Re: [PATCH] mm/vmalloc: Introduce DEBUG_VMALLOCINFO to reduce spinlock contention

[Richard Yao]

On 04/10/2014 12:51 PM, Andi Kleen wrote:
> Richard Yao <ryao@xxxxxxxxxx> writes:
> 
>> Performance analysis of software compilation by Gentoo portage on an
>> Intel E5-2620 with 64GB of RAM revealed that a sizeable amount of time,
>> anywhere from 5% to 15%, was spent in get_vmalloc_info(), with at least
>> 40% of that time spent in the _raw_spin_lock() invoked by it.
> 
> I don't think that's the right fix. We want to be able 
> to debug kernels without having to recompile them.

There are plenty of other features for debugging the VM subsystem that
are disabled in production kernels because they are too expensive. I see
no reason why this should not be one of them.

If someone reading this has a use for this functionality in production
systems, I would love to hear about it. I am having trouble finding uses
for this in production.

That being said, we are clearly spending plenty of time blocked on list
traversal. I imagine that we could use an extent tree to track free
space for even bigger gains, but I have difficulty seeing why
/proc/vmallocinfo should be available on non-debug kernels. Allowing
userland to hold a critical lock indefinitely on production systems is a
deadlock waiting to happen.

> And switching locking around dynamically like this is very
> ugly and hard to maintain.

I welcome suggestions on how to make the changes I have made in this
patch more maintainable.

> Besides are you sure the spin lock is not needed elsewhere?
> 
> How are writers to the list protected?

The spinlock is needed elsewhere, but not to protect this list.
Modifications to this list are done under RCU. The only thing stopping
RCU from being enough to avoid a spinlock is /proc/vmallocinfo, which
does locking to prevent modification while userland is reading the list.

Attachment: signature.asc
Description: OpenPGP digital signature