1 /*-
2 * Copyright (c) 2016-2018 Netflix, Inc.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 *
25 */
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD$");
28
29 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_tcpdebug.h"
32 /**
33 * Some notes about usage.
34 *
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).
39 *
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:
48 *
49 * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550));
50 *
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:
55 *
56 * if (inp->inp_in_hpts)
57 * return;
58 *
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.
62 *
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
69 * system as well.
70 *
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.
76 *
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.
86 *
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).
99 *
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
108 * described above.
109 *
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:
114 *
115 * tcp_set_inp_to_drop(tp, ETIMEDOUT)
116 *
117 * To schedule the tcp_hpts system to call
118 *
119 * tcp_drop(tp, drop_reason)
120 *
121 * at a future point. This is quite handy to prevent locking
122 * issues when dropping connections.
123 *
124 */
125
126 #include <sys/param.h>
127 #include <sys/bus.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>
142 #include <sys/smp.h>
143 #include <sys/counter.h>
144 #include <sys/time.h>
145 #include <sys/kthread.h>
146 #include <sys/kern_prefetch.h>
147
148 #include <vm/uma.h>
149
150 #include <net/route.h>
151 #include <net/vnet.h>
152
153 #define TCPSTATES /* for logging */
154
155 #include <netinet/in.h>
156 #include <netinet/in_kdtrace.h>
157 #include <netinet/in_pcb.h>
158 #include <netinet/ip.h>
159 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
160 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
161 #include <netinet/ip_var.h>
162 #include <netinet/ip6.h>
163 #include <netinet6/in6_pcb.h>
164 #include <netinet6/ip6_var.h>
165 #include <netinet/tcp.h>
166 #include <netinet/tcp_fsm.h>
167 #include <netinet/tcp_seq.h>
168 #include <netinet/tcp_timer.h>
169 #include <netinet/tcp_var.h>
170 #include <netinet/tcpip.h>
171 #include <netinet/cc/cc.h>
172 #include <netinet/tcp_hpts.h>
173 #include <netinet/tcp_log_buf.h>
174
175 #ifdef tcpdebug
176 #include <netinet/tcp_debug.h>
177 #endif /* tcpdebug */
178 #ifdef tcp_offload
179 #include <netinet/tcp_offload.h>
180 #endif
181
182 #include "opt_rss.h"
183
184 MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts");
185 #ifdef RSS
186 static int tcp_bind_threads = 1;
187 #else
188 static int tcp_bind_threads = 0;
189 #endif
190 TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads);
191
192 static struct tcp_hptsi tcp_pace;
193 static int hpts_does_tp_logging = 0;
194
195 static void tcp_wakehpts(struct tcp_hpts_entry *p);
196 static void tcp_wakeinput(struct tcp_hpts_entry *p);
197 static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv);
198 static void tcp_hptsi(struct tcp_hpts_entry *hpts);
199 static void tcp_hpts_thread(void *ctx);
200 static void tcp_init_hptsi(void *st);
201
202 int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP;
203 static int32_t tcp_hpts_callout_skip_swi = 0;
204
205 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW, 0, "TCP Hpts controls");
206
207 #define timersub(tvp, uvp, vvp) \
208 do { \
209 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \
210 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \
211 if ((vvp)->tv_usec < 0) { \
212 (vvp)->tv_sec--; \
213 (vvp)->tv_usec += 1000000; \
214 } \
215 } while (0)
216
217 static int32_t tcp_hpts_precision = 120;
218
219 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW,
220 &tcp_hpts_precision, 120,
221 "Value for PRE() precision of callout");
222
223 counter_u64_t hpts_hopelessly_behind;
224
225 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, hopeless, CTLFLAG_RD,
226 &hpts_hopelessly_behind,
227 "Number of times hpts could not catch up and was behind hopelessly");
228
229 counter_u64_t hpts_loops;
230
231 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD,
232 &hpts_loops, "Number of times hpts had to loop to catch up");
233
234
235 counter_u64_t back_tosleep;
236
237 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD,
238 &back_tosleep, "Number of times hpts found no tcbs");
239
240 counter_u64_t combined_wheel_wrap;
241
242 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD,
243 &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
244
245 counter_u64_t wheel_wrap;
246
247 SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, wheel_wrap, CTLFLAG_RD,
248 &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap");
249
250 static int32_t out_ts_percision = 0;
251
252 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW,
253 &out_ts_percision, 0,
254 "Do we use a percise timestamp for every output cts");
255 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW,
256 &hpts_does_tp_logging, 0,
257 "Do we add to any tp that has logging on pacer logs");
258
259 static int32_t max_pacer_loops = 10;
260 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW,
261 &max_pacer_loops, 10,
262 "What is the maximum number of times the pacer will loop trying to catch up");
263
264 #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2)
265
266 static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED;
267
268
269 static int
270 sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS)
271 {
272 int error;
273 uint32_t new;
274
275 new = hpts_sleep_max;
276 error = sysctl_handle_int(oidp, &new, 0, req);
277 if (error == 0 && req->newptr) {
278 if ((new < (NUM_OF_HPTSI_SLOTS / 4)) ||
279 (new > HPTS_MAX_SLEEP_ALLOWED))
280 error = EINVAL;
281 else
282 hpts_sleep_max = new;
283 }
284 return (error);
285 }
286
287 SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep,
288 CTLTYPE_UINT | CTLFLAG_RW,
289 &hpts_sleep_max, 0,
290 &sysctl_net_inet_tcp_hpts_max_sleep, "IU",
291 "Maximum time hpts will sleep");
292
293 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW,
294 &tcp_min_hptsi_time, 0,
295 "The minimum time the hpts must sleep before processing more slots");
296
297 SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW,
298 &tcp_hpts_callout_skip_swi, 0,
299 "Do we have the callout call directly to the hpts?");
300
301 static void
302 tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv,
303 int ticks_to_run, int idx)
304 {
305 union tcp_log_stackspecific log;
306
307 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
308 log.u_bbr.flex1 = hpts->p_nxt_slot;
309 log.u_bbr.flex2 = hpts->p_cur_slot;
310 log.u_bbr.flex3 = hpts->p_prev_slot;
311 log.u_bbr.flex4 = idx;
312 log.u_bbr.flex5 = hpts->p_curtick;
313 log.u_bbr.flex6 = hpts->p_on_queue_cnt;
314 log.u_bbr.use_lt_bw = 1;
315 log.u_bbr.inflight = ticks_to_run;
316 log.u_bbr.applimited = hpts->overidden_sleep;
317 log.u_bbr.delivered = hpts->saved_curtick;
318 log.u_bbr.timeStamp = tcp_tv_to_usectick(tv);
319 log.u_bbr.epoch = hpts->saved_curslot;
320 log.u_bbr.lt_epoch = hpts->saved_prev_slot;
321 log.u_bbr.pkts_out = hpts->p_delayed_by;
322 log.u_bbr.lost = hpts->p_hpts_sleep_time;
323 log.u_bbr.cur_del_rate = hpts->p_runningtick;
324 TCP_LOG_EVENTP(tp, NULL,
325 &tp->t_inpcb->inp_socket->so_rcv,
326 &tp->t_inpcb->inp_socket->so_snd,
327 BBR_LOG_HPTSDIAG, 0,
328 0, &log, false, tv);
329 }
330
331 static void
332 hpts_timeout_swi(void *arg)
333 {
334 struct tcp_hpts_entry *hpts;
335
336 hpts = (struct tcp_hpts_entry *)arg;
337 swi_sched(hpts->ie_cookie, 0);
338 }
339
340 static void
341 hpts_timeout_dir(void *arg)
342 {
343 tcp_hpts_thread(arg);
344 }
345
346 static inline void
