[RFC] problem with grow_dev_page()/readpage()

[Alexander Viro]

	Linus, looks like the logics with "if we got stale buffer_heads -
no problem, grow_dev_page() will get rid of them" is broken.

	Look at block_read_full_page().  If it sees ->buffers != NULL, it
assumes that buffer size corresponds to ->i_blkbits.  IOW, if we ever
trigger the "let's free old buffer heads" path in grow_dev_page() - we
were in trouble; if somebody would read(2) from that place he'd end up
calling block_read_full_page() and it would screw up.

	So either we need to replace if (!try_to_free_buffers(page, GFP_NOFS))
with unconditional BUG() and make sure that it doesn't get triggered, or
we need to start pulling similar bh resizing code into block_read_full_page()
and friends.

	Comments?