CNFS offset problems

[Russ Allbery]

Russ Allbery <rra at Stanford.EDU> writes:

> Hm.  Not sure where to start looking for this one.  I guess I can go try
> to understand the code that detects that a cycbuff is full.

I hate C integer promotion.

In the comparison:

  if (article.len > cnfs->length - cnfs->free - CNFS_BLOCKSIZE - 1)

if cnfs->length - cnfs->free (which are both off_t), representing the free
space left in the cycbuff, is greater than CNFS_BLOCKSIZE + 1 (if, in
other words, there aren't at least two blocks free), the right hand side
is negative, since off_t is signed.

If off_t is larger than size_t (the type of article.len), which is the
common case for large files since off_t is then a 64-bit type, everything
gets promoted to off_t and this works fine.

If off_t and size_t are both the same size (32 bits or 64 bits), then the
signed type (off_t) is promoted to unsigned to match, and the negative
value on the right hand side becomes a large positive number, thus making
INN think there's room left in the cycbuff.

Bleh.

I'm committing a patch now.  This is why those "comparison between signed
and unsigned" warnings are important.

Hopefully I'll have a chance to clean up the CNFS code in INN 2.5.

-- 
Russ Allbery (rra at stanford.edu)             <http://www.eyrie.org/~eagle/>

    Please send questions to the list rather than mailing me directly.
     <http://www.eyrie.org/~eagle/faqs/questions.html> explains why.