xref: /dragonfly/sys/net/if_poll.c (revision ed183f8c)
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 #include <sys/microtime_pcpu.h>
40 
41 #include <sys/thread2.h>
42 #include <sys/msgport2.h>
43 
44 #include <net/if.h>
45 #include <net/if_var.h>
46 #include <net/if_poll.h>
47 #include <net/netmsg2.h>
48 #include <net/netisr2.h>
49 
50 /*
51  * Polling support for network device drivers.
52  *
53  * Drivers which support this feature try to register one status polling
54  * handler and several TX/RX polling handlers with the polling code.
55  * If interface's if_npoll is called with non-NULL second argument, then
56  * a register operation is requested, else a deregister operation is
57  * requested.  If the requested operation is "register", driver should
58  * setup the ifpoll_info passed in accoding its own needs:
59  *   ifpoll_info.ifpi_status.status_func == NULL
60  *     No status polling handler will be installed on CPU(0)
61  *   ifpoll_info.ifpi_rx[n].poll_func == NULL
62  *     No RX polling handler will be installed on CPU(n)
63  *   ifpoll_info.ifpi_tx[n].poll_func == NULL
64  *     No TX polling handler will be installed on CPU(n)
65  *
66  * Serializer field of ifpoll_info.ifpi_status and ifpoll_info.ifpi_tx[n]
67  * must _not_ be NULL.  The serializer will be held before the status_func
68  * and poll_func being called.  Serializer field of ifpoll_info.ifpi_rx[n]
69  * can be NULL, but the interface's if_flags must have IFF_IDIRECT set,
70  * which indicates that the network processing of the input packets is
71  * running directly instead of being redispatched.
72  *
73  * RX is polled at the specified polling frequency (net.ifpoll.X.pollhz).
74  * TX and status polling could be done at lower frequency than RX frequency
75  * (net.ifpoll.0.status_frac and net.ifpoll.X.tx_frac).  To avoid systimer
76  * staggering at high frequency, RX systimer gives TX and status polling a
77  * piggyback (XXX).
78  *
79  * All of the registered polling handlers are called only if the interface
80  * is marked as IFF_UP, IFF_RUNNING and IFF_NPOLLING.  However, the
81  * interface's register and deregister function (ifnet.if_npoll) will be
82  * called even if interface is not marked with IFF_RUNNING or IFF_UP.
83  *
84  * If registration is successful, the driver must disable interrupts,
85  * and further I/O is performed through the TX/RX polling handler, which
86  * are invoked (at least once per clock tick) with 3 arguments: the "arg"
87  * passed at register time, a struct ifnet pointer, and a "count" limit.
88  * The registered serializer will be held before calling the related
89  * polling handler.
90  *
91  * The count limit specifies how much work the handler can do during the
92  * call -- typically this is the number of packets to be received, or
93  * transmitted, etc. (drivers are free to interpret this number, as long
94  * as the max time spent in the function grows roughly linearly with the
95  * count).
96  *
97  * A second variable controls the sharing of CPU between polling/kernel
98  * network processing, and other activities (typically userlevel tasks):
99  * net.ifpoll.X.{rx,tx}.user_frac (between 0 and 100, default 50) sets the
100  * share of CPU allocated to user tasks.  CPU is allocated proportionally
101  * to the shares, by dynamically adjusting the "count" (poll_burst).
102  *
103  * Other parameters can should be left to their default values.
104  * The following constraints hold
105  *
106  *	1 <= poll_burst <= poll_burst_max
107  *	1 <= poll_each_burst <= poll_burst_max
108  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
109  */
110 
111 #define IFPOLL_LIST_LEN		128
112 #define IFPOLL_FREQ_MAX		30000
113 
114 #define MIN_IOPOLL_BURST_MAX	10
115 #define MAX_IOPOLL_BURST_MAX	5000
116 #define IOPOLL_BURST_MAX	250	/* good for 1000Mbit net and HZ=6000 */
117 
118 #define IOPOLL_EACH_BURST	50
119 #define IOPOLL_USER_FRAC	50
120 
121 #define IFPOLL_FREQ_DEFAULT	6000
122 
123 #define IFPOLL_TXFRAC_DEFAULT	1	/* 1/1 of the pollhz */
124 #define IFPOLL_STFRAC_DEFAULT	120	/* 1/120 of the pollhz */
125 
126 #define IFPOLL_RX		0x1
127 #define IFPOLL_TX		0x2
128 
129 struct iopoll_rec {
130 	struct lwkt_serialize	*serializer;
131 	struct ifnet		*ifp;
132 	void			*arg;
133 	ifpoll_iofn_t		poll_func;
134 };
135 
136 struct iopoll_ctx {
137 	union microtime_pcpu	prev_t;
138 	u_long			short_ticks;		/* statistics */
139 	u_long			lost_polls;		/* statistics */
140 	u_long			suspect;		/* statistics */
141 	u_long			stalled;		/* statistics */
142 	uint32_t		pending_polls;		/* state */
143 
144 	struct netmsg_base	poll_netmsg;
145 	struct netmsg_base	poll_more_netmsg;
146 
147 	int			poll_cpuid;
148 	int			pollhz;
149 	uint32_t		phase;			/* state */
150 	int			residual_burst;		/* state */
151 	uint32_t		poll_each_burst;	/* tunable */
152 	union microtime_pcpu	poll_start_t;		/* state */
153 
154 	uint32_t		poll_burst;		/* state */
155 	uint32_t		poll_burst_max;		/* tunable */
156 	uint32_t		user_frac;		/* tunable */
157 	uint32_t		kern_frac;		/* state */
158 
159 	uint32_t		poll_handlers; /* next free entry in pr[]. */
160 	struct iopoll_rec	pr[IFPOLL_LIST_LEN];
