xref: /freebsd/sys/kern/kern_poll.c (revision ca9ac06c99bfd0150b85d4d83c396ce6237c0e05)
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 <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/socket.h>			/* needed by net/if.h		*/
35 #include <sys/sysctl.h>
36 
37 #include <net/if.h>			/* for IFF_* flags		*/
38 #include <net/netisr.h>			/* for NETISR_POLL		*/
39 
40 #include <sys/proc.h>
41 #include <sys/resourcevar.h>
42 #include <sys/kthread.h>
43 
44 static void netisr_poll(void);		/* the two netisr handlers      */
45 static void netisr_pollmore(void);
46 
47 void hardclock_device_poll(void);	/* hook from hardclock		*/
48 void ether_poll(int);			/* polling while in trap	*/
49 
50 /*
51  * Polling support for [network] device drivers.
52  *
53  * Drivers which support this feature try to register with the
54  * polling code.
55  *
56  * If registration is successful, the driver must disable interrupts,
57  * and further I/O is performed through the handler, which is invoked
58  * (at least once per clock tick) with 3 arguments: the "arg" passed at
59  * register time (a struct ifnet pointer), a command, and a "count" limit.
60  *
61  * The command can be one of the following:
62  *  POLL_ONLY: quick move of "count" packets from input/output queues.
63  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
64  *	other more expensive operations. This command is issued periodically
65  *	but less frequently than POLL_ONLY.
66  *  POLL_DEREGISTER: deregister and return to interrupt mode.
67  *
68  * The first two commands are only issued if the interface is marked as
69  * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
70  *
71  * The count limit specifies how much work the handler can do during the
72  * call -- typically this is the number of packets to be received, or
73  * transmitted, etc. (drivers are free to interpret this number, as long
74  * as the max time spent in the function grows roughly linearly with the
75  * count).
76  *
77  * Deregistration can be requested by the driver itself (typically in the
78  * *_stop() routine), or by the polling code, by invoking the handler.
79  *
80  * Polling can be globally enabled or disabled with the sysctl variable
81  * kern.polling.enable (default is 0, disabled)
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  *	0 <= poll_in_trap <= poll_each_burst
94  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
95  */
96 
97 #define MIN_POLL_BURST_MAX	10
98 #define MAX_POLL_BURST_MAX	1000
99 
100 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
101 	"Device polling parameters");
102 
103 static u_int32_t poll_burst = 5;
104 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
105 	&poll_burst, 0, "Current polling burst size");
106 
107 static u_int32_t poll_each_burst = 5;
108 SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
109 	&poll_each_burst, 0, "Max size of each burst");
110 
111 static u_int32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
112 SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
113 	&poll_burst_max, 0, "Max Polling burst size");
114 
115 static u_int32_t poll_in_idle_loop=0;	/* do we poll in idle loop ? */
116 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
117 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
118 
119 u_int32_t poll_in_trap;			/* used in trap.c */
120 SYSCTL_UINT(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW,
121 	&poll_in_trap, 0, "Poll burst size during a trap");
122 
123 static u_int32_t user_frac = 50;
124 SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
125 	&user_frac, 0, "Desired user fraction of cpu time");
126 
127 static u_int32_t reg_frac = 20 ;
128 SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
129 	&reg_frac, 0, "Every this many cycles poll register");
130 
131 static u_int32_t short_ticks;
132 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
133 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
134 
135 static u_int32_t lost_polls;
136 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
137 	&lost_polls, 0, "How many times we would have lost a poll tick");
138 
139 static u_int32_t pending_polls;
140 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
141 	&pending_polls, 0, "Do we need to poll again");
142 
143 static int residual_burst = 0;
144 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
145 	&residual_burst, 0, "# of residual cycles in burst");
146 
147 static u_int32_t poll_handlers; /* next free entry in pr[]. */
148 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
149 	&poll_handlers, 0, "Number of registered poll handlers");
150 
151 static int polling = 0;		/* global polling enable */
152 SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
153 	&polling, 0, "Polling enabled");
154 
155 static u_int32_t phase;
156 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
157 	&phase, 0, "Polling phase");
158 
159 static u_int32_t suspect;
160 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
161 	&suspect, 0, "suspect event");
162 
163 static u_int32_t stalled;
164 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
165 	&stalled, 0, "potential stalls");
166 
167 static u_int32_t idlepoll_sleeping; /* idlepoll is sleeping */
168 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
169 	&idlepoll_sleeping, 0, "idlepoll is sleeping");
170 
171 
172 #define POLL_LIST_LEN  128
173 struct pollrec {
174 	poll_handler_t	*handler;
175 	struct ifnet	*ifp;
176 };
177 
178 static struct pollrec pr[POLL_LIST_LEN];
179 
180 static void
181 init_device_poll(void)
182 {
183 
184 	netisr_register(NETISR_POLL, (netisr_t *)netisr_poll, NULL, 0);
185 	netisr_register(NETISR_POLLMORE, (netisr_t *)netisr_pollmore, NULL, 0);
186 }
187 SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL)
188 
189 
190 /*
191  * Hook from hardclock. Tries to schedule a netisr, but keeps track
192  * of lost ticks due to the previous handler taking too long.
