Re: [PATCH 2.6.20-rc1 00/10] Kernel memory leak detector 0.13

[Ingo Molnar <mingo@elte.hu>]

* Catalin Marinas <catalin.marinas@gmail.com> wrote:

> >> [...] It could be so simple that it would never need to free any
> >> pages, just grow the size as required and reuse the freed memleak
> >> objects from a list.
> >
> >sounds good to me. Please make it a per-CPU pool.
> 
> Isn't there a risk for the pools to become imbalanced? A lot of 
> allocations would initially happen on the first CPU.

hm, what's the problem with imbalance? These are trees and imbalance 
isnt a big issue.

> >[...] (Add a memleak_object->cpu pointer so that freeing can be done 
> >on any other CPU as well.)
> 
> We could add the freed objects to the CPU pool where they were freed 
> and not use a memleak_object->cpu pointer.

i mean totally per-CPU locking and per-CPU radix trees, etc.

> > We'll have to fix the locking too, to be per-CPU - memleak_lock is 
> > quite a scalability problem right now.
> 
> The memleak_lock is indeed too coarse (but it was easier to track the 
> locking dependencies). With a new allocator, however, I could do a 
> finer grain locking. It probably still needs a (rw)lock for the hash 
> table. Having per-CPU hash tables is inefficient as we would have to 
> look up all the tables at every freeing or scanning for the 
> corresponding memleak_object.

at freeing we only have to look up the tree belonging to object->cpu. 
Scanning overhead does not matter in comparison to runtime tracking 
overhead. (but i doubt it would be much different - scanning overhead 
scales with size of tree)

> There is a global object_list as well covered by memleak_lock (only 
> for insertions/deletions as traversing is RCU). [...]

yeah, that would have to become per-CPU too.

> [...] List insertion/deletion is very small compared to the hash-table 
> look-up and it wouldn't introduce a scalability problem.

it's a common misconception to think that 'small' critical sections are 
fine. That's not the issue. The pure fact of having globally modified 
resource is the problem, the lock cacheline would ping-pong, etc.

	Ingo