161 
162 	struct sysctl_ctx_list	poll_sysctl_ctx;
163 	struct sysctl_oid	*poll_sysctl_tree;
164 };
165 
166 struct poll_comm {
167 	struct systimer		pollclock;
168 	int			poll_cpuid;
169 
170 	int			stfrac_count;		/* state */
171 	int			poll_stfrac;		/* tunable */
172 
173 	int			txfrac_count;		/* state */
174 	int			poll_txfrac;		/* tunable */
175 
176 	int			pollhz;			/* tunable */
177 
178 	struct sysctl_ctx_list	sysctl_ctx;
179 	struct sysctl_oid	*sysctl_tree;
180 };
181 
182 struct stpoll_rec {
183 	struct lwkt_serialize	*serializer;
184 	struct ifnet		*ifp;
185 	ifpoll_stfn_t		status_func;
186 };
187 
188 struct stpoll_ctx {
189 	struct netmsg_base	poll_netmsg;
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 static 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 static int	sysctl_compat_npoll_stfrac(SYSCTL_HANDLER_ARGS);
252 static int	sysctl_compat_npoll_cpuid(SYSCTL_HANDLER_ARGS);
253 
254 static struct stpoll_ctx	stpoll_context;
255 static struct poll_comm		*poll_common[MAXCPU];
256 static struct iopoll_ctx	*rxpoll_context[MAXCPU];
257 static struct iopoll_ctx	*txpoll_context[MAXCPU];
258 
259 SYSCTL_NODE(_net, OID_AUTO, ifpoll, CTLFLAG_RW, 0,
260 	    "Network device polling parameters");
261 
262 static int	iopoll_burst_max = IOPOLL_BURST_MAX;
263 static int	iopoll_each_burst = IOPOLL_EACH_BURST;
264 static int	iopoll_user_frac = IOPOLL_USER_FRAC;
265 
266 static int	ifpoll_pollhz = IFPOLL_FREQ_DEFAULT;
267 static int	ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
268 static int	ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
269 
270 TUNABLE_INT("net.ifpoll.burst_max", &iopoll_burst_max);
271 TUNABLE_INT("net.ifpoll.each_burst", &iopoll_each_burst);
272 TUNABLE_INT("net.ifpoll.user_frac", &iopoll_user_frac);
273 TUNABLE_INT("net.ifpoll.pollhz", &ifpoll_pollhz);
274 TUNABLE_INT("net.ifpoll.status_frac", &ifpoll_stfrac);
275 TUNABLE_INT("net.ifpoll.tx_frac", &ifpoll_txfrac);
276 
277 #if !defined(KTR_IF_POLL)
278 #define  KTR_IF_POLL		KTR_ALL
279 #endif
280 KTR_INFO_MASTER(if_poll);
281 KTR_INFO(KTR_IF_POLL, if_poll, rx_start, 0, "rx start");
282 KTR_INFO(KTR_IF_POLL, if_poll, rx_end, 1, "rx end");
283 KTR_INFO(KTR_IF_POLL, if_poll, tx_start, 2, "tx start");
284 KTR_INFO(KTR_IF_POLL, if_poll, tx_end, 3, "tx end");
285 KTR_INFO(KTR_IF_POLL, if_poll, rx_mstart, 4, "rx more start");
286 KTR_INFO(KTR_IF_POLL, if_poll, rx_mend, 5, "rx more end");
287 KTR_INFO(KTR_IF_POLL, if_poll, tx_mstart, 6, "tx more start");
288 KTR_INFO(KTR_IF_POLL, if_poll, tx_mend, 7, "tx more end");
289 KTR_INFO(KTR_IF_POLL, if_poll, ioclock_start, 8, "ioclock start");
290 KTR_INFO(KTR_IF_POLL, if_poll, ioclock_end, 9, "ioclock end");
291 #define logpoll(name)	KTR_LOG(if_poll_ ## name)
292 
293 #define IFPOLL_FREQ_ADJ(comm)	(((comm)->poll_cpuid * 3) % 50)
294 
295 static __inline int
296 poll_comm_pollhz_div(const struct poll_comm *comm, int pollhz)
297 {
298 	return pollhz + IFPOLL_FREQ_ADJ(comm);
299 }
300 
301 static __inline int
302 poll_comm_pollhz_conv(const struct poll_comm *comm, int pollhz)
303 {
304 	return pollhz - IFPOLL_FREQ_ADJ(comm);
305 }
306 
307 static __inline void
308 ifpoll_sendmsg_oncpu(netmsg_t msg)
309 {
310 	if (msg->lmsg.ms_flags & MSGF_DONE)
311 		netisr_sendmsg_oncpu(&msg->base);
312 }
313 
314 static __inline void
315 sched_stpoll(struct stpoll_ctx *st_ctx)
316 {
317 	ifpoll_sendmsg_oncpu((netmsg_t)&st_ctx->poll_netmsg);
318 }
319 
320 static __inline void
321 sched_iopoll(struct iopoll_ctx *io_ctx)
322 {
323 	ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_netmsg);
324 }
325 
326 static __inline void
327 sched_iopollmore(struct iopoll_ctx *io_ctx, boolean_t direct)
328 {
329 
330 	if (!direct) {
331 		ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_more_netmsg);
332 	} else {
333 		struct netmsg_base *nmsg = &io_ctx->poll_more_netmsg;
334 
335 		nmsg->lmsg.ms_flags &= ~(MSGF_REPLY | MSGF_DONE);
336 		nmsg->lmsg.ms_flags |= MSGF_SYNC;
337 		nmsg->nm_dispatch((netmsg_t)nmsg);
338 		KKASSERT(nmsg->lmsg.ms_flags & MSGF_DONE);
339 	}
340 }
341 
342 /*
343  * Initialize per-cpu polling(4) context.
344  */
345 static void
346 ifpoll_init_pcpu(int cpuid)
347 {
348 
349 	poll_comm_init(cpuid);
350 
351 	if (cpuid == 0)
352 		stpoll_init();
353 	iopoll_init(cpuid);
354 
355 	poll_comm_start(cpuid);
356 }
357 
358 static void
359 ifpoll_init_handler(netmsg_t msg)
360 {
361 	int cpu = mycpuid;
362 
363 	ifpoll_init_pcpu(cpu);
364 	netisr_forwardmsg(&msg->base, cpu + 1);
365 }
366 
367 static void
368 ifpoll_sysinit(void *dummy __unused)
369 {
370 	struct netmsg_base msg;
371 
372 	netmsg_init(&msg, NULL, &curthread->td_msgport, 0, ifpoll_init_handler);
373 	netisr_domsg_global(&msg);
374 }
375 SYSINIT(ifpoll, SI_SUB_PRE_DRIVERS, SI_ORDER_ANY, ifpoll_sysinit, NULL);
376 
377 int
378 ifpoll_register(struct ifnet *ifp)
379 {
380 	struct ifpoll_info *info;
381 	struct netmsg_base nmsg;
382 	int error;
383 
384 	if (ifp->if_npoll == NULL) {
385 		/* Device does not support polling */
386 		return EOPNOTSUPP;
387 	}
388 
389 	info = kmalloc(sizeof(*info), M_TEMP, M_WAITOK | M_ZERO);
390 
391 	/*
392 	 * Attempt to register.  Interlock with IFF_NPOLLING.
