Re: [PATCH v2 2/3] scsi: ufs: Modulize ufs-bsg

[cang]

On 2019-12-17 01:22, Bart Van Assche wrote:
> On 12/15/19 8:36 PM, cang@xxxxxxxxxxxxxx wrote:
> > On 2019-12-16 05:49, Bart Van Assche wrote:
> > > On 2019-12-11 22:37, Bjorn Andersson wrote:
> > > > It's the asymmetry that I don't like.
> > > >
> > > > Perhaps if you instead make ufshcd platform_device_register_data() 
> > > > the
> > > > bsg device you would solve the probe ordering, the remove will be
> > > > symmetric and module autoloading will work as well (although then 
> > > > you
> > > > need a MODULE_ALIAS of platform:device-name).
> > >
> > > Hi Bjorn,
> > >
> > > From Documentation/driver-api/driver-model/platform.rst:
> > > "Platform devices are devices that typically appear as autonomous
> > > entities in the system. This includes legacy port-based devices and
> > > host bridges to peripheral buses, and most controllers integrated
> > > into system-on-chip platforms. What they usually have in common
> > > is direct addressing from a CPU bus. Rarely, a platform_device will
> > > be connected through a segment of some other kind of bus; but its
> > > registers will still be directly addressable."
> > >
> > > Do you agree that the above description is not a good match for the
> > > ufs-bsg kernel module?
> >
> > I missed this one.
> > How about making it a plain device and add it from ufs driver?
>
> Hi Can,
>
> Since the ufs_bsg kernel module already creates one device node under
> /dev/bsg for each UFS host I don't think that we need to create any
> additional device nodes for ufs-bsg devices. My proposal is to modify
> the original patch 2/3 from this series as follows:
> * Use module_init() instead of late_initcall_sync().
> * Remove the ufshcd_get_hba_list_lock() and
>   ufshcd_put_hba_list_unlock() functions.
> * Implement a notification mechanism in the UFS core that invokes a
>   callback function after an UFS host has been created and also after 
> an
>   UFS host has been removed.
> * Register for these notifications from inside the ufs-bsg driver.
> * During registration for notifications, invoke the UFS host creation
>   callback function for all known UFS hosts.
> * If the UFS core is unloaded, invoke the UFS host removal callback
>   function for all known UFS hosts.
>
> I think there are several examples of similar notification mechanisms
> in the Linux kernel, e.g. the probe and remove callback functions in
> struct pci_driver.
>
> Bart.

Hi Bart,

Even in the current ufs_bsg.c, it creates two devices, one is ufs-bsg,
one is the char dev node under /dev/bsg. Why this becomes a problem
after make it a module?

I took a look into the pci_driver, it is no different than making 
ufs-bsg
a plain device. The only special place about pci_driver is that it has 
its
own probe() and remove(), and the probe() in its bus_type calls the
probe() in pci_driver. Meaning the bus->probe() is an intermediate call
used to pass whatever needed by pci_driver->probe().

Of course we can also do this, but isn't it too much for ufs-bsg?
For our case, calling set_dev_drvdata(bsg_dev, hba) to pass hba to
ufs_bsg.c would be enough.

If you take a look at the V3 patch, the change makes the ufs_bsg.c
much conciser. platform_device_register_data() does everything for us,
initialize the device, set device name, provide the match func,
bus type and release func.

Since ufs-bsg is somewhat not a platform device, we can still add it
as a plain device, just need a few more lines to get it initialized.
This allows us leverage kernel's device driver model. Just like Greg
commented, we don't need to re-implement the mechanism again.

Thanks,
Can Guo.