1
0
mirror of https://git.FreeBSD.org/src.git synced 2024-12-28 11:57:28 +00:00
freebsd/sys/kern/kern_poll.c
Luigi Rizzo 4b62214f4a Increase the (arbitrary) limit for the number of packets per tick
from 1k to 20k The previous value was good 10 years ago, but not
anymore now.

More importantly, lots of good surprises:
polling is incredibly effective under virtualization, and not only
prevents livelock but also saves most of the VM exit overhead in
receive mode.

Using polling, a FreeBSD instance under qemu-kvm remains perfectly
responsive even when bombed with 10 Mpps over an emulated e1000,
and happily processes 1.7 Mpps through ipfw.

Note that some incompatibilities still remain: e.g. polling is not
(yet) compatible with netmap, and seems to freeze the guest when
kern.polling.idle_poll=1

MFC after:	3 days
2013-05-22 16:32:18 +00:00

568 lines
16 KiB
C

/*-
* Copyright (c) 2001-2002 Luigi Rizzo
*
* Supported by: the Xorp Project (www.xorp.org)
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_device_polling.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/kthread.h>
#include <sys/proc.h>
#include <sys/eventhandler.h>
#include <sys/resourcevar.h>
#include <sys/socket.h> /* needed by net/if.h */
#include <sys/sockio.h>
#include <sys/sysctl.h>
#include <sys/syslog.h>
#include <net/if.h> /* for IFF_* flags */
#include <net/netisr.h> /* for NETISR_POLL */
#include <net/vnet.h>
void hardclock_device_poll(void); /* hook from hardclock */
static struct mtx poll_mtx;
/*
* Polling support for [network] device drivers.
*
* Drivers which support this feature can register with the
* polling code.
*
* If registration is successful, the driver must disable interrupts,
* and further I/O is performed through the handler, which is invoked
* (at least once per clock tick) with 3 arguments: the "arg" passed at
* register time (a struct ifnet pointer), a command, and a "count" limit.
*
* The command can be one of the following:
* POLL_ONLY: quick move of "count" packets from input/output queues.
* POLL_AND_CHECK_STATUS: as above, plus check status registers or do
* other more expensive operations. This command is issued periodically
* but less frequently than POLL_ONLY.
*
* The count limit specifies how much work the handler can do during the
* call -- typically this is the number of packets to be received, or
* transmitted, etc. (drivers are free to interpret this number, as long
* as the max time spent in the function grows roughly linearly with the
* count).
*
* Polling is enabled and disabled via setting IFCAP_POLLING flag on
* the interface. The driver ioctl handler should register interface
* with polling and disable interrupts, if registration was successful.
*
* A second variable controls the sharing of CPU between polling/kernel
* network processing, and other activities (typically userlevel tasks):
* kern.polling.user_frac (between 0 and 100, default 50) sets the share
* of CPU allocated to user tasks. CPU is allocated proportionally to the
* shares, by dynamically adjusting the "count" (poll_burst).
*
* Other parameters can should be left to their default values.
* The following constraints hold
*
* 1 <= poll_each_burst <= poll_burst <= poll_burst_max
* MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
*/
#define MIN_POLL_BURST_MAX 10
#define MAX_POLL_BURST_MAX 20000
static uint32_t poll_burst = 5;
static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
static uint32_t poll_each_burst = 5;
static SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
"Device polling parameters");
SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD,
&poll_burst, 0, "Current polling burst size");
static int netisr_poll_scheduled;
static int netisr_pollmore_scheduled;
static int poll_shutting_down;
static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS)
{
uint32_t val = poll_burst_max;
int error;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error || !req->newptr )
return (error);
if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX)
return (EINVAL);
mtx_lock(&poll_mtx);
poll_burst_max = val;
if (poll_burst > poll_burst_max)
poll_burst = poll_burst_max;
if (poll_each_burst > poll_burst_max)
poll_each_burst = MIN_POLL_BURST_MAX;
mtx_unlock(&poll_mtx);
return (0);
}
SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW,
0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size");
static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS)
{
uint32_t val = poll_each_burst;
int error;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error || !req->newptr )
return (error);
if (val < 1)
return (EINVAL);
mtx_lock(&poll_mtx);
if (val > poll_burst_max) {
mtx_unlock(&poll_mtx);
return (EINVAL);
}
poll_each_burst = val;
mtx_unlock(&poll_mtx);
return (0);
}
SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW,
0, sizeof(uint32_t), poll_each_burst_sysctl, "I",
"Max size of each burst");
