Re: [RFC PATCH v2 14/15] mm: Add alloc-free handshake to triggermemory region compaction
From: Cody P Schafer
Date: Wed Apr 10 2013 - 19:26:22 EST
On 04/09/2013 02:48 PM, Srivatsa S. Bhat wrote:
We need a way to decide when to trigger the worker threads to perform
region evacuation/compaction. So the strategy used is as follows:
Alloc path of page allocator:
----------------------------
This accurately tracks the allocations and detects the first allocation
in a new region and notes down that region number. Performing compaction
rightaway is not going to be helpful because we need free pages in the
lower regions to be able to do that. And the page allocator allocated in
this region precisely because there was no memory available in lower regions.
So the alloc path just notes down the freshly used region's id.
Free path of page allocator:
---------------------------
When we enter this path, we know that some memory is being freed. Here we
check if the alloc path had noted down any region for compaction. If so,
we trigger the worker function that tries to compact that memory.
Also, we avoid any locking/synchronization overhead over this worker
function in the alloc/free path, by attaching appropriate semantics to the
available status flags etc, such that we won't need any special locking
around them.
Can you explain why avoiding locking works in this case?
It appears the lack of locking is only on the worker side, and the
mem_power_ctrl is implicitly protected by zone->lock on the alloc & free
side.
In the previous patch I see smp_mb(), but no explanation is provided for
why they are needed. Are they related to/necessary for this lack of locking?
What happens when a region is passed over for compaction because the
worker is already compacting another region? Can this occur? Will the
compaction re-trigger appropriately?
I recommend combining this patch and the previous patch to make the
interface more clear, or make functions that explicitly handle the
interface for accessing mem_power_ctrl.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/