Re: [PATCH 0/5] driver core, of: support for reserved devices

[Zev Weiss <zev@bewilderbeest.net>]

On Thu, Oct 21, 2021 at 07:58:46PM PDT, Rob Herring wrote:
>On Thu, Oct 21, 2021 at 9:00 PM Zev Weiss <zev@bewilderbeest.net> wrote:
>>
>> Hello all,
>>
>> This series is another incarnation of a couple other patchsets I've
>> posted recently [0, 1], but again different enough in overall
>> structure that I'm not sure it's exactly a v2 (or v3).
>>
>> As compared to [1], it abandons the writable binary sysfs files and at
>> Frank's suggestion returns to an approach more akin to [0], though
>> without any driver-specific (aspeed-smc) changes, which I figure might
>> as well be done later in a separate series once appropriate
>> infrastructure is in place.
>
>I skimmed this, and overall I like the approach.
>
>> The basic idea is to implement support for a status property value
>> that's documented in the DT spec [2], but thus far not used at all in
>> the kernel (or anywhere else I'm aware of): "reserved".  According to
>> the spec (section 2.3.4, Table 2.4), this status:
>>
>>   Indicates that the device is operational, but should not be used.
>>   Typically this is used for devices that are controlled by another
>>   software component, such as platform firmware.
>>
>> With these changes, devices marked as reserved are (at least in some
>> cases, more on this later) instantiated, but will not have drivers
>> bound to them unless and until userspace explicitly requests it by
>> writing the device's name to the driver's sysfs 'bind' file.  This
>> enables appropriate handling of hardware arrangements that can arise
>> in contexts like OpenBMC, where a device may be shared with another
>> external controller not under the kernel's control (for example, the
>> flash chip storing the host CPU's firmware, shared by the BMC and the
>> host CPU and exclusively under the control of the latter by default).
>> Such a device can be marked as reserved so that the kernel refrains
>> from touching it until appropriate preparatory steps have been taken
>> (e.g. BMC userspace coordinating with the host CPU to arbitrate which
>> processor has control of the firmware flash).
>>
>> Patches 1-3 provide some basic plumbing for checking the "reserved"
>> status of a device, patch 4 is the main driver-core change, and patch
>> 5 tweaks the OF platform code to not skip reserved devices so that
>> they can actually be instantiated.
>>
>> One shortcoming of this series is that it doesn't automatically apply
>> universally across all busses and drivers -- patch 5 enables support
>> for platform devices, but similar changes would be required for
>> support in other busses (e.g. in of_register_spi_devices(),
>> of_i2c_register_devices(), etc.) and drivers that instantiate DT
>> devices.  Since at present a "reserved" status is treated as
>> equivalent to "disabled" and this series preserves that status quo in
>> those cases I'd hope this wouldn't be considered a deal-breaker, but
>> a thing to be aware of at least.
>>
>> Greg: I know on [1] you had commented nack-ing the addition of boolean
>> function parameters; patch 4 adds a flags mask instead in an analogous
>> situation.  I'm not certain how much of an improvement you'd consider
>> that (hopefully at least slightly better, in that the arguments passed
>> at the call site are more self-explanatory); if that's still
>> unsatisfactory I'd welcome any suggested alternatives.
>
>Can't we add a flag bit in struct device to reflect manual binding?
>bind will set it and unbind clears it.
>

I considered this (and actually drafted up a version that did exactly 
that), but it seemed like turning a parameter-passing problem into a 
state-maintenance problem (finding all the places that flag would need 
to be cleared and ensuring newly-added ones don't get missed, which 
unlike a function parameter the compiler can't really check for us).  
Given that the live range (definition to use) of that value is entirely 
within the codepath of a single call-chain (bind_store() -> 
device_driver_attach() -> __driver_probe_device()), continuing to 
maintain that state beyond that call chain seemed like unnecessary 
complexity to me, but if there's a consensus that that would actually be 
preferable I can certainly do it that way instead.


Zev