[lkml]   [2004]   [Nov]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[patch] Real-Time Preemption, -RT-2.6.10-rc1-mm3-V0.7.25-0

i have released the -V0.7.25-0 Real-Time Preemption patch, which can be
downloaded from the usual place:

this release includes fixes, new features and latency improvements.

the biggest change is the threading of the lone remaining non-threaded
external interrupt: the timer interrupt (IRQ#0).

It was not threaded until now because unlike device interrupts the timer
interrupt is quite deeply attached to Linux's architecture, driving
process timing, profiling and the scheduler. I separated these into
stuff that can be done from a thread context and stuff that needs to
execute directly.

Fortunately most of the expensive and latency-generating stuff could be
pushed into the irq thread. As a side-effect this enabled the turning of
rtc_lock and xtime_lock into a mutex. Also, it removed some ugly hacks
from rtc.c and should improve worst-case latencies.

the other bigger change is the reworking of the .config space. Instead
of the deep hierarchy of hard-to-understand technical PREEMPT options,
there's now a flattened out choice of 4 preemption models:

( ) No Forced Preemption (Server)
( ) Voluntary Kernel Preemption (Desktop)
( ) Preemptible Kernel (Low-Latency Desktop)
(X) Complete Preemption (Real-Time)

and updated help texts. Plus for the preemption models where it can be
freely turned off, softirq and hardirq threading is an additional

the third bigger change is the reworking of the tracer to make it easier
to drive it from user-space. There are 3 runtime flags now:

/proc/sys/kernel/trace_enabled (default: 1)
/proc/sys/kernel/trace_user_triggered (default: 0)
/proc/sys/kernel/trace_freerunning (default: 0)

semantics of user-triggered tracing: if enabled then any active tracing
of wakeup and/or critical sections stops and userspace drives start/stop
events via gettimeofday(0,1)/gettimeofday(0,0). The latter saves the
current trace into /proc/latency_trace, the former clears the trace
buffer and starts tracing anew. Doing another gettimeofday(0,1) on an
already running tracer clears the trace and restarts it without saving
the current trace into /proc/latency_trace. Doing a gettimeofday(0,0) on
an already stopped tracer has no effect (i.e. /proc/latency_trace wont
be saved a second time). The tracer does timing for userspace
automatically the same way it does it for the built-in timing
mechanisms, and it can be configured via the preempt_max_latency and
preempt_tresh tunables.

also, wakeup-timing, irq-off and preempt-off critical section timing can
now be done at once again, the /proc/sys/kernel/preempt_wakeup_timing
flag switches between the modes. (default: 1)

other changes since -V0.7.24:

- debug feature: added the RTC-debug patch sanitized by K.R. Foley,
plus further cleanups.

- added upstream fix for kobject related crash, pointed out by Shane

- cleanup: Kconfig help text fixes from Amit Shah

- latency improvement: on UP-IOAPIC, when redirecting an interrupt, do
not ack the APIC. This is the method used for direct interrupts and
on UP it might as well work out fine. It is certainly faster than
masking/unmasking, making UP-IOAPIC the fastest PIC mode again.

- livelock fix: the timer-irq threading unearthed a seqlock related
livelock scenario, which triggered in do_gettimeoffset() big-time.
The solution is to serialize seqlock readers with writers _iff_ the
seqlock status is 'invalid'. This is a rare event, but when it
happens it saves the day.

- debugging helper: the /proc/sys/kernel/debug_direct_keyboard flag
(default: 0) will hack the keyboard IRQ into being direct. NOTE: the
keyboard in this mode should only be used to access SysRq
functionality that is not possible via the threaded keyboard handler.
The direct keyboard IRQ can crash the system.

- new kernel profiling features: added profile=preempt to profile
whether interrupts hit the kernel in preemptible mode or in a
critical section. Added /proc/sys/kernel/prof_pid to profile a
specific PID only. (default: -1, meaning all tasks profiled) Added
/proc/irq/prof_cpu_mask back.

- robustness improvement: do not report atomicity-warnings during
kernel oopses - it's more important to get the oops out to the

to create a -V0.7.25-0 tree from scratch, the patching order is:

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 14:07    [W:0.537 / U:3.348 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site