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 1.38 2007/10/01 09:03:28 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 *);
171 /* Systimer handler */
172 static void pollclock(systimer_t, struct intrframe *);
174 /* Sysctl handlers */
175 static int sysctl_pollhz(SYSCTL_HANDLER_ARGS);
176 static int sysctl_polling(SYSCTL_HANDLER_ARGS);
177 static int sysctl_regfrac(SYSCTL_HANDLER_ARGS);
178 static void poll_add_sysctl(struct sysctl_ctx_list *,
179 struct sysctl_oid_list *, struct pollctx *);
181 static void schedpoll_oncpu(struct pollctx *, struct netmsg *, netisr_fn_t);
183 void init_device_poll_pcpu(int); /* per-cpu init routine */
186 * Initialize per-cpu polling(4) context. Called from kern_clock.c:
189 init_device_poll_pcpu(int cpuid)
191 struct pollctx *pctx;
194 if (cpuid >= POLLCTX_MAX)
197 if (((1 << cpuid) & poll_cpumask0) == 0)
200 poll_cpumask |= (1 << cpuid);
202 pctx = kmalloc(sizeof(*pctx), M_DEVBUF, M_WAITOK | M_ZERO);
204 pctx->poll_burst = 5;
205 pctx->poll_each_burst = 5;
206 pctx->poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
207 pctx->user_frac = 50;
209 pctx->polling_enabled = polling_enabled;
210 pctx->pollhz = pollhz;
211 pctx->poll_cpuid = cpuid;
212 netmsg_init(&pctx->poll_netmsg, &netisr_adone_rport, 0, NULL);
213 netmsg_init(&pctx->poll_more_netmsg, &netisr_adone_rport, 0, NULL);
215 KASSERT(cpuid < POLLCTX_MAX, ("cpu id must < %d", cpuid));
216 poll_context[cpuid] = pctx;
218 if (poll_defcpu < 0) {
222 * Initialize global sysctl nodes, for compat
224 poll_add_sysctl(NULL, SYSCTL_STATIC_CHILDREN(_kern_polling),
229 * Initialize per-cpu sysctl nodes
231 ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", pctx->poll_cpuid);
233 sysctl_ctx_init(&pctx->poll_sysctl_ctx);
234 pctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&pctx->poll_sysctl_ctx,
235 SYSCTL_STATIC_CHILDREN(_kern_polling),
236 OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
237 poll_add_sysctl(&pctx->poll_sysctl_ctx,
238 SYSCTL_CHILDREN(pctx->poll_sysctl_tree), pctx);
241 * Initialize systimer
243 systimer_init_periodic_nq(&pctx->pollclock, pollclock, pctx, 1);
247 schedpoll(struct pollctx *pctx)
249 schedpoll_oncpu(pctx, &pctx->poll_netmsg, netisr_poll);
253 schedpollmore(struct pollctx *pctx)
255 schedpoll_oncpu(pctx, &pctx->poll_more_netmsg, netisr_pollmore);
259 * Set the polling frequency
262 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
264 struct pollctx *pctx = arg1;
270 error = sysctl_handle_int(oidp, &phz, 0, req);
271 if (error || req->newptr == NULL)
275 else if (phz > DEVICE_POLLING_FREQ_MAX)
276 phz = DEVICE_POLLING_FREQ_MAX;
278 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_pollhz);
279 msg.nm_lmsg.u.ms_result = phz;
281 port = cpu_portfn(pctx->poll_cpuid);
282 lwkt_domsg(port, &msg.nm_lmsg, 0);
290 sysctl_polling(SYSCTL_HANDLER_ARGS)
292 struct pollctx *pctx = arg1;
297 enabled = pctx->polling_enabled;
298 error = sysctl_handle_int(oidp, &enabled, 0, req);
299 if (error || req->newptr == NULL)
302 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_polling);
303 msg.nm_lmsg.u.ms_result = enabled;
305 port = cpu_portfn(pctx->poll_cpuid);
306 lwkt_domsg(port, &msg.nm_lmsg, 0);
311 sysctl_regfrac(SYSCTL_HANDLER_ARGS)
313 struct pollctx *pctx = arg1;
319 reg_frac = pctx->reg_frac;
320 error = sysctl_handle_int(oidp, ®_frac, 0, req);
321 if (error || req->newptr == NULL)
324 netmsg_init(&msg, &curthread->td_msgport, 0, poll_sysctl_regfrac);
325 msg.nm_lmsg.u.ms_result = reg_frac;
327 port = cpu_portfn(pctx->poll_cpuid);
328 lwkt_domsg(port, &msg.nm_lmsg, 0);
333 * Hook from polling systimer. Tries to schedule a netisr, but keeps
334 * track of lost ticks due to the previous handler taking too long.
