2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2001-2002 Luigi Rizzo
6 * Supported by: the Xorp Project (www.xorp.org)
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 #include "opt_device_polling.h"
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/kthread.h>
38 #include <sys/epoch.h>
39 #include <sys/eventhandler.h>
40 #include <sys/resourcevar.h>
41 #include <sys/socket.h> /* needed by net/if.h */
42 #include <sys/sockio.h>
43 #include <sys/sysctl.h>
44 #include <sys/syslog.h>
47 #include <net/if_var.h>
48 #include <net/netisr.h> /* for NETISR_POLL */
51 void hardclock_device_poll(void); /* hook from hardclock */
53 static struct mtx poll_mtx;
56 * Polling support for [network] device drivers.
58 * Drivers which support this feature can 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.
72 * The count limit specifies how much work the handler can do during the
73 * call -- typically this is the number of packets to be received, or
74 * transmitted, etc. (drivers are free to interpret this number, as long
75 * as the max time spent in the function grows roughly linearly with the
78 * Polling is enabled and disabled via setting IFCAP_POLLING flag on
79 * the interface. The driver ioctl handler should register interface
80 * with polling and disable interrupts, if registration was successful.
82 * A second variable controls the sharing of CPU between polling/kernel
83 * network processing, and other activities (typically userlevel tasks):
84 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
85 * of CPU allocated to user tasks. CPU is allocated proportionally to the
86 * shares, by dynamically adjusting the "count" (poll_burst).
88 * Other parameters can should be left to their default values.
89 * The following constraints hold
91 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
92 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
95 #define MIN_POLL_BURST_MAX 10
96 #define MAX_POLL_BURST_MAX 20000
98 static uint32_t poll_burst = 5;
99 static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
100 static uint32_t poll_each_burst = 5;
102 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103 "Device polling parameters");
105 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
106 &poll_burst, 0, "Current polling burst size");
108 static int netisr_poll_scheduled;
109 static int netisr_pollmore_scheduled;
110 static int poll_shutting_down;
112 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
114 uint32_t val = poll_burst_max;
117 error = sysctl_handle_int(oidp, &val, 0, req);
118 if (error || !req->newptr )
120 if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
124 poll_burst_max = val;
125 if (poll_burst > poll_burst_max)
126 poll_burst = poll_burst_max;
127 if (poll_each_burst > poll_burst_max)
128 poll_each_burst = MIN_POLL_BURST_MAX;
129 mtx_unlock(&poll_mtx);
133 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max,
134 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
135 poll_burst_max_sysctl, "I",
136 "Max Polling burst size");
138 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
140 uint32_t val = poll_each_burst;
143 error = sysctl_handle_int(oidp, &val, 0, req);
144 if (error || !req->newptr )
150 if (val > poll_burst_max) {
151 mtx_unlock(&poll_mtx);
154 poll_each_burst = val;
155 mtx_unlock(&poll_mtx);
159 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst,
160 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
161 poll_each_burst_sysctl, "I",
162 "Max size of each burst");
164 static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */
165 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
166 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
168 static uint32_t user_frac = 50;
169 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
171 uint32_t val = user_frac;
174 error = sysctl_handle_int(oidp, &val, 0, req);
175 if (error || !req->newptr )
182 mtx_unlock(&poll_mtx);
186 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac,
187 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
188 user_frac_sysctl, "I",
189 "Desired user fraction of cpu time");
191 static uint32_t reg_frac_count = 0;
192 static uint32_t reg_frac = 20 ;
193 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
195 uint32_t val = reg_frac;
198 error = sysctl_handle_int(oidp, &val, 0, req);
199 if (error || !req->newptr )
201 if (val < 1 || val > hz)
206 if (reg_frac_count >= reg_frac)
208 mtx_unlock(&poll_mtx);
212 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac,
213 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
214 reg_frac_sysctl, "I",
215 "Every this many cycles check registers");
217 static uint32_t short_ticks;
218 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
219 &short_ticks, 0, "Hardclock ticks shorter than they should be");
