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
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include "opt_device_polling.h"
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/socket.h> /* needed by net/if.h */
37 #include <sys/sockio.h>
38 #include <sys/sysctl.h>
39 #include <sys/syslog.h>
41 #include <net/if.h> /* for IFF_* flags */
42 #include <net/netisr.h> /* for NETISR_POLL */
45 #include <sys/resourcevar.h>
46 #include <sys/kthread.h>
48 static void netisr_poll(void); /* the two netisr handlers */
49 static void netisr_pollmore(void);
50 static int poll_switch(SYSCTL_HANDLER_ARGS);
52 void hardclock_device_poll(void); /* hook from hardclock */
53 void ether_poll(int); /* polling in idle loop */
55 static struct mtx poll_mtx;
58 * Polling support for [network] device drivers.
60 * Drivers which support this feature can register with the
63 * If registration is successful, the driver must disable interrupts,
64 * and further I/O is performed through the handler, which is invoked
65 * (at least once per clock tick) with 3 arguments: the "arg" passed at
66 * register time (a struct ifnet pointer), a command, and a "count" limit.
68 * The command can be one of the following:
69 * POLL_ONLY: quick move of "count" packets from input/output queues.
70 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
71 * other more expensive operations. This command is issued periodically
72 * but less frequently than POLL_ONLY.
74 * The count limit specifies how much work the handler can do during the
75 * call -- typically this is the number of packets to be received, or
76 * transmitted, etc. (drivers are free to interpret this number, as long
77 * as the max time spent in the function grows roughly linearly with the
80 * Polling is enabled and disabled via setting IFCAP_POLLING flag on
81 * the interface. The driver ioctl handler should register interface
82 * with polling and disable interrupts, if registration was successful.
84 * A second variable controls the sharing of CPU between polling/kernel
85 * network processing, and other activities (typically userlevel tasks):
86 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
87 * of CPU allocated to user tasks. CPU is allocated proportionally to the
88 * shares, by dynamically adjusting the "count" (poll_burst).
90 * Other parameters can should be left to their default values.
91 * The following constraints hold
93 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
94 * 0 <= poll_each_burst
95 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
98 #define MIN_POLL_BURST_MAX 10
99 #define MAX_POLL_BURST_MAX 1000
101 static uint32_t poll_burst = 5;
102 static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
103 static uint32_t poll_each_burst = 5;
105 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
106 "Device polling parameters");
108 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
109 &poll_burst, 0, "Current polling burst size");
111 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
113 uint32_t val = poll_burst_max;
116 error = sysctl_handle_int(oidp, &val, sizeof(int), req);
117 if (error || !req->newptr )
119 if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
123 poll_burst_max = val;
124 if (poll_burst > poll_burst_max)
125 poll_burst = poll_burst_max;
126 if (poll_each_burst > poll_burst_max)
127 poll_each_burst = MIN_POLL_BURST_MAX;
128 mtx_unlock(&poll_mtx);
132 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
133 0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
135 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
137 uint32_t val = poll_each_burst;
140 error = sysctl_handle_int(oidp, &val, sizeof(int), req);
141 if (error || !req->newptr )
147 if (val > poll_burst_max) {
148 mtx_unlock(&poll_mtx);
151 poll_each_burst = val;
152 mtx_unlock(&poll_mtx);
156 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
157 0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
158 "Max size of each burst");
160 static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */
161 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
162 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
164 static uint32_t user_frac = 50;
165 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
167 uint32_t val = user_frac;
170 error = sysctl_handle_int(oidp, &val, sizeof(int), req);
171 if (error || !req->newptr )
173 if (val < 0 || val > 99)
178 mtx_unlock(&poll_mtx);
182 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
183 0, sizeof(uint32_t), user_frac_sysctl, "I",
184 "Desired user fraction of cpu time");
186 static uint32_t reg_frac_count = 0;
187 static uint32_t reg_frac = 20 ;
188 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
190 uint32_t val = reg_frac;
193 error = sysctl_handle_int(oidp, &val, sizeof(int), req);
194 if (error || !req->newptr )
196 if (val < 1 || val > hz)
201 if (reg_frac_count >= reg_frac)
203 mtx_unlock(&poll_mtx);
207 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
