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 1.12 2004/04/21 18:13:47 dillon Exp $
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/socket.h> /* needed by net/if.h */
35 #include <sys/sysctl.h>
37 #include <sys/thread2.h>
38 #include <sys/msgport2.h>
40 #include <i386/include/md_var.h> /* for vm_page_zero_idle() */
41 #include <net/if.h> /* for IFF_* flags */
42 #include <net/netisr.h> /* for NETISR_POLL */
46 #ifndef COMPILING_LINT
47 #error DEVICE_POLLING is not compatible with SMP
51 /* the two netisr handlers */
52 static int netisr_poll(struct netmsg *);
53 static int netisr_pollmore(struct netmsg *);
55 void init_device_poll(void); /* init routine */
56 void hardclock_device_poll(void); /* hook from hardclock */
57 void ether_poll(int); /* polling while in trap */
58 int idle_poll(void); /* poll while in idle loop */
61 * Polling support for [network] device drivers.
63 * Drivers which support this feature try to register with the
66 * If registration is successful, the driver must disable interrupts,
67 * and further I/O is performed through the handler, which is invoked
68 * (at least once per clock tick) with 3 arguments: the "arg" passed at
69 * register time (a struct ifnet pointer), a command, and a "count" limit.
71 * The command can be one of the following:
72 * POLL_ONLY: quick move of "count" packets from input/output queues.
73 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
74 * other more expensive operations. This command is issued periodically
75 * but less frequently than POLL_ONLY.
76 * POLL_DEREGISTER: deregister and return to interrupt mode.
78 * The first two commands are only issued if the interface is marked as
79 * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
81 * The count limit specifies how much work the handler can do during the
82 * call -- typically this is the number of packets to be received, or
83 * transmitted, etc. (drivers are free to interpret this number, as long
84 * as the max time spent in the function grows roughly linearly with the
87 * Deregistration can be requested by the driver itself (typically in the
88 * *_stop() routine), or by the polling code, by invoking the handler.
90 * Polling can be globally enabled or disabled with the sysctl variable
91 * kern.polling.enable (default is 0, disabled)
93 * A second variable controls the sharing of CPU between polling/kernel
94 * network processing, and other activities (typically userlevel tasks):
95 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
96 * of CPU allocated to user tasks. CPU is allocated proportionally to the
97 * shares, by dynamically adjusting the "count" (poll_burst).
99 * Other parameters can should be left to their default values.
100 * The following constraints hold
102 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
103 * 0 <= poll_in_trap <= poll_each_burst
104 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
107 #define MIN_POLL_BURST_MAX 10
108 #define MAX_POLL_BURST_MAX 1000
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 poll_in_idle_loop=1; /* do we poll in idle loop ? */
126 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
127 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
129 u_int32_t poll_in_trap; /* used in trap.c */
130 SYSCTL_UINT(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW,
131 &poll_in_trap, 0, "Poll burst size during a trap");
133 static u_int32_t user_frac = 50;
134 SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
135 &user_frac, 0, "Desired user fraction of cpu time");
137 static u_int32_t reg_frac = 20 ;
138 SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
139 ®_frac, 0, "Every this many cycles poll register");
141 static u_int32_t short_ticks;
142 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
143 &short_ticks, 0, "Hardclock ticks shorter than they should be");
145 static u_int32_t lost_polls;
146 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
147 &lost_polls, 0, "How many times we would have lost a poll tick");
149 static u_int32_t pending_polls;
150 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
151 &pending_polls, 0, "Do we need to poll again");
153 static int residual_burst = 0;
154 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
155 &residual_burst, 0, "# of residual cycles in burst");
157 static u_int32_t poll_handlers; /* next free entry in pr[]. */
158 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
159 &poll_handlers, 0, "Number of registered poll handlers");
161 static int polling = 0; /* global polling enable */
162 SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
163 &polling, 0, "Polling enabled");
165 static u_int32_t phase;
166 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
167 &phase, 0, "Polling phase");
169 static u_int32_t suspect;
170 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
171 &suspect, 0, "suspect event");
173 static u_int32_t stalled;
174 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
175 &stalled, 0, "potential stalls");
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 netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
193 netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
197 * Hook from hardclock. Tries to schedule a netisr, but keeps track
198 * of lost ticks due to the previous handler taking too long.
