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
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/socket.h> /* needed by net/if.h */
34 #include <sys/sysctl.h>
36 #include <net/if.h> /* for IFF_* flags */
37 #include <net/netisr.h> /* for NETISR_POLL */
40 #include <sys/resourcevar.h>
41 #include <sys/kthread.h>
44 #ifndef COMPILING_LINT
45 #error DEVICE_POLLING is not compatible with SMP
49 static void netisr_poll(void); /* the two netisr handlers */
50 void netisr_pollmore(void);
52 void init_device_poll(void); /* init routine */
53 void hardclock_device_poll(void); /* hook from hardclock */
54 void ether_poll(int); /* polling while in trap */
57 * Polling support for [network] device drivers.
59 * Drivers which support this feature try to register with the
62 * If registration is successful, the driver must disable interrupts,
63 * and further I/O is performed through the handler, which is invoked
64 * (at least once per clock tick) with 3 arguments: the "arg" passed at
65 * register time (a struct ifnet pointer), a command, and a "count" limit.
67 * The command can be one of the following:
68 * POLL_ONLY: quick move of "count" packets from input/output queues.
69 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
70 * other more expensive operations. This command is issued periodically
71 * but less frequently than POLL_ONLY.
72 * POLL_DEREGISTER: deregister and return to interrupt mode.
74 * The first two commands are only issued if the interface is marked as
75 * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
77 * The count limit specifies how much work the handler can do during the
78 * call -- typically this is the number of packets to be received, or
79 * transmitted, etc. (drivers are free to interpret this number, as long
80 * as the max time spent in the function grows roughly linearly with the
83 * Deregistration can be requested by the driver itself (typically in the
84 * *_stop() routine), or by the polling code, by invoking the handler.
86 * Polling can be globally enabled or disabled with the sysctl variable
87 * kern.polling.enable (default is 0, disabled)
89 * A second variable controls the sharing of CPU between polling/kernel
90 * network processing, and other activities (typically userlevel tasks):
91 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
92 * of CPU allocated to user tasks. CPU is allocated proportionally to the
93 * shares, by dynamically adjusting the "count" (poll_burst).
95 * Other parameters can should be left to their default values.
96 * The following constraints hold
98 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
99 * 0 <= poll_in_trap <= poll_each_burst
100 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
103 #define MIN_POLL_BURST_MAX 10
104 #define MAX_POLL_BURST_MAX 1000
106 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
107 "Device polling parameters");
109 static u_int32_t poll_burst = 5;
110 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
111 &poll_burst, 0, "Current polling burst size");
113 static u_int32_t poll_each_burst = 5;
114 SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
115 &poll_each_burst, 0, "Max size of each burst");
117 static u_int32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
118 SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
119 &poll_burst_max, 0, "Max Polling burst size");
121 static u_int32_t poll_in_idle_loop=1; /* do we poll in idle loop ? */
122 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
123 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
125 u_int32_t poll_in_trap; /* used in trap.c */
126 SYSCTL_UINT(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW,
127 &poll_in_trap, 0, "Poll burst size during a trap");
129 static u_int32_t user_frac = 50;
130 SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
131 &user_frac, 0, "Desired user fraction of cpu time");
133 static u_int32_t reg_frac = 20 ;
134 SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
135 ®_frac, 0, "Every this many cycles poll register");
137 static u_int32_t short_ticks;
138 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
139 &short_ticks, 0, "Hardclock ticks shorter than they should be");
141 static u_int32_t lost_polls;
142 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
143 &lost_polls, 0, "How many times we would have lost a poll tick");
145 static u_int32_t pending_polls;
146 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
147 &pending_polls, 0, "Do we need to poll again");
149 static int residual_burst = 0;
150 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
151 &residual_burst, 0, "# of residual cycles in burst");
153 static u_int32_t poll_handlers; /* next free entry in pr[]. */
154 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
155 &poll_handlers, 0, "Number of registered poll handlers");
157 static int polling = 0; /* global polling enable */
158 SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
159 &polling, 0, "Polling enabled");
161 static u_int32_t phase;
162 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
163 &phase, 0, "Polling phase");
165 static u_int32_t suspect;
166 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
167 &suspect, 0, "suspect event");
169 static u_int32_t stalled;
170 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
171 &stalled, 0, "potential stalls");
173 static u_int32_t idlepoll_sleeping; /* idlepoll is sleeping */
174 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
175 &idlepoll_sleeping, 0, "idlepoll is sleeping");
178 #define POLL_LIST_LEN 128
180 poll_handler_t *handler;
184 static struct pollrec pr[POLL_LIST_LEN];
187 * register relevant netisr. Called from kern_clock.c:
190 init_device_poll(void)
192 register_netisr(NETISR_POLL, netisr_poll);
196 * Hook from hardclock. Tries to schedule a netisr, but keeps track
197 * of lost ticks due to the previous handler taking too long.