208 0, sizeof(uint32_t), reg_frac_sysctl, "I",
209 "Every this many cycles check registers");
211 static uint32_t short_ticks;
212 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
213 &short_ticks, 0, "Hardclock ticks shorter than they should be");
215 static uint32_t lost_polls;
216 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
217 &lost_polls, 0, "How many times we would have lost a poll tick");
219 static uint32_t pending_polls;
220 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
221 &pending_polls, 0, "Do we need to poll again");
223 static int residual_burst = 0;
224 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
225 &residual_burst, 0, "# of residual cycles in burst");
227 static uint32_t poll_handlers; /* next free entry in pr[]. */
228 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
229 &poll_handlers, 0, "Number of registered poll handlers");
231 static int polling = 0;
232 SYSCTL_PROC(_kern_polling, OID_AUTO, enable, CTLTYPE_UINT | CTLFLAG_RW,
233 0, sizeof(int), poll_switch, "I", "Switch polling for all interfaces");
235 static uint32_t phase;
236 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
237 &phase, 0, "Polling phase");
239 static uint32_t suspect;
240 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
241 &suspect, 0, "suspect event");
243 static uint32_t stalled;
244 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
245 &stalled, 0, "potential stalls");
247 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
248 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
249 &idlepoll_sleeping, 0, "idlepoll is sleeping");
252 #define POLL_LIST_LEN 128
254 poll_handler_t *handler;
258 static struct pollrec pr[POLL_LIST_LEN];
261 init_device_poll(void)
264 mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
265 netisr_register(NETISR_POLL, (netisr_t *)netisr_poll, NULL,
267 netisr_register(NETISR_POLLMORE, (netisr_t *)netisr_pollmore, NULL,
270 SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL)
274 * Hook from hardclock. Tries to schedule a netisr, but keeps track
275 * of lost ticks due to the previous handler taking too long.
276 * Normally, this should not happen, because polling handler should
277 * run for a short time. However, in some cases (e.g. when there are
278 * changes in link status etc.) the drivers take a very long time
279 * (even in the order of milliseconds) to reset and reconfigure the
280 * device, causing apparent lost polls.
282 * The first part of the code is just for debugging purposes, and tries
283 * to count how often hardclock ticks are shorter than they should,
284 * meaning either stray interrupts or delayed events.
287 hardclock_device_poll(void)
289 static struct timeval prev_t, t;
292 if (poll_handlers == 0)
296 delta = (t.tv_usec - prev_t.tv_usec) +
297 (t.tv_sec - prev_t.tv_sec)*1000000;
298 if (delta * hz < 500000)
303 if (pending_polls > 100) {
305 * Too much, assume it has stalled (not always true
306 * see comment above).
317 schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
320 if (pending_polls++ > 0)
325 * ether_poll is called from the idle loop.
328 ether_poll(int count)
335 if (count > poll_each_burst)
336 count = poll_each_burst;
338 for (i = 0 ; i < poll_handlers ; i++)
339 pr[i].handler(pr[i].ifp, POLL_ONLY, count);
341 mtx_unlock(&poll_mtx);
346 * netisr_pollmore is called after other netisr's, possibly scheduling
347 * another NETISR_POLL call, or adapting the burst size for the next cycle.
349 * It is very bad to fetch large bursts of packets from a single card at once,
350 * because the burst could take a long time to be completely processed, or
351 * could saturate the intermediate queue (ipintrq or similar) leading to
352 * losses or unfairness. To reduce the problem, and also to account better for
353 * time spent in network-related processing, we split the burst in smaller
354 * chunks of fixed size, giving control to the other netisr's between chunks.
355 * This helps in improving the fairness, reducing livelock (because we
356 * emulate more closely the "process to completion" that we have with
357 * fastforwarding) and accounting for the work performed in low level
358 * handling and forwarding.
361 static struct timeval poll_start_t;
373 if (residual_burst > 0) {
374 schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
375 mtx_unlock(&poll_mtx);
376 /* will run immediately on return, followed by netisrs */
379 /* here we can account time spent in netisr's in this tick */
381 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
382 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
383 kern_load = (kern_load * hz) / 10000; /* 0..100 */
384 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
388 if (poll_burst < poll_burst_max)
393 if (pending_polls == 0) /* we are done */
397 * Last cycle was long and caused us to miss one or more
398 * hardclock ticks. Restart processing again, but slightly
399 * reduce the burst size to prevent that this happens again.
