2 * Copyright (c) 2001-2002 Luigi Rizzo
4 * Supported by: the Xorp Project (www.xorp.org)
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
27 * $FreeBSD: src/sys/kern/kern_poll.c,v 1.2.2.4 2002/06/27 23:26:33 luigi Exp $
28 * $DragonFly: src/sys/kern/kern_poll.c,v 1.40 2007/10/01 10:03:49 sephe Exp $
31 #include "opt_polling.h"
33 #include <sys/param.h>
34 #include <sys/kernel.h>
35 #include <sys/socket.h> /* needed by net/if.h */
36 #include <sys/sysctl.h>
38 #include <sys/thread2.h>
39 #include <sys/msgport2.h>
41 #include <net/if.h> /* for IFF_* flags */
42 #include <net/netmsg2.h>
45 * Polling support for [network] device drivers.
47 * Drivers which support this feature try to register with the
50 * If registration is successful, the driver must disable interrupts,
51 * and further I/O is performed through the handler, which is invoked
52 * (at least once per clock tick) with 3 arguments: the "arg" passed at
53 * register time (a struct ifnet pointer), a command, and a "count" limit.
55 * The command can be one of the following:
56 * POLL_ONLY: quick move of "count" packets from input/output queues.
57 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
58 * other more expensive operations. This command is issued periodically
59 * but less frequently than POLL_ONLY.
60 * POLL_DEREGISTER: deregister and return to interrupt mode.
61 * POLL_REGISTER: register and disable interrupts
63 * The first two commands are only issued if the interface is marked as
64 * 'IFF_UP, IFF_RUNNING and IFF_POLLING', the last two only if IFF_RUNNING
67 * The count limit specifies how much work the handler can do during the
68 * call -- typically this is the number of packets to be received, or
69 * transmitted, etc. (drivers are free to interpret this number, as long
70 * as the max time spent in the function grows roughly linearly with the
73 * Deregistration can be requested by the driver itself (typically in the
74 * *_stop() routine), or by the polling code, by invoking the handler.
76 * Polling can be enabled or disabled on particular CPU_X with the sysctl
77 * variable kern.polling.X.enable (default is 1, enabled)
79 * A second variable controls the sharing of CPU between polling/kernel
80 * network processing, and other activities (typically userlevel tasks):
81 * kern.polling.X.user_frac (between 0 and 100, default 50) sets the share
82 * of CPU allocated to user tasks. CPU is allocated proportionally to the
83 * shares, by dynamically adjusting the "count" (poll_burst).
85 * Other parameters can should be left to their default values.
86 * The following constraints hold
88 * 1 <= poll_burst <= poll_burst_max
89 * 1 <= poll_each_burst <= poll_burst_max
90 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
93 #define MIN_POLL_BURST_MAX 10
94 #define MAX_POLL_BURST_MAX 1000
96 #ifndef DEVICE_POLLING_FREQ_MAX
97 #define DEVICE_POLLING_FREQ_MAX 30000
99 #define DEVICE_POLLING_FREQ_DEFAULT 2000
101 #define POLL_LIST_LEN 128
106 #define POLLCTX_MAX 32
109 struct sysctl_ctx_list poll_sysctl_ctx;
110 struct sysctl_oid *poll_sysctl_tree;
113 uint32_t poll_each_burst;
114 uint32_t poll_burst_max;
118 uint32_t short_ticks;
120 uint32_t pending_polls;
125 struct timeval poll_start_t;
126 struct timeval prev_t;
128 uint32_t poll_handlers; /* next free entry in pr[]. */
129 struct pollrec pr[POLL_LIST_LEN];
132 struct systimer pollclock;
136 struct netmsg poll_netmsg;
137 struct netmsg poll_more_netmsg;
140 static struct pollctx *poll_context[POLLCTX_MAX];
142 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
