From: cgf AT cygnus DOT com (Christopher Faylor)
Subject: Re: SIGKILL and blocking
15 Jun 1998 13:06:09 -0700
Message-ID: <199806151923.PAA28485.cygnus.cygwin32.developers@kramden.cygnus.com>
To: cgf AT cygnus DOT com, newsham AT lava DOT net
Cc: cygwin32-developers AT cygnus DOT com

>From: newsham AT lava DOT net (Tim Newsham)
>Date: Mon, 15 Jun 1998 08:49:43 -1000 (HST)
>
>> Finally the light dawns.  I never actually looked at the code, and
>> assumed that WSACleanup (despite Sergey's explicit reference) was being
>> called unconditionally in exit().  I can see now that that is not the
>> case.  It was only being called for implicit calls, as he repeatedly
>> tried to tell me.
>> 
>> So, a winsock call *does* block and the WSACleanup call is essentially
>> demonstrating that the winsock code is not 100% thread safe.
>> 
>> In my version of cygwin, all reads occur in a separate thread.  That
>> means that the main thread is always interruptible.  This could be
>> extended to blocking socket operations as well.  Besides read what other
>> operations block?  listen?  select?
>
>scanning over i_* pointers in net.cc:
>   accept, recv, recvfrom, connect, gethostbyname, gethostbyaddr
>
>   send and sendto might block in some situations (ie. TCP flow control
>     window full, and write performed).
>
>   rcmd and rexec probably block, I haven't used the winsock versions yet.
>
>   (select blocks, but we always call socket select in a seperate thread)

Hmm.  That's a lot of calls.

What I was thinking about doing was having one extra thread per user
thread.  If the main thread needs to perform an operation that would block,
it would either start or continue a separate thread which would handle
the blocking operation.  The main thread would wait for either the
secondary thread to finish or for a signal to occur.  If a signal occurs
it terminates the secondary thread, cleans up its stack space and
returns an EINTR.

I've got this working in my version of cygwin but I haven't had time to
really run it through its paces.

Does anyone have experience with threads and signals under UNIX?  How
does a non-main thread deal with signals?  Are non-main threads interrupted
at all?  Does signal() effect threads on a per-thread basis?

cgf