static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */
SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
&poll_in_idle_loop, 0, "Enable device polling in idle loop");
static uint32_t user_frac = 50;
static int user_frac_sysctl(SYSCTL_HANDLER_ARGS)
{
uint32_t val = user_frac;
int error;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error || !req->newptr )
return (error);
if (val < 0 || val > 99)
return (EINVAL);
mtx_lock(&poll_mtx);
user_frac = val;
mtx_unlock(&poll_mtx);
return (0);
}
SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW,
0, sizeof(uint32_t), user_frac_sysctl, "I",
"Desired user fraction of cpu time");
static uint32_t reg_frac_count = 0;
static uint32_t reg_frac = 20 ;
static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS)
{
uint32_t val = reg_frac;
int error;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error || !req->newptr )
return (error);
if (val < 1 || val > hz)
return (EINVAL);
mtx_lock(&poll_mtx);
reg_frac = val;
if (reg_frac_count >= reg_frac)
reg_frac_count = 0;
mtx_unlock(&poll_mtx);
return (0);
}
SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW,
0, sizeof(uint32_t), reg_frac_sysctl, "I",
"Every this many cycles check registers");
static uint32_t short_ticks;
SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD,
&short_ticks, 0, "Hardclock ticks shorter than they should be");
static uint32_t lost_polls;
SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD,
&lost_polls, 0, "How many times we would have lost a poll tick");
static uint32_t pending_polls;
SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD,
&pending_polls, 0, "Do we need to poll again");
static int residual_burst = 0;
SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD,
&residual_burst, 0, "# of residual cycles in burst");
static uint32_t poll_handlers; /* next free entry in pr[]. */
SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
&poll_handlers, 0, "Number of registered poll handlers");
static uint32_t phase;
SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD,
&phase, 0, "Polling phase");
static uint32_t suspect;
SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD,
&suspect, 0, "suspect event");
static uint32_t stalled;
SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD,
&stalled, 0, "potential stalls");
static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */
SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD,
&idlepoll_sleeping, 0, "idlepoll is sleeping");
#define POLL_LIST_LEN 128
struct pollrec {
poll_handler_t *handler;
struct ifnet *ifp;
};
static struct pollrec pr[POLL_LIST_LEN];
static void
poll_shutdown(void *arg, int howto)
{
poll_shutting_down = 1;
}
static void
init_device_poll(void)
{
mtx_init(&poll_mtx, "polling", NULL, MTX_DEF);
EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL,
SHUTDOWN_PRI_LAST);
}
SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL);
/*
* Hook from hardclock. Tries to schedule a netisr, but keeps track
* of lost ticks due to the previous handler taking too long.
* Normally, this should not happen, because polling handler should
* run for a short time. However, in some cases (e.g. when there are
* changes in link status etc.) the drivers take a very long time
* (even in the order of milliseconds) to reset and reconfigure the
* device, causing apparent lost polls.
*
* The first part of the code is just for debugging purposes, and tries
* to count how often hardclock ticks are shorter than they should,
* meaning either stray interrupts or delayed events.
*/
void
hardclock_device_poll(void)
{
static struct timeval prev_t, t;
int delta;
if (poll_handlers == 0 || poll_shutting_down)
return;
microuptime(&t);
delta = (t.tv_usec - prev_t.tv_usec) +
(t.tv_sec - prev_t.tv_sec)*1000000;
if (delta * hz < 500000)
short_ticks++;
else
prev_t = t;
if (pending_polls > 100) {
/*
* Too much, assume it has stalled (not always true
* see comment above).
*/
stalled++;
pending_polls = 0;
phase = 0;
}
if (phase <= 2) {
if (phase != 0)
suspect++;
phase = 1;
netisr_poll_scheduled = 1;
netisr_pollmore_scheduled = 1;
netisr_sched_poll();
phase = 2;
}
if (pending_polls++ > 0)
lost_polls++;
}
/*
* ether_poll is called from the idle loop.
*/
static void
ether_poll(int count)
{
int i;
mtx_lock(&poll_mtx);
if (count > poll_each_burst)
count = poll_each_burst;
for (i = 0 ; i < poll_handlers ; i++)
pr[i].handler(pr[i].ifp, POLL_ONLY, count);
mtx_unlock(&poll_mtx);
}
/*
* netisr_pollmore is called after other netisr's, possibly scheduling
* another NETISR_POLL call, or adapting the burst size for the next cycle.
*
* It is very bad to fetch large bursts of packets from a single card at once,
* because the burst could take a long time to be completely processed, or
* could saturate the intermediate queue (ipintrq or similar) leading to
* losses or unfairness. To reduce the problem, and also to account better for
* time spent in network-related processing, we split the burst in smaller
