1. Current code can't always detect sequential reading, in case
   when read size is not PAGE_CACHE_SIZE aligned.

   If application reads the file by 4096+512 chunks, we have:
   1st read: first read detected, prev_page = 2.
   2nd read: offset == 2, the read is considered random.

   page_cache_readahead() should treat prev_page == offset as
   sequential access. In this case it is better to ++offset,
   because of blockable_page_cache_readahead(offset, size).

2. If application reads 4096 bytes with *ppos == 512, we have to
   read 2 pages, but req_size == 1 in do_generic_mapping_read().

   Usually it's not a problem. But in random read case it results
   in unnecessary page cache misses.

~$ time dd conv=notrunc if=/tmp/GIG of=/tmp/dummy bs=$((4096+512))

2.6.11-clean:	real=370.35 user=0.16 sys=14.66
2.6.11-patched:	real=234.49 user=0.19 sys=12.41
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>