347 hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear)
348 {
349 #ifdef INVARIANTS
350 if (mtx_owned(&hpts->p_mtx) == 0) {
351 /* We don't own the mutex? */
352 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
353 }
354 if (hpts->p_cpu != inp->inp_hpts_cpu) {
355 /* It is not the right cpu/mutex? */
356 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
357 }
358 if (inp->inp_in_hpts == 0) {
359 /* We are not on the hpts? */
360 panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp);
361 }
362 #endif
363 TAILQ_REMOVE(head, inp, inp_hpts);
364 hpts->p_on_queue_cnt--;
365 if (hpts->p_on_queue_cnt < 0) {
366 /* Count should not go negative .. */
367 #ifdef INVARIANTS
368 panic("Hpts goes negative inp:%p hpts:%p",
369 inp, hpts);
370 #endif
371 hpts->p_on_queue_cnt = 0;
372 }
373 if (clear) {
374 inp->inp_hpts_request = 0;
375 inp->inp_in_hpts = 0;
376 }
377 }
378
379 static inline void
380 hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref)
381 {
382 #ifdef INVARIANTS
383 if (mtx_owned(&hpts->p_mtx) == 0) {
384 /* We don't own the mutex? */
385 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
386 }
387 if (hpts->p_cpu != inp->inp_hpts_cpu) {
388 /* It is not the right cpu/mutex? */
389 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
390 }
391 if ((noref == 0) && (inp->inp_in_hpts == 1)) {
392 /* We are already on the hpts? */
393 panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp);
394 }
395 #endif
396 TAILQ_INSERT_TAIL(head, inp, inp_hpts);
397 inp->inp_in_hpts = 1;
398 hpts->p_on_queue_cnt++;
399 if (noref == 0) {
400 in_pcbref(inp);
401 }
402 }
403
404 static inline void
405 hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear)
406 {
407 #ifdef INVARIANTS
408 if (mtx_owned(&hpts->p_mtx) == 0) {
409 /* We don't own the mutex? */
410 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
411 }
412 if (hpts->p_cpu != inp->inp_input_cpu) {
413 /* It is not the right cpu/mutex? */
414 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
415 }
416 if (inp->inp_in_input == 0) {
417 /* We are not on the input hpts? */
418 panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp);
419 }
420 #endif
421 TAILQ_REMOVE(&hpts->p_input, inp, inp_input);
422 hpts->p_on_inqueue_cnt--;
423 if (hpts->p_on_inqueue_cnt < 0) {
424 #ifdef INVARIANTS
425 panic("Hpts in goes negative inp:%p hpts:%p",
426 inp, hpts);
427 #endif
428 hpts->p_on_inqueue_cnt = 0;
429 }
430 #ifdef INVARIANTS
431 if (TAILQ_EMPTY(&hpts->p_input) &&
432 (hpts->p_on_inqueue_cnt != 0)) {
433 /* We should not be empty with a queue count */
434 panic("%s hpts:%p in_hpts input empty but cnt:%d",
435 __FUNCTION__, hpts, hpts->p_on_inqueue_cnt);
436 }
437 #endif
438 if (clear)
439 inp->inp_in_input = 0;
440 }
441
442 static inline void
443 hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line)
444 {
445 #ifdef INVARIANTS
446 if (mtx_owned(&hpts->p_mtx) == 0) {
447 /* We don't own the mutex? */
448 panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp);
449 }
450 if (hpts->p_cpu != inp->inp_input_cpu) {
451 /* It is not the right cpu/mutex? */
452 panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp);
453 }
454 if (inp->inp_in_input == 1) {
455 /* We are already on the input hpts? */
456 panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp);
457 }
458 #endif
459 TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input);
460 inp->inp_in_input = 1;
461 hpts->p_on_inqueue_cnt++;
462 in_pcbref(inp);
463 }
464
465 static void
466 tcp_wakehpts(struct tcp_hpts_entry *hpts)
467 {
468 HPTS_MTX_ASSERT(hpts);
469 if (hpts->p_hpts_wake_scheduled == 0) {
470 hpts->p_hpts_wake_scheduled = 1;
471 swi_sched(hpts->ie_cookie, 0);
472 }
473 }
474
475 static void
476 tcp_wakeinput(struct tcp_hpts_entry *hpts)
477 {
478 HPTS_MTX_ASSERT(hpts);
479 if (hpts->p_hpts_wake_scheduled == 0) {
480 hpts->p_hpts_wake_scheduled = 1;
481 swi_sched(hpts->ie_cookie, 0);
482 }
483 }
484
485 struct tcp_hpts_entry *
486 tcp_cur_hpts(struct inpcb *inp)
487 {
488 int32_t hpts_num;
489 struct tcp_hpts_entry *hpts;
490
491 hpts_num = inp->inp_hpts_cpu;
492 hpts = tcp_pace.rp_ent[hpts_num];
493 return (hpts);
494 }
495
496 struct tcp_hpts_entry *
497 tcp_hpts_lock(struct inpcb *inp)
498 {
499 struct tcp_hpts_entry *hpts;
500 int32_t hpts_num;
501
502 again:
503 hpts_num = inp->inp_hpts_cpu;
504 hpts = tcp_pace.rp_ent[hpts_num];
505 #ifdef INVARIANTS
506 if (mtx_owned(&hpts->p_mtx)) {
507 panic("Hpts:%p owns mtx prior-to lock line:%d",
508 hpts, __LINE__);
509 }
510 #endif
511 mtx_lock(&hpts->p_mtx);
512 if (hpts_num != inp->inp_hpts_cpu) {
513 mtx_unlock(&hpts->p_mtx);
514 goto again;
515 }
516 return (hpts);
517 }
518
519 struct tcp_hpts_entry *
520 tcp_input_lock(struct inpcb *inp)
521 {
522 struct tcp_hpts_entry *hpts;
523 int32_t hpts_num;
524
525 again:
526 hpts_num = inp->inp_input_cpu;
527 hpts = tcp_pace.rp_ent[hpts_num];
528 #ifdef INVARIANTS
529 if (mtx_owned(&hpts->p_mtx)) {
530 panic("Hpts:%p owns mtx prior-to lock line:%d",
531 hpts, __LINE__);
532 }
533 #endif
534 mtx_lock(&hpts->p_mtx);
535 if (hpts_num != inp->inp_input_cpu) {
536 mtx_unlock(&hpts->p_mtx);
537 goto again;
538 }
539 return (hpts);
540 }
541
542 static void
543 tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line)
544 {
545 int32_t add_freed;
546
547 if (inp->inp_flags2 & INP_FREED) {
548 /*
549 * Need to play a special trick so that in_pcbrele_wlocked
550 * does not return 1 when it really should have returned 0.
551 */
552 add_freed = 1;
553 inp->inp_flags2 &= ~INP_FREED;
554 } else {
555 add_freed = 0;
556 }
557 #ifndef INP_REF_DEBUG
558 if (in_pcbrele_wlocked(inp)) {
559 /*
560 * This should not happen. We have the inpcb referred to by
561 * the main socket (why we are called) and the hpts. It
562 * should always return 0.
563 */
564 panic("inpcb:%p release ret 1",
565 inp);
566 }
567 #else
568 if (__in_pcbrele_wlocked(inp, line)) {
569 /*
570 * This should not happen. We have the inpcb referred to by
571 * the main socket (why we are called) and the hpts. It
572 * should always return 0.
573 */
574 panic("inpcb:%p release ret 1",
575 inp);
576 }
577 #endif
578 if (add_freed) {
579 inp->inp_flags2 |= INP_FREED;
580 }
581 }
582
583 static void
584 tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
585 {
586 if (inp->inp_in_hpts) {
587 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1);
588 tcp_remove_hpts_ref(inp, hpts, line);
589 }
590 }
591
592 static void
593 tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line)
594 {
595 HPTS_MTX_ASSERT(hpts);
596 if (inp->inp_in_input) {
597 hpts_sane_input_remove(hpts, inp, 1);
598 tcp_remove_hpts_ref(inp, hpts, line);
599 }
600 }
601
602 /*
603 * Called normally with the INP_LOCKED but it
604 * does not matter, the hpts lock is the key
605 * but the lock order allows us to hold the
606 * INP lock and then get the hpts lock.
607 *
608 * Valid values in the flags are
609 * HPTS_REMOVE_OUTPUT - remove from the output of the hpts.
610 * HPTS_REMOVE_INPUT - remove from the input of the hpts.
611 * Note that you can use one or both values together
612 * and get two actions.
613 */
614 void
615 __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line)
616 {
617 struct tcp_hpts_entry *hpts;
618
619 INP_WLOCK_ASSERT(inp);
620 if (flags & HPTS_REMOVE_OUTPUT) {
621 hpts = tcp_hpts_lock(inp);
622 tcp_hpts_remove_locked_output(hpts, inp, flags, line);
623 mtx_unlock(&hpts->p_mtx);
624 }
625 if (flags & HPTS_REMOVE_INPUT) {
626 hpts = tcp_input_lock(inp);
627 tcp_hpts_remove_locked_input(hpts, inp, flags, line);
628 mtx_unlock(&hpts->p_mtx);
629 }
630 }
631
632 static inline int
633 hpts_tick(uint32_t wheel_tick, uint32_t plus)
634 {
635 /*
636 * Given a slot on the wheel, what slot
637 * is that plus ticks out?
