Re: [LSF/MM/BPF TOPIC] Allowing linkat() to replace the destination

[Al Viro <viro@zeniv.linux.org.uk>]

On Fri, Jan 17, 2020 at 02:56:05PM +0000, Trond Myklebust wrote:

> It sounds to me like we rather need a meta-topic about "How do we get
> simple things done in the Linux fs community?"
> 
> It shouldn't take a ticket to Palm Springs to perform something simple
> like adding a new flag to a syscall.

Sure - adding a new flag is trivial.  Coming up with sane semantics for
it, OTOH, can be rather non-trivial and in this case it is, unfortunately.
"something like link(2), only it tolerates the existing target and
atomically replaces it" does _not_ specify the semantics.  Try to sit
down for a few minutes and come up with the cases when behaviour is
undefined by the above; it won't take longer than that.

We can do it by asking the proponent to come up with full description to
be included into the proposal, then have at it on fsdevel/linux-abi (as
well as security lists).  Doable, but not a small amount of PITA for
original poster and dealing with questions/objections/etc. is certain
to grow a large thread with many branches (and lots of bikeshedding
thrown in) _and_ would include tons of roundtrips, so the latency
(especially early on, while the proposal is still raw) will be a factor.

It's not the question of how to implement it; it's what should it _do_.
And "we'll tweak the behaviour in corner cases later on" is good in
a lot of situations, but not for userland ABI.  I'd been guilty of
such fuckups several times and they are not cheap to fix afterwards ;-/