xref: /dflybsd-src/sys/net/if_poll.c (revision 16348f3142ec62b5d9ebc02695e606859d8b41f7)
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  * $FreeBSD: src/sys/kern/kern_poll.c,v 1.2.2.4 2002/06/27 23:26:33 luigi Exp $
28  */
29 
30 #include "opt_ifpoll.h"
31 
32 #include <sys/param.h>
33 #include <sys/kernel.h>
34 #include <sys/ktr.h>
35 #include <sys/malloc.h>
36 #include <sys/serialize.h>
37 #include <sys/socket.h>
38 #include <sys/sysctl.h>
39 
40 #include <sys/thread2.h>
41 #include <sys/msgport2.h>
42 
43 #include <machine/atomic.h>
44 #include <machine/clock.h>
45 #include <machine/smp.h>
46 
47 #include <net/if.h>
48 #include <net/if_poll.h>
49 #include <net/netmsg2.h>
50 
51 /*
52  * Polling support for network device drivers.
53  *
54  * Drivers which support this feature try to register one status polling
55  * handler and several TX/RX polling handlers with the polling code.
56  * If interface's if_npoll is called with non-NULL second argument, then
57  * a register operation is requested, else a deregister operation is
58  * requested.  If the requested operation is "register", driver should
59  * setup the ifpoll_info passed in accoding its own needs:
60  *   ifpoll_info.ifpi_status.status_func == NULL
61  *     No status polling handler will be installed on CPU(0)
62  *   ifpoll_info.ifpi_rx[n].poll_func == NULL
63  *     No RX polling handler will be installed on CPU(n)
64  *   ifpoll_info.ifpi_tx[n].poll_func == NULL
65  *     No TX polling handler will be installed on CPU(n)
66  *
67  * RX is polled at the specified polling frequency (net.ifpoll.X.pollhz).
68  * TX and status polling could be done at lower frequency than RX frequency
69  * (net.ifpoll.0.status_frac and net.ifpoll.X.tx_frac).  To avoid systimer
70  * staggering at high frequency, RX systimer gives TX and status polling a
71  * piggyback (XXX).
72  *
73  * All of the registered polling handlers are called only if the interface
74  * is marked as 'IFF_RUNNING and IFF_NPOLLING'.  However, the interface's
75  * register and deregister function (ifnet.if_npoll) will be called even
76  * if interface is not marked with 'IFF_RUNNING'.
77  *
78  * If registration is successful, the driver must disable interrupts,
79  * and further I/O is performed through the TX/RX polling handler, which
80  * are invoked (at least once per clock tick) with 3 arguments: the "arg"
81  * passed at register time, a struct ifnet pointer, and a "count" limit.
82  * The registered serializer will be held before calling the related
83  * polling handler.
84  *
85  * The count limit specifies how much work the handler can do during the
86  * call -- typically this is the number of packets to be received, or
87  * transmitted, etc. (drivers are free to interpret this number, as long
88  * as the max time spent in the function grows roughly linearly with the
89  * count).
90  *
91  * A second variable controls the sharing of CPU between polling/kernel
92  * network processing, and other activities (typically userlevel tasks):
93  * net.ifpoll.X.{rx,tx}.user_frac (between 0 and 100, default 50) sets the
94  * share of CPU allocated to user tasks.  CPU is allocated proportionally
95  * to the shares, by dynamically adjusting the "count" (poll_burst).
96  *
97  * Other parameters can should be left to their default values.
98  * The following constraints hold
99  *
100  *	1 <= poll_burst <= poll_burst_max
101  *	1 <= poll_each_burst <= poll_burst_max
102  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
103  */
104 
105 #define IFPOLL_LIST_LEN		128
106 #define IFPOLL_FREQ_MAX		30000
107 
108 #define MIN_IOPOLL_BURST_MAX	10
109 #define MAX_IOPOLL_BURST_MAX	1000
110 #define IOPOLL_BURST_MAX	150	/* good for 100Mbit net and HZ=1000 */
111 
112 #define IOPOLL_EACH_BURST	5
113 
114 #define IFPOLL_FREQ_DEFAULT	2000
115 
116 #define IFPOLL_TXFRAC_DEFAULT	1	/* 1/2 of the pollhz */
117 #define IFPOLL_STFRAC_DEFAULT	19	/* 1/20 of the pollhz */
118 
119 #define IFPOLL_RX		0x1
120 #define IFPOLL_TX		0x2
121 
122 union ifpoll_time {
123 	struct timeval		tv;
124 	uint64_t		tsc;
125 };
126 
127 struct iopoll_rec {
128 	struct lwkt_serialize	*serializer;
129 	struct ifnet		*ifp;
130 	void			*arg;
131 	ifpoll_iofn_t		poll_func;
132 };
133 
134 struct iopoll_ctx {
135 	union ifpoll_time	prev_t;
136 	u_long			short_ticks;		/* statistics */
137 	u_long			lost_polls;		/* statistics */
138 	u_long			suspect;		/* statistics */
139 	u_long			stalled;		/* statistics */
140 	uint32_t		pending_polls;		/* state */
141 
142 	struct netmsg_base	poll_netmsg;
143 	struct netmsg_base	poll_more_netmsg;
144 
145 	int			poll_cpuid;
146 	int			pollhz;
147 	uint32_t		phase;			/* state */
148 	int			residual_burst;		/* state */
149 	uint32_t		poll_each_burst;	/* tunable */
150 	union ifpoll_time	poll_start_t;		/* state */
151 
152 	uint32_t		poll_burst;		/* state */
153 	uint32_t		poll_burst_max;		/* tunable */
154 	uint32_t		user_frac;		/* tunable */
155 	uint32_t		kern_frac;		/* state */
156 
157 	uint32_t		poll_handlers; /* next free entry in pr[]. */
158 	struct iopoll_rec	pr[IFPOLL_LIST_LEN];
159 
160 	struct sysctl_ctx_list	poll_sysctl_ctx;
161 	struct sysctl_oid	*poll_sysctl_tree;
162 } __cachealign;
163 
164 struct poll_comm {
165 	struct systimer		pollclock;
166 	int			poll_cpuid;
167 
168 	int			stfrac_count;		/* state */
169 	int			poll_stfrac;		/* tunable */
170 
171 	int			txfrac_count;		/* state */
172 	int			poll_txfrac;		/* tunable */
