[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200326214005.GB9894@blackberry>
Date: Fri, 27 Mar 2020 08:40:05 +1100
From: Paul Mackerras <paulus@...abs.org>
To: Leonardo Bras <leonardo@...ux.ibm.com>
Cc: Benjamin Herrenschmidt <benh@...nel.crashing.org>,
Michael Ellerman <mpe@...erman.id.au>,
Allison Randal <allison@...utok.net>,
Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
Thomas Gleixner <tglx@...utronix.de>,
linuxppc-dev@...ts.ozlabs.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH 1/1] ppc/smp: Replace unnecessary 'while' by 'if'
On Thu, Mar 26, 2020 at 05:37:52PM -0300, Leonardo Bras wrote:
> spin_until_cond() will wait until nmi_ipi_busy == false, and
> nmi_ipi_lock_start() does not seem to change nmi_ipi_busy, so there is
> no way this while will ever repeat.
>
> Replace this 'while' by an 'if', so it does not look like it can repeat.
Nack, it can repeat. The scenario is that cpu A is in this code,
inside spin_until_cond(); cpu B has previously set nmi_ipi_busy, and
cpu C is also waiting for nmi_ipi_busy to be cleared, like cpu A.
When cpu B clears nmi_ipi_busy, both cpu A and cpu C will see that and
will race inside nmi_ipi_lock_start(). One of them, say cpu C, will
take the lock and proceed to set nmi_ipi_busy and then call
nmi_ipi_unlock(). Then the other cpu (cpu A) will then take the lock
and return from nmi_ipi_lock_start() and find nmi_ipi_busy == true.
At that point it needs to go through the while loop body once more.
Paul.
Powered by blists - more mailing lists