638 */
639 KASSERT(wheel_tick < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_tick));
640 return ((wheel_tick + plus) % NUM_OF_HPTSI_SLOTS);
641 }
642
643 static inline int
644 tick_to_wheel(uint32_t cts_in_wticks)
645 {
646 /*
647 * Given a timestamp in wheel ticks (10usec inc's)
648 * map it to our limited space wheel.
649 */
650 return (cts_in_wticks % NUM_OF_HPTSI_SLOTS);
651 }
652
653 static inline int
654 hpts_ticks_diff(int prev_tick, int tick_now)
655 {
656 /*
657 * Given two ticks that are someplace
658 * on our wheel. How far are they apart?
659 */
660 if (tick_now > prev_tick)
661 return (tick_now - prev_tick);
662 else if (tick_now == prev_tick)
663 /*
664 * Special case, same means we can go all of our
665 * wheel less one slot.
666 */
667 return (NUM_OF_HPTSI_SLOTS - 1);
668 else
669 return ((NUM_OF_HPTSI_SLOTS - prev_tick) + tick_now);
670 }
671
672 /*
673 * Given a tick on the wheel that is the current time
674 * mapped to the wheel (wheel_tick), what is the maximum
675 * distance forward that can be obtained without
676 * wrapping past either prev_tick or running_tick
677 * depending on the htps state? Also if passed
678 * a uint32_t *, fill it with the tick location.
679 *
680 * Note if you do not give this function the current
681 * time (that you think it is) mapped to the wheel
682 * then the results will not be what you expect and
683 * could lead to invalid inserts.
684 */
685 static inline int32_t
686 max_ticks_available(struct tcp_hpts_entry *hpts, uint32_t wheel_tick, uint32_t *target_tick)
687 {
688 uint32_t dis_to_travel, end_tick, pacer_to_now, avail_on_wheel;
689
690 if ((hpts->p_hpts_active == 1) &&
691 (hpts->p_wheel_complete == 0)) {
692 end_tick = hpts->p_runningtick;
693 /* Back up one tick */
694 if (end_tick == 0)
695 end_tick = NUM_OF_HPTSI_SLOTS - 1;
696 else
697 end_tick--;
698 if (target_tick)
699 *target_tick = end_tick;
700 } else {
701 /*
702 * For the case where we are
703 * not active, or we have
704 * completed the pass over
705 * the wheel, we can use the
706 * prev tick and subtract one from it. This puts us
707 * as far out as possible on the wheel.
708 */
709 end_tick = hpts->p_prev_slot;
710 if (end_tick == 0)
711 end_tick = NUM_OF_HPTSI_SLOTS - 1;
712 else
713 end_tick--;
714 if (target_tick)
715 *target_tick = end_tick;
716 /*
717 * Now we have close to the full wheel left minus the
718 * time it has been since the pacer went to sleep. Note
719 * that wheel_tick, passed in, should be the current time
720 * from the perspective of the caller, mapped to the wheel.
721 */
722 if (hpts->p_prev_slot != wheel_tick)
723 dis_to_travel = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
724 else
725 dis_to_travel = 1;
726 /*
727 * dis_to_travel in this case is the space from when the
728 * pacer stopped (p_prev_slot) and where our wheel_tick
729 * is now. To know how many slots we can put it in we
730 * subtract from the wheel size. We would not want
731 * to place something after p_prev_slot or it will
732 * get ran too soon.
733 */
734 return (NUM_OF_HPTSI_SLOTS - dis_to_travel);
735 }
736 /*
737 * So how many slots are open between p_runningtick -> p_cur_slot
738 * that is what is currently un-available for insertion. Special
739 * case when we are at the last slot, this gets 1, so that
740 * the answer to how many slots are available is all but 1.
741 */
742 if (hpts->p_runningtick == hpts->p_cur_slot)
743 dis_to_travel = 1;
744 else
745 dis_to_travel = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
746 /*
747 * How long has the pacer been running?
748 */
749 if (hpts->p_cur_slot != wheel_tick) {
750 /* The pacer is a bit late */
751 pacer_to_now = hpts_ticks_diff(hpts->p_cur_slot, wheel_tick);
752 } else {
753 /* The pacer is right on time, now == pacers start time */
754 pacer_to_now = 0;
755 }
756 /*
757 * To get the number left we can insert into we simply
758 * subtract the distance the pacer has to run from how
759 * many slots there are.
760 */
761 avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel;
762 /*
763 * Now how many of those we will eat due to the pacer's
764 * time (p_cur_slot) of start being behind the
765 * real time (wheel_tick)?
766 */
767 if (avail_on_wheel <= pacer_to_now) {
768 /*
769 * Wheel wrap, we can't fit on the wheel, that
770 * is unusual the system must be way overloaded!
771 * Insert into the assured tick, and return special
772 * "".
773 */
774 counter_u64_add(combined_wheel_wrap, 1);
775 *target_tick = hpts->p_nxt_slot;
776 return (0);
777 } else {
778 /*
779 * We know how many slots are open
780 * on the wheel (the reverse of what
781 * is left to run. Take away the time
782 * the pacer started to now (wheel_tick)
783 * and that tells you how many slots are
784 * open that can be inserted into that won't
785 * be touched by the pacer until later.
786 */
787 return (avail_on_wheel - pacer_to_now);
788 }
789 }
790
791 static int
792 tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref)
793 {
794 uint32_t need_wake = 0;
795
796 HPTS_MTX_ASSERT(hpts);
797 if (inp->inp_in_hpts == 0) {
798 /* Ok we need to set it on the hpts in the current slot */
799 inp->inp_hpts_request = 0;
800 if ((hpts->p_hpts_active == 0) ||
801 (hpts->p_wheel_complete)) {
802 /*
803 * A sleeping hpts we want in next slot to run
804 * note that in this state p_prev_slot == p_cur_slot
805 */
806 inp->inp_hptsslot = hpts_tick(hpts->p_prev_slot, 1);
807 if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0))
808 need_wake = 1;
809 } else if ((void *)inp == hpts->p_inp) {
810 /*
811 * The hpts system is running and the caller
812 * was awoken by the hpts system.
813 * We can't allow you to go into the same slot we
814 * are in (we don't want a loop :-D).
815 */
816 inp->inp_hptsslot = hpts->p_nxt_slot;
817 } else
818 inp->inp_hptsslot = hpts->p_runningtick;
819 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref);
820 if (need_wake) {
821 /*
822 * Activate the hpts if it is sleeping and its
823 * timeout is not 1.
824 */
825 hpts->p_direct_wake = 1;
826 tcp_wakehpts(hpts);
827 }
828 }
829 return (need_wake);
830 }
831
832 int
833 __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line)
834 {
835 int32_t ret;
836 struct tcp_hpts_entry *hpts;
837
838 INP_WLOCK_ASSERT(inp);
839 hpts = tcp_hpts_lock(inp);
840 ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
841 mtx_unlock(&hpts->p_mtx);
842 return (ret);
843 }
844
845 #ifdef INVARIANTS
846 static void
847 check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line)
848 {
849 /*
850 * Sanity checks for the pacer with invariants
851 * on insert.
852 */
853 if (inp_hptsslot >= NUM_OF_HPTSI_SLOTS)
854 panic("hpts:%p inp:%p slot:%d > max",
855 hpts, inp, inp_hptsslot);
856 if ((hpts->p_hpts_active) &&
857 (hpts->p_wheel_complete == 0)) {
858 /*
859 * If the pacer is processing a arc
860 * of the wheel, we need to make
861 * sure we are not inserting within
862 * that arc.