173 
174 	int			pollhz;			/* tunable */
175 
176 	struct sysctl_ctx_list	sysctl_ctx;
177 	struct sysctl_oid	*sysctl_tree;
178 } __cachealign;
179 
180 struct stpoll_rec {
181 	struct lwkt_serialize	*serializer;
182 	struct ifnet		*ifp;
183 	ifpoll_stfn_t		status_func;
184 };
185 
186 struct stpoll_ctx {
187 	struct netmsg_base	poll_netmsg;
188 
189 	int			pollhz;
190 
191 	uint32_t		poll_handlers; /* next free entry in pr[]. */
192 	struct stpoll_rec	pr[IFPOLL_LIST_LEN];
193 
194 	struct sysctl_ctx_list	poll_sysctl_ctx;
195 	struct sysctl_oid	*poll_sysctl_tree;
196 } __cachealign;
197 
198 struct iopoll_sysctl_netmsg {
199 	struct netmsg_base	base;
200 	struct iopoll_ctx	*ctx;
201 };
202 
203 void		ifpoll_init_pcpu(int);
204 static void	ifpoll_register_handler(netmsg_t);
205 static void	ifpoll_deregister_handler(netmsg_t);
206 
207 /*
208  * Status polling
209  */
210 static void	stpoll_init(void);
211 static void	stpoll_handler(netmsg_t);
212 static void	stpoll_clock(struct stpoll_ctx *);
213 static int	stpoll_register(struct ifnet *, const struct ifpoll_status *);
214 static int	stpoll_deregister(struct ifnet *);
215 
216 /*
217  * RX/TX polling
218  */
219 static struct iopoll_ctx *iopoll_ctx_create(int, int);
220 static void	iopoll_init(int);
221 static void	rxpoll_handler(netmsg_t);
222 static void	txpoll_handler(netmsg_t);
223 static void	rxpollmore_handler(netmsg_t);
224 static void	txpollmore_handler(netmsg_t);
225 static void	iopoll_clock(struct iopoll_ctx *);
226 static int	iopoll_register(struct ifnet *, struct iopoll_ctx *,
227 		    const struct ifpoll_io *);
228 static int	iopoll_deregister(struct ifnet *, struct iopoll_ctx *);
229 
230 static void	iopoll_add_sysctl(struct sysctl_ctx_list *,
231 		    struct sysctl_oid_list *, struct iopoll_ctx *, int);
232 static void	sysctl_burstmax_handler(netmsg_t);
233 static int	sysctl_burstmax(SYSCTL_HANDLER_ARGS);
234 static void	sysctl_eachburst_handler(netmsg_t);
235 static int	sysctl_eachburst(SYSCTL_HANDLER_ARGS);
236 
237 /*
238  * Common functions
239  */
240 static void	poll_comm_init(int);
241 static void	poll_comm_start(int);
242 static void	poll_comm_adjust_pollhz(struct poll_comm *);
243 static void	poll_comm_systimer0(systimer_t, int, struct intrframe *);
244 static void	poll_comm_systimer(systimer_t, int, struct intrframe *);
245 static void	sysctl_pollhz_handler(netmsg_t);
246 static void	sysctl_stfrac_handler(netmsg_t);
247 static void	sysctl_txfrac_handler(netmsg_t);
248 static int	sysctl_pollhz(SYSCTL_HANDLER_ARGS);
249 static int	sysctl_stfrac(SYSCTL_HANDLER_ARGS);
250 static int	sysctl_txfrac(SYSCTL_HANDLER_ARGS);
251 
252 static struct stpoll_ctx	stpoll_context;
253 static struct poll_comm		*poll_common[MAXCPU];
254 static struct iopoll_ctx	*rxpoll_context[MAXCPU];
255 static struct iopoll_ctx	*txpoll_context[MAXCPU];
256 
257 SYSCTL_NODE(_net, OID_AUTO, ifpoll, CTLFLAG_RW, 0,
258 	    "Network device polling parameters");
259 
260 static int	iopoll_burst_max = IOPOLL_BURST_MAX;
261 static int	iopoll_each_burst = IOPOLL_EACH_BURST;
262 
263 static int	ifpoll_pollhz = IFPOLL_FREQ_DEFAULT;
264 static int	ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
265 static int	ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
266 
267 TUNABLE_INT("net.ifpoll.burst_max", &iopoll_burst_max);
268 TUNABLE_INT("net.ifpoll.each_burst", &iopoll_each_burst);
269 TUNABLE_INT("net.ifpoll.pollhz", &ifpoll_pollhz);
270 TUNABLE_INT("net.ifpoll.status_frac", &ifpoll_stfrac);
271 TUNABLE_INT("net.ifpoll.tx_frac", &ifpoll_txfrac);
272 
273 static __inline void
274 ifpoll_sendmsg_oncpu(netmsg_t msg)
275 {
276 	if (msg->lmsg.ms_flags & MSGF_DONE)
277 		lwkt_sendmsg(netisr_portfn(mycpuid), &msg->lmsg);
278 }
279 
280 static __inline void
281 sched_stpoll(struct stpoll_ctx *st_ctx)
282 {
283 	ifpoll_sendmsg_oncpu((netmsg_t)&st_ctx->poll_netmsg);
284 }
285 
286 static __inline void
287 sched_iopoll(struct iopoll_ctx *io_ctx)
288 {
289 	ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_netmsg);
290 }
291 
292 static __inline void
293 sched_iopollmore(struct iopoll_ctx *io_ctx)
294 {
295 	ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_more_netmsg);
296 }
297 
298 static __inline void
299 ifpoll_time_get(union ifpoll_time *t)
300 {
301 	if (__predict_true(tsc_present))
302 		t->tsc = rdtsc();
303 	else
304 		microuptime(&t->tv);
305 }
306 
307 /* Return time diff in us */
308 static __inline int
309 ifpoll_time_diff(const union ifpoll_time *s, const union ifpoll_time *e)
310 {
311 	if (__predict_true(tsc_present)) {
312 		return (((e->tsc - s->tsc) * 1000000) / tsc_frequency);
313 	} else {
314 		return ((e->tv.tv_usec - s->tv.tv_usec) +
315 			(e->tv.tv_sec - s->tv.tv_sec) * 1000000);
316 	}
317 }
318 
319 /*
320  * Initialize per-cpu qpolling(4) context.  Called from kern_clock.c:
321  */
322 void
323 ifpoll_init_pcpu(int cpuid)
324 {
325 	if (cpuid >= ncpus2)
326 		return;
327 
328 	poll_comm_init(cpuid);
329 
330 	if (cpuid == 0)
331 		stpoll_init();
332 	iopoll_init(cpuid);
333 
334 	poll_comm_start(cpuid);
335 }
336 
337 int
338 ifpoll_register(struct ifnet *ifp)
339 {
340 	struct ifpoll_info *info;
341 	struct netmsg_base nmsg;
342 	int error;
343 
344 	if (ifp->if_npoll == NULL) {
345 		/* Device does not support polling */
346 		return EOPNOTSUPP;
347 	}
348 
349 	info = kmalloc(sizeof(*info), M_TEMP, M_WAITOK | M_ZERO);
350 
351 	/*
352 	 * Attempt to register.  Interlock with IFF_NPOLLING.
