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 1.39 2007/10/01 09:37:30 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 *);
172 /* Systimer handler */
173 static void pollclock(systimer_t, struct intrframe *);
175 /* Sysctl handlers */
176 static int sysctl_pollhz(SYSCTL_HANDLER_ARGS);
177 static int sysctl_polling(SYSCTL_HANDLER_ARGS);
178 static int sysctl_regfrac(SYSCTL_HANDLER_ARGS);
179 static int sysctl_burstmax(SYSCTL_HANDLER_ARGS);
180 static void poll_add_sysctl(struct sysctl_ctx_list *,
181 struct sysctl_oid_list *, struct pollctx *);
183 static void schedpoll_oncpu(struct pollctx *, struct netmsg *, netisr_fn_t);
185 void init_device_poll_pcpu(int); /* per-cpu init routine */
188 * Initialize per-cpu polling(4) context. Called from kern_clock.c:
191 init_device_poll_pcpu(int cpuid)
193 struct pollctx *pctx;
196 if (cpuid >= POLLCTX_MAX)
199 if (((1 << cpuid) & poll_cpumask0) == 0)
202 poll_cpumask |= (1 << cpuid);
204 pctx = kmalloc(sizeof(*pctx), M_DEVBUF, M_WAITOK | M_ZERO);
206 pctx->poll_burst = 5;
207 pctx->poll_each_burst = 5;
208 pctx->poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
209 pctx->user_frac = 50;
211 pctx->polling_enabled = polling_enabled;
212 pctx->pollhz = pollhz;
213 pctx->poll_cpuid = cpuid;
214 netmsg_init(&pctx->poll_netmsg, &netisr_adone_rport, 0, NULL);
215 netmsg_init(&pctx->poll_more_netmsg, &netisr_adone_rport, 0, NULL);
217 KASSERT(cpuid < POLLCTX_MAX, ("cpu id must < %d", cpuid));
218 poll_context[cpuid] = pctx;
220 if (poll_defcpu < 0) {
224 * Initialize global sysctl nodes, for compat
226 poll_add_sysctl(NULL, SYSCTL_STATIC_CHILDREN(_kern_polling),
231 * Initialize per-cpu sysctl nodes
233 ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", pctx->poll_cpuid);
235 sysctl_ctx_init(&pctx->poll_sysctl_ctx);
236 pctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&pctx->poll_sysctl_ctx,
237 SYSCTL_STATIC_CHILDREN(_kern_polling),
238 OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
239 poll_add_sysctl(&pctx->poll_sysctl_ctx,
240 SYSCTL_CHILDREN(pctx->poll_sysctl_tree), pctx);
243 * Initialize systimer
245 systimer_init_periodic_nq(&pctx->pollclock, pollclock, pctx, 1);
249 schedpoll(struct pollctx *pctx)
251 schedpoll_oncpu(pctx, &pctx->poll_netmsg, netisr_poll);
255 schedpollmore(struct pollctx *pctx)
257 schedpoll_oncpu(pctx, &pctx->poll_more_netmsg, netisr_pollmore);
261 * Set the polling frequency
264 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
266 struct pollctx *pctx = arg1;
272 error = sysctl_handle_int(oidp, &phz, 0, req);
273 if (error || req->newptr == NULL)
277 else if (phz > DEVICE_POLLING_FREQ_MAX)
278 phz = DEVICE_POLLING_FREQ_MAX;
280 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_pollhz);
281 msg.nm_lmsg.u.ms_result = phz;
283 port = cpu_portfn(pctx->poll_cpuid);
284 lwkt_domsg(port, &msg.nm_lmsg, 0);
292 sysctl_polling(SYSCTL_HANDLER_ARGS)
294 struct pollctx *pctx = arg1;
299 enabled = pctx->polling_enabled;
300 error = sysctl_handle_int(oidp, &enabled, 0, req);
301 if (error || req->newptr == NULL)
304 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_polling);
305 msg.nm_lmsg.u.ms_result = enabled;
