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, struct intrframe *);
243 static void poll_comm_systimer(systimer_t, 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[IFPOLL_CTX_MAX];
253 static struct iopoll_ctx *rxpoll_context[IFPOLL_CTX_MAX];
254 static struct iopoll_ctx *txpoll_context[IFPOLL_CTX_MAX];
256 SYSCTL_NODE(_net, OID_AUTO, ifpoll, CTLFLAG_RW, 0,
257 "Network device polling parameters");
259 static int ifpoll_ncpus = IFPOLL_CTX_MAX;
261 static int iopoll_burst_max = IOPOLL_BURST_MAX;
262 static int iopoll_each_burst = IOPOLL_EACH_BURST;
264 static int ifpoll_pollhz = IFPOLL_FREQ_DEFAULT;
265 static int ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
266 static int ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
268 TUNABLE_INT("net.ifpoll.burst_max", &iopoll_burst_max);
269 TUNABLE_INT("net.ifpoll.each_burst", &iopoll_each_burst);
270 TUNABLE_INT("net.ifpoll.pollhz", &ifpoll_pollhz);
271 TUNABLE_INT("net.ifpoll.status_frac", &ifpoll_stfrac);
272 TUNABLE_INT("net.ifpoll.tx_frac", &ifpoll_txfrac);
275 ifpoll_sendmsg_oncpu(netmsg_t msg)
277 if (msg->lmsg.ms_flags & MSGF_DONE)
278 ifnet_sendmsg(&msg->lmsg, mycpuid);
282 sched_stpoll(struct stpoll_ctx *st_ctx)
284 ifpoll_sendmsg_oncpu((netmsg_t)&st_ctx->poll_netmsg);
288 sched_iopoll(struct iopoll_ctx *io_ctx)
290 ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_netmsg);
294 sched_iopollmore(struct iopoll_ctx *io_ctx)
296 ifpoll_sendmsg_oncpu((netmsg_t)&io_ctx->poll_more_netmsg);
300 ifpoll_time_get(union ifpoll_time *t)
308 /* Return time diff in us */
310 ifpoll_time_diff(const union ifpoll_time *s, const union ifpoll_time *e)
313 return (((e->tsc - s->tsc) * 1000000) / tsc_frequency);
315 return ((e->tv.tv_usec - s->tv.tv_usec) +
316 (e->tv.tv_sec - s->tv.tv_sec) * 1000000);
321 * Initialize per-cpu qpolling(4) context. Called from kern_clock.c:
324 ifpoll_init_pcpu(int cpuid)
326 if (cpuid >= IFPOLL_CTX_MAX)
330 if (ifpoll_ncpus > ncpus)
331 ifpoll_ncpus = ncpus;
333 kprintf("ifpoll_ncpus %d\n", ifpoll_ncpus);
336 poll_comm_init(cpuid);
342 poll_comm_start(cpuid);
346 ifpoll_register(struct ifnet *ifp)
348 struct ifpoll_info info;
349 struct netmsg_base nmsg;
352 if (ifp->if_qpoll == NULL) {
353 /* Device does not support polling */
358 * Attempt to register. Interlock with IFF_NPOLLING.
