[lkml]   [2010]   [Jan]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [RFC][PATCH] PM: Force GFP_NOIO during suspend/resume (was: Re: [linux-pm] Memory allocations in .suspend became very unreliable)
    On Thursday 21 January 2010, KOSAKI Motohiro wrote:
    > > > Hi Rafael,
    > > >
    > > > Do you mean this is the unrelated issue of nVidia bug?
    > >
    > > The nvidia driver _is_ buggy, but Maxim said he couldn't reproduce the
    > > problem if all the allocations made by the nvidia driver during suspend
    > > were changed to GFP_ATOMIC.
    > >
    > > > Probably I haven't catch your point. I don't find Maxim's original bug
    > > > report. Can we share the test-case and your analysis detail?
    > >
    > > The Maxim's original report is here:
    > >
    > >
    > > and the message I'm referring to is at:
    > >
    > Hmmm...
    > Usually, Increasing I/O isn't caused MM change. either subsystem change
    > memory alloc/free pattern and another subsystem receive such effect ;)
    > I don't think this message indicate MM fault.
    > And, 2.6.33 MM change is not much. if the fault is in MM change
    > (note: my guess is no), The most doubtful patch is my "killing shrink_all_zones"
    > patch. If old shrink_all_zones reclaimed memory much rather than required.
    > The patch fixed it. IOW, the patch can reduce available free memory to be used
    > buggy .suspend of the driver. but I don't think it is MM fault.
    > As I said, drivers can't use memory freely as their demand in suspend method.
    > It's obvious. They should stop such unrealistic assumption. but How should we fix
    > this?
    > - Gurantee suspend I/O device at last?
    > - Make much much free memory before calling .suspend method? even though
    > typical drivers don't need.

    That doesn't help already. Maxim tried to increase SPARE_PAGES (in
    kernel/power/power.h) and that had no effect.

    > - Ask all drivers how much they require memory before starting suspend and
    > Make enough free memory at first?

    That's equivalent to reworking all drivers to allocate memory before suspend
    eg. with the help of PM notifiers. Which IMHO is unrealistic.

    > - Or, do we have an alternative way?

    The $subject patch?

    > Probably we have multiple option. but I don't think GFP_NOIO is good
    > option. It assume the system have lots non-dirty cache memory and it isn't
    > guranteed.

    Basically nothing is guaranteed in this case. However, does it actually make
    things _worse_? What _exactly_ does happen without the $subject patch if the
    system doesn't have non-dirty cache memory and someone makes a GFP_KERNEL
    allocation during suspend?


     \ /
      Last update: 2010-01-21 21:23    [W:0.023 / U:181.928 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site