2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
33 #include "opt_device_polling.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
40 #include <sys/epoch.h>
41 #include <sys/eventhandler.h>
42 #include <sys/resourcevar.h>
43 #include <sys/socket.h> /* needed by net/if.h */
44 #include <sys/sockio.h>
45 #include <sys/sysctl.h>
46 #include <sys/syslog.h>
49 #include <net/if_var.h>
50 #include <net/netisr.h> /* for NETISR_POLL */
53 void hardclock_device_poll(void); /* hook from hardclock */
55 static struct mtx poll_mtx;
58 * Polling support for [network] device drivers.
60 * Drivers which support this feature can register with the
63 * If registration is successful, the driver must disable interrupts,
64 * and further I/O is performed through the handler, which is invoked
65 * (at least once per clock tick) with 3 arguments: the "arg" passed at
66 * register time (a struct ifnet pointer), a command, and a "count" limit.
68 * The command can be one of the following:
69 * POLL_ONLY: quick move of "count" packets from input/output queues.
70 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
71 * other more expensive operations. This command is issued periodically
72 * but less frequently than POLL_ONLY.
74 * The count limit specifies how much work the handler can do during the
75 * call -- typically this is the number of packets to be received, or
76 * transmitted, etc. (drivers are free to interpret this number, as long
77 * as the max time spent in the function grows roughly linearly with the
80 * Polling is enabled and disabled via setting IFCAP_POLLING flag on
81 * the interface. The driver ioctl handler should register interface
82 * with polling and disable interrupts, if registration was successful.
84 * A second variable controls the sharing of CPU between polling/kernel
85 * network processing, and other activities (typically userlevel tasks):
86 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
87 * of CPU allocated to user tasks. CPU is allocated proportionally to the
88 * shares, by dynamically adjusting the "count" (poll_burst).
90 * Other parameters can should be left to their default values.
91 * The following constraints hold
93 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
94 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
97 #define MIN_POLL_BURST_MAX 10
98 #define MAX_POLL_BURST_MAX 20000
100 static uint32_t poll_burst = 5;
101 static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
102 static uint32_t poll_each_burst = 5;
104 static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
105 "Device polling parameters");
107 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
108 &poll_burst, 0, "Current polling burst size");
110 static int netisr_poll_scheduled;
111 static int netisr_pollmore_scheduled;
112 static int poll_shutting_down;
114 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
116 uint32_t val = poll_burst_max;
119 error = sysctl_handle_int(oidp, &val, 0, req);
120 if (error || !req->newptr )
122 if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
126 poll_burst_max = val;
127 if (poll_burst > poll_burst_max)
128 poll_burst = poll_burst_max;
129 if (poll_each_burst > poll_burst_max)
130 poll_each_burst = MIN_POLL_BURST_MAX;
131 mtx_unlock(&poll_mtx);
135 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max,
136 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
137 poll_burst_max_sysctl, "I",
138 "Max Polling burst size");
140 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
142 uint32_t val = poll_each_burst;
145 error = sysctl_handle_int(oidp, &val, 0, req);
146 if (error || !req->newptr )
152 if (val > poll_burst_max) {
153 mtx_unlock(&poll_mtx);
156 poll_each_burst = val;
157 mtx_unlock(&poll_mtx);
161 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst,
162 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
163 poll_each_burst_sysctl, "I",
164 "Max size of each burst");
166 static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */
167 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
168 &poll_in_idle_loop, 0, "Enable device polling in idle loop");
170 static uint32_t user_frac = 50;
171 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
173 uint32_t val = user_frac;
176 error = sysctl_handle_int(oidp, &val, 0, req);
177 if (error || !req->newptr )
184 mtx_unlock(&poll_mtx);
188 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac,
189 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
190 user_frac_sysctl, "I",
191 "Desired user fraction of cpu time");
193 static uint32_t reg_frac_count = 0;
194 static uint32_t reg_frac = 20 ;
195 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
197 uint32_t val = reg_frac;
200 error = sysctl_handle_int(oidp, &val, 0, req);
201 if (error || !req->newptr )
203 if (val < 1 || val > hz)
208 if (reg_frac_count >= reg_frac)
210 mtx_unlock(&poll_mtx);
214 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac,
215 CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, sizeof(uint32_t),