361 ifnet_serialize_all(ifp);
363 if (ifp->if_flags & IFF_NPOLLING) {
364 /* Already polling */
365 ifnet_deserialize_all(ifp);
369 bzero(&info, sizeof(info));
372 ifp->if_flags |= IFF_NPOLLING;
373 ifp->if_qpoll(ifp, &info);
375 ifnet_deserialize_all(ifp);
377 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
378 0, ifpoll_register_handler);
379 nmsg.lmsg.u.ms_resultp = &info;
381 error = ifnet_domsg(&nmsg.lmsg, 0);
383 if (!ifpoll_deregister(ifp)) {
384 if_printf(ifp, "ifpoll_register: "
385 "ifpoll_deregister failed!\n");
392 ifpoll_deregister(struct ifnet *ifp)
394 struct netmsg_base nmsg;
397 if (ifp->if_qpoll == NULL)
400 ifnet_serialize_all(ifp);
402 if ((ifp->if_flags & IFF_NPOLLING) == 0) {
403 ifnet_deserialize_all(ifp);
406 ifp->if_flags &= ~IFF_NPOLLING;
408 ifnet_deserialize_all(ifp);
410 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
411 0, ifpoll_deregister_handler);
412 nmsg.lmsg.u.ms_resultp = ifp;
414 error = ifnet_domsg(&nmsg.lmsg, 0);
416 ifnet_serialize_all(ifp);
417 ifp->if_qpoll(ifp, NULL);
418 ifnet_deserialize_all(ifp);
424 ifpoll_register_handler(netmsg_t nmsg)
426 const struct ifpoll_info *info = nmsg->lmsg.u.ms_resultp;
427 int cpuid = mycpuid, nextcpu;
430 KKASSERT(cpuid < ifpoll_ncpus);
431 KKASSERT(&curthread->td_msgport == ifnet_portfn(cpuid));
434 error = stpoll_register(info->ifpi_ifp, &info->ifpi_status);
439 error = iopoll_register(info->ifpi_ifp, rxpoll_context[cpuid],
440 &info->ifpi_rx[cpuid]);
444 error = iopoll_register(info->ifpi_ifp, txpoll_context[cpuid],
445 &info->ifpi_tx[cpuid]);
449 /* Adjust polling frequency, after all registration is done */
450 poll_comm_adjust_pollhz(poll_common[cpuid]);
453 if (nextcpu < ifpoll_ncpus)
454 ifnet_forwardmsg(&nmsg->lmsg, nextcpu);
456 lwkt_replymsg(&nmsg->lmsg, 0);
459 lwkt_replymsg(&nmsg->lmsg, error);
463 ifpoll_deregister_handler(netmsg_t nmsg)
465 struct ifnet *ifp = nmsg->lmsg.u.ms_resultp;
466 int cpuid = mycpuid, nextcpu;
468 KKASSERT(cpuid < ifpoll_ncpus);
469 KKASSERT(&curthread->td_msgport == ifnet_portfn(cpuid));
473 stpoll_deregister(ifp);
474 iopoll_deregister(ifp, rxpoll_context[cpuid]);
475 iopoll_deregister(ifp, txpoll_context[cpuid]);
477 /* Adjust polling frequency, after all deregistration is done */
478 poll_comm_adjust_pollhz(poll_common[cpuid]);
481 if (nextcpu < ifpoll_ncpus)
482 ifnet_forwardmsg(&nmsg->lmsg, nextcpu);
484 lwkt_replymsg(&nmsg->lmsg, 0);
490 struct stpoll_ctx *st_ctx = &stpoll_context;
491 const struct poll_comm *comm = poll_common[0];
493 st_ctx->pollhz = comm->pollhz / (comm->poll_stfrac + 1);
495 sysctl_ctx_init(&st_ctx->poll_sysctl_ctx);
496 st_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&st_ctx->poll_sysctl_ctx,
497 SYSCTL_CHILDREN(comm->sysctl_tree),
498 OID_AUTO, "status", CTLFLAG_RD, 0, "");
500 SYSCTL_ADD_UINT(&st_ctx->poll_sysctl_ctx,
501 SYSCTL_CHILDREN(st_ctx->poll_sysctl_tree),
502 OID_AUTO, "handlers", CTLFLAG_RD,
503 &st_ctx->poll_handlers, 0,
504 "Number of registered status poll handlers");
506 netmsg_init(&st_ctx->poll_netmsg, NULL, &netisr_adone_rport,
511 * stpoll_handler is scheduled by sched_stpoll when appropriate, typically
512 * once per polling systimer tick.
515 stpoll_handler(netmsg_t msg)
517 struct stpoll_ctx *st_ctx = &stpoll_context;
518 struct thread *td = curthread;
521 KKASSERT(&td->td_msgport == ifnet_portfn(0));
523 crit_enter_quick(td);
526 lwkt_replymsg(&msg->lmsg, 0);
528 if (st_ctx->poll_handlers == 0) {
533 for (i = 0; i < st_ctx->poll_handlers; ++i) {
534 const struct stpoll_rec *rec = &st_ctx->pr[i];
535 struct ifnet *ifp = rec->ifp;
537 if (!lwkt_serialize_try(rec->serializer))
540 if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
541 (IFF_RUNNING | IFF_NPOLLING))
542 rec->status_func(ifp, st_ctx->pollhz);
544 lwkt_serialize_exit(rec->serializer);
551 * Hook from status poll systimer. Tries to schedule an status poll.
