On Wed, Feb 15, 2012 at 07:57:50AM +0000, Al Viro wrote:

On Wed, Feb 15, 2012 at 07:11:19AM +0000, Al Viro wrote:

quoted

One more thing - unless I'm misreading the code it looks like your ->sb becomes
a dangling pointer once fill_super finishes and ->statfs() is dereferencing
it...

BTW, what happens if root directory grows to more than one level of
pointers?  AFAICS, qnx6_checkroot() assumes that it won't happen...
And why not do that after getting the root inode, anyway?

Ho-hum...  I've added conversion of qnx6_iget() to pagecache and
checkroot move past getting root; result:
	* qnx6_bread/qnx6_getblk are gone
	* qnx6_block_map() has only one caller left
	* s_inodebits/s_dirent_blk/s_inodes_blk gone
s_ptrbits remains, of course, but that one has to survive - it determines
the way indirect blocks' tree is branched.

Remaining issues I see:
	* sbi->sb dangling pointer (those brelse() in success case of
fill_super()).  Probably only one of them needs to be dropped at that
point, leaving the other in sbi->sb_bh to be dropped by put_super
	* sb_set_blocksize() can and will invalidate buffer_heads of the
wrong size.  qnx6_fill_super() probably needs to redo sb_bread() on the
primary superblock after it has done the second sb_set_blocksize().
	* readdir handling of errors (especially in longname case) needs
to be done better.

Other than that... this stuff could use some tidying up, but that's not
a big deal.  I really wonder if these config options are worthwhile -
might be better to stick endianness and mmi to y unconditionally.  And
of course all those incrementals are completely untested and might very
well br broken.