18 Nov
2021
18 Nov
'21
7:38 p.m.
Unix 5th?? Gee, I guess I really must have been doing it...
Doug is not the only person who has pointed this out; that signals are
not 'reliable'.? I always thought some of the behavior was pure
baloney.? If somebody sends you a signal, then you should get it,
period, end of story.? And it shouldn't break your current system call,
either.? We're not talking about de-glitching a button.
I think the problem is really that what was implemented almost--but not
quite--filled a need, so it got used anyway. There is a place for short,
fast, conceptually simple to use messages.? All other kinds of IPC have
setup overhead and most have execution overhead.? Shared memory comes to
mind.? Not everything needs to be novel length if you are just alerting
a condition.
I was absolutely _delighted_ with reliable signals.? I like them even
more when you can send a byte or two with the signal.
I stop listening when I hear anything about not depending on them.? I
really do.? It's a purist position.
------------------------------------------------------------------------
On 11/18/21 1:04 PM, Dan Cross wrote:
The use of SIGHUP to reread a config file comes from, at least, Unix 5th Edition. I
suspect that the practice originated with the `init` process (usually PID 1) which, on
receipt of SIGHUP, would reread the ttys file to figure out what typewriters were attached
to the system. Prior to that, I suspect they'd have to reboot to add a new terminal
to the system, but things from first through 4th edition are not as well preserved as 5th
onwards, so its hard to tell.
By the time Berkeley was writing things like inetd, the practice was well known. Does
that make it a "standard?" Not in any formal sense, but a daemon couldn't
receive a SIGHUP in the usual way an interactive process might, so piggybacking some
special meaning on it makes a certain amount of sense.
Has it fallen out of a favor? A bit. As Doug McIllroy put it on the Unix Heritage
Society list a few years ago, signal() primarily existed to support SIGKILL, not as a
principled basis for IPC. It's racy and doing Real Work in the context of an upcall
from the kernel like a signal handler is fraught.
- Dan C.
> ------------------------------------------------------------------------
> On Thu, Nov 18, 2021 at 12:20 PM Dave McGuire<mcguire at neurotica.com>
wrote:
>
>
> Well, since the last release of SunOS (not Solaris) was in 1994, and the last
release of Ultrix was in 1995..
>> ------------------------------------------------------------------------
>> On 11/18/21 12:12 PM, Thomas DeBellis wrote:
>> They do it _today_.