Re: [PATCH] futex: Support smaller futexes of one byte or two byte size.

[Peter Zijlstra]

On Fri, Dec 06, 2019 at 04:31:29PM +0100, Peter Zijlstra wrote:
> > +		case FUTEX_WAKE:
> > +		case FUTEX_REQUEUE:
> > +			/*
> > +			 * these instructions work with sized mutexes, but you
> > +			 * don't need to pass the size. we could silently
> > +			 * ignore the size argument, but the code won't verify
> > +			 * that the correct size is used, so it's preferable
> > +			 * to make that clear to the caller.
> > +			 *
> > +			 * for requeue the meaning would also be ambiguous: do
> > +			 * both of them have to be the same size or not? they
> > +			 * don't, and that's clearer when you just don't pass
> > +			 * a size argument.
> > +			 */
> > +			return -EINVAL;
> 
> Took me a while to figure out this relies on FUTEX_NONE to avoid the
> alignment tests.

And thikning more on that, I really _realy_ hate this.

You're basically saying WAKE is size-less, but that means we must
consider what it means to have a u32 WAIT on @ptr, and a u8 WAKE on
@ptr+1. If the wake really is size-less that should match.

I'd be much happier with requiring strict sizing. Because conversely,
what happens if you have a u32-WAIT at @ptr paired with a u8-WAKE at
@ptr? If we demand strict size we can say that should not match. This
does however mean we should include the size in the hash-match function.

Your Changelog did not consider these implications at all.