xref: /freebsd/sys/kern/kern_poll.c (revision 195ebc7e9e4b129de810833791a19dfb4349d6a9)
1 /*-
2  * Copyright (c) 2001-2002 Luigi Rizzo
3  *
4  * Supported by: the Xorp Project (www.xorp.org)
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_route.h"
32 #include "opt_device_polling.h"
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/kthread.h>
38 #include <sys/proc.h>
39 #include <sys/eventhandler.h>
40 #include <sys/resourcevar.h>
41 #include <sys/socket.h>			/* needed by net/if.h		*/
42 #include <sys/sockio.h>
43 #include <sys/sysctl.h>
44 #include <sys/syslog.h>
45 #include <sys/vimage.h>
46 
47 #include <net/if.h>			/* for IFF_* flags		*/
48 #include <net/netisr.h>			/* for NETISR_POLL		*/
49 #include <net/route.h>
50 #include <net/vnet.h>
51 
52 static int poll_switch(SYSCTL_HANDLER_ARGS);
53 
54 void hardclock_device_poll(void);	/* hook from hardclock		*/
55 
56 static struct mtx	poll_mtx;
57 
58 /*
59  * Polling support for [network] device drivers.
60  *
61  * Drivers which support this feature can register with the
62  * polling code.
63  *
64  * If registration is successful, the driver must disable interrupts,
65  * and further I/O is performed through the handler, which is invoked
66  * (at least once per clock tick) with 3 arguments: the "arg" passed at
67  * register time (a struct ifnet pointer), a command, and a "count" limit.
68  *
69  * The command can be one of the following:
70  *  POLL_ONLY: quick move of "count" packets from input/output queues.
71  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
72  *	other more expensive operations. This command is issued periodically
73  *	but less frequently than POLL_ONLY.
74  *
75  * The count limit specifies how much work the handler can do during the
76  * call -- typically this is the number of packets to be received, or
77  * transmitted, etc. (drivers are free to interpret this number, as long
78  * as the max time spent in the function grows roughly linearly with the
79  * count).
80  *
81  * Polling is enabled and disabled via setting IFCAP_POLLING flag on
82  * the interface. The driver ioctl handler should register interface
83  * with polling and disable interrupts, if registration was successful.
84  *
85  * A second variable controls the sharing of CPU between polling/kernel
86  * network processing, and other activities (typically userlevel tasks):
87  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
88  * of CPU allocated to user tasks. CPU is allocated proportionally to the
89  * shares, by dynamically adjusting the "count" (poll_burst).
90  *
91  * Other parameters can should be left to their default values.
92  * The following constraints hold
93  *
94  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
95  *	0 <= poll_each_burst
96  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
97  */
98 
99 #define MIN_POLL_BURST_MAX	10
100 #define MAX_POLL_BURST_MAX	1000
101 
102 static uint32_t poll_burst = 5;
103 static uint32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
104 static uint32_t poll_each_burst = 5;
105 
106 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
107 	"Device polling parameters");
108 
109 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
110 	&poll_burst, 0, "Current polling burst size");
111 
112 static int	netisr_poll_scheduled;
113 static int	netisr_pollmore_scheduled;
114 static int	poll_shutting_down;
115 
116 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
117 {
118 	uint32_t val = poll_burst_max;
119 	int error;
120 
121 	error = sysctl_handle_int(oidp, &val, 0, req);
122 	if (error || !req->newptr )
123 		return (error);
124 	if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
125 		return (EINVAL);
126 
127 	mtx_lock(&poll_mtx);
128 	poll_burst_max = val;
129 	if (poll_burst > poll_burst_max)
130 		poll_burst = poll_burst_max;
131 	if (poll_each_burst > poll_burst_max)
132 		poll_each_burst = MIN_POLL_BURST_MAX;
133 	mtx_unlock(&poll_mtx);
134 
135 	return (0);
136 }
137 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
138 	0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
139 
140 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
141 {
142 	uint32_t val = poll_each_burst;
143 	int error;
144 
145 	error = sysctl_handle_int(oidp, &val, 0, req);
146 	if (error || !req->newptr )
147 		return (error);
148 	if (val < 1)
149 		return (EINVAL);
150 
151 	mtx_lock(&poll_mtx);
152 	if (val > poll_burst_max) {
153 		mtx_unlock(&poll_mtx);
154 		return (EINVAL);
155 	}
156 	poll_each_burst = val;
157 	mtx_unlock(&poll_mtx);
158 
159 	return (0);
160 }
161 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
162 	0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
163 	"Max size of each burst");
164 
165 static uint32_t poll_in_idle_loop=0;	/* do we poll in idle loop ? */
166 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
167 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
168 