552 * NOTE: Caller should hold critical section.
555 stpoll_clock(struct stpoll_ctx *st_ctx)
557 KKASSERT(mycpuid == 0);
559 if (st_ctx->poll_handlers == 0)
561 sched_stpoll(st_ctx);
565 stpoll_register(struct ifnet *ifp, const struct ifpoll_status *st_rec)
567 struct stpoll_ctx *st_ctx = &stpoll_context;
570 KKASSERT(&curthread->td_msgport == ifnet_portfn(0));
572 if (st_rec->status_func == NULL)
576 * Check if there is room.
578 if (st_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
580 * List full, cannot register more entries.
581 * This should never happen; if it does, it is probably a
582 * broken driver trying to register multiple times. Checking
583 * this at runtime is expensive, and won't solve the problem
584 * anyways, so just report a few times and then give up.
586 static int verbose = 10; /* XXX */
589 kprintf("status poll handlers list full, "
590 "maybe a broken driver ?\n");
595 struct stpoll_rec *rec = &st_ctx->pr[st_ctx->poll_handlers];
598 rec->serializer = st_rec->serializer;
599 rec->status_func = st_rec->status_func;
601 st_ctx->poll_handlers++;
608 stpoll_deregister(struct ifnet *ifp)
610 struct stpoll_ctx *st_ctx = &stpoll_context;
613 KKASSERT(&curthread->td_msgport == ifnet_portfn(0));
615 for (i = 0; i < st_ctx->poll_handlers; ++i) {
616 if (st_ctx->pr[i].ifp == ifp) /* Found it */
619 if (i == st_ctx->poll_handlers) {
620 kprintf("stpoll_deregister: ifp not found!!!\n");
623 st_ctx->poll_handlers--;
624 if (i < st_ctx->poll_handlers) {
625 /* Last entry replaces this one. */
626 st_ctx->pr[i] = st_ctx->pr[st_ctx->poll_handlers];
634 iopoll_reset_state(struct iopoll_ctx *io_ctx)
637 io_ctx->poll_burst = 5;
638 io_ctx->pending_polls = 0;
639 io_ctx->residual_burst = 0;
641 io_ctx->kern_frac = 0;
642 bzero(&io_ctx->poll_start_t, sizeof(io_ctx->poll_start_t));
643 bzero(&io_ctx->prev_t, sizeof(io_ctx->prev_t));
648 iopoll_init(int cpuid)
650 KKASSERT(cpuid < IFPOLL_CTX_MAX);
652 rxpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_RX);
653 txpoll_context[cpuid] = iopoll_ctx_create(cpuid, IFPOLL_TX);
656 static struct iopoll_ctx *
657 iopoll_ctx_create(int cpuid, int poll_type)
659 struct poll_comm *comm;
660 struct iopoll_ctx *io_ctx;
661 const char *poll_type_str;
663 KKASSERT(poll_type == IFPOLL_RX || poll_type == IFPOLL_TX);
666 * Make sure that tunables are in sane state
668 if (iopoll_burst_max < MIN_IOPOLL_BURST_MAX)
669 iopoll_burst_max = MIN_IOPOLL_BURST_MAX;
670 else if (iopoll_burst_max > MAX_IOPOLL_BURST_MAX)
671 iopoll_burst_max = MAX_IOPOLL_BURST_MAX;
673 if (iopoll_each_burst > iopoll_burst_max)
674 iopoll_each_burst = iopoll_burst_max;
676 comm = poll_common[cpuid];
679 * Create the per-cpu polling context
681 io_ctx = kmalloc(sizeof(*io_ctx), M_DEVBUF, M_WAITOK | M_ZERO);
683 io_ctx->poll_each_burst = iopoll_each_burst;
684 io_ctx->poll_burst_max = iopoll_burst_max;
685 io_ctx->user_frac = 50;
686 if (poll_type == IFPOLL_RX)
687 io_ctx->pollhz = comm->pollhz;
689 io_ctx->pollhz = comm->pollhz / (comm->poll_txfrac + 1);
690 io_ctx->poll_cpuid = cpuid;
691 iopoll_reset_state(io_ctx);
693 netmsg_init(&io_ctx->poll_netmsg, NULL, &netisr_adone_rport,
695 io_ctx->poll_netmsg.lmsg.u.ms_resultp = io_ctx;
697 netmsg_init(&io_ctx->poll_more_netmsg, NULL, &netisr_adone_rport,
698 0, iopollmore_handler);
