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 $
30 #include "opt_ifpoll.h"
32 #include <sys/param.h>
33 #include <sys/kernel.h>
35 #include <sys/malloc.h>
36 #include <sys/serialize.h>
37 #include <sys/socket.h>
38 #include <sys/sysctl.h>
40 #include <sys/thread2.h>
41 #include <sys/msgport2.h>
43 #include <machine/atomic.h>
44 #include <machine/clock.h>
45 #include <machine/smp.h>
48 #include <net/if_poll.h>
49 #include <net/netmsg2.h>
52 * Polling support for network device drivers.
54 * Drivers which support this feature try to register one status polling
55 * handler and several TX/RX polling handlers with the polling code.
56 * If interface's if_qpoll is called with non-NULL second argument, then
57 * a register operation is requested, else a deregister operation is
58 * requested. If the requested operation is "register", driver should
59 * setup the ifpoll_info passed in accoding its own needs:
60 * ifpoll_info.ifpi_status.status_func == NULL
61 * No status polling handler will be installed on CPU(0)
62 * ifpoll_info.ifpi_rx[n].poll_func == NULL
63 * No RX polling handler will be installed on CPU(n)
64 * ifpoll_info.ifpi_tx[n].poll_func == NULL
65 * No TX polling handler will be installed on CPU(n)
67 * RX is polled at the specified polling frequency (net.ifpoll.X.pollhz).
68 * TX and status polling could be done at lower frequency than RX frequency
69 * (net.ifpoll.0.status_frac and net.ifpoll.X.tx_frac). To avoid systimer
70 * staggering at high frequency, RX systimer gives TX and status polling a
73 * All of the registered polling handlers are called only if the interface
74 * is marked as 'IFF_RUNNING and IFF_NPOLLING'. However, the interface's
75 * register and deregister function (ifnet.if_qpoll) will be called even
76 * if interface is not marked with 'IFF_RUNNING'.
78 * If registration is successful, the driver must disable interrupts,
79 * and further I/O is performed through the TX/RX polling handler, which
80 * are invoked (at least once per clock tick) with 3 arguments: the "arg"
81 * passed at register time, a struct ifnet pointer, and a "count" limit.
82 * The registered serializer will be held before calling the related
85 * The count limit specifies how much work the handler can do during the
86 * call -- typically this is the number of packets to be received, or
87 * transmitted, etc. (drivers are free to interpret this number, as long
88 * as the max time spent in the function grows roughly linearly with the
91 * A second variable controls the sharing of CPU between polling/kernel
92 * network processing, and other activities (typically userlevel tasks):
93 * net.ifpoll.X.{rx,tx}.user_frac (between 0 and 100, default 50) sets the
94 * share of CPU allocated to user tasks. CPU is allocated proportionally
95 * to the shares, by dynamically adjusting the "count" (poll_burst).
97 * Other parameters can should be left to their default values.
98 * The following constraints hold
100 * 1 <= poll_burst <= poll_burst_max
101 * 1 <= poll_each_burst <= poll_burst_max
102 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
105 #define IFPOLL_LIST_LEN 128
106 #define IFPOLL_FREQ_MAX 30000
108 #define MIN_IOPOLL_BURST_MAX 10
109 #define MAX_IOPOLL_BURST_MAX 1000
110 #define IOPOLL_BURST_MAX 150 /* good for 100Mbit net and HZ=1000 */
112 #define IOPOLL_EACH_BURST 5
114 #define IFPOLL_FREQ_DEFAULT 2000
116 #define IFPOLL_TXFRAC_DEFAULT 1 /* 1/2 of the pollhz */
117 #define IFPOLL_STFRAC_DEFAULT 19 /* 1/20 of the pollhz */
119 #define IFPOLL_RX 0x1
120 #define IFPOLL_TX 0x2
128 struct lwkt_serialize *serializer;
131 ifpoll_iofn_t poll_func;
135 union ifpoll_time prev_t;
136 uint32_t short_ticks; /* statistics */
137 uint32_t lost_polls; /* statistics */
138 uint32_t suspect; /* statistics */
139 uint32_t stalled; /* statistics */
140 uint32_t pending_polls; /* state */
142 struct netmsg_base poll_netmsg;
146 uint32_t phase; /* state */
147 int residual_burst; /* state */
148 uint32_t poll_each_burst; /* tunable */
149 union ifpoll_time poll_start_t; /* state */
151 uint32_t poll_handlers; /* next free entry in pr[]. */
152 struct iopoll_rec pr[IFPOLL_LIST_LEN];
154 struct netmsg_base poll_more_netmsg;
156 uint32_t poll_burst; /* state */
157 uint32_t poll_burst_max; /* tunable */
158 uint32_t user_frac; /* tunable */
159 uint32_t kern_frac; /* state */
161 struct sysctl_ctx_list poll_sysctl_ctx;
162 struct sysctl_oid *poll_sysctl_tree;
