On 08/02/2016 07:59 AM, Christoph Lameter wrote:
Hmm.... What SLUB does is: 1. Keep a count of the total number of allocated slab pages per node. This counter only needs to be updated when a slab page is allocated from the page allocator or when it is freed to the page allocator. At that point we already hold the per node lock, page allocator operations are extremely costly anyways and so that is ok. 2. Keep a count of the number of partially allocated slab pages per node. At that point we have to access the partial list and take a per node lock. Placing the counter into the same cacheline and the increment/decrement into the period when the lock has been taken avoids the overhead.
As Joonsoo mentioned in his previous comment, the partial list is pretty small anyway. And we cannot avoid traversal of the partial list - we have to count the number of active objects in each partial slab:
active_objs += page->active; So keeping a count of partially allocated slabs seems unnecessary to me.
The number of full pages is then total - partial If both allocators would use the same scheme here then the code to maintain the counter can be moved into mm/slab_common.c. Plus the per node structures could be mostly harmonized between both allocators. Maybe even the page allocator operations could become common code. Aruna: Could you work on a solution like that?
Yup, I'll replace the 3 counters with one counter for number of slabs per node and send out a new patch. I'll try to make the counter management as similar as possible, between SLAB and SLUB.
Thanks, Aruna -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>