353 	 */
354 
355 	ifnet_serialize_all(ifp);
356 
357 	if (ifp->if_flags & IFF_NPOLLING) {
358 		/* Already polling */
359 		ifnet_deserialize_all(ifp);
360 		kfree(info, M_TEMP);
361 		return EBUSY;
362 	}
363 
364 	info->ifpi_ifp = ifp;
365 
366 	ifp->if_flags |= IFF_NPOLLING;
367 	ifp->if_npoll(ifp, info);
368 	KASSERT(ifp->if_npoll_cpuid >= 0, ("invalid npoll cpuid"));
369 
370 	ifnet_deserialize_all(ifp);
371 
372 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
373 		    0, ifpoll_register_handler);
374 	nmsg.lmsg.u.ms_resultp = info;
375 
376 	error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
377 	if (error) {
378 		if (!ifpoll_deregister(ifp)) {
379 			if_printf(ifp, "ifpoll_register: "
380 				  "ifpoll_deregister failed!\n");
381 		}
382 	}
383 
384 	kfree(info, M_TEMP);
385 	return error;
386 }
387 
388 int
389 ifpoll_deregister(struct ifnet *ifp)
390 {
391 	struct netmsg_base nmsg;
392 	int error;
393 
394 	if (ifp->if_npoll == NULL)
395 		return EOPNOTSUPP;
396 
397 	ifnet_serialize_all(ifp);
398 
399 	if ((ifp->if_flags & IFF_NPOLLING) == 0) {
400 		ifnet_deserialize_all(ifp);
401 		return EINVAL;
402 	}
403 	ifp->if_flags &= ~IFF_NPOLLING;
404 
405 	ifnet_deserialize_all(ifp);
406 
407 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
408 		    0, ifpoll_deregister_handler);
409 	nmsg.lmsg.u.ms_resultp = ifp;
410 
411 	error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
412 	if (!error) {
413 		ifnet_serialize_all(ifp);
414 		ifp->if_npoll(ifp, NULL);
415 		KASSERT(ifp->if_npoll_cpuid < 0, ("invalid npoll cpuid"));
416 		ifnet_deserialize_all(ifp);
417 	}
418 	return error;
419 }
420 
421 static void
422 ifpoll_register_handler(netmsg_t nmsg)
423 {
424 	const struct ifpoll_info *info = nmsg->lmsg.u.ms_resultp;
425 	int cpuid = mycpuid, nextcpu;
426 	int error;
427 
428 	KKASSERT(cpuid < ncpus2);
429 	KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
430 
431 	if (cpuid == 0) {
432 		error = stpoll_register(info->ifpi_ifp, &info->ifpi_status);
433 		if (error)
434 			goto failed;
435 	}
436 
437 	error = iopoll_register(info->ifpi_ifp, rxpoll_context[cpuid],
438 				&info->ifpi_rx[cpuid]);
439 	if (error)
440 		goto failed;
441 
442 	error = iopoll_register(info->ifpi_ifp, txpoll_context[cpuid],
443 				&info->ifpi_tx[cpuid]);
444 	if (error)
445 		goto failed;
446 
447 	/* Adjust polling frequency, after all registration is done */
448 	poll_comm_adjust_pollhz(poll_common[cpuid]);
449 
450 	nextcpu = cpuid + 1;
451 	if (nextcpu < ncpus2)
452 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
453 	else
454 		lwkt_replymsg(&nmsg->lmsg, 0);
455 	return;
456 failed:
457 	lwkt_replymsg(&nmsg->lmsg, error);
458 }
459 
460 static void
461 ifpoll_deregister_handler(netmsg_t nmsg)
462 {
463 	struct ifnet *ifp = nmsg->lmsg.u.ms_resultp;
464 	int cpuid = mycpuid, nextcpu;
465 
466 	KKASSERT(cpuid < ncpus2);
467 	KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
468 
469 	/* Ignore errors */
470 	if (cpuid == 0)
471 		stpoll_deregister(ifp);
472 	iopoll_deregister(ifp, rxpoll_context[cpuid]);
473 	iopoll_deregister(ifp, txpoll_context[cpuid]);
474 
475 	/* Adjust polling frequency, after all deregistration is done */
476 	poll_comm_adjust_pollhz(poll_common[cpuid]);
477 
478 	nextcpu = cpuid + 1;
479 	if (nextcpu < ncpus2)
480 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
481 	else
482 		lwkt_replymsg(&nmsg->lmsg, 0);
483 }
484 
485 static void
486 stpoll_init(void)
487 {
488 	struct stpoll_ctx *st_ctx = &stpoll_context;
489 	const struct poll_comm *comm = poll_common[0];
490 
491 	st_ctx->pollhz = comm->pollhz / (comm->poll_stfrac + 1);
492 
493 	sysctl_ctx_init(&st_ctx->poll_sysctl_ctx);
494 	st_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&st_ctx->poll_sysctl_ctx,
495 				   SYSCTL_CHILDREN(comm->sysctl_tree),
496 				   OID_AUTO, "status", CTLFLAG_RD, 0, "");
497 
498 	SYSCTL_ADD_UINT(&st_ctx->poll_sysctl_ctx,
499 			SYSCTL_CHILDREN(st_ctx->poll_sysctl_tree),
500 			OID_AUTO, "handlers", CTLFLAG_RD,
501 			&st_ctx->poll_handlers, 0,
502 			"Number of registered status poll handlers");
503 
504 	netmsg_init(&st_ctx->poll_netmsg, NULL, &netisr_adone_rport,
505 		    0, stpoll_handler);
506 }
507 
508 /*
509  * stpoll_handler is scheduled by sched_stpoll when appropriate, typically
510  * once per polling systimer tick.
511  */
512 static void
513 stpoll_handler(netmsg_t msg)
514 {
515 	struct stpoll_ctx *st_ctx = &stpoll_context;
516 	struct thread *td = curthread;
517 	int i;
518 
519 	KKASSERT(&td->td_msgport == netisr_portfn(0));
520 
521 	crit_enter_quick(td);
522 
523 	/* Reply ASAP */
524 	lwkt_replymsg(&msg->lmsg, 0);
525 
526 	if (st_ctx->poll_handlers == 0) {
527 		crit_exit_quick(td);
528 		return;
529 	}
530 
531 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
532 		const struct stpoll_rec *rec = &st_ctx->pr[i];
533 		struct ifnet *ifp = rec->ifp;
534 
535 		if (!lwkt_serialize_try(rec->serializer))
536 			continue;
537 
538 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
539 		    (IFF_RUNNING | IFF_NPOLLING))
540 			rec->status_func(ifp, st_ctx->pollhz);
541 
542 		lwkt_serialize_exit(rec->serializer);
543 	}
544 
545 	crit_exit_quick(td);
546 }
547 
548 /*
549  * Hook from status poll systimer.  Tries to schedule an status poll.
