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 1.18 2005/06/06 15:02:28 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 */
44 /* the two netisr handlers */
45 static int netisr_poll(struct netmsg *);
46 static int netisr_pollmore(struct netmsg *);
48 void init_device_poll(void); /* init routine */
49 void hardclock_device_poll(void); /* hook from hardclock */
52 * Polling support for [network] device drivers.
54 * Drivers which support this feature try to register with the
57 * If registration is successful, the driver must disable interrupts,
58 * and further I/O is performed through the handler, which is invoked
59 * (at least once per clock tick) with 3 arguments: the "arg" passed at
60 * register time (a struct ifnet pointer), a command, and a "count" limit.
62 * The command can be one of the following:
63 * POLL_ONLY: quick move of "count" packets from input/output queues.
64 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
65 * other more expensive operations. This command is issued periodically
66 * but less frequently than POLL_ONLY.
67 * POLL_DEREGISTER: deregister and return to interrupt mode.
68 * POLL_REGISTER: register and disable interrupts
70 * The first two commands are only issued if the interface is marked as
71 * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
73 * The count limit specifies how much work the handler can do during the
74 * call -- typically this is the number of packets to be received, or
75 * transmitted, etc. (drivers are free to interpret this number, as long
76 * as the max time spent in the function grows roughly linearly with the
79 * Deregistration can be requested by the driver itself (typically in the
80 * *_stop() routine), or by the polling code, by invoking the handler.
82 * Polling can be globally enabled or disabled with the sysctl variable
83 * kern.polling.enable (default is 0, disabled)
85 * A second variable controls the sharing of CPU between polling/kernel
86 * network processing, and other activities (typically userlevel tasks):
87 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
88 * of CPU allocated to user tasks. CPU is allocated proportionally to the
89 * shares, by dynamically adjusting the "count" (poll_burst).
91 * Other parameters can should be left to their default values.
92 * The following constraints hold
94 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
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 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
102 "Device polling parameters");
104 static u_int32_t poll_burst = 5;
105 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
106 &poll_burst, 0, "Current polling burst size");
108 static u_int32_t poll_each_burst = 5;
109 SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
110 &poll_each_burst, 0, "Max size of each burst");
112 static u_int32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
113 SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
114 &poll_burst_max, 0, "Max Polling burst size");
116 static u_int32_t user_frac = 50;
117 SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
118 &user_frac, 0, "Desired user fraction of cpu time");
120 static u_int32_t reg_frac = 20 ;
121 SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
122 ®_frac, 0, "Every this many cycles poll register");
124 static u_int32_t short_ticks;
125 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
126 &short_ticks, 0, "Hardclock ticks shorter than they should be");
128 static u_int32_t lost_polls;
129 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
130 &lost_polls, 0, "How many times we would have lost a poll tick");
132 static u_int32_t pending_polls;
133 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
134 &pending_polls, 0, "Do we need to poll again");
136 static int residual_burst = 0;
137 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
138 &residual_burst, 0, "# of residual cycles in burst");
140 static u_int32_t poll_handlers; /* next free entry in pr[]. */
141 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
142 &poll_handlers, 0, "Number of registered poll handlers");
144 static int polling = 0; /* global polling enable */
145 SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
146 &polling, 0, "Polling enabled");
148 static u_int32_t phase;
149 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
150 &phase, 0, "Polling phase");
152 static u_int32_t suspect;
153 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
154 &suspect, 0, "suspect event");
156 static u_int32_t stalled;
157 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
158 &stalled, 0, "potential stalls");
161 #define POLL_LIST_LEN 128
166 static struct pollrec pr[POLL_LIST_LEN];
169 * register relevant netisr. Called from kern_clock.c:
172 init_device_poll(void)
174 netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
175 netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
179 * Hook from hardclock. Tries to schedule a netisr, but keeps track
180 * of lost ticks due to the previous handler taking too long.
181 * Normally, this should not happen, because polling handler should
182 * run for a short time. However, in some cases (e.g. when there are
183 * changes in link status etc.) the drivers take a very long time
184 * (even in the order of milliseconds) to reset and reconfigure the
185 * device, causing apparent lost polls.
187 * The first part of the code is just for debugging purposes, and tries
188 * to count how often hardclock ticks are shorter than they should,
189 * meaning either stray interrupts or delayed events.
191 * WARNING! called from fastint or IPI, the MP lock might not be held.
194 hardclock_device_poll(void)
196 static struct timeval prev_t, t;
199 if (poll_handlers == 0)
203 delta = (t.tv_usec - prev_t.tv_usec) +
204 (t.tv_sec - prev_t.tv_sec)*1000000;
205 if (delta * hz < 500000)
210 if (pending_polls > 100) {
212 * Too much, assume it has stalled (not always true
213 * see comment above).
224 schednetisr(NETISR_POLL);
227 if (pending_polls++ > 0)
232 * netisr_pollmore is called after other netisr's, possibly scheduling
233 * another NETISR_POLL call, or adapting the burst size for the next cycle.
235 * It is very bad to fetch large bursts of packets from a single card at once,
236 * because the burst could take a long time to be completely processed, or
237 * could saturate the intermediate queue (ipintrq or similar) leading to
238 * losses or unfairness. To reduce the problem, and also to account better for
239 * time spent in network-related processing, we split the burst in smaller
240 * chunks of fixed size, giving control to the other netisr's between chunks.
241 * This helps in improving the fairness, reducing livelock (because we
242 * emulate more closely the "process to completion" that we have with
243 * fastforwarding) and accounting for the work performed in low level
244 * handling and forwarding.
