2 * Copyright (c) 2001-2002 Luigi Rizzo
4 * Supported by: the Xorp Project (www.xorp.org)
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * $FreeBSD: src/sys/kern/kern_poll.c,v 1.2.2.4 2002/06/27 23:26:33 luigi Exp $
28 * $DragonFly: src/sys/kern/kern_poll.c,v 1.24 2006/12/23 00:35:04 swildner Exp $
31 #include "opt_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/sysctl.h>
39 #include <sys/thread2.h>
40 #include <sys/msgport2.h>
42 #include <machine/md_var.h> /* for vm_page_zero_idle() */
43 #include <net/if.h> /* for IFF_* flags */
44 #include <net/netisr.h> /* for NETISR_POLL */
46 /* the two netisr handlers */
47 static int sysctl_pollhz(SYSCTL_HANDLER_ARGS);
48 static int sysctl_polling(SYSCTL_HANDLER_ARGS);
49 static int netisr_poll(struct netmsg *);
50 static int netisr_pollmore(struct netmsg *);
51 static void pollclock(systimer_t, struct intrframe *);
53 void init_device_poll(void); /* init routine */
56 * Polling support for [network] device drivers.
58 * Drivers which support this feature try to register with the
61 * If registration is successful, the driver must disable interrupts,
62 * and further I/O is performed through the handler, which is invoked
63 * (at least once per clock tick) with 3 arguments: the "arg" passed at
64 * register time (a struct ifnet pointer), a command, and a "count" limit.
66 * The command can be one of the following:
67 * POLL_ONLY: quick move of "count" packets from input/output queues.
68 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
69 * other more expensive operations. This command is issued periodically
70 * but less frequently than POLL_ONLY.
71 * POLL_DEREGISTER: deregister and return to interrupt mode.
72 * POLL_REGISTER: register and disable interrupts
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 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
102 #define MIN_POLL_BURST_MAX 10
103 #define MAX_POLL_BURST_MAX 1000
105 #ifndef DEVICE_POLLING_FREQ_MAX
106 #define DEVICE_POLLING_FREQ_MAX 30000
108 #define DEVICE_POLLING_FREQ_DEFAULT 2000
110 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
111 "Device polling parameters");
113 static u_int32_t poll_burst = 5;
114 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
115 &poll_burst, 0, "Current polling burst size");
117 static u_int32_t poll_each_burst = 5;
118 SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
119 &poll_each_burst, 0, "Max size of each burst");
121 static u_int32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
122 SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
123 &poll_burst_max, 0, "Max Polling burst size");
125 static u_int32_t user_frac = 50;
126 SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
127 &user_frac, 0, "Desired user fraction of cpu time");
129 static u_int32_t reg_frac = 20 ;
130 SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
131 ®_frac, 0, "Every this many cycles poll register");
133 static u_int32_t short_ticks;
134 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
135 &short_ticks, 0, "Hardclock ticks shorter than they should be");
137 static u_int32_t lost_polls;
138 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
139 &lost_polls, 0, "How many times we would have lost a poll tick");
141 static u_int32_t pending_polls;
142 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
143 &pending_polls, 0, "Do we need to poll again");
145 static int residual_burst = 0;
146 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
147 &residual_burst, 0, "# of residual cycles in burst");
149 static u_int32_t poll_handlers; /* next free entry in pr[]. */
150 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
151 &poll_handlers, 0, "Number of registered poll handlers");
153 static int polling_enabled = 0; /* global polling enable */
154 TUNABLE_INT("kern.polling.enable", &polling_enabled);
155 SYSCTL_PROC(_kern_polling, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
156 0, 0, sysctl_polling, "I", "Polling enabled");
158 static u_int32_t phase;
159 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
160 &phase, 0, "Polling phase");
162 static u_int32_t suspect;
163 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
164 &suspect, 0, "suspect event");
166 static u_int32_t stalled;
167 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
168 &stalled, 0, "potential stalls");
170 static int pollhz = DEVICE_POLLING_FREQ_DEFAULT;
171 TUNABLE_INT("kern.polling.pollhz", &pollhz);
172 SYSCTL_PROC(_kern_polling, OID_AUTO, pollhz, CTLTYPE_INT | CTLFLAG_RW,
173 0, 0, sysctl_pollhz, "I", "Device polling frequency");
175 #define POLL_LIST_LEN 128
180 static struct pollrec pr[POLL_LIST_LEN];
181 static struct systimer gd0_pollclock;
184 * register relevant netisr. Called from kern_clock.c:
187 init_device_poll(void)
189 netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
190 netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
191 systimer_init_periodic_nq(&gd0_pollclock, pollclock, NULL,
192 polling_enabled ? pollhz : 1);
196 * Set the polling frequency
199 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
204 error = sysctl_handle_int(oidp, &phz, 0, req);
205 if (error || req->newptr == NULL)
209 else if (phz > DEVICE_POLLING_FREQ_MAX)
210 phz = DEVICE_POLLING_FREQ_MAX;
215 gd0_pollclock.periodic = sys_cputimer->fromhz(phz);
221 * Master enable. If polling is disabled, cut the polling systimer
225 sysctl_polling(SYSCTL_HANDLER_ARGS)
229 enabled = polling_enabled;
230 error = sysctl_handle_int(oidp, &enabled, 0, req);
231 if (error || req->newptr == NULL)
233 polling_enabled = enabled;
235 gd0_pollclock.periodic = sys_cputimer->fromhz(pollhz);
237 gd0_pollclock.periodic = sys_cputimer->fromhz(1);
242 * Hook from hardclock. Tries to schedule a netisr, but keeps track
243 * of lost ticks due to the previous handler taking too long.