169 static uint32_t user_frac = 50;
170 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
171 {
172 	uint32_t val = user_frac;
173 	int error;
174 
175 	error = sysctl_handle_int(oidp, &val, 0, req);
176 	if (error || !req->newptr )
177 		return (error);
178 	if (val < 0 || val > 99)
179 		return (EINVAL);
180 
181 	mtx_lock(&poll_mtx);
182 	user_frac = val;
183 	mtx_unlock(&poll_mtx);
184 
185 	return (0);
186 }
187 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
188 	0, sizeof(uint32_t), user_frac_sysctl, "I",
189 	"Desired user fraction of cpu time");
190 
191 static uint32_t reg_frac_count = 0;
192 static uint32_t reg_frac = 20 ;
193 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
194 {
195 	uint32_t val = reg_frac;
196 	int error;
197 
198 	error = sysctl_handle_int(oidp, &val, 0, req);
199 	if (error || !req->newptr )
200 		return (error);
201 	if (val < 1 || val > hz)
202 		return (EINVAL);
203 
204 	mtx_lock(&poll_mtx);
205 	reg_frac = val;
206 	if (reg_frac_count >= reg_frac)
207 		reg_frac_count = 0;
208 	mtx_unlock(&poll_mtx);
209 
210 	return (0);
211 }
212 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
213 	0, sizeof(uint32_t), reg_frac_sysctl, "I",
214 	"Every this many cycles check registers");
215 
216 static uint32_t short_ticks;
217 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
218 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
219 
220 static uint32_t lost_polls;
221 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
222 	&lost_polls, 0, "How many times we would have lost a poll tick");
223 
224 static uint32_t pending_polls;
225 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
226 	&pending_polls, 0, "Do we need to poll again");
227 
228 static int residual_burst = 0;
229 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
230 	&residual_burst, 0, "# of residual cycles in burst");
231 
232 static uint32_t poll_handlers; /* next free entry in pr[]. */
233 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
234 	&poll_handlers, 0, "Number of registered poll handlers");
235 
236 static int polling = 0;
237 SYSCTL_PROC(_kern_polling, OID_AUTO, enable, CTLTYPE_UINT | CTLFLAG_RW,
238 	0, sizeof(int), poll_switch, "I", "Switch polling for all interfaces");
239 
240 static uint32_t phase;
241 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
242 	&phase, 0, "Polling phase");
243 
244 static uint32_t suspect;
245 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
246 	&suspect, 0, "suspect event");
247 
248 static uint32_t stalled;
249 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
250 	&stalled, 0, "potential stalls");
251 
252 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
253 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
254 	&idlepoll_sleeping, 0, "idlepoll is sleeping");
255 
256 
257 #define POLL_LIST_LEN  128
258 struct pollrec {
259 	poll_handler_t	*handler;
260 	struct ifnet	*ifp;
261 };
262 
263 static struct pollrec pr[POLL_LIST_LEN];
264 
265 static void
266 poll_shutdown(void *arg, int howto)
267 {
268 
269 	poll_shutting_down = 1;
270 }
271 
272 static void
273 init_device_poll(void)
274 {
275 
276 	mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
277 	EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
278 	    SHUTDOWN_PRI_LAST);
279 }
280 SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL);
281 
282 
283 /*
284  * Hook from hardclock. Tries to schedule a netisr, but keeps track
285  * of lost ticks due to the previous handler taking too long.
286  * Normally, this should not happen, because polling handler should
287  * run for a short time. However, in some cases (e.g. when there are
288  * changes in link status etc.) the drivers take a very long time
289  * (even in the order of milliseconds) to reset and reconfigure the
290  * device, causing apparent lost polls.
291  *
292  * The first part of the code is just for debugging purposes, and tries
293  * to count how often hardclock ticks are shorter than they should,
294  * meaning either stray interrupts or delayed events.
295  */
296 void
297 hardclock_device_poll(void)
298 {
299 	static struct timeval prev_t, t;
300 	int delta;
301 
302 	if (poll_handlers == 0 || poll_shutting_down)
303 		return;
304 
305 	microuptime(&t);
306 	delta = (t.tv_usec - prev_t.tv_usec) +
307 		(t.tv_sec - prev_t.tv_sec)*1000000;
308 	if (delta * hz < 500000)
309 		short_ticks++;
310 	else
311 		prev_t = t;
312 
313 	if (pending_polls > 100) {
314 		/*
315 		 * Too much, assume it has stalled (not always true
316 		 * see comment above).
317 		 */
318 		stalled++;
319 		pending_polls = 0;
320 		phase = 0;
321 	}
322 
323 	if (phase <= 2) {
324 		if (phase != 0)
325 			suspect++;
326 		phase = 1;
327 		netisr_poll_scheduled = 1;
328 		netisr_pollmore_scheduled = 1;
329 		netisr_sched_poll();
330 		phase = 2;
331 	}
332 	if (pending_polls++ > 0)
333 		lost_polls++;
334 }
335 
336 /*
337  * ether_poll is called from the idle loop.
