xref: /dflybsd-src/sys/net/if_poll.c (revision f00eae149c338528630cd778fd4de222713daa11)
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	5000
110 #define IOPOLL_BURST_MAX	250	/* good for 1000Mbit net and HZ=6000 */
111 
112 #define IOPOLL_EACH_BURST	50
113 #define IOPOLL_USER_FRAC	50
114 
115 #define IFPOLL_FREQ_DEFAULT	6000
116 
117 #define IFPOLL_TXFRAC_DEFAULT	1	/* 1/1 of the pollhz */
118 #define IFPOLL_STFRAC_DEFAULT	120	/* 1/120 of the pollhz */
119 
120 #define IFPOLL_RX		0x1
121 #define IFPOLL_TX		0x2
122 
123 union ifpoll_time {
124 	struct timeval		tv;
125 	uint64_t		tsc;
126 };
127 
128 struct iopoll_rec {
129 	struct lwkt_serialize	*serializer;
130 	struct ifnet		*ifp;
131 	void			*arg;
132 	ifpoll_iofn_t		poll_func;
133 };
134 
135 struct iopoll_ctx {
136 	union ifpoll_time	prev_t;
137 	u_long			short_ticks;		/* statistics */
138 	u_long			lost_polls;		/* statistics */
139 	u_long			suspect;		/* statistics */
140 	u_long			stalled;		/* statistics */
141 	uint32_t		pending_polls;		/* state */
142 
143 	struct netmsg_base	poll_netmsg;
144 	struct netmsg_base	poll_more_netmsg;
145 
146 	int			poll_cpuid;
147 	int			pollhz;
148 	uint32_t		phase;			/* state */
149 	int			residual_burst;		/* state */
150 	uint32_t		poll_each_burst;	/* tunable */
151 	union ifpoll_time	poll_start_t;		/* state */
152 
153 	uint32_t		poll_burst;		/* state */
154 	uint32_t		poll_burst_max;		/* tunable */
155 	uint32_t		user_frac;		/* tunable */
156 	uint32_t		kern_frac;		/* state */
157 
158 	uint32_t		poll_handlers; /* next free entry in pr[]. */
159 	struct iopoll_rec	pr[IFPOLL_LIST_LEN];
160 
161 	struct sysctl_ctx_list	poll_sysctl_ctx;
162 	struct sysctl_oid	*poll_sysctl_tree;
163 } __cachealign;
164 
165 struct poll_comm {
166 	struct systimer		pollclock;
167 	int			poll_cpuid;
168 
169 	int			stfrac_count;		/* state */
170 	int			poll_stfrac;		/* tunable */
171 
172 	int			txfrac_count;		/* state */
173 	int			poll_txfrac;		/* tunable */
174 
175 	int			pollhz;			/* tunable */
176 
177 	struct sysctl_ctx_list	sysctl_ctx;
178 	struct sysctl_oid	*sysctl_tree;
179 } __cachealign;
180 
181 struct stpoll_rec {
182 	struct lwkt_serialize	*serializer;
183 	struct ifnet		*ifp;
184 	ifpoll_stfn_t		status_func;
185 };
186 
187 struct stpoll_ctx {
188 	struct netmsg_base	poll_netmsg;
189 
190 	uint32_t		poll_handlers; /* next free entry in pr[]. */
191 	struct stpoll_rec	pr[IFPOLL_LIST_LEN];
192 
193 	struct sysctl_ctx_list	poll_sysctl_ctx;
194 	struct sysctl_oid	*poll_sysctl_tree;
195 } __cachealign;
196 
197 struct iopoll_sysctl_netmsg {
198 	struct netmsg_base	base;
199 	struct iopoll_ctx	*ctx;
200 };
201 
202 void		ifpoll_init_pcpu(int);
203 static void	ifpoll_register_handler(netmsg_t);
204 static void	ifpoll_deregister_handler(netmsg_t);
205 
206 /*
207  * Status polling
208  */
209 static void	stpoll_init(void);
210 static void	stpoll_handler(netmsg_t);
211 static void	stpoll_clock(struct stpoll_ctx *);
212 static int	stpoll_register(struct ifnet *, const struct ifpoll_status *);
213 static int	stpoll_deregister(struct ifnet *);
214 
215 /*
216  * RX/TX polling
217  */
218 static struct iopoll_ctx *iopoll_ctx_create(int, int);
219 static void	iopoll_init(int);
220 static void	rxpoll_handler(netmsg_t);
221 static void	txpoll_handler(netmsg_t);
222 static void	rxpollmore_handler(netmsg_t);
223 static void	txpollmore_handler(netmsg_t);
224 static void	iopoll_clock(struct iopoll_ctx *);
225 static int	iopoll_register(struct ifnet *, struct iopoll_ctx *,
226 		    const struct ifpoll_io *);
227 static int	iopoll_deregister(struct ifnet *, struct iopoll_ctx *);
228 
229 static void	iopoll_add_sysctl(struct sysctl_ctx_list *,
230 		    struct sysctl_oid_list *, struct iopoll_ctx *, int);
231 static void	sysctl_burstmax_handler(netmsg_t);
232 static int	sysctl_burstmax(SYSCTL_HANDLER_ARGS);
233 static void	sysctl_eachburst_handler(netmsg_t);
234 static int	sysctl_eachburst(SYSCTL_HANDLER_ARGS);
235 
236 /*
237  * Common functions
238  */
239 static void	poll_comm_init(int);
240 static void	poll_comm_start(int);
241 static void	poll_comm_adjust_pollhz(struct poll_comm *);
242 static void	poll_comm_systimer0(systimer_t, int, struct intrframe *);
243 static void	poll_comm_systimer(systimer_t, int, struct intrframe *);
244 static void	sysctl_pollhz_handler(netmsg_t);
245 static void	sysctl_stfrac_handler(netmsg_t);
246 static void	sysctl_txfrac_handler(netmsg_t);
247 static int	sysctl_pollhz(SYSCTL_HANDLER_ARGS);
248 static int	sysctl_stfrac(SYSCTL_HANDLER_ARGS);
249 static int	sysctl_txfrac(SYSCTL_HANDLER_ARGS);
250 static int	sysctl_compat_npoll_stfrac(SYSCTL_HANDLER_ARGS);
251 static int	sysctl_compat_npoll_cpuid(SYSCTL_HANDLER_ARGS);
