circular pivot_root, is this possible ?

From: David Bryson (david_at_tsumego.com)
Date: 03/02/04

  • Next message: Mike: "Re: [PATCH] Re: 2.6.3-bk9 QA testing: firewire good, USB printing dead" 
    Date:	Mon, 1 Mar 2004 18:30:08 -0800
To: linux-kernel@vger.kernel.org

    Greetings!
    I'm working on some embedded applications and I am wondering if
    something like the following is possible. I have implemented most of
    this but the last few steps are a bit of a mystery to me... and after
    reading many threads in the lkml archives I don't believe it's
    possible, but I thought I'd ask anyway.

    flash media contains two filesystem images, an initrd.gz, a
    sys.img(cramfs), and a kernel

    1) kernel boots with initrd.gz as root, loading into ram0
    2) initrd mounts the flash filesystem, and loads sys.img via loopback
    3) a tmpfs is created, contents of sys.img are copied into the tmpfs
    4) system does a pivot_root on the tmpfs, passing the initrd as
    well(somewhere on the tmpfs filesytem)
    5) system begins normal operation

    This works find and is fairly easy. The next part is what I am having
    trouble with.

    Say the system needs an upgrade.... I want to

    1) pivot_root _back out_ of the tmpfs, onto the initrd again
    2) obtain via network a new sys.img, write it to the flash
    3) wipe tmpfs, and recopy the contents of the new sys.img into memory
    4) pivot_root back into the tmpfs and start the higher level system
    again

    pivot_root circles around root filesystems as various tasks are
    completed.

    Somehow init isn't behaving properly with me on this. My first
    attempts included an init on both initrd.gz and sys.img. Having two
    init's running at normal system startup. This did not work well but
    got the job done, as init(PID:1 on the initrd) was running as well as
    init(some other PID on sys.img)

    Eventually I tried only running a script on the initrd.gz to get the
    system pivot_root'd to tmpfs. But if I attempted to start the
    circular pivot_root behavior(back to the inird) I still ended up with
    init(PID:1) on the tmpfs, while my rootfs was in the initrd again.
    Additionally I couldn't unmount devfs because I am assuming the old
    processes running under init are still using it.

    I also tried killing init with various singals, to see if it would
    'reload' the config on the current rootfs(kill -1 1, kill -HUP 1) but
    no success there.

    So the question remains, is it possible to have circular pivot_roots ?
    thanks
    Dave

    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

    • application/pgp-signature attachment: stored
  • Next message: Mike: "Re: [PATCH] Re: 2.6.3-bk9 QA testing: firewire good, USB printing dead"

    Relevant Pages