The deferred probe mechanism uses a successful driver probe/attach as a trigger to revisit the list of deferred probe devices. This works in most cases, except when the probe success is not a valid indicator of resources being available.
In that case, a race condition may occur, where the device/driver core framework will attempt to probe a device that depends on resources before those resources are available, resulting in a -EPROBE_DEFER error and a deferred probe device that will never be initialized.
The example provided in this RFC relies on the probe workqueue used for the HDaudio support where we simultaneously: a) need to use request_module() b) cannot use an async probe due to the use of request_module() c) cannot block the probe of other drivers In this example, the deferred probe can be kicked when the workqueue completes.
The use of request_firmware_nowait() is another conceptual example, where a domain-specific callback can enable resources *after* the probe returns, for example by downloading the firmware, booting a processor and waiting for the processor to be ready for interaction with the Linux host. In this second example, the deferred probe could be kicked when the 'cont' callback completes.
This patchset suggests a 7-line change to solve race conditions in these examples with delayed work.
Discussion:
a) During Intel internal reviews, Andy Shevchenko pointed out another known issue with deferred probe [1]. This patchset is unrelated and does not claim to solve the problem raised by Andy.
b) one possible objection is that this patchset does not suppress a possibly unnecessary round of evaluation of deferred probe devices. It did not feel necessary to any of us to minimize the occurrences of EPROBE_DEFER but instead to make sure the device waiting for resources successfully probes in the end.
c) another objection might be that the driver core should know about such dependencies. It would be desirable but in the cases we've encountered such dependencies are highly domain-specific and not necessarily straightforward to describe. There's been multiple endeavors to improve the description of dependencies, this patchset only focuses on the deferred probe framework, with an improvement when the provider of resources makes these resources available after its probe returns.
[1] https://lore.kernel.org/lkml/20200324175719.62496-1-andriy.shevchenko@linux....
Pierre-Louis Bossart (2): driver core: export driver_deferred_probe_trigger() ASoC: SOF: trigger re-probing of deferred devices from workqueue
drivers/base/dd.c | 3 ++- include/linux/device/driver.h | 1 + sound/soc/sof/core.c | 3 +++ 3 files changed, 6 insertions(+), 1 deletion(-)
base-commit: 8d1998893cd5e3488cd95529f60a187e3009d14b