FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_poll.c
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2001-2002 Luigi Rizzo
5 *
6 * Supported by: the Xorp Project (www.xorp.org)
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD: releng/12.0/sys/kern/kern_poll.c 326271 2017-11-27 15:20:12Z pfg $");
32
33 #include "opt_device_polling.h"
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/proc.h>
40 #include <sys/eventhandler.h>
41 #include <sys/resourcevar.h>
42 #include <sys/socket.h> /* needed by net/if.h */
43 #include <sys/sockio.h>
44 #include <sys/sysctl.h>
45 #include <sys/syslog.h>
46
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/netisr.h> /* for NETISR_POLL */
50 #include <net/vnet.h>
51
52 void hardclock_device_poll(void); /* hook from hardclock */
53
54 static struct mtx poll_mtx;
55
56 /*
57 * Polling support for [network] device drivers.
58 *
59 * Drivers which support this feature can register with the
60 * polling code.
61 *
62 * If registration is successful, the driver must disable interrupts,
63 * and further I/O is performed through the handler, which is invoked
64 * (at least once per clock tick) with 3 arguments: the "arg" passed at
65 * register time (a struct ifnet pointer), a command, and a "count" limit.
66 *
67 * The command can be one of the following:
68 * POLL_ONLY: quick move of "count" packets from input/output queues.
69 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
70 * other more expensive operations. This command is issued periodically
71 * but less frequently than POLL_ONLY.
72 *
73 * The count limit specifies how much work the handler can do during the
74 * call -- typically this is the number of packets to be received, or
75 * transmitted, etc. (drivers are free to interpret this number, as long
76 * as the max time spent in the function grows roughly linearly with the
77 * count).
78 *
79 * Polling is enabled and disabled via setting IFCAP_POLLING flag on
80 * the interface. The driver ioctl handler should register interface
81 * with polling and disable interrupts, if registration was successful.
82 *
83 * A second variable controls the sharing of CPU between polling/kernel
84 * network processing, and other activities (typically userlevel tasks):
85 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
86 * of CPU allocated to user tasks. CPU is allocated proportionally to the
87 * shares, by dynamically adjusting the "count" (poll_burst).
88 *
89 * Other parameters can should be left to their default values.
90 * The following constraints hold
91 *
92 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
93 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
94 */
95
96 #define MIN_POLL_BURST_MAX 10
97 #define MAX_POLL_BURST_MAX 20000
98
99 static uint32_t poll_burst = 5;
100 static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
101 static uint32_t poll_each_burst = 5;
102
103 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
104 "Device polling parameters");
105
106 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
107 &poll_burst, 0, "Current polling burst size");
108
109 static int netisr_poll_scheduled;
110 static int netisr_pollmore_scheduled;
111 static int poll_shutting_down;
112
113 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
114 {
115 uint32_t val = poll_burst_max;
116 int error;
117
118 error = sysctl_handle_int(oidp, &val, 0, req);
119 if (error || !req->newptr )
120 return (error);
121 if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
122 return (EINVAL);
123
124 mtx_lock(&poll_mtx);
125 poll_burst_max = val;
126 if (poll_burst > poll_burst_max)
127 poll_burst = poll_burst_max;
128 if (poll_each_burst > poll_burst_max)
129 poll_each_burst = MIN_POLL_BURST_MAX;
130 mtx_unlock(&poll_mtx);
131
132 return (0);
133 }
134 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
135 0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
136
137 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
138 {
139 uint32_t val = poll_each_burst;
140 int error;
141
142 error = sysctl_handle_int(oidp, &val, 0, req);
143 if (error || !req->newptr )
144 return (error);
145 if (val < 1)
146 return (EINVAL);
147
148 mtx_lock(&poll_mtx);
149 if (val > poll_burst_max) {
150 mtx_unlock(&poll_mtx);
151 return (EINVAL);
152 }
153 poll_each_burst = val;
154 mtx_unlock(&poll_mtx);
155
156 return (0);
157 }
158 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
159 0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