247 static struct timeval poll_start_t;
251 netisr_pollmore(struct netmsg *msg)
257 lwkt_replymsg(&msg->nm_lmsg, 0);
259 if (residual_burst > 0) {
260 schednetisr(NETISR_POLL);
261 /* will run immediately on return, followed by netisrs */
264 /* here we can account time spent in netisr's in this tick */
266 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
267 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
268 kern_load = (kern_load * hz) / 10000; /* 0..100 */
269 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
273 if (poll_burst < poll_burst_max)
278 if (pending_polls == 0) { /* we are done */
282 * Last cycle was long and caused us to miss one or more
283 * hardclock ticks. Restart processing again, but slightly
284 * reduce the burst size to prevent that this happens again.
286 poll_burst -= (poll_burst / 8);
289 schednetisr(NETISR_POLL);
298 * netisr_poll is scheduled by schednetisr when appropriate, typically once
301 * Note that the message is replied immediately in order to allow a new
302 * ISR to be scheduled in the handler.
304 * XXX each registration should indicate whether it needs a critical
305 * section to operate.
309 netisr_poll(struct netmsg *msg)
311 static int reg_frac_count;
313 enum poll_cmd arg = POLL_ONLY;
315 lwkt_replymsg(&msg->nm_lmsg, 0);
318 if (residual_burst == 0) { /* first call in this tick */
319 microuptime(&poll_start_t);
321 * Check that paremeters are consistent with runtime
322 * variables. Some of these tests could be done at sysctl
323 * time, but the savings would be very limited because we
324 * still have to check against reg_frac_count and
325 * poll_each_burst. So, instead of writing separate sysctl
326 * handlers, we do all here.
331 else if (reg_frac < 1)
333 if (reg_frac_count > reg_frac)
334 reg_frac_count = reg_frac - 1;
335 if (reg_frac_count-- == 0) {
336 arg = POLL_AND_CHECK_STATUS;
337 reg_frac_count = reg_frac - 1;
339 if (poll_burst_max < MIN_POLL_BURST_MAX)
340 poll_burst_max = MIN_POLL_BURST_MAX;
341 else if (poll_burst_max > MAX_POLL_BURST_MAX)
342 poll_burst_max = MAX_POLL_BURST_MAX;
344 if (poll_each_burst < 1)
346 else if (poll_each_burst > poll_burst_max)
347 poll_each_burst = poll_burst_max;
349 residual_burst = poll_burst;
351 cycles = (residual_burst < poll_each_burst) ?
352 residual_burst : poll_each_burst;
353 residual_burst -= cycles;
356 for (i = 0 ; i < poll_handlers ; i++) {
357 struct pollrec *p = &pr[i];
358 if ((p->ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING)) == (IFF_UP|IFF_RUNNING|IFF_POLLING)) {
359 p->ifp->if_poll(p->ifp, arg, cycles);
362 } else { /* unregister */
363 for (i = 0 ; i < poll_handlers ; i++) {
364 struct pollrec *p = &pr[i];
365 if (p->ifp->if_flags & IFF_POLLING) {
366 p->ifp->if_flags &= ~IFF_POLLING;
368 * Only call the interface deregistration
369 * function if the interface is still
372 if (p->ifp->if_flags & IFF_RUNNING)
373 p->ifp->if_poll(p->ifp, POLL_DEREGISTER, 1);
379 schednetisr(NETISR_POLLMORE);
386 * Try to register routine for polling. Returns 1 if successful
387 * (and polling should be enabled), 0 otherwise.
389 * Called from mainline code only, not called from an interrupt.
392 ether_poll_register(struct ifnet *ifp)
396 if (polling == 0) /* polling disabled, cannot register */
398 if ((ifp->if_flags & IFF_UP) == 0) /* must be up */
400 if (ifp->if_flags & IFF_POLLING) /* already polling */
402 if (ifp->if_poll == NULL) /* no polling support */
406 * Attempt to register. Interlock with IFF_POLLING.
408 crit_enter(); /* XXX MP - not mp safe */
409 ifp->if_flags |= IFF_POLLING;
410 ifp->if_poll(ifp, POLL_REGISTER, 0);
411 if ((ifp->if_flags & IFF_POLLING) == 0) {
417 * Check if there is room. If there isn't, deregister.
419 if (poll_handlers >= POLL_LIST_LEN) {
421 * List full, cannot register more entries.
422 * This should never happen; if it does, it is probably a
423 * broken driver trying to register multiple times. Checking
424 * this at runtime is expensive, and won't solve the problem
425 * anyways, so just report a few times and then give up.
427 static int verbose = 10 ;
429 printf("poll handlers list full, "
430 "maybe a broken driver ?\n");
433 ifp->if_flags &= ~IFF_POLLING;
434 ifp->if_poll(ifp, POLL_DEREGISTER, 0);
437 pr[poll_handlers].ifp = ifp;
446 * Remove interface from the polling list. Occurs when polling is turned
447 * off. Called from mainline code only, not called from an interrupt.
450 ether_poll_deregister(struct ifnet *ifp)
455 if (ifp == NULL || (ifp->if_flags & IFF_POLLING) == 0) {
459 for (i = 0 ; i < poll_handlers ; i++) {
460 if (pr[i].ifp == ifp) /* found it */
463 ifp->if_flags &= ~IFF_POLLING; /* found or not... */
464 if (i == poll_handlers) {
466 printf("ether_poll_deregister: ifp not found!!!\n");
470 if (i < poll_handlers) { /* Last entry replaces this one. */
471 pr[i].ifp = pr[poll_handlers].ifp;
476 * Only call the deregistration function if the interface is still
479 if (ifp->if_flags & IFF_RUNNING)
480 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
485 emergency_poll_enable(const char *name)
489 printf("%s forced polling on\n", name);