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
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
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/eventhandler.h>
39 #include <sys/resourcevar.h>
40 #include <sys/socket.h> /* needed by net/if.h */
41 #include <sys/sockio.h>
42 #include <sys/sysctl.h>
43 #include <sys/syslog.h>
46 #include <net/if_var.h>
47 #include <net/netisr.h> /* for NETISR_POLL */
50 void hardclock_device_poll(void); /* hook from hardclock */
52 static struct mtx poll_mtx
;
55 * Polling support for [network] device drivers.
57 * Drivers which support this feature can register with the
60 * If registration is successful, the driver must disable interrupts,
61 * and further I/O is performed through the handler, which is invoked
62 * (at least once per clock tick) with 3 arguments: the "arg" passed at
63 * register time (a struct ifnet pointer), a command, and a "count" limit.
65 * The command can be one of the following:
66 * POLL_ONLY: quick move of "count" packets from input/output queues.
67 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do
68 * other more expensive operations. This command is issued periodically
69 * but less frequently than POLL_ONLY.
71 * The count limit specifies how much work the handler can do during the
72 * call -- typically this is the number of packets to be received, or
73 * transmitted, etc. (drivers are free to interpret this number, as long
74 * as the max time spent in the function grows roughly linearly with the
77 * Polling is enabled and disabled via setting IFCAP_POLLING flag on
78 * the interface. The driver ioctl handler should register interface
79 * with polling and disable interrupts, if registration was successful.
81 * A second variable controls the sharing of CPU between polling/kernel
82 * network processing, and other activities (typically userlevel tasks):
83 * kern.polling.user_frac (between 0 and 100, default 50) sets the share
84 * of CPU allocated to user tasks. CPU is allocated proportionally to the
85 * shares, by dynamically adjusting the "count" (poll_burst).
87 * Other parameters can should be left to their default values.
88 * The following constraints hold
90 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max
91 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
94 #define MIN_POLL_BURST_MAX 10
95 #define MAX_POLL_BURST_MAX 20000
97 static uint32_t poll_burst
= 5;
98 static uint32_t poll_burst_max
= 150; /* good for 100Mbit net and HZ=1000 */
99 static uint32_t poll_each_burst
= 5;
101 static SYSCTL_NODE(_kern
, OID_AUTO
, polling
, CTLFLAG_RW
, 0,
102 "Device polling parameters");
104 SYSCTL_UINT(_kern_polling
, OID_AUTO
, burst
, CTLFLAG_RD
,
105 &poll_burst
, 0, "Current polling burst size");
107 static int netisr_poll_scheduled
;
108 static int netisr_pollmore_scheduled
;
109 static int poll_shutting_down
;
111 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS
)
113 uint32_t val
= poll_burst_max
;
116 error
= sysctl_handle_int(oidp
, &val
, 0, req
);
117 if (error
|| !req
->newptr
)
119 if (val
< MIN_POLL_BURST_MAX
|| val
> MAX_POLL_BURST_MAX
)
123 poll_burst_max
= val
;
124 if (poll_burst
> poll_burst_max
)
125 poll_burst
= poll_burst_max
;
126 if (poll_each_burst
> poll_burst_max
)
127 poll_each_burst
= MIN_POLL_BURST_MAX
;
128 mtx_unlock(&poll_mtx
);
132 SYSCTL_PROC(_kern_polling
, OID_AUTO
, burst_max
, CTLTYPE_UINT
| CTLFLAG_RW
,
133 0, sizeof(uint32_t), poll_burst_max_sysctl
, "I", "Max Polling burst size");
135 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS
)
137 uint32_t val
