Re: [PATCH net] ice: Do not get coalesce settings while in reset

[Dawid Osuchowski <dawid.osuchowski@linux.intel.com>]

On 06.05.2024 15:30, Dawid Osuchowski wrote:
> On 02.05.2024 04:56, Jakub Kicinski wrote:
>> Did you not add locks around reset to allow waiting instead of returning
>> -EBUSY to user space? I feel like we've been over this...
> 
> Will use the approach with ice_wait_for_reset() in next revision, thanks
> 
> --Dawid

Hey Jakub,

I went ahead with the approach of using ice_wait_for_reset() [1], 
however this resulted in a new problem in the reset flow. I want to 
prove why I think returning immediately with -EBUSY (or perhaps -EAGAIN) 
is the correct way in this particular case.

The issue has to deal with the way both the ethtool handler and the 
adapter reset flow call rtnl_lock() during operation. If we wait for 
reset completion inside of an ethtool handling function such as 
ice_get_coalesce(), the wait will always timeout due to reset being 
blocked by rtnl_lock() inside of ice_queue_set_napi() (which is called 
during reset process), and in turn we will always return -EBUSY anyways, 
with the added hang time of the timeout value (in case of [1] it's 10 
seconds).

There are other places where similar deadlock can occur, not only in 
ice_queue_set_napi() and Larysa is currently working on an extensive 
solution to this problem.

--Dawid

[1] 
https://lore.kernel.org/netdev/20240506153307.114104-1-dawid.osuchowski@linux.intel.com/