550  * NOTE: Caller should hold critical section.
551  */
552 static void
553 stpoll_clock(struct stpoll_ctx *st_ctx)
554 {
555 	KKASSERT(mycpuid == 0);
556 
557 	if (st_ctx->poll_handlers == 0)
558 		return;
559 	sched_stpoll(st_ctx);
560 }
561 
562 static int
563 stpoll_register(struct ifnet *ifp, const struct ifpoll_status *st_rec)
564 {
565 	struct stpoll_ctx *st_ctx = &stpoll_context;
566 	int error;
567 
568 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
569 
570 	if (st_rec->status_func == NULL)
571 		return 0;
572 
573 	/*
574 	 * Check if there is room.
575 	 */
576 	if (st_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
577 		/*
578 		 * List full, cannot register more entries.
579 		 * This should never happen; if it does, it is probably a
580 		 * broken driver trying to register multiple times. Checking
581 		 * this at runtime is expensive, and won't solve the problem
582 		 * anyways, so just report a few times and then give up.
583 		 */
584 		static int verbose = 10; /* XXX */
585 
586 		if (verbose > 0) {
587 			kprintf("status poll handlers list full, "
588 				"maybe a broken driver ?\n");
589 			verbose--;
590 		}
591 		error = ENOENT;
592 	} else {
593 		struct stpoll_rec *rec = &st_ctx->pr[st_ctx->poll_handlers];
594 
595 		rec->ifp = ifp;
596 		rec->serializer = st_rec->serializer;
597 		rec->status_func = st_rec->status_func;
598 
599 		st_ctx->poll_handlers++;
600 		error = 0;
601 	}
602 	return error;
603 }
604 
605 static int
606 stpoll_deregister(struct ifnet *ifp)
607 {
608 	struct stpoll_ctx *st_ctx = &stpoll_context;
609 	int i, error;
610 
611 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
612 
613 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
614 		if (st_ctx->pr[i].ifp == ifp) /* Found it */
615 			break;
616 	}
617 	if (i == st_ctx->poll_handlers) {
618 		kprintf("stpoll_deregister: ifp not found!!!\n");
619 		error = ENOENT;
620 	} else {
621 		st_ctx->poll_handlers--;
622 		if (i < st_ctx->poll_handlers) {
623 			/* Last entry replaces this one. */
624 			st_ctx->pr[i] = st_ctx->pr[st_ctx->poll_handlers];
625 		}
626 		error = 0;
627 	}
628 	return error;
629 }
630 
631 static __inline void
632 iopoll_reset_state(struct iopoll_ctx *io_ctx)
633 {
634 	crit_enter();
635 	io_ctx->poll_burst = 5;
636 	io_ctx->pending_polls = 0;
637 	io_ctx->residual_burst = 0;
638 	io_ctx->phase = 0;
639 	io_ctx->kern_frac = 0;
640 	bzero(&io_ctx->poll_start_t, sizeof(io_ctx->poll_start_t));
641 	bzero(&io_ctx->prev_t, sizeof(io_ctx->prev_t));
642 	crit_exit();
643 }
644 
645 static void
646 iopoll_init(int cpuid)
647 {
648 	KKASSERT(cpuid < ncpus2);
649 
650 	rxpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_RX);
651 	txpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_TX);
652 }
653 
654 static struct iopoll_ctx *
655 iopoll_ctx_create(int cpuid, int poll_type)
656 {
657 	struct poll_comm *comm;
658 	struct iopoll_ctx *io_ctx;
659 	const char *poll_type_str;
660 	netisr_fn_t handler, more_handler;
661 
662 	KKASSERT(poll_type == IFPOLL_RX || poll_type == IFPOLL_TX);
663 
664 	/*
665 	 * Make sure that tunables are in sane state
666 	 */
667 	if (iopoll_burst_max < MIN_IOPOLL_BURST_MAX)
668 		iopoll_burst_max = MIN_IOPOLL_BURST_MAX;
669 	else if (iopoll_burst_max > MAX_IOPOLL_BURST_MAX)
670 		iopoll_burst_max = MAX_IOPOLL_BURST_MAX;
671 
672 	if (iopoll_each_burst > iopoll_burst_max)
673 		iopoll_each_burst = iopoll_burst_max;
674 
675 	comm = poll_common[cpuid];
676 
677 	/*
678 	 * Create the per-cpu polling context
679 	 */
680 	io_ctx = kmalloc_cachealign(sizeof(*io_ctx), M_DEVBUF,
681 	    M_WAITOK | M_ZERO);
682 
683 	io_ctx->poll_each_burst = iopoll_each_burst;
684 	io_ctx->poll_burst_max = iopoll_burst_max;
685 	io_ctx->user_frac = 50;
686 	if (poll_type == IFPOLL_RX)
687 		io_ctx->pollhz = comm->pollhz;
688 	else
689 		io_ctx->pollhz = comm->pollhz / (comm->poll_txfrac + 1);
690 	io_ctx->poll_cpuid = cpuid;
691 	iopoll_reset_state(io_ctx);
692 
693 	if (poll_type == IFPOLL_RX) {
694 		handler = rxpoll_handler;
695 		more_handler = rxpollmore_handler;
696 	} else {
697 		handler = txpoll_handler;
698 		more_handler = txpollmore_handler;
699 	}
700 
701 	netmsg_init(&io_ctx->poll_netmsg, NULL, &netisr_adone_rport,
702 	    0, handler);
703 	io_ctx->poll_netmsg.lmsg.u.ms_resultp = io_ctx;
704 
705 	netmsg_init(&io_ctx->poll_more_netmsg, NULL, &netisr_adone_rport,
706 	    0, more_handler);
707 	io_ctx->poll_more_netmsg.lmsg.u.ms_resultp = io_ctx;
708 
709 	/*
710 	 * Initialize per-cpu sysctl nodes
711 	 */
712 	if (poll_type == IFPOLL_RX)
713 		poll_type_str = "rx";
714 	else
715 		poll_type_str = "tx";
716 
717 	sysctl_ctx_init(&io_ctx->poll_sysctl_ctx);
718 	io_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&io_ctx->poll_sysctl_ctx,
719 				   SYSCTL_CHILDREN(comm->sysctl_tree),
720 				   OID_AUTO, poll_type_str, CTLFLAG_RD, 0, "");
721 	iopoll_add_sysctl(&io_ctx->poll_sysctl_ctx,
722 	    SYSCTL_CHILDREN(io_ctx->poll_sysctl_tree), io_ctx, poll_type);
723 
724 	return io_ctx;
725 }
726 
727 /*
728  * Hook from iopoll systimer.  Tries to schedule an iopoll, but keeps
729  * track of lost ticks due to the previous handler taking too long.