160 "Max size of each burst");
161
162 static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */
163 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
164 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
165
166 static uint32_t user_frac = 50;
167 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
168 {
169 uint32_t val = user_frac;
170 int error;
171
172 error = sysctl_handle_int(oidp, &val, 0, req);
173 if (error || !req->newptr )
174 return (error);
175 if (val > 99)
176 return (EINVAL);
177
178 mtx_lock(&poll_mtx);
179 user_frac = val;
180 mtx_unlock(&poll_mtx);
181
182 return (0);
183 }
184 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
185 0, sizeof(uint32_t), user_frac_sysctl, "I",
186 "Desired user fraction of cpu time");
187
188 static uint32_t reg_frac_count = 0;
189 static uint32_t reg_frac = 20 ;
190 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
191 {
192 uint32_t val = reg_frac;
193 int error;
194
195 error = sysctl_handle_int(oidp, &val, 0, req);
196 if (error || !req->newptr )
197 return (error);
198 if (val < 1 || val > hz)
199 return (EINVAL);
200
201 mtx_lock(&poll_mtx);
202 reg_frac = val;
203 if (reg_frac_count >= reg_frac)
204 reg_frac_count = 0;
205 mtx_unlock(&poll_mtx);
206
207 return (0);
208 }
209 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
210 0, sizeof(uint32_t), reg_frac_sysctl, "I",
211 "Every this many cycles check registers");
212
213 static uint32_t short_ticks;
214 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
215 &short_ticks, 0, "Hardclock ticks shorter than they should be");
216
217 static uint32_t lost_polls;
218 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
219 &lost_polls, 0, "How many times we would have lost a poll tick");
220
221 static uint32_t pending_polls;
222 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
223 &pending_polls, 0, "Do we need to poll again");
224
225 static int residual_burst = 0;
226 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
227 &residual_burst, 0, "# of residual cycles in burst");
228
229 static uint32_t poll_handlers; /* next free entry in pr[]. */
230 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
231 &poll_handlers, 0, "Number of registered poll handlers");
232
233 static uint32_t phase;
234 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
235 &phase, 0, "Polling phase");
236
237 static uint32_t suspect;
238 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
239 &suspect, 0, "suspect event");
240
241 static uint32_t stalled;
242 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
243 &stalled, 0, "potential stalls");
244
245 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
246 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
247 &idlepoll_sleeping, 0, "idlepoll is sleeping");
248
249
250 #define POLL_LIST_LEN 128
251 struct pollrec {
252 poll_handler_t *handler;
253 struct ifnet *ifp;
254 };
255
256 static struct pollrec pr[POLL_LIST_LEN];
257
258 static void
259 poll_shutdown(void *arg, int howto)
260 {
261
262 poll_shutting_down = 1;
263 }
264
265 static void
266 init_device_poll(void)
267 {
268
269 mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
270 EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
271 SHUTDOWN_PRI_LAST);
272 }
273 SYSINIT(device_poll, SI_SUB_SOFTINTR, SI_ORDER_MIDDLE, init_device_poll, NULL);
274
275
276 /*
277 * Hook from hardclock. Tries to schedule a netisr, but keeps track
278 * of lost ticks due to the previous handler taking too long.
279 * Normally, this should not happen, because polling handler should
280 * run for a short time. However, in some cases (e.g. when there are
281 * changes in link status etc.) the drivers take a very long time
282 * (even in the order of milliseconds) to reset and reconfigure the
283 * device, causing apparent lost polls.
284 *
285 * The first part of the code is just for debugging purposes, and tries
286 * to count how often hardclock ticks are shorter than they should,
287 * meaning either stray interrupts or delayed events.
288 */
289 void
290 hardclock_device_poll(void)
291 {
292 static struct timeval prev_t, t;
293 int delta;
294
295 if (poll_handlers == 0 || poll_shutting_down)
296 return;
297
298 microuptime(&t);
299 delta = (t.tv_usec - prev_t.tv_usec) +
300 (t.tv_sec - prev_t.tv_sec)*1000000;
301 if (delta * hz < 500000)
302 short_ticks++;
303 else
304 prev_t = t;
305
306 if (pending_polls > 100) {
307 /*
308 * Too much, assume it has stalled (not always true
309 * see comment above).
