Re: kernel-side thread stack swapping

On 9/2/22, Konstantin Belousov <kostikbel@gmail.com> wrote:
> On Fri, Sep 02, 2022 at 04:11:40PM +0200, Mateusz Guzik wrote:
>> On 9/2/22, Konstantin Belousov <kostikbel@gmail.com> wrote:
>> > On Fri, Sep 02, 2022 at 02:05:37PM +0200, Mateusz Guzik wrote:
>> >> Is this really of practical use today?
>> >>
>> >> I have a WIP patch which needs to temporarily store something on the
>> >> stack and should things go wrong enough it will be accessed by UMA,
>> >> which can't handle the fault nor decide to skip the access.
>> >>
>> >> I can add something like td_pinstack or whatever to keep it around,
>> >> but perhaps the entire machinery can be just whacked?
>> > p_hold already does that.
>> >
>>
>> I only need to protect the one stack and more importantly don't want
>> to take the proc lock to bump p_hold (nor convert it to atomics), it's
>> all thread-local so to speak.
>
> You do not want to take proc lock, or cannot?  Note that only sleeping
> thread' stack can be swapped out.
>

To add some context here I'm looking at reworking vnode batching in
vdrop -> vdbatch_enqueue to remove vnode interlock -> vdbatch lock ->
vnode list lock dependency (and improve scalability of the thing).

Adding a proc lock here would negatively affect performance for
everyone *and* weirdly serialize same-proc consumers.

-- 
Mateusz Guzik <mjguzik gmail.com>