[lc-checkins] CVS: linux/mm oom_kill.c,1.6,NONE

[Rodrigo S. de C. <rc...@us...>]

Update of /cvsroot/linuxcompressed/linux/mm
In directory usw-pr-cvs1:/tmp/cvs-serv10033/mm

Removed Files:
	oom_kill.c 
Log Message:
This batch of changes includes cleanups and code rewrites. The
swap_out_fragments() and find_free_swap_buffer() are much simpler and
also more efficient. Some preliminary tests showed a performance gain
due to these improvements.

- Fragment Freed Bit removed: the current code has been checked by and
the few parts in the code where we could sleep with a comp page in
locked state were rewritten (mainly swap out). It turns out that now
we won't free a fragment without locking its page, and thus we don't
need a bit to tell that the fragment needs to be freed. All special
cases for Freed bit all over the code were removed.

- Fragment SwapBuffer Bit removed: things changed a lot in swap buffer
code. We don't add a "virtual" fragment to the comp cache hash table
to avoid access to fragments which may be in swap out process. So, no
need for this bit currently.

- Fragment IO Bit added: instead of added a "virtual" fragment to the
comp cache hash table, we don't free the fragment we are swapping out
right after the IO function (rw_swap_page) has been called. We free it
only when the IO has finished. But once the IO takes time, this
fragment can be freed (maybe because it's been swapped in) in the
meanwhile and so we have to tell the comp_cache_free() that this
fragment is being written to disk and that its struct should not be
freed. comp_cache_free() also clears this bit since we don't need to
perform the IO (if it has not yet been submitted). In the case it has
already been submitted, we don't have to free this fragment, only its
struct.

- find_free_swp_buffer() function was completely rewritten. There are
two lists linking all the swap buffer pages: used and free. The comp
page field in swp_buffer_t was removed. No array for swap buffers is
needed, since they are linked in these lists (it's the first step to
make it work with dynamic swap buffers, like listed on todo list).

  About its behaviour, there's no variable which counts the number of
free swap buffers like before. So, once all the buffers were used, we
try to move all the unlocked buffers to the free list at once. If we
couldn't move even one, them we wait for one page to have its IO
finished and go get this swap buffer. Our old code didn't perform this
task so efficiently. For example, everytime we had a free fragment, we
checked all the fragments to see if they were locked. With a used list
we avoid this kind of overhead.

- swap_out_fragments() was rewritten too. Part of the code was moved
to decompress_to_swp_buffer() function. Since we don't have to worry
about Freed bit, the code is much simpler.

- find_*_comp_cache() functions don't have special cases for Freed
fragments.

- once there's no variable number_of_free_swp_buffers and we don't
hold a referente to the swap entry in swap_out_fragments(), there's no
need to update swap buffers in end_buffer_io_async().



--- oom_kill.c DELETED ---