338  */
339 static void
340 ether_poll(int count)
341 {
342 	int i;
343 
344 	mtx_lock(&poll_mtx);
345 
346 	if (count > poll_each_burst)
347 		count = poll_each_burst;
348 
349 	for (i = 0 ; i < poll_handlers ; i++)
350 		pr[i].handler(pr[i].ifp, POLL_ONLY, count);
351 
352 	mtx_unlock(&poll_mtx);
353 }
354 
355 /*
356  * netisr_pollmore is called after other netisr's, possibly scheduling
357  * another NETISR_POLL call, or adapting the burst size for the next cycle.
358  *
359  * It is very bad to fetch large bursts of packets from a single card at once,
360  * because the burst could take a long time to be completely processed, or
361  * could saturate the intermediate queue (ipintrq or similar) leading to
362  * losses or unfairness. To reduce the problem, and also to account better for
363  * time spent in network-related processing, we split the burst in smaller
364  * chunks of fixed size, giving control to the other netisr's between chunks.
365  * This helps in improving the fairness, reducing livelock (because we
366  * emulate more closely the "process to completion" that we have with
367  * fastforwarding) and accounting for the work performed in low level
368  * handling and forwarding.
369  */
370 
371 static struct timeval poll_start_t;
372 
373 void
374 netisr_pollmore()
375 {
376 	struct timeval t;
377 	int kern_load;
378 
379 	mtx_lock(&poll_mtx);
380 	if (!netisr_pollmore_scheduled) {
381 		mtx_unlock(&poll_mtx);
382 		return;
383 	}
384 	netisr_pollmore_scheduled = 0;
385 	phase = 5;
386 	if (residual_burst > 0) {
387 		netisr_poll_scheduled = 1;
388 		netisr_pollmore_scheduled = 1;
389 		netisr_sched_poll();
390 		mtx_unlock(&poll_mtx);
391 		/* will run immediately on return, followed by netisrs */
392 		return;
393 	}
394 	/* here we can account time spent in netisr's in this tick */
395 	microuptime(&t);
396 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
397 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
398 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
399 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
400 		if (poll_burst > 1)
401 			poll_burst--;
402 	} else {
403 		if (poll_burst < poll_burst_max)
404 			poll_burst++;
405 	}
406 
407 	pending_polls--;
408 	if (pending_polls == 0) /* we are done */
409 		phase = 0;
410 	else {
411 		/*
412 		 * Last cycle was long and caused us to miss one or more
413 		 * hardclock ticks. Restart processing again, but slightly
414 		 * reduce the burst size to prevent that this happens again.
415 		 */
416 		poll_burst -= (poll_burst / 8);
417 		if (poll_burst < 1)
418 			poll_burst = 1;
419 		netisr_poll_scheduled = 1;
420 		netisr_pollmore_scheduled = 1;
421 		netisr_sched_poll();
422 		phase = 6;
423 	}
424 	mtx_unlock(&poll_mtx);
425 }
426 
427 /*
428  * netisr_poll is typically scheduled once per tick.
429  */
430 void
431 netisr_poll(void)
432 {
433 	int i, cycles;
434 	enum poll_cmd arg = POLL_ONLY;
435 
436 	mtx_lock(&poll_mtx);
437 	if (!netisr_poll_scheduled) {
438 		mtx_unlock(&poll_mtx);
439 		return;
440 	}
441 	netisr_poll_scheduled = 0;
442 	phase = 3;
443 	if (residual_burst == 0) { /* first call in this tick */
444 		microuptime(&poll_start_t);
445 		if (++reg_frac_count == reg_frac) {
446 			arg = POLL_AND_CHECK_STATUS;
447 			reg_frac_count = 0;
448 		}
449 
450 		residual_burst = poll_burst;
451 	}
452 	cycles = (residual_burst < poll_each_burst) ?
453 		residual_burst : poll_each_burst;
454 	residual_burst -= cycles;
455 
456 	for (i = 0 ; i < poll_handlers ; i++)
457 		pr[i].handler(pr[i].ifp, arg, cycles);
458 
459 	phase = 4;
460 	mtx_unlock(&poll_mtx);
461 }
462 
463 /*
464  * Try to register routine for polling. Returns 0 if successful
465  * (and polling should be enabled), error code otherwise.
466  * A device is not supposed to register itself multiple times.
467  *
468  * This is called from within the *_ioctl() functions.
469  */
470 int
471 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
472 {
473 	int i;
474 
475 	KASSERT(h != NULL, ("%s: handler is NULL", __func__));
476 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
477 
478 	mtx_lock(&poll_mtx);
479 	if (poll_handlers >= POLL_LIST_LEN) {
480 		/*
481 		 * List full, cannot register more entries.
