From: eliz AT is DOT elta DOT co DOT il
Message-Id: <200005270638.JAA08011@mailgw1.netvision.net.il>
Date: Sat, 27 May 2000 09:37:26 +0200
To: pavenis AT lanet DOT lv
Cc: djgpp-workers AT delorie DOT com
X-Mailer: WinNT's Blat ver 1.8.5b http://www.interlog.com/~tcharron
In-reply-to: <392EDC2D.15119.171CE80@localhost> (pavenis@lanet.lv)
Subject: Re: gcc djgpp stopped working
References: <392EBD54 DOT 13380 DOT F946F1 AT localhost> (pavenis AT lanet DOT lv) <392EDC2D DOT 15119 DOT 171CE80 AT localhost>
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

--text follows this line--
> From: pavenis AT lanet DOT lv
> Date: Fri, 26 May 2000 20:18:53 +0200
> 
> Let assume gcc tries to create file in TMPDIR=c:/tmp and directory 
> with the same name already exist there. Then it gets EPERM

Sorry, I don't understand: EPERM should never be returned in DJGPP,
except for some very obscured network-related problems (see doserr2e.c
in the library).  Perhaps you meant EACCES?

> That 
> is rather unlikely as biostime() is used to generate first file name (and 
> it's later incremented).

Where do you see the call to biostime?  Is it something that GCC
sources do?  I don't see biostime in mkstemp.

> In this case (BUG) it's assumed that directory is 
> bad and it quits the loop. Function mkstemp() in libc also has this 
> problem (if it gets EPERM it quits the loop)

But if the file or directory by that name exists, mktemp will see it,
because it calls __file_exists, and so it will never return a name of
an existing file.  What you describe can only happen in a race
condition, when a directory is created *after* mktemp returned a name
of a file that didn't exist a moment before.

> Perhaps mkstemp should use stat() if it gets EPERM to verify whether
> file (or directory) exists and and only give up if it's doen't exist.

stat is too expensive.  We could add EACCES to EEXIST, or we could
explicitly check with _chmod to cover the directory case.

In any case, I think GCC should not abort, but print an error message
in these cases.