[lkml]   [2004]   [Apr]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: 2.6.6-rc{1,2} bad VM/NFS interaction in case of dirty page writeback
    On Tue, 2004-04-27 at 01:58, Andrew Morton wrote:

    > > Currently, we use this on virtually anything that trigggers writepage()
    > > with the wbc->for_reclaim flag set.
    > Why? Sorry, I still do not understand the effect which you're trying to
    > achieve? I thought it was an efficiency thing: send more traffic via
    > writepages(). But your other comments seem to indicate that this is not
    > the primary reason.

    That is the primary reason. Under NFS, writing is a 2 stage process:

    - Stage 1: "schedule" the page for writing. Basically this entails
    creating an nfs_page struct for each dirty page and putting this on the
    NFS private list of dirty pages.
    - Stage 2: "flush" the NFS private list of dirty pages. This involves
    coalescing contiguous nfs_page structs into chunks of size "wsize", and
    actually putting them on the wire in the form of RPC requests.

    writepage() only deals with one page at a time, so it will work fine for
    doing stage 1.
    If you also try force it to do stage 2, then you will end up with chunks
    of size <= PAGE_CACHE_SIZE because there will only be 1 page on the NFS
    private list of dirty pages. In practice this again means that you
    usually have to send 8 times as many RPC requests to the server in order
    to process the same amount of data.

    This is why I'd like to try to leave stage 2) to writepages().

    Now normally, that is not a problem, since the actual calls to
    writepage() are in fact made by means of a call to generic_writepages()
    from within nfs_writepages(), so I can do all the correct things once
    I'm done with generic_writepages().

    However shrink_cache() (where the MM calls writepage() directly) needs
    to be treated specially, since that is not wrapped by a writepages()
    call. In that case, we return WRITEPAGE_ACTIVATE and wait for pdflush to
    call writepages().

    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: 2005-03-22 14:02    [W:0.026 / U:0.084 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site