Re: Regression introduced by bfcfaa77bdf0f775263e906015982a608df01c76 (vfs: use 'unsigned long' accesses for dcache name comparison and hashing)

[Al Viro <viro@ZenIV.linux.org.uk>]

On Thu, Mar 22, 2012 at 08:09:19PM +0000, Al Viro wrote:

> Interesting... that's exactly 8 characters.  Oh, I see - hash_name() gets
> an extra multiplication by 9 in this case.  Look: full_name_hash() will
> handle the first word, decrement len by 8, set hash to <first word> and
> bugger off on !len.  hash_name(), OTOH, will go through the loops once,
> with hash and a both 0.  hash stays 0, a becomes <first word>.  No NUL or
> / in it, so in we go again; hash becomes a * 9, i.e. <first word> * 9.
> a becomes the second word, with mask != 0.  And we are out of the loop,
> and proceed to add nothing to hash (the name is over at that point).  As
> the result, we get hash mismatch for names that are 8 bytes long or
> multiple thereof.

OK, full_name_hash()/hash_name() definitely have a mismatch and it's on the
names of length 8*n: trivial experiment shows that we have
name hash_name full_name_hash
a 61 61
ab 6261 6261
abc 636261 636261
abcd 64636261 64636261
abcdabc 64c6c4c2 64c6c4c2
abcdabcd efcead5 c8c6c4c2
abcdabcd9 efceb0e efceb0e

Linus, which way do you prefer to shift it?  Should hash_name() change to
match full_name_hash() or should it be the other way round?

What happens is that you get multiplication by 9 and adding 0 in the former,
after having added the last full word.  In the latter we add the last full
word, see that there's nothing left and bugger off.