252 
253 static struct stpoll_ctx	stpoll_context;
254 static struct poll_comm		*poll_common[MAXCPU];
255 static struct iopoll_ctx	*rxpoll_context[MAXCPU];
256 static struct iopoll_ctx	*txpoll_context[MAXCPU];
257 
258 SYSCTL_NODE(_net, OID_AUTO, ifpoll, CTLFLAG_RW, 0,
259 	    "Network device polling parameters");
260 
261 static int	iopoll_burst_max = IOPOLL_BURST_MAX;
262 static int	iopoll_each_burst = IOPOLL_EACH_BURST;
263 static int	iopoll_user_frac = IOPOLL_USER_FRAC;
264 
265 static int	ifpoll_pollhz = IFPOLL_FREQ_DEFAULT;
266 static int	ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
267 static int	ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
268 
269 TUNABLE_INT("net.ifpoll.burst_max", &iopoll_burst_max);
270 TUNABLE_INT("net.ifpoll.each_burst", &iopoll_each_burst);
271 TUNABLE_INT("net.ifpoll.user_frac", &iopoll_user_frac);
272 TUNABLE_INT("net.ifpoll.pollhz", &ifpoll_pollhz);
273 TUNABLE_INT("net.ifpoll.status_frac", &ifpoll_stfrac);
274 TUNABLE_INT("net.ifpoll.tx_frac", &ifpoll_txfrac);
275 
276 #define IFPOLL_FREQ_ADJ(comm)	(((comm)->poll_cpuid * 3) % 50)
277 
278 static __inline int
279 poll_comm_pollhz_div(const struct poll_comm *comm, int pollhz)
280 {
281 	return pollhz + IFPOLL_FREQ_ADJ(comm);
282 }
283 
284 static __inline int
285 poll_comm_pollhz_conv(const struct poll_comm *comm, int pollhz)
286 {
287 	return pollhz - IFPOLL_FREQ_ADJ(comm);
288 }
289 
290 static __inline void
291 ifpoll_sendmsg_oncpu(netmsg_t msg)
292 {
293 	if (msg->lmsg.ms_flags & MSGF_DONE)
294 		lwkt_sendmsg(netisr_portfn(mycpuid), &msg->lmsg);
295 }
296 
297 static __inline void
298 sched_stpoll(struct stpoll_ctx *st_ctx)
299 {
300 	ifpoll_sendmsg_oncpu((netmsg_t)&st_ctx->poll_netmsg);
301 }
302 
303 static __inline void
304 sched_iopoll(struct iopoll_ctx *io_ctx)
305 {
306 	ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_netmsg);
307 }
308 
309 static __inline void
310 sched_iopollmore(struct iopoll_ctx *io_ctx)
311 {
312 	ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_more_netmsg);
313 }
314 
315 static __inline void
316 ifpoll_time_get(union ifpoll_time *t)
317 {
318 	if (__predict_true(tsc_present))
319 		t->tsc = rdtsc();
320 	else
321 		microuptime(&t->tv);
322 }
323 
324 /* Return time diff in us */
325 static __inline int
326 ifpoll_time_diff(const union ifpoll_time *s, const union ifpoll_time *e)
327 {
328 	if (__predict_true(tsc_present)) {
329 		return (((e->tsc - s->tsc) * 1000000) / tsc_frequency);
330 	} else {
331 		return ((e->tv.tv_usec - s->tv.tv_usec) +
332 			(e->tv.tv_sec - s->tv.tv_sec) * 1000000);
333 	}
334 }
335 
336 /*
337  * Initialize per-cpu qpolling(4) context.  Called from kern_clock.c:
338  */
339 void
340 ifpoll_init_pcpu(int cpuid)
341 {
342 	if (cpuid >= ncpus2)
343 		return;
344 
345 	poll_comm_init(cpuid);
346 
347 	if (cpuid == 0)
348 		stpoll_init();
349 	iopoll_init(cpuid);
350 
351 	poll_comm_start(cpuid);
352 }
353 
354 int
355 ifpoll_register(struct ifnet *ifp)
356 {
357 	struct ifpoll_info *info;
358 	struct netmsg_base nmsg;
359 	int error;
360 
361 	if (ifp->if_npoll == NULL) {
362 		/* Device does not support polling */
363 		return EOPNOTSUPP;
364 	}
365 
366 	info = kmalloc(sizeof(*info), M_TEMP, M_WAITOK | M_ZERO);
367 
368 	/*
369 	 * Attempt to register.  Interlock with IFF_NPOLLING.
370 	 */
371 
372 	ifnet_serialize_all(ifp);
373 
374 	if (ifp->if_flags & IFF_NPOLLING) {
375 		/* Already polling */
376 		ifnet_deserialize_all(ifp);
377 		kfree(info, M_TEMP);
378 		return EBUSY;
379 	}
380 
381 	info->ifpi_ifp = ifp;
382 
383 	ifp->if_flags |= IFF_NPOLLING;
384 	ifp->if_npoll(ifp, info);
385 
386 	ifnet_deserialize_all(ifp);
387 
388 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
389 		    0, ifpoll_register_handler);
390 	nmsg.lmsg.u.ms_resultp = info;
391 
392 	error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
393 	if (error) {
394 		if (!ifpoll_deregister(ifp)) {
395 			if_printf(ifp, "ifpoll_register: "
396 				  "ifpoll_deregister failed!\n");
397 		}
398 	}
399 
400 	kfree(info, M_TEMP);
401 	return error;
402 }
403 
404 int
405 ifpoll_deregister(struct ifnet *ifp)
406 {
407 	struct netmsg_base nmsg;
408 	int error;
409 
410 	if (ifp->if_npoll == NULL)
411 		return EOPNOTSUPP;
412 
413 	ifnet_serialize_all(ifp);
414 
415 	if ((ifp->if_flags & IFF_NPOLLING) == 0) {
416 		ifnet_deserialize_all(ifp);
417 		return EINVAL;
418 	}
419 	ifp->if_flags &= ~IFF_NPOLLING;
420 
421 	ifnet_deserialize_all(ifp);
422 
423 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
424 		    0, ifpoll_deregister_handler);
425 	nmsg.lmsg.u.ms_resultp = ifp;
426 
427 	error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
