On Mon, 5 Aug 1996, Peter P. Eiserloh wrote:
We need to keep a clear the concept of threads. Too many people
seem to confuse a thread with a process. The following discussion
does not reflect the current state of linux, but rather is an
attempt to stay at a high level discussion.
NO!
There is NO reason to think that "threads" and "processes" are separate
entities. That's how it's traditionally done, but I personally think it's a
major mistake to think that way. The only reason to think that way is
historical baggage.
Both threads and processes are really just one thing: a "context of
execution". Trying to artificially distinguish different cases is just
self-limiting.
A "context of execution", hereby called COE, is just the conglomerate of
all the state of that COE. That state includes things like CPU state
(registers etc), MMU state (page mappings), permission state (uid, gid)
and various "communication states" (open files, signal handlers etc).
Traditionally, the difference between a "thread" and a "process" has been
mainly that a threads has CPU state (+ possibly some other minimal state),
while all the other context comes from the process. However, that's just
one way of dividing up the total state of the COE, and there is nothing
that says that it's the right way to do it. Limiting yourself to that kind of
image is just plain stupid.
The way Linux thinks about this (and the way I want things to work) is that
there is no such thing as a "process" or a "thread". There is only the
totality of the COE (called "task" by Linux). Different COE's can share parts
of their context with each other, and one subset of that sharing is the
traditional "thread"/"process" setup, but that should really be seen as ONLY
a subset (it's an important subset, but that importance comes not from
design, but from standards: we obviusly want to run standards-conforming
threads programs on top of Linux too).
In short: do NOT design around the thread/process way of thinking. The
kernel should be designed around the COE way of thinking, and then the
pthreads library can export the limited pthreads interface to users who
want to use that way of looking at COE's.
Just as an example of what becomes possible when you think COE as opposed
to thread/process:
- You can do a external "cd" program, something that is traditionally
impossible in UNIX and/or process/thread (silly example, but the idea
is that you can have these kinds of "modules" that aren't limited to
the traditional UNIX/threads setup). Do a:
clone(CLONE_VM|CLONE_FS);
child: execve("external-cd");
/* the "execve()" will disassociate the VM, so the only reason we
used CLONE_VM was to make the act of cloning faster */ - You can do "vfork()" naturally (it meeds minimal kernel support, but
that support fits the CUA way of thinking perfectly):
clone(CLONE_VM);
child: continue to run, eventually execve()
mother: wait for execve - you can do external "IO deamons":
clone(CLONE_FILES);
child: open file descriptors etc
mother: use the fd's the child opened and vv.
All of the above work because you aren't tied to the thread/process way of
thinking. Think of a web server for example, where the CGI scripts are done
as "threads of execution". You can't do that with traditional threads,
because traditional threads always have to share the whole address space, so
you'd have to link in everything you ever wanted to do in the web server
itself (a "thread" can't run another executable).
Thinking of this as a "context of execution" problem instead, your tasks can
now chose to execute external programs (= separate the address space from the
parent) etc if they want to, or they can for example share everything with
the parent except for the file descriptors (so that the sub-"threads" can
open lots of files without the parent needing to worry about them: they close
automatically when the sub-"thread" exits, and it doesn't use up fd's in the
parent).
Think of a threaded "inetd", for example. You want low overhead fork+exec, so
with the Linux way you can instead of using a "fork()" you write a
multi-threaded inetd where each thread is created with just CLONE_VM (share
address space, but don't share file descriptors etc). Then the child can
execve if it was a external service (rlogind, for example), or maybe it was
one of the internal inetd services (echo, timeofday) in which case it just
does it's thing and exits.
You can't do that with "thread"/"process".
Linus