On Tuesday, June 23, 2015 04:37:57 PM Tomeu Vizoso wrote:

On 23 June 2015 at 16:51, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Tuesday, June 23, 2015 04:17:29 PM Tomeu Vizoso wrote:

quoted

On 23 June 2015 at 16:37, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Monday, June 22, 2015 07:07:08 PM Rob Herring wrote:

quoted

On Fri, Jun 19, 2015 at 6:20 PM, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Friday, June 19, 2015 03:36:46 PM Tomeu Vizoso wrote:

quoted

On 18 June 2015 at 23:50, Rafael J. Wysocki [off-list ref] wrote:

quoted

On Wednesday, June 17, 2015 03:42:12 PM Tomeu Vizoso wrote:

quoted

To decrease the chances of devices deferring their probes because of
dependencies not having probed yet because of their drivers not having
registered yet, delay all probing until the late initcall level.

This will allow us to avoid deferred probes completely later by probing
dependencies on demand, or by probing them in dependency order.

Signed-off-by: Tomeu Vizoso <redacted>
---
 drivers/base/dd.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/base/dd.c b/drivers/base/dd.c
index a638bbb..18438aa 100644
--- a/drivers/base/dd.c
+++ b/drivers/base/dd.c

@@ -407,6 +407,12 @@ int driver_probe_device(struct device_driver *drv, struct device *dev)
      if (!device_is_registered(dev))
              return -ENODEV;

+     /* Defer all probes until we start processing the queue */
+     if (!driver_deferred_probe_enable) {
+             driver_deferred_probe_add(dev);

Do I think correctly that this will effectively force everybody to use deferred
probing?

Guess it depends on the meaning of "using deferred probing". It will
defer the probe of the first device to late_initcall (which will
happen much earlier in time than before), but afterwards all built-in
drivers will be available and depending on the order in which we try
to probe devices, none may actually ask to defer its probe.

So this will break things like the PNP system driver which relies on probing
stuff at the fs_initcall stage for correctness.  It may also break other
things with similar assumptions.

Yes, but I think that this can be done for only OF based devices
rather than globally for all platform devices and solve that problem.
Matching is already dependent of the type of device.

Well, the current patch is not OF-only, though.

Yeah, I'm currently looking at only delaying probing of devices
created from OF data.

I'm not sure if tying it hard to OF is not too restrictive.

Maybe we can use some general opt-in mechanism that OF will just always use?

Would it help if buses called fwnode_driver_match_device() instead of
the existing OF and ACPI variants and we did it in there?

Probably it would, but I'd need to see the actual patch. :-)

I'm still not sure of how fwnode is used in machines with ACPI.

I'm not sure what you mean.  On ACPI systems struct fwnode_handle is embedded
in struct acpi_device and there is a pointer from struct device to that field
in the companion ACPI device object.

But that would be quite a bit of work that I think should be left for
a later series because otherwise this one is going to balloon in size
really quickly.

Well, I'd prefer not to leave anything to a "later series" that may never be
submitted ...

quoted

In fact, we have a similar problem in ACPI where we have the _DEP object which
is used by firmware to describe dependencies between devices.

I would expect that classes/subsystems would be able to use that data
in their class.get_dependencies() callback, if the passed fwnode is a
ACPI node.

Yes, something like that.

But the point is that this really isn't OF-specific.

Thanks,
Rafael