143 "Device polling parameters");
145 static int poll_defcpu = -1;
146 SYSCTL_INT(_kern_polling, OID_AUTO, defcpu, CTLFLAG_RD,
147 &poll_defcpu, 0, "default CPU to run device polling");
149 static uint32_t poll_cpumask0 = 0xffffffff;
150 TUNABLE_INT("kern.polling.cpumask", (int *)&poll_cpumask0);
152 static uint32_t poll_cpumask;
153 SYSCTL_INT(_kern_polling, OID_AUTO, cpumask, CTLFLAG_RD,
154 &poll_cpumask, 0, "CPUs that can run device polling");
156 static int polling_enabled = 1; /* global polling enable */
157 TUNABLE_INT("kern.polling.enable", &polling_enabled);
159 static int pollhz = DEVICE_POLLING_FREQ_DEFAULT;
160 TUNABLE_INT("kern.polling.pollhz", &pollhz);
162 /* Netisr handlers */
163 static void netisr_poll(struct netmsg *);
164 static void netisr_pollmore(struct netmsg *);
165 static void poll_register(struct netmsg *);
166 static void poll_deregister(struct netmsg *);
167 static void poll_sysctl_pollhz(struct netmsg *);
168 static void poll_sysctl_polling(struct netmsg *);
169 static void poll_sysctl_regfrac(struct netmsg *);
170 static void poll_sysctl_burstmax(struct netmsg *);
171 static void poll_sysctl_eachburst(struct netmsg *);
173 /* Systimer handler */
174 static void pollclock(systimer_t, struct intrframe *);
176 /* Sysctl handlers */
177 static int sysctl_pollhz(SYSCTL_HANDLER_ARGS);
178 static int sysctl_polling(SYSCTL_HANDLER_ARGS);
179 static int sysctl_regfrac(SYSCTL_HANDLER_ARGS);
180 static int sysctl_burstmax(SYSCTL_HANDLER_ARGS);
181 static int sysctl_eachburst(SYSCTL_HANDLER_ARGS);
182 static void poll_add_sysctl(struct sysctl_ctx_list *,
183 struct sysctl_oid_list *, struct pollctx *);
185 static void schedpoll_oncpu(struct pollctx *, struct netmsg *, netisr_fn_t);
187 void init_device_poll_pcpu(int); /* per-cpu init routine */
190 * Initialize per-cpu polling(4) context. Called from kern_clock.c:
193 init_device_poll_pcpu(int cpuid)
195 struct pollctx *pctx;
198 if (cpuid >= POLLCTX_MAX)
201 if (((1 << cpuid) & poll_cpumask0) == 0)
204 poll_cpumask |= (1 << cpuid);
206 pctx = kmalloc(sizeof(*pctx), M_DEVBUF, M_WAITOK | M_ZERO);
208 pctx->poll_burst = 5;
209 pctx->poll_each_burst = 5;
210 pctx->poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
211 pctx->user_frac = 50;
213 pctx->polling_enabled = polling_enabled;
214 pctx->pollhz = pollhz;
215 pctx->poll_cpuid = cpuid;
216 netmsg_init(&pctx->poll_netmsg, &netisr_adone_rport, 0, NULL);
217 netmsg_init(&pctx->poll_more_netmsg, &netisr_adone_rport, 0, NULL);
219 KASSERT(cpuid < POLLCTX_MAX, ("cpu id must < %d", cpuid));
220 poll_context[cpuid] = pctx;
222 if (poll_defcpu < 0) {
226 * Initialize global sysctl nodes, for compat
228 poll_add_sysctl(NULL, SYSCTL_STATIC_CHILDREN(_kern_polling),
233 * Initialize per-cpu sysctl nodes
235 ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", pctx->poll_cpuid);
237 sysctl_ctx_init(&pctx->poll_sysctl_ctx);
238 pctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&pctx->poll_sysctl_ctx,
239 SYSCTL_STATIC_CHILDREN(_kern_polling),
240 OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
241 poll_add_sysctl(&pctx->poll_sysctl_ctx,
242 SYSCTL_CHILDREN(pctx->poll_sysctl_tree), pctx);
245 * Initialize systimer
247 systimer_init_periodic_nq(&pctx->pollclock, pollclock, pctx, 1);
251 schedpoll(struct pollctx *pctx)
253 schedpoll_oncpu(pctx, &pctx->poll_netmsg, netisr_poll);
257 schedpollmore(struct pollctx *pctx)
259 schedpoll_oncpu(pctx, &pctx->poll_more_netmsg, netisr_pollmore);
263 * Set the polling frequency
266 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