393 	 */
394 
395 	ifnet_serialize_all(ifp);
396 
397 	if (ifp->if_flags & IFF_NPOLLING) {
398 		/* Already polling */
399 		ifnet_deserialize_all(ifp);
400 		kfree(info, M_TEMP);
401 		return EBUSY;
402 	}
403 
404 	info->ifpi_ifp = ifp;
405 
406 	ifp->if_flags |= IFF_NPOLLING;
407 	ifp->if_npoll(ifp, info);
408 
409 	ifnet_deserialize_all(ifp);
410 
411 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
412 		    0, ifpoll_register_handler);
413 	nmsg.lmsg.u.ms_resultp = info;
414 
415 	error = netisr_domsg_global(&nmsg);
416 	if (error) {
417 		if (!ifpoll_deregister(ifp)) {
418 			if_printf(ifp, "ifpoll_register: "
419 				  "ifpoll_deregister failed!\n");
420 		}
421 	}
422 
423 	kfree(info, M_TEMP);
424 	return error;
425 }
426 
427 int
428 ifpoll_deregister(struct ifnet *ifp)
429 {
430 	struct netmsg_base nmsg;
431 	int error;
432 
433 	if (ifp->if_npoll == NULL)
434 		return EOPNOTSUPP;
435 
436 	ifnet_serialize_all(ifp);
437 
438 	if ((ifp->if_flags & IFF_NPOLLING) == 0) {
439 		ifnet_deserialize_all(ifp);
440 		return EINVAL;
441 	}
442 	ifp->if_flags &= ~IFF_NPOLLING;
443 
444 	ifnet_deserialize_all(ifp);
445 
446 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
447 		    0, ifpoll_deregister_handler);
448 	nmsg.lmsg.u.ms_resultp = ifp;
449 
450 	error = netisr_domsg_global(&nmsg);
451 	if (!error) {
452 		ifnet_serialize_all(ifp);
453 		ifp->if_npoll(ifp, NULL);
454 		ifnet_deserialize_all(ifp);
455 	}
456 	return error;
457 }
458 
459 static void
460 ifpoll_register_handler(netmsg_t nmsg)
461 {
462 	const struct ifpoll_info *info = nmsg->lmsg.u.ms_resultp;
463 	int cpuid = mycpuid;
464 	int error;
465 
466 	KKASSERT(cpuid < netisr_ncpus);
467 	KKASSERT(&curthread->td_msgport == netisr_cpuport(cpuid));
468 
469 	if (cpuid == 0) {
470 		error = stpoll_register(info->ifpi_ifp, &info->ifpi_status);
471 		if (error)
472 			goto failed;
473 	}
474 
475 	error = iopoll_register(info->ifpi_ifp, rxpoll_context[cpuid],
476 				&info->ifpi_rx[cpuid]);
477 	if (error)
478 		goto failed;
479 
480 	error = iopoll_register(info->ifpi_ifp, txpoll_context[cpuid],
481 				&info->ifpi_tx[cpuid]);
482 	if (error)
483 		goto failed;
484 
485 	/* Adjust polling frequency, after all registration is done */
486 	poll_comm_adjust_pollhz(poll_common[cpuid]);
487 
488 	netisr_forwardmsg(&nmsg->base, cpuid + 1);
489 	return;
490 failed:
491 	netisr_replymsg(&nmsg->base, error);
492 }
493 
494 static void
495 ifpoll_deregister_handler(netmsg_t nmsg)
496 {
497 	struct ifnet *ifp = nmsg->lmsg.u.ms_resultp;
498 	int cpuid = mycpuid;
499 
500 	KKASSERT(cpuid < netisr_ncpus);
501 	KKASSERT(&curthread->td_msgport == netisr_cpuport(cpuid));
502 
503 	/* Ignore errors */
504 	if (cpuid == 0)
505 		stpoll_deregister(ifp);
506 	iopoll_deregister(ifp, rxpoll_context[cpuid]);
507 	iopoll_deregister(ifp, txpoll_context[cpuid]);
508 
509 	/* Adjust polling frequency, after all deregistration is done */
510 	poll_comm_adjust_pollhz(poll_common[cpuid]);
511 
512 	netisr_forwardmsg(&nmsg->base, cpuid + 1);
513 }
514 
515 static void
516 stpoll_init(void)
517 {
518 	struct stpoll_ctx *st_ctx = &stpoll_context;
519 	const struct poll_comm *comm = poll_common[0];
520 
521 	sysctl_ctx_init(&st_ctx->poll_sysctl_ctx);
522 	st_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&st_ctx->poll_sysctl_ctx,
523 				   SYSCTL_CHILDREN(comm->sysctl_tree),
524 				   OID_AUTO, "status", CTLFLAG_RD, 0, "");
525 
526 	SYSCTL_ADD_UINT(&st_ctx->poll_sysctl_ctx,
527 			SYSCTL_CHILDREN(st_ctx->poll_sysctl_tree),
528 			OID_AUTO, "handlers", CTLFLAG_RD,
529 			&st_ctx->poll_handlers, 0,
530 			"Number of registered status poll handlers");
531 
532 	netmsg_init(&st_ctx->poll_netmsg, NULL, &netisr_adone_rport,
533 		    0, stpoll_handler);
534 }
535 
536 /*
537  * stpoll_handler is scheduled by sched_stpoll when appropriate, typically
538  * once per polling systimer tick.
539  */
540 static void
541 stpoll_handler(netmsg_t msg)
542 {
543 	struct stpoll_ctx *st_ctx = &stpoll_context;
544 	struct thread *td = curthread;
545 	int i;
546 
547 	ASSERT_NETISR0;
548 
549 	crit_enter_quick(td);
550 
551 	/* Reply ASAP */
552 	netisr_replymsg(&msg->base, 0);
553 
554 	if (st_ctx->poll_handlers == 0) {
555 		crit_exit_quick(td);
556 		return;
557 	}
558 
559 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
560 		const struct stpoll_rec *rec = &st_ctx->pr[i];
561 		struct ifnet *ifp = rec->ifp;
562 
563 		if (!lwkt_serialize_try(rec->serializer))
564 			continue;
565 
566 		if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
567 		    (IFF_RUNNING | IFF_NPOLLING))
568 			rec->status_func(ifp);
569 
570 		lwkt_serialize_exit(rec->serializer);
571 	}
572 
573 	crit_exit_quick(td);
574 }
575 
576 /*
577  * Hook from status poll systimer.  Tries to schedule an status poll.
578  * NOTE: Caller should hold critical section.
579  */
580 static void
581 stpoll_clock(struct stpoll_ctx *st_ctx)
582 {
583 	KKASSERT(mycpuid == 0);
584 
585 	if (st_ctx->poll_handlers == 0)
586 		return;
587 	sched_stpoll(st_ctx);
588 }
589 
590 static int
591 stpoll_register(struct ifnet *ifp, const struct ifpoll_status *st_rec)
592 {
593 	struct stpoll_ctx *st_ctx = &stpoll_context;
594 	int error;
595 
596 	ASSERT_NETISR0;
597 
598 	if (st_rec->status_func == NULL)
599 		return 0;
600 
601 	/*
602 	 * Check if there is room.
603 	 */
604 	if (st_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
605 		/*
606 		 * List full, cannot register more entries.
607 		 * This should never happen; if it does, it is probably a
608 		 * broken driver trying to register multiple times. Checking
609 		 * this at runtime is expensive, and won't solve the problem
610 		 * anyways, so just report a few times and then give up.
