lkml.org 
[lkml]   [2020]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [PATCH v3] driver core: Break infinite loop when deferred probe can't be satisfied
From
Date


On 26/03/2020 16:39, Greg KH wrote:
> On Thu, Mar 26, 2020 at 06:31:10PM +0200, Andy Shevchenko wrote:
>> On Thu, Mar 26, 2020 at 03:01:22PM +0000, Grant Likely wrote:
>>> On 25/03/2020 12:51, Andy Shevchenko wrote:
>>>> On Tue, Mar 24, 2020 at 08:29:01PM -0700, Saravana Kannan wrote:
>>>>> On Tue, Mar 24, 2020 at 5:38 AM Andy Shevchenko <andriy.shevchenko@linux.intel.com> wrote:
>>>>>> Consider the following scenario.
>>>>>>
>>>>>> The main driver of USB OTG controller (dwc3-pci), which has the following
>>>>>> functional dependencies on certain platform:
>>>>>> - ULPI (tusb1210)
>>>>>> - extcon (tested with extcon-intel-mrfld)
>>>>>>
>>>>>> Note, that first driver, tusb1210, is available at the moment of
>>>>>> dwc3-pci probing, while extcon-intel-mrfld is built as a module and
>>>>>> won't appear till user space does something about it.
>>>>>>
>>>>>> This is depicted by kernel configuration excerpt:
>>>>>>
>>>>>> CONFIG_PHY_TUSB1210=y
>>>>>> CONFIG_USB_DWC3=y
>>>>>> CONFIG_USB_DWC3_ULPI=y
>>>>>> CONFIG_USB_DWC3_DUAL_ROLE=y
>>>>>> CONFIG_USB_DWC3_PCI=y
>>>>>> CONFIG_EXTCON_INTEL_MRFLD=m
>>>>>>
>>>>>> In the Buildroot environment the modules are probed by alphabetical ordering
>>>>>> of their modaliases. The latter comes to the case when USB OTG driver will be
>>>>>> probed first followed by extcon one.
>>>>>>
>>>>>> So, if the platform anticipates extcon device to be appeared, in the above case
>>>>>> we will get deferred probe of USB OTG, because of ordering.
>>>>>>
>>>>>> Since current implementation, done by the commit 58b116bce136 ("drivercore:
>>>>>> deferral race condition fix") counts the amount of triggered deferred probe,
>>>>>> we never advance the situation -- the change makes it to be an infinite loop.
>>>>>
>>>>> Hi Andy,
>>>>>
>>>>> I'm trying to understand this sequence of steps. Sorry if the questions
>>>>> are stupid -- I'm not very familiar with USB/PCI stuff.
>>>>
>>>> Thank you for looking into this. My answer below.
>>>>
>>>> As a first thing I would like to tell that there is another example of bad
>>>> behaviour of deferred probe with no relation to USB. The proposed change also
>>>> fixes that one (however, less possible to find in real life).
>>>>
>>>>>> ---8<---8<---
>>>>>>
>>>>>> [ 22.187127] driver_deferred_probe_trigger <<< 1
>>>>>>
>>>>>> ...here is the late initcall triggers deferred probe...
>>>>>>
>>>>>> [ 22.191725] platform dwc3.0.auto: deferred_probe_work_func in deferred list
>>>>>>
>>>>>> ...dwc3.0.auto is the only device in the deferred list...
>>>>>
>>>>> Ok, dwc3.0.auto is the only unprobed device at this point?
>>>>
>>>> Correct.
>>>>
>>>>>> [ 22.198727] platform dwc3.0.auto: deferred_probe_work_func 1 <<< counter 1
>>>>>>
>>>>>> ...the counter before mutex is unlocked is kept the same...
>>>>>>
>>>>>> [ 22.205663] platform dwc3.0.auto: Retrying from deferred list
>>>>>>
>>>>>> ...mutes has been unlocked, we try to re-probe the driver...
>>>>>>
>>>>>> [ 22.211487] bus: 'platform': driver_probe_device: matched device dwc3.0.auto with driver dwc3
>>>>>> [ 22.220060] bus: 'platform': really_probe: probing driver dwc3 with device dwc3.0.auto
>>>>>> [ 22.238735] bus: 'ulpi': driver_probe_device: matched device dwc3.0.auto.ulpi with driver tusb1210
>>>>>> [ 22.247743] bus: 'ulpi': really_probe: probing driver tusb1210 with device dwc3.0.auto.ulpi
>>>>>> [ 22.256292] driver: 'tusb1210': driver_bound: bound to device 'dwc3.0.auto.ulpi'
>>>>>> [ 22.263723] driver_deferred_probe_trigger <<< 2
>>>>>>
>>>>>> ...the dwc3.0.auto probes ULPI, we got successful bound and bumped counter...
>>>>>>
>>>>>> [ 22.268304] bus: 'ulpi': really_probe: bound device dwc3.0.auto.ulpi to driver tusb1210
>>>>>
>>>>> So where did this dwc3.0.auto.ulpi come from?
>>>>
>>>>> Looks like the device is created by dwc3_probe() through this call flow:
>>>>> dwc3_probe() -> dwc3_core_init() -> dwc3_core_ulpi_init() ->
>>>>> dwc3_ulpi_init() -> ulpi_register_interface() -> ulpi_register()
>>>>
>>>> Correct.
>>>>
>>>>>> [ 22.276697] platform dwc3.0.auto: Driver dwc3 requests probe deferral
>>>>>
>>>>> Can you please point me to which code patch actually caused the probe
>>>>> deferral?
>>>>
>>>> Sure, it's in drd.c.
>>>>
>>>> if (device_property_read_string(dev, "linux,extcon-name", &name) == 0) {
>>>> edev = extcon_get_extcon_dev(name);
>>>> if (!edev)
>>>> return ERR_PTR(-EPROBE_DEFER);
>>>> return edev;
>>>> }
>>>>
>>>>>> ...but extcon driver is still missing...
>>>>>>
>>>>>> [ 22.283174] platform dwc3.0.auto: Added to deferred list
>>>>>> [ 22.288513] platform dwc3.0.auto: driver_deferred_probe_add_trigger local counter: 1 new counter 2
>>>>>
>>>>> I'm not fully aware of all the USB implications, but if extcon is
>>>>> needed, why can't that check be done before we add and probe the ulpi
>>>>> device? That'll avoid this whole "fake" probing and avoid the counter
>>>>> increase. And avoid the need for this patch that's touching the code
>>>>> code that's already a bit delicate.
>>>>
>>>>> Also, with my limited experience with all the possible drivers in the
>>>>> kernel, it's weird that the ulpi device is added and probed before we
>>>>> make sure the parent device (dwc3.0.auto) can actually probe
>>>>> successfully.
>>>>
>>>> As I said above the deferred probe trigger has flaw on its own.
>>>> Even if we fix for USB case, there is (and probably will be) others.
>>>
>>> Right here is the driver design bug. A driver's probe() hook should *not*
>>> return -EPROBE_DEFER after already creating child devices which may have
>>> already been probed.
>>
>> Any documentation statement for this requirement?
>
> There shouldn't be. If you return ANY error from a probe function, your
> driver is essencially "dead" when it comes to that device, and it had
> better have cleaned up after itself. >
> That includes defering probe, that's not "special" here at all.