268 struct pollctx *pctx = arg1;
274 error = sysctl_handle_int(oidp, &phz, 0, req);
275 if (error || req->newptr == NULL)
279 else if (phz > DEVICE_POLLING_FREQ_MAX)
280 phz = DEVICE_POLLING_FREQ_MAX;
282 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_pollhz);
283 msg.nm_lmsg.u.ms_result = phz;
285 port = cpu_portfn(pctx->poll_cpuid);
286 lwkt_domsg(port, &msg.nm_lmsg, 0);
294 sysctl_polling(SYSCTL_HANDLER_ARGS)
296 struct pollctx *pctx = arg1;
301 enabled = pctx->polling_enabled;
302 error = sysctl_handle_int(oidp, &enabled, 0, req);
303 if (error || req->newptr == NULL)
306 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_polling);
307 msg.nm_lmsg.u.ms_result = enabled;
309 port = cpu_portfn(pctx->poll_cpuid);
310 lwkt_domsg(port, &msg.nm_lmsg, 0);
315 sysctl_regfrac(SYSCTL_HANDLER_ARGS)
317 struct pollctx *pctx = arg1;
323 reg_frac = pctx->reg_frac;
324 error = sysctl_handle_int(oidp, ®_frac, 0, req);
325 if (error || req->newptr == NULL)
328 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_regfrac);
329 msg.nm_lmsg.u.ms_result = reg_frac;
331 port = cpu_portfn(pctx->poll_cpuid);
332 lwkt_domsg(port, &msg.nm_lmsg, 0);
337 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
339 struct pollctx *pctx = arg1;
345 burst_max = pctx->poll_burst_max;
346 error = sysctl_handle_int(oidp, &burst_max, 0, req);
347 if (error || req->newptr == NULL)
349 if (burst_max < MIN_POLL_BURST_MAX)
350 burst_max = MIN_POLL_BURST_MAX;
351 else if (burst_max > MAX_POLL_BURST_MAX)
352 burst_max = MAX_POLL_BURST_MAX;
354 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_burstmax);
355 msg.nm_lmsg.u.ms_result = burst_max;
357 port = cpu_portfn(pctx->poll_cpuid);
358 lwkt_domsg(port, &msg.nm_lmsg, 0);
363 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
365 struct pollctx *pctx = arg1;
371 each_burst = pctx->poll_each_burst;
372 error = sysctl_handle_int(oidp, &each_burst, 0, req);
373 if (error || req->newptr == NULL)
376 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_eachburst);
377 msg.nm_lmsg.u.ms_result = each_burst;
379 port = cpu_portfn(pctx->poll_cpuid);
380 lwkt_domsg(port, &msg.nm_lmsg, 0);
385 * Hook from polling systimer. Tries to schedule a netisr, but keeps
386 * track of lost ticks due to the previous handler taking too long.
387 * Normally, this should not happen, because polling handler should
388 * run for a short time. However, in some cases (e.g. when there are
389 * changes in link status etc.) the drivers take a very long time
390 * (even in the order of milliseconds) to reset and reconfigure the
391 * device, causing apparent lost polls.
393 * The first part of the code is just for debugging purposes, and tries
394 * to count how often hardclock ticks are shorter than they should,
395 * meaning either stray interrupts or delayed events.
397 * WARNING! called from fastint or IPI, the MP lock might not be held.
400 pollclock(systimer_t info, struct intrframe *frame __unused)
402 struct pollctx *pctx = info->data;
406 if (pctx->poll_handlers == 0)
410 delta = (t.tv_usec - pctx->prev_t.tv_usec) +
411 (t.tv_sec - pctx->prev_t.tv_sec)*1000000;
412 if (delta * pctx->pollhz < 500000)
417 if (pctx->pending_polls > 100) {
419 * Too much, assume it has stalled (not always true
420 * see comment above).
423 pctx->pending_polls = 0;
427 if (pctx->phase <= 2) {
428 if (pctx->phase != 0)
434 if (pctx->pending_polls++ > 0)
439 * netisr_pollmore is called after other netisr's, possibly scheduling
440 * another NETISR_POLL call, or adapting the burst size for the next cycle.