611 		 */
612 		static int verbose = 10; /* XXX */
613 
614 		if (verbose > 0) {
615 			kprintf("status poll handlers list full, "
616 				"maybe a broken driver ?\n");
617 			verbose--;
618 		}
619 		error = ENOENT;
620 	} else {
621 		struct stpoll_rec *rec = &st_ctx->pr[st_ctx->poll_handlers];
622 
623 		rec->ifp = ifp;
624 		rec->serializer = st_rec->serializer;
625 		rec->status_func = st_rec->status_func;
626 
627 		st_ctx->poll_handlers++;
628 		error = 0;
629 	}
630 	return error;
631 }
632 
633 static int
634 stpoll_deregister(struct ifnet *ifp)
635 {
636 	struct stpoll_ctx *st_ctx = &stpoll_context;
637 	int i, error;
638 
639 	ASSERT_NETISR0;
640 
641 	for (i = 0; i < st_ctx->poll_handlers; ++i) {
642 		if (st_ctx->pr[i].ifp == ifp) /* Found it */
643 			break;
644 	}
645 	if (i == st_ctx->poll_handlers) {
646 		error = ENOENT;
647 	} else {
648 		st_ctx->poll_handlers--;
649 		if (i < st_ctx->poll_handlers) {
650 			/* Last entry replaces this one. */
651 			st_ctx->pr[i] = st_ctx->pr[st_ctx->poll_handlers];
652 		}
653 		error = 0;
654 	}
655 	return error;
656 }
657 
658 static __inline void
659 iopoll_reset_state(struct iopoll_ctx *io_ctx)
660 {
661 	crit_enter();
662 	io_ctx->poll_burst = io_ctx->poll_each_burst;
663 	io_ctx->pending_polls = 0;
664 	io_ctx->residual_burst = 0;
665 	io_ctx->phase = 0;
666 	io_ctx->kern_frac = 0;
667 	bzero(&io_ctx->poll_start_t, sizeof(io_ctx->poll_start_t));
668 	bzero(&io_ctx->prev_t, sizeof(io_ctx->prev_t));
669 	crit_exit();
670 }
671 
672 static void
673 iopoll_init(int cpuid)
674 {
675 	KKASSERT(cpuid < netisr_ncpus);
676 
677 	rxpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_RX);
678 	txpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_TX);
679 }
680 
681 static struct iopoll_ctx *
682 iopoll_ctx_create(int cpuid, int poll_type)
683 {
684 	struct poll_comm *comm;
685 	struct iopoll_ctx *io_ctx;
686 	const char *poll_type_str;
687 	netisr_fn_t handler, more_handler;
688 
689 	KKASSERT(poll_type == IFPOLL_RX || poll_type == IFPOLL_TX);
690 
691 	/*
692 	 * Make sure that tunables are in sane state
693 	 */
694 	if (iopoll_burst_max < MIN_IOPOLL_BURST_MAX)
695 		iopoll_burst_max = MIN_IOPOLL_BURST_MAX;
696 	else if (iopoll_burst_max > MAX_IOPOLL_BURST_MAX)
697 		iopoll_burst_max = MAX_IOPOLL_BURST_MAX;
698 
699 	if (iopoll_each_burst > iopoll_burst_max)
700 		iopoll_each_burst = iopoll_burst_max;
701 
702 	comm = poll_common[cpuid];
703 
704 	/*
705 	 * Create the per-cpu polling context
706 	 */
707 	io_ctx = kmalloc(sizeof(*io_ctx), M_DEVBUF, M_WAITOK | M_ZERO);
708 
709 	io_ctx->poll_each_burst = iopoll_each_burst;
710 	io_ctx->poll_burst_max = iopoll_burst_max;
711 	io_ctx->user_frac = iopoll_user_frac;
712 	if (poll_type == IFPOLL_RX)
713 		io_ctx->pollhz = comm->pollhz;
714 	else
715 		io_ctx->pollhz = comm->pollhz / (comm->poll_txfrac + 1);
716 	io_ctx->poll_cpuid = cpuid;
717 	iopoll_reset_state(io_ctx);
718 
719 	if (poll_type == IFPOLL_RX) {
720 		handler = rxpoll_handler;
721 		more_handler = rxpollmore_handler;
722 	} else {
723 		handler = txpoll_handler;
724 		more_handler = txpollmore_handler;
725 	}
726 
727 	netmsg_init(&io_ctx->poll_netmsg, NULL, &netisr_adone_rport,
728 	    0, handler);
729 	io_ctx->poll_netmsg.lmsg.u.ms_resultp = io_ctx;
730 
731 	netmsg_init(&io_ctx->poll_more_netmsg, NULL, &netisr_adone_rport,
732 	    0, more_handler);
733 	io_ctx->poll_more_netmsg.lmsg.u.ms_resultp = io_ctx;
734 
735 	/*
736 	 * Initialize per-cpu sysctl nodes
737 	 */
738 	if (poll_type == IFPOLL_RX)
739 		poll_type_str = "rx";
740 	else
741 		poll_type_str = "tx";
742 
743 	sysctl_ctx_init(&io_ctx->poll_sysctl_ctx);
744 	io_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&io_ctx->poll_sysctl_ctx,
745 				   SYSCTL_CHILDREN(comm->sysctl_tree),
746 				   OID_AUTO, poll_type_str, CTLFLAG_RD, 0, "");
747 	iopoll_add_sysctl(&io_ctx->poll_sysctl_ctx,
748 	    SYSCTL_CHILDREN(io_ctx->poll_sysctl_tree), io_ctx, poll_type);
749 
750 	return io_ctx;
751 }
752 
753 /*
754  * Hook from iopoll systimer.  Tries to schedule an iopoll, but keeps
755  * track of lost ticks due to the previous handler taking too long.
756  * Normally, this should not happen, because polling handler should
757  * run for a short time.  However, in some cases (e.g. when there are
758  * changes in link status etc.) the drivers take a very long time
759  * (even in the order of milliseconds) to reset and reconfigure the
760  * device, causing apparent lost polls.
761  *
762  * The first part of the code is just for debugging purposes, and tries
763  * to count how often hardclock ticks are shorter than they should,
764  * meaning either stray interrupts or delayed events.
765  *
766  * WARNING! called from fastint or IPI, the MP lock might not be held.
767  * NOTE: Caller should hold critical section.
768  */
769 static void
770 iopoll_clock(struct iopoll_ctx *io_ctx)
771 {
772 	union microtime_pcpu t;
773 	int delta;
774 
775 	KKASSERT(mycpuid == io_ctx->poll_cpuid);
776 
777 	if (io_ctx->poll_handlers == 0)
778 		return;
779 
780 	logpoll(ioclock_start);
781 
782 	microtime_pcpu_get(&t);
783 	delta = microtime_pcpu_diff(&io_ctx->prev_t, &t);
784 	if (delta * io_ctx->pollhz < 500000)
785 		io_ctx->short_ticks++;
786 	else
787 		io_ctx->prev_t = t;
788 
789 	if (io_ctx->pending_polls > 100) {
790 		/*
791 		 * Too much, assume it has stalled (not always true
792 		 * see comment above).
793 		 */
794 		io_ctx->stalled++;
795 		io_ctx->pending_polls = 0;
796 		io_ctx->phase = 0;
797 	}
798 
799 	if (io_ctx->phase <= 2) {
800 		if (io_ctx->phase != 0)
801 			io_ctx->suspect++;
802 		io_ctx->phase = 1;
803 		sched_iopoll(io_ctx);
804 		io_ctx->phase = 2;
805 	}
806 	if (io_ctx->pending_polls++ > 0)
807 		io_ctx->lost_polls++;
808 
809 	logpoll(ioclock_end);
810 }
811 
812 /*
813  * rxpoll_handler and txpoll_handler are scheduled by sched_iopoll when
814  * appropriate, typically once per polling systimer tick.
815  *
816  * Note that the message is replied immediately in order to allow a new
817  * ISR to be scheduled in the handler.
818  */
819 static void
820 rxpoll_handler(netmsg_t msg)
821 {
822 	struct iopoll_ctx *io_ctx;
823 	struct thread *td = curthread;
824 	boolean_t direct = TRUE, crit;
825 	int i, cycles;
826 
827 	logpoll(rx_start);
828 
829 	io_ctx = msg->lmsg.u.ms_resultp;
830 	KKASSERT(&td->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
831 
832 	crit = TRUE;
833 	crit_enter_quick(td);
834 
835 	/* Reply ASAP */
836 	netisr_replymsg(&msg->base, 0);
837 
838 	if (io_ctx->poll_handlers == 0) {
839 		crit_exit_quick(td);
840 		logpoll(rx_end);
841 		return;
842 	}
843 
844 	io_ctx->phase = 3;
845 	if (io_ctx->residual_burst == 0) {
846 		/* First call in this tick */
847 		microtime_pcpu_get(&io_ctx->poll_start_t);
848 		io_ctx->residual_burst = io_ctx->poll_burst;
849 	}
850 	cycles = (io_ctx->residual_burst < io_ctx->poll_each_burst) ?