335 * Normally, this should not happen, because polling handler should
336 * run for a short time. However, in some cases (e.g. when there are
337 * changes in link status etc.) the drivers take a very long time
338 * (even in the order of milliseconds) to reset and reconfigure the
339 * device, causing apparent lost polls.
341 * The first part of the code is just for debugging purposes, and tries
342 * to count how often hardclock ticks are shorter than they should,
343 * meaning either stray interrupts or delayed events.
345 * WARNING! called from fastint or IPI, the MP lock might not be held.
348 pollclock(systimer_t info, struct intrframe *frame __unused)
350 struct pollctx *pctx = info->data;
354 if (pctx->poll_handlers == 0)
358 delta = (t.tv_usec - pctx->prev_t.tv_usec) +
359 (t.tv_sec - pctx->prev_t.tv_sec)*1000000;
360 if (delta * pctx->pollhz < 500000)
365 if (pctx->pending_polls > 100) {
367 * Too much, assume it has stalled (not always true
368 * see comment above).
371 pctx->pending_polls = 0;
375 if (pctx->phase <= 2) {
376 if (pctx->phase != 0)
382 if (pctx->pending_polls++ > 0)
387 * netisr_pollmore is called after other netisr's, possibly scheduling
388 * another NETISR_POLL call, or adapting the burst size for the next cycle.
390 * It is very bad to fetch large bursts of packets from a single card at once,
391 * because the burst could take a long time to be completely processed, or
392 * could saturate the intermediate queue (ipintrq or similar) leading to
393 * losses or unfairness. To reduce the problem, and also to account better for
394 * time spent in network-related processing, we split the burst in smaller
395 * chunks of fixed size, giving control to the other netisr's between chunks.
396 * This helps in improving the fairness, reducing livelock (because we
397 * emulate more closely the "process to completion" that we have with
398 * fastforwarding) and accounting for the work performed in low level
399 * handling and forwarding.
404 netisr_pollmore(struct netmsg *msg)
406 struct pollctx *pctx;
408 int kern_load, cpuid;
410 cpuid = mycpu->gd_cpuid;
411 KKASSERT(cpuid < POLLCTX_MAX);
413 pctx = poll_context[cpuid];
414 KKASSERT(pctx != NULL);
415 KKASSERT(pctx->poll_cpuid == cpuid);
416 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
418 lwkt_replymsg(&msg->nm_lmsg, 0);
421 if (pctx->residual_burst > 0) {
423 /* will run immediately on return, followed by netisrs */
426 /* here we can account time spent in netisr's in this tick */
428 kern_load = (t.tv_usec - pctx->poll_start_t.tv_usec) +
429 (t.tv_sec - pctx->poll_start_t.tv_sec)*1000000; /* us */
430 kern_load = (kern_load * pctx->pollhz) / 10000; /* 0..100 */
431 if (kern_load > (100 - pctx->user_frac)) { /* try decrease ticks */
432 if (pctx->poll_burst > 1)
435 if (pctx->poll_burst < pctx->poll_burst_max)
439 pctx->pending_polls--;
440 if (pctx->pending_polls == 0) { /* we are done */
444 * Last cycle was long and caused us to miss one or more
445 * hardclock ticks. Restart processing again, but slightly
446 * reduce the burst size to prevent that this happens again.
448 pctx->poll_burst -= (pctx->poll_burst / 8);
449 if (pctx->poll_burst < 1)
450 pctx->poll_burst = 1;
457 * netisr_poll is scheduled by schedpoll when appropriate, typically once
458 * per polling systimer tick.
460 * Note that the message is replied immediately in order to allow a new
461 * ISR to be scheduled in the handler.
463 * XXX each registration should indicate whether it needs a critical
464 * section to operate.
468 netisr_poll(struct netmsg *msg)
470 struct pollctx *pctx;
471 int i, cycles, cpuid;
472 enum poll_cmd arg = POLL_ONLY;
474 cpuid = mycpu->gd_cpuid;
475 KKASSERT(cpuid < POLLCTX_MAX);
477 pctx = poll_context[cpuid];
478 KKASSERT(pctx != NULL);
479 KKASSERT(pctx->poll_cpuid == cpuid);
480 KKASSERT(pctx == msg->nm_lmsg.u.ms_resultp);
482 lwkt_replymsg(&msg->nm_lmsg, 0);
485 if (pctx->residual_burst == 0) { /* first call in this tick */
486 microuptime(&pctx->poll_start_t);
488 * Check that paremeters are consistent with runtime
489 * variables. Some of these tests could be done at sysctl
490 * time, but the savings would be very limited because we
491 * still have to check against reg_frac_count and
492 * poll_each_burst. So, instead of writing separate sysctl
493 * handlers, we do all here.