166 struct systimer pollclock;
169 int stfrac_count; /* state */
170 int poll_stfrac; /* tunable */
172 int txfrac_count; /* state */
173 int poll_txfrac; /* tunable */
175 int pollhz; /* tunable */
177 struct sysctl_ctx_list sysctl_ctx;
178 struct sysctl_oid *sysctl_tree;
182 struct lwkt_serialize *serializer;
184 ifpoll_stfn_t status_func;
188 struct netmsg_base poll_netmsg;
192 uint32_t poll_handlers; /* next free entry in pr[]. */
193 struct stpoll_rec pr[IFPOLL_LIST_LEN];
195 struct sysctl_ctx_list poll_sysctl_ctx;
196 struct sysctl_oid *poll_sysctl_tree;
199 struct iopoll_sysctl_netmsg {
200 struct netmsg_base base;
201 struct iopoll_ctx *ctx;
204 void ifpoll_init_pcpu(int);
205 static void ifpoll_register_handler(netmsg_t);
206 static void ifpoll_deregister_handler(netmsg_t);
211 static void stpoll_init(void);
212 static void stpoll_handler(netmsg_t);
213 static void stpoll_clock(struct stpoll_ctx *);
214 static int stpoll_register(struct ifnet *, const struct ifpoll_status *);
215 static int stpoll_deregister(struct ifnet *);
220 static struct iopoll_ctx *iopoll_ctx_create(int, int);
221 static void iopoll_init(int);
222 static void iopoll_handler(netmsg_t);
223 static void iopollmore_handler(netmsg_t);
224 static void iopoll_clock(struct iopoll_ctx *);
225 static int iopoll_register(struct ifnet *, struct iopoll_ctx *,
226 const struct ifpoll_io *);
227 static int iopoll_deregister(struct ifnet *, struct iopoll_ctx *);
229 static void iopoll_add_sysctl(struct sysctl_ctx_list *,
230 struct sysctl_oid_list *, struct iopoll_ctx *);
231 static void sysctl_burstmax_handler(netmsg_t);
232 static int sysctl_burstmax(SYSCTL_HANDLER_ARGS);
233 static void sysctl_eachburst_handler(netmsg_t);
234 static int sysctl_eachburst(SYSCTL_HANDLER_ARGS);
239 static void poll_comm_init(int);
240 static void poll_comm_start(int);
241 static void poll_comm_adjust_pollhz(struct poll_comm *);
242 static void poll_comm_systimer0(systimer_t, int, struct intrframe *);
243 static void poll_comm_systimer(systimer_t, int, struct intrframe *);
244 static void sysctl_pollhz_handler(netmsg_t);
245 static void sysctl_stfrac_handler(netmsg_t);
246 static void sysctl_txfrac_handler(netmsg_t);
247 static int sysctl_pollhz(SYSCTL_HANDLER_ARGS);
248 static int sysctl_stfrac(SYSCTL_HANDLER_ARGS);
249 static int sysctl_txfrac(SYSCTL_HANDLER_ARGS);
251 static struct stpoll_ctx stpoll_context;
252 static struct poll_comm *poll_common[MAXCPU];
253 static struct iopoll_ctx *rxpoll_context[MAXCPU];
254 static struct iopoll_ctx *txpoll_context[MAXCPU];
256 SYSCTL_NODE(_net, OID_AUTO, ifpoll, CTLFLAG_RW, 0,
257 "Network device polling parameters");
259 static int iopoll_burst_max = IOPOLL_BURST_MAX;
260 static int iopoll_each_burst = IOPOLL_EACH_BURST;
262 static int ifpoll_pollhz = IFPOLL_FREQ_DEFAULT;
263 static int ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
264 static int ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
266 TUNABLE_INT("net.ifpoll.burst_max", &iopoll_burst_max);
267 TUNABLE_INT("net.ifpoll.each_burst", &iopoll_each_burst);
268 TUNABLE_INT("net.ifpoll.pollhz", &ifpoll_pollhz);
269 TUNABLE_INT("net.ifpoll.status_frac", &ifpoll_stfrac);
270 TUNABLE_INT("net.ifpoll.tx_frac", &ifpoll_txfrac);
273 ifpoll_sendmsg_oncpu(netmsg_t msg)
275 if (msg->lmsg.ms_flags & MSGF_DONE)
276 lwkt_sendmsg(netisr_portfn(mycpuid), &msg->lmsg);
280 sched_stpoll(struct stpoll_ctx *st_ctx)
282 ifpoll_sendmsg_oncpu((netmsg_t)&st_ctx->poll_netmsg);
286 sched_iopoll(struct iopoll_ctx *io_ctx)
288 ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_netmsg);
292 sched_iopollmore(struct iopoll_ctx *io_ctx)
294 ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_more_netmsg);
298 ifpoll_time_get(union ifpoll_time *t)
306 /* Return time diff in us */
308 ifpoll_time_diff(const union ifpoll_time *s, const union ifpoll_time *e)
311 return (((e->tsc - s->tsc) * 1000000) / tsc_frequency);
313 return ((e->tv.tv_usec - s->tv.tv_usec) +
314 (e->tv.tv_sec - s->tv.tv_sec) * 1000000);
319 * Initialize per-cpu qpolling(4) context. Called from kern_clock.c:
322 ifpoll_init_pcpu(int cpuid)
327 poll_comm_init(cpuid);
333 poll_comm_start(cpuid);
337 ifpoll_register(struct ifnet *ifp)
339 struct ifpoll_info *info;
340 struct netmsg_base nmsg;
343 if (ifp->if_qpoll == NULL) {
344 /* Device does not support polling */
348 info = kmalloc(sizeof(*info), M_TEMP, M_WAITOK | M_ZERO);
351 * Attempt to register. Interlock with IFF_NPOLLING.
354 ifnet_serialize_all(ifp);
356 if (ifp->if_flags & IFF_NPOLLING) {
357 /* Already polling */
358 ifnet_deserialize_all(ifp);
363 info->ifpi_ifp = ifp;
365 ifp->if_flags |= IFF_NPOLLING;
366 ifp->if_qpoll(ifp, info);
368 ifnet_deserialize_all(ifp);
370 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
371 0, ifpoll_register_handler);
372 nmsg.lmsg.u.ms_resultp = info;
374 error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
376 if (!ifpoll_deregister(ifp)) {
377 if_printf(ifp, "ifpoll_register: "
378 "ifpoll_deregister failed!\n");
387 ifpoll_deregister(struct ifnet *ifp)
389 struct netmsg_base nmsg;
392 if (ifp->if_qpoll == NULL)
395 ifnet_serialize_all(ifp);
397 if ((ifp->if_flags & IFF_NPOLLING) == 0) {
398 ifnet_deserialize_all(ifp);
401 ifp->if_flags &= ~IFF_NPOLLING;
403 ifnet_deserialize_all(ifp);
405 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
406 0, ifpoll_deregister_handler);
407 nmsg.lmsg.u.ms_resultp = ifp;
409 error = lwkt_domsg(netisr_portfn(0), &nmsg.lmsg, 0);
411 ifnet_serialize_all(ifp);
412 ifp->if_qpoll(ifp, NULL);
413 ifnet_deserialize_all(ifp);
419 ifpoll_register_handler(netmsg_t nmsg)
421 const struct ifpoll_info *info = nmsg->lmsg.u.ms_resultp;
422 int cpuid = mycpuid, nextcpu;
425 KKASSERT(cpuid < ncpus2);
426 KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
429 error = stpoll_register(info->ifpi_ifp, &info->ifpi_status);
434 error = iopoll_register(info->ifpi_ifp, rxpoll_context[cpuid],
435 &info->ifpi_rx[cpuid]);
439 error = iopoll_register(info->ifpi_ifp, txpoll_context[cpuid],
440 &info->ifpi_tx[cpuid]);
444 /* Adjust polling frequency, after all registration is done */
445 poll_comm_adjust_pollhz(poll_common[cpuid]);
448 if (nextcpu < ncpus2)
449 lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
451 lwkt_replymsg(&nmsg->lmsg, 0);
454 lwkt_replymsg(&nmsg->lmsg, error);
458 ifpoll_deregister_handler(netmsg_t nmsg)
460 struct ifnet *ifp = nmsg->lmsg.u.ms_resultp;
461 int cpuid = mycpuid, nextcpu;
463 KKASSERT(cpuid < ncpus2);
464 KKASSERT(&curthread->td_msgport == netisr_portfn(cpuid));
468 stpoll_deregister(ifp);
469 iopoll_deregister(ifp, rxpoll_context[cpuid]);
470 iopoll_deregister(ifp, txpoll_context[cpuid]);
472 /* Adjust polling frequency, after all deregistration is done */
473 poll_comm_adjust_pollhz(poll_common[cpuid]);
476 if (nextcpu < ncpus2)
477 lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->lmsg);
479 lwkt_replymsg(&nmsg->lmsg, 0);
485 struct stpoll_ctx *st_ctx = &stpoll_context;
486 const struct poll_comm *comm = poll_common[0];
488 st_ctx->pollhz = comm->pollhz / (comm->poll_stfrac + 1);
490 sysctl_ctx_init(&st_ctx->poll_sysctl_ctx);
491 st_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&st_ctx->poll_sysctl_ctx,
492 SYSCTL_CHILDREN(comm->sysctl_tree),
493 OID_AUTO, "status", CTLFLAG_RD, 0, "");
495 SYSCTL_ADD_UINT(&st_ctx->poll_sysctl_ctx,
496 SYSCTL_CHILDREN(st_ctx->poll_sysctl_tree),
497 OID_AUTO, "handlers", CTLFLAG_RD,
498 &st_ctx->poll_handlers, 0,
499 "Number of registered status poll handlers");
501 netmsg_init(&st_ctx->poll_netmsg, NULL, &netisr_adone_rport,
506 * stpoll_handler is scheduled by sched_stpoll when appropriate, typically
507 * once per polling systimer tick.
