Re: [PATCH 00/11] Update for linux v5.1

[Szabolcs Nagy <nsz@port70.net>]

* Rich Felker <dalias@libc.org> [2019-05-21 19:11:47 -0400]:
> On Sun, May 12, 2019 at 04:53:24PM +0200, Szabolcs Nagy wrote:
> > From f6cc4cdf345d1785659ed9b59f3a2827c3fd3c53 Mon Sep 17 00:00:00 2001
> > From: Szabolcs Nagy <nsz@port70.net>
> > Date: Thu, 9 May 2019 19:20:17 +0000
> > Subject: [PATCH 06/11] sys/socket.h: add new SO_ definitions from linux v5.1
> > 
> > the new definitions are from commits
> > 
> >   linux commit a9beb86ae6e55bd92f38453c8623de60b8e5a308
> >   sock: Add SO_RCVTIMEO_NEW and SO_SNDTIMEO_NEW
> > 
> >   linux commit 45bdc66159d49bfc7f75fe02d25bc74f5d2660cf
> >   socket: Rename SO_RCVTIMEO/ SO_SNDTIMEO with _OLD suffixes
> > 
> >   linux commit 9718475e69084de15c3930ce35672a7dc6da866b
> >   socket: Add SO_TIMESTAMPING_NEW
> > 
> >   linux commit 887feae36aee6c08e0dafcdaa5ba921abbb2c56b
> >   socket: Add SO_TIMESTAMP[NS]_NEW
> > 
> >   linux commit 7f1bc6e95d7840d4305595b3e4025cddda88cee5
> >   sockopt: Rename SO_TIMESTAMP* to SO_TIMESTAMP*_OLD
> > 
> >   linux commit f5dd3d0c9638a9d9a02b5964c4ad636f06cf7e2c
> >   net: introduce SO_BINDTOIFINDEX sockopt
> > 
> > SO_BINDTOIFINDEX behaves similar to SO_BINDTODEVICE, but takes a
> > network interface index as argument, rather than the network
> > interface name.
> > 
> > _NEW socket options got introduced for 64bit time_t support on 32bit
> > targets, in musl the affected socket options are always the same as
> > the _OLD values since different time_t requires a new abi, so the
> > _OLD vs _NEW dispatch logic was not copied from the uapi headers.
> 
> Are the _OLD and _NEW ones intended to be application-facing APIs? I
> think this is a question both of the kernel's intent and our intent.
> When we add 64-bit time_t, however we do it, we'll need to emulate the
> _NEW version on old kernels that lack it, and it's not clear that we'd
> want to expose the _OLD version at all.
> 
> Perhaps we should split this patch into the time64 stuff and the other
> changes so that the other changes can be applied right away if there's
> further discussion to be had about time64.

ok i can split this.

i think the intent is that the public api is SO_FOO,
the kernel defines SO_FOO_OLD and SO_FOO_NEW for the
old and new time_t abi respectively and then SO_FOO
is defined using those like

#define SO_FOO (sizeof(long)==4 && sizeof(time_t)==8 ? SO_FOO_NEW : SO_FOO_OLD)

or similar, since on glibc the abi is based on feature
test macros, but for us SO_FOO is always SO_FOO_OLD
so it seemed simplest to just introduce the _OLD, _NEW
names and leave the SO_FOO definitions alone.

(i'm not sure what SO_FOO_NEW means on 64bit targets
where time_t is already 64bit)

> > From 01dcbc761ace09ddc7295e14f33d95bc32a15318 Mon Sep 17 00:00:00 2001
> > From: Szabolcs Nagy <nsz@port70.net>
> > Date: Thu, 9 May 2019 22:49:28 +0000
> > Subject: [PATCH 10/11] ipc: prefer SYS_ipc when it is defined
> > 
> > Linux v5.1 introduced ipc syscalls on targets where previously only
> > SYS_ipc was available, change the logic such that the ipc code keeps
> > using SYS_ipc which works backward compatibly on older kernels.
> 
> I think this patch changes behavior at least on microblaze, which
> already had both. A casual grep of arch/*/bits/syscall.h.in for
> __NR_ipc and __NR_semget suggests there are no others affected, so
> perhaps it's no big deal. Thoughts?

i didnt notice that, i guess microblaze works either way.

it's also possible to change to the new calls and fall back
to SYS_ipc when that's available, however i think semtimedop
will only have 64bit time_t variant on 32bit targets and
IPC_64 should not be passed explicitly to the new syscalls,
so dealing with those seemed to be messy.