496 if (pctx->reg_frac_count-- == 0) {
497 arg = POLL_AND_CHECK_STATUS;
498 pctx->reg_frac_count = pctx->reg_frac - 1;
500 if (pctx->poll_burst_max < MIN_POLL_BURST_MAX)
501 pctx->poll_burst_max = MIN_POLL_BURST_MAX;
502 else if (pctx->poll_burst_max > MAX_POLL_BURST_MAX)
503 pctx->poll_burst_max = MAX_POLL_BURST_MAX;
505 if (pctx->poll_each_burst < 1)
506 pctx->poll_each_burst = 1;
507 else if (pctx->poll_each_burst > pctx->poll_burst_max)
508 pctx->poll_each_burst = pctx->poll_burst_max;
510 pctx->residual_burst = pctx->poll_burst;
512 cycles = (pctx->residual_burst < pctx->poll_each_burst) ?
513 pctx->residual_burst : pctx->poll_each_burst;
514 pctx->residual_burst -= cycles;
516 if (pctx->polling_enabled) {
517 for (i = 0 ; i < pctx->poll_handlers ; i++) {
518 struct ifnet *ifp = pctx->pr[i].ifp;
520 if (!lwkt_serialize_try(ifp->if_serializer))
523 if ((ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING))
524 == (IFF_UP|IFF_RUNNING|IFF_POLLING))
525 ifp->if_poll(ifp, arg, cycles);
527 lwkt_serialize_exit(ifp->if_serializer);
529 } else { /* unregister */
530 for (i = 0 ; i < pctx->poll_handlers ; i++) {
531 struct ifnet *ifp = pctx->pr[i].ifp;
533 lwkt_serialize_enter(ifp->if_serializer);
535 if ((ifp->if_flags & IFF_POLLING) == 0) {
536 KKASSERT(ifp->if_poll_cpuid < 0);
537 lwkt_serialize_exit(ifp->if_serializer);
540 ifp->if_flags &= ~IFF_POLLING;
541 ifp->if_poll_cpuid = -1;
544 * Only call the interface deregistration
545 * function if the interface is still
548 if (ifp->if_flags & IFF_RUNNING)
549 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
551 lwkt_serialize_exit(ifp->if_serializer);
553 pctx->residual_burst = 0;
554 pctx->poll_handlers = 0;
561 poll_register(struct netmsg *msg)
563 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
564 struct pollctx *pctx;
567 cpuid = mycpu->gd_cpuid;
568 KKASSERT(cpuid < POLLCTX_MAX);
570 pctx = poll_context[cpuid];
571 KKASSERT(pctx != NULL);
572 KKASSERT(pctx->poll_cpuid == cpuid);
574 if (pctx->polling_enabled == 0) {
575 /* Polling disabled, cannot register */
581 * Check if there is room.
583 if (pctx->poll_handlers >= POLL_LIST_LEN) {
585 * List full, cannot register more entries.
586 * This should never happen; if it does, it is probably a
587 * broken driver trying to register multiple times. Checking
588 * this at runtime is expensive, and won't solve the problem
589 * anyways, so just report a few times and then give up.
591 static int verbose = 10; /* XXX */
593 kprintf("poll handlers list full, "
594 "maybe a broken driver ?\n");
599 pctx->pr[pctx->poll_handlers].ifp = ifp;
600 pctx->poll_handlers++;
603 if (pctx->poll_handlers == 1) {
604 KKASSERT(pctx->polling_enabled);
605 systimer_adjust_periodic(&pctx->pollclock,
610 lwkt_replymsg(&msg->nm_lmsg, rc);
614 * Try to register routine for polling. Returns 1 if successful
615 * (and polling should be enabled), 0 otherwise.
617 * Called from mainline code only, not called from an interrupt.
