Re: [PATCH] [SCSI] bnx2fc: fix build error when !CONFIG_MODULES

[Robert Love]

On Mon, 2011-03-07 at 16:18 -0800, James Bottomley wrote:
> On Mon, 2011-03-07 at 15:54 -0800, Bhanu Gollapudi wrote:
> > On Mon, 2011-03-07 at 12:16 -0800, Mariusz Kozlowski wrote: 
> > > On Wed, Mar 02, 2011 at 11:10:03PM +0100, Mariusz Kozlowski wrote:
> > > > drivers/scsi/bnx2fc/bnx2fc_fcoe.c:1815: error: dereferencing pointer to incomplete type
> > > > drivers/scsi/bnx2fc/bnx2fc_fcoe.c:1815: error: âMODULE_STATE_LIVEâ undeclared (first use in this function)
> > > 
> > > Hm. Still there in next-20110307. Is this patch wrong or..?
> > > 
> > 
> > James,
> > 
> > Here is my ack for this patch.
> 
> OK, so the patch is actually wrong because adding #ifdefs on modules in
> files really impedes readability.  The bug is using a direct deref on
> module state instead of one of the APIs which work in the non-modular
> case, namely try_module_get(). That means the other two need to come out
> and be reworked (plus all the others in fcoe).
> 
> Reworked looks like it might be a bigger item than bnx2fc.  If any of
> those tests is ever relevant, it means we have a race in the
> fcoe_transport because it shouldn't be calling function pointers on a
> dying module (unless it wants to trigger an oops).
> 
> So, why are you trying to do this in the first place?
> 
First, fcoe.c started with these checks. Here is a comment in fcoe.c at
the point of one of the checks.

/*
 * Make sure the module has been initialized, and is not about to be
 * removed.  Module paramter sysfs files are writable before the
 * module_init function is called and after module_exit.
 */

I don't know the correct way to fix that race is, but we may be past the
need to fix it in the LLDs.

Next, the fcoe transport was added. Since it (libfcoe.ko) is now calling
what used to be the fcoe.ko sysfs entry points I don't think the problem
exists in fcoe.c or in bnx2fc_fcoe.c, the problem should be in the fcoe
transport code, as James suggested.

The fcoe transport code already has these checks to protect against
sysfs files being writable before module initialization is complete. It
uses the ft_mutex to protect the list of transports(LLDs) so when
'create' is called it knows that the transport is still there to call
down to. It holds the ft_mutex until the LLD's 'create' routine returns.
The transports(LLDs) should be detaching themselves from the fcoe
transport layer before they exit. fcoe_transport_detach will try to
acquire the ft_mutex and block until the 'create' call returns and
releases the ft_mutex. I think this ensures that the transport(LLD) will
be fine when the fcoe transport calls it.

My feeling is that these checks are still needed in the fcoe transport,
but not in the LLDs. If someone can suggest a better way to protect
against writable sysfs files when the module hasn't finished
initializing, we should do that instead of the ifdefs.

Hope this helps,

//Rob

FYI: mnc asked about this code and the trylock code in fcoe and libfcoe.
I have patches in our internal validation to remove the trylock usage,
but I don't have patches to fix the module state checking.

--
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/