lkml.org 
[lkml]   [2018]   [Dec]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 2/2] ARC: show_regs: fix lockdep splat for good
    On Thu 20-12-18 18:45:48, Vineet Gupta wrote:
    > On 12/20/18 5:04 AM, Michal Hocko wrote:
    > > On Tue 18-12-18 10:53:59, Vineet Gupta wrote:
    > >> signal handling core calls ARCH show_regs() with preemption disabled
    > >> which causes __might_sleep functions such as mmput leading to lockdep
    > >> splat. Workaround by re-enabling preemption temporarily.
    > >>
    > >> This may not be as bad as it sounds since the preemption disabling
    > >> itself was introduced for a supressing smp_processor_id() warning in x86
    > >> code by commit 3a9f84d354ce ("signals, debug: fix BUG: using
    > >> smp_processor_id() in preemptible code in print_fatal_signal()")
    > > The commit you are referring to here sounds dubious in itself.
    >
    > Indeed that was my thought as well, but it did introduce the preemption disabling
    > logic aroung core calling show_regs() !
    >
    > > We do not
    > > want to stick a preempt_disable just to silence a warning.
    >
    > I presume you are referring to original commit, not my anti-change in ARC code,
    > which is actually re-enabling it.

    Yes, but you are building on a broken concept I believe. What
    implications does re-enabling really have? Now you could reschedule and
    you can move to another CPU. Is this really safe? I believe that yes
    because the preemption disabling is simply bogus. Which doesn't sound
    like a proper justification, does it?

    > > show_regs is
    > > called from preemptible context at several places (e.g. __warn).
    >
    > Right, but do we have other reports which show this, perhaps not too many distros
    > have CONFIG__PREEMPT enabled ?

    I do not follow. If there is some path to require show_regs to run with
    preemption disabled while others don't then something is clearly wrong.

    > > Maybe
    > > this was not the case in 2009 when the change was introduced but this
    > > seems like a relict from the past. So can we fix the actual problem
    > > rather than build on top of it instead?
    >
    > The best/correct fix is to remove the preempt diabling in core code, but that
    > affects every arch out there and will likely trip dormant land mines, needed
    > localized fixes like I'm dealing with now.

    Yes, the fix might be more involved but I would much rather prefer a
    correct code which builds on solid assumptions.

    --
    Michal Hocko
    SUSE Labs

    \
     
     \ /
      Last update: 2018-12-21 14:04    [W:2.261 / U:0.028 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site