mtd_info->size again (lengthy)

[Jamie Lokier]

Jörn Engel wrote:
> replacing them with submit_fio (or some equivalent function, if
> people dislike the name).

May I suggest submit_mtdio?  One concept to one name, you know.

F means flash, sure.  But we don't call the devices /dev/flash0, /dev/flash1.
Nor do we call it the Flash subsystem.

Or, maybe we should change the name to flash everywhere.  Most people
recognise that nowadays - flash is a common term.

> Submit_fio has the advantage of allowing asynchronous operations,
> queueing, out-of-order completion and all those goodies.  If you
> have several chips in a device, I'm sure you want those as well.

Yes!  I always wondered why it wasn't done like that from the beginning.

> And while at it, replace the 32bit size with a 64bit blockno and 32bit
> offset.  Unless someone is stupid enough to create devices with 4G
> erasesize, that should last for a human lifetime or something close.

Although a single chip is unlikely to have 4GiB erasesize for a while,
a cluster of chips might, and it should be possible to represent them
as a single device.  I strongly suggest supporting erasesizes
comparable with the device size.  That doesn't mean writesize has to
be that big - writesize should fit in size_t (actually ssize_t), just
like sys_write().

> > It still leaves the problem of converting all driver.  I've only started
> on mtdram - because it is simple and everyone can test it.  Other
> drivers can follow over time.  Users like jffs2 could be converted as
> well, but I'd rather leave them for now.

> We want synchronous wrappers around submit_fio anyway, as those are
> useful for development and for trivial performance-agnostic code.
> Comments?

I'm thinking we want asynchronous wrappers around the synchronous
methods too, using a dispatch thread.  So that drivers can be
converted one by one, and old drivers don't need to be.  Something like
generic_submit_fio.

-- Jamie