221 static uint32_t lost_polls;
222 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
223 &lost_polls, 0, "How many times we would have lost a poll tick");
225 static uint32_t pending_polls;
226 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
227 &pending_polls, 0, "Do we need to poll again");
229 static int residual_burst = 0;
230 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
231 &residual_burst, 0, "# of residual cycles in burst");
233 static uint32_t poll_handlers; /* next free entry in pr[]. */
234 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
235 &poll_handlers, 0, "Number of registered poll handlers");
237 static uint32_t phase;
238 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
239 &phase, 0, "Polling phase");
241 static uint32_t suspect;
242 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
243 &suspect, 0, "suspect event");
245 static uint32_t stalled;
246 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
247 &stalled, 0, "potential stalls");
249 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
250 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
251 &idlepoll_sleeping, 0, "idlepoll is sleeping");
253 #define POLL_LIST_LEN 128
255 poll_handler_t *handler;
259 static struct pollrec pr[POLL_LIST_LEN];
262 poll_shutdown(void *arg, int howto)
265 poll_shutting_down = 1;
269 init_device_poll(void)
272 mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
273 EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
276 SYSINIT(device_poll, SI_SUB_SOFTINTR, SI_ORDER_MIDDLE, init_device_poll, NULL);
279 * Hook from hardclock. Tries to schedule a netisr, but keeps track
280 * of lost ticks due to the previous handler taking too long.
281 * Normally, this should not happen, because polling handler should
282 * run for a short time. However, in some cases (e.g. when there are
283 * changes in link status etc.) the drivers take a very long time
284 * (even in the order of milliseconds) to reset and reconfigure the
285 * device, causing apparent lost polls.
287 * The first part of the code is just for debugging purposes, and tries
288 * to count how often hardclock ticks are shorter than they should,
289 * meaning either stray interrupts or delayed events.
292 hardclock_device_poll(void)
294 static struct timeval prev_t, t;
297 if (poll_handlers == 0 || poll_shutting_down)
301 delta = (t.tv_usec - prev_t.tv_usec) +
302 (t.tv_sec - prev_t.tv_sec)*1000000;
303 if (delta * hz < 500000)
308 if (pending_polls > 100) {
310 * Too much, assume it has stalled (not always true
311 * see comment above).
322 netisr_poll_scheduled = 1;
323 netisr_pollmore_scheduled = 1;
327 if (pending_polls++ > 0)
332 * ether_poll is called from the idle loop.
335 ether_poll(int count)
337 struct epoch_tracker et;
342 if (count > poll_each_burst)
343 count = poll_each_burst;
346 for (i = 0 ; i < poll_handlers ; i++)
347 pr[i].handler(pr[i].ifp, POLL_ONLY, count);
350 mtx_unlock(&poll_mtx);
354 * netisr_pollmore is called after other netisr's, possibly scheduling
355 * another NETISR_POLL call, or adapting the burst size for the next cycle.
357 * It is very bad to fetch large bursts of packets from a single card at once,
358 * because the burst could take a long time to be completely processed, or
359 * could saturate the intermediate queue (ipintrq or similar) leading to
360 * losses or unfairness. To reduce the problem, and also to account better for
361 * time spent in network-related processing, we split the burst in smaller
362 * chunks of fixed size, giving control to the other netisr's between chunks.
363 * This helps in improving the fairness, reducing livelock (because we
364 * emulate more closely the "process to completion" that we have with
365 * fastforwarding) and accounting for the work performed in low level
366 * handling and forwarding.
369 static struct timeval poll_start_t;
372 netisr_pollmore(void)
377 if (poll_handlers == 0)
381 if (!netisr_pollmore_scheduled) {
382 mtx_unlock(&poll_mtx);
385 netisr_pollmore_scheduled = 0;
387 if (residual_burst > 0) {
388 netisr_poll_scheduled = 1;
389 netisr_pollmore_scheduled = 1;
391 mtx_unlock(&poll_mtx);
392 /* will run immediately on return, followed by netisrs */
395 /* here we can account time spent in netisr's in this tick */
397 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
398 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
399 kern_load = (kern_load * hz) / 10000; /* 0..100 */
400 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
404 if (poll_burst < poll_burst_max)
409 if (pending_polls == 0) /* we are done */
413 * Last cycle was long and caused us to miss one or more
414 * hardclock ticks. Restart processing again, but slightly
415 * reduce the burst size to prevent that this happens again.
