Re: [patch 1/3] fastboot: Create a "asynchronous" initlevel

[Rene Herman <rene.herman@keyaccess.nl>]

On 19-07-08 17:44, Arjan van de Ven wrote:

> On Sat, 19 Jul 2008 10:10:09 +0200
> Rene Herman <rene.herman@keyaccess.nl> wrote:

>>> I'm not sure about this comment, being not very sure about the
>>> semantics of late_initcall but shouldn't late_initcall (level 7)
>>> wait for 6s to have completed?
>> 
>> Following up on this myself -- see for example kernel/power/disk.c:
>> power_suspend(). It's a late intitcall so that, as it comments,
>> "all devices are discovered and initialized". However, your first
>> followup patch makes the USB HCI init async meaning that any USB
>> storage device might not be ready yet when it runs, no?
> 
> good spotting/comment.
> 
> you would have a valid point... if it weren't for the case where much
> of this actual "end device" probing is in various cases already
> asynchronous... what you do have found is a bug in the suspend code.
> Unless code does:
>         /* wait for the known devices to complete their probing */
>         while (driver_probe_done() != 0)
>                 msleep(100);
> (taken from init/do_mounts.c)
> 
> ... the assertion in the comment that probing is done is absolutely
> false, with or without my patches.

Yes, I see. Unfortunately, WITH your patches, driver_probe_done() would 
also no longer be safe when run from a late_initcall() it would appear.

driver_probe_done() tests a variable that's incremented just before the 
driver model calls into the driver .probe method and decremented on 
return from it (really_probe).

However, if the entire module_init() is async the probing may not even 
have _started_ yet let alone finished. Let's take ehci_hcd_init() as an 
example both since you changed that one and since it'll fairly often be 
en route to mass storage devices.

Only after ehci_init() calls foo_register_driver() is the driver model 
aware of it and will it start calling the probe methods meaning the 
driver_probe_done() would be racing.

I have the sneaking suspicion that this is a bit of a fundamental issue. 
Turning some of the driver level (6) async basicaly removes the ordering 
between drivers and late_initcall (level 7).

I trust it will completely and utterly destroy the point of this patch 
to flush level 6a before starting level 7?

> (Not that I want the suspend/resume code to call this, because that
> would make the boot even longer ;( )

Well, yes, but bugs are bugs. CCing Pavel and Rafael as well :-)

Rene.