310 */
311 stalled++;
312 pending_polls = 0;
313 phase = 0;
314 }
315
316 if (phase <= 2) {
317 if (phase != 0)
318 suspect++;
319 phase = 1;
320 netisr_poll_scheduled = 1;
321 netisr_pollmore_scheduled = 1;
322 netisr_sched_poll();
323 phase = 2;
324 }
325 if (pending_polls++ > 0)
326 lost_polls++;
327 }
328
329 /*
330 * ether_poll is called from the idle loop.
331 */
332 static void
333 ether_poll(int count)
334 {
335 int i;
336
337 mtx_lock(&poll_mtx);
338
339 if (count > poll_each_burst)
340 count = poll_each_burst;
341
342 for (i = 0 ; i < poll_handlers ; i++)
343 pr[i].handler(pr[i].ifp, POLL_ONLY, count);
344
345 mtx_unlock(&poll_mtx);
346 }
347
348 /*
349 * netisr_pollmore is called after other netisr's, possibly scheduling
350 * another NETISR_POLL call, or adapting the burst size for the next cycle.
351 *
352 * It is very bad to fetch large bursts of packets from a single card at once,
353 * because the burst could take a long time to be completely processed, or
354 * could saturate the intermediate queue (ipintrq or similar) leading to
355 * losses or unfairness. To reduce the problem, and also to account better for
356 * time spent in network-related processing, we split the burst in smaller
357 * chunks of fixed size, giving control to the other netisr's between chunks.
358 * This helps in improving the fairness, reducing livelock (because we
359 * emulate more closely the "process to completion" that we have with
360 * fastforwarding) and accounting for the work performed in low level
361 * handling and forwarding.
362 */
363
364 static struct timeval poll_start_t;
365
366 void
367 netisr_pollmore()
368 {
369 struct timeval t;
370 int kern_load;
371
372 if (poll_handlers == 0)
373 return;
374
375 mtx_lock(&poll_mtx);
376 if (!netisr_pollmore_scheduled) {
377 mtx_unlock(&poll_mtx);
378 return;
379 }
380 netisr_pollmore_scheduled = 0;
381 phase = 5;
382 if (residual_burst > 0) {
383 netisr_poll_scheduled = 1;
384 netisr_pollmore_scheduled = 1;
385 netisr_sched_poll();
386 mtx_unlock(&poll_mtx);
387 /* will run immediately on return, followed by netisrs */
388 return;
389 }
390 /* here we can account time spent in netisr's in this tick */
391 microuptime(&t);
392 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
393 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
394 kern_load = (kern_load * hz) / 10000; /* 0..100 */
395 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
396 if (poll_burst > 1)
397 poll_burst--;
398 } else {
399 if (poll_burst < poll_burst_max)
400 poll_burst++;
401 }
402
403 pending_polls--;
404 if (pending_polls == 0) /* we are done */
405 phase = 0;
406 else {
407 /*
408 * Last cycle was long and caused us to miss one or more
409 * hardclock ticks. Restart processing again, but slightly
410 * reduce the burst size to prevent that this happens again.
411 */
412 poll_burst -= (poll_burst / 8);
413 if (poll_burst < 1)
414 poll_burst = 1;
415 netisr_poll_scheduled = 1;
416 netisr_pollmore_scheduled = 1;
417 netisr_sched_poll();
418 phase = 6;
419 }
420 mtx_unlock(&poll_mtx);
421 }
422
423 /*
424 * netisr_poll is typically scheduled once per tick.
425 */
426 void
427 netisr_poll(void)
428 {
429 int i, cycles;
430 enum poll_cmd arg = POLL_ONLY;
431
432 if (poll_handlers == 0)
433 return;
434
435 mtx_lock(&poll_mtx);
436 if (!netisr_poll_scheduled) {
437 mtx_unlock(&poll_mtx);
438 return;
439 }
440 netisr_poll_scheduled = 0;
441 phase = 3;
442 if (residual_burst == 0) { /* first call in this tick */
443 microuptime(&poll_start_t);
444 if (++reg_frac_count == reg_frac) {
445 arg = POLL_AND_CHECK_STATUS;
446 reg_frac_count = 0;
447 }
448
449 residual_burst = poll_burst;
450 }
451 cycles = (residual_burst < poll_each_burst) ?