307 port = cpu_portfn(pctx->poll_cpuid);
308 lwkt_domsg(port, &msg.nm_lmsg, 0);
313 sysctl_regfrac(SYSCTL_HANDLER_ARGS)
315 struct pollctx *pctx = arg1;
321 reg_frac = pctx->reg_frac;
322 error = sysctl_handle_int(oidp, ®_frac, 0, req);
323 if (error || req->newptr == NULL)
326 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_regfrac);
327 msg.nm_lmsg.u.ms_result = reg_frac;
329 port = cpu_portfn(pctx->poll_cpuid);
330 lwkt_domsg(port, &msg.nm_lmsg, 0);
335 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
337 struct pollctx *pctx = arg1;
343 burst_max = pctx->poll_burst_max;
344 error = sysctl_handle_int(oidp, &burst_max, 0, req);
345 if (error || req->newptr == NULL)
347 if (burst_max < MIN_POLL_BURST_MAX)
348 burst_max = MIN_POLL_BURST_MAX;
349 else if (burst_max > MAX_POLL_BURST_MAX)
350 burst_max = MAX_POLL_BURST_MAX;
352 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_burstmax);
353 msg.nm_lmsg.u.ms_result = burst_max;
355 port = cpu_portfn(pctx->poll_cpuid);
356 lwkt_domsg(port, &msg.nm_lmsg, 0);
361 * Hook from polling systimer. Tries to schedule a netisr, but keeps
362 * track of lost ticks due to the previous handler taking too long.
363 * Normally, this should not happen, because polling handler should
364 * run for a short time. However, in some cases (e.g. when there are
365 * changes in link status etc.) the drivers take a very long time
366 * (even in the order of milliseconds) to reset and reconfigure the
367 * device, causing apparent lost polls.
369 * The first part of the code is just for debugging purposes, and tries
370 * to count how often hardclock ticks are shorter than they should,
371 * meaning either stray interrupts or delayed events.
373 * WARNING! called from fastint or IPI, the MP lock might not be held.
376 pollclock(systimer_t info, struct intrframe *frame __unused)
378 struct pollctx *pctx = info->data;
382 if (pctx->poll_handlers == 0)
386 delta = (t.tv_usec - pctx->prev_t.tv_usec) +
387 (t.tv_sec - pctx->prev_t.tv_sec)*1000000;
388 if (delta * pctx->pollhz < 500000)
393 if (pctx->pending_polls > 100) {
395 * Too much, assume it has stalled (not always true
396 * see comment above).
399 pctx->pending_polls = 0;
403 if (pctx->phase <= 2) {
404 if (pctx->phase != 0)
410 if (pctx->pending_polls++ > 0)
415 * netisr_pollmore is called after other netisr's, possibly scheduling
416 * another NETISR_POLL call, or adapting the burst size for the next cycle.
418 * It is very bad to fetch large bursts of packets from a single card at once,
419 * because the burst could take a long time to be completely processed, or
420 * could saturate the intermediate queue (ipintrq or similar) leading to
421 * losses or unfairness. To reduce the problem, and also to account better for
422 * time spent in network-related processing, we split the burst in smaller
423 * chunks of fixed size, giving control to the other netisr's between chunks.
424 * This helps in improving the fairness, reducing livelock (because we
425 * emulate more closely the "process to completion" that we have with
426 * fastforwarding) and accounting for the work performed in low level
427 * handling and forwarding.
