DragonFly BSD
DragonFly commits List (threaded) for 2005-05
[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]

Re: cvs commit: src/sys/i386/i386 trap.c src/sys/kern kern_poll.c

From: Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>
Date: Wed, 25 May 2005 08:10:19 -0700 (PDT)

:On Wed, May 25, 2005 at 01:06:53AM -0700, Matthew Dillon wrote:
:>     Putting a hook in the trap code does not result in any sort of
:>     deterministic operation.  I don't know who put that hack in there
:>     (inherited from FreeBSDland) and I don't want to know.  That whole mess
:>     was put in because hardclock (the normal polling point) is a totally
:>     inadequate place to trigger the poll.
:The reason it is there doesn't have anything to do with hardclock.
:It's cheap to check there and helps network programs by reducing
:the latency.

    Cheap is a relative term.  But you are missing the point... the problem
    isn't whether it's cheap, it's whether it provides any sort of 
    deterministic, reliable operation.  And the answer is: it doesn't.

    Here's a thought experiment:

    What is the overhead of taking 1000 interrupts per second on a 
    modern machine ?  Just the overhead.  What about 10,000 interrupts per

    Now once you've calculated that, try to think up a good reason to do
    the polling any other way.

					Matthew Dillon 

[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]