Re: [Xen-devel] [PATCH] rwlock: allow recursive read locking when already locked in write mode

From: "Roger Pau Monné" <roger.pau@citrix.com>
To: Jan Beulich <jbeulich@suse.com>
Cc: "Jürgen Groß" <jgross@suse.com>,
	"Stefano Stabellini" <sstabellini@kernel.org>,
	"Julien Grall" <julien@xen.org>, "Wei Liu" <wl@xen.org>,
	"Konrad Rzeszutek Wilk" <konrad.wilk@oracle.com>,
	"George Dunlap" <George.Dunlap@eu.citrix.com>,
	"Andrew Cooper" <andrew.cooper3@citrix.com>,
	"Ian Jackson" <ian.jackson@eu.citrix.com>,
	xen-devel@lists.xenproject.org
Subject: Re: [Xen-devel] [PATCH] rwlock: allow recursive read locking when already locked in write mode
Date: Thu, 20 Feb 2020 16:17:34 +0100	[thread overview]
Message-ID: <20200220151734.GM4679@Air-de-Roger> (raw)
In-Reply-To: <0a3a762e-9a0d-7395-d3c4-aca07c366979@suse.com>

On Thu, Feb 20, 2020 at 04:02:55PM +0100, Jan Beulich wrote:
> On 20.02.2020 15:11, Roger Pau Monné wrote:
> > On Thu, Feb 20, 2020 at 01:48:54PM +0100, Jan Beulich wrote:
> >> Another option is to use the recurse_cpu field of the
> >> associated spin lock: The field is used for recursive locks
> >> only, and hence the only conflict would be with
> >> _spin_is_locked(), which we don't (and in the future then
> >> also shouldn't) use on this lock.
> > 
> > I looked into that also, but things get more complicated AFAICT, as it's
> > not possible to atomically fetch the state of the lock and the owner
> > CPU at the same time. Neither you could set the LOCKED bit and the CPU
> > at the same time.
> 
> There's no need to atomically fetch both afaics: The field is
> valid only if the LOCKED bit is set. And when reading the
> field you only care about the value being equal to
> smp_processor_id(), i.e. it is fine to set LOCKED before
> updating the CPU field on lock, and to reset the CPU field to
> SPINLOCK_NO_CPU (or whatever it's called) before clearing
> LOCKED.

Yes that would require the usage of a sentinel value as 0 would be a
valid processor ID.

I would try to refrain from (abu)sing internal spinlock fields, as I
think we can solve this just by using the cnts field on the current
rwlock implementation.

What issue do you have in mind that would prevent storing the CPU
write locked in the cnts field?

Thanks, Roger.

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel