xref: /freebsd/sys/kern/kern_poll.c (revision b197d4b893974c9eb4d7b38704c6d5c486235d6f)
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("$FreeBSD$");
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/epoch.h>
41 #include <sys/eventhandler.h>
42 #include <sys/resourcevar.h>
43 #include <sys/socket.h>			/* needed by net/if.h		*/
44 #include <sys/sockio.h>
45 #include <sys/sysctl.h>
46 #include <sys/syslog.h>
47 
48 #include <net/if.h>
49 #include <net/if_var.h>
50 #include <net/netisr.h>			/* for NETISR_POLL		*/
51 #include <net/vnet.h>
52 
53 void hardclock_device_poll(void);	/* hook from hardclock		*/
54 
55 static struct mtx	poll_mtx;
56 
57 /*
58  * Polling support for [network] device drivers.
59  *
60  * Drivers which support this feature can register with the
61  * polling code.
62  *
63  * If registration is successful, the driver must disable interrupts,
64  * and further I/O is performed through the handler, which is invoked
65  * (at least once per clock tick) with 3 arguments: the "arg" passed at
66  * register time (a struct ifnet pointer), a command, and a "count" limit.
67  *
68  * The command can be one of the following:
69  *  POLL_ONLY: quick move of "count" packets from input/output queues.
70  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
71  *	other more expensive operations. This command is issued periodically
72  *	but less frequently than POLL_ONLY.
73  *
74  * The count limit specifies how much work the handler can do during the
75  * call -- typically this is the number of packets to be received, or
76  * transmitted, etc. (drivers are free to interpret this number, as long
77  * as the max time spent in the function grows roughly linearly with the
78  * count).
79  *
80  * Polling is enabled and disabled via setting IFCAP_POLLING flag on
81  * the interface. The driver ioctl handler should register interface
82  * with polling and disable interrupts, if registration was successful.
83  *
84  * A second variable controls the sharing of CPU between polling/kernel
85  * network processing, and other activities (typically userlevel tasks):
86  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
87  * of CPU allocated to user tasks. CPU is allocated proportionally to the
88  * shares, by dynamically adjusting the "count" (poll_burst).
89  *
90  * Other parameters can should be left to their default values.
91  * The following constraints hold
92  *
93  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
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	20000
99 
100 static uint32_t poll_burst = 5;
101 static uint32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
102 static uint32_t poll_each_burst = 5;
103 
104 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
105     "Device polling parameters");
106 
107 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
108 	&poll_burst, 0, "Current polling burst size");
109 
110 static int	netisr_poll_scheduled;
111 static int	netisr_pollmore_scheduled;
112 static int	poll_shutting_down;
113 
114 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
115 {
116 	uint32_t val = poll_burst_max;
117 	int error;
118 
119 	error = sysctl_handle_int(oidp, &val, 0, req);
120 	if (error || !req->newptr )
121 		return (error);
122 	if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
123 		return (EINVAL);
124 
125 	mtx_lock(&poll_mtx);
126 	poll_burst_max = val;
127 	if (poll_burst > poll_burst_max)
128 		poll_burst = poll_burst_max;
129 	if (poll_each_burst > poll_burst_max)
130 		poll_each_burst = MIN_POLL_BURST_MAX;
131 	mtx_unlock(&poll_mtx);
132 
133 	return (0);
134 }
135 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max,
136     CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
137     poll_burst_max_sysctl, "I",
138     "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,
162     CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
163     poll_each_burst_sysctl, "I",
164     "Max size of each burst");
165 
166 static uint32_t poll_in_idle_loop=0;	/* do we poll in idle loop ? */
167 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
168 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
169 
170 static uint32_t user_frac = 50;
171 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
172 {
173 	uint32_t val = user_frac;
174 	int error;
175 
176 	error = sysctl_handle_int(oidp, &val, 0, req);
177 	if (error || !req->newptr )
178 		return (error);
179 	if (val > 99)
180 		return (EINVAL);
181 
182 	mtx_lock(&poll_mtx);
183 	user_frac = val;
184 	mtx_unlock(&poll_mtx);
185 
186 	return (0);
187 }
188 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac,
189     CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
190     user_frac_sysctl, "I",
191     "Desired user fraction of cpu time");
192 
193 static uint32_t reg_frac_count = 0;
194 static uint32_t reg_frac = 20 ;
195 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
196 {
197 	uint32_t val = reg_frac;
198 	int error;
199 
200 	error = sysctl_handle_int(oidp, &val, 0, req);
201 	if (error || !req->newptr )
202 		return (error);
203 	if (val < 1 || val > hz)
204 		return (EINVAL);
205 
206 	mtx_lock(&poll_mtx);
207 	reg_frac = val;
208 	if (reg_frac_count >= reg_frac)
209 		reg_frac_count = 0;
210 	mtx_unlock(&poll_mtx);
211 
212 	return (0);
213 }
214 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac,
215     CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
216     reg_frac_sysctl, "I",
217     "Every this many cycles check registers");
218 
219 static uint32_t short_ticks;
220 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
221 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
222 
223 static uint32_t lost_polls;
224 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
225 	&lost_polls, 0, "How many times we would have lost a poll tick");
226 
227 static uint32_t pending_polls;
228 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
229 	&pending_polls, 0, "Do we need to poll again");
230 
231 static int residual_burst = 0;
232 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
233 	&residual_burst, 0, "# of residual cycles in burst");
234 
235 static uint32_t poll_handlers; /* next free entry in pr[]. */
236 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
237 	&poll_handlers, 0, "Number of registered poll handlers");
238 
239 static uint32_t phase;
240 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
241 	&phase, 0, "Polling phase");
242 
243 static uint32_t suspect;
244 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
245 	&suspect, 0, "suspect event");
246 
247 static uint32_t stalled;
248 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
249 	&stalled, 0, "potential stalls");
250 
251 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
252 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
253 	&idlepoll_sleeping, 0, "idlepoll is sleeping");
254 
255 #define POLL_LIST_LEN  128
256 struct pollrec {
257 	poll_handler_t	*handler;
258 	struct ifnet	*ifp;
259 };
260 
261 static struct pollrec pr[POLL_LIST_LEN];
262 
263 static void
264 poll_shutdown(void *arg, int howto)
265 {
266 
267 	poll_shutting_down = 1;
268 }
269 
270 static void
271 init_device_poll(void)
272 {
273 
274 	mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
275 	EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
276 	    SHUTDOWN_PRI_LAST);
277 }
278 SYSINIT(device_poll, SI_SUB_SOFTINTR, SI_ORDER_MIDDLE, init_device_poll, NULL);
279 
280 /*
281  * Hook from hardclock. Tries to schedule a netisr, but keeps track
282  * of lost ticks due to the previous handler taking too long.
283  * Normally, this should not happen, because polling handler should
284  * run for a short time. However, in some cases (e.g. when there are
285  * changes in link status etc.) the drivers take a very long time
286  * (even in the order of milliseconds) to reset and reconfigure the
287  * device, causing apparent lost polls.
288  *
289  * The first part of the code is just for debugging purposes, and tries
290  * to count how often hardclock ticks are shorter than they should,
291  * meaning either stray interrupts or delayed events.
292  */
293 void
294 hardclock_device_poll(void)
295 {
296 	static struct timeval prev_t, t;
297 	int delta;
298 
299 	if (poll_handlers == 0 || poll_shutting_down)
300 		return;
301 
302 	microuptime(&t);
303 	delta = (t.tv_usec - prev_t.tv_usec) +
304 		(t.tv_sec - prev_t.tv_sec)*1000000;
305 	if (delta * hz < 500000)
306 		short_ticks++;
307 	else
308 		prev_t = t;
309 
310 	if (pending_polls > 100) {
311 		/*
312 		 * Too much, assume it has stalled (not always true
313 		 * see comment above).
314 		 */
315 		stalled++;
316 		pending_polls = 0;
317 		phase = 0;
318 	}
319 
320 	if (phase <= 2) {
321 		if (phase != 0)
322 			suspect++;
323 		phase = 1;
324 		netisr_poll_scheduled = 1;
325 		netisr_pollmore_scheduled = 1;
326 		netisr_sched_poll();
327 		phase = 2;
328 	}
329 	if (pending_polls++ > 0)
330 		lost_polls++;
331 }
332 
333 /*
334  * ether_poll is called from the idle loop.
335  */
336 static void
337 ether_poll(int count)
338 {
339 	struct epoch_tracker et;
340 	int i;
341 
342 	mtx_lock(&poll_mtx);
343 
344 	if (count > poll_each_burst)
345 		count = poll_each_burst;
346 
347 	NET_EPOCH_ENTER(et);
348 	for (i = 0 ; i < poll_handlers ; i++)
349 		pr[i].handler(pr[i].ifp, POLL_ONLY, count);
350 	NET_EPOCH_EXIT(et);
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(void)
375 {
376 	struct timeval t;
377 	int kern_load;
378 
379 	if (poll_handlers == 0)
380 		return;
381 
382 	mtx_lock(&poll_mtx);
383 	if (!netisr_pollmore_scheduled) {
384 		mtx_unlock(&poll_mtx);
385 		return;
386 	}
387 	netisr_pollmore_scheduled = 0;
388 	phase = 5;
389 	if (residual_burst > 0) {
390 		netisr_poll_scheduled = 1;
391 		netisr_pollmore_scheduled = 1;
392 		netisr_sched_poll();
393 		mtx_unlock(&poll_mtx);
394 		/* will run immediately on return, followed by netisrs */
395 		return;
396 	}
397 	/* here we can account time spent in netisr's in this tick */
398 	microuptime(&t);
399 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
400 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
401 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
402 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
403 		if (poll_burst > 1)
404 			poll_burst--;
405 	} else {
406 		if (poll_burst < poll_burst_max)
407 			poll_burst++;
408 	}
409 
410 	pending_polls--;
411 	if (pending_polls == 0) /* we are done */
412 		phase = 0;
413 	else {
414 		/*
415 		 * Last cycle was long and caused us to miss one or more
416 		 * hardclock ticks. Restart processing again, but slightly
417 		 * reduce the burst size to prevent that this happens again.
418 		 */
419 		poll_burst -= (poll_burst / 8);
420 		if (poll_burst < 1)
421 			poll_burst = 1;
422 		netisr_poll_scheduled = 1;
423 		netisr_pollmore_scheduled = 1;
424 		netisr_sched_poll();
425 		phase = 6;
426 	}
427 	mtx_unlock(&poll_mtx);
428 }
429 
430 /*
431  * netisr_poll is typically scheduled once per tick.
432  */
433 void
434 netisr_poll(void)
435 {
436 	int i, cycles;
437 	enum poll_cmd arg = POLL_ONLY;
438 
439 	NET_EPOCH_ASSERT();
440 
441 	if (poll_handlers == 0)
442 		return;
443 
444 	mtx_lock(&poll_mtx);
445 	if (!netisr_poll_scheduled) {
446 		mtx_unlock(&poll_mtx);
447 		return;
448 	}
449 	netisr_poll_scheduled = 0;
450 	phase = 3;
451 	if (residual_burst == 0) { /* first call in this tick */
452 		microuptime(&poll_start_t);
453 		if (++reg_frac_count == reg_frac) {
454 			arg = POLL_AND_CHECK_STATUS;
455 			reg_frac_count = 0;
456 		}
457 
458 		residual_burst = poll_burst;
459 	}
460 	cycles = (residual_burst < poll_each_burst) ?
461 		residual_burst : poll_each_burst;
462 	residual_burst -= cycles;
463 
464 	for (i = 0 ; i < poll_handlers ; i++)
465 		pr[i].handler(pr[i].ifp, arg, cycles);
466 
467 	phase = 4;
468 	mtx_unlock(&poll_mtx);
469 }
470 
471 /*
472  * Try to register routine for polling. Returns 0 if successful
473  * (and polling should be enabled), error code otherwise.
474  * A device is not supposed to register itself multiple times.
475  *
476  * This is called from within the *_ioctl() functions.
477  */
478 int
479 ether_poll_register(poll_handler_t *h, if_t ifp)
480 {
481 	int i;
482 
483 	KASSERT(h != NULL, ("%s: handler is NULL", __func__));
484 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
485 
486 	mtx_lock(&poll_mtx);
487 	if (poll_handlers >= POLL_LIST_LEN) {
488 		/*
489 		 * List full, cannot register more entries.
490 		 * This should never happen; if it does, it is probably a
491 		 * broken driver trying to register multiple times. Checking
492 		 * this at runtime is expensive, and won't solve the problem
493 		 * anyways, so just report a few times and then give up.
494 		 */
495 		static int verbose = 10 ;
496 		if (verbose >0) {
497 			log(LOG_ERR, "poll handlers list full, "
498 			    "maybe a broken driver ?\n");
499 			verbose--;
500 		}
501 		mtx_unlock(&poll_mtx);
502 		return (ENOMEM); /* no polling for you */
503 	}
504 
505 	for (i = 0 ; i < poll_handlers ; i++)
506 		if (pr[i].ifp == ifp && pr[i].handler != NULL) {
507 			mtx_unlock(&poll_mtx);
508 			log(LOG_DEBUG, "ether_poll_register: %s: handler"
509 			    " already registered\n", ifp->if_xname);
510 			return (EEXIST);
511 		}
512 
513 	pr[poll_handlers].handler = h;
514 	pr[poll_handlers].ifp = ifp;
515 	poll_handlers++;
516 	mtx_unlock(&poll_mtx);
517 	if (idlepoll_sleeping)
518 		wakeup(&idlepoll_sleeping);
519 	return (0);
520 }
521 
522 /*
523  * Remove interface from the polling list. Called from *_ioctl(), too.
524  */
525 int
526 ether_poll_deregister(if_t ifp)
527 {
528 	int i;
529 
530 	KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
531 
532 	mtx_lock(&poll_mtx);
533 
534 	for (i = 0 ; i < poll_handlers ; i++)
535 		if (pr[i].ifp == ifp) /* found it */
536 			break;
537 	if (i == poll_handlers) {
538 		log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
539 		    ifp->if_xname);
540 		mtx_unlock(&poll_mtx);
541 		return (ENOENT);
542 	}
543 	poll_handlers--;
544 	if (i < poll_handlers) { /* Last entry replaces this one. */
545 		pr[i].handler = pr[poll_handlers].handler;
546 		pr[i].ifp = pr[poll_handlers].ifp;
547 	}
548 	mtx_unlock(&poll_mtx);
549 	return (0);
550 }
551 
552 static void
553 poll_idle(void)
554 {
555 	struct thread *td = curthread;
556 	struct rtprio rtp;
557 
558 	rtp.prio = RTP_PRIO_MAX;	/* lowest priority */
559 	rtp.type = RTP_PRIO_IDLE;
560 	PROC_SLOCK(td->td_proc);
561 	rtp_to_pri(&rtp, td);
562 	PROC_SUNLOCK(td->td_proc);
563 
564 	for (;;) {
565 		if (poll_in_idle_loop && poll_handlers > 0) {
566 			idlepoll_sleeping = 0;
567 			ether_poll(poll_each_burst);
568 			thread_lock(td);
569 			mi_switch(SW_VOL);
570 		} else {
571 			idlepoll_sleeping = 1;
572 			tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
573 		}
574 	}
575 }
576 
577 static struct proc *idlepoll;
578 static struct kproc_desc idlepoll_kp = {
579 	 "idlepoll",
580 	 poll_idle,
581 	 &idlepoll
582 };
583 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
584     &idlepoll_kp);
585