Re: [PATCH RFC] arm64: introduce mm_context_t flags

[Catalin Marinas <catalin.marinas@arm.com>]

On Sat, Aug 05, 2017 at 12:49:19AM +0300, Yury Norov wrote:
> On Fri, Aug 04, 2017 at 06:38:10PM +0100, Catalin Marinas wrote:
> > On Wed, Aug 02, 2017 at 08:29:40PM +0300, Yury Norov wrote:
> > > On Wed, Aug 02, 2017 at 05:39:01PM +0100, Catalin Marinas wrote:
> > > > On Mon, Jul 31, 2017 at 05:48:25PM +0300, Yury Norov wrote:
> > > > > In patch 06beb72fbe23e ("arm64: introduce mm context flag to keep 32 bit task
> > > > > information") you introduce the field flags but use it only for a single flag -
> > > > > TIF_32BIT. It looks hacky to me for three reasons:
> > > > >  - The flag is introduced for the case where it's impossible to get the thread
> > > > >    info structure for the thread associated with mm. So thread_info flags (TIF)
> > > > >    may also be unavailable at place. This is not the case for the only existing
> > > > >    user of if - uprobes, but in general this approach requires to include thread
> > > > >    headers in mm code, which may become unwanted dependency.
> > > > >  - New flag, if it uses TIF bits, for consistency should for example set/clear
> > > > >    TIF_32BIT_AARCH64 for ILP32 tasks. And to be completely consistent, with
> > > > >    current approach we'd mirror thread_info flags to mm_context flags. And keep
> > > > >    it syncronized.
> > > > >  - If we start using TIF flags here, we cannot easily add new mm_context
> > > > >    specific bits because they may mess with TIF ones.
> > > > > 
> > > > > I think that this is not what was intended when you added new field in
> > > > > mm_context_t.
> > > > 
> > > > TIF_32BIT was handy at the time but it indeed denotes AArch32 user
> > > > task. For ILP32 we wouldn't need to set this bit since the instruction
> > > > set is A64 and uprobe should support it (though not sure anyone tried).
> > > > I noticed in your patch introducing binfmt_ilp32.c that SET_PERSONALITY
> > > > actually sets this flag in the mm context.
> > > 
> > > Depending on what will be decided here, I'll change ilp32 patch
> > > accordingly.
> > 
> > Since this was meant to keep track of AArch32 tasks, the ILP32
> > personality macros need to clear it.
>  
> I understand it. I meant that the exact fix will depend on what we
> will figure out here.
> 
> I have also fixed small issue with headers in the patch "arm64: signal:
> share lp64 signal structures and routines to ilp32", so I think I will
> create rc4-based branch that will incorporate all changes. But if you
> need I can also update rc3-based branch. And 4.12 - do you need the
> updated version of it?

Not for 4.12, I'll just take it as it is with a fixup for
SET_PERSONALITY. I will hold off with an 4.13 branch until the final
4.13 is actually released, so you can rebase your -rc3/rc4 branch.

-- 
Catalin