510 stpoll_handler(netmsg_t msg)
512 struct stpoll_ctx *st_ctx = &stpoll_context;
513 struct thread *td = curthread;
516 KKASSERT(&td->td_msgport == netisr_portfn(0));
518 crit_enter_quick(td);
521 lwkt_replymsg(&msg->lmsg, 0);
523 if (st_ctx->poll_handlers == 0) {
528 for (i = 0; i < st_ctx->poll_handlers; ++i) {
529 const struct stpoll_rec *rec = &st_ctx->pr[i];
530 struct ifnet *ifp = rec->ifp;
532 if (!lwkt_serialize_try(rec->serializer))
535 if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
536 (IFF_RUNNING | IFF_NPOLLING))
537 rec->status_func(ifp, st_ctx->pollhz);
539 lwkt_serialize_exit(rec->serializer);
546 * Hook from status poll systimer. Tries to schedule an status poll.
547 * NOTE: Caller should hold critical section.
550 stpoll_clock(struct stpoll_ctx *st_ctx)
552 KKASSERT(mycpuid == 0);
554 if (st_ctx->poll_handlers == 0)
556 sched_stpoll(st_ctx);
560 stpoll_register(struct ifnet *ifp, const struct ifpoll_status *st_rec)
562 struct stpoll_ctx *st_ctx = &stpoll_context;
565 KKASSERT(&curthread->td_msgport == netisr_portfn(0));
567 if (st_rec->status_func == NULL)
571 * Check if there is room.
573 if (st_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
575 * List full, cannot register more entries.
576 * This should never happen; if it does, it is probably a
577 * broken driver trying to register multiple times. Checking
578 * this at runtime is expensive, and won't solve the problem
579 * anyways, so just report a few times and then give up.
581 static int verbose = 10; /* XXX */
584 kprintf("status poll handlers list full, "
585 "maybe a broken driver ?\n");
590 struct stpoll_rec *rec = &st_ctx->pr[st_ctx->poll_handlers];
593 rec->serializer = st_rec->serializer;
594 rec->status_func = st_rec->status_func;
596 st_ctx->poll_handlers++;
603 stpoll_deregister(struct ifnet *ifp)
605 struct stpoll_ctx *st_ctx = &stpoll_context;
608 KKASSERT(&curthread->td_msgport == netisr_portfn(0));
610 for (i = 0; i < st_ctx->poll_handlers; ++i) {
611 if (st_ctx->pr[i].ifp == ifp) /* Found it */
614 if (i == st_ctx->poll_handlers) {
615 kprintf("stpoll_deregister: ifp not found!!!\n");
618 st_ctx->poll_handlers--;
619 if (i < st_ctx->poll_handlers) {
620 /* Last entry replaces this one. */
621 st_ctx->pr[i] = st_ctx->pr[st_ctx->poll_handlers];
629 iopoll_reset_state(struct iopoll_ctx *io_ctx)
632 io_ctx->poll_burst = 5;
633 io_ctx->pending_polls = 0;
634 io_ctx->residual_burst = 0;
636 io_ctx->kern_frac = 0;
637 bzero(&io_ctx->poll_start_t, sizeof(io_ctx->poll_start_t));
638 bzero(&io_ctx->prev_t, sizeof(io_ctx->prev_t));
643 iopoll_init(int cpuid)
645 KKASSERT(cpuid < ncpus2);
647 rxpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_RX);
648 txpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_TX);
651 static struct iopoll_ctx *
652 iopoll_ctx_create(int cpuid, int poll_type)
654 struct poll_comm *comm;
655 struct iopoll_ctx *io_ctx;
656 const char *poll_type_str;
658 KKASSERT(poll_type == IFPOLL_RX || poll_type == IFPOLL_TX);
661 * Make sure that tunables are in sane state
663 if (iopoll_burst_max < MIN_IOPOLL_BURST_MAX)
664 iopoll_burst_max = MIN_IOPOLL_BURST_MAX;
665 else if (iopoll_burst_max > MAX_IOPOLL_BURST_MAX)
666 iopoll_burst_max = MAX_IOPOLL_BURST_MAX;
668 if (iopoll_each_burst > iopoll_burst_max)
669 iopoll_each_burst = iopoll_burst_max;
671 comm = poll_common[cpuid];
674 * Create the per-cpu polling context
676 io_ctx = kmalloc(sizeof(*io_ctx), M_DEVBUF, M_WAITOK | M_ZERO);
678 io_ctx->poll_each_burst = iopoll_each_burst;
679 io_ctx->poll_burst_max = iopoll_burst_max;
680 io_ctx->user_frac = 50;
681 if (poll_type == IFPOLL_RX)
682 io_ctx->pollhz = comm->pollhz;
