Re: [PATCH] kernel: Introduce a write lock/unlock wrapper for tasklist_lock

[Aiqun Yu (Maria)]

On 1/4/2024 2:18 AM, Matthew Wilcox wrote:
> On Wed, Jan 03, 2024 at 10:58:33AM +0800, Aiqun Yu (Maria) wrote:
> > On 1/2/2024 5:14 PM, Matthew Wilcox wrote:
> > > > > -void __lockfunc queued_write_lock_slowpath(struct qrwlock *lock)
> > > > > +void __lockfunc queued_write_lock_slowpath(struct qrwlock *lock, bool irq)
> > > > >     {
> > > > >     	int cnts;
> > > > > @@ -82,7 +83,11 @@ void __lockfunc queued_write_lock_slowpath(struct qrwlock *lock)
> > > > Also a new state showed up after the current design:
> > > > 1. locked flag with _QW_WAITING, while irq enabled.
> > > > 2. And this state will be only in interrupt context.
> > > > 3. lock->wait_lock is hold by the write waiter.
> > > > So per my understanding, a different behavior also needed to be done in
> > > > queued_write_lock_slowpath:
> > > >     when (unlikely(in_interrupt())) , get the lock directly.
> > >
> > > I don't think so.  Remember that write_lock_irq() can only be called in
> > > process context, and when interrupts are enabled.
> > In current kernel drivers, I can see same lock called with write_lock_irq
> > and write_lock_irqsave in different drivers.
> >
> > And this is the scenario I am talking about:
> > 1. cpu0 have task run and called write_lock_irq.(Not in interrupt context)
> > 2. cpu0 hold the lock->wait_lock and re-enabled the interrupt.
>
> Oh, I missed that it was holding the wait_lock.  Yes, we also need to
> release the wait_lock before spinning with interrupts disabled.
>
> > I was thinking to support both write_lock_irq and write_lock_irqsave with
> > interrupt enabled together in queued_write_lock_slowpath.
> >
> > That's why I am suggesting in write_lock_irqsave when (in_interrupt()),
> > instead spin for the lock->wait_lock, spin to get the lock->cnts directly.
>
> Mmm, but the interrupt could come in on a different CPU and that would
> lead to it stealing the wait_lock from the CPU which is merely waiting
> for the readers to go away.
That's right.
The fairness(or queue mechanism) wouldn't be ensured (only in interrupt 
context) if we have the special design when (in_interrupt()) spin to get 
the lock->cnts directly. When in interrupt context, the later 
write_lock_irqsave may get the lock earlier than the write_lock_irq() 
which is not in interrupt context.

This is a side effect of the design, while similar unfairness design in 
read lock as well. I think it is reasonable to have in_interrupt() 
waiters get lock earlier from the whole system's performance of view.
>

--
Thx and BRs,
Aiqun(Maria) Yu