216 reg_frac_sysctl, "I",
217 "Every this many cycles check registers");
219 static uint32_t short_ticks;
220 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
221 &short_ticks, 0, "Hardclock ticks shorter than they should be");
223 static uint32_t lost_polls;
224 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
225 &lost_polls, 0, "How many times we would have lost a poll tick");
227 static uint32_t pending_polls;
228 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
229 &pending_polls, 0, "Do we need to poll again");
231 static int residual_burst = 0;
232 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
233 &residual_burst, 0, "# of residual cycles in burst");
235 static uint32_t poll_handlers; /* next free entry in pr[]. */
236 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
237 &poll_handlers, 0, "Number of registered poll handlers");
239 static uint32_t phase;
240 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
241 &phase, 0, "Polling phase");
243 static uint32_t suspect;
244 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
245 &suspect, 0, "suspect event");
247 static uint32_t stalled;
248 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
249 &stalled, 0, "potential stalls");
251 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
252 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
253 &idlepoll_sleeping, 0, "idlepoll is sleeping");
255 #define POLL_LIST_LEN 128
257 poll_handler_t *handler;
261 static struct pollrec pr[POLL_LIST_LEN];
264 poll_shutdown(void *arg, int howto)
267 poll_shutting_down = 1;
271 init_device_poll(void)
274 mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
275 EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
278 SYSINIT(device_poll, SI_SUB_SOFTINTR, SI_ORDER_MIDDLE, init_device_poll, NULL);
281 * Hook from hardclock. Tries to schedule a netisr, but keeps track
282 * of lost ticks due to the previous handler taking too long.
283 * Normally, this should not happen, because polling handler should
284 * run for a short time. However, in some cases (e.g. when there are
285 * changes in link status etc.) the drivers take a very long time
286 * (even in the order of milliseconds) to reset and reconfigure the
287 * device, causing apparent lost polls.
289 * The first part of the code is just for debugging purposes, and tries
290 * to count how often hardclock ticks are shorter than they should,
291 * meaning either stray interrupts or delayed events.
294 hardclock_device_poll(void)
296 static struct timeval prev_t, t;
299 if (poll_handlers == 0 || poll_shutting_down)
303 delta = (t.tv_usec - prev_t.tv_usec) +
304 (t.tv_sec - prev_t.tv_sec)*1000000;
305 if (delta * hz < 500000)
310 if (pending_polls > 100) {
312 * Too much, assume it has stalled (not always true
313 * see comment above).
324 netisr_poll_scheduled = 1;
325 netisr_pollmore_scheduled = 1;
329 if (pending_polls++ > 0)
334 * ether_poll is called from the idle loop.
337 ether_poll(int count)
339 struct epoch_tracker et;
344 if (count > poll_each_burst)
345 count = poll_each_burst;
348 for (i = 0 ; i < poll_handlers ; i++)
349 pr[i].handler(pr[i].ifp, POLL_ONLY, count);
352 mtx_unlock(&poll_mtx);
356 * netisr_pollmore is called after other netisr's, possibly scheduling
357 * another NETISR_POLL call, or adapting the burst size for the next cycle.
359 * It is very bad to fetch large bursts of packets from a single card at once,
360 * because the burst could take a long time to be completely processed, or
361 * could saturate the intermediate queue (ipintrq or similar) leading to
362 * losses or unfairness. To reduce the problem, and also to account better for
363 * time spent in network-related processing, we split the burst in smaller
364 * chunks of fixed size, giving control to the other netisr's between chunks.
365 * This helps in improving the fairness, reducing livelock (because we
366 * emulate more closely the "process to completion" that we have with
367 * fastforwarding) and accounting for the work performed in low level
368 * handling and forwarding.
371 static struct timeval poll_start_t;
379 if (poll_handlers == 0)
383 if (!netisr_pollmore_scheduled) {
384 mtx_unlock(&poll_mtx);
387 netisr_pollmore_scheduled = 0;
389 if (residual_burst > 0) {
390 netisr_poll_scheduled = 1;
391 netisr_pollmore_scheduled = 1;
393 mtx_unlock(&poll_mtx);
394 /* will run immediately on return, followed by netisrs */
397 /* here we can account time spent in netisr's in this tick */
399 kern_load = (t.tv_usec - poll_start_t.tv_usec) +
400 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
401 kern_load = (kern_load * hz) / 10000; /* 0..100 */
402 if (kern_load > (100 - user_frac)) { /* try decrease ticks */
406 if (poll_burst < poll_burst_max)
411 if (pending_polls == 0) /* we are done */
415 * Last cycle was long and caused us to miss one or more
416 * hardclock ticks. Restart processing again, but slightly
417 * reduce the burst size to prevent that this happens again.
