xref: /freebsd/sys/kern/kern_poll.c (revision 38f0b757fd84d17d0fc24739a7cda160c4516d81)
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_device_polling.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/kthread.h>
37 #include <sys/proc.h>
38 #include <sys/eventhandler.h>
39 #include <sys/resourcevar.h>
40 #include <sys/socket.h>			/* needed by net/if.h		*/
41 #include <sys/sockio.h>
42 #include <sys/sysctl.h>
43 #include <sys/syslog.h>
44 
45 #include <net/if.h>
46 #include <net/if_var.h>
47 #include <net/netisr.h>			/* for NETISR_POLL		*/
48 #include <net/vnet.h>
49 
50 void hardclock_device_poll(void);	/* hook from hardclock		*/
51 
52 static struct mtx	poll_mtx;
53 
54 /*
55  * Polling support for [network] device drivers.
56  *
57  * Drivers which support this feature can register with the
58  * polling code.
59  *
60  * If registration is successful, the driver must disable interrupts,
61  * and further I/O is performed through the handler, which is invoked
62  * (at least once per clock tick) with 3 arguments: the "arg" passed at
63  * register time (a struct ifnet pointer), a command, and a "count" limit.
64  *
65  * The command can be one of the following:
66  *  POLL_ONLY: quick move of "count" packets from input/output queues.
67  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
68  *	other more expensive operations. This command is issued periodically
69  *	but less frequently than POLL_ONLY.
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  * Polling is enabled and disabled via setting IFCAP_POLLING flag on
78  * the interface. The driver ioctl handler should register interface
79  * with polling and disable interrupts, if registration was successful.
80  *
81  * A second variable controls the sharing of CPU between polling/kernel
82  * network processing, and other activities (typically userlevel tasks):
83  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
84  * of CPU allocated to user tasks. CPU is allocated proportionally to the
85  * shares, by dynamically adjusting the "count" (poll_burst).
86  *
87  * Other parameters can should be left to their default values.
88  * The following constraints hold
89  *
90  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
91  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
92  */
93 
94 #define MIN_POLL_BURST_MAX	10
95 #define MAX_POLL_BURST_MAX	20000
96 
97 static uint32_t poll_burst = 5;
98 static uint32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
99 static uint32_t poll_each_burst = 5;
100 
101 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
102 	"Device polling parameters");
103 
104 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
105 	&poll_burst, 0, "Current polling burst size");
106 
107 static int	netisr_poll_scheduled;
108 static int	netisr_pollmore_scheduled;
109 static int	poll_shutting_down;
110 
111 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
112 {
113 	uint32_t val = poll_burst_max;
114 	int error;
115 
116 	error = sysctl_handle_int(oidp, &val, 0, req);
117 	if (error || !req->newptr )
118 		return (error);
119 	if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
120 		return (EINVAL);
121 
122 	mtx_lock(&poll_mtx);
123 	poll_burst_max = val;
124 	if (poll_burst > poll_burst_max)
125 		poll_burst = poll_burst_max;
126 	if (poll_each_burst > poll_burst_max)
127 		poll_each_burst = MIN_POLL_BURST_MAX;
128 	mtx_unlock(&poll_mtx);
129 
130 	return (0);
131 }
132 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
133 	0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
134 
135 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
136 {
137 	uint32_t val = poll_each_burst;
138 	int error;
139 
140 	error = sysctl_handle_int(oidp, &val, 0, req);
141 	if (error || !req->newptr )
142 		return (error);
143 	if (val < 1)
144 		return (EINVAL);
145 
146 	mtx_lock(&poll_mtx);
147 	if (val > poll_burst_max) {
148 		mtx_unlock(&poll_mtx);
149 		return (EINVAL);
150 	}
151 	poll_each_burst = val;
152 	mtx_unlock(&poll_mtx);
153 
154 	return (0);
155 }
156 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
157 	0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
158 	"Max size of each burst");
159 
160 static uint32_t poll_in_idle_loop=0;	/* do we poll in idle loop ? */
161 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
162 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
163 
164 static uint32_t user_frac = 50;
165 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
166 {
167 	uint32_t val = user_frac;
168 	int error;
169 
170 	error = sysctl_handle_int(oidp, &val, 0, req);
171 	if (error || !req->newptr )
172 		return (error);
173 	if (val > 99)
174 		return (EINVAL);
175 
176 	mtx_lock(&poll_mtx);
177 	user_frac = val;
178 	mtx_unlock(&poll_mtx);
179 
180 	return (0);
181 }
182 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
183 	0, sizeof(uint32_t), user_frac_sysctl, "I",
184 	"Desired user fraction of cpu time");
185 
186 static uint32_t reg_frac_count = 0;
187 static uint32_t reg_frac = 20 ;
188 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
189 {
190 	uint32_t val = reg_frac;
191 	int error;
192 
193 	error = sysctl_handle_int(oidp, &val, 0, req);
194 	if (error || !req->newptr )
195 		return (error);
196 	if (val < 1 || val > hz)
197 		return (EINVAL);
198 
199 	mtx_lock(&poll_mtx);
200 	reg_frac = val;
201 	if (reg_frac_count >= reg_frac)
202 		reg_frac_count = 0;
203 	mtx_unlock(&poll_mtx);
204 
205 	return (0);
206 }
207 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
208 	0, sizeof(uint32_t), reg_frac_sysctl, "I",
209 	"Every this many cycles check registers");
210 
211 static uint32_t short_ticks;
212 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
213 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
214 
215 static uint32_t lost_polls;
216 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
217 	&lost_polls, 0, "How many times we would have lost a poll tick");
218 
219 static uint32_t pending_polls;
220 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
221 	&pending_polls, 0, "Do we need to poll again");
222 
223 static int residual_burst = 0;
224 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
225 	&residual_burst, 0, "# of residual cycles in burst");
226 
227 static uint32_t poll_handlers; /* next free entry in pr[]. */
228 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
229 	&poll_handlers, 0, "Number of registered poll handlers");
230 
231 static uint32_t phase;
232 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
233 	&phase, 0, "Polling phase");
234 
235 static uint32_t suspect;
236 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
237 	&suspect, 0, "suspect event");
238 
239 static uint32_t stalled;
240 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
241 	&stalled, 0, "potential stalls");
242 
243 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
244 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
245 	&idlepoll_sleeping, 0, "idlepoll is sleeping");
246 
247 
248 #define POLL_LIST_LEN  128
249 struct pollrec {
250 	poll_handler_t	*handler;
251 	struct ifnet	*ifp;
252 };
253 
254 static struct pollrec pr[POLL_LIST_LEN];
255 
256 static void
257 poll_shutdown(void *arg, int howto)
258 {
259 
260 	poll_shutting_down = 1;
261 }
262 
263 static void
264 init_device_poll(void)
265 {
266 
267 	mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
268 	EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
269 	    SHUTDOWN_PRI_LAST);
270 }
271 SYSINIT(device_poll, SI_SUB_SOFTINTR, SI_ORDER_MIDDLE, init_device_poll, NULL);
272 
273 
274 /*
275  * Hook from hardclock. Tries to schedule a netisr, but keeps track
276  * of lost ticks due to the previous handler taking too long.
277  * Normally, this should not happen, because polling handler should
278  * run for a short time. However, in some cases (e.g. when there are
279  * changes in link status etc.) the drivers take a very long time
280  * (even in the order of milliseconds) to reset and reconfigure the
281  * device, causing apparent lost polls.
282  *
283  * The first part of the code is just for debugging purposes, and tries
284  * to count how often hardclock ticks are shorter than they should,
285  * meaning either stray interrupts or delayed events.
286  */
287 void
288 hardclock_device_poll(void)
289 {
290 	static struct timeval prev_t, t;
291 	int delta;
292 
293 	if (poll_handlers == 0 || poll_shutting_down)
294 		return;
295 
296 	microuptime(&t);
297 	delta = (t.tv_usec - prev_t.tv_usec) +
298 		(t.tv_sec - prev_t.tv_sec)*1000000;
299 	if (delta * hz < 500000)
300 		short_ticks++;
301 	else
302 		prev_t = t;
303 
304 	if (pending_polls > 100) {
305 		/*
306 		 * Too much, assume it has stalled (not always true
307 		 * see comment above).
