Re: [PATCH next v2] OF: convert devtree lock from rw_lock to raw spinlock

[Stephen Warren <swarren-3lzwWm7+Weoh9ZMKESR00Q@public.gmane.org>]

On 02/08/2013 04:09 PM, Rob Herring wrote:
> On 02/06/2013 02:30 PM, Paul Gortmaker wrote:
>> From: Thomas Gleixner <tglx-hfZtesqFncYOwBW4kG4KsQ@public.gmane.org>
>>
>> With the locking cleanup in place (from "OF: Fixup resursive
>> locking code paths"), we can now do the conversion from the
>> rw_lock to a raw spinlock as required for preempt-rt.
>>
>> The previous cleanup and this conversion were originally
>> separate since they predated when mainline got raw spinlock (in
>> commit c2f21ce2e31286a "locking: Implement new raw_spinlock").
>>
>> So, at that point in time, the cleanup was considered plausible
>> for mainline, but not this conversion.  In any case, we've kept
>> them separate as it makes for easier review and better bisection.
>>
>> Signed-off-by: Thomas Gleixner <tglx-hfZtesqFncYOwBW4kG4KsQ@public.gmane.org>
>> [PG: taken from preempt-rt, update subject & add a commit log]
>> Signed-off-by: Paul Gortmaker <paul.gortmaker-CWA4WttNNZF54TAoqtyWWQ@public.gmane.org>
>> ---
>>
>> [v2: recent commit e81b329 ("powerpc+of: Add /proc device tree
>>  updating to of node add/remove") added two more instances of
>>  write_unlock that also needed converting to raw_spin_unlock.
>>  Retested (boot) on sbc8548, defconfig builds on arm/sparc; no
>>  new warnings observed.]
>>
>>  arch/sparc/kernel/prom_common.c |   4 +-
>>  drivers/of/base.c               | 100 ++++++++++++++++++++++------------------
>>  include/linux/of.h              |   2 +-
>>  3 files changed, 59 insertions(+), 47 deletions(-)
> 
> Applied.

This commit is present in next-20130211, and causes a boot failure
(hang) early while booting on Tegra. Reverting just this one commit
solves the issue.

I'll see if I can track down where the issue is. Given the commit
description, I assume there's some new recursive lock issue that snuck
in between the previous fix for them and this commit? Any hints welcome.

One thing I wonder looking at the patch: Most paths use
raw_spin_lock_irqsave() but a few use just raw_spin_lock(). I wonder how
that decision was made?