432 netisr_pollmore(struct netmsg *msg)
434 struct pollctx *pctx;
436 int kern_load, cpuid;
438 cpuid = mycpu->gd_cpuid;
439 KKASSERT(cpuid < POLLCTX_MAX);
441 pctx = poll_context[cpuid];
442 KKASSERT(pctx != NULL);
443 KKASSERT(pctx->poll_cpuid == cpuid);
444 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
446 lwkt_replymsg(&msg->nm_lmsg, 0);
449 if (pctx->residual_burst > 0) {
451 /* will run immediately on return, followed by netisrs */
454 /* here we can account time spent in netisr's in this tick */
456 kern_load = (t.tv_usec - pctx->poll_start_t.tv_usec) +
457 (t.tv_sec - pctx->poll_start_t.tv_sec)*1000000; /* us */
458 kern_load = (kern_load * pctx->pollhz) / 10000; /* 0..100 */
459 if (kern_load > (100 - pctx->user_frac)) { /* try decrease ticks */
460 if (pctx->poll_burst > 1)
463 if (pctx->poll_burst < pctx->poll_burst_max)
467 pctx->pending_polls--;
468 if (pctx->pending_polls == 0) { /* we are done */
472 * Last cycle was long and caused us to miss one or more
473 * hardclock ticks. Restart processing again, but slightly
474 * reduce the burst size to prevent that this happens again.
476 pctx->poll_burst -= (pctx->poll_burst / 8);
477 if (pctx->poll_burst < 1)
478 pctx->poll_burst = 1;
485 * netisr_poll is scheduled by schedpoll when appropriate, typically once
486 * per polling systimer tick.
488 * Note that the message is replied immediately in order to allow a new
489 * ISR to be scheduled in the handler.
491 * XXX each registration should indicate whether it needs a critical
492 * section to operate.
496 netisr_poll(struct netmsg *msg)
498 struct pollctx *pctx;
499 int i, cycles, cpuid;
500 enum poll_cmd arg = POLL_ONLY;
502 cpuid = mycpu->gd_cpuid;
503 KKASSERT(cpuid < POLLCTX_MAX);
505 pctx = poll_context[cpuid];
506 KKASSERT(pctx != NULL);
507 KKASSERT(pctx->poll_cpuid == cpuid);
508 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
510 lwkt_replymsg(&msg->nm_lmsg, 0);
513 if (pctx->residual_burst == 0) { /* first call in this tick */
514 microuptime(&pctx->poll_start_t);
516 * Check that paremeters are consistent with runtime
517 * variables. Some of these tests could be done at sysctl
518 * time, but the savings would be very limited because we
519 * still have to check against reg_frac_count and
520 * poll_each_burst. So, instead of writing separate sysctl
521 * handlers, we do all here.
524 if (pctx->reg_frac_count-- == 0) {
525 arg = POLL_AND_CHECK_STATUS;
526 pctx->reg_frac_count = pctx->reg_frac - 1;
529 if (pctx->poll_each_burst < 1)
530 pctx->poll_each_burst = 1;
531 else if (pctx->poll_each_burst > pctx->poll_burst_max)
532 pctx->poll_each_burst = pctx->poll_burst_max;
534 pctx->residual_burst = pctx->poll_burst;
536 cycles = (pctx->residual_burst < pctx->poll_each_burst) ?
537 pctx->residual_burst : pctx->poll_each_burst;
538 pctx->residual_burst -= cycles;
540 if (pctx->polling_enabled) {
541 for (i = 0 ; i < pctx->poll_handlers ; i++) {
542 struct ifnet *ifp = pctx->pr[i].ifp;
544 if (!lwkt_serialize_try(ifp->if_serializer))
547 if ((ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING))
548 == (IFF_UP|IFF_RUNNING|IFF_POLLING))
549 ifp->if_poll(ifp, arg, cycles);
551 lwkt_serialize_exit(ifp->if_serializer);
553 } else { /* unregister */
554 for (i = 0 ; i < pctx->poll_handlers ; i++) {
555 struct ifnet *ifp = pctx->pr[i].ifp;
557 lwkt_serialize_enter(ifp->if_serializer);
559 if ((ifp->if_flags & IFF_POLLING) == 0) {
560 KKASSERT(ifp->if_poll_cpuid < 0);
561 lwkt_serialize_exit(ifp->if_serializer);
564 ifp->if_flags &= ~IFF_POLLING;
565 ifp->if_poll_cpuid = -1;
568 * Only call the interface deregistration
569 * function if the interface is still
572 if (ifp->if_flags & IFF_RUNNING)
573 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
575 lwkt_serialize_exit(ifp->if_serializer);
577 pctx->residual_burst = 0;
578 pctx->poll_handlers = 0;
585 poll_register(struct netmsg *msg)
587 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
588 struct pollctx *pctx;
591 cpuid = mycpu->gd_cpuid;
592 KKASSERT(cpuid < POLLCTX_MAX);
594 pctx = poll_context[cpuid];
595 KKASSERT(pctx != NULL);
596 KKASSERT(pctx->poll_cpuid == cpuid);
598 if (pctx->polling_enabled == 0) {
599 /* Polling disabled, cannot register */
605 * Check if there is room.
