Messages in this thread | | | Date | Thu, 14 Oct 2010 00:23:13 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [PATCH net-next] net: allocate skbs on local node |
| |
On Thu, 14 Oct 2010, Pekka Enberg wrote:
> Argh. Why would we want to introduce something that's effectively a new > allocator based on SLUB? If there's something controversial in the current > patch series, lets just keep it out of mainline. A "rewrite" is the reason > we're in this mess so lets not repeat the same mistake again! >
SLUB is a good base framework for developing just about any slab allocator you can imagine, in part because of its enhanced debugging facilities. Nick originally developed SLQB with much of the same SLUB framework and the queueing changes that Christoph is proposing in his new unified allocator builds upon SLUB.
Instead of the slab.c, slab_queue.c, and slab_nonqueue.c trifecta, I suggested building as much of the core allocator into a single file as possible and then extending that with a config option such as CONFIG_SLAB_QUEUEING, if possible. Christoph knows his allocator better than anybody so he'd be the person to ask if this was indeed feasible and, if so, I think it's in the best interest of a long-term maintainable kernel.
I care about how this is organized because I think the current config option demanding users select between SLAB and SLUB without really understanding the differences (especially for users who run a very wide range of applications and the pros and cons of better microbenchmark results for one allocator over another isn't at all convincing) is detrimental.
| |