419 poll_burst -= (poll_burst / 8);
422 netisr_poll_scheduled = 1;
423 netisr_pollmore_scheduled = 1;
427 mtx_unlock(&poll_mtx);
431 * netisr_poll is typically scheduled once per tick.
437 enum poll_cmd arg = POLL_ONLY;
441 if (poll_handlers == 0)
445 if (!netisr_poll_scheduled) {
446 mtx_unlock(&poll_mtx);
449 netisr_poll_scheduled = 0;
451 if (residual_burst == 0) { /* first call in this tick */
452 microuptime(&poll_start_t);
453 if (++reg_frac_count == reg_frac) {
454 arg = POLL_AND_CHECK_STATUS;
458 residual_burst = poll_burst;
460 cycles = (residual_burst < poll_each_burst) ?
461 residual_burst : poll_each_burst;
462 residual_burst -= cycles;
464 for (i = 0 ; i < poll_handlers ; i++)
465 pr[i].handler(pr[i].ifp, arg, cycles);
468 mtx_unlock(&poll_mtx);
472 * Try to register routine for polling. Returns 0 if successful
473 * (and polling should be enabled), error code otherwise.
474 * A device is not supposed to register itself multiple times.
476 * This is called from within the *_ioctl() functions.
479 ether_poll_register(poll_handler_t *h, if_t ifp)
483 KASSERT(h != NULL, ("%s: handler is NULL", __func__));
484 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
487 if (poll_handlers >= POLL_LIST_LEN) {
489 * List full, cannot register more entries.
490 * This should never happen; if it does, it is probably a
491 * broken driver trying to register multiple times. Checking
492 * this at runtime is expensive, and won't solve the problem
493 * anyways, so just report a few times and then give up.
495 static int verbose = 10 ;
497 log(LOG_ERR, "poll handlers list full, "
498 "maybe a broken driver ?\n");
501 mtx_unlock(&poll_mtx);
502 return (ENOMEM); /* no polling for you */
505 for (i = 0 ; i < poll_handlers ; i++)
506 if (pr[i].ifp == ifp && pr[i].handler != NULL) {
507 mtx_unlock(&poll_mtx);
508 log(LOG_DEBUG, "ether_poll_register: %s: handler"
509 " already registered\n", ifp->if_xname);
513 pr[poll_handlers].handler = h;
514 pr[poll_handlers].ifp = ifp;
516 mtx_unlock(&poll_mtx);
517 if (idlepoll_sleeping)
518 wakeup(&idlepoll_sleeping);
523 * Remove interface from the polling list. Called from *_ioctl(), too.
526 ether_poll_deregister(if_t ifp)
530 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
534 for (i = 0 ; i < poll_handlers ; i++)
535 if (pr[i].ifp == ifp) /* found it */
537 if (i == poll_handlers) {
538 log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
540 mtx_unlock(&poll_mtx);
544 if (i < poll_handlers) { /* Last entry replaces this one. */
545 pr[i].handler = pr[poll_handlers].handler;
546 pr[i].ifp = pr[poll_handlers].ifp;
548 mtx_unlock(&poll_mtx);
555 struct thread *td = curthread;
558 rtp.prio = RTP_PRIO_MAX; /* lowest priority */
559 rtp.type = RTP_PRIO_IDLE;
560 PROC_SLOCK(td->td_proc);
561 rtp_to_pri(&rtp, td);
562 PROC_SUNLOCK(td->td_proc);
565 if (poll_in_idle_loop && poll_handlers > 0) {
566 idlepoll_sleeping = 0;
567 ether_poll(poll_each_burst);
571 idlepoll_sleeping = 1;
572 tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
577 static struct proc *idlepoll;
578 static struct kproc_desc idlepoll_kp = {
583 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,