863 */
864 int distance, yet_to_run;
865
866 distance = hpts_ticks_diff(hpts->p_runningtick, inp_hptsslot);
867 if (hpts->p_runningtick != hpts->p_cur_slot)
868 yet_to_run = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot);
869 else
870 yet_to_run = 0; /* processing last slot */
871 if (yet_to_run > distance) {
872 panic("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d",
873 hpts, inp, inp_hptsslot,
874 distance, yet_to_run,
875 hpts->p_runningtick, hpts->p_cur_slot);
876 }
877 }
878 }
879 #endif
880
881 static void
882 tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, int32_t line,
883 struct hpts_diag *diag, struct timeval *tv)
884 {
885 uint32_t need_new_to = 0;
886 uint32_t wheel_cts, last_tick;
887 int32_t wheel_tick, maxticks;
888 int8_t need_wakeup = 0;
889
890 HPTS_MTX_ASSERT(hpts);
891 if (diag) {
892 memset(diag, 0, sizeof(struct hpts_diag));
893 diag->p_hpts_active = hpts->p_hpts_active;
894 diag->p_prev_slot = hpts->p_prev_slot;
895 diag->p_runningtick = hpts->p_runningtick;
896 diag->p_nxt_slot = hpts->p_nxt_slot;
897 diag->p_cur_slot = hpts->p_cur_slot;
898 diag->p_curtick = hpts->p_curtick;
899 diag->p_lasttick = hpts->p_lasttick;
900 diag->slot_req = slot;
901 diag->p_on_min_sleep = hpts->p_on_min_sleep;
902 diag->hpts_sleep_time = hpts->p_hpts_sleep_time;
903 }
904 if (inp->inp_in_hpts == 0) {
905 if (slot == 0) {
906 /* Immediate */
907 tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0);
908 return;
909 }
910 /* Get the current time relative to the wheel */
911 wheel_cts = tcp_tv_to_hptstick(tv);
912 /* Map it onto the wheel */
913 wheel_tick = tick_to_wheel(wheel_cts);
914 /* Now what's the max we can place it at? */
915 maxticks = max_ticks_available(hpts, wheel_tick, &last_tick);
916 if (diag) {
917 diag->wheel_tick = wheel_tick;
918 diag->maxticks = maxticks;
919 diag->wheel_cts = wheel_cts;
920 }
921 if (maxticks == 0) {
922 /* The pacer is in a wheel wrap behind, yikes! */
923 if (slot > 1) {
924 /*
925 * Reduce by 1 to prevent a forever loop in
926 * case something else is wrong. Note this
927 * probably does not hurt because the pacer
928 * if its true is so far behind we will be
929 * > 1second late calling anyway.
930 */
931 slot--;
932 }
933 inp->inp_hptsslot = last_tick;
934 inp->inp_hpts_request = slot;
935 } else if (maxticks >= slot) {
936 /* It all fits on the wheel */
937 inp->inp_hpts_request = 0;
938 inp->inp_hptsslot = hpts_tick(wheel_tick, slot);
939 } else {
940 /* It does not fit */
941 inp->inp_hpts_request = slot - maxticks;
942 inp->inp_hptsslot = last_tick;
943 }
944 if (diag) {
945 diag->slot_remaining = inp->inp_hpts_request;
946 diag->inp_hptsslot = inp->inp_hptsslot;
947 }
948 #ifdef INVARIANTS
949 check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line);
950 #endif
951 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0);
952 if ((hpts->p_hpts_active == 0) &&
953 (inp->inp_hpts_request == 0) &&
954 (hpts->p_on_min_sleep == 0)) {
955 /*
956 * The hpts is sleeping and not on a minimum
957 * sleep time, we need to figure out where
958 * it will wake up at and if we need to reschedule
959 * its time-out.
960 */
961 uint32_t have_slept, yet_to_sleep;
962
963 /* Now do we need to restart the hpts's timer? */
964 have_slept = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick);
965 if (have_slept < hpts->p_hpts_sleep_time)
966 yet_to_sleep = hpts->p_hpts_sleep_time - have_slept;
967 else {
968 /* We are over-due */
969 yet_to_sleep = 0;
970 need_wakeup = 1;
971 }
972 if (diag) {
973 diag->have_slept = have_slept;
974 diag->yet_to_sleep = yet_to_sleep;
975 }
976 if (yet_to_sleep &&
977 (yet_to_sleep > slot)) {
978 /*
979 * We need to reschedule the hpts's time-out.
980 */
981 hpts->p_hpts_sleep_time = slot;
982 need_new_to = slot * HPTS_TICKS_PER_USEC;
983 }
984 }
985 /*
986 * Now how far is the hpts sleeping to? if active is 1, its
987 * up and ticking we do nothing, otherwise we may need to
988 * reschedule its callout if need_new_to is set from above.
989 */
990 if (need_wakeup) {
991 hpts->p_direct_wake = 1;
992 tcp_wakehpts(hpts);
993 if (diag) {
994 diag->need_new_to = 0;
995 diag->co_ret = 0xffff0000;
996 }
997 } else if (need_new_to) {
998 int32_t co_ret;
999 struct timeval tv;
1000 sbintime_t sb;
1001
1002 tv.tv_sec = 0;
1003 tv.tv_usec = 0;
1004 while (need_new_to > HPTS_USEC_IN_SEC) {
1005 tv.tv_sec++;
1006 need_new_to -= HPTS_USEC_IN_SEC;
1007 }
1008 tv.tv_usec = need_new_to;
1009 sb = tvtosbt(tv);
1010 if (tcp_hpts_callout_skip_swi == 0) {
1011 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1012 hpts_timeout_swi, hpts, hpts->p_cpu,
1013 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1014 } else {
1015 co_ret = callout_reset_sbt_on(&hpts->co, sb, 0,
1016 hpts_timeout_dir, hpts,
1017 hpts->p_cpu,
1018 C_PREL(tcp_hpts_precision));
1019 }
1020 if (diag) {
1021 diag->need_new_to = need_new_to;
1022 diag->co_ret = co_ret;
1023 }
1024 }
1025 } else {
1026 #ifdef INVARIANTS
1027 panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp);
1028 #endif
1029 }
1030 }
1031
1032 uint32_t
1033 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag)
1034 {
1035 struct tcp_hpts_entry *hpts;
1036 uint32_t slot_on;
1037 struct timeval tv;
1038
1039 /*
1040 * We now return the next-slot the hpts will be on, beyond its
1041 * current run (if up) or where it was when it stopped if it is
1042 * sleeping.
1043 */
1044 INP_WLOCK_ASSERT(inp);
1045 hpts = tcp_hpts_lock(inp);
1046 microuptime(&tv);
1047 tcp_hpts_insert_locked(hpts, inp, slot, line, diag, &tv);
1048 slot_on = hpts->p_nxt_slot;
1049 mtx_unlock(&hpts->p_mtx);
1050 return (slot_on);
1051 }
1052
1053 uint32_t
1054 __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){
1055 return (tcp_hpts_insert_diag(inp, slot, line, NULL));
1056 }
1057 int
1058 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line)
1059 {
1060 int32_t retval = 0;
1061
1062 HPTS_MTX_ASSERT(hpts);
1063 if (inp->inp_in_input == 0) {
1064 /* Ok we need to set it on the hpts in the current slot */
1065 hpts_sane_input_insert(hpts, inp, line);
1066 retval = 1;
1067 if (hpts->p_hpts_active == 0) {
1068 /*
1069 * Activate the hpts if it is sleeping.
1070 */
1071 retval = 2;
1072 hpts->p_direct_wake = 1;
1073 tcp_wakeinput(hpts);
1074 }
1075 } else if (hpts->p_hpts_active == 0) {
1076 retval = 4;
1077 hpts->p_direct_wake = 1;
1078 tcp_wakeinput(hpts);
1079 }
1080 return (retval);
1081 }
1082
1083 int32_t
1084 __tcp_queue_to_input(struct inpcb *inp, int line)
1085 {
1086 struct tcp_hpts_entry *hpts;
1087 int32_t ret;
1088
1089 hpts = tcp_input_lock(inp);
1090 ret = __tcp_queue_to_input_locked(inp, hpts, line);
1091 mtx_unlock(&hpts->p_mtx);
1092 return (ret);
1093 }
1094
1095 void
1096 __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line)
1097 {
1098 struct tcp_hpts_entry *hpts;
1099 struct tcpcb *tp;
1100
1101 tp = intotcpcb(inp);
1102 hpts = tcp_input_lock(tp->t_inpcb);
1103 if (inp->inp_in_input == 0) {
1104 /* Ok we need to set it on the hpts in the current slot */
1105 hpts_sane_input_insert(hpts, inp, line);
1106 if (hpts->p_hpts_active == 0) {
1107 /*
1108 * Activate the hpts if it is sleeping.
