lkml.org 
[lkml]   [2001]   [Aug]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: kswap spinning


    On Sat, 18 Aug 2001, Mark Hemment wrote:

    > Hi,
    >
    > Jumping from 2.4.7 to 2.4.9 has shown up a problem with the VM balancing
    > code.
    > Under load, I've seen kswapd become a CPU hog (on a 5GB box). Now that
    > I've got a theory, I cannot reproduce the problem for confirmation, but
    > here is the theory anyway.
    >
    > The tests in free_shortage() and inactive_shortage() assume that the
    > memory state for a zone can be improved by calling refill_inactive_scan(),
    > page_launder(), the inode/dcache shrinking functions, and the general
    > slab-cache reaping func. Usually, some combination of the reaping
    > functions will improve a zone's state - but not always.
    >
    > The problem is the DMA zone. As it is (relatively) small on some archs
    > (IA-32 for example), it is possible that almost all pages from this zone
    > are being used by a sub-systems which simply won't give them up. Examples
    > of use are; vmalloc()ed page for modules, pre-allocated socket buffs
    > for NICs, task-structs/kernel-stack - you get the idea.
    >
    > In the case I believe I'm seeing, both free_shortage() and
    > inactive_shortage() are returning a shortage for the DMA zone which
    > triggers all the work in do_try_to_free_pages(). But as there aren't any
    > DMA pages left in the page-cache, being used as anonymous pages, or being
    > used in the other places the code looks, do_try_to_free_pages() returns
    > non-zero to kswapd() and off we go again. This also causes callers of
    > try_to_free_pages() (from __alloc_pages()) to suck CPU cycles as well.
    >
    > On HIGHMEM boxes, it is possible for the NORMAL zone to get into the
    > same state (although v unlikely).
    >
    > I'd rather not get into having specialist code in mm/vmscan.c (or arch
    > specific code) to handle a small DMA zone - so what are the other
    > solutions? (Assuming the above theory holds true.)
    >
    > One possible solution is not to give DMA memory out, except for;
    > o explicit DMA allocations
    > o page-cache, anonymous page, allocations
    > assuming explicit DMA allocations are low, we'll at least know the
    > remaining DMA pages are somewhere we can get at them - would need to be
    > trigger by an arch specific flag for those that don't have a "tiny" zone.
    > However, I don't like this, feels like fixing the wrong problem. :(

    The DMA zone problem _can_ happen, but it should also happen with 2.4.7 I
    think. (I'm not completly sure, though. Maybe Linus changes to
    try_to_free_pages() are the reason...)

    Lets first find out the actual problem.

    Could you please boot with profile=2 and use readprofile to find out where
    kswapd is spending its time?

    -
    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/

    \
     
     \ /
      Last update: 2005-03-22 12:57    [W:0.026 / U:64.768 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site