Re: [RFC 00/12] On-demand device registration

From: Alexander Holler
Date: Thu Apr 30 2015 - 03:44:49 EST


Am 29.04.2015 um 11:46 schrieb Alexander Holler:
Am 29.04.2015 um 08:58 schrieb Tomeu Vizoso:
On 28 April 2015 at 20:17, Alexander Holler <holler@xxxxxxxxxxxxx> wrote:
Am 28.04.2015 um 14:49 schrieb Tomeu Vizoso:

On 25 April 2015 at 01:15, Alexander Holler <holler@xxxxxxxxxxxxx>
wrote:

Am 24.04.2015 um 16:47 schrieb Tomeu Vizoso:

Hi,

while reading the thread [0] that Alexander Holler started with his
series to make probing order deterministic, it occurred to me that
it should
be possible to achieve the same by probing devices as they are
referenced by
other devices.

This basically reuses the information that is already embedded in the
probe() implementations, saving us from refactoring existing
drivers or
adding information to DTBs.

The main issue I see is that the registration code path in some
subsystems may not be reentrant, so some refactoring of the
locking will be
needed. In my testing I have found this problem with regulators,
as the
supply of a regulator might end up being registered during the
registration
of the first one.

Something I'm not completely happy with is that I have had to move
the
population of the device tree after all platform drivers have been
registered. Otherwise I don't see how I could register drivers on
demand as
we don't have yet each driver's compatible strings.

I have done my testing on a Tegra124-based Chromebook, and these
patches
were enough to eliminate all the deferred probes.


First you have to solve a problem which is totally unrelated to DT or
ACPI or x86 or ARM:

I think as long as drivers don't register themself whithout any side
effect, this problem isn't solvable. In order to make an ordered
list of
drivers to start, you need to know which drivers are actually
available.


Yeah, I kind of side-stepped that issue by waiting until all drivers
have been registered before registering devices. I think someone
suggested doing so in your thread (maybe Grant?).


That doesn't work. As said above, several drivers doing a lot more
than just
registering in their initcall. They might even crash if some
prerequisits
aren't given. And several of these prerequisits (init orders) are
hardcoded
by various means.

But aren't those dependencies being taken care currently by the
initcall level the driver is placed in? That remains the same in this
approach.

In short, no. There are various very ugly things done in several drivers
to enforce some order.

To explain it a bit more:

The case with the regulator you've described above is just one of these ugly things done on some driver initcalls. That's why I've decided to introduce a new initcall level which only contains drivers which don't have any side effect but just registering themself. Ideally all drivers would end up in that level, and thus the special level would not be necessary at all. Because this means that several drivers have to change, which is a lot of work, an intermediate workaround is to make a whitelist. That's easier than a blacklist which would mean you have to examine every driver. And that whitelist is exactly what I did by introducing those "well done" drivers.

And just in case of, I haven't looked at your patches at all. I've just read the introductory mail and the subjects of the patches and then concluded what I've written.

That means if you've a special use case in mind, your solution might work. But as an overall solution the problem with identifying drivers (mapping initcalls to drivers) has to be solved (which I've tried with those "well done" driver list).

Regards,

Alexander Holler
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/