1109 */
1110 hpts->p_direct_wake = 1;
1111 tcp_wakeinput(hpts);
1112 }
1113 } else if (hpts->p_hpts_active == 0) {
1114 hpts->p_direct_wake = 1;
1115 tcp_wakeinput(hpts);
1116 }
1117 inp->inp_hpts_drop_reas = reason;
1118 mtx_unlock(&hpts->p_mtx);
1119 }
1120
1121 static uint16_t
1122 hpts_random_cpu(struct inpcb *inp){
1123 /*
1124 * No flow type set distribute the load randomly.
1125 */
1126 uint16_t cpuid;
1127 uint32_t ran;
1128
1129 /*
1130 * If one has been set use it i.e. we want both in and out on the
1131 * same hpts.
1132 */
1133 if (inp->inp_input_cpu_set) {
1134 return (inp->inp_input_cpu);
1135 } else if (inp->inp_hpts_cpu_set) {
1136 return (inp->inp_hpts_cpu);
1137 }
1138 /* Nothing set use a random number */
1139 ran = arc4random();
1140 cpuid = (ran & 0xffff) % mp_ncpus;
1141 return (cpuid);
1142 }
1143
1144 static uint16_t
1145 hpts_cpuid(struct inpcb *inp){
1146 u_int cpuid;
1147
1148
1149 /*
1150 * If one has been set use it i.e. we want both in and out on the
1151 * same hpts.
1152 */
1153 if (inp->inp_input_cpu_set) {
1154 return (inp->inp_input_cpu);
1155 } else if (inp->inp_hpts_cpu_set) {
1156 return (inp->inp_hpts_cpu);
1157 }
1158 /* If one is set the other must be the same */
1159 #ifdef RSS
1160 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
1161 if (cpuid == NETISR_CPUID_NONE)
1162 return (hpts_random_cpu(inp));
1163 else
1164 return (cpuid);
1165 #else
1166 /*
1167 * We don't have a flowid -> cpuid mapping, so cheat and just map
1168 * unknown cpuids to curcpu. Not the best, but apparently better
1169 * than defaulting to swi 0.
1170 */
1171 if (inp->inp_flowtype != M_HASHTYPE_NONE) {
1172 cpuid = inp->inp_flowid % mp_ncpus;
1173 return (cpuid);
1174 }
1175 cpuid = hpts_random_cpu(inp);
1176 return (cpuid);
1177 #endif
1178 }
1179
1180 static void
1181 tcp_drop_in_pkts(struct tcpcb *tp)
1182 {
1183 struct mbuf *m, *n;
1184
1185 m = tp->t_in_pkt;
1186 if (m)
1187 n = m->m_nextpkt;
1188 else
1189 n = NULL;
1190 tp->t_in_pkt = NULL;
1191 while (m) {
1192 m_freem(m);
1193 m = n;
1194 if (m)
1195 n = m->m_nextpkt;
1196 }
1197 }
1198
1199 /*
1200 * Do NOT try to optimize the processing of inp's
1201 * by first pulling off all the inp's into a temporary
1202 * list (e.g. TAILQ_CONCAT). If you do that the subtle
1203 * interactions of switching CPU's will kill because of
1204 * problems in the linked list manipulation. Basically
1205 * you would switch cpu's with the hpts mutex locked
1206 * but then while you were processing one of the inp's
1207 * some other one that you switch will get a new
1208 * packet on the different CPU. It will insert it
1209 * on the new hpts's input list. Creating a temporary
1210 * link in the inp will not fix it either, since
1211 * the other hpts will be doing the same thing and
1212 * you will both end up using the temporary link.
1213 *
1214 * You will die in an ASSERT for tailq corruption if you
1215 * run INVARIANTS or you will die horribly without
1216 * INVARIANTS in some unknown way with a corrupt linked
1217 * list.
1218 */
1219 static void
1220 tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv)
1221 {
1222 struct tcpcb *tp;
1223 struct inpcb *inp;
1224 uint16_t drop_reason;
1225 int16_t set_cpu;
1226 uint32_t did_prefetch = 0;
1227 int dropped;
1228 struct epoch_tracker et;
1229
1230 HPTS_MTX_ASSERT(hpts);
1231 #ifndef VIMAGE
1232 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1233 #endif
1234 while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) {
1235 HPTS_MTX_ASSERT(hpts);
1236 hpts_sane_input_remove(hpts, inp, 0);
1237 if (inp->inp_input_cpu_set == 0) {
1238 set_cpu = 1;
1239 } else {
1240 set_cpu = 0;
1241 }
1242 hpts->p_inp = inp;
1243 drop_reason = inp->inp_hpts_drop_reas;
1244 inp->inp_in_input = 0;
1245 mtx_unlock(&hpts->p_mtx);
1246 INP_WLOCK(inp);
1247 #ifdef VIMAGE
1248 CURVNET_SET(inp->inp_vnet);
1249 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1250 #endif
1251 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1252 (inp->inp_flags2 & INP_FREED)) {
1253 out:
1254 hpts->p_inp = NULL;
1255 if (in_pcbrele_wlocked(inp) == 0) {
1256 INP_WUNLOCK(inp);
1257 }
1258 #ifdef VIMAGE
1259 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1260 CURVNET_RESTORE();
1261 #endif
1262 mtx_lock(&hpts->p_mtx);
1263 continue;
1264 }
1265 tp = intotcpcb(inp);
1266 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1267 goto out;
1268 }
1269 if (drop_reason) {
1270 /* This tcb is being destroyed for drop_reason */
1271 tcp_drop_in_pkts(tp);
1272 tp = tcp_drop(tp, drop_reason);
1273 if (tp == NULL) {
1274 INP_WLOCK(inp);
1275 }
1276 if (in_pcbrele_wlocked(inp) == 0)
1277 INP_WUNLOCK(inp);
1278 #ifdef VIMAGE
1279 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1280 CURVNET_RESTORE();
1281 #endif
1282 mtx_lock(&hpts->p_mtx);
1283 continue;
1284 }
1285 if (set_cpu) {
1286 /*
1287 * Setup so the next time we will move to the right
1288 * CPU. This should be a rare event. It will
1289 * sometimes happens when we are the client side
1290 * (usually not the server). Somehow tcp_output()
1291 * gets called before the tcp_do_segment() sets the
1292 * intial state. This means the r_cpu and r_hpts_cpu
1293 * is 0. We get on the hpts, and then tcp_input()
1294 * gets called setting up the r_cpu to the correct
1295 * value. The hpts goes off and sees the mis-match.
1296 * We simply correct it here and the CPU will switch
1297 * to the new hpts nextime the tcb gets added to the
1298 * the hpts (not this time) :-)
1299 */
1300 tcp_set_hpts(inp);
1301 }
1302 if (tp->t_fb_ptr != NULL) {
1303 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1304 did_prefetch = 1;
1305 }
1306 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1307 if (inp->inp_in_input)
1308 tcp_hpts_remove(inp, HPTS_REMOVE_INPUT);
1309 dropped = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1310 if (dropped) {
1311 /* Re-acquire the wlock so we can release the reference */
1312 INP_WLOCK(inp);
1313 }
1314 } else if (tp->t_in_pkt) {
1315 /*
1316 * We reach here only if we had a
1317 * stack that supported INP_SUPPORTS_MBUFQ
1318 * and then somehow switched to a stack that
1319 * does not. The packets are basically stranded
1320 * and would hang with the connection until
1321 * cleanup without this code. Its not the
1322 * best way but I know of no other way to
1323 * handle it since the stack needs functions
1324 * it does not have to handle queued packets.