699 io_ctx->poll_more_netmsg.lmsg.u.ms_resultp = io_ctx;
702 * Initialize per-cpu sysctl nodes
704 if (poll_type == IFPOLL_RX)
705 poll_type_str = "rx";
707 poll_type_str = "tx";
709 sysctl_ctx_init(&io_ctx->poll_sysctl_ctx);
710 io_ctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&io_ctx->poll_sysctl_ctx,
711 SYSCTL_CHILDREN(comm->sysctl_tree),
712 OID_AUTO, poll_type_str, CTLFLAG_RD, 0, "");
713 iopoll_add_sysctl(&io_ctx->poll_sysctl_ctx,
714 SYSCTL_CHILDREN(io_ctx->poll_sysctl_tree), io_ctx);
720 * Hook from iopoll systimer. Tries to schedule an iopoll, but keeps
721 * track of lost ticks due to the previous handler taking too long.
722 * Normally, this should not happen, because polling handler should
723 * run for a short time. However, in some cases (e.g. when there are
724 * changes in link status etc.) the drivers take a very long time
725 * (even in the order of milliseconds) to reset and reconfigure the
726 * device, causing apparent lost polls.
728 * The first part of the code is just for debugging purposes, and tries
729 * to count how often hardclock ticks are shorter than they should,
730 * meaning either stray interrupts or delayed events.
732 * WARNING! called from fastint or IPI, the MP lock might not be held.
733 * NOTE: Caller should hold critical section.
736 iopoll_clock(struct iopoll_ctx *io_ctx)
741 KKASSERT(mycpuid == io_ctx->poll_cpuid);
743 if (io_ctx->poll_handlers == 0)
747 delta = ifpoll_time_diff(&io_ctx->prev_t, &t);
748 if (delta * io_ctx->pollhz < 500000)
749 io_ctx->short_ticks++;
753 if (io_ctx->pending_polls > 100) {
755 * Too much, assume it has stalled (not always true
756 * see comment above).
759 io_ctx->pending_polls = 0;
763 if (io_ctx->phase <= 2) {
764 if (io_ctx->phase != 0)
767 sched_iopoll(io_ctx);
770 if (io_ctx->pending_polls++ > 0)
771 io_ctx->lost_polls++;
775 * iopoll_handler is scheduled by sched_iopoll when appropriate, typically
776 * once per polling systimer tick.
778 * Note that the message is replied immediately in order to allow a new
779 * ISR to be scheduled in the handler.
782 iopoll_handler(netmsg_t msg)
784 struct iopoll_ctx *io_ctx;
785 struct thread *td = curthread;
788 io_ctx = msg->lmsg.u.ms_resultp;
789 KKASSERT(&td->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
791 crit_enter_quick(td);
794 lwkt_replymsg(&msg->lmsg, 0);
796 if (io_ctx->poll_handlers == 0) {
802 if (io_ctx->residual_burst == 0) {
803 /* First call in this tick */
804 ifpoll_time_get(&io_ctx->poll_start_t);
805 io_ctx->residual_burst = io_ctx->poll_burst;
807 cycles = (io_ctx->residual_burst < io_ctx->poll_each_burst) ?
808 io_ctx->residual_burst : io_ctx->poll_each_burst;
809 io_ctx->residual_burst -= cycles;
811 for (i = 0; i < io_ctx->poll_handlers; i++) {
812 const struct iopoll_rec *rec = &io_ctx->pr[i];
813 struct ifnet *ifp = rec->ifp;
815 if (!lwkt_serialize_try(rec->serializer))
818 if ((ifp->if_flags & (IFF_RUNNING | IFF_NPOLLING)) ==
819 (IFF_RUNNING | IFF_NPOLLING))
820 rec->poll_func(ifp, rec->arg, cycles);
822 lwkt_serialize_exit(rec->serializer);
826 * Do a quick exit/enter to catch any higher-priority
830 crit_enter_quick(td);
832 sched_iopollmore(io_ctx);
839 * iopollmore_handler is called after other netisr's, possibly scheduling
840 * another iopoll_handler call, or adapting the burst size for the next cycle.