851 		 io_ctx->residual_burst : io_ctx->poll_each_burst;
852 	io_ctx->residual_burst -= cycles;
853 
854 	for (i = 0; i < io_ctx->poll_handlers; i++) {
855 		const struct iopoll_rec *rec = &io_ctx->pr[i];
856 		struct ifnet *ifp = rec->ifp;
857 
858 		if (rec->serializer != NULL) {
859 			if (!crit) {
860 				crit = TRUE;
861 				crit_enter_quick(td);
862 			}
863 			if (__predict_false(
864 			    !lwkt_serialize_try(rec->serializer))) {
865 				/* RX serializer generally will not fail. */
866 				continue;
867 			}
868 		} else if (crit) {
869 			/*
870 			 * Exit critical section, if the RX polling
871 			 * handler does not require serialization,
872 			 * i.e. RX polling is doing direct input.
873 			 */
874 			crit_exit_quick(td);
875 			crit = FALSE;
876 		}
877 
878 		if ((ifp->if_flags & IFF_IDIRECT) == 0) {
879 			direct = FALSE;
880 			KASSERT(rec->serializer != NULL,
881 			    ("rx polling handler is not serialized"));
882 		}
883 #ifdef INVARIANTS
884 		else {
885 			KASSERT(rec->serializer == NULL,
886 			    ("serialized direct input"));
887 		}
888 #endif
889 
890 		if ((ifp->if_flags & (IFF_UP | IFF_RUNNING | IFF_NPOLLING)) ==
891 		    (IFF_UP | IFF_RUNNING | IFF_NPOLLING))
892 			rec->poll_func(ifp, rec->arg, cycles);
893 
894 		if (rec->serializer != NULL)
895 			lwkt_serialize_exit(rec->serializer);
896 	}
897 
898 	if (crit) {
899 		/*
900 		 * Do a quick exit/enter to catch any higher-priority
901 		 * interrupt sources.
902 		 */
903 		crit_exit_quick(td);
904 	}
905 	crit_enter_quick(td);
906 
907 	io_ctx->phase = 4;
908 	sched_iopollmore(io_ctx, direct);
909 
910 	crit_exit_quick(td);
911 
912 	logpoll(rx_end);
913 }
914 
915 static void
916 txpoll_handler(netmsg_t msg)
917 {
918 	struct iopoll_ctx *io_ctx;
919 	struct thread *td = curthread;
920 	int i;
921 
922 	logpoll(tx_start);
923 
924 	io_ctx = msg->lmsg.u.ms_resultp;
925 	KKASSERT(&td->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
926 
927 	crit_enter_quick(td);
928 
929 	/* Reply ASAP */
930 	netisr_replymsg(&msg->base, 0);
931 
932 	if (io_ctx->poll_handlers == 0) {
933 		crit_exit_quick(td);
934 		logpoll(tx_end);
935 		return;
936 	}
937 
938 	io_ctx->phase = 3;
939 
940 	for (i = 0; i < io_ctx->poll_handlers; i++) {
941 		const struct iopoll_rec *rec = &io_ctx->pr[i];
942 		struct ifnet *ifp = rec->ifp;
943 
944 		if (!lwkt_serialize_try(rec->serializer))
945 			continue;
946 
947 		if ((ifp->if_flags & (IFF_UP | IFF_RUNNING | IFF_NPOLLING)) ==
948 		    (IFF_UP | IFF_RUNNING | IFF_NPOLLING))
949 			rec->poll_func(ifp, rec->arg, -1);
950 
951 		lwkt_serialize_exit(rec->serializer);
952 	}
953 
954 	/*
955 	 * Do a quick exit/enter to catch any higher-priority
956 	 * interrupt sources.
957 	 */
958 	crit_exit_quick(td);
959 	crit_enter_quick(td);
960 
961 	io_ctx->phase = 4;
962 	sched_iopollmore(io_ctx, TRUE);
963 
964 	crit_exit_quick(td);
965 
966 	logpoll(tx_end);
967 }
968 
969 /*
970  * rxpollmore_handler and txpollmore_handler are called after other netisr's,
971  * possibly scheduling another rxpoll_handler or txpoll_handler call, or
972  * adapting the burst size for the next cycle.
973  *
974  * It is very bad to fetch large bursts of packets from a single card at once,
975  * because the burst could take a long time to be completely processed leading
976  * to unfairness.  To reduce the problem, and also to account better for time
977  * spent in network-related processing, we split the burst in smaller chunks
978  * of fixed size, giving control to the other netisr's between chunks.  This
979  * helps in improving the fairness, reducing livelock and accounting for the
980  * work performed in low level handling.
981  */
982 static void
983 rxpollmore_handler(netmsg_t msg)
984 {
985 	struct thread *td = curthread;
986 	struct iopoll_ctx *io_ctx;
987 	union microtime_pcpu t;
988 	int kern_load;
989 	uint32_t pending_polls;
990 
991 	logpoll(rx_mstart);
992 
993 	io_ctx = msg->lmsg.u.ms_resultp;
994 	KKASSERT(&td->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
995 
996 	crit_enter_quick(td);
997 
998 	/* Replay ASAP */
999 	netisr_replymsg(&msg->base, 0);
1000 
1001 	if (io_ctx->poll_handlers == 0) {
1002 		crit_exit_quick(td);
1003 		logpoll(rx_mend);
1004 		return;
1005 	}
1006 
1007 	io_ctx->phase = 5;
1008 	if (io_ctx->residual_burst > 0) {
1009 		sched_iopoll(io_ctx);
1010 		crit_exit_quick(td);
1011 		/* Will run immediately on return, followed by netisrs */
1012 		logpoll(rx_mend);
1013 		return;
1014 	}
1015 
1016 	/* Here we can account time spent in iopoll's in this tick */
1017 	microtime_pcpu_get(&t);
1018 	kern_load = microtime_pcpu_diff(&io_ctx->poll_start_t, &t);
1019 	kern_load = (kern_load * io_ctx->pollhz) / 10000; /* 0..100 */
1020 	io_ctx->kern_frac = kern_load;
1021 
1022 	if (kern_load > (100 - io_ctx->user_frac)) {
1023 		/* Try decrease ticks */
1024 		if (io_ctx->poll_burst > 1)
1025 			io_ctx->poll_burst--;
1026 	} else {
1027 		if (io_ctx->poll_burst < io_ctx->poll_burst_max)
1028 			io_ctx->poll_burst++;
1029 	}
1030 
1031 	io_ctx->pending_polls--;
1032 	pending_polls = io_ctx->pending_polls;
1033 
1034 	if (pending_polls == 0) {
1035 		/* We are done */
1036 		io_ctx->phase = 0;
1037 	} else {
1038 		/*
1039 		 * Last cycle was long and caused us to miss one or more
1040 		 * hardclock ticks.  Restart processing again, but slightly
1041 		 * reduce the burst size to prevent that this happens again.