452 residual_burst : poll_each_burst;
453 residual_burst -= cycles;
454
455 for (i = 0 ; i < poll_handlers ; i++)
456 pr[i].handler(pr[i].ifp, arg, cycles);
457
458 phase = 4;
459 mtx_unlock(&poll_mtx);
460 }
461
462 /*
463 * Try to register routine for polling. Returns 0 if successful
464 * (and polling should be enabled), error code otherwise.
465 * A device is not supposed to register itself multiple times.
466 *
467 * This is called from within the *_ioctl() functions.
468 */
469 int
470 ether_poll_register(poll_handler_t *h, if_t ifp)
471 {
472 int i;
473
474 KASSERT(h != NULL, ("%s: handler is NULL", __func__));
475 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
476
477 mtx_lock(&poll_mtx);
478 if (poll_handlers >= POLL_LIST_LEN) {
479 /*
480 * List full, cannot register more entries.
481 * This should never happen; if it does, it is probably a
482 * broken driver trying to register multiple times. Checking
483 * this at runtime is expensive, and won't solve the problem
484 * anyways, so just report a few times and then give up.
485 */
486 static int verbose = 10 ;
487 if (verbose >0) {
488 log(LOG_ERR, "poll handlers list full, "
489 "maybe a broken driver ?\n");
490 verbose--;
491 }
492 mtx_unlock(&poll_mtx);
493 return (ENOMEM); /* no polling for you */
494 }
495
496 for (i = 0 ; i < poll_handlers ; i++)
497 if (pr[i].ifp == ifp && pr[i].handler != NULL) {
498 mtx_unlock(&poll_mtx);
499 log(LOG_DEBUG, "ether_poll_register: %s: handler"
500 " already registered\n", ifp->if_xname);
501 return (EEXIST);
502 }
503
504 pr[poll_handlers].handler = h;
505 pr[poll_handlers].ifp = ifp;
506 poll_handlers++;
507 mtx_unlock(&poll_mtx);
508 if (idlepoll_sleeping)
509 wakeup(&idlepoll_sleeping);
510 return (0);
511 }
512
513 /*
514 * Remove interface from the polling list. Called from *_ioctl(), too.
515 */
516 int
517 ether_poll_deregister(if_t ifp)
518 {
519 int i;
520
521 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
522
523 mtx_lock(&poll_mtx);
524
525 for (i = 0 ; i < poll_handlers ; i++)
526 if (pr[i].ifp == ifp) /* found it */
527 break;
528 if (i == poll_handlers) {
529 log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
530 ifp->if_xname);
531 mtx_unlock(&poll_mtx);
532 return (ENOENT);
533 }
534 poll_handlers--;
535 if (i < poll_handlers) { /* Last entry replaces this one. */
536 pr[i].handler = pr[poll_handlers].handler;
537 pr[i].ifp = pr[poll_handlers].ifp;
538 }
539 mtx_unlock(&poll_mtx);
540 return (0);
541 }
542
543 static void
544 poll_idle(void)
545 {
546 struct thread *td = curthread;
547 struct rtprio rtp;
548
549 rtp.prio = RTP_PRIO_MAX; /* lowest priority */
550 rtp.type = RTP_PRIO_IDLE;
551 PROC_SLOCK(td->td_proc);
552 rtp_to_pri(&rtp, td);
553 PROC_SUNLOCK(td->td_proc);
554
555 for (;;) {
556 if (poll_in_idle_loop && poll_handlers > 0) {
557 idlepoll_sleeping = 0;
558 ether_poll(poll_each_burst);
559 thread_lock(td);
560 mi_switch(SW_VOL, NULL);
561 thread_unlock(td);
562 } else {
563 idlepoll_sleeping = 1;
564 tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
565 }
566 }
567 }
568
569 static struct proc *idlepoll;
570 static struct kproc_desc idlepoll_kp = {
571 "idlepoll",
572 poll_idle,
573 &idlepoll
574 };
575 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
576 &idlepoll_kp);
Cache object: 7ef29b9ea412a72f3c405b4988f14f39
|