842 * It is very bad to fetch large bursts of packets from a single card at once,
843 * because the burst could take a long time to be completely processed leading
844 * to unfairness. To reduce the problem, and also to account better for time
845 * spent in network-related processing, we split the burst in smaller chunks
846 * of fixed size, giving control to the other netisr's between chunks. This
847 * helps in improving the fairness, reducing livelock and accounting for the
848 * work performed in low level handling.
851 iopollmore_handler(netmsg_t msg)
853 struct thread *td = curthread;
854 struct iopoll_ctx *io_ctx;
857 uint32_t pending_polls;
859 io_ctx = msg->lmsg.u.ms_resultp;
860 KKASSERT(&td->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
862 crit_enter_quick(td);
865 lwkt_replymsg(&msg->lmsg, 0);
867 if (io_ctx->poll_handlers == 0) {
873 if (io_ctx->residual_burst > 0) {
874 sched_iopoll(io_ctx);
876 /* Will run immediately on return, followed by netisrs */
880 /* Here we can account time spent in iopoll's in this tick */
882 kern_load = ifpoll_time_diff(&io_ctx->poll_start_t, &t);
883 kern_load = (kern_load * io_ctx->pollhz) / 10000; /* 0..100 */
884 io_ctx->kern_frac = kern_load;
886 if (kern_load > (100 - io_ctx->user_frac)) {
887 /* Try decrease ticks */
888 if (io_ctx->poll_burst > 1)
889 io_ctx->poll_burst--;
891 if (io_ctx->poll_burst < io_ctx->poll_burst_max)
892 io_ctx->poll_burst++;
895 io_ctx->pending_polls--;
896 pending_polls = io_ctx->pending_polls;
898 if (pending_polls == 0) {
903 * Last cycle was long and caused us to miss one or more
904 * hardclock ticks. Restart processing again, but slightly
905 * reduce the burst size to prevent that this happens again.
907 io_ctx->poll_burst -= (io_ctx->poll_burst / 8);
908 if (io_ctx->poll_burst < 1)
909 io_ctx->poll_burst = 1;
910 sched_iopoll(io_ctx);
918 iopoll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
919 struct iopoll_ctx *io_ctx)
921 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "burst_max",
922 CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_burstmax,
923 "IU", "Max Polling burst size");
925 SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "each_burst",
926 CTLTYPE_UINT | CTLFLAG_RW, io_ctx, 0, sysctl_eachburst,
927 "IU", "Max size of each burst");
929 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
930 &io_ctx->phase, 0, "Polling phase");
932 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
933 &io_ctx->suspect, 0, "suspect event");
935 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
936 &io_ctx->stalled, 0, "potential stalls");
938 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RD,
939 &io_ctx->poll_burst, 0, "Current polling burst size");
941 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
942 &io_ctx->user_frac, 0,
943 "Desired user fraction of cpu time");
945 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "kern_frac", CTLFLAG_RD,
946 &io_ctx->kern_frac, 0,
947 "Kernel fraction of cpu time");
949 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
950 &io_ctx->short_ticks, 0,
951 "Hardclock ticks shorter than they should be");
953 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
954 &io_ctx->lost_polls, 0,