1042 		 */
1043 		io_ctx->poll_burst -= (io_ctx->poll_burst / 8);
1044 		if (io_ctx->poll_burst < 1)
1045 			io_ctx->poll_burst = 1;
1046 		sched_iopoll(io_ctx);
1047 		io_ctx->phase = 6;
1048 	}
1049 
1050 	crit_exit_quick(td);
1051 
1052 	logpoll(rx_mend);
1053 }
1054 
1055 static void
1056 txpollmore_handler(netmsg_t msg)
1057 {
1058 	struct thread *td = curthread;
1059 	struct iopoll_ctx *io_ctx;
1060 	uint32_t pending_polls;
1061 
1062 	logpoll(tx_mstart);
1063 
1064 	io_ctx = msg->lmsg.u.ms_resultp;
1065 	KKASSERT(&td->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
1066 
1067 	crit_enter_quick(td);
1068 
1069 	/* Replay ASAP */
1070 	netisr_replymsg(&msg->base, 0);
1071 
1072 	if (io_ctx->poll_handlers == 0) {
1073 		crit_exit_quick(td);
1074 		logpoll(tx_mend);
1075 		return;
1076 	}
1077 
1078 	io_ctx->phase = 5;
1079 
1080 	io_ctx->pending_polls--;
1081 	pending_polls = io_ctx->pending_polls;
1082 
1083 	if (pending_polls == 0) {
1084 		/* We are done */
1085 		io_ctx->phase = 0;
1086 	} else {
1087 		/*
1088 		 * Last cycle was long and caused us to miss one or more
1089 		 * hardclock ticks.  Restart processing again.
1090 		 */
1091 		sched_iopoll(io_ctx);
1092 		io_ctx->phase = 6;
1093 	}
1094 
1095 	crit_exit_quick(td);
1096 
1097 	logpoll(tx_mend);
1098 }
1099 
1100 static void
1101 iopoll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
1102     struct iopoll_ctx *io_ctx, int poll_type)
1103 {
1104 	if (poll_type == IFPOLL_RX) {
1105 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
1106 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_burstmax,
1107 		    "IU", "Max Polling burst size");
1108 
1109 		SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
1110 		    CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_eachburst,
1111 		    "IU", "Max size of each burst");
1112 
1113 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
1114 		    &io_ctx->poll_burst, 0, "Current polling burst size");
1115 
1116 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
1117 		    &io_ctx->user_frac, 0, "Desired user fraction of cpu time");
1118 
1119 		SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "kern_frac", CTLFLAG_RD,
1120 		    &io_ctx->kern_frac, 0, "Kernel fraction of cpu time");
1121 
1122 		SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
1123 		    &io_ctx->residual_burst, 0,
1124 		    "# of residual cycles in burst");
1125 	}
1126 
1127 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
1128 	    &io_ctx->phase, 0, "Polling phase");
1129 
1130 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
1131 	    &io_ctx->suspect, "Suspected events");
1132 
1133 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
1134 	    &io_ctx->stalled, "Potential stalls");
1135 
1136 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
1137 	    &io_ctx->short_ticks,
1138 	    "Hardclock ticks shorter than they should be");
1139 
1140 	SYSCTL_ADD_ULONG(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
1141 	    &io_ctx->lost_polls,
1142 	    "How many times we would have lost a poll tick");
1143 
1144 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
1145 	    &io_ctx->pending_polls, 0, "Do we need to poll again");
1146 
1147 	SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
1148 	    &io_ctx->poll_handlers, 0, "Number of registered poll handlers");
1149 }
1150 
1151 static void
1152 sysctl_burstmax_handler(netmsg_t nmsg)
1153 {
1154 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1155 	struct iopoll_ctx *io_ctx;
1156 
1157 	io_ctx = msg->ctx;
1158 	KKASSERT(&curthread->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
1159 
1160 	io_ctx->poll_burst_max = nmsg->lmsg.u.ms_result;
1161 	if (io_ctx->poll_each_burst > io_ctx->poll_burst_max)
1162 		io_ctx->poll_each_burst = io_ctx->poll_burst_max;
1163 	if (io_ctx->poll_burst > io_ctx->poll_burst_max)
1164 		io_ctx->poll_burst = io_ctx->poll_burst_max;
1165 	if (io_ctx->residual_burst > io_ctx->poll_burst_max)
1166 		io_ctx->residual_burst = io_ctx->poll_burst_max;
1167 
1168 	netisr_replymsg(&nmsg->base, 0);
1169 }
1170 
1171 static int
1172 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
1173 {
1174 	struct iopoll_ctx *io_ctx = arg1;
1175 	struct iopoll_sysctl_netmsg msg;
1176 	uint32_t burst_max;
1177 	int error;
1178 
1179 	burst_max = io_ctx->poll_burst_max;
1180 	error = sysctl_handle_int(oidp, &burst_max, 0, req);
1181 	if (error || req->newptr == NULL)
1182 		return error;
1183 	if (burst_max < MIN_IOPOLL_BURST_MAX)
1184 		burst_max = MIN_IOPOLL_BURST_MAX;
1185 	else if (burst_max > MAX_IOPOLL_BURST_MAX)
1186 		burst_max = MAX_IOPOLL_BURST_MAX;
1187 
1188 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1189 		    0, sysctl_burstmax_handler);
1190 	msg.base.lmsg.u.ms_result = burst_max;
1191 	msg.ctx = io_ctx;
1192 
1193 	return netisr_domsg(&msg.base, io_ctx->poll_cpuid);
1194 }
1195 
1196 static void
1197 sysctl_eachburst_handler(netmsg_t nmsg)
1198 {
1199 	struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1200 	struct iopoll_ctx *io_ctx;
1201 	uint32_t each_burst;
1202 
1203 	io_ctx = msg->ctx;
1204 	KKASSERT(&curthread->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
1205 
1206 	each_burst = nmsg->lmsg.u.ms_result;
1207 	if (each_burst > io_ctx->poll_burst_max)
1208 		each_burst = io_ctx->poll_burst_max;
1209 	else if (each_burst < 1)
1210 		each_burst = 1;
1211 	io_ctx->poll_each_burst = each_burst;
1212 
1213 	netisr_replymsg(&nmsg->base, 0);
1214 }
1215 
1216 static int
1217 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
1218 {
1219 	struct iopoll_ctx *io_ctx = arg1;
1220 	struct iopoll_sysctl_netmsg msg;
1221 	uint32_t each_burst;
1222 	int error;
1223 
1224 	each_burst = io_ctx->poll_each_burst;
1225 	error = sysctl_handle_int(oidp, &each_burst, 0, req);
1226 	if (error || req->newptr == NULL)
1227 		return error;
1228 
1229 	netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1230 		    0, sysctl_eachburst_handler);
1231 	msg.base.lmsg.u.ms_result = each_burst;
1232 	msg.ctx = io_ctx;
1233 
1234 	return netisr_domsg(&msg.base, io_ctx->poll_cpuid);
1235 }
1236 
1237 static int
1238 iopoll_register(struct ifnet *ifp, struct iopoll_ctx *io_ctx,
1239 		const struct ifpoll_io *io_rec)
1240 {
1241 	int error;
1242 
1243 	KKASSERT(&curthread->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
1244 
1245 	if (io_rec->poll_func == NULL)
1246 		return 0;
1247 
1248 	/*
1249 	 * Check if there is room.
1250 	 */
1251 	if (io_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
1252 		/*
1253 		 * List full, cannot register more entries.
1254 		 * This should never happen; if it does, it is probably a
1255 		 * broken driver trying to register multiple times. Checking
1256 		 * this at runtime is expensive, and won't solve the problem
1257 		 * anyways, so just report a few times and then give up.