199 * Normally, this should not happen, because polling handler should
200 * run for a short time. However, in some cases (e.g. when there are
201 * changes in link status etc.) the drivers take a very long time
202 * (even in the order of milliseconds) to reset and reconfigure the
203 * device, causing apparent lost polls.
205 * The first part of the code is just for debugging purposes, and tries
206 * to count how often hardclock ticks are shorter than they should,
207 * meaning either stray interrupts or delayed events.
209 * WARNING! called from fastint or IPI, the MP lock might not be held.
212 hardclock_device_poll(void)
214 static struct timeval prev_t, t;
217 if (poll_handlers == 0)
221 delta = (t.tv_usec - prev_t.tv_usec) +
222 (t.tv_sec - prev_t.tv_sec)*1000000;
223 if (delta * hz < 500000)
228 if (pending_polls > 100) {
230 * Too much, assume it has stalled (not always true
231 * see comment above).
242 schednetisr(NETISR_POLL);
245 if (pending_polls++ > 0)
250 * ether_poll is called from the idle loop or from the trap handler.
253 ether_poll(int count)
258 if (count > poll_each_burst)
259 count = poll_each_burst;
260 for (i = 0 ; i < poll_handlers ; i++)
261 if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
262 (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
263 pr[i].handler(pr[i].ifp, 0, count); /* quick check */
268 * idle_poll is replaces the body of the idle loop when DEVICE_POLLING
269 * is used. YYY not currently implemented.
274 if (poll_in_idle_loop && poll_handlers > 0) {
277 ether_poll(poll_each_burst);
283 return vm_page_zero_idle();
287 * netisr_pollmore is called after other netisr's, possibly scheduling
288 * another NETISR_POLL call, or adapting the burst size for the next cycle.
290 * It is very bad to fetch large bursts of packets from a single card at once,
291 * because the burst could take a long time to be completely processed, or
292 * could saturate the intermediate queue (ipintrq or similar) leading to
293 * losses or unfairness. To reduce the problem, and also to account better for
294 * time spent in network-related processing, we split the burst in smaller
295 * chunks of fixed size, giving control to the other netisr's between chunks.
296 * This helps in improving the fairness, reducing livelock (because we
297 * emulate more closely the "process to completion" that we have with
298 * fastforwarding) and accounting for the work performed in low level
299 * handling and forwarding.
302 static struct timeval poll_start_t;
306 netisr_pollmore(struct netmsg *msg)
313 if (residual_burst > 0) {
314 schednetisr(NETISR_POLL);
315 /* will run immediately on return, followed by netisrs */
318 /* here we can account time spent in netisr's in this tick */
320 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
321 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
322 kern_load = (kern_load * hz) / 10000; /* 0..100 */
323 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
327 if (poll_burst < poll_burst_max)
332 if (pending_polls == 0) /* we are done */
336 * Last cycle was long and caused us to miss one or more
337 * hardclock ticks. Restart processing again, but slightly
338 * reduce the burst size to prevent that this happens again.
340 poll_burst -= (poll_burst / 8);
343 schednetisr(NETISR_POLL);
348 lwkt_replymsg(&msg->nm_lmsg, 0);
353 * netisr_poll is scheduled by schednetisr when appropriate, typically once
354 * per tick. It is called at splnet() so first thing to do is to upgrade to
355 * splimp(), and call all registered handlers.