193  * Normally, this should not happen, because polling handler should
194  * run for a short time. However, in some cases (e.g. when there are
195  * changes in link status etc.) the drivers take a very long time
196  * (even in the order of milliseconds) to reset and reconfigure the
197  * device, causing apparent lost polls.
198  *
199  * The first part of the code is just for debugging purposes, and tries
200  * to count how often hardclock ticks are shorter than they should,
201  * meaning either stray interrupts or delayed events.
202  */
203 void
204 hardclock_device_poll(void)
205 {
206 	static struct timeval prev_t, t;
207 	int delta;
208 
209 	if (poll_handlers == 0)
210 		return;
211 
212 	microuptime(&t);
213 	delta = (t.tv_usec - prev_t.tv_usec) +
214 		(t.tv_sec - prev_t.tv_sec)*1000000;
215 	if (delta * hz < 500000)
216 		short_ticks++;
217 	else
218 		prev_t = t;
219 
220 	if (pending_polls > 100) {
221 		/*
222 		 * Too much, assume it has stalled (not always true
223 		 * see comment above).
224 		 */
225 		stalled++;
226 		pending_polls = 0;
227 		phase = 0;
228 	}
229 
230 	if (phase <= 2) {
231 		if (phase != 0)
232 			suspect++;
233 		phase = 1;
234 		schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
235 		phase = 2;
236 	}
237 	if (pending_polls++ > 0)
238 		lost_polls++;
239 }
240 
241 /*
242  * ether_poll is called from the idle loop or from the trap handler.
243  */
244 void
245 ether_poll(int count)
246 {
247 	int i;
248 
249 	mtx_lock(&Giant);
250 
251 	if (count > poll_each_burst)
252 		count = poll_each_burst;
253 	for (i = 0 ; i < poll_handlers ; i++)
254 		if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
255 		    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
256 			pr[i].handler(pr[i].ifp, 0, count); /* quick check */
257 	mtx_unlock(&Giant);
258 }
259 
260 /*
261  * netisr_pollmore is called after other netisr's, possibly scheduling
262  * another NETISR_POLL call, or adapting the burst size for the next cycle.
263  *
264  * It is very bad to fetch large bursts of packets from a single card at once,
265  * because the burst could take a long time to be completely processed, or
266  * could saturate the intermediate queue (ipintrq or similar) leading to
267  * losses or unfairness. To reduce the problem, and also to account better for
268  * time spent in network-related processing, we split the burst in smaller
269  * chunks of fixed size, giving control to the other netisr's between chunks.
270  * This helps in improving the fairness, reducing livelock (because we
271  * emulate more closely the "process to completion" that we have with
272  * fastforwarding) and accounting for the work performed in low level
273  * handling and forwarding.
