Re: [PATCH v2 1/2] sysfs: handle 'parent deleted before child added'

["Williams, Dan J" <dan.j.williams@xxxxxxxxx>]

On Fri, Apr 6, 2012 at 2:17 PM, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Apr 06, 2012 at 02:06:50PM -0700, Williams, Dan J wrote:
>> On Fri, Apr 6, 2012 at 1:45 PM, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>> > On Fri, Apr 06, 2012 at 01:41:06PM -0700, Dan Williams wrote:
>> >> In scsi at least two cases of the parent device being deleted before the
>> >> child is added have been observed.
>> >>
>> >> 1/ scsi is performing async scans and the device is removed prior to the
>> >>    async can thread running (can happen with an in-opportune / unlikely
>> >>    unplug during initial scan).
>> >
>> > That sounds like a bug in the scsi code, doesn't it?
>> >
>> >> 2/ libsas discovery event running after the parent port has been torn
>> >>    down (this is a bug in libsas).
>> >
>> > Is this fixed somewhere?
>>
>> Yes, these two issues have pending fixes that are posted to linux-scsi:
>>
>> http://marc.info/?l=linux-scsi&m=133239707903443&w=2
>> http://marc.info/?l=linux-scsi&m=133239709603452&w=2
>>
>> > I don't want to paper over bugs like this by changing the sysfs core.
>> > We went through this a lot years ago when scsi changed to use the driver
>> > core, and I thought we had fixed all of these types of errors properly.
>>
>> Hotplug lifetime rules are still transport specific.  So in this case
>> scsi-core is innocent these are bugs from libsas and
>> scsi_transport_sas.
>
> Ok, thanks for the explaination.
>
>> > So, any chance to fix these properly as well?
>>
>> This patch doesn't really paper over anything.  It turns a NULL
>> pointer crash into an explicit warning from kobject_add_internal.  For
>> the libsas/scsi case this device_add() failure is still fatal.
>> Regardless of whether sysfs changes the above two fixes are still
>> required.
>>
>> Since the -EEXIST case is just a KERN_ERR and not a BUG_ON I figured
>> it was worthwhile to post a patch to do the same for this 'parent
>> deleted' case.  But if crashing is the expectation then this patch can
>> be dropped.
>
> No, crashing is not the expectation :)

I thought not, but sometimes the kernel likes to teach people that
bollix an api a hard lesson :).

>
> But, without that crash, would the above fixes ever have been noticed
> and fixed?  The device_add() most likely would have quietly failed and
> no one would have been the wiser.
>
> Or would something else have caused this to be an obvious problem?
>

We still have the big red flag dump_stack() in kobject_add_internal()
(which patch 2 turns into a real WARN()), and for scsi our hotplug
tests still crash later on because libsas makes assumptions about the
device path.  I understand the paranoia here,  "check for NULL" is
usually a band-aid, but in this case this is just a softer
introduction to a debug session.  No less vocal than before as far as
I can see.

--
Dan
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/