428 	if (!error) {
429 		ifnet_serialize_all(ifp);
430 		ifp->if_npoll(ifp, NULL);
431 		ifnet_deserialize_all(ifp);
432 	}
433 	return error;
434 }
435 
436 static void
437 ifpoll_register_handler(netmsg_t nmsg)
438 {
439 	const struct ifpoll_info *info = nmsg->lmsg.u.ms_resultp;
440 	int cpuid = mycpuid, nextcpu;
441 	int error;
442 
443 	KKASSERT(cpuid < ncpus2);
444 	KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
445 
446 	if (cpuid == 0) {
447 		error = stpoll_register(info->ifpi_ifp, &info->ifpi_status);
448 		if (error)
449 			goto failed;
450 	}
451 
452 	error = iopoll_register(info->ifpi_ifp, rxpoll_context[cpuid],
453 				&info->ifpi_rx[cpuid]);
454 	if (error)
455 		goto failed;
456 
457 	error = iopoll_register(info->ifpi_ifp, txpoll_context[cpuid],
458 				&info->ifpi_tx[cpuid]);
459 	if (error)
460 		goto failed;
461 
462 	/* Adjust polling frequency, after all registration is done */
463 	poll_comm_adjust_pollhz(poll_common[cpuid]);
464 
465 	nextcpu = cpuid + 1;
466 	if (nextcpu < ncpus2)
467 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
468 	else
469 		lwkt_replymsg(&nmsg->lmsg, 0);
470 	return;
471 failed:
472 	lwkt_replymsg(&nmsg->lmsg, error);
473 }
474 
475 static void
476 ifpoll_deregister_handler(netmsg_t nmsg)
477 {
478 	struct ifnet *ifp = nmsg->lmsg.u.ms_resultp;
479 	int cpuid = mycpuid, nextcpu;
480 
481 	KKASSERT(cpuid < ncpus2);
482 	KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
483 
484 	/* Ignore errors */
485 	if (cpuid == 0)
486 		stpoll_deregister(ifp);
487 	iopoll_deregister(ifp, rxpoll_context[cpuid]);
488 	iopoll_deregister(ifp, txpoll_context[cpuid]);
489 
490 	/* Adjust polling frequency, after all deregistration is done */
491 	poll_comm_adjust_pollhz(poll_common[cpuid]);
492 
493 	nextcpu = cpuid + 1;
494 	if (nextcpu < ncpus2)
495 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
496 	else
497 		lwkt_replymsg(&nmsg->lmsg, 0);
498 }
499 
500 static void
501 stpoll_init(void)
502 {
503 	struct stpoll_ctx *st_ctx = &stpoll_context;
504 	const struct poll_comm *comm = poll_common[0];
505 
506 	sysctl_ctx_init(&st_ctx->poll_sysctl_ctx);
507 	st_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&st_ctx->poll_sysctl_ctx,
508 				   SYSCTL_CHILDREN(comm->sysctl_tree),
509 				   OID_AUTO, "status", CTLFLAG_RD, 0, "");
510 
511 	SYSCTL_ADD_UINT(&st_ctx->poll_sysctl_ctx,
512 			SYSCTL_CHILDREN(st_ctx->poll_sysctl_tree),
513 			OID_AUTO, "handlers", CTLFLAG_RD,
514 			&st_ctx->poll_handlers, 0,
515 			"Number of registered status poll handlers");
516 
517 	netmsg_init(&st_ctx->poll_netmsg, NULL, &netisr_adone_rport,
518 		    0, stpoll_handler);
519 }
520 
521 /*
522  * stpoll_handler is scheduled by sched_stpoll when appropriate, typically
523  * once per polling systimer tick.
524  */
525 static void
526 stpoll_handler(netmsg_t msg)
527 {
528 	struct stpoll_ctx *st_ctx = &stpoll_context;
529 	struct thread *td = curthread;
530 	int i;
531 
532 	KKASSERT(&td->td_msgport == netisr_portfn(0));
533 
534 	crit_enter_quick(td);
535 
536 	/* Reply ASAP */
537 	lwkt_replymsg(&msg->lmsg, 0);
538 
539 	if (st_ctx->poll_handlers == 0) {
540 		crit_exit_quick(td);
541 		return;
542 	}
543 
544 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
545 		const struct stpoll_rec *rec = &st_ctx->pr[i];
546 		struct ifnet *ifp = rec->ifp;
547 
548 		if (!lwkt_serialize_try(rec->serializer))
549 			continue;
550 
551 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
552 		    (IFF_RUNNING | IFF_NPOLLING))
553 			rec->status_func(ifp);
554 
555 		lwkt_serialize_exit(rec->serializer);
556 	}
557 
558 	crit_exit_quick(td);
559 }
560 
561 /*
562  * Hook from status poll systimer.  Tries to schedule an status poll.
563  * NOTE: Caller should hold critical section.
564  */
565 static void
566 stpoll_clock(struct stpoll_ctx *st_ctx)
567 {
568 	KKASSERT(mycpuid == 0);
569 
570 	if (st_ctx->poll_handlers == 0)
571 		return;
572 	sched_stpoll(st_ctx);
573 }
574 
575 static int
576 stpoll_register(struct ifnet *ifp, const struct ifpoll_status *st_rec)
577 {
578 	struct stpoll_ctx *st_ctx = &stpoll_context;
579 	int error;
580 
581 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
582 
583 	if (st_rec->status_func == NULL)
584 		return 0;
585 
586 	/*
587 	 * Check if there is room.
588 	 */
589 	if (st_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
590 		/*
591 		 * List full, cannot register more entries.
592 		 * This should never happen; if it does, it is probably a
593 		 * broken driver trying to register multiple times. Checking
594 		 * this at runtime is expensive, and won't solve the problem
595 		 * anyways, so just report a few times and then give up.
