[PATCH v12 3/3] mtd: Add driver for concatenating devices

[Mahapatra, Amit Kumar]

[AMD Official Use Only - AMD Internal Distribution Only]

> On 13/05/2025 at 14:45:39 GMT, "Mahapatra, Amit Kumar" <amit.kumar-
> mahapatra at amd.com> wrote:
>
> > [AMD Official Use Only - AMD Internal Distribution Only]
> >
> > Hello Miquel,
> >
> >> >> > +           mtd->dev.parent = concat->subdev[0]->dev.parent;
> >> >> > +           mtd->dev = concat->subdev[0]->dev;
> >> >> > +
> >> >> > +           /* Register the platform device */
> >> >> > +           ret = mtd_device_register(mtd, NULL, 0);
> >> >> > +           if (ret)
> >> >> > +                   goto destroy_concat;
> >> >> > +   }
> >> >> > +
> >> >> > +   return 0;
> >> >> > +
> >> >> > +destroy_concat:
> >> >> > +   mtd_concat_destroy(mtd);
> >> >> > +
> >> >> > +   return ret;
> >> >> > +}
> >> >> > +
> >> >> > +late_initcall(mtd_virt_concat_create_join);
> >> >>
> >> >> The current implementation does not support probe deferrals, I
> >> >> believe it should be handled.
> >> >
> >> > I see that the parse_mtd_partitions() API can return -EPROBE_DEFER
> >> > during MTD device registration, but this behavior is specific to
> >> > the parse_qcomsmem_part parser. None of the other parsers appear to
> >> > support probe deferral. As discussed in RFC [1], the virtual concat
> >> > feature is purely a fixed-partition capability, and based on my
> >> > understanding, the fixed-partition parser does not support probe deferral.
> >> > Please let me know if you can think of any other probe deferral
> >> > scenarios that might impact the virtual concat driver.
> >>
> >> That's true, but I kind of dislike the late_initcall, I fear it might break in creative
> ways.
> >
> > I understand, but since we require the partition information to be
> > available, late_initcall seems to be the most suitable choice among
> > the initcall levels—if we decide to proceed with using an initcall.
> > Regarding potential failures, as far as I can tell, the operation
> > would fail if, at the time of concatenation, one or more of the MTD
> > devices involved in the concat are not yet available. In such a
> > scenario, we can issue a kernel warning and exit gracefully. But,
> > However, if you prefer to move away from using initcalls and have an
> > alternative implementation approach in mind, please let us know.
>
> I am sorry but this does not work with modules, and we cannot ignore this case I
> believe. More specifically, if a controller probe is deferred (with EPROBE_DEFER
> or just prevented because some dependencies are not yet satisfied), you'll get
> incorrectly defined mtd devices.

Ok, an alternative solution could be to remove the initcall registration
and instead invoke mtd_virt_concat_create_join()—which was previously
registered as a late_initcall—directly from mtd_device_parse_register().
I believe this approach would address both of your concerns regarding
module support and probe deferral. Additionally, we could consider
moving the entire code from mtd_virt_concat.c into mtdconcat.c.
Please let us know your take on this.

Regards,
Amit
>
> Thanks,
> Miquèl