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>
190 MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts");
192 static int tcp_bind_threads = 1;
194 static int tcp_bind_threads = 2;
196 TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads);
198 static struct tcp_hptsi tcp_pace;
199 static int hpts_does_tp_logging = 0;
201 static void tcp_wakehpts(struct tcp_hpts_entry *p);
202 static void tcp_wakeinput(struct tcp_hpts_entry *p);
203 static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv);
204 static void tcp_hptsi(struct tcp_hpts_entry *hpts);
205 static void tcp_hpts_thread(void *ctx);
206 static void tcp_init_hptsi(void *st);
208 int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP;
209 static int32_t tcp_hpts_callout_skip_swi = 0;
211 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
212 "TCP Hpts controls");
214 #define timersub(tvp, uvp, vvp) \
216 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \
217 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \
218 if ((vvp)->tv_usec < 0) { \
220 (vvp)->tv_usec += 1000000; \
224 static int32_t tcp_hpts_precision = 120;
226 struct hpts_domain_info {
231 struct hpts_domain_info hpts_domains[MAXMEMDOM];
233 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW,
234 &tcp_hpts_precision, 120,
235 "Value for PRE() precision of callout");
237 counter_u64_t hpts_hopelessly_behind;
239 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, hopeless, CTLFLAG_RD,
240 &hpts_hopelessly_behind,
241 "Number of times hpts could not catch up and was behind hopelessly");
243 counter_u64_t hpts_loops;
245 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD,
246 &hpts_loops, "Number of times hpts had to loop to catch up");
248 counter_u64_t back_tosleep;
250 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD,
251 &back_tosleep, "Number of times hpts found no tcbs");
253 counter_u64_t combined_wheel_wrap;
255 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD,
256 &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
258 counter_u64_t wheel_wrap;
260 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, wheel_wrap, CTLFLAG_RD,
261 &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
263 static int32_t out_ts_percision = 0;
265 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW,
266 &out_ts_percision, 0,
267 "Do we use a percise timestamp for every output cts");
268 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW,
269 &hpts_does_tp_logging, 0,
270 "Do we add to any tp that has logging on pacer logs");
272 static int32_t max_pacer_loops = 10;
273 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW,
274 &max_pacer_loops, 10,
275 "What is the maximum number of times the pacer will loop trying to catch up");
277 #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2)
279 static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED;
282 sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS)
287 new = hpts_sleep_max;
288 error = sysctl_handle_int(oidp, &new, 0, req);
289 if (error == 0 && req->newptr) {
290 if ((new < (NUM_OF_HPTSI_SLOTS / 4)) ||
291 (new > HPTS_MAX_SLEEP_ALLOWED))
294 hpts_sleep_max = new;
299 SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep,
300 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
302 &sysctl_net_inet_tcp_hpts_max_sleep, "IU",
303 "Maximum time hpts will sleep");
305 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW,
306 &tcp_min_hptsi_time, 0,
307 "The minimum time the hpts must sleep before processing more slots");
309 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW,
310 &tcp_hpts_callout_skip_swi, 0,
311 "Do we have the callout call directly to the hpts?");
314 tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv,
315 int ticks_to_run, int idx)
317 union tcp_log_stackspecific log;
319 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
320 log.u_bbr.flex1 = hpts->p_nxt_slot;
321 log.u_bbr.flex2 = hpts->p_cur_slot;
322 log.u_bbr.flex3 = hpts->p_prev_slot;
323 log.u_bbr.flex4 = idx;
324 log.u_bbr.flex5 = hpts->p_curtick;
325 log.u_bbr.flex6 = hpts->p_on_queue_cnt;
326 log.u_bbr.use_lt_bw = 1;
327 log.u_bbr.inflight = ticks_to_run;
328 log.u_bbr.applimited = hpts->overidden_sleep;
329 log.u_bbr.delivered = hpts->saved_curtick;
330 log.u_bbr.timeStamp = tcp_tv_to_usectick(tv);
331 log.u_bbr.epoch = hpts->saved_curslot;
332 log.u_bbr.lt_epoch = hpts->saved_prev_slot;
333 log.u_bbr.pkts_out = hpts->p_delayed_by;
334 log.u_bbr.lost = hpts->p_hpts_sleep_time;
335 log.u_bbr.cur_del_rate = hpts->p_runningtick;
336 TCP_LOG_EVENTP(tp, NULL,
337 &tp->t_inpcb->inp_socket->so_rcv,
338 &tp->t_inpcb->inp_socket->so_snd,
344 hpts_timeout_swi(void *arg)
346 struct tcp_hpts_entry *hpts;
348 hpts = (struct tcp_hpts_entry *)arg;
349 swi_sched(hpts->ie_cookie, 0);
353 hpts_timeout_dir(void *arg)
355 tcp_hpts_thread(arg);
359 hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear)
362 if (mtx_owned(&hpts->p_mtx) == 0) {
363 /* We don't own the mutex? */
364 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
366 if (hpts->p_cpu != inp->inp_hpts_cpu) {
367 /* It is not the right cpu/mutex? */
368 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
370 if (inp->inp_in_hpts == 0) {
371 /* We are not on the hpts? */
372 panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp);
375 TAILQ_REMOVE(head, inp, inp_hpts);
376 hpts->p_on_queue_cnt--;
377 if (hpts->p_on_queue_cnt < 0) {
378 /* Count should not go negative .. */
380 panic("Hpts goes negative inp:%p hpts:%p",
383 hpts->p_on_queue_cnt = 0;
386 inp->inp_hpts_request = 0;
387 inp->inp_in_hpts = 0;
392 hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref)
395 if (mtx_owned(&hpts->p_mtx) == 0) {
396 /* We don't own the mutex? */
397 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
399 if (hpts->p_cpu != inp->inp_hpts_cpu) {
400 /* It is not the right cpu/mutex? */