1258 		 */
1259 		static int verbose = 10; /* XXX */
1260 		if (verbose > 0) {
1261 			kprintf("io poll handlers list full, "
1262 				"maybe a broken driver ?\n");
1263 			verbose--;
1264 		}
1265 		error = ENOENT;
1266 	} else {
1267 		struct iopoll_rec *rec = &io_ctx->pr[io_ctx->poll_handlers];
1268 
1269 		rec->ifp = ifp;
1270 		rec->serializer = io_rec->serializer;
1271 		rec->arg = io_rec->arg;
1272 		rec->poll_func = io_rec->poll_func;
1273 
1274 		io_ctx->poll_handlers++;
1275 		error = 0;
1276 	}
1277 	return error;
1278 }
1279 
1280 static int
1281 iopoll_deregister(struct ifnet *ifp, struct iopoll_ctx *io_ctx)
1282 {
1283 	int i, error;
1284 
1285 	KKASSERT(&curthread->td_msgport == netisr_cpuport(io_ctx->poll_cpuid));
1286 
1287 	for (i = 0; i < io_ctx->poll_handlers; ++i) {
1288 		if (io_ctx->pr[i].ifp == ifp) /* Found it */
1289 			break;
1290 	}
1291 	if (i == io_ctx->poll_handlers) {
1292 		error = ENOENT;
1293 	} else {
1294 		io_ctx->poll_handlers--;
1295 		if (i < io_ctx->poll_handlers) {
1296 			/* Last entry replaces this one. */
1297 			io_ctx->pr[i] = io_ctx->pr[io_ctx->poll_handlers];
1298 		}
1299 
1300 		if (io_ctx->poll_handlers == 0)
1301 			iopoll_reset_state(io_ctx);
1302 		error = 0;
1303 	}
1304 	return error;
1305 }
1306 
1307 static void
1308 poll_comm_init(int cpuid)
1309 {
1310 	struct poll_comm *comm;
1311 	char cpuid_str[16];
1312 
1313 	comm = kmalloc(sizeof(*comm), M_DEVBUF, M_WAITOK | M_ZERO);
1314 
1315 	if (ifpoll_stfrac < 1)
1316 		ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
1317 	if (ifpoll_txfrac < 1)
1318 		ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
1319 
1320 	comm->poll_cpuid = cpuid;
1321 	comm->pollhz = poll_comm_pollhz_div(comm, ifpoll_pollhz);
1322 	comm->poll_stfrac = ifpoll_stfrac - 1;
1323 	comm->poll_txfrac = ifpoll_txfrac - 1;
1324 
1325 	ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", cpuid);
1326 
1327 	sysctl_ctx_init(&comm->sysctl_ctx);
1328 	comm->sysctl_tree = SYSCTL_ADD_NODE(&comm->sysctl_ctx,
1329 			    SYSCTL_STATIC_CHILDREN(_net_ifpoll),
1330 			    OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
1331 
1332 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1333 			OID_AUTO, "pollhz", CTLTYPE_INT | CTLFLAG_RW,
1334 			comm, 0, sysctl_pollhz,
1335 			"I", "Device polling frequency");
1336 
1337 	if (cpuid == 0) {
1338 		SYSCTL_ADD_PROC(&comm->sysctl_ctx,
1339 				SYSCTL_CHILDREN(comm->sysctl_tree),
1340 				OID_AUTO, "status_frac",
1341 				CTLTYPE_INT | CTLFLAG_RW,
1342 				comm, 0, sysctl_stfrac,
1343 				"I", "# of cycles before status is polled");
1344 	}
1345 	SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1346 			OID_AUTO, "tx_frac", CTLTYPE_INT | CTLFLAG_RW,
1347 			comm, 0, sysctl_txfrac,
1348 			"I", "# of cycles before TX is polled");
1349 
1350 	poll_common[cpuid] = comm;
1351 }
1352 
1353 static void
1354 poll_comm_start(int cpuid)
1355 {
1356 	struct poll_comm *comm = poll_common[cpuid];
1357 	systimer_func_t func;
1358 
1359 	/*
1360 	 * Initialize systimer
1361 	 */
1362 	if (cpuid == 0)
1363 		func = poll_comm_systimer0;
1364 	else
1365 		func = poll_comm_systimer;
1366 	systimer_init_periodic_nq(&comm->pollclock, func, comm, 1);
1367 }
1368 
1369 static void
1370 _poll_comm_systimer(struct poll_comm *comm)
1371 {
1372 	iopoll_clock(rxpoll_context[comm->poll_cpuid]);
1373 	if (comm->txfrac_count-- == 0) {
1374 		comm->txfrac_count = comm->poll_txfrac;
1375 		iopoll_clock(txpoll_context[comm->poll_cpuid]);
1376 	}
1377 }
1378 
1379 static void
1380 poll_comm_systimer0(systimer_t info, int in_ipi __unused,
1381     struct intrframe *frame __unused)
1382 {
1383 	struct poll_comm *comm = info->data;
1384 	globaldata_t gd = mycpu;
1385 
1386 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid == 0);
1387 
1388 	crit_enter_gd(gd);
1389 
1390 	if (comm->stfrac_count-- == 0) {
1391 		comm->stfrac_count = comm->poll_stfrac;
1392 		stpoll_clock(&stpoll_context);
1393 	}
1394 	_poll_comm_systimer(comm);
1395 
1396 	crit_exit_gd(gd);
1397 }
1398 
1399 static void
1400 poll_comm_systimer(systimer_t info, int in_ipi __unused,
1401     struct intrframe *frame __unused)
1402 {
1403 	struct poll_comm *comm = info->data;
1404 	globaldata_t gd = mycpu;
1405 
1406 	KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid != 0);
1407 
1408 	crit_enter_gd(gd);
1409 	_poll_comm_systimer(comm);
1410 	crit_exit_gd(gd);
1411 }
1412 
1413 static void
1414 poll_comm_adjust_pollhz(struct poll_comm *comm)
1415 {
1416 	uint32_t handlers;
1417 	int pollhz = 1;
1418 
1419 	KKASSERT(&curthread->td_msgport == netisr_cpuport(comm->poll_cpuid));
1420 
1421 	/*
1422 	 * If there is no polling handler registered, set systimer
1423 	 * frequency to the lowest value.  Polling systimer frequency
1424 	 * will be adjusted to the requested value, once there are
1425 	 * registered handlers.