730  * Normally, this should not happen, because polling handler should
731  * run for a short time.  However, in some cases (e.g. when there are
732  * changes in link status etc.) the drivers take a very long time
733  * (even in the order of milliseconds) to reset and reconfigure the
734  * device, causing apparent lost polls.
735  *
736  * The first part of the code is just for debugging purposes, and tries
737  * to count how often hardclock ticks are shorter than they should,
738  * meaning either stray interrupts or delayed events.
739  *
740  * WARNING! called from fastint or IPI, the MP lock might not be held.
741  * NOTE: Caller should hold critical section.
742  */
743 static void
744 iopoll_clock(struct iopoll_ctx *io_ctx)
745 {
746 	union ifpoll_time t;
747 	int delta;
748 
749 	KKASSERT(mycpuid == io_ctx->poll_cpuid);
750 
751 	if (io_ctx->poll_handlers == 0)
752 		return;
753 
754 	ifpoll_time_get(&t);
755 	delta = ifpoll_time_diff(&io_ctx->prev_t, &t);
756 	if (delta * io_ctx->pollhz < 500000)
757 		io_ctx->short_ticks++;
758 	else
759 		io_ctx->prev_t = t;
760 
761 	if (io_ctx->pending_polls > 100) {
762 		/*
763 		 * Too much, assume it has stalled (not always true
764 		 * see comment above).
765 		 */
766 		io_ctx->stalled++;
767 		io_ctx->pending_polls = 0;
768 		io_ctx->phase = 0;
769 	}
770 
771 	if (io_ctx->phase <= 2) {
772 		if (io_ctx->phase != 0)
773 			io_ctx->suspect++;
774 		io_ctx->phase = 1;
775 		sched_iopoll(io_ctx);
776 		io_ctx->phase = 2;
777 	}
778 	if (io_ctx->pending_polls++ > 0)
779 		io_ctx->lost_polls++;
780 }
781 
782 /*
783  * rxpoll_handler and txpoll_handler are scheduled by sched_iopoll when
784  * appropriate, typically once per polling systimer tick.
785  *
786  * Note that the message is replied immediately in order to allow a new
787  * ISR to be scheduled in the handler.
788  */
789 static void
790 rxpoll_handler(netmsg_t msg)
791 {
792 	struct iopoll_ctx *io_ctx;
793 	struct thread *td = curthread;
794 	int i, cycles;
795 
796 	io_ctx = msg->lmsg.u.ms_resultp;
797 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
798 
799 	crit_enter_quick(td);
800 
801 	/* Reply ASAP */
802 	lwkt_replymsg(&msg->lmsg, 0);
803 
804 	if (io_ctx->poll_handlers == 0) {
805 		crit_exit_quick(td);
806 		return;
807 	}
808 
809 	io_ctx->phase = 3;
810 	if (io_ctx->residual_burst == 0) {
811 		/* First call in this tick */
812 		ifpoll_time_get(&io_ctx->poll_start_t);
813 		io_ctx->residual_burst = io_ctx->poll_burst;
814 	}
815 	cycles = (io_ctx->residual_burst < io_ctx->poll_each_burst) ?
816 		 io_ctx->residual_burst : io_ctx->poll_each_burst;
817 	io_ctx->residual_burst -= cycles;
818 
819 	for (i = 0; i < io_ctx->poll_handlers; i++) {
820 		const struct iopoll_rec *rec = &io_ctx->pr[i];
821 		struct ifnet *ifp = rec->ifp;
822 
823 		if (!lwkt_serialize_try(rec->serializer))
824 			continue;
825 
826 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
827 		    (IFF_RUNNING | IFF_NPOLLING))
828 			rec->poll_func(ifp, rec->arg, cycles);
829 
830 		lwkt_serialize_exit(rec->serializer);
831 	}
832 
833 	/*
834 	 * Do a quick exit/enter to catch any higher-priority
835 	 * interrupt sources.
836 	 */
837 	crit_exit_quick(td);
838 	crit_enter_quick(td);
839 
840 	sched_iopollmore(io_ctx);
841 	io_ctx->phase = 4;
842 
843 	crit_exit_quick(td);
844 }
845 
846 static void
847 txpoll_handler(netmsg_t msg)
848 {
849 	struct iopoll_ctx *io_ctx;
850 	struct thread *td = curthread;
851 	int i;
852 
853 	io_ctx = msg->lmsg.u.ms_resultp;
854 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
855 
856 	crit_enter_quick(td);
857 
858 	/* Reply ASAP */
859 	lwkt_replymsg(&msg->lmsg, 0);
860 
861 	if (io_ctx->poll_handlers == 0) {
862 		crit_exit_quick(td);
863 		return;
864 	}
865 
866 	io_ctx->phase = 3;
867 
868 	for (i = 0; i < io_ctx->poll_handlers; i++) {
869 		const struct iopoll_rec *rec = &io_ctx->pr[i];
870 		struct ifnet *ifp = rec->ifp;
871 
872 		if (!lwkt_serialize_try(rec->serializer))
873 			continue;
874 
875 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
876 		    (IFF_RUNNING | IFF_NPOLLING))
877 			rec->poll_func(ifp, rec->arg, -1);
878 
879 		lwkt_serialize_exit(rec->serializer);
880 	}
881 
882 	/*
883 	 * Do a quick exit/enter to catch any higher-priority
884 	 * interrupt sources.
885 	 */
886 	crit_exit_quick(td);
887 	crit_enter_quick(td);
888 
889 	sched_iopollmore(io_ctx);
890 	io_ctx->phase = 4;
891 
892 	crit_exit_quick(td);
893 }
894 
895 /*
896  * rxpollmore_handler and txpollmore_handler are called after other netisr's,
897  * possibly scheduling another rxpoll_handler or txpoll_handler call, or
898  * adapting the burst size for the next cycle.
899  *
900  * It is very bad to fetch large bursts of packets from a single card at once,
901  * because the burst could take a long time to be completely processed leading
902  * to unfairness.  To reduce the problem, and also to account better for time
903  * spent in network-related processing, we split the burst in smaller chunks
904  * of fixed size, giving control to the other netisr's between chunks.  This
905  * helps in improving the fairness, reducing livelock and accounting for the
906  * work performed in low level handling.
