2015-11-03 02:08:20

by Yunhong Jiang

[permalink] [raw]
Subject: Can srcu_read_lock() be called from interrupt context?

Hi, Paul
I have a question to the srcu_read_lock(). Can it be invoked on
interrupt context? According to
http://lxr.free-electrons.com/source/kernel/rcu/srcu.c#L292 it can only be
called from process context, but according to
http://lxr.free-electrons.com/source/include/linux/srcu.h#L213 seems it's ok
from irq context only if it matches with the srcu_read_unlock. Can you
please give some hints?

Thanks
--jyh


2015-11-03 09:55:47

by Paul E. McKenney

[permalink] [raw]
Subject: Re: Can srcu_read_lock() be called from interrupt context?

On Mon, Nov 02, 2015 at 05:56:56PM -0800, Yunhong Jiang wrote:
> Hi, Paul
> I have a question to the srcu_read_lock(). Can it be invoked on
> interrupt context? According to
> http://lxr.free-electrons.com/source/kernel/rcu/srcu.c#L292 it can only be
> called from process context, but according to
> http://lxr.free-electrons.com/source/include/linux/srcu.h#L213 seems it's ok
> from irq context only if it matches with the srcu_read_unlock. Can you
> please give some hints?

Adding Lai Jianshan for his thoughts.

I believe that srcu.h is correct, at least assuming that interrupts do
not nest too deeply. (If they were to nest four billion deep, then the
->seq[] counter could overflow, defeating the checks, but the CPU stack
would have overflowed long before.)

Lai, am I missing anything here?

Thanx, Paul

2015-11-03 14:47:22

by Peter Zijlstra

[permalink] [raw]
Subject: Re: Can srcu_read_lock() be called from interrupt context?

On Tue, Nov 03, 2015 at 01:55:39AM -0800, Paul E. McKenney wrote:
> On Mon, Nov 02, 2015 at 05:56:56PM -0800, Yunhong Jiang wrote:
> > Hi, Paul
> > I have a question to the srcu_read_lock(). Can it be invoked on
> > interrupt context? According to
> > http://lxr.free-electrons.com/source/kernel/rcu/srcu.c#L292 it can only be
> > called from process context, but according to
> > http://lxr.free-electrons.com/source/include/linux/srcu.h#L213 seems it's ok
> > from irq context only if it matches with the srcu_read_unlock. Can you
> > please give some hints?
>
> Adding Lai Jianshan for his thoughts.
>
> I believe that srcu.h is correct, at least assuming that interrupts do
> not nest too deeply. (If they were to nest four billion deep, then the
> ->seq[] counter could overflow, defeating the checks, but the CPU stack
> would have overflowed long before.)

It seems like a strange constraint to me; not being able to use
srcu_read_lock() from IRQ (or even NMI) context. And looking at the
various implementations of it nothing ever prohibited this.

While srcu _allows_ for sleeping while holding the read side primitives,
it is not required at all.

So I would suggest amending the comment and RCU/checklist.txt.

Even call_srcu() should be IRQ-safe.

2015-11-03 15:36:23

by Paul E. McKenney

[permalink] [raw]
Subject: Re: Can srcu_read_lock() be called from interrupt context?

On Tue, Nov 03, 2015 at 03:47:17PM +0100, Peter Zijlstra wrote:
> On Tue, Nov 03, 2015 at 01:55:39AM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 02, 2015 at 05:56:56PM -0800, Yunhong Jiang wrote:
> > > Hi, Paul
> > > I have a question to the srcu_read_lock(). Can it be invoked on
> > > interrupt context? According to
> > > http://lxr.free-electrons.com/source/kernel/rcu/srcu.c#L292 it can only be
> > > called from process context, but according to
> > > http://lxr.free-electrons.com/source/include/linux/srcu.h#L213 seems it's ok
> > > from irq context only if it matches with the srcu_read_unlock. Can you
> > > please give some hints?
> >
> > Adding Lai Jianshan for his thoughts.
> >
> > I believe that srcu.h is correct, at least assuming that interrupts do
> > not nest too deeply. (If they were to nest four billion deep, then the
> > ->seq[] counter could overflow, defeating the checks, but the CPU stack
> > would have overflowed long before.)
>
> It seems like a strange constraint to me; not being able to use
> srcu_read_lock() from IRQ (or even NMI) context. And looking at the
> various implementations of it nothing ever prohibited this.
>
> While srcu _allows_ for sleeping while holding the read side primitives,
> it is not required at all.
>
> So I would suggest amending the comment and RCU/checklist.txt.

We need to hear from Lai. I -think- that it is OK for one srcu_read_lock()
to interrupt another, but I could easily be missing something.

> Even call_srcu() should be IRQ-safe.

Looks plausible to me. Lai?

Thanx, Paul