1325 */
1326 tcp_drop_in_pkts(tp);
1327 }
1328 if (in_pcbrele_wlocked(inp) == 0)
1329 INP_WUNLOCK(inp);
1330 INP_UNLOCK_ASSERT(inp);
1331 #ifdef VIMAGE
1332 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1333 CURVNET_RESTORE();
1334 #endif
1335 mtx_lock(&hpts->p_mtx);
1336 hpts->p_inp = NULL;
1337 }
1338 #ifndef VIMAGE
1339 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1340 INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
1341 #endif
1342 }
1343
1344 static void
1345 tcp_hptsi(struct tcp_hpts_entry *hpts)
1346 {
1347 struct epoch_tracker et;
1348 struct tcpcb *tp;
1349 struct inpcb *inp = NULL, *ninp;
1350 struct timeval tv;
1351 int32_t ticks_to_run, i, error;
1352 int32_t loop_cnt = 0;
1353 int32_t did_prefetch = 0;
1354 int32_t prefetch_ninp = 0;
1355 int32_t prefetch_tp = 0;
1356 int32_t wrap_loop_cnt = 0;
1357 int16_t set_cpu;
1358
1359 HPTS_MTX_ASSERT(hpts);
1360 /* record previous info for any logging */
1361 hpts->saved_lasttick = hpts->p_lasttick;
1362 hpts->saved_curtick = hpts->p_curtick;
1363 hpts->saved_curslot = hpts->p_cur_slot;
1364 hpts->saved_prev_slot = hpts->p_prev_slot;
1365
1366 hpts->p_lasttick = hpts->p_curtick;
1367 hpts->p_curtick = tcp_gethptstick(&tv);
1368 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1369 if ((hpts->p_on_queue_cnt == 0) ||
1370 (hpts->p_lasttick == hpts->p_curtick)) {
1371 /*
1372 * No time has yet passed,
1373 * or nothing to do.
1374 */
1375 hpts->p_prev_slot = hpts->p_cur_slot;
1376 hpts->p_lasttick = hpts->p_curtick;
1377 goto no_run;
1378 }
1379 again:
1380 hpts->p_wheel_complete = 0;
1381 HPTS_MTX_ASSERT(hpts);
1382 ticks_to_run = hpts_ticks_diff(hpts->p_prev_slot, hpts->p_cur_slot);
1383 if (((hpts->p_curtick - hpts->p_lasttick) > ticks_to_run) &&
1384 (hpts->p_on_queue_cnt != 0)) {
1385 /*
1386 * Wheel wrap is occuring, basically we
1387 * are behind and the distance between
1388 * run's has spread so much it has exceeded
1389 * the time on the wheel (1.024 seconds). This
1390 * is ugly and should NOT be happening. We
1391 * need to run the entire wheel. We last processed
1392 * p_prev_slot, so that needs to be the last slot
1393 * we run. The next slot after that should be our
1394 * reserved first slot for new, and then starts
1395 * the running position. Now the problem is the
1396 * reserved "not to yet" place does not exist
1397 * and there may be inp's in there that need
1398 * running. We can merge those into the
1399 * first slot at the head.
1400 */
1401 wrap_loop_cnt++;
1402 hpts->p_nxt_slot = hpts_tick(hpts->p_prev_slot, 1);
1403 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 2);
1404 /*
1405 * Adjust p_cur_slot to be where we are starting from
1406 * hopefully we will catch up (fat chance if something
1407 * is broken this bad :( )
1408 */
1409 hpts->p_cur_slot = hpts->p_prev_slot;
1410 /*
1411 * The next slot has guys to run too, and that would
1412 * be where we would normally start, lets move them into
1413 * the next slot (p_prev_slot + 2) so that we will
1414 * run them, the extra 10usecs of late (by being
1415 * put behind) does not really matter in this situation.
1416 */
1417 #ifdef INVARIANTS
1418 /*
1419 * To prevent a panic we need to update the inpslot to the
1420 * new location. This is safe since it takes both the
1421 * INP lock and the pacer mutex to change the inp_hptsslot.
1422 */
1423 TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts) {
1424 inp->inp_hptsslot = hpts->p_runningtick;
1425 }
1426 #endif
1427 TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningtick],
1428 &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts);
1429 ticks_to_run = NUM_OF_HPTSI_SLOTS - 1;
1430 counter_u64_add(wheel_wrap, 1);
1431 } else {
1432 /*
1433 * Nxt slot is always one after p_runningtick though
1434 * its not used usually unless we are doing wheel wrap.
1435 */
1436 hpts->p_nxt_slot = hpts->p_prev_slot;
1437 hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 1);
1438 }
1439 #ifdef INVARIANTS
1440 if (TAILQ_EMPTY(&hpts->p_input) &&
1441 (hpts->p_on_inqueue_cnt != 0)) {
1442 panic("tp:%p in_hpts input empty but cnt:%d",
1443 hpts, hpts->p_on_inqueue_cnt);
1444 }
1445 #endif
1446 HPTS_MTX_ASSERT(hpts);
1447 if (hpts->p_on_queue_cnt == 0) {
1448 goto no_one;
1449 }
1450 HPTS_MTX_ASSERT(hpts);
1451 #ifndef VIMAGE
1452 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1453 #endif
1454 for (i = 0; i < ticks_to_run; i++) {
1455 /*
1456 * Calculate our delay, if there are no extra ticks there
1457 * was not any (i.e. if ticks_to_run == 1, no delay).
1458 */
1459 hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC;
1460 HPTS_MTX_ASSERT(hpts);
1461 while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1462 /* For debugging */
1463 hpts->p_inp = inp;
1464 #ifdef INVARIANTS
1465 if (hpts->p_runningtick != inp->inp_hptsslot) {
1466 panic("Hpts:%p inp:%p slot mis-aligned %u vs %u",
1467 hpts, inp, hpts->p_runningtick, inp->inp_hptsslot);
1468 }
1469 #endif
1470 /* Now pull it */
1471 if (inp->inp_hpts_cpu_set == 0) {
1472 set_cpu = 1;
1473 } else {
1474 set_cpu = 0;
1475 }
1476 hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningtick], 0);
1477 if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) {
1478 /* We prefetch the next inp if possible */
1479 kern_prefetch(ninp, &prefetch_ninp);
1480 prefetch_ninp = 1;
1481 }
1482 if (inp->inp_hpts_request) {
1483 /*
1484 * This guy is deferred out further in time
1485 * then our wheel had available on it.
1486 * Push him back on the wheel or run it
1487 * depending.
1488 */
1489 uint32_t maxticks, last_tick, remaining_slots;
1490
1491 remaining_slots = ticks_to_run - (i + 1);
1492 if (inp->inp_hpts_request > remaining_slots) {
1493 /*
1494 * How far out can we go?
1495 */
1496 maxticks = max_ticks_available(hpts, hpts->p_cur_slot, &last_tick);
1497 if (maxticks >= inp->inp_hpts_request) {
1498 /* we can place it finally to be processed */
1499 inp->inp_hptsslot = hpts_tick(hpts->p_runningtick, inp->inp_hpts_request);
1500 inp->inp_hpts_request = 0;
1501 } else {
1502 /* Work off some more time */
1503 inp->inp_hptsslot = last_tick;
1504 inp->inp_hpts_request-= maxticks;
1505 }
1506 hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], __LINE__, 1);
1507 hpts->p_inp = NULL;
1508 continue;
1509 }
1510 inp->inp_hpts_request = 0;
1511 /* Fall through we will so do it now */
1512 }
1513 /*
1514 * We clear the hpts flag here after dealing with
1515 * remaining slots. This way anyone looking with the
1516 * TCB lock will see its on the hpts until just
1517 * before we unlock.
1518 */
1519 inp->inp_in_hpts = 0;
1520 mtx_unlock(&hpts->p_mtx);
1521 INP_WLOCK(inp);
1522 if (in_pcbrele_wlocked(inp)) {
1523 mtx_lock(&hpts->p_mtx);
1524 hpts->p_inp = NULL;
1525 continue;
1526 }
1527 if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) ||
1528 (inp->inp_flags2 & INP_FREED)) {
1529 out_now:
1530 #ifdef INVARIANTS
1531 if (mtx_owned(&hpts->p_mtx)) {
1532 panic("Hpts:%p owns mtx prior-to lock line:%d",
1533 hpts, __LINE__);
1534 }
1535 #endif
1536 INP_WUNLOCK(inp);
1537 mtx_lock(&hpts->p_mtx);
1538 hpts->p_inp = NULL;
1539 continue;
1540 }
1541 tp = intotcpcb(inp);
1542 if ((tp == NULL) || (tp->t_inpcb == NULL)) {
1543 goto out_now;
1544 }
1545 if (set_cpu) {
1546 /*
1547 * Setup so the next time we will move to
1548 * the right CPU. This should be a rare
1549 * event. It will sometimes happens when we
1550 * are the client side (usually not the
1551 * server). Somehow tcp_output() gets called
1552 * before the tcp_do_segment() sets the
1553 * intial state. This means the r_cpu and
1554 * r_hpts_cpu is 0. We get on the hpts, and
1555 * then tcp_input() gets called setting up
1556 * the r_cpu to the correct value. The hpts
1557 * goes off and sees the mis-match. We
1558 * simply correct it here and the CPU will
1559 * switch to the new hpts nextime the tcb
1560 * gets added to the the hpts (not this one)
1561 * :-)
1562 */
1563 tcp_set_hpts(inp);
1564 }
1565 #ifdef VIMAGE
1566 CURVNET_SET(inp->inp_vnet);
1567 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1568 #endif
1569 /* Lets do any logging that we might want to */
1570 if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
1571 tcp_hpts_log(hpts, tp, &tv, ticks_to_run, i);
1572 }
1573 /*
1574 * There is a hole here, we get the refcnt on the
1575 * inp so it will still be preserved but to make
1576 * sure we can get the INP we need to hold the p_mtx
1577 * above while we pull out the tp/inp, as long as
1578 * fini gets the lock first we are assured of having
1579 * a sane INP we can lock and test.
