lkml.org 
[lkml]   [2017]   [Jul]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v6] workqueue: Fix edge cases for calc of pool's cpumask
Hello, Michael.

On Thu, Jul 27, 2017 at 02:07:53PM -0500, Michael Bringmann wrote:
> The problem lies with the ordering of events with respect to the order in
> which we add (or remove) CPUs to NUMA systems, and make use of that knowledge.

Isn't the root cause that the upper layers including workqueue expect
cpu <-> node mapping to be static but powerpc doesn't follow that? I
don't get why ordering matters here.

> The CPUs present are assigned to nodes, and workqueues and their infrastructure
> are created to use the CPUs in a node. Workqueues are created at boot time
> and updated or created as CPUs are added or removed. However, there is little
> or no synchronization or ordering of these events, and the data structures

What I meant was that there's no synchronization construct protection
cpu <-> node mapping. If arch code changes it during hot plug, it's
changing it underneath anybody who might be using that association.

> mapping CPUs to nodes may not be updated before the workqueue infrastructure
> is built for a node. Thus we have the possibility of an invalid CPU mask
> attribute being attached to a newly created workqueue before the CPUs have
> been properly registered and published to a node.
>
> This patch attempts to provide a partial ordering of events within workqueue
> by delaying the use of newly calculated CPU masks as the value for a workqueue
> attribute until they have valid content. Instead the workqueue code must delay
> creating new workqueues until this function succeeds, or it can use a previously
> calculated cpumask attribute that is known to be valid.
>
> This patch attempts to ensure that a valid, usable cpumask is used to set up
> newly created pools for workqueues. This patch provides a fix for NUMA systems
> which can add/subtract processors dynamically. The patch is expected to be an
> intermediate one while developers find and correct any underlying issues.

And what this patch does is adding a bandaid so that we at least don't
crash immediately when this condition triggers until the arch code can
be fixed properly.

Thanks.

--
tejun

\
 
 \ /
  Last update: 2017-07-27 21:24    [W:0.076 / U:16.524 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site