Re: [PATCH v3] driver core: Break infinite loop when deferred probe can't be satisfied

[Saravana Kannan <saravanak@google.com>]

On Thu, Mar 26, 2020 at 1:39 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
>
> On Wed, Mar 25, 2020 at 11:09 PM Saravana Kannan <saravanak@google.com> wrote:
> >
> > On Wed, Mar 25, 2020 at 5:51 AM Andy Shevchenko
> > <andriy.shevchenko@linux.intel.com> wrote:
> > >
>
> [cut]
>
> > >
> > > Yes, it's (unlikely) possible (*), but it will give one more iteration per such
> > > case. It's definitely better than infinite loop. Do you agree?
> >
> > Sorry I wasn't being clear (I was in a rush). I'm saying this patch
> > can reintroduce the bug where the deferred probe isn't triggered when
> > it should be.
> >
> > Let's take a simple execution flow.
> >
> > probe_okay is at 10.
> >
> > Thread-A
> >   really_probe(Device-A)
> >     local_probe_okay_count = 10
> >     Device-A probe function is running...
> >
> > Thread-B
> >   really_probe(Device-B)
> >     Device-B probes successfully.
> >     probe_okay incremented to 11
> >
> > Thread-C
> >   Device-C (which had bound earlier) is unbound (say module is
> > unloaded or a million other reasons).
> >   probe_okay is decremented to 10.
> >
> > Thread-A continues
> >   Device-A probe function returns -EPROBE_DEFER
> >   driver_deferred_probe_add_trigger() doesn't do anything because
> >     local_probe_okay_count == probe_okay
> >   But Device-A might have deferred probe waiting on Device-B.
> >   Device-A never probes.
> >
> > > *) It means during probe you have _intensive_ removing, of course you may keep
> > > kernel busy with iterations, but it has no practical sense. DoS attacks more
> > > effective in different ways.
> >
> > I wasn't worried about DoS attacks. More of a functional correctness
> > issue what I explained above.
>
> The code is functionally incorrect as is already AFAICS.
>
> > Anyway, if your issue and similar issues can be handles in driver core
> > in a clean way without breaking other cases, I don't have any problem
> > with that. Just that, I think the current solution breaks other cases.
>
> OK, so the situation right now is that commit 58b116bce136 has
> introduced a regression and so it needs to be fixed or reverted.  The
> cases that were previously broken and were unbroken by that commit
> don't matter here, so you cannot argue that they would be "broken".
>
> It looks to me like the original issue fixed by the commit in question
> needs to be addressed differently, so I would vote for reverting it
> and starting over.

I'm fine with whatever approach. My only point is that code that's
been there for 5+ years might be preventing that race in a multitude
of platforms. So I'm just reviewing to make sure fixes aren't
introducing regressions. I'm all for anyone cleaning up/redoing
deferred probe.

> > As an alternate solution, assuming "linux,extcon-name" is coming
> > from some firmware, you might want to look into the fw_devlink
> > feature.
>
> That would be a workaround for a driver core issue, though, wouldn't it?

I'm not saying don't fix it in the driver core if it can be done
without adding regressions.

> > That feature allows driver core to add device links from firmware
> > information. If you can get that feature to create device links from
> > your dwc3.0.auto (or its parent pci_dev?) to the extcon supplier
> > device, all of this can be sidestepped and your dwc3.0.auto's (or the
> > dwc pci_dev's) probe will be triggered only after extcon is probed.
> >
> > I have very little familiarity with PCI/ACPI. I spent about an hour or
> > two poking at ACPI scan/property code. The relationship between a
> > pci_dev and an acpi_device is a bit confusing to me because I see:
> >
> > static int dwc3_pci_probe(struct pci_dev *pci, const struct pci_device_id *id)
> > {
> >         struct property_entry *p = (struct property_entry *)id->driver_data;
> >         struct dwc3_pci         *dwc;
> >         struct resource         res[2];
> >         int                     ret;
> >         struct device           *dev = &pci->dev;
> > ....
> >         dwc->dwc3 = platform_device_alloc("dwc3", PLATFORM_DEVID_AUTO);
> > ....
> >         ACPI_COMPANION_SET(&dwc->dwc3->dev, ACPI_COMPANION(dev));
> >
> > And ACPI_COMPANION returns an acpi_device by looking at dev->fwnode.
> > So how the heck is a pci_device.dev.fwnode pointing to an
> > acpi_device.fwnode?
>
> acpi_device is an of_node counterpart (or it is an fwnode itself if you will).

If I understand correctly, you are saying it's similar to struct
device_node for OF -- as in, a data struct that stores the unpacked
ACPI firmware data. That helps me understand what is going on with
ACPI_COMPANION_SET() in the PCI driver.

But then, why does it have a "struct device dev" field embedded in it?
Does the acpi_device.dev ever get registered with driver core?

Thanks,
Saravana