308 		 */
309 		stalled++;
310 		pending_polls = 0;
311 		phase = 0;
312 	}
313 
314 	if (phase <= 2) {
315 		if (phase != 0)
316 			suspect++;
317 		phase = 1;
318 		netisr_poll_scheduled = 1;
319 		netisr_pollmore_scheduled = 1;
320 		netisr_sched_poll();
321 		phase = 2;
322 	}
323 	if (pending_polls++ > 0)
324 		lost_polls++;
325 }
326 
327 /*
328  * ether_poll is called from the idle loop.
329  */
330 static void
331 ether_poll(int count)
332 {
333 	int i;
334 
335 	mtx_lock(&poll_mtx);
336 
337 	if (count > poll_each_burst)
338 		count = poll_each_burst;
339 
340 	for (i = 0 ; i < poll_handlers ; i++)
341 		pr[i].handler(pr[i].ifp, POLL_ONLY, count);
342 
343 	mtx_unlock(&poll_mtx);
344 }
345 
346 /*
347  * netisr_pollmore is called after other netisr's, possibly scheduling
348  * another NETISR_POLL call, or adapting the burst size for the next cycle.
349  *
350  * It is very bad to fetch large bursts of packets from a single card at once,
351  * because the burst could take a long time to be completely processed, or
352  * could saturate the intermediate queue (ipintrq or similar) leading to
353  * losses or unfairness. To reduce the problem, and also to account better for
354  * time spent in network-related processing, we split the burst in smaller
355  * chunks of fixed size, giving control to the other netisr's between chunks.
356  * This helps in improving the fairness, reducing livelock (because we
357  * emulate more closely the "process to completion" that we have with
358  * fastforwarding) and accounting for the work performed in low level
359  * handling and forwarding.
360  */
361 
362 static struct timeval poll_start_t;
363 
364 void
365 netisr_pollmore()
366 {
367 	struct timeval t;
368 	int kern_load;
369 
370 	mtx_lock(&poll_mtx);
371 	if (!netisr_pollmore_scheduled) {
372 		mtx_unlock(&poll_mtx);
373 		return;
374 	}
375 	netisr_pollmore_scheduled = 0;
376 	phase = 5;
377 	if (residual_burst > 0) {
378 		netisr_poll_scheduled = 1;
379 		netisr_pollmore_scheduled = 1;
380 		netisr_sched_poll();
381 		mtx_unlock(&poll_mtx);
382 		/* will run immediately on return, followed by netisrs */
383 		return;
384 	}
385 	/* here we can account time spent in netisr's in this tick */
386 	microuptime(&t);
387 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
388 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
389 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
390 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
391 		if (poll_burst > 1)
392 			poll_burst--;
393 	} else {
394 		if (poll_burst < poll_burst_max)
395 			poll_burst++;
396 	}
397 
398 	pending_polls--;
399 	if (pending_polls == 0) /* we are done */
400 		phase = 0;
401 	else {
402 		/*
403 		 * Last cycle was long and caused us to miss one or more
404 		 * hardclock ticks. Restart processing again, but slightly
405 		 * reduce the burst size to prevent that this happens again.
406 		 */
407 		poll_burst -= (poll_burst / 8);
408 		if (poll_burst < 1)
409 			poll_burst = 1;
410 		netisr_poll_scheduled = 1;
411 		netisr_pollmore_scheduled = 1;
412 		netisr_sched_poll();
413 		phase = 6;
414 	}
415 	mtx_unlock(&poll_mtx);
416 }
417 
418 /*
419  * netisr_poll is typically scheduled once per tick.
420  */
421 void
422 netisr_poll(void)
423 {
424 	int i, cycles;
425 	enum poll_cmd arg = POLL_ONLY;
426 
427 	mtx_lock(&poll_mtx);
428 	if (!netisr_poll_scheduled) {
429 		mtx_unlock(&poll_mtx);
430 		return;
431 	}
432 	netisr_poll_scheduled = 0;
433 	phase = 3;
434 	if (residual_burst == 0) { /* first call in this tick */
435 		microuptime(&poll_start_t);
436 		if (++reg_frac_count == reg_frac) {
437 			arg = POLL_AND_CHECK_STATUS;
438 			reg_frac_count = 0;
439 		}
440 
441 		residual_burst = poll_burst;
442 	}
443 	cycles = (residual_burst < poll_each_burst) ?
