RE: [PATCH 3/4] zsmalloc: add details to zs_map_object boiler plate

[Dan Magenheimer <dan.magenheimer@oracle.com>]

> From: Dan Magenheimer
> Subject: RE: [PATCH 3/4] zsmalloc: add details to zs_map_object boiler plate
> 
> > From: Minchan Kim [mailto:minchan@kernel.org]
> > Subject: Re: [PATCH 3/4] zsmalloc: add details to zs_map_object boiler plate
> >
> > On Wed, Jul 11, 2012 at 09:15:43AM -0500, Seth Jennings wrote:
> > > On 07/11/2012 02:42 AM, Minchan Kim wrote:
> > > > On 07/11/2012 12:17 AM, Seth Jennings wrote:
> > > >> On 07/09/2012 09:35 PM, Minchan Kim wrote:
> > > >>> Maybe we need local_irq_save/restore in zs_[un]map_object path.
> > > >>
> > > >> I'd rather not disable interrupts since that will create
> > > >> unnecessary interrupt latency for all users, even if they
> > > >
> > > > Agreed.
> > > > Although we guide k[un]map atomic is so fast, it isn't necessary
> > > > to force irq_[enable|disable]. Okay.
> > > >
> > > >> don't need interrupt protection.  If a particular user uses
> > > >> zs_map_object() in an interrupt path, it will be up to that
> > > >> user to disable interrupts to ensure safety.
> > > >
> > > > Nope. It shouldn't do that.
> > > > Any user in interrupt context can't assume that there isn't any other user using per-cpu buffer
> > > > right before interrupt happens.
> > > >
> > > > The concern is that if such bug happens, it's very hard to find a bug.
> > > > So, how about adding this?
> > > >
> > > > void zs_map_object(...)
> > > > {
> > > > 	BUG_ON(in_interrupt());
> > > > }
> > >
> > > I not completely following you, but I think I'm following
> > > enough.  Your point is that the per-cpu buffers are shared
> > > by all zsmalloc users and one user doesn't know if another
> > > user is doing a zs_map_object() in an interrupt path.
> >
> > And vise versa is yes.
> >
> > > However, I think what you are suggesting is to disallow
> > > mapping in interrupt context.  This is a problem for zcache
> > > as it already does mapping in interrupt context, namely for
> > > page decompression in the page fault handler.
> >
> > I don't get it.
> > Page fault handler isn't interrupt context.
> >
> > > What do you think about making the per-cpu buffers local to
> > > each zsmalloc pool? That way each user has their own per-cpu
> > > buffers and don't step on each other's toes.
> >
> > Maybe, It could be a solution if you really need it in interrupt context.
> > But the concern is it could hurt zsmalloc's goal which is memory
> > space efficiency if your system has lots of CPUs.
> 
> Sorry to be so far behind on this thread.
> 
> For frontswap and zram, the "put" calls are not in interrupt
> context.  For cleancache, the put call IS in interrupt context.
> So if you want to use zsmalloc for zcache+cleancache, interrupt
> context is a concern.  As discussed previously in a separate
> thread though, zsmalloc will take a lot of work to support the full
> needs of zcache.  So, pick your poison.

Oops, correction.  Cleancache puts are not in interrupt context
but do have interrupts disabled.  That's quite different of
course.  So Minchan's BUG_ON(in_interrupt()) should be fine for
now.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>