401 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
403 if ((noref == 0) && (inp->inp_in_hpts == 1)) {
404 /* We are already on the hpts? */
405 panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp);
408 TAILQ_INSERT_TAIL(head, inp, inp_hpts);
409 inp->inp_in_hpts = 1;
410 hpts->p_on_queue_cnt++;
417 hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear)
420 if (mtx_owned(&hpts->p_mtx) == 0) {
421 /* We don't own the mutex? */
422 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
424 if (hpts->p_cpu != inp->inp_input_cpu) {
425 /* It is not the right cpu/mutex? */
426 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
428 if (inp->inp_in_input == 0) {
429 /* We are not on the input hpts? */
430 panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp);
433 TAILQ_REMOVE(&hpts->p_input, inp, inp_input);
434 hpts->p_on_inqueue_cnt--;
435 if (hpts->p_on_inqueue_cnt < 0) {
437 panic("Hpts in goes negative inp:%p hpts:%p",
440 hpts->p_on_inqueue_cnt = 0;
443 if (TAILQ_EMPTY(&hpts->p_input) &&
444 (hpts->p_on_inqueue_cnt != 0)) {
445 /* We should not be empty with a queue count */
446 panic("%s hpts:%p in_hpts input empty but cnt:%d",
447 __FUNCTION__, hpts, hpts->p_on_inqueue_cnt);
451 inp->inp_in_input = 0;
455 hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line)
458 if (mtx_owned(&hpts->p_mtx) == 0) {
459 /* We don't own the mutex? */
460 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
462 if (hpts->p_cpu != inp->inp_input_cpu) {
463 /* It is not the right cpu/mutex? */
464 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
466 if (inp->inp_in_input == 1) {
467 /* We are already on the input hpts? */
468 panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp);
471 TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input);
472 inp->inp_in_input = 1;
473 hpts->p_on_inqueue_cnt++;
478 tcp_wakehpts(struct tcp_hpts_entry *hpts)
480 HPTS_MTX_ASSERT(hpts);
481 if (hpts->p_hpts_wake_scheduled == 0) {
482 hpts->p_hpts_wake_scheduled = 1;
483 swi_sched(hpts->ie_cookie, 0);
488 tcp_wakeinput(struct tcp_hpts_entry *hpts)
490 HPTS_MTX_ASSERT(hpts);
491 if (hpts->p_hpts_wake_scheduled == 0) {
492 hpts->p_hpts_wake_scheduled = 1;
493 swi_sched(hpts->ie_cookie, 0);
497 struct tcp_hpts_entry *
498 tcp_cur_hpts(struct inpcb *inp)
501 struct tcp_hpts_entry *hpts;
503 hpts_num = inp->inp_hpts_cpu;
504 hpts = tcp_pace.rp_ent[hpts_num];
508 struct tcp_hpts_entry *
509 tcp_hpts_lock(struct inpcb *inp)
511 struct tcp_hpts_entry *hpts;
515 hpts_num = inp->inp_hpts_cpu;
516 hpts = tcp_pace.rp_ent[hpts_num];
518 if (mtx_owned(&hpts->p_mtx)) {
519 panic("Hpts:%p owns mtx prior-to lock line:%d",
523 mtx_lock(&hpts->p_mtx);
524 if (hpts_num != inp->inp_hpts_cpu) {
525 mtx_unlock(&hpts->p_mtx);
531 struct tcp_hpts_entry *
532 tcp_input_lock(struct inpcb *inp)
534 struct tcp_hpts_entry *hpts;
538 hpts_num = inp->inp_input_cpu;
539 hpts = tcp_pace.rp_ent[hpts_num];
541 if (mtx_owned(&hpts->p_mtx)) {
542 panic("Hpts:%p owns mtx prior-to lock line:%d",
546 mtx_lock(&hpts->p_mtx);
547 if (hpts_num != inp->inp_input_cpu) {
548 mtx_unlock(&hpts->p_mtx);
555 tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line)
559 if (inp->inp_flags2 & INP_FREED) {
561 * Need to play a special trick so that in_pcbrele_wlocked
562 * does not return 1 when it really should have returned 0.
565 inp->inp_flags2 &= ~INP_FREED;
569 #ifndef INP_REF_DEBUG
570 if (in_pcbrele_wlocked(inp)) {
572 * This should not happen. We have the inpcb referred to by
573 * the main socket (why we are called) and the hpts. It
574 * should always return 0.
576 panic("inpcb:%p release ret 1",
580 if (__in_pcbrele_wlocked(inp, line)) {
582 * This should not happen. We have the inpcb referred to by
583 * the main socket (why we are called) and the hpts. It
584 * should always return 0.
586 panic("inpcb:%p release ret 1",
591 inp->inp_flags2 |= INP_FREED;
596 tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
598 if (inp->inp_in_hpts) {
599 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1);
600 tcp_remove_hpts_ref(inp, hpts, line);
605 tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
607 HPTS_MTX_ASSERT(hpts);
608 if (inp->inp_in_input) {
609 hpts_sane_input_remove(hpts, inp, 1);
610 tcp_remove_hpts_ref(inp, hpts, line);
615 * Called normally with the INP_LOCKED but it
616 * does not matter, the hpts lock is the key
617 * but the lock order allows us to hold the
618 * INP lock and then get the hpts lock.
620 * Valid values in the flags are
621 * HPTS_REMOVE_OUTPUT - remove from the output of the hpts.
622 * HPTS_REMOVE_INPUT - remove from the input of the hpts.
623 * Note that you can use one or both values together
624 * and get two actions.
627 __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line)
629 struct tcp_hpts_entry *hpts;
631 INP_WLOCK_ASSERT(inp);
632 if (flags & HPTS_REMOVE_OUTPUT) {
633 hpts = tcp_hpts_lock(inp);
634 tcp_hpts_remove_locked_output(hpts, inp, flags, line);
635 mtx_unlock(&hpts->p_mtx);
637 if (flags & HPTS_REMOVE_INPUT) {
638 hpts = tcp_input_lock(inp);
639 tcp_hpts_remove_locked_input(hpts, inp, flags, line);
640 mtx_unlock(&hpts->p_mtx);
645 hpts_tick(uint32_t wheel_tick, uint32_t plus)
648 * Given a slot on the wheel, what slot
649 * is that plus ticks out?
651 KASSERT(wheel_tick < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_tick));
652 return ((wheel_tick + plus) % NUM_OF_HPTSI_SLOTS);
656 tick_to_wheel(uint32_t cts_in_wticks)
659 * Given a timestamp in wheel ticks (10usec inc's)
660 * map it to our limited space wheel.
662 return (cts_in_wticks % NUM_OF_HPTSI_SLOTS);
666 hpts_ticks_diff(int prev_tick, int tick_now)
669 * Given two ticks that are someplace
670 * on our wheel. How far are they apart?
672 if (tick_now > prev_tick)
673 return (tick_now - prev_tick);
674 else if (tick_now == prev_tick)
676 * Special case, same means we can go all of our
677 * wheel less one slot.