274  */
275 
276 static struct timeval poll_start_t;
277 
278 void
279 netisr_pollmore()
280 {
281 	struct timeval t;
282 	int kern_load;
283 	/* XXX run at splhigh() or equivalent */
284 
285 	phase = 5;
286 	if (residual_burst > 0) {
287 		schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
288 		/* will run immediately on return, followed by netisrs */
289 		return;
290 	}
291 	/* here we can account time spent in netisr's in this tick */
292 	microuptime(&t);
293 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
294 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
295 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
296 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
297 		if (poll_burst > 1)
298 			poll_burst--;
299 	} else {
300 		if (poll_burst < poll_burst_max)
301 			poll_burst++;
302 	}
303 
304 	pending_polls--;
305 	if (pending_polls == 0) /* we are done */
306 		phase = 0;
307 	else {
308 		/*
309 		 * Last cycle was long and caused us to miss one or more
310 		 * hardclock ticks. Restart processing again, but slightly
311 		 * reduce the burst size to prevent that this happens again.
312 		 */
313 		poll_burst -= (poll_burst / 8);
314 		if (poll_burst < 1)
315 			poll_burst = 1;
316 		schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
317 		phase = 6;
318 	}
319 }
320 
321 /*
322  * netisr_poll is scheduled by schednetisr when appropriate, typically once
323  * per tick. It is called at splnet() so first thing to do is to upgrade to
324  * splimp(), and call all registered handlers.
325  */
326 static void
327 netisr_poll(void)
328 {
329 	static int reg_frac_count;
330 	int i, cycles;
331 	enum poll_cmd arg = POLL_ONLY;
332 	mtx_lock(&Giant);
333 
334 	phase = 3;
335 	if (residual_burst == 0) { /* first call in this tick */
336 		microuptime(&poll_start_t);
337 		/*
338 		 * Check that paremeters are consistent with runtime
339 		 * variables. Some of these tests could be done at sysctl
340 		 * time, but the savings would be very limited because we
341 		 * still have to check against reg_frac_count and
342 		 * poll_each_burst. So, instead of writing separate sysctl
343 		 * handlers, we do all here.
344 		 */
345 
346 		if (reg_frac > hz)
347 			reg_frac = hz;
348 		else if (reg_frac < 1)
349 			reg_frac = 1;
350 		if (reg_frac_count > reg_frac)
351 			reg_frac_count = reg_frac - 1;
352 		if (reg_frac_count-- == 0) {
353 			arg = POLL_AND_CHECK_STATUS;
354 			reg_frac_count = reg_frac - 1;
355 		}
356 		if (poll_burst_max < MIN_POLL_BURST_MAX)
357 			poll_burst_max = MIN_POLL_BURST_MAX;
358 		else if (poll_burst_max > MAX_POLL_BURST_MAX)
359 			poll_burst_max = MAX_POLL_BURST_MAX;
360 
361 		if (poll_each_burst < 1)
362 			poll_each_burst = 1;
363 		else if (poll_each_burst > poll_burst_max)
364 			poll_each_burst = poll_burst_max;
365 
366 		if (poll_burst > poll_burst_max)
367 			poll_burst = poll_burst_max;
368 		residual_burst = poll_burst;
369 	}
370 	cycles = (residual_burst < poll_each_burst) ?
371 		residual_burst : poll_each_burst;
372 	residual_burst -= cycles;
373 
374 	if (polling) {
375 		for (i = 0 ; i < poll_handlers ; i++)
376 			if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
377 			    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
378 				pr[i].handler(pr[i].ifp, arg, cycles);
379 	} else {	/* unregister */
380 		for (i = 0 ; i < poll_handlers ; i++) {
381 			if (pr[i].handler &&
382 			    pr[i].ifp->if_flags & IFF_RUNNING) {
383 				pr[i].ifp->if_flags &= ~IFF_POLLING;
384 				pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1);
385 			}
386 			pr[i].handler=NULL;
387 		}
388 		residual_burst = 0;
389 		poll_handlers = 0;
390 	}
391 	/* on -stable, schednetisr(NETISR_POLLMORE); */
392 	phase = 4;
393 	mtx_unlock(&Giant);
394 }
395 
396 /*
397  * Try to register routine for polling. Returns 1 if successful
398  * (and polling should be enabled), 0 otherwise.
399  * A device is not supposed to register itself multiple times.
400  *
401  * This is called from within the *_intr() functions, so we do not need
402  * further locking.
