Re: [PATCH net] net: wan: fsl_uhdlc_hdlc: fix dma_rmb usage in hdlc_rx_done

["Christophe Leroy (CS GROUP)" <chleroy@kernel.org>]

Le 05/05/2026 à 10:14, Holger Brunck a écrit :
>>
>> Le 04/05/2026 à 17:56, Holger Brunck a écrit :
>>> If dma_rmb is used it has to be done after reading bd_status and
>>> checking if R_E_S is zero. Therefore we need to move it into the while loop.
>>
>> Can you give more details ? Why does dma_rmb() has to be done after reading
>> bd_status and checking if R_E_S is zero ?
>>
> 
> when R_E_S is zero in the status of the buffer descriptor it means the buffer is
> filled with data from the device.  Now the CPU owns the descriptor. Now we
> should execute the dma_rmb to be sure that we read the data correctly.
> And this we need to redo for each buffer descriptor which is filled with data,
> that’s why it must be done within the for loop and not before and after.

We enter hdlc_rx_done() after an interrupt which triggers scheduling of 
ucc_hdlc_poll(). I think dma_rmb() is needed _before_ reading the first 
status, otherwise it might read an erroneous status.

Once we are here the interrupt has been cleared so any new buffer will 
trigger a new interrupt and call again this function. Therefore I don't 
think it is worth the cost of a dma_rmb() inside the loop.

Christophe