442 * It is very bad to fetch large bursts of packets from a single card at once,
443 * because the burst could take a long time to be completely processed, or
444 * could saturate the intermediate queue (ipintrq or similar) leading to
445 * losses or unfairness. To reduce the problem, and also to account better for
446 * time spent in network-related processing, we split the burst in smaller
447 * chunks of fixed size, giving control to the other netisr's between chunks.
448 * This helps in improving the fairness, reducing livelock (because we
449 * emulate more closely the "process to completion" that we have with
450 * fastforwarding) and accounting for the work performed in low level
451 * handling and forwarding.
456 netisr_pollmore(struct netmsg *msg)
458 struct pollctx *pctx;
460 int kern_load, cpuid;
462 cpuid = mycpu->gd_cpuid;
463 KKASSERT(cpuid < POLLCTX_MAX);
465 pctx = poll_context[cpuid];
466 KKASSERT(pctx != NULL);
467 KKASSERT(pctx->poll_cpuid == cpuid);
468 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
470 lwkt_replymsg(&msg->nm_lmsg, 0);
473 if (pctx->residual_burst > 0) {
475 /* will run immediately on return, followed by netisrs */
478 /* here we can account time spent in netisr's in this tick */
480 kern_load = (t.tv_usec - pctx->poll_start_t.tv_usec) +
481 (t.tv_sec - pctx->poll_start_t.tv_sec)*1000000; /* us */
482 kern_load = (kern_load * pctx->pollhz) / 10000; /* 0..100 */
483 if (kern_load > (100 - pctx->user_frac)) { /* try decrease ticks */
484 if (pctx->poll_burst > 1)
487 if (pctx->poll_burst < pctx->poll_burst_max)
491 pctx->pending_polls--;
492 if (pctx->pending_polls == 0) { /* we are done */
496 * Last cycle was long and caused us to miss one or more
497 * hardclock ticks. Restart processing again, but slightly
498 * reduce the burst size to prevent that this happens again.
500 pctx->poll_burst -= (pctx->poll_burst / 8);
501 if (pctx->poll_burst < 1)
502 pctx->poll_burst = 1;
509 * netisr_poll is scheduled by schedpoll when appropriate, typically once
510 * per polling systimer tick.
512 * Note that the message is replied immediately in order to allow a new
513 * ISR to be scheduled in the handler.
515 * XXX each registration should indicate whether it needs a critical
516 * section to operate.
520 netisr_poll(struct netmsg *msg)
522 struct pollctx *pctx;
523 int i, cycles, cpuid;
524 enum poll_cmd arg = POLL_ONLY;
526 cpuid = mycpu->gd_cpuid;
527 KKASSERT(cpuid < POLLCTX_MAX);
529 pctx = poll_context[cpuid];
530 KKASSERT(pctx != NULL);
531 KKASSERT(pctx->poll_cpuid == cpuid);
532 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
534 lwkt_replymsg(&msg->nm_lmsg, 0);
537 if (pctx->residual_burst == 0) { /* first call in this tick */
538 microuptime(&pctx->poll_start_t);
540 if (pctx->reg_frac_count-- == 0) {
541 arg = POLL_AND_CHECK_STATUS;
542 pctx->reg_frac_count = pctx->reg_frac - 1;
545 pctx->residual_burst = pctx->poll_burst;
547 cycles = (pctx->residual_burst < pctx->poll_each_burst) ?
