On Tuesday 28 January 2014, Ulf Hansson wrote:

On 28 January 2014 01:59, Tomasz Figa [off-list ref] wrote:

quoted

On 27.01.2014 11:19, Ulf Hansson wrote:

quoted

There is already a host capability that I think we could use to handle
this. MMC_CAP_NONREMOVABLE, the corresponding DT binding string is
"non-removable", and it may be set per host device.

Using this cap means the mmc_rescan process that runs to detect new
cards, will only be executed once and during boot. So, we need to make
sure all resources and powers are provided to the card at this point.
Otherwise the card will not be detected.

I don't quite like this requirement, especially if you consider
multi-platform kernels where a lot of drivers is going to be provided as
modules. WLAN drivers are especially good candidates. This means that even
if the card is powered off at boot-up, if user (or init system) loads
appropriate module, which powers the chip on, MMC core must be able to
notice this.

To be able to detect the card, the WLAN driver doesn't have to be
probed, only the "power controller" driver. I suppose this is were it
becomes a bit tricky.

Somehow the mmc core needs to be involved in the probe process of the
power controller driver. Could perhaps the power controller bus be
located in the mmc core and thus the power controller driver needs to
register itself by using a new API from the mmc core? Similar how SDIO
func driver's register themselves.

I think there is another option, which does have its own pros and cons:
We could move all the power handling back into the sdio function driver
if we allow a secondary detection path using DT rather than the probing
of the SDIO bus. Essentially you'd have to list the class/vendor/device
ID for each function that cannot be autodetected in DT, and have the
SDIO core pretend that it found the device just by looking at the
device nodes, and register the struct sdio_func so it can be bound to
the driver. The driver then does all the power handling in a device
specific way. At some point the hardware gets registered at the 
mmc host, and the sdio core connects the bus state to the already present
sdio_func, possibly notifying the function driver that it has become
usable.

Obviously, this can only work for CAP_NONREMOVABLE devices, but those
are exactly the ones we are worried about here. The advantage is that
the power sequencing for a particular device can then be in device
specific code and can have an arbitrarily complex in the driver without
needing the mmc code to handle all possible corner cases.

	Arnd