1580 */
1581 #ifdef INVARIANTS
1582 if (mtx_owned(&hpts->p_mtx)) {
1583 panic("Hpts:%p owns mtx before tcp-output:%d",
1584 hpts, __LINE__);
1585 }
1586 #endif
1587 if (tp->t_fb_ptr != NULL) {
1588 kern_prefetch(tp->t_fb_ptr, &did_prefetch);
1589 did_prefetch = 1;
1590 }
1591 if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) {
1592 error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0);
1593 if (error) {
1594 /* The input killed the connection */
1595 goto skip_pacing;
1596 }
1597 }
1598 inp->inp_hpts_calls = 1;
1599 error = tp->t_fb->tfb_tcp_output(tp);
1600 inp->inp_hpts_calls = 0;
1601 if (ninp && ninp->inp_ppcb) {
1602 /*
1603 * If we have a nxt inp, see if we can
1604 * prefetch its ppcb. Note this may seem
1605 * "risky" since we have no locks (other
1606 * than the previous inp) and there no
1607 * assurance that ninp was not pulled while
1608 * we were processing inp and freed. If this
1609 * occured it could mean that either:
1610 *
1611 * a) Its NULL (which is fine we won't go
1612 * here) <or> b) Its valid (which is cool we
1613 * will prefetch it) <or> c) The inp got
1614 * freed back to the slab which was
1615 * reallocated. Then the piece of memory was
1616 * re-used and something else (not an
1617 * address) is in inp_ppcb. If that occurs
1618 * we don't crash, but take a TLB shootdown
1619 * performance hit (same as if it was NULL
1620 * and we tried to pre-fetch it).
1621 *
1622 * Considering that the likelyhood of <c> is
1623 * quite rare we will take a risk on doing
1624 * this. If performance drops after testing
1625 * we can always take this out. NB: the
1626 * kern_prefetch on amd64 actually has
1627 * protection against a bad address now via
1628 * the DMAP_() tests. This will prevent the
1629 * TLB hit, and instead if <c> occurs just
1630 * cause us to load cache with a useless
1631 * address (to us).
1632 */
1633 kern_prefetch(ninp->inp_ppcb, &prefetch_tp);
1634 prefetch_tp = 1;
1635 }
1636 INP_WUNLOCK(inp);
1637 skip_pacing:
1638 #ifdef VIMAGE
1639 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1640 CURVNET_RESTORE();
1641 #endif
1642 INP_UNLOCK_ASSERT(inp);
1643 #ifdef INVARIANTS
1644 if (mtx_owned(&hpts->p_mtx)) {
1645 panic("Hpts:%p owns mtx prior-to lock line:%d",
1646 hpts, __LINE__);
1647 }
1648 #endif
1649 mtx_lock(&hpts->p_mtx);
1650 hpts->p_inp = NULL;
1651 }
1652 HPTS_MTX_ASSERT(hpts);
1653 hpts->p_inp = NULL;
1654 hpts->p_runningtick++;
1655 if (hpts->p_runningtick >= NUM_OF_HPTSI_SLOTS) {
1656 hpts->p_runningtick = 0;
1657 }
1658 }
1659 #ifndef VIMAGE
1660 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1661 #endif
1662 no_one:
1663 HPTS_MTX_ASSERT(hpts);
1664 hpts->p_delayed_by = 0;
1665 /*
1666 * Check to see if we took an excess amount of time and need to run
1667 * more ticks (if we did not hit eno-bufs).
1668 */
1669 #ifdef INVARIANTS
1670 if (TAILQ_EMPTY(&hpts->p_input) &&
1671 (hpts->p_on_inqueue_cnt != 0)) {
1672 panic("tp:%p in_hpts input empty but cnt:%d",
1673 hpts, hpts->p_on_inqueue_cnt);
1674 }
1675 #endif
1676 hpts->p_prev_slot = hpts->p_cur_slot;
1677 hpts->p_lasttick = hpts->p_curtick;
1678 if (loop_cnt > max_pacer_loops) {
1679 /*
1680 * Something is serious slow we have
1681 * looped through processing the wheel
1682 * and by the time we cleared the
1683 * needs to run max_pacer_loops time
1684 * we still needed to run. That means
1685 * the system is hopelessly behind and
1686 * can never catch up :(
1687 *
1688 * We will just lie to this thread
1689 * and let it thing p_curtick is
1690 * correct. When it next awakens
1691 * it will find itself further behind.
1692 */
1693 counter_u64_add(hpts_hopelessly_behind, 1);
1694 goto no_run;
1695 }
1696 hpts->p_curtick = tcp_gethptstick(&tv);
1697 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1698 if ((wrap_loop_cnt < 2) &&
1699 (hpts->p_lasttick != hpts->p_curtick)) {
1700 counter_u64_add(hpts_loops, 1);
1701 loop_cnt++;
1702 goto again;
1703 }
1704 no_run:
1705 /*
1706 * Set flag to tell that we are done for
1707 * any slot input that happens during
1708 * input.
1709 */
1710 hpts->p_wheel_complete = 1;
1711 /*
1712 * Run any input that may be there not covered
1713 * in running data.
1714 */
1715 if (!TAILQ_EMPTY(&hpts->p_input)) {
1716 tcp_input_data(hpts, &tv);
1717 /*
1718 * Now did we spend too long running
1719 * input and need to run more ticks?
1720 */
1721 KASSERT(hpts->p_prev_slot == hpts->p_cur_slot,
1722 ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts,
1723 hpts->p_prev_slot, hpts->p_cur_slot));
1724 KASSERT(hpts->p_lasttick == hpts->p_curtick,
1725 ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts,
1726 hpts->p_lasttick, hpts->p_curtick));
1727 hpts->p_curtick = tcp_gethptstick(&tv);
1728 if (hpts->p_lasttick != hpts->p_curtick) {
1729 counter_u64_add(hpts_loops, 1);
1730 hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1731 goto again;
1732 }
1733 }
1734 {
1735 uint32_t t = 0, i, fnd = 0;
1736
1737 if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) {
1738 /*
1739 * Find next slot that is occupied and use that to
1740 * be the sleep time.