244 * Normally, this should not happen, because polling handler should
245 * run for a short time. However, in some cases (e.g. when there are
246 * changes in link status etc.) the drivers take a very long time
247 * (even in the order of milliseconds) to reset and reconfigure the
248 * device, causing apparent lost polls.
250 * The first part of the code is just for debugging purposes, and tries
251 * to count how often hardclock ticks are shorter than they should,
252 * meaning either stray interrupts or delayed events.
254 * WARNING! called from fastint or IPI, the MP lock might not be held.
257 pollclock(systimer_t info __unused, struct intrframe *frame __unused)
259 static struct timeval prev_t, t;
262 if (poll_handlers == 0)
266 delta = (t.tv_usec - prev_t.tv_usec) +
267 (t.tv_sec - prev_t.tv_sec)*1000000;
268 if (delta * hz < 500000)
273 if (pending_polls > 100) {
275 * Too much, assume it has stalled (not always true
276 * see comment above).
287 schednetisr(NETISR_POLL);
290 if (pending_polls++ > 0)
295 * netisr_pollmore is called after other netisr's, possibly scheduling
296 * another NETISR_POLL call, or adapting the burst size for the next cycle.
298 * It is very bad to fetch large bursts of packets from a single card at once,
299 * because the burst could take a long time to be completely processed, or
300 * could saturate the intermediate queue (ipintrq or similar) leading to
301 * losses or unfairness. To reduce the problem, and also to account better for
302 * time spent in network-related processing, we split the burst in smaller
303 * chunks of fixed size, giving control to the other netisr's between chunks.
304 * This helps in improving the fairness, reducing livelock (because we
305 * emulate more closely the "process to completion" that we have with
306 * fastforwarding) and accounting for the work performed in low level
307 * handling and forwarding.
310 static struct timeval poll_start_t;
314 netisr_pollmore(struct netmsg *msg)
320 lwkt_replymsg(&msg->nm_lmsg, 0);
322 if (residual_burst > 0) {
323 schednetisr(NETISR_POLL);
324 /* will run immediately on return, followed by netisrs */
327 /* here we can account time spent in netisr's in this tick */
329 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
330 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
331 kern_load = (kern_load * hz) / 10000; /* 0..100 */
332 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
336 if (poll_burst < poll_burst_max)
341 if (pending_polls == 0) { /* we are done */
345 * Last cycle was long and caused us to miss one or more
346 * hardclock ticks. Restart processing again, but slightly
347 * reduce the burst size to prevent that this happens again.
349 poll_burst -= (poll_burst / 8);
352 schednetisr(NETISR_POLL);
361 * netisr_poll is scheduled by schednetisr when appropriate, typically once
364 * Note that the message is replied immediately in order to allow a new
365 * ISR to be scheduled in the handler.
367 * XXX each registration should indicate whether it needs a critical
368 * section to operate.
372 netisr_poll(struct netmsg *msg)
374 static int reg_frac_count;
376 enum poll_cmd arg = POLL_ONLY;
378 lwkt_replymsg(&msg->nm_lmsg, 0);
381 if (residual_burst == 0) { /* first call in this tick */
382 microuptime(&poll_start_t);
384 * Check that paremeters are consistent with runtime
385 * variables. Some of these tests could be done at sysctl
386 * time, but the savings would be very limited because we
387 * still have to check against reg_frac_count and
388 * poll_each_burst. So, instead of writing separate sysctl
389 * handlers, we do all here.
