>There've been troubles (particularly in the logging system) where
>deleting a file, creating a new one, and then crashing before the
>directory updates themselves get out to disk causes havoc such as you
>describe.
>
>Search the bug database. In any event, Casper's right. It's possible
>to make applications that can deal with system halts at any point, and
>still keep the data self-consistent. That some applications don't get
>it right doesn't mean that none can do it.
And there was buggy code in Solaris; e.g., this code in drvsubr.c
which manipulates some of the /etc/*_* files:
340 /* Make sure that the file is on disk */
341 if (fflush(newfp) != 0 || fsync(fileno(newfp)) != 0)
342 status = ERROR;
343 else
344 rv = NOERR;
345
346 (void) fclose(newfp);
used to look like this:
(void) fsync(fileno(newfp));
(void) fclose(newfp);
Considering that most of these files are less than 8K in size, no output
was ever written before the fsync().
So no surprise that some files in /etc sometimes turned up with bogus
content.
Casper
