Re: [PATCH RFC v3 0/7] epoll: Introduce new syscalls, epoll_ctl_batch and epoll_pwait1

[Fam Zheng <famz@redhat.com>]

On Wed, 02/18 19:49, Ingo Molnar wrote:
> 
> * Fam Zheng <famz@redhat.com> wrote:
> 
> > On Sun, 02/15 15:00, Jonathan Corbet wrote:
> > > On Fri, 13 Feb 2015 17:03:56 +0800
> > > Fam Zheng <famz@redhat.com> wrote:
> > > 
> > > > SYNOPSIS
> > > > 
> > > >        #include <sys/epoll.h>
> > > > 
> > > >        int epoll_pwait1(int epfd, int flags,
> > > >                         struct epoll_event *events,
> > > >                         int maxevents,
> > > >                         struct epoll_wait_params *params);
> > > 
> > > Quick, possibly dumb question: might it make sense to also pass in 
> > > sizeof(struct epoll_wait_params)?  That way, when somebody wants to add
> > > another parameter in the future, the kernel can tell which version is in
> > > use and they won't have to do an epoll_pwait2()?
> > > 
> > 
> > Flags can be used for that, if the change is not 
> > radically different.
> 
> Passing in size is generally better than flags, because 
> that way an extension of the ABI (new field[s]) 
> automatically signals towards the kernel what to do with 
> old binaries - while extending the functionality of new 
> binaries, without sacrificing functionality.
> 
> With flags you are either limited to the same structure 
> size - or have to decode a 'size' value from the flags 
> value - which is fragile (and in which case a real 'size' 
> parameter is better).
> 
> in the perf ABI we use something like that: there's a 
> perf_attr.size parameter that iterates the ABI forward, 
> while still being binary compatible with older software.
> 
> If old binaries pass in a smaller structure to a newer 
> kernel then the kernel pads the new fields with zero by 
> default - that way the kernel internals are never burdened 
> with compatibility details and data format versions.
> 
> If new user-space passes in a large structure than the 
> kernel can handle then the kernel returns an error - this 
> way user-space can transparently support conditional 
> features and fallback logic.
> 
> It works really well, we've done literally a hundred perf 
> ABI extensions this way in the last 4+ years, in a pretty 
> natural fashion, without littering the kernel (or 
> user-space) with version legacies and without breaking 
> existing perf tooling.
> 
> Other syscall ABIs already get painful when trying to 
> handle 2-3 data structure versions, so people either give 
> up, or add flags kludges or go to new syscall entries: 
> which is painful in its own fashion and adds unnecessary 
> latency to feature introduction as well.
> 

Excellent. This now makes a lot of sense to me, thanks to your explanations,
Ingo.

I'll add the "size" field in the next revision.

Thanks,
Fam