548 pctx->residual_burst : pctx->poll_each_burst;
549 pctx->residual_burst -= cycles;
551 if (pctx->polling_enabled) {
552 for (i = 0 ; i < pctx->poll_handlers ; i++) {
553 struct ifnet *ifp = pctx->pr[i].ifp;
555 if (!lwkt_serialize_try(ifp->if_serializer))
558 if ((ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING))
559 == (IFF_UP|IFF_RUNNING|IFF_POLLING))
560 ifp->if_poll(ifp, arg, cycles);
562 lwkt_serialize_exit(ifp->if_serializer);
564 } else { /* unregister */
565 for (i = 0 ; i < pctx->poll_handlers ; i++) {
566 struct ifnet *ifp = pctx->pr[i].ifp;
568 lwkt_serialize_enter(ifp->if_serializer);
570 if ((ifp->if_flags & IFF_POLLING) == 0) {
571 KKASSERT(ifp->if_poll_cpuid < 0);
572 lwkt_serialize_exit(ifp->if_serializer);
575 ifp->if_flags &= ~IFF_POLLING;
576 ifp->if_poll_cpuid = -1;
579 * Only call the interface deregistration
580 * function if the interface is still
583 if (ifp->if_flags & IFF_RUNNING)
584 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
586 lwkt_serialize_exit(ifp->if_serializer);
588 pctx->residual_burst = 0;
589 pctx->poll_handlers = 0;
596 poll_register(struct netmsg *msg)
598 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
599 struct pollctx *pctx;
602 cpuid = mycpu->gd_cpuid;
603 KKASSERT(cpuid < POLLCTX_MAX);
605 pctx = poll_context[cpuid];
606 KKASSERT(pctx != NULL);
607 KKASSERT(pctx->poll_cpuid == cpuid);
609 if (pctx->polling_enabled == 0) {
610 /* Polling disabled, cannot register */
616 * Check if there is room.
618 if (pctx->poll_handlers >= POLL_LIST_LEN) {
620 * List full, cannot register more entries.
621 * This should never happen; if it does, it is probably a
622 * broken driver trying to register multiple times. Checking
623 * this at runtime is expensive, and won't solve the problem
624 * anyways, so just report a few times and then give up.
626 static int verbose = 10; /* XXX */
628 kprintf("poll handlers list full, "
629 "maybe a broken driver ?\n");
634 pctx->pr[pctx->poll_handlers].ifp = ifp;
635 pctx->poll_handlers++;
638 if (pctx->poll_handlers == 1) {
639 KKASSERT(pctx->polling_enabled);
640 systimer_adjust_periodic(&pctx->pollclock,
645 lwkt_replymsg(&msg->nm_lmsg, rc);
649 * Try to register routine for polling. Returns 1 if successful
650 * (and polling should be enabled), 0 otherwise.
652 * Called from mainline code only, not called from an interrupt.
655 ether_poll_register(struct ifnet *ifp)
659 KKASSERT(poll_defcpu < POLLCTX_MAX);
661 return ether_pollcpu_register(ifp, poll_defcpu);
665 ether_pollcpu_register(struct ifnet *ifp, int cpuid)
671 if (ifp->if_poll == NULL) {
672 /* Device does not support polling */
676 if (cpuid < 0 || cpuid >= POLLCTX_MAX)
679 if (((1 << cpuid) & poll_cpumask) == 0) {
680 /* Polling is not supported on 'cpuid' */
683 KKASSERT(poll_context[cpuid] != NULL);
686 * Attempt to register. Interlock with IFF_POLLING.
688 crit_enter(); /* XXX MP - not mp safe */
690 lwkt_serialize_enter(ifp->if_serializer);
691 if (ifp->if_flags & IFF_POLLING) {
692 /* Already polling */
693 KKASSERT(ifp->if_poll_cpuid >= 0);
694 lwkt_serialize_exit(ifp->if_serializer);
698 KKASSERT(ifp->if_poll_cpuid < 0);
699 ifp->if_flags |= IFF_POLLING;
700 ifp->if_poll_cpuid = cpuid;
701 if (ifp->if_flags & IFF_RUNNING)
702 ifp->if_poll(ifp, POLL_REGISTER, 0);
703 lwkt_serialize_exit(ifp->if_serializer);
705 netmsg_init(&msg, &curthread->td_msgport, 0, poll_register);
706 msg.nm_lmsg.u.ms_resultp = ifp;
708 port = cpu_portfn(cpuid);
709 lwkt_domsg(port, &msg.nm_lmsg, 0);
711 if (msg.nm_lmsg.ms_error) {
712 lwkt_serialize_enter(ifp->if_serializer);
713 ifp->if_flags &= ~IFF_POLLING;
714 ifp->if_poll_cpuid = -1;
715 if (ifp->if_flags & IFF_RUNNING)
716 ifp->if_poll(ifp, POLL_DEREGISTER, 0);
717 lwkt_serialize_exit(ifp->if_serializer);
728 poll_deregister(struct netmsg *msg)
730 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
731 struct pollctx *pctx;
734 cpuid = mycpu->gd_cpuid;
735 KKASSERT(cpuid < POLLCTX_MAX);
737 pctx = poll_context[cpuid];
738 KKASSERT(pctx != NULL);
739 KKASSERT(pctx->poll_cpuid == cpuid);
741 for (i = 0 ; i < pctx->poll_handlers ; i++) {
742 if (pctx->pr[i].ifp == ifp) /* Found it */
745 if (i == pctx->poll_handlers) {
746 kprintf("ether_poll_deregister: ifp not found!!!\n");
749 pctx->poll_handlers--;
750 if (i < pctx->poll_handlers) {
751 /* Last entry replaces this one. */
752 pctx->pr[i].ifp = pctx->pr[pctx->poll_handlers].ifp;
755 if (pctx->poll_handlers == 0)
756 systimer_adjust_periodic(&pctx->pollclock, 1);
759 lwkt_replymsg(&msg->nm_lmsg, rc);
763 * Remove interface from the polling list. Occurs when polling is turned
764 * off. Called from mainline code only, not called from an interrupt.