596 		 */
597 		static int verbose = 10; /* XXX */
598 
599 		if (verbose > 0) {
600 			kprintf("status poll handlers list full, "
601 				"maybe a broken driver ?\n");
602 			verbose--;
603 		}
604 		error = ENOENT;
605 	} else {
606 		struct stpoll_rec *rec = &st_ctx->pr[st_ctx->poll_handlers];
607 
608 		rec->ifp = ifp;
609 		rec->serializer = st_rec->serializer;
610 		rec->status_func = st_rec->status_func;
611 
612 		st_ctx->poll_handlers++;
613 		error = 0;
614 	}
615 	return error;
616 }
617 
618 static int
619 stpoll_deregister(struct ifnet *ifp)
620 {
621 	struct stpoll_ctx *st_ctx = &stpoll_context;
622 	int i, error;
623 
624 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
625 
626 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
627 		if (st_ctx->pr[i].ifp == ifp) /* Found it */
628 			break;
629 	}
630 	if (i == st_ctx->poll_handlers) {
631 		error = ENOENT;
632 	} else {
633 		st_ctx->poll_handlers--;
634 		if (i < st_ctx->poll_handlers) {
635 			/* Last entry replaces this one. */
636 			st_ctx->pr[i] = st_ctx->pr[st_ctx->poll_handlers];
637 		}
638 		error = 0;
639 	}
640 	return error;
641 }
642 
643 static __inline void
644 iopoll_reset_state(struct iopoll_ctx *io_ctx)
645 {
646 	crit_enter();
647 	io_ctx->poll_burst = io_ctx->poll_each_burst;
648 	io_ctx->pending_polls = 0;
649 	io_ctx->residual_burst = 0;
650 	io_ctx->phase = 0;
651 	io_ctx->kern_frac = 0;
652 	bzero(&io_ctx->poll_start_t, sizeof(io_ctx->poll_start_t));
653 	bzero(&io_ctx->prev_t, sizeof(io_ctx->prev_t));
654 	crit_exit();
655 }
656 
657 static void
658 iopoll_init(int cpuid)
659 {
660 	KKASSERT(cpuid < ncpus2);
661 
662 	rxpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_RX);
663 	txpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_TX);
664 }
665 
666 static struct iopoll_ctx *
667 iopoll_ctx_create(int cpuid, int poll_type)
668 {
669 	struct poll_comm *comm;
670 	struct iopoll_ctx *io_ctx;
671 	const char *poll_type_str;
672 	netisr_fn_t handler, more_handler;
673 
674 	KKASSERT(poll_type == IFPOLL_RX || poll_type == IFPOLL_TX);
675 
676 	/*
677 	 * Make sure that tunables are in sane state
678 	 */
679 	if (iopoll_burst_max < MIN_IOPOLL_BURST_MAX)
680 		iopoll_burst_max = MIN_IOPOLL_BURST_MAX;
681 	else if (iopoll_burst_max > MAX_IOPOLL_BURST_MAX)
682 		iopoll_burst_max = MAX_IOPOLL_BURST_MAX;
683 
684 	if (iopoll_each_burst > iopoll_burst_max)
685 		iopoll_each_burst = iopoll_burst_max;
686 
687 	comm = poll_common[cpuid];
688 
689 	/*
690 	 * Create the per-cpu polling context
691 	 */
692 	io_ctx = kmalloc_cachealign(sizeof(*io_ctx), M_DEVBUF,
693 	    M_WAITOK | M_ZERO);
694 
695 	io_ctx->poll_each_burst = iopoll_each_burst;
696 	io_ctx->poll_burst_max = iopoll_burst_max;
697 	io_ctx->user_frac = iopoll_user_frac;
698 	if (poll_type == IFPOLL_RX)
699 		io_ctx->pollhz = comm->pollhz;
700 	else
701 		io_ctx->pollhz = comm->pollhz / (comm->poll_txfrac + 1);
702 	io_ctx->poll_cpuid = cpuid;
703 	iopoll_reset_state(io_ctx);
704 
705 	if (poll_type == IFPOLL_RX) {
706 		handler = rxpoll_handler;
707 		more_handler = rxpollmore_handler;
708 	} else {
709 		handler = txpoll_handler;
710 		more_handler = txpollmore_handler;
711 	}
712 
713 	netmsg_init(&io_ctx->poll_netmsg, NULL, &netisr_adone_rport,
714 	    0, handler);
715 	io_ctx->poll_netmsg.lmsg.u.ms_resultp = io_ctx;
716 
717 	netmsg_init(&io_ctx->poll_more_netmsg, NULL, &netisr_adone_rport,
718 	    0, more_handler);
719 	io_ctx->poll_more_netmsg.lmsg.u.ms_resultp = io_ctx;
720 
721 	/*
722 	 * Initialize per-cpu sysctl nodes
723 	 */
724 	if (poll_type == IFPOLL_RX)
725 		poll_type_str = "rx";
726 	else
727 		poll_type_str = "tx";
728 
729 	sysctl_ctx_init(&io_ctx->poll_sysctl_ctx);
730 	io_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&io_ctx->poll_sysctl_ctx,
731 				   SYSCTL_CHILDREN(comm->sysctl_tree),
732 				   OID_AUTO, poll_type_str, CTLFLAG_RD, 0, "");
733 	iopoll_add_sysctl(&io_ctx->poll_sysctl_ctx,
734 	    SYSCTL_CHILDREN(io_ctx->poll_sysctl_tree), io_ctx, poll_type);
735 
736 	return io_ctx;
737 }
738 
739 /*
740  * Hook from iopoll systimer.  Tries to schedule an iopoll, but keeps
741  * track of lost ticks due to the previous handler taking too long.
742  * Normally, this should not happen, because polling handler should
743  * run for a short time.  However, in some cases (e.g. when there are
744  * changes in link status etc.) the drivers take a very long time
745  * (even in the order of milliseconds) to reset and reconfigure the
746  * device, causing apparent lost polls.
747  *
748  * The first part of the code is just for debugging purposes, and tries
749  * to count how often hardclock ticks are shorter than they should,
750  * meaning either stray interrupts or delayed events.
751  *
752  * WARNING! called from fastint or IPI, the MP lock might not be held.
753  * NOTE: Caller should hold critical section.
754  */
755 static void
756 iopoll_clock(struct iopoll_ctx *io_ctx)
757 {
758 	union ifpoll_time t;
759 	int delta;
760 
761 	KKASSERT(mycpuid == io_ctx->poll_cpuid);
762 
763 	if (io_ctx->poll_handlers == 0)
764 		return;
765 
766 	ifpoll_time_get(&t);
767 	delta = ifpoll_time_diff(&io_ctx->prev_t, &t);
768 	if (delta * io_ctx->pollhz < 500000)
769 		io_ctx->short_ticks++;
770 	else
771 		io_ctx->prev_t = t;
772 
773 	if (io_ctx->pending_polls > 100) {
774 		/*
775 		 * Too much, assume it has stalled (not always true
776 		 * see comment above).
777 		 */
778 		io_ctx->stalled++;
779 		io_ctx->pending_polls = 0;
780 		io_ctx->phase = 0;
781 	}
782 
783 	if (io_ctx->phase <= 2) {
784 		if (io_ctx->phase != 0)
785 			io_ctx->suspect++;
786 		io_ctx->phase = 1;
787 		sched_iopoll(io_ctx);
788 		io_ctx->phase = 2;
789 	}
790 	if (io_ctx->pending_polls++ > 0)
791 		io_ctx->lost_polls++;
792 }
793 
794 /*
795  * rxpoll_handler and txpoll_handler are scheduled by sched_iopoll when
796  * appropriate, typically once per polling systimer tick.
797  *
798  * Note that the message is replied immediately in order to allow a new
799  * ISR to be scheduled in the handler.
800  */
801 static void
802 rxpoll_handler(netmsg_t msg)
803 {
804 	struct iopoll_ctx *io_ctx;
805 	struct thread *td = curthread;
806 	int i, cycles;
807 
808 	io_ctx = msg->lmsg.u.ms_resultp;
809 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
810 
811 	crit_enter_quick(td);
812 
813 	/* Reply ASAP */
814 	lwkt_replymsg(&msg->lmsg, 0);
815 
816 	if (io_ctx->poll_handlers == 0) {
817 		crit_exit_quick(td);
818 		return;
819 	}
820 
821 	io_ctx->phase = 3;
822 	if (io_ctx->residual_burst == 0) {
823 		/* First call in this tick */
824 		ifpoll_time_get(&io_ctx->poll_start_t);
825 		io_ctx->residual_burst = io_ctx->poll_burst;
826 	}
827 	cycles = (io_ctx->residual_burst < io_ctx->poll_each_burst) ?