198 * The first part of the code is just for debugging purposes, and tries
199 * to count how often hardclock ticks are shorter than they should,
200 * meaning either stray interrupts or delayed events.
203 hardclock_device_poll(void)
205 static struct timeval prev_t, t;
208 if (poll_handlers == 0)
212 delta = (t.tv_usec - prev_t.tv_usec) +
213 (t.tv_sec - prev_t.tv_sec)*1000000;
214 if (delta * hz < 500000)
219 if (pending_polls > 100) {
220 /* too much, assume it has stalled */
222 printf("poll stalled [%d] in phase %d\n",
232 schednetisr(NETISR_POLL);
235 if (pending_polls++ > 0)
240 * ether_poll is called from the idle loop or from the trap handler.
243 ether_poll(int count)
249 if (count > poll_each_burst)
250 count = poll_each_burst;
251 for (i = 0 ; i < poll_handlers ; i++)
252 if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
253 (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
254 pr[i].handler(pr[i].ifp, 0, count); /* quick check */
259 * netisr_pollmore is called after other netisr's, possibly scheduling
260 * another NETISR_POLL call, or adapting the burst size for the next cycle.
262 * It is very bad to fetch large bursts of packets from a single card at once,
263 * because the burst could take a long time to be completely processed, or
264 * could saturate the intermediate queue (ipintrq or similar) leading to
265 * losses or unfairness. To reduce the problem, and also to account better for
266 * time spent in network-related processing, we split the burst in smaller
267 * chunks of fixed size, giving control to the other netisr's between chunks.
268 * This helps in improving the fairness, reducing livelock (because we
269 * emulate more closely the "process to completion" that we have with
270 * fastforwarding) and accounting for the work performed in low level
271 * handling and forwarding.
274 static struct timeval poll_start_t;
281 /* XXX run at splhigh() or equivalent */
284 if (residual_burst > 0) {
285 schednetisr(NETISR_POLL);
286 /* will run immediately on return, followed by netisrs */
289 /* here we can account time spent in netisr's in this tick */
291 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
292 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
293 kern_load = (kern_load * hz) / 10000; /* 0..100 */
294 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
298 if (poll_burst < poll_burst_max)
303 if (pending_polls == 0) /* we are done */
307 * Last cycle was long and caused us to miss one or more
308 * hardclock ticks. Restart processing again, but slightly
309 * reduce the burst size to prevent that this happens again.
311 poll_burst -= (poll_burst / 8);
314 schednetisr(NETISR_POLL);
320 * netisr_poll is scheduled by schednetisr when appropriate, typically once
321 * per tick. It is called at splnet() so first thing to do is to upgrade to
322 * splimp(), and call all registered handlers.
327 static int reg_frac_count;
329 enum poll_cmd arg = POLL_ONLY;
333 if (residual_burst == 0) { /* first call in this tick */
334 microuptime(&poll_start_t);
336 * Check that paremeters are consistent with runtime
337 * variables. Some of these tests could be done at sysctl
338 * time, but the savings would be very limited because we
339 * still have to check against reg_frac_count and
340 * poll_each_burst. So, instead of writing separate sysctl
341 * handlers, we do all here.