955 "How many times we would have lost a poll tick");
957 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
958 &io_ctx->pending_polls, 0, "Do we need to poll again");
960 SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RD,
961 &io_ctx->residual_burst, 0,
962 "# of residual cycles in burst");
964 SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
965 &io_ctx->poll_handlers, 0,
966 "Number of registered poll handlers");
970 sysctl_burstmax_handler(netmsg_t nmsg)
972 struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
973 struct iopoll_ctx *io_ctx;
976 KKASSERT(&curthread->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
978 io_ctx->poll_burst_max = nmsg->lmsg.u.ms_result;
979 if (io_ctx->poll_each_burst > io_ctx->poll_burst_max)
980 io_ctx->poll_each_burst = io_ctx->poll_burst_max;
981 if (io_ctx->poll_burst > io_ctx->poll_burst_max)
982 io_ctx->poll_burst = io_ctx->poll_burst_max;
983 if (io_ctx->residual_burst > io_ctx->poll_burst_max)
984 io_ctx->residual_burst = io_ctx->poll_burst_max;
986 lwkt_replymsg(&nmsg->lmsg, 0);
990 sysctl_burstmax(SYSCTL_HANDLER_ARGS)
992 struct iopoll_ctx *io_ctx = arg1;
993 struct iopoll_sysctl_netmsg msg;
997 burst_max = io_ctx->poll_burst_max;
998 error = sysctl_handle_int(oidp, &burst_max, 0, req);
999 if (error || req->newptr == NULL)
1001 if (burst_max < MIN_IOPOLL_BURST_MAX)
1002 burst_max = MIN_IOPOLL_BURST_MAX;
1003 else if (burst_max > MAX_IOPOLL_BURST_MAX)
1004 burst_max = MAX_IOPOLL_BURST_MAX;
1006 netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1007 0, sysctl_burstmax_handler);
1008 msg.base.lmsg.u.ms_result = burst_max;
1011 return ifnet_domsg(&msg.base.lmsg, io_ctx->poll_cpuid);
1015 sysctl_eachburst_handler(netmsg_t nmsg)
1017 struct iopoll_sysctl_netmsg *msg = (struct iopoll_sysctl_netmsg *)nmsg;
1018 struct iopoll_ctx *io_ctx;
1019 uint32_t each_burst;
1022 KKASSERT(&curthread->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
1024 each_burst = nmsg->lmsg.u.ms_result;
1025 if (each_burst > io_ctx->poll_burst_max)
1026 each_burst = io_ctx->poll_burst_max;
1027 else if (each_burst < 1)
1029 io_ctx->poll_each_burst = each_burst;
1031 lwkt_replymsg(&nmsg->lmsg, 0);
1035 sysctl_eachburst(SYSCTL_HANDLER_ARGS)
1037 struct iopoll_ctx *io_ctx = arg1;
1038 struct iopoll_sysctl_netmsg msg;
1039 uint32_t each_burst;
1042 each_burst = io_ctx->poll_each_burst;
1043 error = sysctl_handle_int(oidp, &each_burst, 0, req);
1044 if (error || req->newptr == NULL)
1047 netmsg_init(&msg.base, NULL, &curthread->td_msgport,
1048 0, sysctl_eachburst_handler);
1049 msg.base.lmsg.u.ms_result = each_burst;
1052 return ifnet_domsg(&msg.base.lmsg, io_ctx->poll_cpuid);
1056 iopoll_register(struct ifnet *ifp, struct iopoll_ctx *io_ctx,
1057 const struct ifpoll_io *io_rec)
1061 KKASSERT(&curthread->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
1063 if (io_rec->poll_func == NULL)
1067 * Check if there is room.
1069 if (io_ctx->poll_handlers >= IFPOLL_LIST_LEN) {
1071 * List full, cannot register more entries.
1072 * This should never happen; if it does, it is probably a
1073 * broken driver trying to register multiple times. Checking
1074 * this at runtime is expensive, and won't solve the problem
1075 * anyways, so just report a few times and then give up.