444 		residual_burst : poll_each_burst;
445 	residual_burst -= cycles;
446 
447 	for (i = 0 ; i < poll_handlers ; i++)
448 		pr[i].handler(pr[i].ifp, arg, cycles);
449 
450 	phase = 4;
451 	mtx_unlock(&poll_mtx);
452 }
453 
454 /*
455  * Try to register routine for polling. Returns 0 if successful
456  * (and polling should be enabled), error code otherwise.
457  * A device is not supposed to register itself multiple times.
458  *
459  * This is called from within the *_ioctl() functions.
460  */
461 int
462 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
463 {
464 	int i;
465 
466 	KASSERT(h != NULL, ("%s: handler is NULL", __func__));
467 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
468 
469 	mtx_lock(&poll_mtx);
470 	if (poll_handlers >= POLL_LIST_LEN) {
471 		/*
472 		 * List full, cannot register more entries.
473 		 * This should never happen; if it does, it is probably a
474 		 * broken driver trying to register multiple times. Checking
475 		 * this at runtime is expensive, and won't solve the problem
476 		 * anyways, so just report a few times and then give up.
477 		 */
478 		static int verbose = 10 ;
479 		if (verbose >0) {
480 			log(LOG_ERR, "poll handlers list full, "
481 			    "maybe a broken driver ?\n");
482 			verbose--;
483 		}
484 		mtx_unlock(&poll_mtx);
485 		return (ENOMEM); /* no polling for you */
486 	}
487 
488 	for (i = 0 ; i < poll_handlers ; i++)
489 		if (pr[i].ifp == ifp && pr[i].handler != NULL) {
490 			mtx_unlock(&poll_mtx);
491 			log(LOG_DEBUG, "ether_poll_register: %s: handler"
492 			    " already registered\n", ifp->if_xname);
493 			return (EEXIST);
494 		}
495 
496 	pr[poll_handlers].handler = h;
497 	pr[poll_handlers].ifp = ifp;
498 	poll_handlers++;
499 	mtx_unlock(&poll_mtx);
500 	if (idlepoll_sleeping)
501 		wakeup(&idlepoll_sleeping);
502 	return (0);
503 }
504 
505 /*
506  * Remove interface from the polling list. Called from *_ioctl(), too.
507  */
508 int
509 ether_poll_deregister(struct ifnet *ifp)
510 {
511 	int i;
512 
513 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
514 
515 	mtx_lock(&poll_mtx);
516 
517 	for (i = 0 ; i < poll_handlers ; i++)
518 		if (pr[i].ifp == ifp) /* found it */
519 			break;
520 	if (i == poll_handlers) {
521 		log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
522 		    ifp->if_xname);
523 		mtx_unlock(&poll_mtx);
524 		return (ENOENT);
525 	}
526 	poll_handlers--;
527 	if (i < poll_handlers) { /* Last entry replaces this one. */
528 		pr[i].handler = pr[poll_handlers].handler;
529 		pr[i].ifp = pr[poll_handlers].ifp;
530 	}
531 	mtx_unlock(&poll_mtx);
532 	return (0);
533 }
534 
535 static void
536 poll_idle(void)
537 {
538 	struct thread *td = curthread;
539 	struct rtprio rtp;
540 
541 	rtp.prio = RTP_PRIO_MAX;	/* lowest priority */
542 	rtp.type = RTP_PRIO_IDLE;
543 	PROC_SLOCK(td->td_proc);
544 	rtp_to_pri(&rtp, td);
545 	PROC_SUNLOCK(td->td_proc);
546 
547 	for (;;) {
548 		if (poll_in_idle_loop && poll_handlers > 0) {
549 			idlepoll_sleeping = 0;
550 			ether_poll(poll_each_burst);
551 			thread_lock(td);
552 			mi_switch(SW_VOL, NULL);
553 			thread_unlock(td);
554 		} else {
555 			idlepoll_sleeping = 1;
556 			tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
557 		}
558 	}
559 }
560 
561 static struct proc *idlepoll;
562 static struct kproc_desc idlepoll_kp = {
563 	 "idlepoll",
564 	 poll_idle,
565 	 &idlepoll
566 };
567 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
568     &idlepoll_kp);
569