684 io_ctx->pollhz = comm->pollhz / (comm->poll_txfrac + 1);
685 io_ctx->poll_cpuid = cpuid;
686 iopoll_reset_state(io_ctx);
688 netmsg_init(&io_ctx->poll_netmsg, NULL, &netisr_adone_rport,
690 io_ctx->poll_netmsg.lmsg.u.ms_resultp = io_ctx;
692 netmsg_init(&io_ctx->poll_more_netmsg, NULL, &netisr_adone_rport,
693 0, iopollmore_handler);
694 io_ctx->poll_more_netmsg.lmsg.u.ms_resultp = io_ctx;
697 * Initialize per-cpu sysctl nodes
699 if (poll_type == IFPOLL_RX)
700 poll_type_str = "rx";
702 poll_type_str = "tx";
704 sysctl_ctx_init(&io_ctx->poll_sysctl_ctx);
705 io_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&io_ctx->poll_sysctl_ctx,
706 SYSCTL_CHILDREN(comm->sysctl_tree),
707 OID_AUTO, poll_type_str, CTLFLAG_RD, 0, "");
708 iopoll_add_sysctl(&io_ctx->poll_sysctl_ctx,
709 SYSCTL_CHILDREN(io_ctx->poll_sysctl_tree), io_ctx);
715 * Hook from iopoll systimer. Tries to schedule an iopoll, but keeps
716 * track of lost ticks due to the previous handler taking too long.
717 * Normally, this should not happen, because polling handler should
718 * run for a short time. However, in some cases (e.g. when there are
719 * changes in link status etc.) the drivers take a very long time
720 * (even in the order of milliseconds) to reset and reconfigure the
721 * device, causing apparent lost polls.
723 * The first part of the code is just for debugging purposes, and tries
724 * to count how often hardclock ticks are shorter than they should,
725 * meaning either stray interrupts or delayed events.
727 * WARNING! called from fastint or IPI, the MP lock might not be held.
728 * NOTE: Caller should hold critical section.
731 iopoll_clock(struct iopoll_ctx *io_ctx)
736 KKASSERT(mycpuid == io_ctx->poll_cpuid);
738 if (io_ctx->poll_handlers == 0)
742 delta = ifpoll_time_diff(&io_ctx->prev_t, &t);
743 if (delta * io_ctx->pollhz < 500000)
744 io_ctx->short_ticks++;
748 if (io_ctx->pending_polls > 100) {
750 * Too much, assume it has stalled (not always true
751 * see comment above).
754 io_ctx->pending_polls = 0;
758 if (io_ctx->phase <= 2) {
759 if (io_ctx->phase != 0)
762 sched_iopoll(io_ctx);
765 if (io_ctx->pending_polls++ > 0)
766 io_ctx->lost_polls++;
770 * iopoll_handler is scheduled by sched_iopoll when appropriate, typically
771 * once per polling systimer tick.
773 * Note that the message is replied immediately in order to allow a new
774 * ISR to be scheduled in the handler.
777 iopoll_handler(netmsg_t msg)
779 struct iopoll_ctx *io_ctx;
780 struct thread *td = curthread;
783 io_ctx = msg->lmsg.u.ms_resultp;
784 KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
786 crit_enter_quick(td);
789 lwkt_replymsg(&msg->lmsg, 0);
791 if (io_ctx->poll_handlers == 0) {
797 if (io_ctx->residual_burst == 0) {
798 /* First call in this tick */
799 ifpoll_time_get(&io_ctx->poll_start_t);
800 io_ctx->residual_burst = io_ctx->poll_burst;
802 cycles = (io_ctx->residual_burst < io_ctx->poll_each_burst) ?
803 io_ctx->residual_burst : io_ctx->poll_each_burst;
804 io_ctx->residual_burst -= cycles;
806 for (i = 0; i < io_ctx->poll_handlers; i++) {
807 const struct iopoll_rec *rec = &io_ctx->pr[i];
808 struct ifnet *ifp = rec->ifp;
810 if (!lwkt_serialize_try(rec->serializer))
813 if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
814 (IFF_RUNNING | IFF_NPOLLING))
815 rec->poll_func(ifp, rec->arg, cycles);
817 lwkt_serialize_exit(rec->serializer);
821 * Do a quick exit/enter to catch any higher-priority
825 crit_enter_quick(td);
827 sched_iopollmore(io_ctx);
834 * iopollmore_handler is called after other netisr's, possibly scheduling
835 * another iopoll_handler call, or adapting the burst size for the next cycle.