620 ether_poll_register(struct ifnet *ifp)
624 KKASSERT(poll_defcpu < POLLCTX_MAX);
626 return ether_pollcpu_register(ifp, poll_defcpu);
630 ether_pollcpu_register(struct ifnet *ifp, int cpuid)
636 if (ifp->if_poll == NULL) {
637 /* Device does not support polling */
641 if (cpuid < 0 || cpuid >= POLLCTX_MAX)
644 if (((1 << cpuid) & poll_cpumask) == 0) {
645 /* Polling is not supported on 'cpuid' */
648 KKASSERT(poll_context[cpuid] != NULL);
651 * Attempt to register. Interlock with IFF_POLLING.
653 crit_enter(); /* XXX MP - not mp safe */
655 lwkt_serialize_enter(ifp->if_serializer);
656 if (ifp->if_flags & IFF_POLLING) {
657 /* Already polling */
658 KKASSERT(ifp->if_poll_cpuid >= 0);
659 lwkt_serialize_exit(ifp->if_serializer);
663 KKASSERT(ifp->if_poll_cpuid < 0);
664 ifp->if_flags |= IFF_POLLING;
665 ifp->if_poll_cpuid = cpuid;
666 if (ifp->if_flags & IFF_RUNNING)
667 ifp->if_poll(ifp, POLL_REGISTER, 0);
668 lwkt_serialize_exit(ifp->if_serializer);
670 netmsg_init(&msg, &curthread->td_msgport, 0, poll_register);
671 msg.nm_lmsg.u.ms_resultp = ifp;
673 port = cpu_portfn(cpuid);
674 lwkt_domsg(port, &msg.nm_lmsg, 0);
676 if (msg.nm_lmsg.ms_error) {
677 lwkt_serialize_enter(ifp->if_serializer);
678 ifp->if_flags &= ~IFF_POLLING;
679 ifp->if_poll_cpuid = -1;
680 if (ifp->if_flags & IFF_RUNNING)
681 ifp->if_poll(ifp, POLL_DEREGISTER, 0);
682 lwkt_serialize_exit(ifp->if_serializer);
693 poll_deregister(struct netmsg *msg)
695 struct ifnet *ifp = msg->nm_lmsg.u.ms_resultp;
696 struct pollctx *pctx;
699 cpuid = mycpu->gd_cpuid;
700 KKASSERT(cpuid < POLLCTX_MAX);
702 pctx = poll_context[cpuid];
703 KKASSERT(pctx != NULL);
704 KKASSERT(pctx->poll_cpuid == cpuid);
706 for (i = 0 ; i < pctx->poll_handlers ; i++) {
707 if (pctx->pr[i].ifp == ifp) /* Found it */
710 if (i == pctx->poll_handlers) {
711 kprintf("ether_poll_deregister: ifp not found!!!\n");
714 pctx->poll_handlers--;
715 if (i < pctx->poll_handlers) {
716 /* Last entry replaces this one. */
717 pctx->pr[i].ifp = pctx->pr[pctx->poll_handlers].ifp;
720 if (pctx->poll_handlers == 0)
721 systimer_adjust_periodic(&pctx->pollclock, 1);
724 lwkt_replymsg(&msg->nm_lmsg, rc);
728 * Remove interface from the polling list. Occurs when polling is turned
729 * off. Called from mainline code only, not called from an interrupt.
732 ether_poll_deregister(struct ifnet *ifp)
738 KKASSERT(ifp != NULL);
740 if (ifp->if_poll == NULL)
745 lwkt_serialize_enter(ifp->if_serializer);
746 if ((ifp->if_flags & IFF_POLLING) == 0) {
747 KKASSERT(ifp->if_poll_cpuid < 0);
748 lwkt_serialize_exit(ifp->if_serializer);
753 cpuid = ifp->if_poll_cpuid;
754 KKASSERT(cpuid >= 0);
755 KKASSERT(poll_context[cpuid] != NULL);
757 ifp->if_flags &= ~IFF_POLLING;
758 ifp->if_poll_cpuid = -1;
759 lwkt_serialize_exit(ifp->if_serializer);
761 netmsg_init(&msg, &curthread->td_msgport, 0, poll_deregister);
762 msg.nm_lmsg.u.ms_resultp = ifp;
764 port = cpu_portfn(cpuid);
765 lwkt_domsg(port, &msg.nm_lmsg, 0);
767 if (!msg.nm_lmsg.ms_error) {
768 lwkt_serialize_enter(ifp->if_serializer);
769 if (ifp->if_flags & IFF_RUNNING)
770 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
771 lwkt_serialize_exit(ifp->if_serializer);
782 poll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
783 struct pollctx *pctx)
785 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "enable",
786 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_polling,
787 "I", "Polling enabled");
789 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "pollhz",
790 CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_pollhz,
791 "I", "Device polling frequency");
793 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "reg_frac",
794 CTLTYPE_UINT | CTLFLAG_RW, pctx, 0, sysctl_regfrac,
795 "IU", "Every this many cycles poll register");
