Re: [PATCH v4] lib: cpu_rmap: avoid flushing all workqueues

[Ben Hutchings <bhutchings@solarflare.com>]

On Wed, 2013-01-02 at 15:12 -0800, Andrew Morton wrote:
> On Wed,  2 Jan 2013 13:52:25 -0800
> David Decotigny <decot@googlers.com> wrote:
> 
> > In some cases, free_irq_cpu_rmap() is called while holding a lock
> > (eg. rtnl). This can lead to deadlocks, because it invokes
> > flush_scheduled_work() which ends up waiting for whole system
> > workqueue to flush, but some pending works might try to acquire the
> > lock we are already holding.
> > 
> > This commit uses reference-counting to replace
> > irq_run_affinity_notifiers(). It also removes
> > irq_run_affinity_notifiers() altogether.
> 
> I can't say that I've ever noticed cpu_rmap.c before :( Is is too late
> to review it?
> 
> - The naming is chaotic.  At least these:
> 
> 	EXPORT_SYMBOL(alloc_cpu_rmap);
> 	EXPORT_SYMBOL(free_cpu_rmap);
> 	EXPORT_SYMBOL(cpu_rmap_add);
> 	EXPORT_SYMBOL(cpu_rmap_update);
> 	EXPORT_SYMBOL(free_irq_cpu_rmap);
> 	EXPORT_SYMBOL(irq_cpu_rmap_add);
> 
>   should be consistently named cpu_rmap_foo()

There is a common practice of defining alloc_foo() and free_foo()
alongside foo_do_this() and foo_do_that().  I deliberately chose to
follow that.  If this is deprecated then it should be documented
somewhere.

There is also a separation between functions that are specific to IRQ
affinity (last 2) and those that are not (first 4).

> - What's the locking model?  It appears to be caller-provided, but
>   it is undocumented.

I think caller-provided can be assumed as the default for library code.
And IRQ setup and teardown need to be properly serialised in the driver
already.

>   drivers/net/ethernet/mellanox/mlx4/ appears to be using
>   msix_ctl.pool_lock for exclusion, but I didn't check for coverage.
>
>   drivers/net/ethernet/sfc/efx.c seems to not need locking because
>   all its cpu_rmap operations are at module_init() time.
>
>   The cpu_rmap code would be less of a hand grenade if each of its
>   interface functions documented the caller's locking requirements.

This particular 'hand grenade' *was* documented.  So I don't think
documentation is the problem.

> As for this patch: there's no cc:stable here but it does appear that
> the problem is sufficiently serious to justify a backport, agree?
[...]

Not sure.  So far as I can see, nothing called free_irq_cpu_rmap() while
holding the RTNL lock before v3.8-rc1.  If there can be work items on a
global workqueue that lock a PCI device (perhaps EEH?) then stable
versions may also be affected.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.