837 * It is very bad to fetch large bursts of packets from a single card at once,
838 * because the burst could take a long time to be completely processed leading
839 * to unfairness. To reduce the problem, and also to account better for time
840 * spent in network-related processing, we split the burst in smaller chunks
841 * of fixed size, giving control to the other netisr's between chunks. This
842 * helps in improving the fairness, reducing livelock and accounting for the
843 * work performed in low level handling.
846 iopollmore_handler(netmsg_t msg)
848 struct thread *td = curthread;
849 struct iopoll_ctx *io_ctx;
852 uint32_t pending_polls;
854 io_ctx = msg->lmsg.u.ms_resultp;
855 KKASSERT(&td->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
857 crit_enter_quick(td);
860 lwkt_replymsg(&msg->lmsg, 0);
862 if (io_ctx->poll_handlers == 0) {
868 if (io_ctx->residual_burst > 0) {
869 sched_iopoll(io_ctx);
871 /* Will run immediately on return, followed by netisrs */
875 /* Here we can account time spent in iopoll's in this tick */
877 kern_load = ifpoll_time_diff(&io_ctx->poll_start_t, &t);
878 kern_load = (kern_load * io_ctx->pollhz) / 10000; /* 0..100 */
879 io_ctx->kern_frac = kern_load;
881 if (kern_load > (100 - io_ctx->user_frac)) {
882 /* Try decrease ticks */
883 if (io_ctx->poll_burst > 1)
884 io_ctx->poll_burst--;
886 if (io_ctx->poll_burst < io_ctx->poll_burst_max)
887 io_ctx->poll_burst++;
890 io_ctx->pending_polls--;
891 pending_polls = io_ctx->pending_polls;
893 if (pending_polls == 0) {
898 * Last cycle was long and caused us to miss one or more
899 * hardclock ticks. Restart processing again, but slightly
900 * reduce the burst size to prevent that this happens again.
902 io_ctx->poll_burst -= (io_ctx->poll_burst / 8);
903 if (io_ctx->poll_burst < 1)
904 io_ctx->poll_burst = 1;
905 sched_iopoll(io_ctx);
913 iopoll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
914 struct iopoll_ctx *io_ctx)
916 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
917 CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_burstmax,
918 "IU", "Max Polling burst size");
920 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
921 CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_eachburst,
922 "IU", "Max size of each burst");
924 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
925 &io_ctx->phase, 0, "Polling phase");
927 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
928 &io_ctx->suspect, 0, "suspect event");
930 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
931 &io_ctx->stalled, 0, "potential stalls");
933 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
934 &io_ctx->poll_burst, 0, "Current polling burst size");
936 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
937 &io_ctx->user_frac, 0,
938 "Desired user fraction of cpu time");
940 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "kern_frac", CTLFLAG_RD,
941 &io_ctx->kern_frac, 0,
942 "Kernel fraction of cpu time");
944 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
945 &io_ctx->short_ticks, 0,
946 "Hardclock ticks shorter than they should be");
948 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
949 &io_ctx->lost_polls, 0,
950 "How many times we would have lost a poll tick");
952 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
953 &io_ctx->pending_polls, 0, "Do we need to poll again");
955 SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
956 &io_ctx->residual_burst, 0,
957 "# of residual cycles in burst");
959 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
960 &io_ctx->poll_handlers, 0,
961 "Number of registered poll handlers");
965 sysctl_burstmax_handler(netmsg_t nmsg)
967 struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
968 struct iopoll_ctx *io_ctx;
971 KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
973 io_ctx->poll_burst_max = nmsg->lmsg.u.ms_result;
974 if (io_ctx->poll_each_burst > io_ctx->poll_burst_max)
975 io_ctx->poll_each_burst = io_ctx->poll_burst_max;
976 if (io_ctx->poll_burst > io_ctx->poll_burst_max)
977 io_ctx->poll_burst = io_ctx->poll_burst_max;
978 if (io_ctx->residual_burst > io_ctx->poll_burst_max)
979 io_ctx->residual_burst = io_ctx->poll_burst_max;
981 lwkt_replymsg(&nmsg->lmsg, 0);
985 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
987 struct iopoll_ctx *io_ctx = arg1;
988 struct iopoll_sysctl_netmsg msg;
992 burst_max = io_ctx->poll_burst_max;
993 error = sysctl_handle_int(oidp, &burst_max, 0, req);
994 if (error || req->newptr == NULL)
996 if (burst_max < MIN_IOPOLL_BURST_MAX)
997 burst_max = MIN_IOPOLL_BURST_MAX;
998 else if (burst_max > MAX_IOPOLL_BURST_MAX)
999 burst_max = MAX_IOPOLL_BURST_MAX;
1001 netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1002 0, sysctl_burstmax_handler);
1003 msg.base.lmsg.u.ms_result = burst_max;
1006 return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1010 sysctl_eachburst_handler(netmsg_t nmsg)
1012 struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1013 struct iopoll_ctx *io_ctx;
1014 uint32_t each_burst;
1017 KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1019 each_burst = nmsg->lmsg.u.ms_result;
1020 if (each_burst > io_ctx->poll_burst_max)
1021 each_burst = io_ctx->poll_burst_max;
1022 else if (each_burst < 1)
1024 io_ctx->poll_each_burst = each_burst;
1026 lwkt_replymsg(&nmsg->lmsg, 0);
1030 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
1032 struct iopoll_ctx *io_ctx = arg1;
1033 struct iopoll_sysctl_netmsg msg;
1034 uint32_t each_burst;
1037 each_burst = io_ctx->poll_each_burst;
1038 error = sysctl_handle_int(oidp, &each_burst, 0, req);
1039 if (error || req->newptr == NULL)
1042 netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1043 0, sysctl_eachburst_handler);
1044 msg.base.lmsg.u.ms_result = each_burst;
1047 return lwkt_domsg(netisr_portfn(io_ctx->poll_cpuid), &msg.base.lmsg, 0);
1051 iopoll_register(struct ifnet *ifp, struct iopoll_ctx *io_ctx,
1052 const struct ifpoll_io *io_rec)
1056 KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1058 if (io_rec->poll_func == NULL)
1062 * Check if there is room.