767 ether_poll_deregister(struct ifnet *ifp)
773 KKASSERT(ifp != NULL);
775 if (ifp->if_poll == NULL)
780 lwkt_serialize_enter(ifp->if_serializer);
781 if ((ifp->if_flags & IFF_POLLING) == 0) {
782 KKASSERT(ifp->if_poll_cpuid < 0);
783 lwkt_serialize_exit(ifp->if_serializer);
788 cpuid = ifp->if_poll_cpuid;
789 KKASSERT(cpuid >= 0);
790 KKASSERT(poll_context[cpuid] != NULL);
792 ifp->if_flags &= ~IFF_POLLING;
793 ifp->if_poll_cpuid = -1;
794 lwkt_serialize_exit(ifp->if_serializer);
796 netmsg_init(&msg, &curthread->td_msgport, 0, poll_deregister);
797 msg.nm_lmsg.u.ms_resultp = ifp;
799 port = cpu_portfn(cpuid);
800 lwkt_domsg(port, &msg.nm_lmsg, 0);
802 if (!msg.nm_lmsg.ms_error) {
803 lwkt_serialize_enter(ifp->if_serializer);
804 if (ifp->if_flags & IFF_RUNNING)
805 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
806 lwkt_serialize_exit(ifp->if_serializer);
817 poll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
818 struct pollctx *pctx)
820 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "enable",
821 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_polling,
822 "I", "Polling enabled");
824 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "pollhz",
825 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_pollhz,
826 "I", "Device polling frequency");
828 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "reg_frac",
829 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_regfrac,
830 "IU", "Every this many cycles poll register");
832 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
833 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_burstmax,
834 "IU", "Max Polling burst size");
836 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
837 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_eachburst,
838 "IU", "Max size of each burst");
840 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
841 &pctx->phase, 0, "Polling phase");
843 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
844 &pctx->suspect, 0, "suspect event");
846 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
847 &pctx->stalled, 0, "potential stalls");
849 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
850 &pctx->poll_burst, 0, "Current polling burst size");
852 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
854 "Desired user fraction of cpu time");
856 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
857 &pctx->short_ticks, 0,
858 "Hardclock ticks shorter than they should be");
860 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
861 &pctx->lost_polls, 0,
862 "How many times we would have lost a poll tick");
864 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
865 &pctx->pending_polls, 0, "Do we need to poll again");
867 SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
868 &pctx->residual_burst, 0,
869 "# of residual cycles in burst");
871 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
872 &pctx->poll_handlers, 0,
873 "Number of registered poll handlers");
877 schedpoll_oncpu(struct pollctx *pctx, struct netmsg *msg, netisr_fn_t handler)
879 if (msg->nm_lmsg.ms_flags & MSGF_DONE) {
882 netmsg_init(msg, &netisr_adone_rport, 0, handler);
884 msg->nm_lmsg.u.ms_resultp = pctx;
886 port = cpu_portfn(mycpu->gd_cpuid);
887 lwkt_sendmsg(port, &msg->nm_lmsg);
892 poll_sysctl_pollhz(struct netmsg *msg)
894 struct pollctx *pctx;
897 cpuid = mycpu->gd_cpuid;
898 KKASSERT(cpuid < POLLCTX_MAX);
900 pctx = poll_context[cpuid];
901 KKASSERT(pctx != NULL);
902 KKASSERT(pctx->poll_cpuid == cpuid);
905 * If polling is disabled or there is no device registered,
906 * don't adjust polling systimer frequency.