1077 static int verbose = 10; /* XXX */
1079 kprintf("io poll handlers list full, "
1080 "maybe a broken driver ?\n");
1085 struct iopoll_rec *rec = &io_ctx->pr[io_ctx->poll_handlers];
1088 rec->serializer = io_rec->serializer;
1089 rec->arg = io_rec->arg;
1090 rec->poll_func = io_rec->poll_func;
1092 io_ctx->poll_handlers++;
1099 iopoll_deregister(struct ifnet *ifp, struct iopoll_ctx *io_ctx)
1103 KKASSERT(&curthread->td_msgport == ifnet_portfn(io_ctx->poll_cpuid));
1105 for (i = 0; i < io_ctx->poll_handlers; ++i) {
1106 if (io_ctx->pr[i].ifp == ifp) /* Found it */
1109 if (i == io_ctx->poll_handlers) {
1112 io_ctx->poll_handlers--;
1113 if (i < io_ctx->poll_handlers) {
1114 /* Last entry replaces this one. */
1115 io_ctx->pr[i] = io_ctx->pr[io_ctx->poll_handlers];
1118 if (io_ctx->poll_handlers == 0)
1119 iopoll_reset_state(io_ctx);
1126 poll_comm_init(int cpuid)
1128 struct poll_comm *comm;
1131 comm = kmalloc(sizeof(*comm), M_DEVBUF, M_WAITOK | M_ZERO);
1133 if (ifpoll_stfrac < 0)
1134 ifpoll_stfrac = IFPOLL_STFRAC_DEFAULT;
1135 if (ifpoll_txfrac < 0)
1136 ifpoll_txfrac = IFPOLL_TXFRAC_DEFAULT;
1138 comm->pollhz = ifpoll_pollhz;
1139 comm->poll_cpuid = cpuid;
1140 comm->poll_stfrac = ifpoll_stfrac;
1141 comm->poll_txfrac = ifpoll_txfrac;
1143 ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", cpuid);
1145 sysctl_ctx_init(&comm->sysctl_ctx);
1146 comm->sysctl_tree = SYSCTL_ADD_NODE(&comm->sysctl_ctx,
1147 SYSCTL_STATIC_CHILDREN(_net_ifpoll),
1148 OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
1150 SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1151 OID_AUTO, "pollhz", CTLTYPE_INT | CTLFLAG_RW,
1152 comm, 0, sysctl_pollhz,
1153 "I", "Device polling frequency");
1156 SYSCTL_ADD_PROC(&comm->sysctl_ctx,
1157 SYSCTL_CHILDREN(comm->sysctl_tree),
1158 OID_AUTO, "status_frac",
1159 CTLTYPE_INT | CTLFLAG_RW,
1160 comm, 0, sysctl_stfrac,
1161 "I", "# of cycles before status is polled");
1163 SYSCTL_ADD_PROC(&comm->sysctl_ctx, SYSCTL_CHILDREN(comm->sysctl_tree),
1164 OID_AUTO, "tx_frac", CTLTYPE_INT | CTLFLAG_RW,
1165 comm, 0, sysctl_txfrac,
1166 "I", "# of cycles before TX is polled");
1168 poll_common[cpuid] = comm;
1172 poll_comm_start(int cpuid)
1174 struct poll_comm *comm = poll_common[cpuid];
1175 void (*func)(systimer_t, struct intrframe *);
1178 * Initialize systimer
1181 func = poll_comm_systimer0;
1183 func = poll_comm_systimer;
1184 systimer_init_periodic_nq(&comm->pollclock, func, comm, 1);
1188 _poll_comm_systimer(struct poll_comm *comm)
1190 if (comm->txfrac_count-- == 0) {
1191 comm->txfrac_count = comm->poll_txfrac;
1192 iopoll_clock(txpoll_context[comm->poll_cpuid]);
1194 iopoll_clock(rxpoll_context[comm->poll_cpuid]);
1198 poll_comm_systimer0(systimer_t info, struct intrframe *frame __unused)
1200 struct poll_comm *comm = info->data;
1201 globaldata_t gd = mycpu;
1203 KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid == 0);
1207 if (comm->stfrac_count-- == 0) {
1208 comm->stfrac_count = comm->poll_stfrac;
1209 stpoll_clock(&stpoll_context);
1211 _poll_comm_systimer(comm);
1217 poll_comm_systimer(systimer_t info, struct intrframe *frame __unused)
1219 struct poll_comm *comm = info->data;
1220 globaldata_t gd = mycpu;
1222 KKASSERT(comm->poll_cpuid == gd->gd_cpuid && gd->gd_cpuid != 0);
1225 _poll_comm_systimer(comm);
1230 poll_comm_adjust_pollhz(struct poll_comm *comm)
1235 KKASSERT(&curthread->td_msgport == ifnet_portfn(comm->poll_cpuid));
1238 * If there is no polling handler registered, set systimer
1239 * frequency to the lowest value. Polling systimer frequency
1240 * will be adjusted to the requested value, once there are
1241 * registered handlers.