907  */
908 static void
909 rxpollmore_handler(netmsg_t msg)
910 {
911 	struct thread *td = curthread;
912 	struct iopoll_ctx *io_ctx;
913 	union ifpoll_time t;
914 	int kern_load;
915 	uint32_t pending_polls;
916 
917 	io_ctx = msg->lmsg.u.ms_resultp;
918 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
919 
920 	crit_enter_quick(td);
921 
922 	/* Replay ASAP */
923 	lwkt_replymsg(&msg->lmsg, 0);
924 
925 	if (io_ctx->poll_handlers == 0) {
926 		crit_exit_quick(td);
927 		return;
928 	}
929 
930 	io_ctx->phase = 5;
931 	if (io_ctx->residual_burst > 0) {
932 		sched_iopoll(io_ctx);
933 		crit_exit_quick(td);
934 		/* Will run immediately on return, followed by netisrs */
935 		return;
936 	}
937 
938 	/* Here we can account time spent in iopoll's in this tick */
939 	ifpoll_time_get(&t);
940 	kern_load = ifpoll_time_diff(&io_ctx->poll_start_t, &t);
941 	kern_load = (kern_load * io_ctx->pollhz) / 10000; /* 0..100 */
942 	io_ctx->kern_frac = kern_load;
943 
944 	if (kern_load > (100 - io_ctx->user_frac)) {
945 		/* Try decrease ticks */
946 		if (io_ctx->poll_burst > 1)
947 			io_ctx->poll_burst--;
948 	} else {
949 		if (io_ctx->poll_burst < io_ctx->poll_burst_max)
950 			io_ctx->poll_burst++;
951 	}
952 
953 	io_ctx->pending_polls--;
954 	pending_polls = io_ctx->pending_polls;
955 
956 	if (pending_polls == 0) {
957 		/* We are done */
958 		io_ctx->phase = 0;
959 	} else {
960 		/*
961 		 * Last cycle was long and caused us to miss one or more
962 		 * hardclock ticks.  Restart processing again, but slightly
963 		 * reduce the burst size to prevent that this happens again.
964 		 */
965 		io_ctx->poll_burst -= (io_ctx->poll_burst / 8);
966 		if (io_ctx->poll_burst < 1)
967 			io_ctx->poll_burst = 1;
968 		sched_iopoll(io_ctx);
969 		io_ctx->phase = 6;
970 	}
971 
972 	crit_exit_quick(td);
973 }
974 
975 static void
976 txpollmore_handler(netmsg_t msg)
977 {
978 	struct thread *td = curthread;
979 	struct iopoll_ctx *io_ctx;
980 	uint32_t pending_polls;
981 
982 	io_ctx = msg->lmsg.u.ms_resultp;
983 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
984 
985 	crit_enter_quick(td);
986 
987 	/* Replay ASAP */
988 	lwkt_replymsg(&msg->lmsg, 0);
989 
990 	if (io_ctx->poll_handlers == 0) {
991 		crit_exit_quick(td);
992 		return;
993 	}
994 
995 	io_ctx->phase = 5;
996 
997 	io_ctx->pending_polls--;
998 	pending_polls = io_ctx->pending_polls;
999 
1000 	if (pending_polls == 0) {
1001 		/* We are done */
1002 		io_ctx->phase = 0;
1003 	} else {
1004 		/*
1005 		 * Last cycle was long and caused us to miss one or more
1006 		 * hardclock ticks.  Restart processing again.
1007 		 */
1008 		sched_iopoll(io_ctx);
1009 		io_ctx->phase = 6;
1010 	}
1011 
1012 	crit_exit_quick(td);
1013 }
1014 
1015 static void
1016 iopoll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
1017     struct iopoll_ctx *io_ctx, int poll_type)
1018 {
1019 	if (poll_type == IFPOLL_RX) {
1020 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
1021 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_burstmax,
1022 		    "IU", "Max Polling burst size");
1023 
1024 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
1025 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_eachburst,
1026 		    "IU", "Max size of each burst");
1027 
1028 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
1029 		    &io_ctx->poll_burst, 0, "Current polling burst size");
1030 
1031 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
1032 		    &io_ctx->user_frac, 0, "Desired user fraction of cpu time");
1033 
1034 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "kern_frac", CTLFLAG_RD,
1035 		    &io_ctx->kern_frac, 0, "Kernel fraction of cpu time");
1036 
1037 		SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
1038 		    &io_ctx->residual_burst, 0,
1039 		    "# of residual cycles in burst");
1040 	}
1041 
1042 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
1043 	    &io_ctx->phase, 0, "Polling phase");
1044 
1045 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
1046 	    &io_ctx->suspect, "Suspected events");
1047 
1048 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
1049 	    &io_ctx->stalled, "Potential stalls");
1050 
1051 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
1052 	    &io_ctx->short_ticks,
1053 	    "Hardclock ticks shorter than they should be");
1054 
1055 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
1056 	    &io_ctx->lost_polls,
1057 	    "How many times we would have lost a poll tick");
1058 
1059 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
1060 	    &io_ctx->pending_polls, 0, "Do we need to poll again");
1061 
1062 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
1063 	    &io_ctx->poll_handlers, 0, "Number of registered poll handlers");
1064 }
1065 
1066 static void
1067 sysctl_burstmax_handler(netmsg_t nmsg)
1068 {
1069 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1070 	struct iopoll_ctx *io_ctx;
1071 
1072 	io_ctx = msg->ctx;
1073 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1074 
1075 	io_ctx->poll_burst_max = nmsg->lmsg.u.ms_result;
1076 	if (io_ctx->poll_each_burst > io_ctx->poll_burst_max)
1077 		io_ctx->poll_each_burst = io_ctx->poll_burst_max;
1078 	if (io_ctx->poll_burst > io_ctx->poll_burst_max)
1079 		io_ctx->poll_burst = io_ctx->poll_burst_max;
1080 	if (io_ctx->residual_burst > io_ctx->poll_burst_max)
1081 		io_ctx->residual_burst = io_ctx->poll_burst_max;
1082 
1083 	lwkt_replymsg(&nmsg->lmsg, 0);
1084 }
1085 
1086 static int
1087 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
1088 {
1089 	struct iopoll_ctx *io_ctx = arg1;
1090 	struct iopoll_sysctl_netmsg msg;
1091 	uint32_t burst_max;
1092 	int error;
1093 
1094 	burst_max = io_ctx->poll_burst_max;
1095 	error = sysctl_handle_int(oidp, &burst_max, 0, req);
1096 	if (error || req->newptr == NULL)
1097 		return error;
1098 	if (burst_max < MIN_IOPOLL_BURST_MAX)
1099 		burst_max = MIN_IOPOLL_BURST_MAX;
1100 	else if (burst_max > MAX_IOPOLL_BURST_MAX)
1101 		burst_max = MAX_IOPOLL_BURST_MAX;
1102 
1103 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1104 		    0, sysctl_burstmax_handler);
1105 	msg.base.lmsg.u.ms_result = burst_max;
1106 	msg.ctx = io_ctx;
1107 
1108 	return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1109 }
1110 
1111 static void
1112 sysctl_eachburst_handler(netmsg_t nmsg)
1113 {
1114 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1115 	struct iopoll_ctx *io_ctx;
1116 	uint32_t each_burst;
1117 
1118 	io_ctx = msg->ctx;
