xref: /freebsd/sys/kern/kern_poll.c (revision 74036c4de983981ee2b23f91e285363293aa7c3c)
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>			/* for IFF_* flags		*/
46 #include <net/netisr.h>			/* for NETISR_POLL		*/
47 #include <net/vnet.h>
48 
49 void hardclock_device_poll(void);	/* hook from hardclock		*/
50 
51 static struct mtx	poll_mtx;
52 
53 /*
54  * Polling support for [network] device drivers.
55  *
56  * Drivers which support this feature can register with the
57  * polling code.
58  *
59  * If registration is successful, the driver must disable interrupts,
60  * and further I/O is performed through the handler, which is invoked
61  * (at least once per clock tick) with 3 arguments: the "arg" passed at
62  * register time (a struct ifnet pointer), a command, and a "count" limit.
63  *
64  * The command can be one of the following:
65  *  POLL_ONLY: quick move of "count" packets from input/output queues.
66  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
67  *	other more expensive operations. This command is issued periodically
68  *	but less frequently than POLL_ONLY.
69  *
70  * The count limit specifies how much work the handler can do during the
71  * call -- typically this is the number of packets to be received, or
72  * transmitted, etc. (drivers are free to interpret this number, as long
73  * as the max time spent in the function grows roughly linearly with the
74  * count).
75  *
76  * Polling is enabled and disabled via setting IFCAP_POLLING flag on
77  * the interface. The driver ioctl handler should register interface
78  * with polling and disable interrupts, if registration was successful.
79  *
80  * A second variable controls the sharing of CPU between polling/kernel
81  * network processing, and other activities (typically userlevel tasks):
82  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
83  * of CPU allocated to user tasks. CPU is allocated proportionally to the
84  * shares, by dynamically adjusting the "count" (poll_burst).
85  *
86  * Other parameters can should be left to their default values.
87  * The following constraints hold
88  *
89  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
90  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
91  */
92 
93 #define MIN_POLL_BURST_MAX	10
94 #define MAX_POLL_BURST_MAX	20000
95 
96 static uint32_t poll_burst = 5;
97 static uint32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
98 static uint32_t poll_each_burst = 5;
99 
100 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
101 	"Device polling parameters");
102 
103 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
104 	&poll_burst, 0, "Current polling burst size");
105 
106 static int	netisr_poll_scheduled;
107 static int	netisr_pollmore_scheduled;
108 static int	poll_shutting_down;
109 
110 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
111 {
112 	uint32_t val = poll_burst_max;
113 	int error;
114 
115 	error = sysctl_handle_int(oidp, &val, 0, req);
116 	if (error || !req->newptr )
117 		return (error);
118 	if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
119 		return (EINVAL);
120 
121 	mtx_lock(&poll_mtx);
122 	poll_burst_max = val;
123 	if (poll_burst > poll_burst_max)
124 		poll_burst = poll_burst_max;
125 	if (poll_each_burst > poll_burst_max)
126 		poll_each_burst = MIN_POLL_BURST_MAX;
127 	mtx_unlock(&poll_mtx);
128 
129 	return (0);
130 }
131 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
132 	0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
133 
134 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
135 {
136 	uint32_t val = poll_each_burst;
137 	int error;
138 
139 	error = sysctl_handle_int(oidp, &val, 0, req);
140 	if (error || !req->newptr )
141 		return (error);
142 	if (val < 1)
143 		return (EINVAL);
144 
145 	mtx_lock(&poll_mtx);
146 	if (val > poll_burst_max) {
147 		mtx_unlock(&poll_mtx);
148 		return (EINVAL);
149 	}
150 	poll_each_burst = val;
151 	mtx_unlock(&poll_mtx);
152 
153 	return (0);
154 }
155 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
156 	0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
157 	"Max size of each burst");
158 
159 static uint32_t poll_in_idle_loop=0;	/* do we poll in idle loop ? */
160 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
161 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
162 
163 static uint32_t user_frac = 50;