403  */
404 int
405 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
406 {
407 	int s;
408 
409 	if (polling == 0) /* polling disabled, cannot register */
410 		return 0;
411 	if (h == NULL || ifp == NULL)		/* bad arguments	*/
412 		return 0;
413 	if ( !(ifp->if_flags & IFF_UP) )	/* must be up		*/
414 		return 0;
415 	if (ifp->if_flags & IFF_POLLING)	/* already polling	*/
416 		return 0;
417 
418 	s = splhigh();
419 	if (poll_handlers >= POLL_LIST_LEN) {
420 		/*
421 		 * List full, cannot register more entries.
422 		 * This should never happen; if it does, it is probably a
423 		 * broken driver trying to register multiple times. Checking
424 		 * this at runtime is expensive, and won't solve the problem
425 		 * anyways, so just report a few times and then give up.
426 		 */
427 		static int verbose = 10 ;
428 		splx(s);
429 		if (verbose >0) {
430 			printf("poll handlers list full, "
431 				"maybe a broken driver ?\n");
432 			verbose--;
433 		}
434 		return 0; /* no polling for you */
435 	}
436 
437 	pr[poll_handlers].handler = h;
438 	pr[poll_handlers].ifp = ifp;
439 	poll_handlers++;
440 	ifp->if_flags |= IFF_POLLING;
441 	splx(s);
442 	if (idlepoll_sleeping)
443 		wakeup(&idlepoll_sleeping);
444 	return 1; /* polling enabled in next call */
445 }
446 
447 /*
448  * Remove interface from the polling list. Normally called by *_stop().
449  * It is not an error to call it with IFF_POLLING clear, the call is
450  * sufficiently rare to be preferable to save the space for the extra
451  * test in each driver in exchange of one additional function call.
452  */
453 int
454 ether_poll_deregister(struct ifnet *ifp)
455 {
456 	int i;
457 
458 	mtx_lock(&Giant);
459 	if ( !ifp || !(ifp->if_flags & IFF_POLLING) ) {
460 		mtx_unlock(&Giant);
461 		return 0;
462 	}
463 	for (i = 0 ; i < poll_handlers ; i++)
464 		if (pr[i].ifp == ifp) /* found it */
465 			break;
466 	ifp->if_flags &= ~IFF_POLLING; /* found or not... */
467 	if (i == poll_handlers) {
468 		mtx_unlock(&Giant);
469 		printf("ether_poll_deregister: ifp not found!!!\n");
470 		return 0;
471 	}
472 	poll_handlers--;
473 	if (i < poll_handlers) { /* Last entry replaces this one. */
474 		pr[i].handler = pr[poll_handlers].handler;
475 		pr[i].ifp = pr[poll_handlers].ifp;
476 	}
477 	mtx_unlock(&Giant);
478 	return 1;
479 }
480 
481 static void
482 poll_idle(void)
483 {
484 	struct thread *td = curthread;
485 	struct rtprio rtp;
486 	int pri;
487 
488 	rtp.prio = RTP_PRIO_MAX;	/* lowest priority */
489 	rtp.type = RTP_PRIO_IDLE;
490 	mtx_lock_spin(&sched_lock);
491 	rtp_to_pri(&rtp, td->td_ksegrp);
492 	pri = td->td_priority;
493 	mtx_unlock_spin(&sched_lock);
494 
495 	for (;;) {
496 		if (poll_in_idle_loop && poll_handlers > 0) {
497 			idlepoll_sleeping = 0;
498 			mtx_lock(&Giant);
499 			ether_poll(poll_each_burst);
500 			mtx_unlock(&Giant);
501 			mtx_assert(&Giant, MA_NOTOWNED);
502 			mtx_lock_spin(&sched_lock);
503 			mi_switch(SW_VOL, NULL);
504 			mtx_unlock_spin(&sched_lock);
505 		} else {
506 			idlepoll_sleeping = 1;
507 			tsleep(&idlepoll_sleeping, pri, "pollid", hz * 3);
508 		}
509 	}
510 }
511 
512 static struct proc *idlepoll;
513 static struct kproc_desc idlepoll_kp = {
514 	 "idlepoll",
515 	 poll_idle,
516 	 &idlepoll
517 };
518 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start, &idlepoll_kp)
519