359 netisr_poll(struct netmsg *msg)
361 static int reg_frac_count;
363 enum poll_cmd arg = POLL_ONLY;
367 if (residual_burst == 0) { /* first call in this tick */
368 microuptime(&poll_start_t);
370 * Check that paremeters are consistent with runtime
371 * variables. Some of these tests could be done at sysctl
372 * time, but the savings would be very limited because we
373 * still have to check against reg_frac_count and
374 * poll_each_burst. So, instead of writing separate sysctl
375 * handlers, we do all here.
380 else if (reg_frac < 1)
382 if (reg_frac_count > reg_frac)
383 reg_frac_count = reg_frac - 1;
384 if (reg_frac_count-- == 0) {
385 arg = POLL_AND_CHECK_STATUS;
386 reg_frac_count = reg_frac - 1;
388 if (poll_burst_max < MIN_POLL_BURST_MAX)
389 poll_burst_max = MIN_POLL_BURST_MAX;
390 else if (poll_burst_max > MAX_POLL_BURST_MAX)
391 poll_burst_max = MAX_POLL_BURST_MAX;
393 if (poll_each_burst < 1)
395 else if (poll_each_burst > poll_burst_max)
396 poll_each_burst = poll_burst_max;
398 residual_burst = poll_burst;
400 cycles = (residual_burst < poll_each_burst) ?
401 residual_burst : poll_each_burst;
402 residual_burst -= cycles;
405 for (i = 0 ; i < poll_handlers ; i++)
406 if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
407 (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
408 pr[i].handler(pr[i].ifp, arg, cycles);
409 } else { /* unregister */
410 for (i = 0 ; i < poll_handlers ; i++) {
412 pr[i].ifp->if_flags & IFF_RUNNING) {
413 pr[i].ifp->if_flags &= ~IFF_POLLING;
414 pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1);
421 schednetisr(NETISR_POLLMORE);
424 lwkt_replymsg(&msg->nm_lmsg, 0);
429 * Try to register routine for polling. Returns 1 if successful
430 * (and polling should be enabled), 0 otherwise.
431 * A device is not supposed to register itself multiple times.
433 * This is called from within the *_intr() functions, so we do not need
437 ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
441 if (polling == 0) /* polling disabled, cannot register */
443 if (h == NULL || ifp == NULL) /* bad arguments */
445 if ( !(ifp->if_flags & IFF_UP) ) /* must be up */
447 if (ifp->if_flags & IFF_POLLING) /* already polling */
451 if (poll_handlers >= POLL_LIST_LEN) {
453 * List full, cannot register more entries.
454 * This should never happen; if it does, it is probably a
455 * broken driver trying to register multiple times. Checking
456 * this at runtime is expensive, and won't solve the problem
457 * anyways, so just report a few times and then give up.
459 static int verbose = 10 ;
462 printf("poll handlers list full, "
463 "maybe a broken driver ?\n");
466 return 0; /* no polling for you */
469 pr[poll_handlers].handler = h;
470 pr[poll_handlers].ifp = ifp;
472 ifp->if_flags |= IFF_POLLING;
474 return 1; /* polling enabled in next call */
478 * Remove interface from the polling list. Normally called by *_stop().
479 * It is not an error to call it with IFF_POLLING clear, the call is
480 * sufficiently rare to be preferable to save the space for the extra
481 * test in each driver in exchange of one additional function call.
484 ether_poll_deregister(struct ifnet *ifp)
489 if ( !ifp || !(ifp->if_flags & IFF_POLLING) ) {
493 for (i = 0 ; i < poll_handlers ; i++)
494 if (pr[i].ifp == ifp) /* found it */
496 ifp->if_flags &= ~IFF_POLLING; /* found or not... */
497 if (i == poll_handlers) {
499 printf("ether_poll_deregister: ifp not found!!!\n");
503 if (i < poll_handlers) { /* Last entry replaces this one. */
504 pr[i].handler = pr[poll_handlers].handler;
505 pr[i].ifp = pr[poll_handlers].ifp;