679 return (NUM_OF_HPTSI_SLOTS - 1);
681 return ((NUM_OF_HPTSI_SLOTS - prev_tick) + tick_now);
685 * Given a tick on the wheel that is the current time
686 * mapped to the wheel (wheel_tick), what is the maximum
687 * distance forward that can be obtained without
688 * wrapping past either prev_tick or running_tick
689 * depending on the htps state? Also if passed
690 * a uint32_t *, fill it with the tick location.
692 * Note if you do not give this function the current
693 * time (that you think it is) mapped to the wheel
694 * then the results will not be what you expect and
695 * could lead to invalid inserts.
697 static inline int32_t
698 max_ticks_available(struct tcp_hpts_entry *hpts, uint32_t wheel_tick, uint32_t *target_tick)
700 uint32_t dis_to_travel, end_tick, pacer_to_now, avail_on_wheel;
702 if ((hpts->p_hpts_active == 1) &&
703 (hpts->p_wheel_complete == 0)) {
704 end_tick = hpts->p_runningtick;
705 /* Back up one tick */
707 end_tick = NUM_OF_HPTSI_SLOTS - 1;
711 *target_tick = end_tick;
714 * For the case where we are
715 * not active, or we have
716 * completed the pass over
717 * the wheel, we can use the
718 * prev tick and subtract one from it. This puts us
719 * as far out as possible on the wheel.
721 end_tick = hpts->p_prev_slot;
723 end_tick = NUM_OF_HPTSI_SLOTS - 1;
727 *target_tick = end_tick;
729 * Now we have close to the full wheel left minus the
730 * time it has been since the pacer went to sleep. Note
731 * that wheel_tick, passed in, should be the current time
732 * from the perspective of the caller, mapped to the wheel.
734 if (hpts->p_prev_slot != wheel_tick)
735 dis_to_travel = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
739 * dis_to_travel in this case is the space from when the
740 * pacer stopped (p_prev_slot) and where our wheel_tick
741 * is now. To know how many slots we can put it in we
742 * subtract from the wheel size. We would not want
743 * to place something after p_prev_slot or it will
746 return (NUM_OF_HPTSI_SLOTS - dis_to_travel);
749 * So how many slots are open between p_runningtick -> p_cur_slot
750 * that is what is currently un-available for insertion. Special
751 * case when we are at the last slot, this gets 1, so that
752 * the answer to how many slots are available is all but 1.
754 if (hpts->p_runningtick == hpts->p_cur_slot)
757 dis_to_travel = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
759 * How long has the pacer been running?
761 if (hpts->p_cur_slot != wheel_tick) {
762 /* The pacer is a bit late */
763 pacer_to_now = hpts_ticks_diff(hpts->p_cur_slot, wheel_tick);
765 /* The pacer is right on time, now == pacers start time */
769 * To get the number left we can insert into we simply
770 * subract the distance the pacer has to run from how
771 * many slots there are.
773 avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel;
775 * Now how many of those we will eat due to the pacer's
776 * time (p_cur_slot) of start being behind the
777 * real time (wheel_tick)?
779 if (avail_on_wheel <= pacer_to_now) {
781 * Wheel wrap, we can't fit on the wheel, that
782 * is unusual the system must be way overloaded!
783 * Insert into the assured tick, and return special
786 counter_u64_add(combined_wheel_wrap, 1);
787 *target_tick = hpts->p_nxt_slot;
791 * We know how many slots are open
792 * on the wheel (the reverse of what
793 * is left to run. Take away the time
794 * the pacer started to now (wheel_tick)
795 * and that tells you how many slots are
796 * open that can be inserted into that won't
797 * be touched by the pacer until later.
799 return (avail_on_wheel - pacer_to_now);
804 tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref)
806 uint32_t need_wake = 0;
808 HPTS_MTX_ASSERT(hpts);
809 if (inp->inp_in_hpts == 0) {
810 /* Ok we need to set it on the hpts in the current slot */
811 inp->inp_hpts_request = 0;
812 if ((hpts->p_hpts_active == 0) ||
813 (hpts->p_wheel_complete)) {
815 * A sleeping hpts we want in next slot to run
816 * note that in this state p_prev_slot == p_cur_slot
818 inp->inp_hptsslot = hpts_tick(hpts->p_prev_slot, 1);
819 if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0))
821 } else if ((void *)inp == hpts->p_inp) {
823 * The hpts system is running and the caller
824 * was awoken by the hpts system.
825 * We can't allow you to go into the same slot we
826 * are in (we don't want a loop :-D).
828 inp->inp_hptsslot = hpts->p_nxt_slot;
830 inp->inp_hptsslot = hpts->p_runningtick;
831 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref);
834 * Activate the hpts if it is sleeping and its
837 hpts->p_direct_wake = 1;
845 __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line)
848 struct tcp_hpts_entry *hpts;
850 INP_WLOCK_ASSERT(inp);
851 hpts = tcp_hpts_lock(inp);
852 ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
853 mtx_unlock(&hpts->p_mtx);
859 check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line)
862 * Sanity checks for the pacer with invariants
865 if (inp_hptsslot >= NUM_OF_HPTSI_SLOTS)
866 panic("hpts:%p inp:%p slot:%d > max",
867 hpts, inp, inp_hptsslot);
868 if ((hpts->p_hpts_active) &&
869 (hpts->p_wheel_complete == 0)) {
871 * If the pacer is processing a arc
872 * of the wheel, we need to make
873 * sure we are not inserting within
876 int distance, yet_to_run;
878 distance = hpts_ticks_diff(hpts->p_runningtick, inp_hptsslot);
879 if (hpts->p_runningtick != hpts->p_cur_slot)
880 yet_to_run = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
882 yet_to_run = 0; /* processing last slot */
883 if (yet_to_run > distance) {
884 panic("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d",
885 hpts, inp, inp_hptsslot,
886 distance, yet_to_run,
887 hpts->p_runningtick, hpts->p_cur_slot);
894 tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, int32_t line,
895 struct hpts_diag *diag, struct timeval *tv)
897 uint32_t need_new_to = 0;
898 uint32_t wheel_cts, last_tick;
899 int32_t wheel_tick, maxticks;
900 int8_t need_wakeup = 0;
902 HPTS_MTX_ASSERT(hpts);
904 memset(diag, 0, sizeof(struct hpts_diag));
905 diag->p_hpts_active = hpts->p_hpts_active;
906 diag->p_prev_slot = hpts->p_prev_slot;
907 diag->p_runningtick = hpts->p_runningtick;
908 diag->p_nxt_slot = hpts->p_nxt_slot;
909 diag->p_cur_slot = hpts->p_cur_slot;
910 diag->p_curtick = hpts->p_curtick;
911 diag->p_lasttick = hpts->p_lasttick;
912 diag->slot_req = slot;
913 diag->p_on_min_sleep = hpts->p_on_min_sleep;
914 diag->hpts_sleep_time = hpts->p_hpts_sleep_time;
916 if (inp->inp_in_hpts == 0) {
919 tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
922 /* Get the current time relative to the wheel */
923 wheel_cts = tcp_tv_to_hptstick(tv);
924 /* Map it onto the wheel */
925 wheel_tick = tick_to_wheel(wheel_cts);
926 /* Now what's the max we can place it at? */
927 maxticks = max_ticks_available(hpts, wheel_tick, &last_tick);
929 diag->wheel_tick = wheel_tick;
930 diag->maxticks = maxticks;
931 diag->wheel_cts = wheel_cts;
934 /* The pacer is in a wheel wrap behind, yikes! */
937 * Reduce by 1 to prevent a forever loop in
938 * case something else is wrong. Note this
939 * probably does not hurt because the pacer
940 * if its true is so far behind we will be
941 * > 1second late calling anyway.
945 inp->inp_hptsslot = last_tick;
946 inp->inp_hpts_request = slot;
947 } else if (maxticks >= slot) {
948 /* It all fits on the wheel */
949 inp->inp_hpts_request = 0;
950 inp->inp_hptsslot = hpts_tick(wheel_tick, slot);
952 /* It does not fit */
953 inp->inp_hpts_request = slot - maxticks;
954 inp->inp_hptsslot = last_tick;
957 diag->slot_remaining = inp->inp_hpts_request;
958 diag->inp_hptsslot = inp->inp_hptsslot;
961 check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line);
963 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0);
964 if ((hpts->p_hpts_active == 0) &&
965 (inp->inp_hpts_request == 0) &&
966 (hpts->p_on_min_sleep == 0)) {
968 * The hpts is sleeping and not on a minimum
969 * sleep time, we need to figure out where
970 * it will wake up at and if we need to reschedule
973 uint32_t have_slept, yet_to_sleep;
975 /* Now do we need to restart the hpts's timer? */
976 have_slept = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
977 if (have_slept < hpts->p_hpts_sleep_time)
978 yet_to_sleep = hpts->p_hpts_sleep_time - have_slept;
980 /* We are over-due */
985 diag->have_slept = have_slept;
986 diag->yet_to_sleep = yet_to_sleep;
989 (yet_to_sleep > slot)) {
991 * We need to reschedule the hpts's time-out.
993 hpts->p_hpts_sleep_time = slot;
994 need_new_to = slot * HPTS_TICKS_PER_USEC;
998 * Now how far is the hpts sleeping to? if active is 1, its
999 * up and ticking we do nothing, otherwise we may need to
1000 * reschedule its callout if need_new_to is set from above.
1003 hpts->p_direct_wake = 1;
1006 diag->need_new_to = 0;
1007 diag->co_ret = 0xffff0000;
1009 } else if (need_new_to) {
1016 while (need_new_to > HPTS_USEC_IN_SEC) {
1018 need_new_to -= HPTS_USEC_IN_SEC;
1020 tv.tv_usec = need_new_to;
1022 if (tcp_hpts_callout_skip_swi == 0) {
1023 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1024 hpts_timeout_swi, hpts, hpts->p_cpu,
1025 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1027 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1028 hpts_timeout_dir, hpts,
1030 C_PREL(tcp_hpts_precision));
1033 diag->need_new_to = need_new_to;
1034 diag->co_ret = co_ret;
1039 panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp);
1045 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag)
1047 struct tcp_hpts_entry *hpts;
1052 * We now return the next-slot the hpts will be on, beyond its
1053 * current run (if up) or where it was when it stopped if it is
1056 INP_WLOCK_ASSERT(inp);
1057 hpts = tcp_hpts_lock(inp);
1059 tcp_hpts_insert_locked(hpts, inp, slot, line, diag, &tv);
1060 slot_on = hpts->p_nxt_slot;
1061 mtx_unlock(&hpts->p_mtx);
1066 __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){
1067 return (tcp_hpts_insert_diag(inp, slot, line, NULL));
1070 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line)
1074 HPTS_MTX_ASSERT(hpts);
1075 if (inp->inp_in_input == 0) {
1076 /* Ok we need to set it on the hpts in the current slot */
1077 hpts_sane_input_insert(hpts, inp, line);
1079 if (hpts->p_hpts_active == 0) {
1081 * Activate the hpts if it is sleeping.
1084 hpts->p_direct_wake = 1;
1085 tcp_wakeinput(hpts);
1087 } else if (hpts->p_hpts_active == 0) {
1089 hpts->p_direct_wake = 1;
1090 tcp_wakeinput(hpts);
1096 __tcp_queue_to_input(struct inpcb *inp, int line)
1098 struct tcp_hpts_entry *hpts;
1101 hpts = tcp_input_lock(inp);
1102 ret = __tcp_queue_to_input_locked(inp, hpts, line);
1103 mtx_unlock(&hpts->p_mtx);
1108 __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line)
1110 struct tcp_hpts_entry *hpts;
1113 tp = intotcpcb(inp);
1114 hpts = tcp_input_lock(tp->t_inpcb);
1115 if (inp->inp_in_input == 0) {
1116 /* Ok we need to set it on the hpts in the current slot */
1117 hpts_sane_input_insert(hpts, inp, line);
1118 if (hpts->p_hpts_active == 0) {
1120 * Activate the hpts if it is sleeping.
1122 hpts->p_direct_wake = 1;
1123 tcp_wakeinput(hpts);
1125 } else if (hpts->p_hpts_active == 0) {
1126 hpts->p_direct_wake = 1;
1127 tcp_wakeinput(hpts);
1129 inp->inp_hpts_drop_reas = reason;
1130 mtx_unlock(&hpts->p_mtx);
1134 hpts_random_cpu(struct inpcb *inp){
1136 * No flow type set distribute the load randomly.
1142 * If one has been set use it i.e. we want both in and out on the
1145 if (inp->inp_input_cpu_set) {
1146 return (inp->inp_input_cpu);
1147 } else if (inp->inp_hpts_cpu_set) {
1148 return (inp->inp_hpts_cpu);
1150 /* Nothing set use a random number */
1152 cpuid = (ran & 0xffff) % mp_ncpus;
1157 hpts_cpuid(struct inpcb *inp)
1160 #if !defined(RSS) && defined(NUMA)
1161 struct hpts_domain_info *di;
1165 * If one has been set use it i.e. we want both in and out on the
1168 if (inp->inp_input_cpu_set) {
1169 return (inp->inp_input_cpu);
1170 } else if (inp->inp_hpts_cpu_set) {
1171 return (inp->inp_hpts_cpu);
1173 /* If one is set the other must be the same */
1175 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
1176 if (cpuid == NETISR_CPUID_NONE)
1177 return (hpts_random_cpu(inp));
1182 * We don't have a flowid -> cpuid mapping, so cheat and just map
1183 * unknown cpuids to curcpu. Not the best, but apparently better
1184 * than defaulting to swi 0.
1187 if (inp->inp_flowtype == M_HASHTYPE_NONE)
1188 return (hpts_random_cpu(inp));
1190 * Hash to a thread based on the flowid. If we are using numa,
1191 * then restrict the hash to the numa domain where the inp lives.
1194 if (tcp_bind_threads == 2 && inp->inp_numa_domain != M_NODOM) {
1195 di = &hpts_domains[inp->inp_numa_domain];
1196 cpuid = di->cpu[inp->inp_flowid % di->count];
1199 cpuid = inp->inp_flowid % mp_ncpus;
1206 tcp_drop_in_pkts(struct tcpcb *tp)
1215 tp->t_in_pkt = NULL;
1225 * Do NOT try to optimize the processing of inp's
1226 * by first pulling off all the inp's into a temporary
1227 * list (e.g. TAILQ_CONCAT). If you do that the subtle
1228 * interactions of switching CPU's will kill because of
1229 * problems in the linked list manipulation. Basically
1230 * you would switch cpu's with the hpts mutex locked
1231 * but then while you were processing one of the inp's
1232 * some other one that you switch will get a new
1233 * packet on the different CPU. It will insert it
1234 * on the new hpts's input list. Creating a temporary
1235 * link in the inp will not fix it either, since
1236 * the other hpts will be doing the same thing and
1237 * you will both end up using the temporary link.
1239 * You will die in an ASSERT for tailq corruption if you
1240 * run INVARIANTS or you will die horribly without
1241 * INVARIANTS in some unknown way with a corrupt linked
1245 tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv)
1249 uint16_t drop_reason;
1251 uint32_t did_prefetch = 0;
1254 HPTS_MTX_ASSERT(hpts);
1257 while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) {
1258 HPTS_MTX_ASSERT(hpts);
1259 hpts_sane_input_remove(hpts, inp, 0);
1260 if (inp->inp_input_cpu_set == 0) {
1266 drop_reason = inp->inp_hpts_drop_reas;
1267 inp->inp_in_input = 0;
1268 mtx_unlock(&hpts->p_mtx);
1271 CURVNET_SET(inp->inp_vnet);
1273 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1274 (inp->inp_flags2 & INP_FREED)) {
1277 if (in_pcbrele_wlocked(inp) == 0) {
1283 mtx_lock(&hpts->p_mtx);
1286 tp = intotcpcb(inp);
1287 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1291 /* This tcb is being destroyed for drop_reason */
1292 tcp_drop_in_pkts(tp);
1293 tp = tcp_drop(tp, drop_reason);
1297 if (in_pcbrele_wlocked(inp) == 0)
1302 mtx_lock(&hpts->p_mtx);
1307 * Setup so the next time we will move to the right
1308 * CPU. This should be a rare event. It will
1309 * sometimes happens when we are the client side
1310 * (usually not the server). Somehow tcp_output()
1311 * gets called before the tcp_do_segment() sets the
1312 * intial state. This means the r_cpu and r_hpts_cpu
1313 * is 0. We get on the hpts, and then tcp_input()
1314 * gets called setting up the r_cpu to the correct
1315 * value. The hpts goes off and sees the mis-match.
1316 * We simply correct it here and the CPU will switch
1317 * to the new hpts nextime the tcb gets added to the
1318 * the hpts (not this time) :-)
1322 if (tp->t_fb_ptr != NULL) {
1323 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1326 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1327 if (inp->inp_in_input)
1328 tcp_hpts_remove(inp, HPTS_REMOVE_INPUT);
1329 dropped = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1331 /* Re-acquire the wlock so we can release the reference */
1334 } else if (tp->t_in_pkt) {
1336 * We reach here only if we had a
1337 * stack that supported INP_SUPPORTS_MBUFQ
1338 * and then somehow switched to a stack that
1339 * does not. The packets are basically stranded
1340 * and would hang with the connection until
1341 * cleanup without this code. Its not the
1342 * best way but I know of no other way to
1343 * handle it since the stack needs functions
1344 * it does not have to handle queued packets.
1346 tcp_drop_in_pkts(tp);
1348 if (in_pcbrele_wlocked(inp) == 0)
1350 INP_UNLOCK_ASSERT(inp);
1354 mtx_lock(&hpts->p_mtx);
1360 tcp_hptsi(struct tcp_hpts_entry *hpts)
1363 struct inpcb *inp = NULL, *ninp;
1365 int32_t ticks_to_run, i, error;
1366 int32_t paced_cnt = 0;
1367 int32_t loop_cnt = 0;
1368 int32_t did_prefetch = 0;
1369 int32_t prefetch_ninp = 0;
1370 int32_t prefetch_tp = 0;
1371 int32_t wrap_loop_cnt = 0;
1374 HPTS_MTX_ASSERT(hpts);
1377 /* record previous info for any logging */
1378 hpts->saved_lasttick = hpts->p_lasttick;
1379 hpts->saved_curtick = hpts->p_curtick;
1380 hpts->saved_curslot = hpts->p_cur_slot;
1381 hpts->saved_prev_slot = hpts->p_prev_slot;
1383 hpts->p_lasttick = hpts->p_curtick;
1384 hpts->p_curtick = tcp_gethptstick(&tv);
1385 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1386 if ((hpts->p_on_queue_cnt == 0) ||
1387 (hpts->p_lasttick == hpts->p_curtick)) {
1389 * No time has yet passed,
1392 hpts->p_prev_slot = hpts->p_cur_slot;
1393 hpts->p_lasttick = hpts->p_curtick;
1397 hpts->p_wheel_complete = 0;
1398 HPTS_MTX_ASSERT(hpts);
1399 ticks_to_run = hpts_ticks_diff(hpts->p_prev_slot, hpts->p_cur_slot);
1400 if (((hpts->p_curtick - hpts->p_lasttick) > ticks_to_run) &&
1401 (hpts->p_on_queue_cnt != 0)) {
1403 * Wheel wrap is occuring, basically we
1404 * are behind and the distance between
1405 * run's has spread so much it has exceeded
1406 * the time on the wheel (1.024 seconds). This
1407 * is ugly and should NOT be happening. We
1408 * need to run the entire wheel. We last processed
1409 * p_prev_slot, so that needs to be the last slot
1410 * we run. The next slot after that should be our
1411 * reserved first slot for new, and then starts
1412 * the running postion. Now the problem is the
1413 * reserved "not to yet" place does not exist
1414 * and there may be inp's in there that need
1415 * running. We can merge those into the
1416 * first slot at the head.
