Re: Causing a process switch to test a theory.

From: Stephan Uphoff (ups_at_tree.com)
Date: 03/21/05

  • Next message: Dan Nelson: "Re: mysql threads" 
    To: matthew@digitalstratum.com
Date: Mon, 21 Mar 2005 10:16:26 -0500

    On Mon, 2005-03-21 at 09:16, Matthew Hagerty wrote:
    > Zera William Holladay wrote:
    >
    > >If you post the section(s) of code in question, then you'll probably
    > >elicit some responses. PIPE_BUF is a POSIX defined minimum, so you might
    > >grep for sections of code that contain fpathconf(*, _PC_PIPE_BUF) to
    > >determine if the programmers took this into consideration. At least
    > >you'll be able to follow the logical flow of the program from fpathconf()
    > >forward.
    > >
    > >Further, if you do some fancy programming (like preempting a process
    > >unnaturally) to determine if there is an error in this particular aspect
    > >of Apache, then you'll also have to show that you have not inflicted the
    > >error too, which will probably be harder than what you set out to solve or
    > >figure out.
    > >
    > >Good luck, Zera Holladay
    > >_______________________________________________
    > >freebsd-hackers@freebsd.org mailing list
    > >http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
    > >To unsubscribe, send any mail to "freebsd-hackers-unsubscribe@freebsd.org"
    > >
    > >
    >
    > I was not wanting to preempt any of the processes unnaturally, I was
    > looking more for suggestions as to how I can reliably cause preemption
    > for testing. Without knowing the details of how FreeBSD's scheduler
    > works, I could only assume that making a heavy load might cause such
    > preemption to happen, but that is just an under educated guess.
    >
    > I'm not familiar with fpathconf(), I'll go look that up for sure and
    > check Apache's code for it's use. Thanks!
    >
    > As for code, below is the portion that does to actual writing to the log
    > file, opened as either a normal file or a pipe depending on the config
    > file. Also note that BUFFERED_LOGS is not normally defined (must be
    > passed to make at compile time), and even then it only makes sure there
    > is *at least* LOG_BUFSIZE bytes to write before writing. It does not
    > prevent longer than LOG_BUFSIZE bytes from being written.
    >
    > Matthew
    >
    > < SNIP -- CODE DELETED >

    Why not just replace the write() call with a small test function?
    There is really no need to modify the kernel for your experiments.

    size_t test_pipe_write(int d, const void *buf, size_t nbytes)
    {
            size_t n;
            size_t done=0;

            while (nbytes > ATOMIC_WRITE_MAX) {
                     n = write(d,buf,ATOMIC_WRITE_MAX);
                    if (n <= 0)
                            return done;
                    buf = (char *) buf + n;
                    nbytes -= n;
                    usleep(NAP_TIME);
             }

            n = write(d,buf,nbytes);
            if (n <= 0)
                    return done;
            else
                    return n + done;
    }

    Stephan

    _______________________________________________
    freebsd-hackers@freebsd.org mailing list
    http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
    To unsubscribe, send any mail to "freebsd-hackers-unsubscribe@freebsd.org"

  • Next message: Dan Nelson: "Re: mysql threads"