What is special in this case is that if a .probe() hook had registered a
child device, then removed that child device (so it did clean up after
itself) and then return -EPROBE_DEFER, then we end up in an endless
probe loop.

But this is unusual behaviour. Normally a .probe() hook checks all
required resources are available before registering any child devices.
This driver doesn't do that. Arguably this is indeed an additional
requirement beyond "clean up after yourself". I cannot find anyplace
where it is documented. In fact, I cannot find any documentation on
EPROBE_DEFER in the Documentation/ tree. How about the below?

>> By the way, I may imagine other mechanisms that probe the driver on other CPU
>> at the same time (let's consider parallel modprobes). The current code has a
>> flaw with that.
>
> That can't happen, the driver core prevents that.

Greg's right, that can't happen. At worst a driver will get an
additional defer event; but it all still works.

g.

---
diff --git a/Documentation/driver-api/driver-model/driver.rst
b/Documentation/driver-api/driver-model/driver.rst
index baa6a85c8287..46adede13aba 100644
--- a/Documentation/driver-api/driver-model/driver.rst
+++ b/Documentation/driver-api/driver-model/driver.rst
@@ -167,7 +167,17 @@ the driver to that device.

A driver's probe() may return a negative errno value to indicate that
the driver did not bind to this device, in which case it should have
-released all resources it allocated::
+released all resources it allocated. Optionally, probe() may return
+-EPROBE_DEFER if the driver depends on resources that are not yet
+available (e.g., supplied by a driver that hasn't initialized yet).
+The driver core will put the device onto the deferred probe list and
+will try to call it again later. Important: -EPROBE_DEFER must not be
+returned if probe() has already created child devices, even if those
+child devices have were removed again in a cleanup path. If -EPROBE_DEFER
+is returned after a child device has been registered, it may result in an
+infinite loop of .probe() calls to the same driver.
+
+::

void (*sync_state)(struct device *dev);
\
 
 \ /
  Last update: 2020-03-26 19:07    [W:0.068 / U:18.612 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site