607 if (pctx->poll_handlers >= POLL_LIST_LEN) {
609 * List full, cannot register more entries.
610 * This should never happen; if it does, it is probably a
611 * broken driver trying to register multiple times. Checking
612 * this at runtime is expensive, and won't solve the problem
613 * anyways, so just report a few times and then give up.
615 static int verbose = 10; /* XXX */
617 kprintf("poll handlers list full, "
618 "maybe a broken driver ?\n");
623 pctx->pr[pctx->poll_handlers].ifp = ifp;
624 pctx->poll_handlers++;
627 if (pctx->poll_handlers == 1) {
628 KKASSERT(pctx->polling_enabled);
629 systimer_adjust_periodic(&pctx->pollclock,
634 lwkt_replymsg(&msg->nm_lmsg, rc);
638 * Try to register routine for polling. Returns 1 if successful
639 * (and polling should be enabled), 0 otherwise.
641 * Called from mainline code only, not called from an interrupt.
644 ether_poll_register(struct ifnet *ifp)
648 KKASSERT(poll_defcpu < POLLCTX_MAX);
650 return ether_pollcpu_register(ifp, poll_defcpu);
654 ether_pollcpu_register(struct ifnet *ifp, int cpuid)
660 if (ifp->if_poll == NULL) {
661 /* Device does not support polling */
665 if (cpuid < 0 || cpuid >= POLLCTX_MAX)
668 if (((1 << cpuid) & poll_cpumask) == 0) {
669 /* Polling is not supported on 'cpuid' */
672 KKASSERT(poll_context[cpuid] != NULL);
675 * Attempt to register. Interlock with IFF_POLLING.
677 crit_enter(); /* XXX MP - not mp safe */
679 lwkt_serialize_enter(ifp->if_serializer);
680 if (ifp->if_flags & IFF_POLLING) {
681 /* Already polling */
682 KKASSERT(ifp->if_poll_cpuid >= 0);
683 lwkt_serialize_exit(ifp->if_serializer);
687 KKASSERT(ifp->if_poll_cpuid < 0);
688 ifp->if_flags |= IFF_POLLING;
689 ifp->if_poll_cpuid = cpuid;
690 if (ifp->if_flags & IFF_RUNNING)
691 ifp->if_poll(ifp, POLL_REGISTER, 0);
692 lwkt_serialize_exit(ifp->if_serializer);
694 netmsg_init(&msg, &curthread->td_msgport, 0, poll_register);
695 msg.nm_lmsg.u.ms_resultp = ifp;
697 port = cpu_portfn(cpuid);
698 lwkt_domsg(port, &msg.nm_lmsg, 0);
700 if (msg.nm_lmsg.ms_error) {
701 lwkt_serialize_enter(ifp->if_serializer);
702 ifp->if_flags &= ~IFF_POLLING;
703 ifp->if_poll_cpuid = -1;
704 if (ifp->if_flags & IFF_RUNNING)
705 ifp->if_poll(ifp, POLL_DEREGISTER, 0);
706 lwkt_serialize_exit(ifp->if_serializer);
717 poll_deregister(struct netmsg *msg)
719 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
720 struct pollctx *pctx;
723 cpuid = mycpu->gd_cpuid;
724 KKASSERT(cpuid < POLLCTX_MAX);
726 pctx = poll_context[cpuid];
727 KKASSERT(pctx != NULL);
728 KKASSERT(pctx->poll_cpuid == cpuid);
730 for (i = 0 ; i < pctx->poll_handlers ; i++) {
731 if (pctx->pr[i].ifp == ifp) /* Found it */
734 if (i == pctx->poll_handlers) {
735 kprintf("ether_poll_deregister: ifp not found!!!\n");
738 pctx->poll_handlers--;
739 if (i < pctx->poll_handlers) {
740 /* Last entry replaces this one. */
741 pctx->pr[i].ifp = pctx->pr[pctx->poll_handlers].ifp;
744 if (pctx->poll_handlers == 0)
745 systimer_adjust_periodic(&pctx->pollclock, 1);
748 lwkt_replymsg(&msg->nm_lmsg, rc);
752 * Remove interface from the polling list. Occurs when polling is turned
753 * off. Called from mainline code only, not called from an interrupt.
