On Mon, 10 Feb 2014, Raghavendra K T wrote:
As you rightly pointed , I 'll drop remote memory term and use
something like :
"* Ensure readahead success on a memoryless node cpu. But we limit
* the readahead to 4k pages to avoid trashing page cache." ..
I don't know how to proceed here after pointing it out twice, I'm afraid.
numa_mem_id() is local memory for a memoryless node. node_present_pages()
has no place in your patch.
Regarding ACCESS_ONCE, since we will have to add
inside the function and still there is nothing that could prevent us
getting run on different cpu with a different node (as Andrew ponted), I have
not included in current patch that I am posting.
Moreover this case is hopefully not fatal since it is just a hint for
readahead we can do.
I have no idea why you think the ACCESS_ONCE() is a problem. It's relying
on gcc's implementation to ensure that the equation is done only for one
node. It has absolutely nothing to do with the fact that the process may
be moved to another cpu upon returning or even immediately after the
calculation is done. Is it possible that node0 has 80% of memory free and
node1 has 80% of memory inactive? Well, then your equation doesn't work
quite so well if the process moves.
There is no downside whatsoever to using it, I have no idea why you think
it's better without it.
So there are many possible implementation:
(1) use numa_mem_id(), apply freepage limit and use 4k page limit for all
case
(Jan had reservation about this case)
(2)for normal case: use free memory calculation and do not apply 4k
limit (no change).
for memoryless cpu case: use numa_mem_id for more accurate
calculation of limit and also apply 4k limit.
(3) for normal case: use free memory calculation and do not apply 4k
limit (no change).
for memoryless case: apply 4k page limit
(4) use numa_mem_id() and apply only free page limit..
So, I ll be resending the patch with changelog and comment changes
based on your and Andrew's feedback (type (3) implementation).
It's frustrating to have to say something three times. Ask yourself what
happens if ALL NODES WITH CPUS DO NOT HAVE MEMORY?