On 10/7/24 19:51, Andreas Hindborg wrote:
> Hi Damien,
>
> Damien Le Moal <dlemoal@xxxxxxxxxx> writes:
>
>> Commit 734e1a860312 ("block: Prevent deadlocks when switching
>> elevators") introduced the function elv_iosched_load_module() to allow
>> loading an elevator module outside of elv_iosched_store() with the
>> target device queue not frozen, to avoid deadlocks. However, the "none"
>> scheduler does not have a module and as a result,
>> elv_iosched_load_module() always returns an error when trying to switch
>> to this valid scheduler.
>
> The commit message here is a bit misleading. The problem is not that
> `request_module` can fail, the problem is that some failure modes cause
> this function to return a positive integer. This is not caught by
> callers and it ends up causing all kinds of problems in user space.
>
> Perhaps it makes sense to check for a positive return value at the call
> site of the `load_module` pointer in `queue_attr_store`, so this does
> not repeat at some point in the future?
Well, the patch version that went upstream totally ignores the return value of
request_module(), as it did before. So I do not think there are any issues
anymore... ? Might be good to add a comment about it above that request_module()
call in elv_iosched_load_module().
>
> Or maybe document that `load_module` implementations should not return a
> positive value unless it actually wants to send this to user space as
> the result of a write to the `scheduler` sysfs file?
Yes, ->load_module() should return 0 on success and negative error code on
error. Otherwise, a positive error may be interpreted by user space as a success
write to the sysfs attribute. Adding a comment for that will be good too.
Care to send a patch ?
--
Damien Le Moal
Western Digital Research
