Hi,

On 10/13/21 8:48 PM, Rafael J. Wysocki wrote:

On Wed, Oct 13, 2021 at 8:23 PM Hans de Goede [off-list ref] wrote:

quoted

Hi,

On 10/13/21 7:29 PM, Rafael J. Wysocki wrote:

quoted

On Sun, Oct 10, 2021 at 8:57 PM Hans de Goede [off-list ref] wrote:

quoted

The clk and regulator frameworks expect clk/regulator consumer-devices
to have info about the consumed clks/regulators described in the device's
fw_node.

To work around cases where this info is not present in the firmware tables,
which is often the case on x86/ACPI devices, both frameworks allow the
provider-driver to attach info about consumers to the clks/regulators
when registering these.

This causes problems with the probe ordering wrt drivers for consumers
of these clks/regulators. Since the lookups are only registered when the
provider-driver binds, trying to get these clks/regulators before then
results in a -ENOENT error for clks and a dummy regulator for regulators.

One case where we hit this issue is camera sensors such as e.g. the OV8865
sensor found on the Microsoft Surface Go. The sensor uses clks, regulators
and GPIOs provided by a TPS68470 PMIC which is described in an INT3472
ACPI device. There is special platform code handling this and setting
platform_data with the necessary consumer info on the MFD cells
instantiated for the PMIC under: drivers/platform/x86/intel/int3472.

For this to work properly the ov8865 driver must not bind to the I2C-client
for the OV8865 sensor until after the TPS68470 PMIC gpio, regulator and
clk MFD cells have all been fully setup.

The OV8865 on the Microsoft Surface Go is just one example, all X86
devices using the Intel IPU3 camera block found on recent Intel SoCs
have similar issues where there is an INT3472 HID ACPI-device, which
describes the clks and regulators, and the driver for this INT3472 device
must be fully initialized before the sensor driver (any sensor driver)
binds for things to work properly.

On these devices the ACPI nodes describing the sensors all have a _DEP
dependency on the matching INT3472 ACPI device (there is one per sensor).

This allows solving the probe-ordering problem by delaying the enumeration
(instantiation of the I2C-client in the ov8865 example) of ACPI-devices
which have a _DEP dependency on an INT3472 device.

The new acpi_dev_ready_for_enumeration() helper used for this is also
exported because for devices, which have the enumeration_by_parent flag
set, the parent-driver will do its own scan of child ACPI devices and
it will try to enumerate those during its probe(). Code doing this such
as e.g. the i2c-core-acpi.c code must call this new helper to ensure
that it too delays the enumeration until all the _DEP dependencies are
met on devices which have the new honor_deps flag set.

Signed-off-by: Hans de Goede <redacted>
---
 drivers/acpi/scan.c     | 36 ++++++++++++++++++++++++++++++++++--
 include/acpi/acpi_bus.h |  5 ++++-
 2 files changed, 38 insertions(+), 3 deletions(-)

diff --git a/drivers/acpi/scan.c b/drivers/acpi/scan.c
index 5b54c80b9d32..efee6ee91c8f 100644
--- a/drivers/acpi/scan.c
+++ b/drivers/acpi/scan.c

@@ -796,6 +796,12 @@ static const char * const acpi_ignore_dep_ids[] = {
        NULL
 };

+/* List of HIDs for which we honor deps of matching ACPI devs, when checking _DEP lists. */
+static const char * const acpi_honor_dep_ids[] = {
+       "INT3472", /* Camera sensor PMIC / clk and regulator info */
+       NULL
+};
+
 static struct acpi_device *acpi_bus_get_parent(acpi_handle handle)
 {
        struct acpi_device *device = NULL;

@@ -1757,8 +1763,12 @@ static void acpi_scan_dep_init(struct acpi_device *adev)
        struct acpi_dep_data *dep;

        list_for_each_entry(dep, &acpi_dep_list, node) {
-               if (dep->consumer == adev->handle)
+               if (dep->consumer == adev->handle) {
+                       if (dep->honor_dep)
+                               adev->flags.honor_deps = 1;

Any concerns about doing

adev->flags.honor_deps = dep->honor_dep;

here?

The idea is to set adev->flags.honor_deps even if the device has
multiple deps and only one of them has the honor_dep flag set.

If we just do:

        adev->flags.honor_deps = dep->honor_dep;

Then adev->flags.honor_deps ends up having the honor_dep
flag of the last dependency checked.