482 		 * This should never happen; if it does, it is probably a
483 		 * broken driver trying to register multiple times. Checking
484 		 * this at runtime is expensive, and won't solve the problem
485 		 * anyways, so just report a few times and then give up.
486 		 */
487 		static int verbose = 10 ;
488 		if (verbose >0) {
489 			log(LOG_ERR, "poll handlers list full, "
490 			    "maybe a broken driver ?\n");
491 			verbose--;
492 		}
493 		mtx_unlock(&poll_mtx);
494 		return (ENOMEM); /* no polling for you */
495 	}
496 
497 	for (i = 0 ; i < poll_handlers ; i++)
498 		if (pr[i].ifp == ifp && pr[i].handler != NULL) {
499 			mtx_unlock(&poll_mtx);
500 			log(LOG_DEBUG, "ether_poll_register: %s: handler"
501 			    " already registered\n", ifp->if_xname);
502 			return (EEXIST);
503 		}
504 
505 	pr[poll_handlers].handler = h;
506 	pr[poll_handlers].ifp = ifp;
507 	poll_handlers++;
508 	mtx_unlock(&poll_mtx);
509 	if (idlepoll_sleeping)
510 		wakeup(&idlepoll_sleeping);
511 	return (0);
512 }
513 
514 /*
515  * Remove interface from the polling list. Called from *_ioctl(), too.
516  */
517 int
518 ether_poll_deregister(struct ifnet *ifp)
519 {
520 	int i;
521 
522 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
523 
524 	mtx_lock(&poll_mtx);
525 
526 	for (i = 0 ; i < poll_handlers ; i++)
527 		if (pr[i].ifp == ifp) /* found it */
528 			break;
529 	if (i == poll_handlers) {
530 		log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
531 		    ifp->if_xname);
532 		mtx_unlock(&poll_mtx);
533 		return (ENOENT);
534 	}
535 	poll_handlers--;
536 	if (i < poll_handlers) { /* Last entry replaces this one. */
537 		pr[i].handler = pr[poll_handlers].handler;
538 		pr[i].ifp = pr[poll_handlers].ifp;
539 	}
540 	mtx_unlock(&poll_mtx);
541 	return (0);
542 }
543 
544 /*
545  * Legacy interface for turning polling on all interfaces at one time.
546  */
547 static int
548 poll_switch(SYSCTL_HANDLER_ARGS)
549 {
550 	INIT_VNET_NET(curvnet);
551 	struct ifnet *ifp;
552 	int error;
553 	int val = polling;
554 
555 	error = sysctl_handle_int(oidp, &val, 0, req);
556 	if (error || !req->newptr )
557 		return (error);
558 
559 	if (val == polling)
560 		return (0);
561 
562 	if (val < 0 || val > 1)
563 		return (EINVAL);
564 
565 	polling = val;
566 
567 	IFNET_RLOCK();
568 	TAILQ_FOREACH(ifp, &V_ifnet, if_link) {
569 		if (ifp->if_capabilities & IFCAP_POLLING) {
570 			struct ifreq ifr;
571 
572 			if (val == 1)
573 				ifr.ifr_reqcap =
574 				    ifp->if_capenable | IFCAP_POLLING;
575 			else
576 				ifr.ifr_reqcap =
577 				    ifp->if_capenable & ~IFCAP_POLLING;
578 			(void) (*ifp->if_ioctl)(ifp, SIOCSIFCAP, (caddr_t)&ifr);
579 		}
580 	}
581 	IFNET_RUNLOCK();
582 
583 	log(LOG_ERR, "kern.polling.enable is deprecated. Use ifconfig(8)");
584 
585 	return (0);
586 }
587 
588 static void
589 poll_idle(void)
590 {
591 	struct thread *td = curthread;
592 	struct rtprio rtp;
593 
594 	rtp.prio = RTP_PRIO_MAX;	/* lowest priority */
595 	rtp.type = RTP_PRIO_IDLE;
596 	PROC_SLOCK(td->td_proc);
597 	rtp_to_pri(&rtp, td);
598 	PROC_SUNLOCK(td->td_proc);
599 
600 	for (;;) {
601 		if (poll_in_idle_loop && poll_handlers > 0) {
602 			idlepoll_sleeping = 0;
603 			ether_poll(poll_each_burst);
604 			thread_lock(td);
605 			mi_switch(SW_VOL, NULL);
606 			thread_unlock(td);
607 		} else {
608 			idlepoll_sleeping = 1;
609 			tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
610 		}
611 	}
612 }
613 
614 static struct proc *idlepoll;
615 static struct kproc_desc idlepoll_kp = {
616 	 "idlepoll",
617 	 poll_idle,
618 	 &idlepoll
619 };
620 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
621     &idlepoll_kp);
622