Re: [PATCH v3 1/9] staging: fsl-mc: move bus driver out of staging

[Greg KH]

On Fri, Dec 09, 2016 at 12:36:26AM +0000, Stuart Yoder wrote:
> > -----Original Message-----
> > From: Greg KH [mailto:gregkh@xxxxxxxxxxxxxxxxxxx]
> > Sent: Thursday, December 08, 2016 10:05 AM
> > To: Stuart Yoder <stuart.yoder@xxxxxxx>
> > Cc: devel@xxxxxxxxxxxxxxxxxxxx; agraf@xxxxxxx; arnd@xxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; Leo Li
> > <leoyang.li@xxxxxxx>; Catalin Horghidan <catalin.horghidan@xxxxxxx>; Ioana Ciornei
> > <ioana.ciornei@xxxxxxx>; Laurentiu Tudor <laurentiu.tudor@xxxxxxx>
> > Subject: Re: [PATCH v3 1/9] staging: fsl-mc: move bus driver out of staging
> > 
> > On Wed, Dec 07, 2016 at 08:19:20PM +0000, Stuart Yoder wrote:
> > > > -----Original Message-----
> > > > From: Greg KH [mailto:gregkh@xxxxxxxxxxxxxxxxxxx]
> > > > Sent: Wednesday, December 07, 2016 9:53 AM
> > > > To: Stuart Yoder <stuart.yoder@xxxxxxx>
> > > > Cc: devel@xxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; agraf@xxxxxxx; arnd@xxxxxxxx; Leo Li
> > > > <leoyang.li@xxxxxxx>; Ioana Ciornei <ioana.ciornei@xxxxxxx>; Catalin Horghidan
> > > > <catalin.horghidan@xxxxxxx>; Laurentiu Tudor <laurentiu.tudor@xxxxxxx>; Ruxandra Ioana Radulescu
> > > > <ruxandra.radulescu@xxxxxxx>
> > > > Subject: Re: [PATCH v3 1/9] staging: fsl-mc: move bus driver out of staging
> > > >
> > > > On Thu, Dec 01, 2016 at 04:41:26PM -0600, Stuart Yoder wrote:
> > > > > Move the source files out of staging into their final locations:
> > > > >   -include files in drivers/staging/fsl-mc/include go to include/linux/fsl
> > > > >   -irq-gic-v3-its-fsl-mc-msi.c goes to drivers/irqchip
> > > > >   -source in drivers/staging/fsl-mc/bus goes to drivers/bus/fsl-mc
> > > > >   -README.txt, providing and overview of DPAA goes to
> > > > >    Documentation/dpaa2/overview.txt
> > > > >   -update MAINTAINERS with new location
> > > > >
> > > > > Delete other remaining staging files-- Makefile, Kconfig, TODO
> > > >
> > > > Ok, given that I haven't ever reviewed this code, I had a few questions
> > > > that I couldn't easily figure out by looking at your code:
> > > > 	- what is the lifecycle of your 'struct device' usage?  Who
> > > > 	  creates it, who frees it, and who accesses it?
> > >
> > > We embed a 'struct device' inside our bus specific device struct
> > > 'struct fsl_mc_device'.  So, when a new fsl-mc object is discovered
> > > on the bus during initial enumeration or hotplug we create a new
> > > 'struct fsl_mc_device' and do a device_initialize()/device_add().
> > > (see fsl_mc_device_add() for where this is done)
> > >
> > > 'struct device' is freed when a device is removed-- the reverse
> > > of the above.
> > 
> > Where is the device freed?  I see you trying to do some "odd" stuff in
> > fsl_mc_device_remove() by deleting and then putting a device structure.
> > I can't find a "release()" callback anywhere for your bus, where is it?
> > 
> > What happens when the reference count falls to 0 for your struct device?
> 
> Hrm...something seems wrong in free path, and I think this needs to
> be refactored.
> 
> IIRC, when German (former maintainer) wrote that code he loosely based
> it on the register/unregister platform bus code:
> 
> int platform_device_register(struct platform_device *pdev)
> {
>         device_initialize(&pdev->dev);
>         arch_setup_pdev_archdata(pdev);
>         return platform_device_add(pdev);
> }
> void platform_device_unregister(struct platform_device *pdev)
> {
>         platform_device_del(pdev);
>         platform_device_put(pdev);
> }
> 
> ...I'm puzzling over how that code handles a refcount of zero
> as I see no 'release' callback anywhere, but I must be missing
> something.
> 
> In any case, we'll get this refactored.

Have you tried removing a device?  The kernel should complain loudly
about there not being a release function for your device.

> > > > 	- root_dprc_count, why are you using an atomic variable for
> > > > 	  this?  What is it for other than "look, I'm running!"?
> > >
> > > There can be multiple root buses, and this variable simply tracks the count
> > > of them.
> > 
> > Why does it matter?
> > 
> > > It's is atomic there might be a theoretical race condition where 2
> > > buses might be added at the same time.  The root buses are found in
> > > the device tree and so if there is no chance that device tree
> > > processing happens in parallel on multiple cores then we could remove
> > > the atomic.
> > 
> > Why not just use a lock, or better yet, not care about a "count" at all?
> > I don't see you doing anything with the count, other than emitting a
> > WARN() if it drops down below 0 for some reason, or when you call
> > fsl_mc_bus_exists() which for some reason is exported yet no one uses
> > it...
> 
> We can drop this count.  At one time I think there was envisioned an 
> external user who needed it, but it's no longer the case.

Please do, we are trying to get rid of atomic_t abuse on other mailing
lists, and this one fits the pattern of "no real need for it" :)

> Given the additional refactoring, I think the fsl-mc bus driver needs
> to stay in staging for a bit.  In order to facilitate further review
> I'm going to refactor the patch series:
>   staging: fsl-mc: move bus driver out of staging, add dpio
> 
> ...to just add dpio (into staging).  This will allow the Eth driver
> series sent earlier this week to go into staging:
>   staging: Introduce Freescale DPAA2 Ethernet driver
> 
> With all that in staging we'll have a fully functional Ethernet
> driver.

Ok, that sounds reasonable.

thanks,

greg k-h