[lkml]   [2006]   [Dec]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCH] mm: fix page_mkclean_one (was: 2.6.19 file content corruption on ext3)

    On Wed, 20 Dec 2006, Linus Torvalds wrote:
    > >
    > > There's also redirty_page_for_writepage().
    > _dirtying_ a page makes sense in any situation. You can always dirty them.
    > I'm just saying that you can't just mark them *clean*.
    > If your point was that the filesystem had better be able to take care of
    > "redirty_page_for_writepage()", then yes, of course. But since it's the
    > filesystem itself that does it, it had _better_ be able to take care of
    > the situation it puts itself into.

    Btw, as an example of something where this may NOT be ok, look at

    I'm not at all convinced that "migrate_page_copy()" can work at all. It

    if (PageDirty(page)) {

    which is an example of what NOT to do, because it claims to clear the page
    for IO, but doesn't actually _do_ any IO.

    And this is wrong, for many reasons.

    For example, it's very possible that the old page is not actually
    up-to-date, and is only partially dirty using some FS-specific dirty data
    queues (like NFS does with its dirty data, or buffer-heads can do for
    local filesystems). When you do

    if (clear_dirty(page))

    in generic VM code, that is a BUG. It's an insane operation. It cannot
    work. It's exactly what I'm trying to avoid.

    So page migration is probably broken, but it's no less broken than it
    always has been. And I don't think many people use it anyway. It might
    work "by accident" in a lot of situations, but to actually be solid, it
    really would need to do something fundamentally different, like:

    - have a per-mapping "migrate()" function that actually knows HOW to
    migrate the dirty state from one page to another.

    - or, preferably, by just not migrating dirty pages, and just actually
    doing the writeback on them first.

    Again, this is an example of just _incorrect_ code, that thinks that it
    can "just clear the dirty bit". You can't do that. It's wrong. And it is
    not wrong just because I say so, but because the operations itself simply
    is FUNDAMENTALLY not a sensible one.

    This is why I keep harping on this issue: there are two cases, and two
    cases only, when you can clear a page. And no, "migrating the data to
    another page" was not one of those two cases. The cases are, and will
    _always_ be: (a) full writeback IO of _all_ the dirty data on the page
    (and that can only be done by the low-level filesystem, since it's the
    only one that knows what rules it has followed for marking things dirty)
    and (b) cancelling dirty data that got truncated and literally removed
    from the filesystem.

    So I don't claim that I fixed all the cases. mm/migrate.c is still broken.
    Maybe somebody else also uses "clear_page_dirty_for_io()" even though the
    name very clearly says FOR IO. I didn't check, but I think they're mostly
    right now.

    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: 2009-11-18 23:46    [W:0.024 / U:5.096 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site