756 ether_poll_deregister(struct ifnet *ifp)
762 KKASSERT(ifp != NULL);
764 if (ifp->if_poll == NULL)
769 lwkt_serialize_enter(ifp->if_serializer);
770 if ((ifp->if_flags & IFF_POLLING) == 0) {
771 KKASSERT(ifp->if_poll_cpuid < 0);
772 lwkt_serialize_exit(ifp->if_serializer);
777 cpuid = ifp->if_poll_cpuid;
778 KKASSERT(cpuid >= 0);
779 KKASSERT(poll_context[cpuid] != NULL);
781 ifp->if_flags &= ~IFF_POLLING;
782 ifp->if_poll_cpuid = -1;
783 lwkt_serialize_exit(ifp->if_serializer);
785 netmsg_init(&msg, &curthread->td_msgport, 0, poll_deregister);
786 msg.nm_lmsg.u.ms_resultp = ifp;
788 port = cpu_portfn(cpuid);
789 lwkt_domsg(port, &msg.nm_lmsg, 0);
791 if (!msg.nm_lmsg.ms_error) {
792 lwkt_serialize_enter(ifp->if_serializer);
793 if (ifp->if_flags & IFF_RUNNING)
794 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
795 lwkt_serialize_exit(ifp->if_serializer);
806 poll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
807 struct pollctx *pctx)
809 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "enable",
810 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_polling,
811 "I", "Polling enabled");
813 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "pollhz",
814 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_pollhz,
815 "I", "Device polling frequency");
817 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "reg_frac",
818 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_regfrac,
819 "IU", "Every this many cycles poll register");
821 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
822 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_burstmax,
823 "IU", "Max Polling burst size");
825 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
826 &pctx->phase, 0, "Polling phase");
828 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
829 &pctx->suspect, 0, "suspect event");
831 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
832 &pctx->stalled, 0, "potential stalls");
834 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
835 &pctx->poll_burst, 0, "Current polling burst size");
837 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "each_burst", CTLFLAG_RW,
838 &pctx->poll_each_burst, 0, "Max size of each burst");
840 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
842 "Desired user fraction of cpu time");
844 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
845 &pctx->short_ticks, 0,
846 "Hardclock ticks shorter than they should be");
848 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
849 &pctx->lost_polls, 0,
850 "How many times we would have lost a poll tick");
852 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
853 &pctx->pending_polls, 0, "Do we need to poll again");
855 SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
856 &pctx->residual_burst, 0,
857 "# of residual cycles in burst");
859 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
860 &pctx->poll_handlers, 0,
861 "Number of registered poll handlers");
865 schedpoll_oncpu(struct pollctx *pctx, struct netmsg *msg, netisr_fn_t handler)
867 if (msg->nm_lmsg.ms_flags & MSGF_DONE) {
870 netmsg_init(msg, &netisr_adone_rport, 0, handler);
872 msg->nm_lmsg.u.ms_resultp = pctx;
874 port = cpu_portfn(mycpu->gd_cpuid);
875 lwkt_sendmsg(port, &msg->nm_lmsg);
880 poll_sysctl_pollhz(struct netmsg *msg)
882 struct pollctx *pctx;
885 cpuid = mycpu->gd_cpuid;
886 KKASSERT(cpuid < POLLCTX_MAX);
888 pctx = poll_context[cpuid];
889 KKASSERT(pctx != NULL);
890 KKASSERT(pctx->poll_cpuid == cpuid);
893 * If polling is disabled or there is no device registered,
894 * don't adjust polling systimer frequency.
