Message-Id: <1233351412.908.69.camel@lts-notebook>
Date:	Fri, 30 Jan 2009 16:36:52 -0500
From:	Lee Schermerhorn <Lee.Schermerhorn@...com>
To:	Hugh Dickins <hugh@...itas.com>
Cc:	Linus Torvalds <torvalds@...ux-foundation.org>,
	Greg KH <gregkh@...e.de>,
	Maksim Yevmenkin <maksim.yevmenkin@...il.com>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	Nick Piggin <npiggin@...e.de>,
	Andrew Morton <akpm@...ux-foundation.org>,
	will@...wder-design.com, Rik van Riel <riel@...hat.com>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>,
	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
	Mikos Szeredi <miklos@...redi.hu>
Subject: Re: [PATCH] Fix OOPS in mmap_region() when merging adjacent
	VM_LOCKED file segments

On Fri, 2009-01-30 at 21:12 +0000, Hugh Dickins wrote:
> On Fri, 30 Jan 2009, Linus Torvalds wrote:
> > On Fri, 30 Jan 2009, Lee Schermerhorn wrote:
> > > 
> > > So happens, I'm mapping with MAP_SHARED, so the VM_ACCOUNT flag gets
> > > cleared later in mmap_region().  Comments say that this is for checking
> > > memory availability during shmem_file_setup().  Maybe we can move the
> > > temporary setting of VM_ACCOUNT until just before the call to
> > > shmem_zero_setup()?
> > 
> > Yeah, that would probably fix it, and looks like the right thing to do. 
> 
> I do need to refresh my memory on that in a moment...
> 
> > 
> > It all looks pretty confused wrong to set the whole VM_ACCOUNT flag for a 
> > file-backed file AT ALL in the first place, but the code knows that it 
> > won't matter for a shared file, and will be cleared again later.
> > 
> > So it plays these temporary games with vm_flags, and it didn't matter 
> > because of how we used to call "vma_merge()" either early only for the 
> > anonymous memory case (that had VM_ACCOUNT stable and didn't have that 
> > temporary case at all) or much later (after having undone the temporary 
> > flag setting) for files.
> 
> I'm to blame for those games, and now they've given trouble,
> the right thing may be to put an end to them.
> 
> > 
> > Why do we pass in that "accountable" flag, btw? It's only ever set to 0 by 
> > a MAP_PRIVATE mapping that hits is_file_hugepages() (see do_mmap_pgoff), 
> > and we could just do that decision all inside mmap_region(). So the flag 
> > doesn't really seem to have any real meaning, and is just passed around 
> > for some odd historical reason?
> 
> It looks like the "accountable" flag dates from before Miklos separated
> mmap_region() out from do_mmap_pgoff(): so he just passed it on down to
> mmap_region() as an additional argument, preferring to leave the more
> complex MAP_PRIVATE/is_file_hugepages test behind in do_mmap_pgoff().
> 
> It seemed rather a random refactoring to me.  Looking at it again,
> I wonder if we should be getting do_brk() to use mmap_region() too;
> but my appetite for cleanups is low at present, bugs we have enough.
> 
> By the way, there's an argument to say that you should add
> VM_MIXEDMAP to VM_CAN_NONLINEAR in VM_MERGEABLE_FLAGS: I don't
> really care whether we merge the odd filemap_xip vma or not,
> but it used to do so and now won't.
> 
> By the same (used to merge, now won't) argument, one could say
> VM_INSERTPAGE should be there too; but whereas VM_MIXEDMAP is used
> in one place only, quite a lot of drivers use vm_insert_page(), so
> I feel more comfortable with the idea that it's stopping merges -
> though in that case, shouldn't we add it to VM_SPECIAL?
> 
> But I'm caring more about that VM_ACCOUNT...

I just verified that adding VM_ACCOUNT to VM_MERGEABLE does allow the
merge to happen with the test program.  And the system didn't come
crashing down around me.  But, I wouldn't trust that simple test as the
last word.  A short run of a stress load I use held up/still running,
but I can't tell whether it's merging as expected there.   

I am running a slightly modified version of Maksim's test program under
the harness.  I modified it to mmap the entire region to reserve space,
then MAP_FIXED at each page address in the range returned by the first
mmap.  I saw that it was leaving holes between some of the pages w/o
this.  I'm going to automate the check for merging [read map and verify
a single segment at expected range] and leave that running with the
load.

Lee

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists