Re: [PATCH 7/10] ACPI / hotplug: Move container-specific code out of the core

[Rafael J. Wysocki]

On Saturday, December 14, 2013 06:07:06 AM Rafael J. Wysocki wrote:
> On Friday, December 13, 2013 02:17:32 PM Yasuaki Ishimatsu wrote:
> > (2013/12/13 13:56), Rafael J. Wysocki wrote:
> > > On Friday, December 13, 2013 11:56:32 AM Yasuaki Ishimatsu wrote:
> > >> Hi Rafael,
> > >
> > > Hi,
> > >
> > >> Please share your more detailed idea. I started to implement the following
> > >> idea. But the idea has one problem.
> > >>
> > >>>>> The eject work flow can be:
> > >>>>>      (1) an eject event occurs,
> > >>>>>      (2) the container "physical" device fails offline in acpi_scan_hot_remove()
> > >>>>>          emmitting, say, KOBJ_CHANGE for the "physical" device,
> > >>>>>      (3) user space notices the KOBJ_CHANGE and does the cleanup as needed,
> > >>>>>      (4) user space changes the "physical" container device flag controlling
> > >>>>>          offline to 0,
> > >>>>>      (5) user space uses the sysfs "eject" attribute of the ACPI container object
> > >>>>>          to finally eject the container,
> > >>>>>      (6) the offline in acpi_scan_hot_remove() is now successful, because the
> > >>>>>          flag controlling it has been set to 0 in step (4),
> > >>>>>      (7) the "physical" container device goes away before executing _EJ0,
> > >>>>>      (8) the container is ejected.
> > >>
> > >> I want to emit KOBJ_CHANGE before offlining devices on container device at (2).
> > >> But acpi_scan_hot_remove() offlines devices on container device at first.
> > >> So when offline container device, devices on container has been offlined.
> > >>
> > >> Thus the idea cannot fulfill my necessary feature.
> > >
> > > Well, in that case we need to treat containers in a special way at the ACPI
> > > level.  Which is a bit unfortunate so to speak.
> > >
> > > To that end I'd try to add a new flag to struct acpi_hotplug_profile, say
> > > .verify_offline, such that if set, it would cause acpi_scan_hot_remove() to
> > > check if all of the "physical" companions of the top-level device are offline
> > > to start with, and if not, it would just emit KOBJ_CHANGE for the companions
> > > that are not offline and bail out.
> > >
> > > So the above algorithm would become:
> > >
> > > (1) an eject event occurs,
> > > (2) acpi_scan_hot_remove() checks the verify_offline flag in the target device's
> > >      scan_handler structure,
> > > (3) if set (it would always be set for containers), acpi_scan_hot_remove()
> > >      checks the status of the target device's "physical" companions; if at least
> > >      one of them is offline, KOBJ_CHANGE is emitted for that "physical" device,
> > >      and acpi_scan_hot_remove() returns, [I guess we can just emit KOBJ_CHANGE
> > >      for the first companion that is not offline at this point.]
> > > (4) user space notices the KOBJ_CHANGE and does the cleanup as needed; in the
> > >      process it carries out the offline operation for the container's "physical"
> > >      companion (there's only one such companion for each container), [That
> > >      operation for the container itself is trivial, but to succeed it requires
> > >      all devices below the container to be taken offline in advance.]
> > > (5) user space uses the sysfs "eject" attribute of the ACPI container object
> > >      to finally eject the container,
> > > (6) acpi_scan_hot_remove() is now successful, because the container's "physical"
> > >      companion is now offline,
> > > (7) the "physical" container device goes away before executing _EJ0,
> > > (8) the container is ejected.
> > >
> > > I think that should work for you.
> > 
> > This idea seems to same as your previous work.
> > http://lkml.org/lkml/2013/2/23/97
> 
> No, it is not.  That one didn't involve physical device representations.
> 
> > How about add autoremove flag into acpi_hotplug_profile and check it as follow:
> 
> This is very similar to "enable" except that it generates the uevent and
> "enable" doesn't.  You might as well modify "enable" to trigger a uevent if
> eject is not enabled (note that with the latest patches in linux-next "enable"
> only applies to eject).
> 
> That said I don't think we should generate any uevents for struct acpi_device
> objects, because they are not devices.
> 
> > ---
> >   drivers/acpi/scan.c | 5 +++++
> >   1 file changed, 5 insertions(+)
> > 
> > diff --git a/drivers/acpi/scan.c b/drivers/acpi/scan.c
> > index 5383c81..c43d110 100644
> > --- a/drivers/acpi/scan.c
> > +++ b/drivers/acpi/scan.c
> > @@ -409,6 +409,11 @@ static void acpi_hotplug_notify_cb(acpi_handle handle, u32 type, void *data)
> >   			ost_code = ACPI_OST_SC_EJECT_NOT_SUPPORTED;
> >   			goto err_out;
> >   		}
> > +		if (!handler->hotplug.autoremove) {
> > +			kobject_uevent(&device->dev.kobj, KOBJ_CHANGE);
> > +			ost_code = ACPI_OST_SC_EJECT_NON_SPECIFIC_FAILURE;
> > +			goto err_out;
> > +		}
> >   		acpi_evaluate_hotplug_ost(handle, ACPI_NOTIFY_EJECT_REQUEST,
> >   					  ACPI_OST_SC_EJECT_IN_PROGRESS, NULL);
> >   		break;
> > 
> > Adding the check into "acpi_hotplug_notify_cb()", user need not change the
> > flag for removing container device by "sysfs eject".
> 
> Which is utterly confusing.  There is no reason whatsoever why the sysfs eject
> attribute should work differently from the event-triggered eject.  Quite the
> opposite is the case: it should work in the same way in my opinion so that it
> is possible to test the eject code path using that attribute.
> 
> I'm traveling now, but when I get back home (next week), I'll try to implement
> the thing I was talking about above.

