Re: [PATCH] switch rangeset's lock to rwlock

[Keir Fraser <keir.xen@gmail.com>]

Jan Beulich wrote:
>>>> On 30.09.14 at 22:53,<keir.xen@gmail.com>  wrote:
>> Do the searches ever get long enough that a read lock helps? If any of
>> the rangesets is getting large and making the searches slow then it
>> would be quite easy to switch from linked list to red-black tree?
>
> As noted elsewhere, even very brief locking periods can cause
> convoys for many-vCPU guests. One case where we observe
> this is hvm_get_guest_time_fixed() (without clear route for
> mitigation as converting to rw lock is not an option here, and I
> didn't get around to try out whether eliminating the lock
> altogether in favor of atomic CPU operations would make this
> any better).

Given how rarely most of these rangesets get updated, it would be nice 
to let the guest itself have lock-free access at the expense of having 
to pause it to make modifications. Although, does a guest ever modify 
any of its own rangesets? I'm not sure that ever happens.

  -- Keir

> Jan
>