401 poll_burst -= (poll_burst / 8);
404 schednetisrbits(1 << NETISR_POLL | 1 << NETISR_POLLMORE);
407 mtx_unlock(&poll_mtx);
411 * netisr_poll is scheduled by schednetisr when appropriate, typically once
418 enum poll_cmd arg = POLL_ONLY;
424 if (residual_burst == 0) { /* first call in this tick */
425 microuptime(&poll_start_t);
426 if (++reg_frac_count == reg_frac) {
427 arg = POLL_AND_CHECK_STATUS;
431 residual_burst = poll_burst;
433 cycles = (residual_burst < poll_each_burst) ?
434 residual_burst : poll_each_burst;
435 residual_burst -= cycles;
437 for (i = 0 ; i < poll_handlers ; i++)
438 pr[i].handler(pr[i].ifp, arg, cycles);
441 mtx_unlock(&poll_mtx);
445 * Try to register routine for polling. Returns 0 if successful
446 * (and polling should be enabled), error code otherwise.
447 * A device is not supposed to register itself multiple times.
449 * This is called from within the *_ioctl() functions.
452 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
456 KASSERT(h != NULL, ("%s: handler is NULL", __func__));
457 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
462 if (poll_handlers >= POLL_LIST_LEN) {
464 * List full, cannot register more entries.
465 * This should never happen; if it does, it is probably a
466 * broken driver trying to register multiple times. Checking
467 * this at runtime is expensive, and won't solve the problem
468 * anyways, so just report a few times and then give up.
470 static int verbose = 10 ;
472 log(LOG_ERR, "poll handlers list full, "
473 "maybe a broken driver ?\n");
476 mtx_unlock(&poll_mtx);
477 return (ENOMEM); /* no polling for you */
480 for (i = 0 ; i < poll_handlers ; i++)
481 if (pr[i].ifp == ifp && pr[i].handler != NULL) {
482 mtx_unlock(&poll_mtx);
483 log(LOG_DEBUG, "ether_poll_register: %s: handler"
484 " already registered\n", ifp->if_xname);
488 pr[poll_handlers].handler = h;
489 pr[poll_handlers].ifp = ifp;
491 mtx_unlock(&poll_mtx);
492 if (idlepoll_sleeping)
493 wakeup(&idlepoll_sleeping);
498 * Remove interface from the polling list. Called from *_ioctl(), too.
501 ether_poll_deregister(struct ifnet *ifp)
505 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
510 for (i = 0 ; i < poll_handlers ; i++)
511 if (pr[i].ifp == ifp) /* found it */
513 if (i == poll_handlers) {
514 log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
516 mtx_unlock(&poll_mtx);
520 if (i < poll_handlers) { /* Last entry replaces this one. */
521 pr[i].handler = pr[poll_handlers].handler;
522 pr[i].ifp = pr[poll_handlers].ifp;
524 mtx_unlock(&poll_mtx);
529 * Legacy interface for turning polling on all interfaces at one time.
532 poll_switch(SYSCTL_HANDLER_ARGS)
538 error = sysctl_handle_int(oidp, &val, sizeof(int), req);
539 if (error || !req->newptr )
545 if (val < 0 || val > 1)
552 TAILQ_FOREACH(ifp, &ifnet, if_link) {
553 if (ifp->if_capabilities & IFCAP_POLLING) {
558 ifp->if_capenable | IFCAP_POLLING;
561 ifp->if_capenable & ~IFCAP_POLLING;
562 IFF_LOCKGIANT(ifp); /* LOR here */
563 (void) (*ifp->if_ioctl)(ifp, SIOCSIFCAP, (caddr_t)&ifr);
564 IFF_UNLOCKGIANT(ifp);
570 log(LOG_ERR, "kern.polling.enable is deprecated. Use ifconfig(8)");
578 struct thread *td = curthread;
581 rtp.prio = RTP_PRIO_MAX; /* lowest priority */
582 rtp.type = RTP_PRIO_IDLE;
583 mtx_lock_spin(&sched_lock);
584 rtp_to_pri(&rtp, td);
585 mtx_unlock_spin(&sched_lock);
588 if (poll_in_idle_loop && poll_handlers > 0) {
589 idlepoll_sleeping = 0;
590 ether_poll(poll_each_burst);
591 mtx_lock_spin(&sched_lock);
592 mi_switch(SW_VOL, NULL);
593 mtx_unlock_spin(&sched_lock);
595 idlepoll_sleeping = 1;
596 tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
601 static struct proc *idlepoll;
602 static struct kproc_desc idlepoll_kp = {
607 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start, &idlepoll_kp)