Re: [PATCH] drivers/base: error handling fixes

[Heiko Carstens <heiko.carstens@de.ibm.com>]

> >>If someone wants to provide a better fix, let's see the patch...
> >If sysfs_remove_group() would also work for non-created (-existent) groups
> >then the patch below would work. Unfortunately that is not the case. So one
> >would have to remember if sysfs_create_group() was done and succeeded before
> >calling sysfs_remove_group()...
> >There must be an easier way.
> >diff --git a/drivers/base/topology.c b/drivers/base/topology.c
> >index 3ef9d51..d0056c3 100644
> >--- a/drivers/base/topology.c
> >+++ b/drivers/base/topology.c
> 
> ACK

Ah, you probably got me wrong. The patch doesn't work, because of the
sysfs_remove_group() stuff. That's why there is no Signed-off-by:...

What _could_ happen with this patch applied: some code fails on
CPU_UP_PREPARE and then all notifiers get called with CPU_UP_CANCELLED.

That would cause the call of topology_remove_dev(sys_dev) and therefore
sysfs_remove_group(&sys_dev->kobj, &topology_attr_group) which will crash,
because sysfs_create_group() wasn't even called.
To fix this one has to remember if sysfs_create_group() succeeded or was
even called. That would be a per-cpu array. Now, I think it's just
overkill to introduce a per-cpu array for error-checking.

IMHO it would make sense to change sysfs_remove_group() so it will survive
if being asked to remove groups that don't exist.