Shared libraries are not necessary

[Barry Margolin]

In article <211 at titccy.cc.titech.ac.jp> mohta at necom830.cc.titech.ac.jp (Masataka Ohta) writes:
>In article <1991May17.075555.29787 at Think.COM> barmar at think.com writes:
>>There are two kinds of software upgrades: those which change the interface
>>to the library, and those which only change the implementation.
>And, my claim is that the second kind is almost nothing.
...
>Most software upgrade is a little more complex than can be processed by
>mere library change.

How many standard C and Unix library functions have had changes to their
interface?  I'm sure there have been some changes over the years to the
implementations of most of these library routines, but the interfaces have
stayed nearly constant.

>>I believe they also enhanced
>>the /etc/hosts version of the routine to return multiple addresses (our
>>/etc/hosts file has several host names with multiple addresses).
>
>You believe what? No. They didn't do so.

Thank you for correcting me.

>You should not have said what you believe. 

If I shouldn't say what I believe, what should I say?  Everything else is
things that I don't believe, and saying them would be lying.

>					    Instead, see
>lib/libc/net/hosttable/gethostnamadr.c of 4.3BSD source. There is only
>one IP address entry provided for gethostbyname().

I think I may have been confusing the /etc/hosts version with Sun's YP
version.  We don't use the DNS version of gethostbyname(), but I've been
told that our Suns will use both addresses.

>As you don't know enough about host name look up and accompanied complexities
>of its management, don't say:
>
>>Note that an upgrade to use the DNS didn't *have* to include support for
>>multiple addresses.
>
>unless you want to be dishonest.

What I meant was just as the /etc/hosts version ignores multiple addresses,
the designers of the DNS version could have decided to return only one
address, in order to leave the interface the same.

Note that the predecessor to the Domain system, the standard HOSTS.TXT
file, supported multiple host addresses; it's not new to the DNS, except in
the BSD implementation.  The original gethostbyname() designers clearly
screwed up when they had it only return one address.

>>In Multics, library design guidelines also make the first kind easy.  Most
>>library routines with non-trivial parameters take most of the parameters as
>>a pointer to a structure.  By convention, the first member of the structure
>>is a version number or version string
>
>The only parameter for gethostbyname() is, trivially, a string.

If the structure to be filled in were an in-out parameter, as we generally
did in Multics, then the version number could be put there.  For library
routines that allocate their output structure we would pass the expected
version as a separate parameter.  The Multics calling sequence also
includes the number of arguments, and often includes the argument types
(whenever the function is "varargs" or any of the arguments has variable
length), so a function could look at the number and types of arguments and
infer the version of the caller.

No, it's not easy to retrofit these calling conventions onto an established
system.  These problems are, in my opinion, a legacy of Unix's simplistic
original design.  On Multics, dynamic linking is used extensively, with
many users having personal versions of various library routines, and there
were few problems from these interface enhancements.  It just requires
reasonable system design.

-- 
Barry Margolin, Thinking Machines Corp.

barmar at think.com
{uunet,harvard}!think!barmar