164 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
165 {
166 	uint32_t val = user_frac;
167 	int error;
168 
169 	error = sysctl_handle_int(oidp, &val, 0, req);
170 	if (error || !req->newptr )
171 		return (error);
172 	if (val < 0 || val > 99)
173 		return (EINVAL);
174 
175 	mtx_lock(&poll_mtx);
176 	user_frac = val;
177 	mtx_unlock(&poll_mtx);
178 
179 	return (0);
180 }
181 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
182 	0, sizeof(uint32_t), user_frac_sysctl, "I",
183 	"Desired user fraction of cpu time");
184 
185 static uint32_t reg_frac_count = 0;
186 static uint32_t reg_frac = 20 ;
187 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
188 {
189 	uint32_t val = reg_frac;
190 	int error;
191 
192 	error = sysctl_handle_int(oidp, &val, 0, req);
193 	if (error || !req->newptr )
194 		return (error);
195 	if (val < 1 || val > hz)
196 		return (EINVAL);
197 
198 	mtx_lock(&poll_mtx);
199 	reg_frac = val;
200 	if (reg_frac_count >= reg_frac)
201 		reg_frac_count = 0;
202 	mtx_unlock(&poll_mtx);
203 
204 	return (0);
205 }
206 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
207 	0, sizeof(uint32_t), reg_frac_sysctl, "I",
208 	"Every this many cycles check registers");
209 
210 static uint32_t short_ticks;
211 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
212 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
213 
214 static uint32_t lost_polls;
215 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
216 	&lost_polls, 0, "How many times we would have lost a poll tick");
217 
218 static uint32_t pending_polls;
219 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
220 	&pending_polls, 0, "Do we need to poll again");
221 
222 static int residual_burst = 0;
223 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
224 	&residual_burst, 0, "# of residual cycles in burst");
225 
226 static uint32_t poll_handlers; /* next free entry in pr[]. */
227 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
228 	&poll_handlers, 0, "Number of registered poll handlers");
229 
230 static uint32_t phase;
231 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
232 	&phase, 0, "Polling phase");
233 
234 static uint32_t suspect;
235 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
236 	&suspect, 0, "suspect event");
237 
238 static uint32_t stalled;
239 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
240 	&stalled, 0, "potential stalls");
241 
242 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
243 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
244 	&idlepoll_sleeping, 0, "idlepoll is sleeping");
245 
246 
247 #define POLL_LIST_LEN  128
248 struct pollrec {
249 	poll_handler_t	*handler;
250 	struct ifnet	*ifp;
251 };
252 
253 static struct pollrec pr[POLL_LIST_LEN];
254 
255 static void
256 poll_shutdown(void *arg, int howto)
257 {
258 
259 	poll_shutting_down = 1;
260 }
261 
262 static void
263 init_device_poll(void)
264 {
265 
266 	mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
267 	EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
268 	    SHUTDOWN_PRI_LAST);
269 }
270 SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL);
271 
272 
273 /*
274  * Hook from hardclock. Tries to schedule a netisr, but keeps track
275  * of lost ticks due to the previous handler taking too long.
276  * Normally, this should not happen, because polling handler should
277  * run for a short time. However, in some cases (e.g. when there are
278  * changes in link status etc.) the drivers take a very long time
279  * (even in the order of milliseconds) to reset and reconfigure the
280  * device, causing apparent lost polls.
281  *
282  * The first part of the code is just for debugging purposes, and tries
283  * to count how often hardclock ticks are shorter than they should,
284  * meaning either stray interrupts or delayed events.
285  */
286 void
287 hardclock_device_poll(void)
288 {
289 	static struct timeval prev_t, t;
290 	int delta;
291 
292 	if (poll_handlers == 0 || poll_shutting_down)
293 		return;
294 
295 	microuptime(&t);
296 	delta = (t.tv_usec - prev_t.tv_usec) +
297 		(t.tv_sec - prev_t.tv_sec)*1000000;
298 	if (delta * hz < 500000)
299 		short_ticks++;
300 	else
301 		prev_t = t;
302 
303 	if (pending_polls > 100) {
304 		/*
305 		 * Too much, assume it has stalled (not always true
306 		 * see comment above).
