diff options
Diffstat (limited to 'share/man/man9/scheduler.9')
| -rw-r--r-- | share/man/man9/scheduler.9 | 276 |
1 files changed, 276 insertions, 0 deletions
diff --git a/share/man/man9/scheduler.9 b/share/man/man9/scheduler.9 new file mode 100644 index 000000000000..c56fd0b6c653 --- /dev/null +++ b/share/man/man9/scheduler.9 @@ -0,0 +1,276 @@ +.\" Copyright (c) 2000-2001 John H. Baldwin <jhb@FreeBSD.org> +.\" All rights reserved. +.\" +.\" 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 DEVELOPERS ``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 DEVELOPERS 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. +.\" +.\" $FreeBSD$ +.\" +.Dd November 3, 2000 +.Dt SCHEDULER 9 +.Os +.Sh NAME +.Nm curpriority_cmp , +.Nm maybe_resched , +.Nm resetpriority , +.Nm roundrobin , +.Nm roundrobin_interval , +.Nm sched_setup , +.Nm schedclock , +.Nm schedcpu , +.Nm setrunnable , +.Nm updatepri +.Nd perform round-robin scheduling of runnable processes +.Sh SYNOPSIS +.In sys/param.h +.In sys/proc.h +.Ft int +.Fn curpriority_cmp "struct proc *p" +.Ft void +.Fn maybe_resched "struct thread *td" +.Ft void +.Fn propagate_priority "struct proc *p" +.Ft void +.Fn resetpriority "struct ksegrp *kg" +.Ft void +.Fn roundrobin "void *arg" +.Ft int +.Fn roundrobin_interval "void" +.Ft void +.Fn sched_setup "void *dummy" +.Ft void +.Fn schedclock "struct thread *td" +.Ft void +.Fn schedcpu "void *arg" +.Ft void +.Fn setrunnable "struct thread *td" +.Ft void +.Fn updatepri "struct thread *td" +.Sh DESCRIPTION +Each process has three different priorities stored in +.Vt "struct proc" : +.Va p_usrpri , +.Va p_nativepri , +and +.Va p_priority . +.Pp +The +.Va p_usrpri +member is the user priority of the process calculated from a process' +estimated CPU time and nice level. +.Pp +The +.Va p_nativepri +member is the saved priority used by +.Fn propagate_priority . +When a process obtains a mutex, its priority is saved in +.Va p_nativepri . +While it holds the mutex, the process's priority may be bumped by another +process that blocks on the mutex. +When the process releases the mutex, then its priority is restored to the +priority saved in +.Va p_nativepri . +.Pp +The +.Va p_priority +member is the actual priority of the process and is used to determine what +.Xr runqueue 9 +it runs on, for example. +.Pp +The +.Fn curpriority_cmp +function compares the cached priority of the currently running process with +process +.Fa p . +If the currently running process has a higher priority, then it will return +a value less than zero. +If the current process has a lower priority, then it will return a value +greater than zero. +If the current process has the same priority as +.Fa p , +then +.Fn curpriority_cmp +will return zero. +The cached priority of the currently running process is updated when a process +resumes from +.Xr tsleep 9 +or returns to userland in +.Fn userret +and is stored in the private variable +.Va curpriority . +.Pp +The +.Fn maybe_resched +function compares the priorities of the current thread and +.Fa td . +If +.Fa td +has a higher priority than the current thread, then a context switch is +needed, and +.Dv KEF_NEEDRESCHED +is set. +.Pp +The +.Fn propagate_priority +looks at the process that owns the mutex +.Fa p +is blocked on. +That process's priority is bumped to the priority of +.Fa p +if needed. +If the process is currently running, then the function returns. +If the process is on a +.Xr runqueue 9 , +then the process is moved to the appropriate +.Xr runqueue 9 +for its new priority. +If the process is blocked on a mutex, its position in the list of +processes blocked on the mutex in question is updated to reflect its new +priority. +Then, the function repeats the procedure using the process that owns the +mutex just encountered. +Note that a process's priorities are only bumped to the priority of the +original process +.Fa p , +not to the priority of the previously encountered process. +.Pp +The +.Fn resetpriority +function recomputes the user priority of the ksegrp +.Fa kg +(stored in +.Va kg_user_pri ) +and calls +.Fn maybe_resched +to force a reschedule of each thread in the group if needed. +.Pp +The +.Fn roundrobin +function is used as a +.Xr timeout 9 +function to force a reschedule every +.Va sched_quantum +ticks. +.Pp +The +.Fn roundrobin_interval +function simply returns the number of clock ticks in between reschedules +triggered by +.Fn roundrobin . +Thus, all it does is return the current value of +.Va sched_quantum . +.Pp +The +.Fn sched_setup +function is a +.Xr SYSINIT 9 +that is called to start the callout driven scheduler functions. +It just calls the +.Fn roundrobin +and +.Fn schedcpu +functions for the first time. +After the initial call, the two functions will propagate themselves by +registering their callout event again at the completion of the respective +function. +.Pp +The +.Fn schedclock +function is called by +.Fn statclock +to adjust the priority of the currently running thread's ksegrp. +It updates the group's estimated CPU time and then adjusts the priority via +.Fn resetpriority . +.Pp +The +.Fn schedcpu +function updates all process priorities. +First, it updates statistics that track how long processes have been in various +process states. +Secondly, it updates the estimated CPU time for the current process such +that about 90% of the CPU usage is forgotten in 5 * load average seconds. +For example, if the load average is 2.00, +then at least 90% of the estimated CPU time for the process should be based +on the amount of CPU time the process has had in the last 10 seconds. +It then recomputes the priority of the process and moves it to the +appropriate +.Xr runqueue 9 +if necessary. +Thirdly, it updates the %CPU estimate used by utilities such as +.Xr ps 1 +and +.Xr top 1 +so that 95% of the CPU usage is forgotten in 60 seconds. +Once all process priorities have been updated, +.Fn schedcpu +calls +.Fn vmmeter +to update various other statistics including the load average. +Finally, it schedules itself to run again in +.Va hz +clock ticks. +.Pp +The +.Fn setrunnable +function is used to change a process's state to be runnable. +The process is placed on a +.Xr runqueue 9 +if needed, and the swapper process is woken up and told to swap the process in +if the process is swapped out. +If the process has been asleep for at least one run of +.Fn schedcpu , +then +.Fn updatepri +is used to adjust the priority of the process. +.Pp +The +.Fn updatepri +function is used to adjust the priority of a process that has been asleep. +It retroactively decays the estimated CPU time of the process for each +.Fn schedcpu +event that the process was asleep. +Finally, it calls +.Fn resetpriority +to adjust the priority of the process. +.Sh SEE ALSO +.Xr mi_switch 9 , +.Xr runqueue 9 , +.Xr sleepqueue 9 , +.Xr tsleep 9 +.Sh BUGS +The +.Va curpriority +variable really should be per-CPU. +In addition, +.Fn maybe_resched +should compare the priority of +.Fa chk +with that of each CPU, and then send an IPI to the processor with the lowest +priority to trigger a reschedule if needed. +.Pp +Priority propagation is broken and is thus disabled by default. +The +.Va p_nativepri +variable is only updated if a process does not obtain a sleep mutex on the +first try. +Also, if a process obtains more than one sleep mutex in this manner, and +had its priority bumped in between, then +.Va p_nativepri +will be clobbered. |