1064 if (io_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
1066 * List full, cannot register more entries.
1067 * This should never happen; if it does, it is probably a
1068 * broken driver trying to register multiple times. Checking
1069 * this at runtime is expensive, and won't solve the problem
1070 * anyways, so just report a few times and then give up.
1072 static int verbose = 10; /* XXX */
1074 kprintf("io poll handlers list full, "
1075 "maybe a broken driver ?\n");
1080 struct iopoll_rec *rec = &io_ctx->pr[io_ctx->poll_handlers];
1083 rec->serializer = io_rec->serializer;
1084 rec->arg = io_rec->arg;
1085 rec->poll_func = io_rec->poll_func;
1087 io_ctx->poll_handlers++;
1094 iopoll_deregister(struct ifnet *ifp, struct iopoll_ctx *io_ctx)
1098 KKASSERT(&curthread->td_msgport == netisr_portfn(io_ctx->poll_cpuid));
1100 for (i = 0; i < io_ctx->poll_handlers; ++i) {
1101 if (io_ctx->pr[i].ifp == ifp) /* Found it */
1104 if (i == io_ctx->poll_handlers) {
1107 io_ctx->poll_handlers--;
1108 if (i < io_ctx->poll_handlers) {
1109 /* Last entry replaces this one. */
1110 io_ctx->pr[i] = io_ctx->pr[io_ctx->poll_handlers];
1113 if (io_ctx->poll_handlers == 0)
1114 iopoll_reset_state(io_ctx);
1121 poll_comm_init(int cpuid)
1123 struct poll_comm *comm;
1126 comm = kmalloc(sizeof(*comm), M_DEVBUF, M_WAITOK | M_ZERO);
1128 if (ifpoll_stfrac < 0)
1129 ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
1130 if (ifpoll_txfrac < 0)
1131 ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
1133 comm->pollhz = ifpoll_pollhz;
1134 comm->poll_cpuid = cpuid;
1135 comm->poll_stfrac = ifpoll_stfrac;
1136 comm->poll_txfrac = ifpoll_txfrac;
1138 ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", cpuid);
1140 sysctl_ctx_init(&comm->sysctl_ctx);
1141 comm->sysctl_tree = SYSCTL_ADD_NODE(&comm->sysctl_ctx,
1142 SYSCTL_STATIC_CHILDREN(_net_ifpoll),
1143 OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
1145 SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1146 OID_AUTO, "pollhz", CTLTYPE_INT | CTLFLAG_RW,
1147 comm, 0, sysctl_pollhz,
1148 "I", "Device polling frequency");
1151 SYSCTL_ADD_PROC(&comm->sysctl_ctx,
1152 SYSCTL_CHILDREN(comm->sysctl_tree),
1153 OID_AUTO, "status_frac",
1154 CTLTYPE_INT | CTLFLAG_RW,
1155 comm, 0, sysctl_stfrac,
1156 "I", "# of cycles before status is polled");
1158 SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1159 OID_AUTO, "tx_frac", CTLTYPE_INT | CTLFLAG_RW,
1160 comm, 0, sysctl_txfrac,
1161 "I", "# of cycles before TX is polled");
1163 poll_common[cpuid] = comm;
1167 poll_comm_start(int cpuid)
1169 struct poll_comm *comm = poll_common[cpuid];
1170 systimer_func_t func;
1173 * Initialize systimer
1176 func = poll_comm_systimer0;
1178 func = poll_comm_systimer;
1179 systimer_init_periodic_nq(&comm->pollclock, func, comm, 1);
1183 _poll_comm_systimer(struct poll_comm *comm)
1185 if (comm->txfrac_count-- == 0) {
1186 comm->txfrac_count = comm->poll_txfrac;
1187 iopoll_clock(txpoll_context[comm->poll_cpuid]);
1189 iopoll_clock(rxpoll_context[comm->poll_cpuid]);
1193 poll_comm_systimer0(systimer_t info, int in_ipi __unused,
1194 struct intrframe *frame __unused)
1196 struct poll_comm *comm = info->data;
1197 globaldata_t gd = mycpu;
1199 KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid == 0);
1203 if (comm->stfrac_count-- == 0) {
1204 comm->stfrac_count = comm->poll_stfrac;
1205 stpoll_clock(&stpoll_context);
1207 _poll_comm_systimer(comm);
1213 poll_comm_systimer(systimer_t info, int in_ipi __unused,
1214 struct intrframe *frame __unused)
1216 struct poll_comm *comm = info->data;
1217 globaldata_t gd = mycpu;
1219 KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid != 0);
1222 _poll_comm_systimer(comm);
1227 poll_comm_adjust_pollhz(struct poll_comm *comm)
1232 KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1235 * If there is no polling handler registered, set systimer
1236 * frequency to the lowest value. Polling systimer frequency
1237 * will be adjusted to the requested value, once there are
1238 * registered handlers.