1243 handlers = rxpoll_context[mycpuid]->poll_handlers +
1244 txpoll_context[mycpuid]->poll_handlers;
1245 if (comm->poll_cpuid == 0)
1246 handlers += stpoll_context.poll_handlers;
1248 pollhz = comm->pollhz;
1249 systimer_adjust_periodic(&comm->pollclock, pollhz);
1253 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
1255 struct poll_comm *comm = arg1;
1256 struct netmsg_base nmsg;
1260 error = sysctl_handle_int(oidp, &phz, 0, req);
1261 if (error || req->newptr == NULL)
1265 else if (phz > IFPOLL_FREQ_MAX)
1266 phz = IFPOLL_FREQ_MAX;
1268 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1269 0, sysctl_pollhz_handler);
1270 nmsg.lmsg.u.ms_result = phz;
1272 return ifnet_domsg(&nmsg.lmsg, comm->poll_cpuid);
1276 sysctl_pollhz_handler(netmsg_t nmsg)
1278 struct poll_comm *comm = poll_common[mycpuid];
1280 KKASSERT(&curthread->td_msgport == ifnet_portfn(comm->poll_cpuid));
1282 /* Save polling frequency */
1283 comm->pollhz = nmsg->lmsg.u.ms_result;
1286 * Adjust cached pollhz
1288 rxpoll_context[mycpuid]->pollhz = comm->pollhz;
1289 txpoll_context[mycpuid]->pollhz =
1290 comm->pollhz / (comm->poll_txfrac + 1);
1292 stpoll_context.pollhz = comm->pollhz / (comm->poll_stfrac + 1);
1295 * Adjust polling frequency
1297 poll_comm_adjust_pollhz(comm);
1299 lwkt_replymsg(&nmsg->lmsg, 0);
1303 sysctl_stfrac(SYSCTL_HANDLER_ARGS)
1305 struct poll_comm *comm = arg1;
1306 struct netmsg_base nmsg;
1309 KKASSERT(comm->poll_cpuid == 0);
1311 stfrac = comm->poll_stfrac;
1312 error = sysctl_handle_int(oidp, &stfrac, 0, req);
1313 if (error || req->newptr == NULL)
1318 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1319 0, sysctl_stfrac_handler);
1320 nmsg.lmsg.u.ms_result = stfrac;
1322 return ifnet_domsg(&nmsg.lmsg, comm->poll_cpuid);
1326 sysctl_stfrac_handler(netmsg_t nmsg)
1328 struct poll_comm *comm = poll_common[mycpuid];
1329 int stfrac = nmsg->lmsg.u.ms_result;
1331 KKASSERT(&curthread->td_msgport == ifnet_portfn(comm->poll_cpuid));
1334 comm->poll_stfrac = stfrac;
1335 if (comm->stfrac_count > comm->poll_stfrac)
1336 comm->stfrac_count = comm->poll_stfrac;
1339 lwkt_replymsg(&nmsg->lmsg, 0);
1343 sysctl_txfrac(SYSCTL_HANDLER_ARGS)
1345 struct poll_comm *comm = arg1;
1346 struct netmsg_base nmsg;
1349 txfrac = comm->poll_txfrac;
1350 error = sysctl_handle_int(oidp, &txfrac, 0, req);
1351 if (error || req->newptr == NULL)
1356 netmsg_init(&nmsg, NULL, &curthread->td_msgport,
1357 0, sysctl_txfrac_handler);
1358 nmsg.lmsg.u.ms_result = txfrac;
1360 return ifnet_domsg(&nmsg.lmsg, comm->poll_cpuid);
1364 sysctl_txfrac_handler(netmsg_t nmsg)
1366 struct poll_comm *comm = poll_common[mycpuid];
1367 int txfrac = nmsg->lmsg.u.ms_result;
1369 KKASSERT(&curthread->td_msgport == ifnet_portfn(comm->poll_cpuid));
1372 comm->poll_txfrac = txfrac;
1373 if (comm->txfrac_count > comm->poll_txfrac)
1374 comm->txfrac_count = comm->poll_txfrac;
1377 lwkt_replymsg(&nmsg->lmsg, 0);