RE: [PATCH 20/22] x86,word-at-a-time: Remove .fixup usage

[David Laight]

From: Peter Zijlstra
> Sent: 10 November 2021 11:10
> 
> On Wed, Nov 10, 2021 at 10:46:42AM +0000, David Laight wrote:
> > From: Peter Zijlstra
> > > Sent: 09 November 2021 21:08
> > ...
> > >
> > > GCC does the same, but I wanted to have the exception stuff be in
> > > .text.cold, but alas it doesn't do that. I left the attribute because of
> > > it's descriptive value.
> > >
> > > >  Unless the cold attribute is helping move
> > > > ("shrink-wrap"?) the basic block to a whole other section
> > > > (.text.cold.)?
> > >
> > > I was hoping it would do that, but it doesn't on gcc-11.
> >
> > Wouldn't moving part of a function to .text.cold (or .text.unlikely)
> > generate the same problems with the stack backtrace code as the
> > .text.fixup section you are removing had??
> 
> GCC can already split a function into func and func.cold today (or
> worse: func, func.isra.N, func.cold, func.isra.N.cold etc..).
> 
> I'm assuming reliable unwind and livepatch know how to deal with this.

They'll have 'proper' function labels at the top - so backtrace
stands a chance.
Indeed you (probably) want it to output "func.irsa.n.cold" rather
than just "func" to help show which copy it is in.

I guess that livepatch will need separate patches for each
version of the function - which might be 'interesting' if
all the copies actually need patching at the same time.
You'd certainly want a warning if there seemed to be multiple
copies of the function.

I'm waiting for the side-channel attack caused by detecting
timing differences caused by TLB misses when speculatively
executing code in the .cold/.unlikely sections.
ISTR recent x86 cpu speculate unknown conditional branches
'randomly' - rather than (say) assuming backwards taken.
So you can't (easily) stop speculative execution into the 'cold'
text.

I don't know if speculative execution will load TLB,
it would speed a lot of code up - with the same downsides
as evicting code from the L1-cache.
A 'half-way house' would be to do the page table walk, but
hold the read value 'pending' the code being needed.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)