1419 hpts->p_nxt_slot = hpts_tick(hpts->p_prev_slot, 1);
1420 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 2);
1422 * Adjust p_cur_slot to be where we are starting from
1423 * hopefully we will catch up (fat chance if something
1424 * is broken this bad :( )
1426 hpts->p_cur_slot = hpts->p_prev_slot;
1428 * The next slot has guys to run too, and that would
1429 * be where we would normally start, lets move them into
1430 * the next slot (p_prev_slot + 2) so that we will
1431 * run them, the extra 10usecs of late (by being
1432 * put behind) does not really matter in this situation.
1436 * To prevent a panic we need to update the inpslot to the
1437 * new location. This is safe since it takes both the
1438 * INP lock and the pacer mutex to change the inp_hptsslot.
1440 TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts) {
1441 inp->inp_hptsslot = hpts->p_runningtick;
1444 TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningtick],
1445 &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts);
1446 ticks_to_run = NUM_OF_HPTSI_SLOTS - 1;
1447 counter_u64_add(wheel_wrap, 1);
1450 * Nxt slot is always one after p_runningtick though
1451 * its not used usually unless we are doing wheel wrap.
1453 hpts->p_nxt_slot = hpts->p_prev_slot;
1454 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 1);
1457 if (TAILQ_EMPTY(&hpts->p_input) &&
1458 (hpts->p_on_inqueue_cnt != 0)) {
1459 panic("tp:%p in_hpts input empty but cnt:%d",
1460 hpts, hpts->p_on_inqueue_cnt);
1463 HPTS_MTX_ASSERT(hpts);
1464 if (hpts->p_on_queue_cnt == 0) {
1467 HPTS_MTX_ASSERT(hpts);
1468 for (i = 0; i < ticks_to_run; i++) {
1470 * Calculate our delay, if there are no extra ticks there
1471 * was not any (i.e. if ticks_to_run == 1, no delay).
1473 hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC;
1474 HPTS_MTX_ASSERT(hpts);
1475 while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1480 if (hpts->p_runningtick != inp->inp_hptsslot) {
1481 panic("Hpts:%p inp:%p slot mis-aligned %u vs %u",
1482 hpts, inp, hpts->p_runningtick, inp->inp_hptsslot);
1486 if (inp->inp_hpts_cpu_set == 0) {
1491 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningtick], 0);
1492 if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1493 /* We prefetch the next inp if possible */
1494 kern_prefetch(ninp, &prefetch_ninp);
1497 if (inp->inp_hpts_request) {
1499 * This guy is deferred out further in time
1500 * then our wheel had available on it.
1501 * Push him back on the wheel or run it
1504 uint32_t maxticks, last_tick, remaining_slots;
1506 remaining_slots = ticks_to_run - (i + 1);
1507 if (inp->inp_hpts_request > remaining_slots) {
1509 * How far out can we go?
1511 maxticks = max_ticks_available(hpts, hpts->p_cur_slot, &last_tick);
1512 if (maxticks >= inp->inp_hpts_request) {
1513 /* we can place it finally to be processed */
1514 inp->inp_hptsslot = hpts_tick(hpts->p_runningtick, inp->inp_hpts_request);
1515 inp->inp_hpts_request = 0;
1517 /* Work off some more time */
1518 inp->inp_hptsslot = last_tick;
1519 inp->inp_hpts_request-= maxticks;
1521 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], __LINE__, 1);
1525 inp->inp_hpts_request = 0;
1526 /* Fall through we will so do it now */
1529 * We clear the hpts flag here after dealing with
1530 * remaining slots. This way anyone looking with the
1531 * TCB lock will see its on the hpts until just
1534 inp->inp_in_hpts = 0;
1535 mtx_unlock(&hpts->p_mtx);
1537 if (in_pcbrele_wlocked(inp)) {
1538 mtx_lock(&hpts->p_mtx);
1542 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1543 (inp->inp_flags2 & INP_FREED)) {
1546 if (mtx_owned(&hpts->p_mtx)) {
1547 panic("Hpts:%p owns mtx prior-to lock line:%d",
1552 mtx_lock(&hpts->p_mtx);
1556 tp = intotcpcb(inp);
1557 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1562 * Setup so the next time we will move to
1563 * the right CPU. This should be a rare
1564 * event. It will sometimes happens when we
1565 * are the client side (usually not the
1566 * server). Somehow tcp_output() gets called
1567 * before the tcp_do_segment() sets the
1568 * intial state. This means the r_cpu and
1569 * r_hpts_cpu is 0. We get on the hpts, and
1570 * then tcp_input() gets called setting up
1571 * the r_cpu to the correct value. The hpts
1572 * goes off and sees the mis-match. We
1573 * simply correct it here and the CPU will
1574 * switch to the new hpts nextime the tcb
1575 * gets added to the the hpts (not this one)
1581 CURVNET_SET(inp->inp_vnet);
1583 /* Lets do any logging that we might want to */
1584 if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1585 tcp_hpts_log(hpts, tp, &tv, ticks_to_run, i);
1588 * There is a hole here, we get the refcnt on the
1589 * inp so it will still be preserved but to make
1590 * sure we can get the INP we need to hold the p_mtx
1591 * above while we pull out the tp/inp, as long as
1592 * fini gets the lock first we are assured of having
1593 * a sane INP we can lock and test.
1596 if (mtx_owned(&hpts->p_mtx)) {
1597 panic("Hpts:%p owns mtx before tcp-output:%d",
1601 if (tp->t_fb_ptr != NULL) {
1602 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1605 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1606 error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1608 /* The input killed the connection */
1612 inp->inp_hpts_calls = 1;
1613 error = tp->t_fb->tfb_tcp_output(tp);
1614 inp->inp_hpts_calls = 0;
1615 if (ninp && ninp->inp_ppcb) {
1617 * If we have a nxt inp, see if we can
1618 * prefetch its ppcb. Note this may seem
1619 * "risky" since we have no locks (other
1620 * than the previous inp) and there no
1621 * assurance that ninp was not pulled while
1622 * we were processing inp and freed. If this
1623 * occured it could mean that either:
1625 * a) Its NULL (which is fine we won't go
1626 * here) <or> b) Its valid (which is cool we
1627 * will prefetch it) <or> c) The inp got
1628 * freed back to the slab which was
1629 * reallocated. Then the piece of memory was
1630 * re-used and something else (not an
1631 * address) is in inp_ppcb. If that occurs
1632 * we don't crash, but take a TLB shootdown
1633 * performance hit (same as if it was NULL
1634 * and we tried to pre-fetch it).
1636 * Considering that the likelyhood of <c> is
1637 * quite rare we will take a risk on doing
1638 * this. If performance drops after testing
1639 * we can always take this out. NB: the
1640 * kern_prefetch on amd64 actually has
1641 * protection against a bad address now via
1642 * the DMAP_() tests. This will prevent the
1643 * TLB hit, and instead if <c> occurs just
1644 * cause us to load cache with a useless
1647 kern_prefetch(ninp->inp_ppcb, &prefetch_tp);
1655 INP_UNLOCK_ASSERT(inp);
1657 if (mtx_owned(&hpts->p_mtx)) {
1658 panic("Hpts:%p owns mtx prior-to lock line:%d",
1662 mtx_lock(&hpts->p_mtx);
1665 HPTS_MTX_ASSERT(hpts);
1667 hpts->p_runningtick++;
1668 if (hpts->p_runningtick >= NUM_OF_HPTSI_SLOTS) {
1669 hpts->p_runningtick = 0;
1673 HPTS_MTX_ASSERT(hpts);
1674 hpts->p_delayed_by = 0;
1676 * Check to see if we took an excess amount of time and need to run
1677 * more ticks (if we did not hit eno-bufs).
1680 if (TAILQ_EMPTY(&hpts->p_input) &&
1681 (hpts->p_on_inqueue_cnt != 0)) {
1682 panic("tp:%p in_hpts input empty but cnt:%d",
1683 hpts, hpts->p_on_inqueue_cnt);
1686 hpts->p_prev_slot = hpts->p_cur_slot;
1687 hpts->p_lasttick = hpts->p_curtick;
1688 if (loop_cnt > max_pacer_loops) {
1690 * Something is serious slow we have
1691 * looped through processing the wheel
1692 * and by the time we cleared the
1693 * needs to run max_pacer_loops time
1694 * we still needed to run. That means
1695 * the system is hopelessly behind and
1696 * can never catch up :(
1698 * We will just lie to this thread
1699 * and let it thing p_curtick is
1700 * correct. When it next awakens
1701 * it will find itself further behind.
1703 counter_u64_add(hpts_hopelessly_behind, 1);
1706 hpts->p_curtick = tcp_gethptstick(&tv);
1707 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1708 if ((wrap_loop_cnt < 2) &&
1709 (hpts->p_lasttick != hpts->p_curtick)) {
1710 counter_u64_add(hpts_loops, 1);
1716 * Set flag to tell that we are done for
1717 * any slot input that happens during
1720 hpts->p_wheel_complete = 1;
1722 * Run any input that may be there not covered
1725 if (!TAILQ_EMPTY(&hpts->p_input)) {
1726 tcp_input_data(hpts, &tv);
1728 * Now did we spend too long running
1729 * input and need to run more ticks?
1731 KASSERT(hpts->p_prev_slot == hpts->p_cur_slot,
1732 ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts,
1733 hpts->p_prev_slot, hpts->p_cur_slot));
1734 KASSERT(hpts->p_lasttick == hpts->p_curtick,
1735 ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts,
1736 hpts->p_lasttick, hpts->p_curtick));
1737 hpts->p_curtick = tcp_gethptstick(&tv);
1738 if (hpts->p_lasttick != hpts->p_curtick) {
1739 counter_u64_add(hpts_loops, 1);
1740 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1745 uint32_t t = 0, i, fnd = 0;
1747 if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) {
1749 * Find next slot that is occupied and use that to
1750 * be the sleep time.
1752 for (i = 0, t = hpts_tick(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) {
1753 if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) {
1757 t = (t + 1) % NUM_OF_HPTSI_SLOTS;
1760 hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max);
1763 panic("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt);
1765 counter_u64_add(back_tosleep, 1);
1766 hpts->p_on_queue_cnt = 0;
1769 } else if (wrap_loop_cnt >= 2) {
1770 /* Special case handling */
1771 hpts->p_hpts_sleep_time = tcp_min_hptsi_time;
1773 /* No one on the wheel sleep for all but 400 slots or sleep max */
1775 hpts->p_hpts_sleep_time = hpts_sleep_max;
1781 __tcp_set_hpts(struct inpcb *inp, int32_t line)
1783 struct tcp_hpts_entry *hpts;
1785 INP_WLOCK_ASSERT(inp);
1786 hpts = tcp_hpts_lock(inp);
1787 if ((inp->inp_in_hpts == 0) &&
1788 (inp->inp_hpts_cpu_set == 0)) {
1789 inp->inp_hpts_cpu = hpts_cpuid(inp);
1790 inp->inp_hpts_cpu_set = 1;
1792 mtx_unlock(&hpts->p_mtx);
1793 hpts = tcp_input_lock(inp);
1794 if ((inp->inp_input_cpu_set == 0) &&
1795 (inp->inp_in_input == 0)) {
1796 inp->inp_input_cpu = hpts_cpuid(inp);
1797 inp->inp_input_cpu_set = 1;
1799 mtx_unlock(&hpts->p_mtx);
1803 tcp_hpts_delayedby(struct inpcb *inp){
1804 return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by);
1808 tcp_hpts_thread(void *ctx)
1810 struct tcp_hpts_entry *hpts;
1811 struct epoch_tracker et;
1815 hpts = (struct tcp_hpts_entry *)ctx;
1816 mtx_lock(&hpts->p_mtx);
1817 if (hpts->p_direct_wake) {
