lkml.org 
[lkml]   [2009]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: RFC for a new Scheduling policy/class in the Linux-kernel
    From
    Date
    On Thu, 2009-07-16 at 08:59 -0400, James H. Anderson wrote:
    >
    > Raistlin wrote:
    > > Also, I'm not sure I can find in the FMLP paper information about the
    > > possibility of a task to suspend itself (e.g., I/O completion) while
    > > holding a short lock... I assume this is not recommended, but may be
    > > wrong, and, in that case, I hope Prof. Anderson and Bjorn will excuse
    > > and correct me. :-)
    > >
    > >
    > This is a really excellent point and something I probably should have
    > mentioned. We developed the FMLP strictly for real-time (only)
    > workloads. We were specifically looking at protecting memory-resident
    > resources (not I/O). The FMLP would have to be significantly extended
    > to work in settings where these assumptions don't hold.

    One thing I've been thinking about is extending lockdep to help verify
    things like this.

    If we were to annotate a syscall/trap with something like:

    lockdep_assume_rt()

    And teach lockdep about non-RT blocking objects, we could validate that
    the callchain down from lockdep_assume_rt() would not indeed contain a
    non-RT resource, but also that we don't take locks which might in other
    another code path.

    That is, suppose:

    sys_foo()
    lockdep_assume_rt()
    mutex_lock(&my_lock)

    vs

    sys_bar()
    mutex_lock(&my_lock)
    down_read(&mm->mmap_sem)

    vs

    page-fault
    down_read(&mm->mmap_sem)
    lock_page(page)

    Would indeed generate a warning because mmap_sem is known to block on
    IO, and there is a dependency (through sys_bar()) between my_lock and
    mmap_sem, therefore sys_foo()'s assumption is invalid.





    \
     
     \ /
      Last update: 2009-07-16 15:39    [W:0.024 / U:30.656 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site