It took some more time than I had expected, but I finally was able to get to that.

The following two patches implement the idea.  This is the minimum (in my opinion)
implementation and it may be extended in some ways.

Patch [1/2] introduces a new demand_offline flag for struct acpi_hotplug_profile
that makes acpi_scan_hot_remove() check the offline status of the device object's
companion physical devices to start with and return -EBUSY if at least one of them
is not offline.

Patch [2/2] uses that flag to implement the container handling.  The details are
in the changelog, but that's how it is supposed to work.

During the initial namespace scan the container ACPI scan handler should create
"physical" system container device under /sys/devices/system/container/ for
each ACPI container object (the sysfs name of that device should be the same as
the sysfs name of the corresponding container object and they should be linked
to each other via the firmware_node and physical_node symbolic links, respectively).
Those system container devices are initially online.

When a container eject event happens, acpi_scan_hot_remove() will notice that
hotplug.demand_offline is set in the device object's scan handler and will
check the online status of its "physical" companion device, which is online
(that is the system container device the above paragraph is about).  That will
cause KOBJ_CHANGE to be emitted for the system container device and -EBUSY to
be returned by acpi_scan_hot_remove().

Now, user space needs to offline the system container device through its online
sysfs attribute (that should be present, because the bus type for containers
provides the online and offline callbacks).  However, the offline for system
container devices will only succeed if the physical devices right below the
container are all offline, so user space will have to offline those devices
before attempting to offline the system container device itself.  When
finished, user space can trigger the container removal with the help of the
eject sysfs attribute of the ACPI container object pointed to by the system
container device's firmware_node link (this time the check in
acpi_scan_hot_remove() will succeed, because the system container device in
question is now offline).

The way it is implemented is a bit hackish (the driver_data pointer is slightly
abused), but that's a special case and I wanted to avoid adding new fields to
struct device just for handling it.

The patches haven't been tested yet.  I'm going to do that later today, but
first I need to take care of some other things, so that has to wait.

Thanks,
Rafael

--
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/