907 * Polling systimer frequency will be adjusted once polling
908 * is enabled and there are registered devices.
910 pctx->pollhz = msg->nm_lmsg.u.ms_result;
911 if (pctx->polling_enabled && pctx->poll_handlers)
912 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
915 * Make sure that reg_frac and reg_frac_count are within valid range.
917 if (pctx->reg_frac > pctx->pollhz) {
918 pctx->reg_frac = pctx->pollhz;
919 if (pctx->reg_frac_count > pctx->reg_frac)
920 pctx->reg_frac_count = pctx->reg_frac - 1;
923 lwkt_replymsg(&msg->nm_lmsg, 0);
927 poll_sysctl_polling(struct netmsg *msg)
929 struct pollctx *pctx;
932 cpuid = mycpu->gd_cpuid;
933 KKASSERT(cpuid < POLLCTX_MAX);
935 pctx = poll_context[cpuid];
936 KKASSERT(pctx != NULL);
937 KKASSERT(pctx->poll_cpuid == cpuid);
940 * If polling is disabled or there is no device registered,
941 * cut the polling systimer frequency to 1hz.
943 pctx->polling_enabled = msg->nm_lmsg.u.ms_result;
944 if (pctx->polling_enabled && pctx->poll_handlers)
945 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
947 systimer_adjust_periodic(&pctx->pollclock, 1);
948 lwkt_replymsg(&msg->nm_lmsg, 0);
952 poll_sysctl_regfrac(struct netmsg *msg)
954 struct pollctx *pctx;
958 cpuid = mycpu->gd_cpuid;
959 KKASSERT(cpuid < POLLCTX_MAX);
961 pctx = poll_context[cpuid];
962 KKASSERT(pctx != NULL);
963 KKASSERT(pctx->poll_cpuid == cpuid);
965 reg_frac = msg->nm_lmsg.u.ms_result;
966 if (reg_frac > pctx->pollhz)
967 reg_frac = pctx->pollhz;
968 else if (reg_frac < 1)
971 pctx->reg_frac = reg_frac;
972 if (pctx->reg_frac_count > pctx->reg_frac)
973 pctx->reg_frac_count = pctx->reg_frac - 1;
975 lwkt_replymsg(&msg->nm_lmsg, 0);
979 poll_sysctl_burstmax(struct netmsg *msg)
981 struct pollctx *pctx;
984 cpuid = mycpu->gd_cpuid;
985 KKASSERT(cpuid < POLLCTX_MAX);
987 pctx = poll_context[cpuid];
988 KKASSERT(pctx != NULL);
989 KKASSERT(pctx->poll_cpuid == cpuid);
991 pctx->poll_burst_max = msg->nm_lmsg.u.ms_result;
992 if (pctx->poll_each_burst > pctx->poll_burst_max)
993 pctx->poll_each_burst = pctx->poll_burst_max;
994 if (pctx->poll_burst > pctx->poll_burst_max)
995 pctx->poll_burst = pctx->poll_burst_max;
996 if (pctx->residual_burst > pctx->poll_burst_max)
997 pctx->residual_burst = pctx->poll_burst_max;
999 lwkt_replymsg(&msg->nm_lmsg, 0);
1003 poll_sysctl_eachburst(struct netmsg *msg)
1005 struct pollctx *pctx;
1006 uint32_t each_burst;
1009 cpuid = mycpu->gd_cpuid;
1010 KKASSERT(cpuid < POLLCTX_MAX);
1012 pctx = poll_context[cpuid];
1013 KKASSERT(pctx != NULL);
1014 KKASSERT(pctx->poll_cpuid == cpuid);
1016 each_burst = msg->nm_lmsg.u.ms_result;
1017 if (each_burst > pctx->poll_burst_max)
1018 each_burst = pctx->poll_burst_max;
1019 else if (each_burst < 1)
1021 pctx->poll_each_burst = each_burst;
1023 lwkt_replymsg(&msg->nm_lmsg, 0);