1119 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1120 
1121 	each_burst = nmsg->lmsg.u.ms_result;
1122 	if (each_burst > io_ctx->poll_burst_max)
1123 		each_burst = io_ctx->poll_burst_max;
1124 	else if (each_burst < 1)
1125 		each_burst = 1;
1126 	io_ctx->poll_each_burst = each_burst;
1127 
1128 	lwkt_replymsg(&nmsg->lmsg, 0);
1129 }
1130 
1131 static int
1132 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
1133 {
1134 	struct iopoll_ctx *io_ctx = arg1;
1135 	struct iopoll_sysctl_netmsg msg;
1136 	uint32_t each_burst;
1137 	int error;
1138 
1139 	each_burst = io_ctx->poll_each_burst;
1140 	error = sysctl_handle_int(oidp, &each_burst, 0, req);
1141 	if (error || req->newptr == NULL)
1142 		return error;
1143 
1144 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1145 		    0, sysctl_eachburst_handler);
1146 	msg.base.lmsg.u.ms_result = each_burst;
1147 	msg.ctx = io_ctx;
1148 
1149 	return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1150 }
1151 
1152 static int
1153 iopoll_register(struct ifnet *ifp, struct iopoll_ctx *io_ctx,
1154 		const struct ifpoll_io *io_rec)
1155 {
1156 	int error;
1157 
1158 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1159 
1160 	if (io_rec->poll_func == NULL)
1161 		return 0;
1162 
1163 	/*
1164 	 * Check if there is room.
1165 	 */
1166 	if (io_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
1167 		/*
1168 		 * List full, cannot register more entries.
1169 		 * This should never happen; if it does, it is probably a
1170 		 * broken driver trying to register multiple times. Checking
1171 		 * this at runtime is expensive, and won't solve the problem
1172 		 * anyways, so just report a few times and then give up.
1173 		 */
1174 		static int verbose = 10; /* XXX */
1175 		if (verbose > 0) {
1176 			kprintf("io poll handlers list full, "
1177 				"maybe a broken driver ?\n");
1178 			verbose--;
1179 		}
1180 		error = ENOENT;
1181 	} else {
1182 		struct iopoll_rec *rec = &io_ctx->pr[io_ctx->poll_handlers];
1183 
1184 		rec->ifp = ifp;
1185 		rec->serializer = io_rec->serializer;
1186 		rec->arg = io_rec->arg;
1187 		rec->poll_func = io_rec->poll_func;
1188 
1189 		io_ctx->poll_handlers++;
1190 		error = 0;
1191 	}
1192 	return error;
1193 }
1194 
1195 static int
1196 iopoll_deregister(struct ifnet *ifp, struct iopoll_ctx *io_ctx)
1197 {
1198 	int i, error;
1199 
1200 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1201 
1202 	for (i = 0; i < io_ctx->poll_handlers; ++i) {
1203 		if (io_ctx->pr[i].ifp == ifp) /* Found it */
1204 			break;
1205 	}
1206 	if (i == io_ctx->poll_handlers) {
1207 		error = ENOENT;
1208 	} else {
1209 		io_ctx->poll_handlers--;
1210 		if (i < io_ctx->poll_handlers) {
1211 			/* Last entry replaces this one. */
1212 			io_ctx->pr[i] = io_ctx->pr[io_ctx->poll_handlers];
1213 		}
1214 
1215 		if (io_ctx->poll_handlers == 0)
1216 			iopoll_reset_state(io_ctx);
1217 		error = 0;
1218 	}
1219 	return error;
1220 }
1221 
1222 static void
1223 poll_comm_init(int cpuid)
1224 {
1225 	struct poll_comm *comm;
1226 	char cpuid_str[16];
1227 
1228 	comm = kmalloc_cachealign(sizeof(*comm), M_DEVBUF, M_WAITOK | M_ZERO);
1229 
1230 	if (ifpoll_stfrac < 0)
1231 		ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
1232 	if (ifpoll_txfrac < 0)
1233 		ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
1234 
1235 	comm->pollhz = ifpoll_pollhz;
1236 	comm->poll_cpuid = cpuid;
1237 	comm->poll_stfrac = ifpoll_stfrac;
1238 	comm->poll_txfrac = ifpoll_txfrac;
1239 
1240 	ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", cpuid);
1241 
1242 	sysctl_ctx_init(&comm->sysctl_ctx);
1243 	comm->sysctl_tree = SYSCTL_ADD_NODE(&comm->sysctl_ctx,
1244 			    SYSCTL_STATIC_CHILDREN(_net_ifpoll),
1245 			    OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
1246 
1247 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1248 			OID_AUTO, "pollhz", CTLTYPE_INT | CTLFLAG_RW,
1249 			comm, 0, sysctl_pollhz,
1250 			"I", "Device polling frequency");
1251 
1252 	if (cpuid == 0) {
1253 		SYSCTL_ADD_PROC(&comm->sysctl_ctx,
1254 				SYSCTL_CHILDREN(comm->sysctl_tree),
1255 				OID_AUTO, "status_frac",
1256 				CTLTYPE_INT | CTLFLAG_RW,
1257 				comm, 0, sysctl_stfrac,
1258 				"I", "# of cycles before status is polled");
1259 	}
1260 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1261 			OID_AUTO, "tx_frac", CTLTYPE_INT | CTLFLAG_RW,
1262 			comm, 0, sysctl_txfrac,
1263 			"I", "# of cycles before TX is polled");
1264 
1265 	poll_common[cpuid] = comm;
1266 }
1267 
1268 static void
1269 poll_comm_start(int cpuid)
1270 {
1271 	struct poll_comm *comm = poll_common[cpuid];
1272 	systimer_func_t func;
1273 
1274 	/*
1275 	 * Initialize systimer
1276 	 */
1277 	if (cpuid == 0)
1278 		func = poll_comm_systimer0;
1279 	else
1280 		func = poll_comm_systimer;
1281 	systimer_init_periodic_nq(&comm->pollclock, func, comm, 1);
1282 }
1283 
1284 static void
1285 _poll_comm_systimer(struct poll_comm *comm)
1286 {
1287 	if (comm->txfrac_count-- == 0) {
1288 		comm->txfrac_count = comm->poll_txfrac;
1289 		iopoll_clock(txpoll_context[comm->poll_cpuid]);
1290 	}
1291 	iopoll_clock(rxpoll_context[comm->poll_cpuid]);
1292 }
1293 
1294 static void
1295 poll_comm_systimer0(systimer_t info, int in_ipi __unused,
1296     struct intrframe *frame __unused)
1297 {
1298 	struct poll_comm *comm = info->data;
1299 	globaldata_t gd = mycpu;
1300 
1301 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid == 0);
1302 
1303 	crit_enter_gd(gd);
1304 
1305 	if (comm->stfrac_count-- == 0) {
1306 		comm->stfrac_count = comm->poll_stfrac;
1307 		stpoll_clock(&stpoll_context);
1308 	}
1309 	_poll_comm_systimer(comm);
1310 
1311 	crit_exit_gd(gd);
1312 }
1313 
1314 static void
1315 poll_comm_systimer(systimer_t info, int in_ipi __unused,
1316     struct intrframe *frame __unused)
1317 {
1318 	struct poll_comm *comm = info->data;
1319 	globaldata_t gd = mycpu;
1320 
1321 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid != 0);
1322 
1323 	crit_enter_gd(gd);
1324 	_poll_comm_systimer(comm);
1325 	crit_exit_gd(gd);
1326 }
1327 
1328 static void
1329 poll_comm_adjust_pollhz(struct poll_comm *comm)
1330 {
1331 	uint32_t handlers;
1332 	int pollhz = 1;
1333 
1334 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1335 
1336 	/*
1337 	 * If there is no polling handler registered, set systimer
1338 	 * frequency to the lowest value.  Polling systimer frequency
1339 	 * will be adjusted to the requested value, once there are
1340 	 * registered handlers.