= poll_each_burst
;
140 error
= sysctl_handle_int(oidp
, &val
, 0, req
);
141 if (error
|| !req
->newptr
)
147 if (val
> poll_burst_max
) {
148 mtx_unlock(&poll_mtx
);
151 poll_each_burst
= val
;
152 mtx_unlock(&poll_mtx
);
156 SYSCTL_PROC(_kern_polling
, OID_AUTO
, each_burst
, CTLTYPE_UINT
| CTLFLAG_RW
,
157 0, sizeof(uint32_t), poll_each_burst_sysctl
, "I",
158 "Max size of each burst");
160 static uint32_t poll_in_idle_loop
=0; /* do we poll in idle loop ? */
161 SYSCTL_UINT(_kern_polling
, OID_AUTO
, idle_poll
, CTLFLAG_RW
,
162 &poll_in_idle_loop
, 0, "Enable device polling in idle loop");
164 static uint32_t user_frac
= 50;
165 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS
)
167 uint32_t val
= user_frac
;
170 error
= sysctl_handle_int(oidp
, &val
, 0, req
);
171 if (error
|| !req
->newptr
)
178 mtx_unlock(&poll_mtx
);
182 SYSCTL_PROC(_kern_polling
, OID_AUTO
, user_frac
, CTLTYPE_UINT
| CTLFLAG_RW
,
183 0, sizeof(uint32_t), user_frac_sysctl
, "I",
184 "Desired user fraction of cpu time");
186 static uint32_t reg_frac_count
= 0;
187 static uint32_t reg_frac
= 20 ;
188 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS
)
190 uint32_t val
= reg_frac
;
193 error
= sysctl_handle_int(oidp
, &val
, 0, req
);
194 if (error
|| !req
->newptr
)
196 if (val
< 1 || val
> hz
)
201 if (reg_frac_count
>= reg_frac
)
203 mtx_unlock(&poll_mtx
);
207 SYSCTL_PROC(_kern_polling
, OID_AUTO
, reg_frac
, CTLTYPE_UINT
| CTLFLAG_RW
,
208 0, sizeof(uint32_t), reg_frac_sysctl
, "I",
209 "Every this many cycles check registers");
211 static uint32_t short_ticks
;
212 SYSCTL_UINT(_kern_polling
, OID_AUTO
, short_ticks
, CTLFLAG_RD
,
213 &short_ticks
, 0, "Hardclock ticks shorter than they should be");
215 static uint32_t lost_polls
;
216 SYSCTL_UINT(_kern_polling
, OID_AUTO
, lost_polls
, CTLFLAG_RD
,
217 &lost_polls
, 0, "How many times we would have lost a poll tick");
219 static uint32_t pending_polls
;
220 SYSCTL_UINT(_kern_polling
, OID_AUTO
, pending_polls
, CTLFLAG_RD
,
221 &pending_polls
, 0, "Do we need to poll again");
223 static int residual_burst
= 0;
224 SYSCTL_INT(_kern_polling
, OID_AUTO
, residual_burst
, CTLFLAG_RD
,
225 &residual_burst
, 0, "# of residual cycles in burst");
227 static uint32_t poll_handlers
; /* next free entry in pr[]. */
228 SYSCTL_UINT(_kern_polling
, OID_AUTO
, handlers
, CTLFLAG_RD
,
229 &poll_handlers
, 0, "Number of registered poll handlers");
231 static uint32_t phase
;
232 SYSCTL_UINT(_kern_polling
, OID_AUTO
, phase
, CTLFLAG_RD
,
233 &phase
, 0, "Polling phase");
235 static uint32_t suspect
;
236 SYSCTL_UINT(_kern_polling
, OID_AUTO
, suspect
, CTLFLAG_RD
,
237 &suspect
, 0, "suspect event");
239 static uint32_t stalled
;
240 SYSCTL_UINT(_kern_polling
, OID_AUTO
, stalled
, CTLFLAG_RD
,
241 &stalled
, 0, "potential stalls");
243 static uint32_t idlepoll_sleeping
; /* idlepoll is sleeping */
244 SYSCTL_UINT(_kern_polling
, OID_AUTO
, idlepoll_sleeping
, CTLFLAG_RD
,
245 &idlepoll_sleeping
, 0, "idlepoll is sleeping");
248 #define POLL_LIST_LEN 128
250 poll_handler_t
*handler
;
254 static struct pollrec pr
[POLL_LIST_LEN
];
257 poll_shutdown(void *arg
, int howto
)
260 poll_shutting_down
= 1;
264 init_device_poll(void)
267 mtx_init(&poll_mtx
, "polling", NULL
, MTX_DEF
);
268 EVENTHANDLER_REGISTER(shutdown_post_sync
, poll_shutdown
, NULL
,
271 SYSINIT(device_poll
, SI_SUB_SOFTINTR
, SI_ORDER_MIDDLE
, init_device_poll
, NULL
);
275 * Hook from hardclock. Tries to schedule a netisr, but keeps track
276 * of lost ticks due to the previous handler taking too long.
