lkml.org 
[lkml]   [2012]   [Nov]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [RFC v3 0/3] vmpressure_fd: Linux VM pressure notifications
    On Wed, Nov 07, 2012 at 01:30:16PM +0200, Pekka Enberg wrote: [...]
    > I love the API and implementation simplifications but I hate the new
    > ABI. It's a specialized, single-purpose syscall and bunch of procfs
    > tunables and I don't see how it's 'extensible' to anything but VM

    It is extensible to VM pressure notifications, yeah. We're probably not
    going to add the raw vmstat values to it (and that's why we changed the
    name). But having three levels is not the best thing we can do -- we can
    do better. As I described here:

    http://lkml.org/lkml/2012/10/25/115

    That is, later we might want to tell the kernel how much reclaimable
    memory userland has. So this can be two-way communication, which to me
    sounds pretty cool. :) And who knows what we'll do after that.

    But these are just plans. We might end up not having this, but we always
    have an option to have it one day.

    > If people object to vmevent_fd() system call, we should consider using
    > something more generic like perf_event_open() instead of inventing our
    > own special purpose ABI.

    Ugh. While I *love* perf, but, IIUC, it was designed for other things:
    handling tons of events, so it has many stuff that are completely
    unnecessary here: we don't need ring buffers, formats, 7+k LOC, etc. Folks
    will complain that we need the whole perf stuff for such a simple thing
    (just like cgroups).

    Also note that for pre-OOM we have to be really fast, i.e. use shortest
    possible path (and, btw, that's why in this version the read() now can be
    blocking -- and so we no longer have to do two poll()+read() syscalls,
    just single read is now possible).

    So I really don't see the need for perf here: it doesn't result in any
    code reuse, but instead it just complicates our task. As for ABI
    maintenance point of view, it is just the same thing as the dedicated
    syscall.

    Thanks,
    Anton.


    \
     
     \ /
      Last update: 2012-11-07 13:41    [W:2.912 / U:0.188 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site