828 		 io_ctx->residual_burst : io_ctx->poll_each_burst;
829 	io_ctx->residual_burst -= cycles;
830 
831 	for (i = 0; i < io_ctx->poll_handlers; i++) {
832 		const struct iopoll_rec *rec = &io_ctx->pr[i];
833 		struct ifnet *ifp = rec->ifp;
834 
835 		if (!lwkt_serialize_try(rec->serializer))
836 			continue;
837 
838 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
839 		    (IFF_RUNNING | IFF_NPOLLING))
840 			rec->poll_func(ifp, rec->arg, cycles);
841 
842 		lwkt_serialize_exit(rec->serializer);
843 	}
844 
845 	/*
846 	 * Do a quick exit/enter to catch any higher-priority
847 	 * interrupt sources.
848 	 */
849 	crit_exit_quick(td);
850 	crit_enter_quick(td);
851 
852 	sched_iopollmore(io_ctx);
853 	io_ctx->phase = 4;
854 
855 	crit_exit_quick(td);
856 }
857 
858 static void
859 txpoll_handler(netmsg_t msg)
860 {
861 	struct iopoll_ctx *io_ctx;
862 	struct thread *td = curthread;
863 	int i;
864 
865 	io_ctx = msg->lmsg.u.ms_resultp;
866 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
867 
868 	crit_enter_quick(td);
869 
870 	/* Reply ASAP */
871 	lwkt_replymsg(&msg->lmsg, 0);
872 
873 	if (io_ctx->poll_handlers == 0) {
874 		crit_exit_quick(td);
875 		return;
876 	}
877 
878 	io_ctx->phase = 3;
879 
880 	for (i = 0; i < io_ctx->poll_handlers; i++) {
881 		const struct iopoll_rec *rec = &io_ctx->pr[i];
882 		struct ifnet *ifp = rec->ifp;
883 
884 		if (!lwkt_serialize_try(rec->serializer))
885 			continue;
886 
887 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
888 		    (IFF_RUNNING | IFF_NPOLLING))
889 			rec->poll_func(ifp, rec->arg, -1);
890 
891 		lwkt_serialize_exit(rec->serializer);
892 	}
893 
894 	/*
895 	 * Do a quick exit/enter to catch any higher-priority
896 	 * interrupt sources.
897 	 */
898 	crit_exit_quick(td);
899 	crit_enter_quick(td);
900 
901 	sched_iopollmore(io_ctx);
902 	io_ctx->phase = 4;
903 
904 	crit_exit_quick(td);
905 }
906 
907 /*
908  * rxpollmore_handler and txpollmore_handler are called after other netisr's,
909  * possibly scheduling another rxpoll_handler or txpoll_handler call, or
910  * adapting the burst size for the next cycle.
911  *
912  * It is very bad to fetch large bursts of packets from a single card at once,
913  * because the burst could take a long time to be completely processed leading
914  * to unfairness.  To reduce the problem, and also to account better for time
915  * spent in network-related processing, we split the burst in smaller chunks
916  * of fixed size, giving control to the other netisr's between chunks.  This
917  * helps in improving the fairness, reducing livelock and accounting for the
918  * work performed in low level handling.
919  */
920 static void
921 rxpollmore_handler(netmsg_t msg)
922 {
923 	struct thread *td = curthread;
924 	struct iopoll_ctx *io_ctx;
925 	union ifpoll_time t;
926 	int kern_load;
927 	uint32_t pending_polls;
928 
929 	io_ctx = msg->lmsg.u.ms_resultp;
930 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
931 
932 	crit_enter_quick(td);
933 
934 	/* Replay ASAP */
935 	lwkt_replymsg(&msg->lmsg, 0);
936 
937 	if (io_ctx->poll_handlers == 0) {
938 		crit_exit_quick(td);
939 		return;
940 	}
941 
942 	io_ctx->phase = 5;
943 	if (io_ctx->residual_burst > 0) {
944 		sched_iopoll(io_ctx);
945 		crit_exit_quick(td);
946 		/* Will run immediately on return, followed by netisrs */
947 		return;
948 	}
949 
950 	/* Here we can account time spent in iopoll's in this tick */
951 	ifpoll_time_get(&t);
952 	kern_load = ifpoll_time_diff(&io_ctx->poll_start_t, &t);
953 	kern_load = (kern_load * io_ctx->pollhz) / 10000; /* 0..100 */
954 	io_ctx->kern_frac = kern_load;
955 
956 	if (kern_load > (100 - io_ctx->user_frac)) {
957 		/* Try decrease ticks */
958 		if (io_ctx->poll_burst > 1)
959 			io_ctx->poll_burst--;
960 	} else {
961 		if (io_ctx->poll_burst < io_ctx->poll_burst_max)
962 			io_ctx->poll_burst++;
963 	}
964 
965 	io_ctx->pending_polls--;
966 	pending_polls = io_ctx->pending_polls;
967 
968 	if (pending_polls == 0) {
969 		/* We are done */
970 		io_ctx->phase = 0;
971 	} else {
972 		/*
973 		 * Last cycle was long and caused us to miss one or more
974 		 * hardclock ticks.  Restart processing again, but slightly
975 		 * reduce the burst size to prevent that this happens again.
976 		 */
977 		io_ctx->poll_burst -= (io_ctx->poll_burst / 8);
978 		if (io_ctx->poll_burst < 1)
979 			io_ctx->poll_burst = 1;
980 		sched_iopoll(io_ctx);
981 		io_ctx->phase = 6;
982 	}
983 
984 	crit_exit_quick(td);
985 }
986 
987 static void
988 txpollmore_handler(netmsg_t msg)
989 {
990 	struct thread *td = curthread;
991 	struct iopoll_ctx *io_ctx;
992 	uint32_t pending_polls;
993 
994 	io_ctx = msg->lmsg.u.ms_resultp;
995 	KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
996 
997 	crit_enter_quick(td);
998 
999 	/* Replay ASAP */
1000 	lwkt_replymsg(&msg->lmsg, 0);
1001 
1002 	if (io_ctx->poll_handlers == 0) {
1003 		crit_exit_quick(td);
1004 		return;
1005 	}
1006 
1007 	io_ctx->phase = 5;
1008 
1009 	io_ctx->pending_polls--;
1010 	pending_polls = io_ctx->pending_polls;
1011 
1012 	if (pending_polls == 0) {
1013 		/* We are done */
1014 		io_ctx->phase = 0;
1015 	} else {
1016 		/*
1017 		 * Last cycle was long and caused us to miss one or more
1018 		 * hardclock ticks.  Restart processing again.
