Re: [PATCH 2/2] fs/kernel_read_file: use usermodehelper_read_trylock() as a stop gap

[Lukas Middendorf]

On 17/04/2021 01:58, Luis Chamberlain wrote:
> The VFS lacks support to do automatic freeze / thaw of filesystems
> on the suspend / resume cycle. This can cause some issues, namely
> stalls when we have reads/writes during the suspend / resume cycle.
>
> Although for module loading / kexec the probability of this happening
> is extremely low, maybe even impossible, its a known real issue with
> the request_firmare() API which it does direct fs read. For this reason
> only be chatty about the issue on the call used by the firmware API.
>
> Lukas Middendorf has reported an easy situation to reproduce, which can
> be caused by questionably buggy drivers which call the request_firmware()
> API on resume.
[snip]
> The VFS fs freeze work fixes this issue, however it requires a bit
> more work, which may take a while to land upstream, and so for now
> this provides a simple stop-gap solution.
>
> We can remove this stop-gap solution once the kernel gets VFS
> fs freeze / thaw support.
>
> Reported-by: Lukas Middendorf <kernel@xxxxxxxxxxx>
> Signed-off-by: Luis Chamberlain <mcgrof@xxxxxxxxxx>

Tested-by: Lukas Middendorf <kernel@xxxxxxxxxxx>


Works as advertised.

This prevents stalls on resume with buggy drivers (e.g. si2168) by 
totally blocking uncached request_firmware() on resume. Uncached 
request_firmware() will fail reliably (also in situations where it by 
accident worked previously without stalling).
If firmware caching has been set up properly before suspend (either 
through firmware_request_cache() or through request_firmware() outside 
of a suspend/resume situation), the call to request_firmware() will 
still work as expected on resume. This should not break properly 
behaving drivers.

A failing firmware load is definitely preferable (and easier to debug 
and fix in the respective drivers) compared to a stall on resume.

Lukas