307 		 */
308 		stalled++;
309 		pending_polls = 0;
310 		phase = 0;
311 	}
312 
313 	if (phase <= 2) {
314 		if (phase != 0)
315 			suspect++;
316 		phase = 1;
317 		netisr_poll_scheduled = 1;
318 		netisr_pollmore_scheduled = 1;
319 		netisr_sched_poll();
320 		phase = 2;
321 	}
322 	if (pending_polls++ > 0)
323 		lost_polls++;
324 }
325 
326 /*
327  * ether_poll is called from the idle loop.
328  */
329 static void
330 ether_poll(int count)
331 {
332 	int i;
333 
334 	mtx_lock(&poll_mtx);
335 
336 	if (count > poll_each_burst)
337 		count = poll_each_burst;
338 
339 	for (i = 0 ; i < poll_handlers ; i++)
340 		pr[i].handler(pr[i].ifp, POLL_ONLY, count);
341 
342 	mtx_unlock(&poll_mtx);
343 }
344 
345 /*
346  * netisr_pollmore is called after other netisr's, possibly scheduling
347  * another NETISR_POLL call, or adapting the burst size for the next cycle.
348  *
349  * It is very bad to fetch large bursts of packets from a single card at once,
350  * because the burst could take a long time to be completely processed, or
351  * could saturate the intermediate queue (ipintrq or similar) leading to
352  * losses or unfairness. To reduce the problem, and also to account better for
353  * time spent in network-related processing, we split the burst in smaller
354  * chunks of fixed size, giving control to the other netisr's between chunks.
355  * This helps in improving the fairness, reducing livelock (because we
356  * emulate more closely the "process to completion" that we have with
357  * fastforwarding) and accounting for the work performed in low level
358  * handling and forwarding.
359  */
360 
361 static struct timeval poll_start_t;
362 
363 void
364 netisr_pollmore()
365 {
366 	struct timeval t;
367 	int kern_load;
368 
369 	mtx_lock(&poll_mtx);
370 	if (!netisr_pollmore_scheduled) {
371 		mtx_unlock(&poll_mtx);
372 		return;
373 	}
374 	netisr_pollmore_scheduled = 0;
375 	phase = 5;
376 	if (residual_burst > 0) {
377 		netisr_poll_scheduled = 1;
378 		netisr_pollmore_scheduled = 1;
379 		netisr_sched_poll();
380 		mtx_unlock(&poll_mtx);
381 		/* will run immediately on return, followed by netisrs */
382 		return;
383 	}
384 	/* here we can account time spent in netisr's in this tick */
385 	microuptime(&t);
386 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
387 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
388 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
389 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
390 		if (poll_burst > 1)
391 			poll_burst--;
392 	} else {
393 		if (poll_burst < poll_burst_max)
394 			poll_burst++;
395 	}
396 
397 	pending_polls--;
398 	if (pending_polls == 0) /* we are done */
399 		phase = 0;
400 	else {
401 		/*
402 		 * Last cycle was long and caused us to miss one or more
403 		 * hardclock ticks. Restart processing again, but slightly
404 		 * reduce the burst size to prevent that this happens again.
405 		 */
406 		poll_burst -= (poll_burst / 8);
407 		if (poll_burst < 1)
408 			poll_burst = 1;
409 		netisr_poll_scheduled = 1;
410 		netisr_pollmore_scheduled = 1;
411 		netisr_sched_poll();
412 		phase = 6;
413 	}
414 	mtx_unlock(&poll_mtx);
415 }
416 
417 /*
418  * netisr_poll is typically scheduled once per tick.
419  */
420 void
421 netisr_poll(void)
422 {
423 	int i, cycles;
424 	enum poll_cmd arg = POLL_ONLY;
425 
426 	mtx_lock(&poll_mtx);
427 	if (!netisr_poll_scheduled) {
428 		mtx_unlock(&poll_mtx);
429 		return;
430 	}
431 	netisr_poll_scheduled = 0;
432 	phase = 3;
433 	if (residual_burst == 0) { /* first call in this tick */
434 		microuptime(&poll_start_t);
435 		if (++reg_frac_count == reg_frac) {
436 			arg = POLL_AND_CHECK_STATUS;
437 			reg_frac_count = 0;
438 		}
439 
440 		residual_burst = poll_burst;
441 	}
442 	cycles = (residual_burst < poll_each_burst) ?
