Path: utzoo!attcan!uunet!wuarchive!udel!mmdf
From: archer%segin4.segin.fr@prime.com (Vincent Archer)
Newsgroups: comp.os.minix
Subject: Re: A few bugs in 1.5.5: named pipes, sh(1) read
Message-ID: <16274@nigel.udel.EDU>
Date: 9 Apr 90 13:28:43 GMT
Sender: mmdf@udel.EDU
Lines: 30

Michael S. Mattone <msm@NUCLEUS.MI.ORG> says:

MSM>1. Pipe limit 7K.  The following fails if 'fd' is a pipe:
MSM>        write(fd, buf, 8000);
MSM>   This bug (feature?) has been annoying me since 1.2.  Anyone else?
MSM>   I posted my modifications to make this problem go away 2 1/2 times
MSM>   before.  Should I try for 3 1/2?  Or is this something POSIX requires?

Better yet, it fails with a specific error code saying that it's too big (8000
that is). That's because Unix requires writes to be atomic. Otherwise, you
could find (assuming two processes writing onto the same pipe) the following
situation:

    P1 says: write(8000)
             the pipe accepts 7K, and P1 waits
    P1 gets: signal -9
             and is killed. FS notices this and purges the remaining 1K
    P2 says: read(8000)
             and gets the 7K, but never the 1K (missing)

In that situation, you get a partial block of data. If you're using your pipe
as a formatted packet stream, you wind up with a non-terminated packet. This
can get annoying sometimes, and Unix therefore tries to forbid a write to be
done in two (or more) chunks. Do not correct this behaviour, it's standard
practice on Unix.


    Vincent Archer

Email: archer%segin4.segin.fr@prime.com     (Yes, indirect address required!)