1341 	 */
1342 	handlers = rxpoll_context[mycpuid]->poll_handlers +
1343 		   txpoll_context[mycpuid]->poll_handlers;
1344 	if (comm->poll_cpuid == 0)
1345 		handlers += stpoll_context.poll_handlers;
1346 	if (handlers)
1347 		pollhz = comm->pollhz;
1348 	systimer_adjust_periodic(&comm->pollclock, pollhz);
1349 }
1350 
1351 static int
1352 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
1353 {
1354 	struct poll_comm *comm = arg1;
1355 	struct netmsg_base nmsg;
1356 	int error, phz;
1357 
1358 	phz = comm->pollhz;
1359 	error = sysctl_handle_int(oidp, &phz, 0, req);
1360 	if (error || req->newptr == NULL)
1361 		return error;
1362 	if (phz <= 0)
1363 		return EINVAL;
1364 	else if (phz > IFPOLL_FREQ_MAX)
1365 		phz = IFPOLL_FREQ_MAX;
1366 
1367 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1368 		    0, sysctl_pollhz_handler);
1369 	nmsg.lmsg.u.ms_result = phz;
1370 
1371 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1372 }
1373 
1374 static void
1375 sysctl_pollhz_handler(netmsg_t nmsg)
1376 {
1377 	struct poll_comm *comm = poll_common[mycpuid];
1378 
1379 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1380 
1381 	/* Save polling frequency */
1382 	comm->pollhz = nmsg->lmsg.u.ms_result;
1383 
1384 	/*
1385 	 * Adjust cached pollhz
1386 	 */
1387 	rxpoll_context[mycpuid]->pollhz = comm->pollhz;
1388 	txpoll_context[mycpuid]->pollhz =
1389 	    comm->pollhz / (comm->poll_txfrac + 1);
1390 	if (mycpuid == 0)
1391 		stpoll_context.pollhz = comm->pollhz / (comm->poll_stfrac + 1);
1392 
1393 	/*
1394 	 * Adjust polling frequency
1395 	 */
1396 	poll_comm_adjust_pollhz(comm);
1397 
1398 	lwkt_replymsg(&nmsg->lmsg, 0);
1399 }
1400 
1401 static int
1402 sysctl_stfrac(SYSCTL_HANDLER_ARGS)
1403 {
1404 	struct poll_comm *comm = arg1;
1405 	struct netmsg_base nmsg;
1406 	int error, stfrac;
1407 
1408 	KKASSERT(comm->poll_cpuid == 0);
1409 
1410 	stfrac = comm->poll_stfrac;
1411 	error = sysctl_handle_int(oidp, &stfrac, 0, req);
1412 	if (error || req->newptr == NULL)
1413 		return error;
1414 	if (stfrac < 0)
1415 		return EINVAL;
1416 
1417 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1418 		    0, sysctl_stfrac_handler);
1419 	nmsg.lmsg.u.ms_result = stfrac;
1420 
1421 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1422 }
1423 
1424 static void
1425 sysctl_stfrac_handler(netmsg_t nmsg)
1426 {
1427 	struct poll_comm *comm = poll_common[mycpuid];
1428 	int stfrac = nmsg->lmsg.u.ms_result;
1429 
1430 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1431 
1432 	crit_enter();
1433 	comm->poll_stfrac = stfrac;
1434 	if (comm->stfrac_count > comm->poll_stfrac)
1435 		comm->stfrac_count = comm->poll_stfrac;
1436 	crit_exit();
1437 
1438 	lwkt_replymsg(&nmsg->lmsg, 0);
1439 }
1440 
1441 static int
1442 sysctl_txfrac(SYSCTL_HANDLER_ARGS)
1443 {
1444 	struct poll_comm *comm = arg1;
1445 	struct netmsg_base nmsg;
1446 	int error, txfrac;
1447 
1448 	txfrac = comm->poll_txfrac;
1449 	error = sysctl_handle_int(oidp, &txfrac, 0, req);
1450 	if (error || req->newptr == NULL)
1451 		return error;
1452 	if (txfrac < 0)
1453 		return EINVAL;
1454 
1455 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1456 		    0, sysctl_txfrac_handler);
1457 	nmsg.lmsg.u.ms_result = txfrac;
1458 
1459 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1460 }
1461 
1462 static void
1463 sysctl_txfrac_handler(netmsg_t nmsg)
1464 {
1465 	struct poll_comm *comm = poll_common[mycpuid];
1466 	int txfrac = nmsg->lmsg.u.ms_result;
1467 
1468 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1469 
1470 	crit_enter();
1471 	comm->poll_txfrac = txfrac;
1472 	if (comm->txfrac_count > comm->poll_txfrac)
1473 		comm->txfrac_count = comm->poll_txfrac;
1474 	crit_exit();
1475 
1476 	lwkt_replymsg(&nmsg->lmsg, 0);
1477 }
1478