Re: [PATCH] Sys V shared memory limited to 8TiB.

[Robin Holt <holt@sgi.com>]

On Thu, Apr 11, 2013 at 02:07:08PM -0700, Andrew Morton wrote:
> On Wed, 10 Apr 2013 21:42:23 -0500 Robin Holt <holt@sgi.com> wrote:
> 
> > On Wed, Apr 10, 2013 at 04:15:07PM -0700, Andrew Morton wrote:
> > > On Tue, 9 Apr 2013 21:39:24 -0500 Robin Holt <holt@sgi.com> wrote:
> > > 
> > > > Trying to run an application which was trying to put data into half
> > > > of memory using shmget(), we found that having a shmall value below
> > > > 8EiB-8TiB would prevent us from using anything more than 8TiB.  By setting
> > > > kernel.shmall greater that 8EiB-8TiB would make the job work.
> > > > 
> > > > In the newseg() function, ns->shm_tot which, at 8TiB is INT_MAX.
> > > 
> > > You have way too much memory.
> > > 
> > > > ipc/shm.c:
> > > >  458 static int newseg(struct ipc_namespace *ns, struct ipc_params *params)
> > > >  459 {
> > > > ...
> > > >  465         int numpages = (size + PAGE_SIZE -1) >> PAGE_SHIFT;
> > > > ...
> > > >  474         if (ns->shm_tot + numpages > ns->shm_ctlall)
> > > >  475                 return -ENOSPC;
> > > >
> > > > ...
> > > >
> > > > --- a/include/linux/ipc_namespace.h
> > > > +++ b/include/linux/ipc_namespace.h
> > > > @@ -43,8 +43,8 @@ struct ipc_namespace {
> > > >  
> > > >  	size_t		shm_ctlmax;
> > > >  	size_t		shm_ctlall;
> > > > +	unsigned long	shm_tot;
> > > >  	int		shm_ctlmni;
> > > > -	int		shm_tot;
> > > >  	/*
> > > >  	 * Defines whether IPC_RMID is forced for _all_ shm segments regardless
> > > >  	 * of shmctl()
> > > 
> > > I reviewed everything for fallout from this and don't see any obvious
> > > issues.
> > > 
> > > I do wonder about the appropriateness of the unsigned long type.  Most
> > > (but by no means all) code in this area uses size_t, and the
> > > above-quoted ns->shm_ctlall is size_t.
> > 
> > The only reason I went with unsigned long instead of size_t was most
> > places in the kernel track stuff I recalled that was tracking stuff
> > in pages used unsigned longs.  Also, I found shm_tot field in shm_info
> > structure was an unsigned long so this felt like a natural fit.  I would
> > happily changed to size_t.  Whatever you feel is right.
> 
> I have no really strong feelings, but let's at least put some thought
> into it.  I do prefer ulong and find size_t to be a PITA.  I guess it
> doesn't matter much.
> 
> > > --- a/ipc/shm.c~ipc-sysv-shared-memory-limited-to-8tib-fix
> > > +++ a/ipc/shm.c
> > > @@ -462,7 +462,7 @@ static int newseg(struct ipc_namespace *
> > >  	size_t size = params->u.size;
> > >  	int error;
> > >  	struct shmid_kernel *shp;
> > > -	int numpages = (size + PAGE_SIZE -1) >> PAGE_SHIFT;
> > > +	size_t numpages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
> > 
> > I was holding off from that change only because I was asking for this
> > to go to stable and this doubles the size of the patch. ;)
> 
> It's a bug, isn't it?  Is there anything else which prevents creation
> of segments which are >=8TB?

Definitely a bug.  Have not tried to create a segment >= 8TiB.  I can
give that a try tomorrow morning when I have a machine again to test with.

Thanks,
Robin