Re: simplify disksort, please review.

From: Poul-Henning Kamp (phk_at_phk.freebsd.dk)
Date: 06/10/05

  • Next message: Eivind Eklund: "Re: Death to toor" 
    To: Jeff Roberson <jroberson@chesapeake.net>
Date: Fri, 10 Jun 2005 10:10:47 +0200

    In message <20050610040302.S16943@mail.chesapeake.net>, Jeff Roberson writes:
    >On Thu, 9 Jun 2005, Poul-Henning Kamp wrote:
    >
    >> In message <20050609193008.GB837@darkness.comp.waw.pl>, Pawel Jakub Dawidek writes:
    >>
    >> >The one example of how the order can be broken (write(offset, size)):
    >> >
    >> > write(1024, 512)
    >> > write(0, 2048)
    >>
    >> If you issue these two requests just like that, you get no guarantee
    >> which order they get written in.
    >>
    >> It's not just disksort which might surprise you, tagged queuing and
    >> write caches may mess up your day as well.
    >
    >Anyway, it's really not important to solve at the disk queue layer,
    >because it's only an issue with raw devices, and it's very irregular to
    >have multiple processes opening the same raw device to issue a lot of io
    >of different sizes.

    It will not be solved at any level, because it would require the GEOM
    layer to either make all writes strictly synchronous, a total no-go,
    or keep very expensive housekeeping of which requests are outstanding
    so we can block conflicting writes etc.

    This is something that belongs 100% in the code issuing the I/O requests
    and if some of that code does something stupid, it looses and that is
    how it should be.

    Jeff is correct that the buffer-cache mostly will prevent people from
    shooting their feet off, but nothing prevents a GEOM class from making
    a fool of itself. 

    -- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk@FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.
_______________________________________________
freebsd-arch@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-arch
To unsubscribe, send any mail to "freebsd-arch-unsubscribe@freebsd.org"
  • Next message: Eivind Eklund: "Re: Death to toor"