OK, but in that case dep_unmet may be blocking the enumeration of the
device even if the one in the acpi_honor_dep_ids[] list has probed
successfully.

Isn't that a concern?

For the devices where we set the dep->honor_dep flag this is
not a concern (based on the DSDTs which I've seen).

I also don't expect it to be a concern for other cases where we may
set that flag in the future either. This is an opt-in thing, so
I expect that in cases where we opt in to this, we also ensure that
any other _DEPs are also met (by having a Linux driver which calls
acpi_dev_clear_dependencies() for them).

And now a days we also have the acpi_ignore_dep_ids[] list so if
in the future there are some _DEP-s which never get fulfilled/met
on a device where we set the adev->flags.honor_deps flag, then we
can always add the ACPI HIDs for those devices to that list.

quoted

quoted

quoted

+
                        adev->dep_unmet++;
+               }
        }
 }

@@ -1962,7 +1972,7 @@ static u32 acpi_scan_check_dep(acpi_handle handle, bool check_dep)
        for (count = 0, i = 0; i < dep_devices.count; i++) {
                struct acpi_device_info *info;
                struct acpi_dep_data *dep;
-               bool skip;
+               bool skip, honor_dep;

                status = acpi_get_object_info(dep_devices.handles[i], &info);
                if (ACPI_FAILURE(status)) {

@@ -1971,6 +1981,7 @@ static u32 acpi_scan_check_dep(acpi_handle handle, bool check_dep)
                }

                skip = acpi_info_matches_ids(info, acpi_ignore_dep_ids);
+              honor_dep = acpi_info_matches_ids(info, acpi_honor_dep_ids);
                kfree(info);

                if (skip)

@@ -1984,6 +1995,7 @@ static u32 acpi_scan_check_dep(acpi_handle handle, bool check_dep)

                dep->supplier = dep_devices.handles[i];
                dep->consumer = handle;
+               dep->honor_dep = honor_dep;

                mutex_lock(&acpi_dep_list_lock);
                list_add_tail(&dep->node , &acpi_dep_list);

@@ -2071,6 +2083,9 @@ static acpi_status acpi_bus_check_add_2(acpi_handle handle, u32 lvl_not_used,

 static void acpi_default_enumeration(struct acpi_device *device)
 {
+       if (!acpi_dev_ready_for_enumeration(device))
+               return;

I'm not sure about this.

First of all, this adds an acpi_device_is_present() check here which
potentially is a change in behavior and I'm not sure how it is related
to the other changes in this patch (it is not mentioned in the
changelog AFAICS).

I'm saying "potentially", because if we get here at all,
acpi_device_is_present() has been evaluated already by
acpi_bus_attach().

Right the idea was that for this code-path the extra
acpi_device_is_present() check is a no-op since the only
caller of acpi_default_enumeration() has already done
that check before calling acpi_default_enumeration(),
where as the is_present check is useful for users outside
of the ACPI core code, like e.g. the i2c ACPI enumeration
code.

Although I see this is also called from
acpi_generic_device_attach which comes into play when there
is devicetree info embedded inside the ACPI tables.

That too, but generally speaking this change should at least be
mentioned in the changelog.

quoted

quoted

Now, IIUC, the new acpi_dev_ready_for_enumeration() is kind of an
extension of acpi_device_is_present(), so shouldn't it be called by
acpi_bus_attach() instead of the latter rather than from here?

That is an interesting proposal. I assume you want this to replace
the current acpi_device_is_present() call in acpi_bus_attach()
then ?

That seems consistent to me.

quoted

For the use-case at hand here that should work fine and it would also
make the honor_deps flag work for devices which bind to the actual
acpi_device (because we delay the device_attach()) or
use an acpi_scan_handler.

This would mean though that we can now have acpi_device-s where
acpi_device_is_present() returns true, but which are not
initialized (do not have device->flags.initialized set)
that would be a new acpi_device state which we have not had
before. I do not immediately forsee this causing issues,
but still...

If you want me to replace the current acpi_device_is_present() call
in acpi_bus_attach() with the new acpi_dev_ready_for_enumeration()
helper, let me know and I'll prepare a new version with this change
(and run some tests with that new version).

I would prefer doing that to making acpi_default_enumeration() special
with respect to the handling of dependencies.

Ok I will make this change in the next version (ETA sometime next week).

Regards,

Hans