* chunks of fixed size, giving control to the other netisr's between chunks.
* This helps in improving the fairness, reducing livelock (because we
* emulate more closely the "process to completion" that we have with
* fastforwarding) and accounting for the work performed in low level
* handling and forwarding.
*/
static struct timeval poll_start_t;
void
netisr_pollmore()
{
struct timeval t;
int kern_load;
mtx_lock(&poll_mtx);
if (!netisr_pollmore_scheduled) {
mtx_unlock(&poll_mtx);
return;
}
netisr_pollmore_scheduled = 0;
phase = 5;
if (residual_burst > 0) {
netisr_poll_scheduled = 1;
netisr_pollmore_scheduled = 1;
netisr_sched_poll();
mtx_unlock(&poll_mtx);
/* will run immediately on return, followed by netisrs */
return;
}
/* here we can account time spent in netisr's in this tick */
microuptime(&t);
kern_load = (t.tv_usec - poll_start_t.tv_usec) +
(t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */
kern_load = (kern_load * hz) / 10000; /* 0..100 */
if (kern_load > (100 - user_frac)) { /* try decrease ticks */
if (poll_burst > 1)
poll_burst--;
} else {
if (poll_burst < poll_burst_max)
poll_burst++;
}
pending_polls--;
if (pending_polls == 0) /* we are done */
phase = 0;
else {
/*
* Last cycle was long and caused us to miss one or more
* hardclock ticks. Restart processing again, but slightly
* reduce the burst size to prevent that this happens again.
*/
poll_burst -= (poll_burst / 8);
if (poll_burst < 1)
poll_burst = 1;
netisr_poll_scheduled = 1;
netisr_pollmore_scheduled = 1;
netisr_sched_poll();
phase = 6;
}
mtx_unlock(&poll_mtx);
}
/*
* netisr_poll is typically scheduled once per tick.
*/
void
netisr_poll(void)
{
int i, cycles;
enum poll_cmd arg = POLL_ONLY;
mtx_lock(&poll_mtx);
if (!netisr_poll_scheduled) {
mtx_unlock(&poll_mtx);
return;
}
netisr_poll_scheduled = 0;
phase = 3;
if (residual_burst == 0) { /* first call in this tick */
microuptime(&poll_start_t);
if (++reg_frac_count == reg_frac) {
arg = POLL_AND_CHECK_STATUS;
reg_frac_count = 0;
}
residual_burst = poll_burst;
}
cycles = (residual_burst < poll_each_burst) ?
residual_burst : poll_each_burst;
residual_burst -= cycles;
for (i = 0 ; i < poll_handlers ; i++)
pr[i].handler(pr[i].ifp, arg, cycles);
phase = 4;
mtx_unlock(&poll_mtx);
}
/*
* Try to register routine for polling. Returns 0 if successful
* (and polling should be enabled), error code otherwise.
* A device is not supposed to register itself multiple times.
*
* This is called from within the *_ioctl() functions.
*/
int
ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
{
int i;
KASSERT(h != NULL, ("%s: handler is NULL", __func__));
KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
mtx_lock(&poll_mtx);
if (poll_handlers >= POLL_LIST_LEN) {
/*
* List full, cannot register more entries.
* This should never happen; if it does, it is probably a
* broken driver trying to register multiple times. Checking
* this at runtime is expensive, and won't solve the problem
* anyways, so just report a few times and then give up.
*/
static int verbose = 10 ;
if (verbose >0) {
log(LOG_ERR, "poll handlers list full, "
"maybe a broken driver ?\n");
verbose--;
}
mtx_unlock(&poll_mtx);
return (ENOMEM); /* no polling for you */
}
for (i = 0 ; i < poll_handlers ; i++)
if (pr[i].ifp == ifp && pr[i].handler != NULL) {
mtx_unlock(&poll_mtx);
log(LOG_DEBUG, "ether_poll_register: %s: handler"
" already registered\n", ifp->if_xname);
return (EEXIST);
}
pr[poll_handlers].handler = h;
pr[poll_handlers].ifp = ifp;
poll_handlers++;
mtx_unlock(&poll_mtx);
if (idlepoll_sleeping)
wakeup(&idlepoll_sleeping);
return (0);
}
/*
* Remove interface from the polling list. Called from *_ioctl(), too.
*/
int
ether_poll_deregister(struct ifnet *ifp)
{
int i;
KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__));
mtx_lock(&poll_mtx);
for (i = 0 ; i < poll_handlers ; i++)
if (pr[i].ifp == ifp) /* found it */
break;
if (i == poll_handlers) {
log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n",
ifp->if_xname);
mtx_unlock(&poll_mtx);
return (ENOENT);
}
poll_handlers--;
if (i < poll_handlers) { /* Last entry replaces this one. */
pr[i].handler = pr[poll_handlers].handler;
pr[i].ifp = pr[poll_handlers].ifp;
}
mtx_unlock(&poll_mtx);
return (0);
}
static void
poll_idle(void)
{
struct thread *td = curthread;
struct rtprio rtp;
rtp.prio = RTP_PRIO_MAX; /* lowest priority */
rtp.type = RTP_PRIO_IDLE;
PROC_SLOCK(td->td_proc);
rtp_to_pri(&rtp, td);
PROC_SUNLOCK(td->td_proc);
for (;;) {
if (poll_in_idle_loop && poll_handlers > 0) {
idlepoll_sleeping = 0;
ether_poll(poll_each_burst);
thread_lock(td);
mi_switch(SW_VOL, NULL);
thread_unlock(td);
} else {
idlepoll_sleeping = 1;
tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3);
}
}
}
static struct proc *idlepoll;
static struct kproc_desc idlepoll_kp = {
"idlepoll",
poll_idle,
&idlepoll
};
SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start,
&idlepoll_kp);