1426 	 */
1427 	handlers = rxpoll_context[mycpuid]->poll_handlers +
1428 		   txpoll_context[mycpuid]->poll_handlers;
1429 	if (comm->poll_cpuid == 0)
1430 		handlers += stpoll_context.poll_handlers;
1431 	if (handlers)
1432 		pollhz = comm->pollhz;
1433 	systimer_adjust_periodic(&comm->pollclock, pollhz);
1434 }
1435 
1436 static int
1437 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
1438 {
1439 	struct poll_comm *comm = arg1;
1440 	struct netmsg_base nmsg;
1441 	int error, phz;
1442 
1443 	phz = poll_comm_pollhz_conv(comm, comm->pollhz);
1444 	error = sysctl_handle_int(oidp, &phz, 0, req);
1445 	if (error || req->newptr == NULL)
1446 		return error;
1447 	if (phz <= 0)
1448 		return EINVAL;
1449 	else if (phz > IFPOLL_FREQ_MAX)
1450 		phz = IFPOLL_FREQ_MAX;
1451 
1452 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1453 		    0, sysctl_pollhz_handler);
1454 	nmsg.lmsg.u.ms_result = phz;
1455 
1456 	return netisr_domsg(&nmsg, comm->poll_cpuid);
1457 }
1458 
1459 static void
1460 sysctl_pollhz_handler(netmsg_t nmsg)
1461 {
1462 	struct poll_comm *comm = poll_common[mycpuid];
1463 
1464 	KKASSERT(&curthread->td_msgport == netisr_cpuport(comm->poll_cpuid));
1465 
1466 	/* Save polling frequency */
1467 	comm->pollhz = poll_comm_pollhz_div(comm, nmsg->lmsg.u.ms_result);
1468 
1469 	/*
1470 	 * Adjust cached pollhz
1471 	 */
1472 	rxpoll_context[mycpuid]->pollhz = comm->pollhz;
1473 	txpoll_context[mycpuid]->pollhz =
1474 	    comm->pollhz / (comm->poll_txfrac + 1);
1475 
1476 	/*
1477 	 * Adjust polling frequency
1478 	 */
1479 	poll_comm_adjust_pollhz(comm);
1480 
1481 	netisr_replymsg(&nmsg->base, 0);
1482 }
1483 
1484 static int
1485 sysctl_stfrac(SYSCTL_HANDLER_ARGS)
1486 {
1487 	struct poll_comm *comm = arg1;
1488 	struct netmsg_base nmsg;
1489 	int error, stfrac;
1490 
1491 	KKASSERT(comm->poll_cpuid == 0);
1492 
1493 	stfrac = comm->poll_stfrac + 1;
1494 	error = sysctl_handle_int(oidp, &stfrac, 0, req);
1495 	if (error || req->newptr == NULL)
1496 		return error;
1497 	if (stfrac < 1)
1498 		return EINVAL;
1499 
1500 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1501 		    0, sysctl_stfrac_handler);
1502 	nmsg.lmsg.u.ms_result = stfrac - 1;
1503 
1504 	return netisr_domsg(&nmsg, comm->poll_cpuid);
1505 }
1506 
1507 static void
1508 sysctl_stfrac_handler(netmsg_t nmsg)
1509 {
1510 	struct poll_comm *comm = poll_common[mycpuid];
1511 	int stfrac = nmsg->lmsg.u.ms_result;
1512 
1513 	KKASSERT(&curthread->td_msgport == netisr_cpuport(comm->poll_cpuid));
1514 
1515 	crit_enter();
1516 	comm->poll_stfrac = stfrac;
1517 	if (comm->stfrac_count > comm->poll_stfrac)
1518 		comm->stfrac_count = comm->poll_stfrac;
1519 	crit_exit();
1520 
1521 	netisr_replymsg(&nmsg->base, 0);
1522 }
1523 
1524 static int
1525 sysctl_txfrac(SYSCTL_HANDLER_ARGS)
1526 {
1527 	struct poll_comm *comm = arg1;
1528 	struct netmsg_base nmsg;
1529 	int error, txfrac;
1530 
1531 	txfrac = comm->poll_txfrac + 1;
1532 	error = sysctl_handle_int(oidp, &txfrac, 0, req);
1533 	if (error || req->newptr == NULL)
1534 		return error;
1535 	if (txfrac < 1)
1536 		return EINVAL;
1537 
1538 	netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1539 		    0, sysctl_txfrac_handler);
1540 	nmsg.lmsg.u.ms_result = txfrac - 1;
1541 
1542 	return netisr_domsg(&nmsg, comm->poll_cpuid);
1543 }
1544 
1545 static void
1546 sysctl_txfrac_handler(netmsg_t nmsg)
1547 {
1548 	struct poll_comm *comm = poll_common[mycpuid];
1549 	int txfrac = nmsg->lmsg.u.ms_result;
1550 
1551 	KKASSERT(&curthread->td_msgport == netisr_cpuport(comm->poll_cpuid));
1552 
1553 	crit_enter();
1554 	comm->poll_txfrac = txfrac;
1555 	if (comm->txfrac_count > comm->poll_txfrac)
1556 		comm->txfrac_count = comm->poll_txfrac;
1557 	crit_exit();
1558 
1559 	netisr_replymsg(&nmsg->base, 0);
1560 }
1561 
1562 void
1563 ifpoll_compat_setup(struct ifpoll_compat *cp,
1564     struct sysctl_ctx_list *sysctl_ctx,
1565     struct sysctl_oid *sysctl_tree,
1566     int unit, struct lwkt_serialize *slz)
1567 {
1568 	cp->ifpc_stcount = 0;
1569 	cp->ifpc_stfrac = ((poll_common[0]->poll_stfrac + 1) *
1570 	    howmany(IOPOLL_BURST_MAX, IOPOLL_EACH_BURST)) - 1;
1571 
1572 	cp->ifpc_cpuid = unit % netisr_ncpus;
1573 	cp->ifpc_serializer = slz;
1574 
1575 	if (sysctl_ctx != NULL && sysctl_tree != NULL) {
1576 		SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1577 		    OID_AUTO, "npoll_stfrac", CTLTYPE_INT | CTLFLAG_RW,
1578 		    cp, 0, sysctl_compat_npoll_stfrac, "I",
1579 		    "polling status frac");
1580 		SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1581 		    OID_AUTO, "npoll_cpuid", CTLTYPE_INT | CTLFLAG_RW,
1582 		    cp, 0, sysctl_compat_npoll_cpuid, "I",
1583 		    "polling cpuid");
1584 	}
1585 }
1586 
1587 static int
1588 sysctl_compat_npoll_stfrac(SYSCTL_HANDLER_ARGS)
1589 {
1590 	struct ifpoll_compat *cp = arg1;
1591 	int error = 0, stfrac;
1592 
1593 	lwkt_serialize_enter(cp->ifpc_serializer);
1594 
1595 	stfrac = cp->ifpc_stfrac + 1;
1596 	error = sysctl_handle_int(oidp, &stfrac, 0, req);
1597 	if (!error && req->newptr != NULL) {
1598 		if (stfrac < 1) {
1599 			error = EINVAL;
1600 		} else {
1601 			cp->ifpc_stfrac = stfrac - 1;
1602 			if (cp->ifpc_stcount > cp->ifpc_stfrac)
1603 				cp->ifpc_stcount = cp->ifpc_stfrac;
1604 		}
1605 	}
1606 
1607 	lwkt_serialize_exit(cp->ifpc_serializer);
1608 	return error;
1609 }
1610 
1611 static int
1612 sysctl_compat_npoll_cpuid(SYSCTL_HANDLER_ARGS)
1613 {
1614 	struct ifpoll_compat *cp = arg1;
1615 	int error = 0, cpuid;
1616 
1617 	lwkt_serialize_enter(cp->ifpc_serializer);
1618 
1619 	cpuid = cp->ifpc_cpuid;
1620 	error = sysctl_handle_int(oidp, &cpuid, 0, req);
1621 	if (!error && req->newptr != NULL) {
1622 		if (cpuid < 0 || cpuid >= netisr_ncpus)
1623 			error = EINVAL;
1624 		else
1625 			cp->ifpc_cpuid = cpuid;
1626 	}
1627 
1628 	lwkt_serialize_exit(cp->ifpc_serializer);
1629 	return error;
1630 }
1631