797 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
798 &pctx->phase, 0, "Polling phase");
800 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
801 &pctx->suspect, 0, "suspect event");
803 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
804 &pctx->stalled, 0, "potential stalls");
806 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
807 &pctx->poll_burst, 0, "Current polling burst size");
809 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "each_burst", CTLFLAG_RW,
810 &pctx->poll_each_burst, 0, "Max size of each burst");
812 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst_max", CTLFLAG_RW,
813 &pctx->poll_burst_max, 0, "Max Polling burst size");
815 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
817 "Desired user fraction of cpu time");
819 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
820 &pctx->short_ticks, 0,
821 "Hardclock ticks shorter than they should be");
823 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
824 &pctx->lost_polls, 0,
825 "How many times we would have lost a poll tick");
827 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
828 &pctx->pending_polls, 0, "Do we need to poll again");
830 SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
831 &pctx->residual_burst, 0,
832 "# of residual cycles in burst");
834 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
835 &pctx->poll_handlers, 0,
836 "Number of registered poll handlers");
840 schedpoll_oncpu(struct pollctx *pctx, struct netmsg *msg, netisr_fn_t handler)
842 if (msg->nm_lmsg.ms_flags & MSGF_DONE) {
845 netmsg_init(msg, &netisr_adone_rport, 0, handler);
847 msg->nm_lmsg.u.ms_resultp = pctx;
849 port = cpu_portfn(mycpu->gd_cpuid);
850 lwkt_sendmsg(port, &msg->nm_lmsg);
855 poll_sysctl_pollhz(struct netmsg *msg)
857 struct pollctx *pctx;
860 cpuid = mycpu->gd_cpuid;
861 KKASSERT(cpuid < POLLCTX_MAX);
863 pctx = poll_context[cpuid];
864 KKASSERT(pctx != NULL);
865 KKASSERT(pctx->poll_cpuid == cpuid);
868 * If polling is disabled or there is no device registered,
869 * don't adjust polling systimer frequency.
870 * Polling systimer frequency will be adjusted once polling
871 * is enabled and there are registered devices.
873 pctx->pollhz = msg->nm_lmsg.u.ms_result;
874 if (pctx->polling_enabled && pctx->poll_handlers)
875 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
878 * Make sure that reg_frac and reg_frac_count are within valid range.
880 if (pctx->reg_frac > pctx->pollhz) {
881 pctx->reg_frac = pctx->pollhz;
882 if (pctx->reg_frac_count > pctx->reg_frac)
883 pctx->reg_frac_count = pctx->reg_frac - 1;
886 lwkt_replymsg(&msg->nm_lmsg, 0);
890 poll_sysctl_polling(struct netmsg *msg)
892 struct pollctx *pctx;
895 cpuid = mycpu->gd_cpuid;
896 KKASSERT(cpuid < POLLCTX_MAX);
898 pctx = poll_context[cpuid];
899 KKASSERT(pctx != NULL);
900 KKASSERT(pctx->poll_cpuid == cpuid);
903 * If polling is disabled or there is no device registered,
904 * cut the polling systimer frequency to 1hz.
906 pctx->polling_enabled = msg->nm_lmsg.u.ms_result;
907 if (pctx->polling_enabled && pctx->poll_handlers)
908 systimer_adjust_periodic(&pctx->pollclock, pctx->pollhz);
910 systimer_adjust_periodic(&pctx->pollclock, 1);
911 lwkt_replymsg(&msg->nm_lmsg, 0);
915 poll_sysctl_regfrac(struct netmsg *msg)
917 struct pollctx *pctx;
921 cpuid = mycpu->gd_cpuid;
922 KKASSERT(cpuid < POLLCTX_MAX);
924 pctx = poll_context[cpuid];
925 KKASSERT(pctx != NULL);
926 KKASSERT(pctx->poll_cpuid == cpuid);
928 reg_frac = msg->nm_lmsg.u.ms_result;
929 if (reg_frac > pctx->pollhz)
930 reg_frac = pctx->pollhz;
931 else if (reg_frac < 1)
934 pctx->reg_frac = reg_frac;
935 if (pctx->reg_frac_count > pctx->reg_frac)
936 pctx->reg_frac_count = pctx->reg_frac - 1;
938 lwkt_replymsg(&msg->nm_lmsg, 0);