277 * Normally, this should not happen, because polling handler should
278 * run for a short time. However, in some cases (e.g. when there are
279 * changes in link status etc.) the drivers take a very long time
280 * (even in the order of milliseconds) to reset and reconfigure the
281 * device, causing apparent lost polls.
283 * The first part of the code is just for debugging purposes, and tries
284 * to count how often hardclock ticks are shorter than they should,
285 * meaning either stray interrupts or delayed events.
288 hardclock_device_poll(void)
290 static struct timeval prev_t
, t
;
293 if (poll_handlers
== 0 || poll_shutting_down
)
297 delta
= (t
.tv_usec
- prev_t
.tv_usec
) +
298 (t
.tv_sec
- prev_t
.tv_sec
)*1000000;
299 if (delta
* hz
< 500000)
304 if (pending_polls
> 100) {
306 * Too much, assume it has stalled (not always true
307 * see comment above).
318 netisr_poll_scheduled
= 1;
319 netisr_pollmore_scheduled
= 1;
323 if (pending_polls
++ > 0)
328 * ether_poll is called from the idle loop.
331 ether_poll(int count
)
337 if (count
> poll_each_burst
)
338 count
= poll_each_burst
;
340 for (i
= 0 ; i
< poll_handlers
; i
++)
341 pr
[i
].handler(pr
[i
].ifp
, POLL_ONLY
, count
);
343 mtx_unlock(&poll_mtx
);
347 * netisr_pollmore is called after other netisr's, possibly scheduling
348 * another NETISR_POLL call, or adapting the burst size for the next cycle.
350 * It is very bad to fetch large bursts of packets from a single card at once,
351 * because the burst could take a long time to be completely processed, or
352 * could saturate the intermediate queue (ipintrq or similar) leading to
353 * losses or unfairness. To reduce the problem, and also to account better for
354 * time spent in network-related processing, we split the burst in smaller
355 * chunks of fixed size, giving control to the other netisr's between chunks.
356 * This helps in improving the fairness, reducing livelock (because we
357 * emulate more closely the "process to completion" that we have with
358 * fastforwarding) and accounting for the work performed in low level
359 * handling and forwarding.
362 static struct timeval poll_start_t
;
370 if (poll_handlers
== 0)
374 if (!netisr_pollmore_scheduled
) {
375 mtx_unlock(&poll_mtx
);
378 netisr_pollmore_scheduled
= 0;
380 if (residual_burst
> 0) {
381 netisr_poll_scheduled
= 1;
382 netisr_pollmore_scheduled
= 1;
384 mtx_unlock(&poll_mtx
);
385 /* will run immediately on return, followed by netisrs */
388 /* here we can account time spent in netisr's in this tick */
390 kern_load
= (t
.tv_usec
- poll_start_t
.tv_usec
) +
391 (t
.tv_sec
- poll_start_t
.tv_sec
)*1000000; /* us */
392 kern_load
= (kern_load
* hz
) / 10000; /* 0..100 */
393 if (kern_load
> (100 - user_frac
)) { /* try decrease ticks */
397 if (poll_burst
< poll_burst_max
)
402 if (pending_polls
== 0) /* we are done */
406 * Last cycle was long and caused us to miss one or more
407 * hardclock ticks. Restart processing again, but slightly
408 * reduce the burst size to prevent that this happens again.
410 poll_burst
-= (poll_burst
/ 8);
413 netisr_poll_scheduled
= 1;
414 netisr_pollmore_scheduled
= 1;
418 mtx_unlock(&poll_mtx
);
422 * netisr_poll is typically scheduled once per tick.
