On Tue, Nov 9, 2021 at 7:12 PM Bjorn Helgaas [off-list ref] wrote:

On Tue, Nov 09, 2021 at 06:18:18PM +0100, Rafael J. Wysocki wrote:

quoted

On Tue, Nov 9, 2021 at 7:59 AM Uwe Kleine-König
[off-list ref] wrote:

quoted

On Mon, Nov 08, 2021 at 08:56:19PM -0600, Bjorn Helgaas wrote:

quoted

[+cc Greg: new device_is_bound() use]

ack, that's what I would have suggested now, too.

quoted

On Mon, Nov 08, 2021 at 10:22:26PM +0100, Uwe Kleine-König wrote:

quoted

pci_pm_runtime_resume() exits early when the device to resume isn't
bound yet:

    if (!to_pci_driver(dev->driver))
            return 0;

This however isn't true when the device currently probes and
local_pci_probe() calls pm_runtime_get_sync() because then the driver
core already setup dev->driver. As a result the driver's resume callback
is called before the driver's probe function is called and so more often
than not required driver data isn't setup yet.

So replace the check for the device being unbound by a check that only
becomes true after .probe() succeeded.

I like the fact that this patch is short and simple.

But there are 30+ users of to_pci_driver().  This patch asserts that
*one* of them, pci_pm_runtime_resume(), is special and needs to test
device_is_bound() instead of using to_pci_driver().

Maybe for the other locations using device_is_bound(&pdev->dev) instead
of to_pci_driver(pdev) != NULL would be nice, too?

I have another doubt: device_is_bound() should (according to its
kernel-doc) be called with the device lock held. For the call stack that
is (maybe) fixed here, the lock is held (by __device_attach). We
probably should check if the lock is also held for the other calls of
pci_pm_runtime_resume().

Hmm, the device lock is a mutex, the pm functions might be called in
atomic context, right?

quoted

It's special because the current PM implementation calls it via
pm_runtime_get_sync() before the driver's .probe() method.  That
connection is a little bit obscure and fragile.  What if the PM
implementation changes?

Maybe a saver bet would be to not use pm_runtime_get_sync() in
local_pci_probe()?

Yes, in principle it might be replaced with pm_runtime_get_noresume().

In theory, that may be problematic if a device is put into a low-power
state on remove and then the driver is bound again to it.

quoted

I wonder if the same problem exists on remove, i.e. pci_device_remove()
calls pm_runtime_put_sync() after the driver's .remove() callback was
called.

If it is called after ->remove() and before clearing the device's
driver pointer, then yes.

Yes, that is the case:

  pci_device_remove
    if (drv->remove) {
      pm_runtime_get_sync
      drv->remove()                # <-- driver ->remove() method
      pm_runtime_put_noidle
    }
    ...
    pm_runtime_put_sync            # <-- after ->remove()

So pm_runtime_put_sync() is called after drv->remove(), and it may
call drv->pm->runtime_idle().  I think the driver may not expect this.

quoted

If this is turned into pm_runtime_put_noidle(), all should work.

pci_device_remove() already calls pm_runtime_put_noidle() immediately
after calling the driver ->remove() method.

Are you saying we should do this, which means pci_device_remove()
would call pm_runtime_put_noidle() twice?

Well, they are both needed to keep the PM-runtime reference counting in balance.

This still has an issue, though, because user space would be able to
trigger a runtime suspend via sysfs after we've dropped the last
reference to the device in pci_device_remove().

So instead, we can drop the pm_runtime_get_sync() and
pm_runtime_put_sync() from local_pci_probe() and pci_device_remove(),
respectively, and add pm_runtine_get_noresume() to pci_pm_init(),
which will prevent PM-runtime from touching the device until it has a
driver that supports PM-runtime.

We'll lose the theoretical ability to put unbound devices into D3 this
way, but we learned some time ago that this isn't safe in all cases
anyway.