394 else if (reg_frac < 1)
396 if (reg_frac_count > reg_frac)
397 reg_frac_count = reg_frac - 1;
398 if (reg_frac_count-- == 0) {
399 arg = POLL_AND_CHECK_STATUS;
400 reg_frac_count = reg_frac - 1;
402 if (poll_burst_max < MIN_POLL_BURST_MAX)
403 poll_burst_max = MIN_POLL_BURST_MAX;
404 else if (poll_burst_max > MAX_POLL_BURST_MAX)
405 poll_burst_max = MAX_POLL_BURST_MAX;
407 if (poll_each_burst < 1)
409 else if (poll_each_burst > poll_burst_max)
410 poll_each_burst = poll_burst_max;
412 residual_burst = poll_burst;
414 cycles = (residual_burst < poll_each_burst) ?
415 residual_burst : poll_each_burst;
416 residual_burst -= cycles;
418 if (polling_enabled) {
419 for (i = 0 ; i < poll_handlers ; i++) {
420 struct pollrec *p = &pr[i];
421 if ((p->ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING)) == (IFF_UP|IFF_RUNNING|IFF_POLLING)) {
422 if (lwkt_serialize_try(p->ifp->if_serializer)) {
423 p->ifp->if_poll(p->ifp, arg, cycles);
424 lwkt_serialize_exit(p->ifp->if_serializer);
428 } else { /* unregister */
429 for (i = 0 ; i < poll_handlers ; i++) {
430 struct pollrec *p = &pr[i];
431 if ((p->ifp->if_flags & IFF_POLLING) == 0)
434 * Only call the interface deregistration
435 * function if the interface is still
438 lwkt_serialize_enter(p->ifp->if_serializer);
439 p->ifp->if_flags &= ~IFF_POLLING;
440 if (p->ifp->if_flags & IFF_RUNNING)
441 p->ifp->if_poll(p->ifp, POLL_DEREGISTER, 1);
442 lwkt_serialize_exit(p->ifp->if_serializer);
447 schednetisr(NETISR_POLLMORE);
454 * Try to register routine for polling. Returns 1 if successful
455 * (and polling should be enabled), 0 otherwise.
457 * Called from mainline code only, not called from an interrupt.
460 ether_poll_register(struct ifnet *ifp)
464 if (polling_enabled == 0) /* polling disabled, cannot register */
466 if ((ifp->if_flags & IFF_UP) == 0) /* must be up */
468 if (ifp->if_flags & IFF_POLLING) /* already polling */
470 if (ifp->if_poll == NULL) /* no polling support */
474 * Attempt to register. Interlock with IFF_POLLING.
476 crit_enter(); /* XXX MP - not mp safe */
477 lwkt_serialize_enter(ifp->if_serializer);
478 ifp->if_flags |= IFF_POLLING;
479 ifp->if_poll(ifp, POLL_REGISTER, 0);
480 lwkt_serialize_exit(ifp->if_serializer);
481 if ((ifp->if_flags & IFF_POLLING) == 0) {
487 * Check if there is room. If there isn't, deregister.
489 if (poll_handlers >= POLL_LIST_LEN) {
491 * List full, cannot register more entries.
492 * This should never happen; if it does, it is probably a
493 * broken driver trying to register multiple times. Checking
494 * this at runtime is expensive, and won't solve the problem
495 * anyways, so just report a few times and then give up.
497 static int verbose = 10 ;
499 kprintf("poll handlers list full, "
500 "maybe a broken driver ?\n");
503 ifp->if_flags &= ~IFF_POLLING;
504 lwkt_serialize_enter(ifp->if_serializer);
505 ifp->if_poll(ifp, POLL_DEREGISTER, 0);
506 lwkt_serialize_exit(ifp->if_serializer);
509 pr[poll_handlers].ifp = ifp;
518 * Remove interface from the polling list. Occurs when polling is turned
519 * off. Called from mainline code only, not called from an interrupt.
522 ether_poll_deregister(struct ifnet *ifp)
527 if (ifp == NULL || (ifp->if_flags & IFF_POLLING) == 0) {
531 for (i = 0 ; i < poll_handlers ; i++) {
532 if (pr[i].ifp == ifp) /* found it */
535 ifp->if_flags &= ~IFF_POLLING; /* found or not... */
536 if (i == poll_handlers) {
538 kprintf("ether_poll_deregister: ifp not found!!!\n");
542 if (i < poll_handlers) { /* Last entry replaces this one. */
543 pr[i].ifp = pr[poll_handlers].ifp;
548 * Only call the deregistration function if the interface is still
551 if (ifp->if_flags & IFF_RUNNING) {
552 lwkt_serialize_enter(ifp->if_serializer);
553 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
554 lwkt_serialize_exit(ifp->if_serializer);
560 emergency_poll_enable(const char *name)
562 if (polling_enabled == 0) {
564 kprintf("%s forced polling on\n", name);