1240 handlers = rxpoll_context[mycpuid]->poll_handlers +
1241 txpoll_context[mycpuid]->poll_handlers;
1242 if (comm->poll_cpuid == 0)
1243 handlers += stpoll_context.poll_handlers;
1245 pollhz = comm->pollhz;
1246 systimer_adjust_periodic(&comm->pollclock, pollhz);
1250 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
1252 struct poll_comm *comm = arg1;
1253 struct netmsg_base nmsg;
1257 error = sysctl_handle_int(oidp, &phz, 0, req);
1258 if (error || req->newptr == NULL)
1262 else if (phz > IFPOLL_FREQ_MAX)
1263 phz = IFPOLL_FREQ_MAX;
1265 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1266 0, sysctl_pollhz_handler);
1267 nmsg.lmsg.u.ms_result = phz;
1269 return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1273 sysctl_pollhz_handler(netmsg_t nmsg)
1275 struct poll_comm *comm = poll_common[mycpuid];
1277 KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1279 /* Save polling frequency */
1280 comm->pollhz = nmsg->lmsg.u.ms_result;
1283 * Adjust cached pollhz
1285 rxpoll_context[mycpuid]->pollhz = comm->pollhz;
1286 txpoll_context[mycpuid]->pollhz =
1287 comm->pollhz / (comm->poll_txfrac + 1);
1289 stpoll_context.pollhz = comm->pollhz / (comm->poll_stfrac + 1);
1292 * Adjust polling frequency
1294 poll_comm_adjust_pollhz(comm);
1296 lwkt_replymsg(&nmsg->lmsg, 0);
1300 sysctl_stfrac(SYSCTL_HANDLER_ARGS)
1302 struct poll_comm *comm = arg1;
1303 struct netmsg_base nmsg;
1306 KKASSERT(comm->poll_cpuid == 0);
1308 stfrac = comm->poll_stfrac;
1309 error = sysctl_handle_int(oidp, &stfrac, 0, req);
1310 if (error || req->newptr == NULL)
1315 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1316 0, sysctl_stfrac_handler);
1317 nmsg.lmsg.u.ms_result = stfrac;
1319 return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1323 sysctl_stfrac_handler(netmsg_t nmsg)
1325 struct poll_comm *comm = poll_common[mycpuid];
1326 int stfrac = nmsg->lmsg.u.ms_result;
1328 KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1331 comm->poll_stfrac = stfrac;
1332 if (comm->stfrac_count > comm->poll_stfrac)
1333 comm->stfrac_count = comm->poll_stfrac;
1336 lwkt_replymsg(&nmsg->lmsg, 0);
1340 sysctl_txfrac(SYSCTL_HANDLER_ARGS)
1342 struct poll_comm *comm = arg1;
1343 struct netmsg_base nmsg;
1346 txfrac = comm->poll_txfrac;
1347 error = sysctl_handle_int(oidp, &txfrac, 0, req);
1348 if (error || req->newptr == NULL)
1353 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1354 0, sysctl_txfrac_handler);
1355 nmsg.lmsg.u.ms_result = txfrac;
1357 return lwkt_domsg(netisr_portfn(comm->poll_cpuid), &nmsg.lmsg, 0);
1361 sysctl_txfrac_handler(netmsg_t nmsg)
1363 struct poll_comm *comm = poll_common[mycpuid];
1364 int txfrac = nmsg->lmsg.u.ms_result;
1366 KKASSERT(&curthread->td_msgport == netisr_portfn(comm->poll_cpuid));
1369 comm->poll_txfrac = txfrac;
1370 if (comm->txfrac_count > comm->poll_txfrac)
1371 comm->txfrac_count = comm->poll_txfrac;
1374 lwkt_replymsg(&nmsg->lmsg, 0);