346 else if (reg_frac < 1)
348 if (reg_frac_count > reg_frac)
349 reg_frac_count = reg_frac - 1;
350 if (reg_frac_count-- == 0) {
351 arg = POLL_AND_CHECK_STATUS;
352 reg_frac_count = reg_frac - 1;
354 if (poll_burst_max < MIN_POLL_BURST_MAX)
355 poll_burst_max = MIN_POLL_BURST_MAX;
356 else if (poll_burst_max > MAX_POLL_BURST_MAX)
357 poll_burst_max = MAX_POLL_BURST_MAX;
359 if (poll_each_burst < 1)
361 else if (poll_each_burst > poll_burst_max)
362 poll_each_burst = poll_burst_max;
364 residual_burst = poll_burst;
366 cycles = (residual_burst < poll_each_burst) ?
367 residual_burst : poll_each_burst;
368 residual_burst -= cycles;
371 for (i = 0 ; i < poll_handlers ; i++)
372 if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
373 (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
374 pr[i].handler(pr[i].ifp, arg, cycles);
375 } else { /* unregister */
376 for (i = 0 ; i < poll_handlers ; i++) {
378 pr[i].ifp->if_flags & IFF_RUNNING) {
379 pr[i].ifp->if_ipending &= ~IFF_POLLING;
380 pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1);
387 /* on -stable, schednetisr(NETISR_POLLMORE); */
393 * Try to register routine for polling. Returns 1 if successful
394 * (and polling should be enabled), 0 otherwise.
395 * A device is not supposed to register itself multiple times.
397 * This is called from within the *_intr() functions, so we do not need
401 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
405 if (polling == 0) /* polling disabled, cannot register */
407 if (h == NULL || ifp == NULL) /* bad arguments */
409 if ( !(ifp->if_flags & IFF_UP) ) /* must be up */
411 if (ifp->if_ipending & IFF_POLLING) /* already polling */
415 if (poll_handlers >= POLL_LIST_LEN) {
417 * List full, cannot register more entries.
418 * This should never happen; if it does, it is probably a
419 * broken driver trying to register multiple times. Checking
420 * this at runtime is expensive, and won't solve the problem
421 * anyways, so just report a few times and then give up.
423 static int verbose = 10 ;
426 printf("poll handlers list full, "
427 "maybe a broken driver ?\n");
430 return 0; /* no polling for you */
433 pr[poll_handlers].handler = h;
434 pr[poll_handlers].ifp = ifp;
436 ifp->if_ipending |= IFF_POLLING;
438 if (idlepoll_sleeping)
439 wakeup(&idlepoll_sleeping);
440 return 1; /* polling enabled in next call */
444 * Remove interface from the polling list. Normally called by *_stop().
445 * It is not an error to call it with IFF_POLLING clear, the call is
446 * sufficiently rare to be preferable to save the space for the extra
447 * test in each driver in exchange of one additional function call.
450 ether_poll_deregister(struct ifnet *ifp)
455 if ( !ifp || !(ifp->if_ipending & IFF_POLLING) ) {
459 for (i = 0 ; i < poll_handlers ; i++)
460 if (pr[i].ifp == ifp) /* found it */
462 ifp->if_ipending &= ~IFF_POLLING; /* found or not... */
463 if (i == poll_handlers) {
465 printf("ether_poll_deregister: ifp not found!!!\n");
469 if (i < poll_handlers) { /* Last entry replaces this one. */
470 pr[i].handler = pr[poll_handlers].handler;
471 pr[i].ifp = pr[poll_handlers].ifp;
480 struct thread *td = curthread;
484 rtp.prio = RTP_PRIO_MAX; /* lowest priority */
485 rtp.type = RTP_PRIO_IDLE;
486 mtx_lock_spin(&sched_lock);
487 rtp_to_pri(&rtp, td->td_ksegrp);
488 pri = td->td_priority;
489 mtx_unlock_spin(&sched_lock);
492 if (poll_in_idle_loop && poll_handlers > 0) {
493 idlepoll_sleeping = 0;
495 ether_poll(poll_each_burst);
497 mtx_assert(&Giant, MA_NOTOWNED);
498 mtx_lock_spin(&sched_lock);
500 td->td_proc->p_stats->p_ru.ru_nvcsw++;
502 mtx_unlock_spin(&sched_lock);
504 idlepoll_sleeping = 1;
505 tsleep(&idlepoll_sleeping, pri, "pollid", hz * 3);
510 static struct proc *idlepoll;
511 static struct kproc_desc idlepoll_kp = {
516 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start, &idlepoll_kp)