895 * Polling systimer frequency will be adjusted once polling
896 * is enabled and there are registered devices.
898 pctx->pollhz = msg->nm_lmsg.u.ms_result;
899 if (pctx->polling_enabled && pctx->poll_handlers)
900 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
903 * Make sure that reg_frac and reg_frac_count are within valid range.
905 if (pctx->reg_frac > pctx->pollhz) {
906 pctx->reg_frac = pctx->pollhz;
907 if (pctx->reg_frac_count > pctx->reg_frac)
908 pctx->reg_frac_count = pctx->reg_frac - 1;
911 lwkt_replymsg(&msg->nm_lmsg, 0);
915 poll_sysctl_polling(struct netmsg *msg)
917 struct pollctx *pctx;
920 cpuid = mycpu->gd_cpuid;
921 KKASSERT(cpuid < POLLCTX_MAX);
923 pctx = poll_context[cpuid];
924 KKASSERT(pctx != NULL);
925 KKASSERT(pctx->poll_cpuid == cpuid);
928 * If polling is disabled or there is no device registered,
929 * cut the polling systimer frequency to 1hz.
931 pctx->polling_enabled = msg->nm_lmsg.u.ms_result;
932 if (pctx->polling_enabled && pctx->poll_handlers)
933 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
935 systimer_adjust_periodic(&pctx->pollclock, 1);
936 lwkt_replymsg(&msg->nm_lmsg, 0);
940 poll_sysctl_regfrac(struct netmsg *msg)
942 struct pollctx *pctx;
946 cpuid = mycpu->gd_cpuid;
947 KKASSERT(cpuid < POLLCTX_MAX);
949 pctx = poll_context[cpuid];
950 KKASSERT(pctx != NULL);
951 KKASSERT(pctx->poll_cpuid == cpuid);
953 reg_frac = msg->nm_lmsg.u.ms_result;
954 if (reg_frac > pctx->pollhz)
955 reg_frac = pctx->pollhz;
956 else if (reg_frac < 1)
959 pctx->reg_frac = reg_frac;
960 if (pctx->reg_frac_count > pctx->reg_frac)
961 pctx->reg_frac_count = pctx->reg_frac - 1;
963 lwkt_replymsg(&msg->nm_lmsg, 0);
967 poll_sysctl_burstmax(struct netmsg *msg)
969 struct pollctx *pctx;
972 cpuid = mycpu->gd_cpuid;
973 KKASSERT(cpuid < POLLCTX_MAX);
975 pctx = poll_context[cpuid];
976 KKASSERT(pctx != NULL);
977 KKASSERT(pctx->poll_cpuid == cpuid);
979 pctx->poll_burst_max = msg->nm_lmsg.u.ms_result;
980 if (pctx->poll_each_burst > pctx->poll_burst_max)
981 pctx->poll_each_burst = pctx->poll_burst_max;
982 if (pctx->poll_burst > pctx->poll_burst_max)
983 pctx->poll_burst = pctx->poll_burst_max;
984 if (pctx->residual_burst > pctx->poll_burst_max)
985 pctx->residual_burst = pctx->poll_burst_max;
987 lwkt_replymsg(&msg->nm_lmsg, 0);