443 		residual_burst : poll_each_burst;
444 	residual_burst -= cycles;
445 
446 	for (i = 0 ; i < poll_handlers ; i++)
447 		pr[i].handler(pr[i].ifp, arg, cycles);
448 
449 	phase = 4;
450 	mtx_unlock(&poll_mtx);
451 }
452 
453 /*
454  * Try to register routine for polling. Returns 0 if successful
455  * (and polling should be enabled), error code otherwise.
456  * A device is not supposed to register itself multiple times.
457  *
458  * This is called from within the *_ioctl() functions.
459  */
460 int
461 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
462 {
463 	int i;
464 
465 	KASSERT(h != NULL, ("%s: handler is NULL", __func__));
466 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
467 
468 	mtx_lock(&poll_mtx);
469 	if (poll_handlers >= POLL_LIST_LEN) {
470 		/*
471 		 * List full, cannot register more entries.
472 		 * This should never happen; if it does, it is probably a
473 		 * broken driver trying to register multiple times. Checking
474 		 * this at runtime is expensive, and won't solve the problem
475 		 * anyways, so just report a few times and then give up.
476 		 */
477 		static int verbose = 10 ;
478 		if (verbose >0) {
479 			log(LOG_ERR, "poll handlers list full, "
480 			    "maybe a broken driver ?\n");
481 			verbose--;
482 		}
483 		mtx_unlock(&poll_mtx);
484 		return (ENOMEM); /* no polling for you */
485 	}
486 
487 	for (i = 0 ; i < poll_handlers ; i++)
488 		if (pr[i].ifp == ifp && pr[i].handler != NULL) {
489 			mtx_unlock(&poll_mtx);
490 			log(LOG_DEBUG, "ether_poll_register: %s: handler"
491 			    " already registered\n", ifp->if_xname);
492 			return (EEXIST);
493 		}
494 
495 	pr[poll_handlers].handler = h;
496 	pr[poll_handlers].ifp = ifp;
497 	poll_handlers++;
498 	mtx_unlock(&poll_mtx);
499 	if (idlepoll_sleeping)
500 		wakeup(&idlepoll_sleeping);
501 	return (0);
502 }
503 
504 /*
505  * Remove interface from the polling list. Called from *_ioctl(), too.
506  */
507 int
508 ether_poll_deregister(struct ifnet *ifp)
509 {
510 	int i;
511 
512 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
513 
514 	mtx_lock(&poll_mtx);
515 
516 	for (i = 0 ; i < poll_handlers ; i++)
517 		if (pr[i].ifp == ifp) /* found it */
518 			break;
519 	if (i == poll_handlers) {
520 		log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
521 		    ifp->if_xname);
522 		mtx_unlock(&poll_mtx);
523 		return (ENOENT);
524 	}
525 	poll_handlers--;
526 	if (i < poll_handlers) { /* Last entry replaces this one. */
527 		pr[i].handler = pr[poll_handlers].handler;
528 		pr[i].ifp = pr[poll_handlers].ifp;
529 	}
530 	mtx_unlock(&poll_mtx);
531 	return (0);
532 }
533 
534 static void
535 poll_idle(void)
536 {
537 	struct thread *td = curthread;
538 	struct rtprio rtp;
539 
540 	rtp.prio = RTP_PRIO_MAX;	/* lowest priority */
541 	rtp.type = RTP_PRIO_IDLE;
542 	PROC_SLOCK(td->td_proc);
543 	rtp_to_pri(&rtp, td);
544 	PROC_SUNLOCK(td->td_proc);
545 
546 	for (;;) {
547 		if (poll_in_idle_loop && poll_handlers > 0) {
548 			idlepoll_sleeping = 0;
549 			ether_poll(poll_each_burst);
550 			thread_lock(td);
551 			mi_switch(SW_VOL, NULL);
552 			thread_unlock(td);
553 		} else {
554 			idlepoll_sleeping = 1;
555 			tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
556 		}
557 	}
558 }
559 
560 static struct proc *idlepoll;
561 static struct kproc_desc idlepoll_kp = {
562 	 "idlepoll",
563 	 poll_idle,
564 	 &idlepoll
565 };
566 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
567     &idlepoll_kp);
568