1019 		 */
1020 		sched_iopoll(io_ctx);
1021 		io_ctx->phase = 6;
1022 	}
1023 
1024 	crit_exit_quick(td);
1025 }
1026 
1027 static void
1028 iopoll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
1029     struct iopoll_ctx *io_ctx, int poll_type)
1030 {
1031 	if (poll_type == IFPOLL_RX) {
1032 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
1033 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_burstmax,
1034 		    "IU", "Max Polling burst size");
1035 
1036 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
1037 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_eachburst,
1038 		    "IU", "Max size of each burst");
1039 
1040 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
1041 		    &io_ctx->poll_burst, 0, "Current polling burst size");
1042 
1043 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
1044 		    &io_ctx->user_frac, 0, "Desired user fraction of cpu time");
1045 
1046 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "kern_frac", CTLFLAG_RD,
1047 		    &io_ctx->kern_frac, 0, "Kernel fraction of cpu time");
1048 
1049 		SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
1050 		    &io_ctx->residual_burst, 0,
1051 		    "# of residual cycles in burst");
1052 	}
1053 
1054 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
1055 	    &io_ctx->phase, 0, "Polling phase");
1056 
1057 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
1058 	    &io_ctx->suspect, "Suspected events");
1059 
1060 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
1061 	    &io_ctx->stalled, "Potential stalls");
1062 
1063 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
1064 	    &io_ctx->short_ticks,
1065 	    "Hardclock ticks shorter than they should be");
1066 
1067 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
1068 	    &io_ctx->lost_polls,
1069 	    "How many times we would have lost a poll tick");
1070 
1071 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
1072 	    &io_ctx->pending_polls, 0, "Do we need to poll again");
1073 
1074 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
1075 	    &io_ctx->poll_handlers, 0, "Number of registered poll handlers");
1076 }
1077 
1078 static void
1079 sysctl_burstmax_handler(netmsg_t nmsg)
1080 {
1081 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1082 	struct iopoll_ctx *io_ctx;
1083 
1084 	io_ctx = msg->ctx;
1085 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1086 
1087 	io_ctx->poll_burst_max = nmsg->lmsg.u.ms_result;
1088 	if (io_ctx->poll_each_burst > io_ctx->poll_burst_max)
1089 		io_ctx->poll_each_burst = io_ctx->poll_burst_max;
1090 	if (io_ctx->poll_burst > io_ctx->poll_burst_max)
1091 		io_ctx->poll_burst = io_ctx->poll_burst_max;
1092 	if (io_ctx->residual_burst > io_ctx->poll_burst_max)
1093 		io_ctx->residual_burst = io_ctx->poll_burst_max;
1094 
1095 	lwkt_replymsg(&nmsg->lmsg, 0);
1096 }
1097 
1098 static int
1099 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
1100 {
1101 	struct iopoll_ctx *io_ctx = arg1;
1102 	struct iopoll_sysctl_netmsg msg;
1103 	uint32_t burst_max;
1104 	int error;
1105 
1106 	burst_max = io_ctx->poll_burst_max;
1107 	error = sysctl_handle_int(oidp, &burst_max, 0, req);
1108 	if (error || req->newptr == NULL)
1109 		return error;
1110 	if (burst_max < MIN_IOPOLL_BURST_MAX)
1111 		burst_max = MIN_IOPOLL_BURST_MAX;
1112 	else if (burst_max > MAX_IOPOLL_BURST_MAX)
1113 		burst_max = MAX_IOPOLL_BURST_MAX;
1114 
1115 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1116 		    0, sysctl_burstmax_handler);
1117 	msg.base.lmsg.u.ms_result = burst_max;
1118 	msg.ctx = io_ctx;
1119 
1120 	return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1121 }
1122 
1123 static void
1124 sysctl_eachburst_handler(netmsg_t nmsg)
1125 {
1126 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1127 	struct iopoll_ctx *io_ctx;
1128 	uint32_t each_burst;
1129 
1130 	io_ctx = msg->ctx;
1131 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1132 
1133 	each_burst = nmsg->lmsg.u.ms_result;
1134 	if (each_burst > io_ctx->poll_burst_max)
1135 		each_burst = io_ctx->poll_burst_max;
1136 	else if (each_burst < 1)
1137 		each_burst = 1;
1138 	io_ctx->poll_each_burst = each_burst;
1139 
1140 	lwkt_replymsg(&nmsg->lmsg, 0);
1141 }
1142 
1143 static int
1144 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
1145 {
1146 	struct iopoll_ctx *io_ctx = arg1;
1147 	struct iopoll_sysctl_netmsg msg;
1148 	uint32_t each_burst;
1149 	int error;
1150 
1151 	each_burst = io_ctx->poll_each_burst;
1152 	error = sysctl_handle_int(oidp, &each_burst, 0, req);
1153 	if (error || req->newptr == NULL)
1154 		return error;
1155 
1156 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1157 		    0, sysctl_eachburst_handler);
1158 	msg.base.lmsg.u.ms_result = each_burst;
1159 	msg.ctx = io_ctx;
1160 
1161 	return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1162 }
1163 
1164 static int
1165 iopoll_register(struct ifnet *ifp, struct iopoll_ctx *io_ctx,
1166 		const struct ifpoll_io *io_rec)
1167 {
1168 	int error;
1169 
1170 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1171 
1172 	if (io_rec->poll_func == NULL)
1173 		return 0;
1174 
1175 	/*
1176 	 * Check if there is room.
1177 	 */
1178 	if (io_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
1179 		/*
1180 		 * List full, cannot register more entries.
1181 		 * This should never happen; if it does, it is probably a
1182 		 * broken driver trying to register multiple times. Checking
1183 		 * this at runtime is expensive, and won't solve the problem
1184 		 * anyways, so just report a few times and then give up.
1185 		 */
1186 		static int verbose = 10; /* XXX */
1187 		if (verbose > 0) {
1188 			kprintf("io poll handlers list full, "
1189 				"maybe a broken driver ?\n");
1190 			verbose--;
1191 		}
1192 		error = ENOENT;
1193 	} else {
1194 		struct iopoll_rec *rec = &io_ctx->pr[io_ctx->poll_handlers];
1195 
1196 		rec->ifp = ifp;
1197 		rec->serializer = io_rec->serializer;
1198 		rec->arg = io_rec->arg;
1199 		rec->poll_func = io_rec->poll_func;
1200 
1201 		io_ctx->poll_handlers++;
1202 		error = 0;
1203 	}
1204 	return error;
1205 }
1206 
1207 static int
1208 iopoll_deregister(struct ifnet *ifp, struct iopoll_ctx *io_ctx)
1209 {
1210 	int i, error;
1211 
1212 	KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1213 
1214 	for (i = 0; i < io_ctx->poll_handlers; ++i) {
1215 		if (io_ctx->pr[i].ifp == ifp) /* Found it */
1216 			break;
1217 	}
1218 	if (i == io_ctx->poll_handlers) {
1219 		error = ENOENT;
1220 	} else {
1221 		io_ctx->poll_handlers--;
1222 		if (i < io_ctx->poll_handlers) {
1223 			/* Last entry replaces this one. */
1224 			io_ctx->pr[i] = io_ctx->pr[io_ctx->poll_handlers];
1225 		}
1226 
1227 		if (io_ctx->poll_handlers == 0)
1228 			iopoll_reset_state(io_ctx);
1229 		error = 0;
1230 	}
1231 	return error;
1232 }
1233 
1234 static void
1235 poll_comm_init(int cpuid)
1236 {
1237 	struct poll_comm *comm;
1238 	char cpuid_str[16];
1239 
1240 	comm = kmalloc_cachealign(sizeof(*comm), M_DEVBUF, M_WAITOK | M_ZERO);
1241 
1242 	if (ifpoll_stfrac < 1)
1243 		ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
1244 	if (ifpoll_txfrac < 1)
1245 		ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
1246 
1247 	comm->poll_cpuid = cpuid;
1248 	comm->pollhz = poll_comm_pollhz_div(comm, ifpoll_pollhz);
1249 	comm->poll_stfrac = ifpoll_stfrac - 1;
1250 	comm->poll_txfrac = ifpoll_txfrac - 1;
1251 
1252 	ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", cpuid);
1253 
1254 	sysctl_ctx_init(&comm->sysctl_ctx);
1255 	comm->sysctl_tree = SYSCTL_ADD_NODE(&comm->sysctl_ctx,
1256 			    SYSCTL_STATIC_CHILDREN(_net_ifpoll),
1257 			    OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
1258 
1259 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1260 			OID_AUTO, "pollhz", CTLTYPE_INT | CTLFLAG_RW,
1261 			comm, 0, sysctl_pollhz,
1262 			"I", "Device polling frequency");
1263 
1264 	if (cpuid == 0) {
1265 		SYSCTL_ADD_PROC(&comm->sysctl_ctx,
1266 				SYSCTL_CHILDREN(comm->sysctl_tree),
1267 				OID_AUTO, "status_frac",
1268 				CTLTYPE_INT | CTLFLAG_RW,
1269 				comm, 0, sysctl_stfrac,
1270 				"I", "# of cycles before status is polled");
1271 	}
1272 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1273 			OID_AUTO, "tx_frac", CTLTYPE_INT | CTLFLAG_RW,
1274 			comm, 0, sysctl_txfrac,
1275 			"I", "# of cycles before TX is polled");
1276 
1277 	poll_common[cpuid] = comm;
1278 }
1279 
1280 static void
1281 poll_comm_start(int cpuid)
1282 {
1283 	struct poll_comm *comm = poll_common[cpuid];
1284 	systimer_func_t func;
1285 
1286 	/*
1287 	 * Initialize systimer
1288 	 */
1289 	if (cpuid == 0)
1290 		func = poll_comm_systimer0;
1291 	else
1292 		func = poll_comm_systimer;
1293 	systimer_init_periodic_nq(&comm->pollclock, func, comm, 1);
1294 }
1295 
1296 static void
1297 _poll_comm_systimer(struct poll_comm *comm)
1298 {
1299 	iopoll_clock(rxpoll_context[comm->poll_cpuid]);
1300 	if (comm->txfrac_count-- == 0) {
1301 		comm->txfrac_count = comm->poll_txfrac;
1302 		iopoll_clock(txpoll_context[comm->poll_cpuid]);
1303 	}
1304 }
1305 
1306 static void
1307 poll_comm_systimer0(systimer_t info, int in_ipi __unused,
1308     struct intrframe *frame __unused)
1309 {
1310 	struct poll_comm *comm = info->data;
1311 	globaldata_t gd = mycpu;
1312 
1313 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid == 0);
1314 
1315 	crit_enter_gd(gd);
1316 
1317 	if (comm->stfrac_count-- == 0) {
1318 		comm->stfrac_count = comm->poll_stfrac;
1319 		stpoll_clock(&stpoll_context);
1320 	}
1321 	_poll_comm_systimer(comm);
1322 
1323 	crit_exit_gd(gd);
1324 }
1325 
1326 static void
1327 poll_comm_systimer(systimer_t info, int in_ipi __unused,
1328     struct intrframe *frame __unused)
1329 {
1330 	struct poll_comm *comm = info->data;
1331 	globaldata_t gd = mycpu;
1332 
1333 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid != 0);
1334 
1335 	crit_enter_gd(gd);
1336 	_poll_comm_systimer(comm);
1337 	crit_exit_gd(gd);
1338 }
1339 
1340 static void
1341 poll_comm_adjust_pollhz(struct poll_comm *comm)
1342 {
1343 	uint32_t handlers;
1344 	int pollhz = 1;
1345 
1346 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1347 
1348 	/*
1349 	 * If there is no polling handler registered, set systimer
1350 	 * frequency to the lowest value.  Polling systimer frequency
1351 	 * will be adjusted to the requested value, once there are
1352 	 * registered handlers.