1818 /* Signaled by input */
1819 callout_stop(&hpts->co);
1822 if (callout_pending(&hpts->co) ||
1823 !callout_active(&hpts->co)) {
1824 mtx_unlock(&hpts->p_mtx);
1827 callout_deactivate(&hpts->co);
1829 hpts->p_hpts_wake_scheduled = 0;
1830 hpts->p_hpts_active = 1;
1831 NET_EPOCH_ENTER(et);
1834 HPTS_MTX_ASSERT(hpts);
1836 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1837 if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) {
1838 hpts->overidden_sleep = tv.tv_usec;
1839 tv.tv_usec = tcp_min_hptsi_time;
1840 hpts->p_on_min_sleep = 1;
1842 /* Clear the min sleep flag */
1843 hpts->overidden_sleep = 0;
1844 hpts->p_on_min_sleep = 0;
1846 hpts->p_hpts_active = 0;
1848 if (tcp_hpts_callout_skip_swi == 0) {
1849 callout_reset_sbt_on(&hpts->co, sb, 0,
1850 hpts_timeout_swi, hpts, hpts->p_cpu,
1851 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1853 callout_reset_sbt_on(&hpts->co, sb, 0,
1854 hpts_timeout_dir, hpts,
1856 C_PREL(tcp_hpts_precision));
1858 hpts->p_direct_wake = 0;
1859 mtx_unlock(&hpts->p_mtx);
1865 tcp_init_hptsi(void *st)
1867 int32_t i, j, error, bound = 0, created = 0;
1871 struct tcp_hpts_entry *hpts;
1875 uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1878 tcp_pace.rp_proc = NULL;
1879 tcp_pace.rp_num_hptss = ncpus;
1880 hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK);
1881 hpts_loops = counter_u64_alloc(M_WAITOK);
1882 back_tosleep = counter_u64_alloc(M_WAITOK);
1883 combined_wheel_wrap = counter_u64_alloc(M_WAITOK);
1884 wheel_wrap = counter_u64_alloc(M_WAITOK);
1885 sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *));
1886 tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO);
1887 asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS;
1888 for (i = 0; i < tcp_pace.rp_num_hptss; i++) {
1889 tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry),
1890 M_TCPHPTS, M_WAITOK | M_ZERO);
1891 tcp_pace.rp_ent[i]->p_hptss = malloc(asz,
1892 M_TCPHPTS, M_WAITOK);
1893 hpts = tcp_pace.rp_ent[i];
1895 * Init all the hpts structures that are not specifically
1896 * zero'd by the allocations. Also lets attach them to the
1897 * appropriate sysctl block as well.
1899 mtx_init(&hpts->p_mtx, "tcp_hpts_lck",
1900 "hpts", MTX_DEF | MTX_DUPOK);
1901 TAILQ_INIT(&hpts->p_input);
1902 for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) {
1903 TAILQ_INIT(&hpts->p_hptss[j]);
1905 sysctl_ctx_init(&hpts->hpts_ctx);
1906 sprintf(unit, "%d", i);
1907 hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx,
1908 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts),
1911 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1913 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1914 SYSCTL_CHILDREN(hpts->hpts_root),
1915 OID_AUTO, "in_qcnt", CTLFLAG_RD,
1916 &hpts->p_on_inqueue_cnt, 0,
1917 "Count TCB's awaiting input processing");
1918 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1919 SYSCTL_CHILDREN(hpts->hpts_root),
1920 OID_AUTO, "out_qcnt", CTLFLAG_RD,
1921 &hpts->p_on_queue_cnt, 0,
1922 "Count TCB's awaiting output processing");
1923 SYSCTL_ADD_U16(&hpts->hpts_ctx,
1924 SYSCTL_CHILDREN(hpts->hpts_root),
1925 OID_AUTO, "active", CTLFLAG_RD,
1926 &hpts->p_hpts_active, 0,
1927 "Is the hpts active");
1928 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1929 SYSCTL_CHILDREN(hpts->hpts_root),
1930 OID_AUTO, "curslot", CTLFLAG_RD,
1931 &hpts->p_cur_slot, 0,
1932 "What the current running pacers goal");
1933 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1934 SYSCTL_CHILDREN(hpts->hpts_root),
1935 OID_AUTO, "runtick", CTLFLAG_RD,
1936 &hpts->p_runningtick, 0,
1937 "What the running pacers current slot is");
1938 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1939 SYSCTL_CHILDREN(hpts->hpts_root),
1940 OID_AUTO, "curtick", CTLFLAG_RD,
1941 &hpts->p_curtick, 0,
1942 "What the running pacers last tick mapped to the wheel was");
1943 hpts->p_hpts_sleep_time = hpts_sleep_max;
1945 hpts->p_curtick = tcp_gethptstick(&tv);
1946 hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1947 hpts->p_cpu = 0xffff;
1948 hpts->p_nxt_slot = hpts_tick(hpts->p_cur_slot, 1);
1949 callout_init(&hpts->co, 1);
1952 /* Don't try to bind to NUMA domains if we don't have any */
1953 if (vm_ndomains == 1 && tcp_bind_threads == 2)
1954 tcp_bind_threads = 0;
1957 * Now lets start ithreads to handle the hptss.
1960 hpts = tcp_pace.rp_ent[i];
1962 error = swi_add(&hpts->ie, "hpts",
1963 tcp_hpts_thread, (void *)hpts,
1964 SWI_NET, INTR_MPSAFE, &hpts->ie_cookie);
1966 panic("Can't add hpts:%p i:%d err:%d",
1970 if (tcp_bind_threads == 1) {
1971 if (intr_event_bind(hpts->ie, i) == 0)
1973 } else if (tcp_bind_threads == 2) {
1975 domain = pc->pc_domain;
1976 CPU_COPY(&cpuset_domain[domain], &cs);
1977 if (intr_event_bind_ithread_cpuset(hpts->ie, &cs)
1980 count = hpts_domains[domain].count;
1981 hpts_domains[domain].cpu[count] = i;
1982 hpts_domains[domain].count++;
1986 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1988 if (tcp_hpts_callout_skip_swi == 0) {
1989 callout_reset_sbt_on(&hpts->co, sb, 0,
1990 hpts_timeout_swi, hpts, hpts->p_cpu,
1991 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1993 callout_reset_sbt_on(&hpts->co, sb, 0,
1994 hpts_timeout_dir, hpts,
1996 C_PREL(tcp_hpts_precision));
2000 * If we somehow have an empty domain, fall back to choosing
2001 * among all htps threads.
2003 for (i = 0; i < vm_ndomains; i++) {
2004 if (hpts_domains[i].count == 0) {
2005 tcp_bind_threads = 0;
2010 printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n",
2012 tcp_bind_threads == 2 ? "NUMA domains" : "cpus");
2015 SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL);
2016 MODULE_VERSION(tcphpts, 1);