428 enum poll_cmd arg
= POLL_ONLY
;
430 if (poll_handlers
== 0)
434 if (!netisr_poll_scheduled
) {
435 mtx_unlock(&poll_mtx
);
438 netisr_poll_scheduled
= 0;
440 if (residual_burst
== 0) { /* first call in this tick */
441 microuptime(&poll_start_t
);
442 if (++reg_frac_count
== reg_frac
) {
443 arg
= POLL_AND_CHECK_STATUS
;
447 residual_burst
= poll_burst
;
449 cycles
= (residual_burst
< poll_each_burst
) ?
450 residual_burst
: poll_each_burst
;
451 residual_burst
-= cycles
;
453 for (i
= 0 ; i
< poll_handlers
; i
++)
454 pr
[i
].handler(pr
[i
].ifp
, arg
, cycles
);
457 mtx_unlock(&poll_mtx
);
461 * Try to register routine for polling. Returns 0 if successful
462 * (and polling should be enabled), error code otherwise.
463 * A device is not supposed to register itself multiple times.
465 * This is called from within the *_ioctl() functions.
468 ether_poll_register(poll_handler_t
*h
, if_t ifp
)
472 KASSERT(h
!= NULL
, ("%s: handler is NULL", __func__
));
473 KASSERT(ifp
!= NULL
, ("%s: ifp is NULL", __func__
));
476 if (poll_handlers
>= POLL_LIST_LEN
) {
478 * List full, cannot register more entries.
479 * This should never happen; if it does, it is probably a
480 * broken driver trying to register multiple times. Checking
481 * this at runtime is expensive, and won't solve the problem
482 * anyways, so just report a few times and then give up.
484 static int verbose
= 10 ;
486 log(LOG_ERR
, "poll handlers list full, "
487 "maybe a broken driver ?\n");
490 mtx_unlock(&poll_mtx
);
491 return (ENOMEM
); /* no polling for you */
494 for (i
= 0 ; i
< poll_handlers
; i
++)
495 if (pr
[i
].ifp
== ifp
&& pr
[i
].handler
!= NULL
) {
496 mtx_unlock(&poll_mtx
);
497 log(LOG_DEBUG
, "ether_poll_register: %s: handler"
498 " already registered\n", ifp
->if_xname
);
502 pr
[poll_handlers
].handler
= h
;
503 pr
[poll_handlers
].ifp
= ifp
;
505 mtx_unlock(&poll_mtx
);
506 if (idlepoll_sleeping
)
507 wakeup(&idlepoll_sleeping
);
512 * Remove interface from the polling list. Called from *_ioctl(), too.
515 ether_poll_deregister(if_t ifp
)
519 KASSERT(ifp
!= NULL
, ("%s: ifp is NULL", __func__
));
523 for (i
= 0 ; i
< poll_handlers
; i
++)
524 if (pr
[i
].ifp
== ifp
) /* found it */
526 if (i
== poll_handlers
) {
527 log(LOG_DEBUG
, "ether_poll_deregister: %s: not found!\n",
529 mtx_unlock(&poll_mtx
);
533 if (i
< poll_handlers
) { /* Last entry replaces this one. */
534 pr
[i
].handler
= pr
[poll_handlers
].handler
;
535 pr
[i
].ifp
= pr
[poll_handlers
].ifp
;
537 mtx_unlock(&poll_mtx
);
544 struct thread
*td
= curthread
;
547 rtp
.prio
= RTP_PRIO_MAX
; /* lowest priority */
548 rtp
.type
= RTP_PRIO_IDLE
;
549 PROC_SLOCK(td
->td_proc
);
550 rtp_to_pri(&rtp
, td
);
551 PROC_SUNLOCK(td
->td_proc
);
554 if (poll_in_idle_loop
&& poll_handlers
> 0) {
555 idlepoll_sleeping
= 0;
556 ether_poll(poll_each_burst
);
558 mi_switch(SW_VOL
, NULL
);
561 idlepoll_sleeping
= 1;
562 tsleep(&idlepoll_sleeping
, 0, "pollid", hz
* 3);
567 static struct proc
*idlepoll
;
568 static struct kproc_desc idlepoll_kp
= {
573 SYSINIT(idlepoll
, SI_SUB_KTHREAD_VM
, SI_ORDER_ANY
, kproc_start
,