INP_WLOCK_ASSERT(inp);
1782 hpts = tcp_hpts_lock(inp);
1783 if ((inp->inp_in_hpts == 0) &&
1784 (inp->inp_hpts_cpu_set == 0)) {
1785 inp->inp_hpts_cpu = hpts_cpuid(inp);
1786 inp->inp_hpts_cpu_set = 1;
1788 mtx_unlock(&hpts->p_mtx);
1789 hpts = tcp_input_lock(inp);
1790 if ((inp->inp_input_cpu_set == 0) &&
1791 (inp->inp_in_input == 0)) {
1792 inp->inp_input_cpu = hpts_cpuid(inp);
1793 inp->inp_input_cpu_set = 1;
1795 mtx_unlock(&hpts->p_mtx);
1799 tcp_hpts_delayedby(struct inpcb *inp){
1800 return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by);
1804 tcp_hpts_thread(void *ctx)
1806 struct tcp_hpts_entry *hpts;
1807 struct epoch_tracker et;
1811 hpts = (struct tcp_hpts_entry *)ctx;
1812 mtx_lock(&hpts->p_mtx);
1813 if (hpts->p_direct_wake) {
1814 /* Signaled by input */
1815 callout_stop(&hpts->co);
1818 if (callout_pending(&hpts->co) ||
1819 !callout_active(&hpts->co)) {
1820 mtx_unlock(&hpts->p_mtx);
1823 callout_deactivate(&hpts->co);
1825 hpts->p_hpts_wake_scheduled = 0;
1826 hpts->p_hpts_active = 1;
1827 NET_EPOCH_ENTER(et);
1830 HPTS_MTX_ASSERT(hpts);
1832 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1833 if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) {
1834 hpts->overidden_sleep = tv.tv_usec;
1835 tv.tv_usec = tcp_min_hptsi_time;
1836 hpts->p_on_min_sleep = 1;
1838 /* Clear the min sleep flag */
1839 hpts->overidden_sleep = 0;
1840 hpts->p_on_min_sleep = 0;
1842 hpts->p_hpts_active = 0;
1844 if (tcp_hpts_callout_skip_swi == 0) {
1845 callout_reset_sbt_on(&hpts->co, sb, 0,
1846 hpts_timeout_swi, hpts, hpts->p_cpu,
1847 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1849 callout_reset_sbt_on(&hpts->co, sb, 0,
1850 hpts_timeout_dir, hpts,
1852 C_PREL(tcp_hpts_precision));
1854 hpts->p_direct_wake = 0;
1855 mtx_unlock(&hpts->p_mtx);
1861 tcp_init_hptsi(void *st)
1863 int32_t i, j, error, bound = 0, created = 0;
1867 struct tcp_hpts_entry *hpts;
1871 uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1874 tcp_pace.rp_proc = NULL;
1875 tcp_pace.rp_num_hptss = ncpus;
1876 hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK);
1877 hpts_loops = counter_u64_alloc(M_WAITOK);
1878 back_tosleep = counter_u64_alloc(M_WAITOK);
1879 combined_wheel_wrap = counter_u64_alloc(M_WAITOK);
1880 wheel_wrap = counter_u64_alloc(M_WAITOK);
1881 sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *));
1882 tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO);
1883 asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS;
1884 for (i = 0; i < tcp_pace.rp_num_hptss; i++) {
1885 tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry),
1886 M_TCPHPTS, M_WAITOK | M_ZERO);
1887 tcp_pace.rp_ent[i]->p_hptss = malloc(asz,
1888 M_TCPHPTS, M_WAITOK);
1889 hpts = tcp_pace.rp_ent[i];
1891 * Init all the hpts structures that are not specifically
1892 * zero'd by the allocations. Also lets attach them to the
1893 * appropriate sysctl block as well.
