Re: Regression - locking (all from 2.6.28)

[Pavel Machek <pavel@ucw.cz>]

On Tue 2009-03-17 09:07:32, KAMEZAWA Hiroyuki wrote:
> On Mon, 16 Mar 2009 23:04:27 +0100
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> 
> > On Friday 06 March 2009, Pavel Machek wrote:
> > > Hi!
> > > 
> > > On Fri 2009-03-06 11:19:46, Dave Hansen wrote:
> > > > On Fri, 2009-03-06 at 18:00 +0000, Catalin Marinas wrote:
> > > > > > I think you should be more worried about consistency rather than missing
> > > > > > entries.  Take these two lines of code:
> > > > > > 
> > > > > >       start_pfn = node->node_start_pfn;
> > > > > >       /* hotplug occurs here */
> > > > > >       end_pfn = start_pfn + node->node_spanned_pages;
> > > > > > 
> > > > > > What if someone comes in and adds memory to the node, at the beginning
> > > > > > of the node, after you have calculated start_pfn?  Try to think of what
> > > > > > value you'll get for end_pfn and whether it is consistent and was *ever*
> > > > > > valid at all.  Would that oops the kernel?
> > > > > 
> > > > > I assume pfn_valid() should handle this and kmemleak wouldn't scan the
> > > > > page, unless we need locks around pfn_valid as well but I haven't seen
> > > > > any used in the kernel.
> > > > 
> > > > You assume incorrectly. :(
> > > > 
> > > > Take my above example, and assume that you have two nodes which are
> > > > right next to each other.  You might run over the end of one node and
> > > > into the next one.  Your pages will be pfn_valid() but you will be on
> > > > the wrong node.
> > > > 
> > > > Please take a look at those locks that I mentioned.  Notice that they
> > > > are lock the pgdat *span*, not the validity of pages inside the pgdat.
> > > > Your code deals with what pages the pgdats *span* and thus needs that
> > > > lock.  Notice that my example also had to do with those two lines of
> > > > code incorrectly guessing the pgdat's *span*.
> > > > 
> > > > We recently went to some pain to make sure that the software suspend
> > > > code (which walks pgdat ranges as well) worked with memory hotplug.
> > > > There really isn't that much code around that actually cares at runtime
> > > > about which physical areas a particular node or zone spans.  Yours is a
> > > > rarity and will require some caution.
> > > > 
> > > > You could probably also use the memory hotplug mutex found here:
> > > > 
> > > > https://lists.linux-foundation.org/pipermail/linux-pm/2008-November/018884.html
> > > > 
> > > > But I'm not sure where those patches have gone.  Hmmm.  Pavel?
> > > 
> > > I don't think they were applied. They probably should... Rafael was
> > > about to look into that, but he lost the patch pointer.
> > 
> > Yes.  In fact, sending the patch again to me would be appreciated.
> > 
> 
> Ah...but as I wrote, for me, testing this is difficult.
> Could some expert of hibernation update and post this again ? For memoyr hotplug.
> it seems enough that sending the patch witch CC: to linux-mm and including
> "Memory Hotplug" in the subject.
> 

Well, testing this is probably difficult for everyone, as
hotplug-capable machines are rare/expensive.

Can youjust retransmit the patch to rafael?
> 

-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html