1741 */
1742 for (i = 0, t = hpts_tick(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) {
1743 if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) {
1744 fnd = 1;
1745 break;
1746 }
1747 t = (t + 1) % NUM_OF_HPTSI_SLOTS;
1748 }
1749 if (fnd) {
1750 hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max);
1751 } else {
1752 #ifdef INVARIANTS
1753 panic("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt);
1754 #endif
1755 counter_u64_add(back_tosleep, 1);
1756 hpts->p_on_queue_cnt = 0;
1757 goto non_found;
1758 }
1759 } else if (wrap_loop_cnt >= 2) {
1760 /* Special case handling */
1761 hpts->p_hpts_sleep_time = tcp_min_hptsi_time;
1762 } else {
1763 /* No one on the wheel sleep for all but 400 slots or sleep max */
1764 non_found:
1765 hpts->p_hpts_sleep_time = hpts_sleep_max;
1766 }
1767 }
1768 }
1769
1770 void
1771 __tcp_set_hpts(struct inpcb *inp, int32_t line)
1772 {
1773 struct tcp_hpts_entry *hpts;
1774
1775 INP_WLOCK_ASSERT(inp);
1776 hpts = tcp_hpts_lock(inp);
1777 if ((inp->inp_in_hpts == 0) &&
1778 (inp->inp_hpts_cpu_set == 0)) {
1779 inp->inp_hpts_cpu = hpts_cpuid(inp);
1780 inp->inp_hpts_cpu_set = 1;
1781 }
1782 mtx_unlock(&hpts->p_mtx);
1783 hpts = tcp_input_lock(inp);
1784 if ((inp->inp_input_cpu_set == 0) &&
1785 (inp->inp_in_input == 0)) {
1786 inp->inp_input_cpu = hpts_cpuid(inp);
1787 inp->inp_input_cpu_set = 1;
1788 }
1789 mtx_unlock(&hpts->p_mtx);
1790 }
1791
1792 uint16_t
1793 tcp_hpts_delayedby(struct inpcb *inp){
1794 return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by);
1795 }
1796
1797 static void
1798 tcp_hpts_thread(void *ctx)
1799 {
1800 struct tcp_hpts_entry *hpts;
1801 struct timeval tv;
1802 sbintime_t sb;
1803
1804 hpts = (struct tcp_hpts_entry *)ctx;
1805 mtx_lock(&hpts->p_mtx);
1806 if (hpts->p_direct_wake) {
1807 /* Signaled by input */
1808 callout_stop(&hpts->co);
1809 } else {
1810 /* Timed out */
1811 if (callout_pending(&hpts->co) ||
1812 !callout_active(&hpts->co)) {
1813 mtx_unlock(&hpts->p_mtx);
1814 return;
1815 }
1816 callout_deactivate(&hpts->co);
1817 }
1818 hpts->p_hpts_wake_scheduled = 0;
1819 hpts->p_hpts_active = 1;
1820 tcp_hptsi(hpts);
1821 HPTS_MTX_ASSERT(hpts);
1822 tv.tv_sec = 0;
1823 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1824 if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) {
1825 hpts->overidden_sleep = tv.tv_usec;
1826 tv.tv_usec = tcp_min_hptsi_time;
1827 hpts->p_on_min_sleep = 1;
1828 } else {
1829 /* Clear the min sleep flag */
1830 hpts->overidden_sleep = 0;
1831 hpts->p_on_min_sleep = 0;
1832 }
1833 hpts->p_hpts_active = 0;
1834 sb = tvtosbt(tv);
1835 if (tcp_hpts_callout_skip_swi == 0) {
1836 callout_reset_sbt_on(&hpts->co, sb, 0,
1837 hpts_timeout_swi, hpts, hpts->p_cpu,
1838 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1839 } else {
1840 callout_reset_sbt_on(&hpts->co, sb, 0,
1841 hpts_timeout_dir, hpts,
1842 hpts->p_cpu,
1843 C_PREL(tcp_hpts_precision));
1844 }
1845 hpts->p_direct_wake = 0;
1846 mtx_unlock(&hpts->p_mtx);
1847 }
1848
1849 #undef timersub
1850
1851 static void
1852 tcp_init_hptsi(void *st)
1853 {
1854 int32_t i, j, error, bound = 0, created = 0;
1855 size_t sz, asz;
1856 struct timeval tv;
1857 sbintime_t sb;
1858 struct tcp_hpts_entry *hpts;
1859 char unit[16];
1860 uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1861
1862 tcp_pace.rp_proc = NULL;
1863 tcp_pace.rp_num_hptss = ncpus;
1864 hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK);
1865 hpts_loops = counter_u64_alloc(M_WAITOK);
1866 back_tosleep = counter_u64_alloc(M_WAITOK);
1867 combined_wheel_wrap = counter_u64_alloc(M_WAITOK);
1868 wheel_wrap = counter_u64_alloc(M_WAITOK);
1869 sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *));
1870 tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO);
1871 asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS;
1872 for (i = 0; i < tcp_pace.rp_num_hptss; i++) {
1873 tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry),
1874 M_TCPHPTS, M_WAITOK | M_ZERO);
1875 tcp_pace.rp_ent[i]->p_hptss = malloc(asz,
1876 M_TCPHPTS, M_WAITOK);
1877 hpts = tcp_pace.rp_ent[i];
1878 /*
1879 * Init all the hpts structures that are not specifically
1880 * zero'd by the allocations. Also lets attach them to the
1881 * appropriate sysctl block as well.
1882 */
1883 mtx_init(&hpts->p_mtx, "tcp_hpts_lck",
1884 "hpts", MTX_DEF | MTX_DUPOK);
1885 TAILQ_INIT(&hpts->p_input);
1886 for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) {
1887 TAILQ_INIT(&hpts->p_hptss[j]);
1888 }
1889 sysctl_ctx_init(&hpts->hpts_ctx);
1890 sprintf(unit, "%d", i);
1891 hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx,
1892 SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts),
1893 OID_AUTO,
1894 unit,
1895 CTLFLAG_RW, 0,
1896 "");
1897 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1898 SYSCTL_CHILDREN(hpts->hpts_root),
1899 OID_AUTO, "in_qcnt", CTLFLAG_RD,
1900 &hpts->p_on_inqueue_cnt, 0,
1901 "Count TCB's awaiting input processing");
1902 SYSCTL_ADD_INT(&hpts->hpts_ctx,
1903 SYSCTL_CHILDREN(hpts->hpts_root),
1904 OID_AUTO, "out_qcnt", CTLFLAG_RD,
1905 &hpts->p_on_queue_cnt, 0,
1906 "Count TCB's awaiting output processing");
1907 SYSCTL_ADD_U16(&hpts->hpts_ctx,
1908 SYSCTL_CHILDREN(hpts->hpts_root),
1909 OID_AUTO, "active", CTLFLAG_RD,
1910 &hpts->p_hpts_active, 0,
1911 "Is the hpts active");
1912 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1913 SYSCTL_CHILDREN(hpts->hpts_root),
1914 OID_AUTO, "curslot", CTLFLAG_RD,
1915 &hpts->p_cur_slot, 0,
1916 "What the current running pacers goal");
1917 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1918 SYSCTL_CHILDREN(hpts->hpts_root),
1919 OID_AUTO, "runtick", CTLFLAG_RD,
1920 &hpts->p_runningtick, 0,
1921 "What the running pacers current slot is");
1922 SYSCTL_ADD_UINT(&hpts->hpts_ctx,
1923 SYSCTL_CHILDREN(hpts->hpts_root),
1924 OID_AUTO, "curtick", CTLFLAG_RD,
1925 &hpts->p_curtick, 0,
1926 "What the running pacers last tick mapped to the wheel was");
1927 hpts->p_hpts_sleep_time = hpts_sleep_max;
1928 hpts->p_num = i;
1929 hpts->p_curtick = tcp_gethptstick(&tv);
1930 hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick);
1931 hpts->p_cpu = 0xffff;
1932 hpts->p_nxt_slot = hpts_tick(hpts->p_cur_slot, 1);
1933 callout_init(&hpts->co, 1);
1934 }
1935 /*
1936 * Now lets start ithreads to handle the hptss.
1937 */
1938 CPU_FOREACH(i) {
1939 hpts = tcp_pace.rp_ent[i];
1940 hpts->p_cpu = i;
1941 error = swi_add(&hpts->ie, "hpts",
1942 tcp_hpts_thread, (void *)hpts,
1943 SWI_NET, INTR_MPSAFE, &hpts->ie_cookie);
1944 if (error) {
1945 panic("Can't add hpts:%p i:%d err:%d",
1946 hpts, i, error);
1947 }
1948 created++;
1949 if (tcp_bind_threads) {
1950 if (intr_event_bind(hpts->ie, i) == 0)
1951 bound++;
1952 }
1953 tv.tv_sec = 0;
1954 tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC;
1955 sb = tvtosbt(tv);
1956 if (tcp_hpts_callout_skip_swi == 0) {
1957 callout_reset_sbt_on(&hpts->co, sb, 0,
1958 hpts_timeout_swi, hpts, hpts->p_cpu,
1959 (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision)));
1960 } else {
1961 callout_reset_sbt_on(&hpts->co, sb, 0,
1962 hpts_timeout_dir, hpts,
1963 hpts->p_cpu,
1964 C_PREL(tcp_hpts_precision));
1965 }
1966 }
1967 printf("TCP Hpts created %d swi interrupt thread and bound %d\n",
1968 created, bound);
1969 return;
1970 }
1971
1972 SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL);
1973 MODULE_VERSION(tcphpts, 1);
Cache object: 3cefd9d64f358190578cacd8e04946f6
|