1353 	 */
1354 	handlers = rxpoll_context[mycpuid]->poll_handlers +
1355 		   txpoll_context[mycpuid]->poll_handlers;
1356 	if (comm->poll_cpuid == 0)
1357 		handlers += stpoll_context.poll_handlers;
1358 	if (handlers)
1359 		pollhz = comm->pollhz;
1360 	systimer_adjust_periodic(&comm->pollclock, pollhz);
1361 }
1362 
1363 static int
1364 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
1365 {
1366 	struct poll_comm *comm = arg1;
1367 	struct netmsg_base nmsg;
1368 	int error, phz;
1369 
1370 	phz = poll_comm_pollhz_conv(comm, comm->pollhz);
1371 	error = sysctl_handle_int(oidp, &phz, 0, req);
1372 	if (error || req->newptr == NULL)
1373 		return error;
1374 	if (phz <= 0)
1375 		return EINVAL;
1376 	else if (phz > IFPOLL_FREQ_MAX)
1377 		phz = IFPOLL_FREQ_MAX;
1378 
1379 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1380 		    0, sysctl_pollhz_handler);
1381 	nmsg.lmsg.u.ms_result = phz;
1382 
1383 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1384 }
1385 
1386 static void
1387 sysctl_pollhz_handler(netmsg_t nmsg)
1388 {
1389 	struct poll_comm *comm = poll_common[mycpuid];
1390 
1391 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1392 
1393 	/* Save polling frequency */
1394 	comm->pollhz = poll_comm_pollhz_div(comm, nmsg->lmsg.u.ms_result);
1395 
1396 	/*
1397 	 * Adjust cached pollhz
1398 	 */
1399 	rxpoll_context[mycpuid]->pollhz = comm->pollhz;
1400 	txpoll_context[mycpuid]->pollhz =
1401 	    comm->pollhz / (comm->poll_txfrac + 1);
1402 
1403 	/*
1404 	 * Adjust polling frequency
1405 	 */
1406 	poll_comm_adjust_pollhz(comm);
1407 
1408 	lwkt_replymsg(&nmsg->lmsg, 0);
1409 }
1410 
1411 static int
1412 sysctl_stfrac(SYSCTL_HANDLER_ARGS)
1413 {
1414 	struct poll_comm *comm = arg1;
1415 	struct netmsg_base nmsg;
1416 	int error, stfrac;
1417 
1418 	KKASSERT(comm->poll_cpuid == 0);
1419 
1420 	stfrac = comm->poll_stfrac + 1;
1421 	error = sysctl_handle_int(oidp, &stfrac, 0, req);
1422 	if (error || req->newptr == NULL)
1423 		return error;
1424 	if (stfrac < 1)
1425 		return EINVAL;
1426 
1427 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1428 		    0, sysctl_stfrac_handler);
1429 	nmsg.lmsg.u.ms_result = stfrac - 1;
1430 
1431 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1432 }
1433 
1434 static void
1435 sysctl_stfrac_handler(netmsg_t nmsg)
1436 {
1437 	struct poll_comm *comm = poll_common[mycpuid];
1438 	int stfrac = nmsg->lmsg.u.ms_result;
1439 
1440 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1441 
1442 	crit_enter();
1443 	comm->poll_stfrac = stfrac;
1444 	if (comm->stfrac_count > comm->poll_stfrac)
1445 		comm->stfrac_count = comm->poll_stfrac;
1446 	crit_exit();
1447 
1448 	lwkt_replymsg(&nmsg->lmsg, 0);
1449 }
1450 
1451 static int
1452 sysctl_txfrac(SYSCTL_HANDLER_ARGS)
1453 {
1454 	struct poll_comm *comm = arg1;
1455 	struct netmsg_base nmsg;
1456 	int error, txfrac;
1457 
1458 	txfrac = comm->poll_txfrac + 1;
1459 	error = sysctl_handle_int(oidp, &txfrac, 0, req);
1460 	if (error || req->newptr == NULL)
1461 		return error;
1462 	if (txfrac < 1)
1463 		return EINVAL;
1464 
1465 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1466 		    0, sysctl_txfrac_handler);
1467 	nmsg.lmsg.u.ms_result = txfrac - 1;
1468 
1469 	return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1470 }
1471 
1472 static void
1473 sysctl_txfrac_handler(netmsg_t nmsg)
1474 {
1475 	struct poll_comm *comm = poll_common[mycpuid];
1476 	int txfrac = nmsg->lmsg.u.ms_result;
1477 
1478 	KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1479 
1480 	crit_enter();
1481 	comm->poll_txfrac = txfrac;
1482 	if (comm->txfrac_count > comm->poll_txfrac)
1483 		comm->txfrac_count = comm->poll_txfrac;
1484 	crit_exit();
1485 
1486 	lwkt_replymsg(&nmsg->lmsg, 0);
1487 }
1488 
1489 void
1490 ifpoll_compat_setup(struct ifpoll_compat *cp,
1491     struct sysctl_ctx_list *sysctl_ctx,
1492     struct sysctl_oid *sysctl_tree,
1493     int unit, struct lwkt_serialize *slz)
1494 {
1495 	cp->ifpc_stcount = 0;
1496 	cp->ifpc_stfrac = ((poll_common[0]->poll_stfrac + 1) *
1497 	    howmany(IOPOLL_BURST_MAX, IOPOLL_EACH_BURST)) - 1;
1498 
1499 	cp->ifpc_cpuid = unit % ncpus2;
1500 	cp->ifpc_serializer = slz;
1501 
1502 	if (sysctl_ctx != NULL && sysctl_tree != NULL) {
1503 		SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1504 		    OID_AUTO, "npoll_stfrac", CTLTYPE_INT | CTLFLAG_RW,
1505 		    cp, 0, sysctl_compat_npoll_stfrac, "I",
1506 		    "polling status frac");
1507 		SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1508 		    OID_AUTO, "npoll_cpuid", CTLTYPE_INT | CTLFLAG_RW,
1509 		    cp, 0, sysctl_compat_npoll_cpuid, "I",
1510 		    "polling cpuid");
1511 	}
1512 }
1513 
1514 static int
1515 sysctl_compat_npoll_stfrac(SYSCTL_HANDLER_ARGS)
1516 {
1517 	struct ifpoll_compat *cp = arg1;
1518 	int error = 0, stfrac;
1519 
1520 	lwkt_serialize_enter(cp->ifpc_serializer);
1521 
1522 	stfrac = cp->ifpc_stfrac + 1;
1523 	error = sysctl_handle_int(oidp, &stfrac, 0, req);
1524 	if (!error && req->newptr != NULL) {
1525 		if (stfrac < 1) {
1526 			error = EINVAL;
1527 		} else {
1528 			cp->ifpc_stfrac = stfrac - 1;
1529 			if (cp->ifpc_stcount > cp->ifpc_stfrac)
1530 				cp->ifpc_stcount = cp->ifpc_stfrac;
1531 		}
1532 	}
1533 
1534 	lwkt_serialize_exit(cp->ifpc_serializer);
1535 	return error;
1536 }
1537 
1538 static int
1539 sysctl_compat_npoll_cpuid(SYSCTL_HANDLER_ARGS)
1540 {
1541 	struct ifpoll_compat *cp = arg1;
1542 	int error = 0, cpuid;
1543 
1544 	lwkt_serialize_enter(cp->ifpc_serializer);
1545 
1546 	cpuid = cp->ifpc_cpuid;
1547 	error = sysctl_handle_int(oidp, &cpuid, 0, req);
1548 	if (!error && req->newptr != NULL) {
1549 		if (cpuid < 0 || cpuid >= ncpus2)
1550 			error = EINVAL;
1551 		else
1552 			cp->ifpc_cpuid = cpuid;
1553 	}
1554 
1555 	lwkt_serialize_exit(cp->ifpc_serializer);
1556 	return error;
1557 }
1558