should we use atomic instructions to manipulate vmstats.v_wire_count?

Nicolas Thery nthery at gmail.com
Sun Jan 13 05:38:09 PST 2008


2008/1/13, Simon 'corecode' Schubert <corecode at fs.ei.tum.de>:
> Nicolas Thery wrote:
> >>> I noticed that vmstats.v_wire_count is incremented and decremented w/o
> >>> protection.  In contrast, atomic instructions (from <atomic.h>) are
> >>> used in freebsd.
> >>> Is it a bug in dfly or is it intentional?
> >> As far as I understand, we're requiring the mplock for vm operations, so that should be covered.
> > Ok but what about zget() called from zalloc() which increments
> > v_wire_count?   I reckon zalloc() can be called w/o holding the
> > mplock.
>
> I don't think you're allowed to call zalloc() without the mplock.  However there is a comment stating exactly this.  But following the function calls, you could wind up like this:
>
> zalloc
> zget
> vm_page_alloc or kmem_alloc3
> ...
>
> Now this for sure isn't mp safe.  I believe the comment is wrong, but of course I'm no authority in this area.

Thanks.  There are also XXX comments in zone.cpp stating that
increment operations on zone_* sysctl variables are not mp-safe.
Let's wait for the Authority to comment on this :-)





More information about the Kernel mailing list