1895 mtx_init(&hpts->p_mtx, "tcp_hpts_lck",
1896 "hpts", MTX_DEF | MTX_DUPOK);
1897 TAILQ_INIT(&hpts->p_input);
1898 for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) {
1899 TAILQ_INIT(&hpts->p_hptss[j]);
1901 sysctl_ctx_init(&hpts->hpts_ctx);
1902 sprintf(unit, "%d", i);
1903 hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx,
1904 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts),
1907 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1909 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1910 SYSCTL_CHILDREN(hpts->hpts_root),
1911 OID_AUTO, "in_qcnt", CTLFLAG_RD,
1912 &hpts->p_on_inqueue_cnt, 0,
1913 "Count TCB's awaiting input processing");
1914 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1915 SYSCTL_CHILDREN(hpts->hpts_root),
1916 OID_AUTO, "out_qcnt", CTLFLAG_RD,
1917 &hpts->p_on_queue_cnt, 0,
1918 "Count TCB's awaiting output processing");
1919 SYSCTL_ADD_U16(&hpts->hpts_ctx,
1920 SYSCTL_CHILDREN(hpts->hpts_root),
1921 OID_AUTO, "active", CTLFLAG_RD,
1922 &hpts->p_hpts_active, 0,
1923 "Is the hpts active");
1924 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1925 SYSCTL_CHILDREN(hpts->hpts_root),
1926 OID_AUTO, "curslot", CTLFLAG_RD,
1927 &hpts->p_cur_slot, 0,
1928 "What the current running pacers goal");
1929 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1930 SYSCTL_CHILDREN(hpts->hpts_root),
1931 OID_AUTO, "runtick", CTLFLAG_RD,
1932 &hpts->p_runningtick, 0,
1933 "What the running pacers current slot is");
1934 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1935 SYSCTL_CHILDREN(hpts->hpts_root),
1936 OID_AUTO, "curtick", CTLFLAG_RD,
1937 &hpts->p_curtick, 0,
1938 "What the running pacers last tick mapped to the wheel was");
1939 hpts->p_hpts_sleep_time = hpts_sleep_max;
1941 hpts->p_curtick = tcp_gethptstick(&tv);
1942 hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1943 hpts->p_cpu = 0xffff;
1944 hpts->p_nxt_slot = hpts_tick(hpts->p_cur_slot, 1);
1945 callout_init(&hpts->co, 1);
1948 /* Don't try to bind to NUMA domains if we don't have any */
1949 if (vm_ndomains == 1 && tcp_bind_threads == 2)
1950 tcp_bind_threads = 0;
1953 * Now lets start ithreads to handle the hptss.
1956 hpts = tcp_pace.rp_ent[i];
1958 error = swi_add(&hpts->ie, "hpts",
1959 tcp_hpts_thread, (void *)hpts,
1960 SWI_NET, INTR_MPSAFE, &hpts->ie_cookie);
1962 panic("Can't add hpts:%p i:%d err:%d",
1966 if (tcp_bind_threads == 1) {
1967 if (intr_event_bind(hpts->ie, i) == 0)
1969 } else if (tcp_bind_threads == 2) {
1971 domain = pc->pc_domain;
1972 CPU_COPY(&cpuset_domain[domain], &cs);
1973 if (intr_event_bind_ithread_cpuset(hpts->ie, &cs)
1976 count = hpts_domains[domain].count;
1977 hpts_domains[domain].cpu[count] = i;
1978 hpts_domains[domain].count++;
1982 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1984 if (tcp_hpts_callout_skip_swi == 0) {
1985 callout_reset_sbt_on(&hpts->co, sb, 0,
1986 hpts_timeout_swi, hpts, hpts->p_cpu,
1987 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1989 callout_reset_sbt_on(&hpts->co, sb, 0,
1990 hpts_timeout_dir, hpts,
1992 C_PREL(tcp_hpts_precision));
1996 * If we somehow have an empty domain, fall back to choosing
1997 * among all htps threads.
1999 for (i = 0; i < vm_ndomains; i++) {
2000 if (hpts_domains[i].count == 0) {
2001 tcp_bind_threads = 0;
2006 printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n",
2008 tcp_bind_threads == 2 ? "NUMA domains" : "cpus");
2011 SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL);
2012 MODULE_VERSION(tcphpts, 1);