417 poll_burst -= (poll_burst / 8);
420 netisr_poll_scheduled = 1;
421 netisr_pollmore_scheduled = 1;
425 mtx_unlock(&poll_mtx);
429 * netisr_poll is typically scheduled once per tick.
435 enum poll_cmd arg = POLL_ONLY;
439 if (poll_handlers == 0)
443 if (!netisr_poll_scheduled) {
444 mtx_unlock(&poll_mtx);
447 netisr_poll_scheduled = 0;
449 if (residual_burst == 0) { /* first call in this tick */
450 microuptime(&poll_start_t);
451 if (++reg_frac_count == reg_frac) {
452 arg = POLL_AND_CHECK_STATUS;
456 residual_burst = poll_burst;
458 cycles = (residual_burst < poll_each_burst) ?
459 residual_burst : poll_each_burst;
460 residual_burst -= cycles;
462 for (i = 0 ; i < poll_handlers ; i++)
463 pr[i].handler(pr[i].ifp, arg, cycles);
466 mtx_unlock(&poll_mtx);
470 * Try to register routine for polling. Returns 0 if successful
471 * (and polling should be enabled), error code otherwise.
472 * A device is not supposed to register itself multiple times.
474 * This is called from within the *_ioctl() functions.
477 ether_poll_register(poll_handler_t *h, if_t ifp)
481 KASSERT(h != NULL, ("%s: handler is NULL", __func__));
482 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
485 if (poll_handlers >= POLL_LIST_LEN) {
487 * List full, cannot register more entries.
488 * This should never happen; if it does, it is probably a
489 * broken driver trying to register multiple times. Checking
490 * this at runtime is expensive, and won't solve the problem
491 * anyways, so just report a few times and then give up.
493 static int verbose = 10 ;
495 log(LOG_ERR, "poll handlers list full, "
496 "maybe a broken driver ?\n");
499 mtx_unlock(&poll_mtx);
500 return (ENOMEM); /* no polling for you */
503 for (i = 0 ; i < poll_handlers ; i++)
504 if (pr[i].ifp == ifp && pr[i].handler != NULL) {
505 mtx_unlock(&poll_mtx);
506 log(LOG_DEBUG, "ether_poll_register: %s: handler"
507 " already registered\n", ifp->if_xname);
511 pr[poll_handlers].handler = h;
512 pr[poll_handlers].ifp = ifp;
514 mtx_unlock(&poll_mtx);
515 if (idlepoll_sleeping)
516 wakeup(&idlepoll_sleeping);
521 * Remove interface from the polling list. Called from *_ioctl(), too.
524 ether_poll_deregister(if_t ifp)
528 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
532 for (i = 0 ; i < poll_handlers ; i++)
533 if (pr[i].ifp == ifp) /* found it */
535 if (i == poll_handlers) {
536 log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
538 mtx_unlock(&poll_mtx);
542 if (i < poll_handlers) { /* Last entry replaces this one. */
543 pr[i].handler = pr[poll_handlers].handler;
544 pr[i].ifp = pr[poll_handlers].ifp;
546 mtx_unlock(&poll_mtx);
553 struct thread *td = curthread;
556 rtp.prio = RTP_PRIO_MAX; /* lowest priority */
557 rtp.type = RTP_PRIO_IDLE;
558 PROC_SLOCK(td->td_proc);
559 rtp_to_pri(&rtp, td);
560 PROC_SUNLOCK(td->td_proc);
563 if (poll_in_idle_loop && poll_handlers > 0) {
564 idlepoll_sleeping = 0;
565 ether_poll(poll_each_burst);
569 idlepoll_sleeping = 1;
570 tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
575 static struct proc *idlepoll;
576 static struct kproc_desc idlepoll_kp = {
581 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,