Re: [PATCH net V2] net: stmmac: Fix for mismatched host/device DMA address width

[Jochen Henneberg]

"Russell King (Oracle)" <linux@xxxxxxxxxxxxxxx> writes:

> On Thu, Mar 16, 2023 at 02:15:03PM +0100, Jochen Henneberg wrote:
>> Currently DMA address width is either read from a RO device register
>> or force set from the platform data. This breaks DMA when the host DMA
>> address width is <=32it but the device is >32bit.
>> 
>> Right now the driver may decide to use a 2nd DMA descriptor for
>> another buffer (happens in case of TSO xmit) assuming that 32bit
>> addressing is used due to platform configuration but the device will
>> still use both descriptor addresses as one address.
>> 
>> This can be observed with the Intel EHL platform driver that sets
>> 32bit for addr64 but the MAC reports 40bit. The TX queue gets stuck in
>> case of TCP with iptables NAT configuration on TSO packets.
>> 
>> The logic should be like this: Whatever we do on the host side (memory
>> allocation GFP flags) should happen with the host DMA width, whenever
>> we decide how to set addresses on the device registers we must use the
>> device DMA address width.
>> 
>> This patch renames the platform address width field from addr64 (term
>> used in device datasheet) to host_addr and uses this value exclusively
>> for host side operations while all chip operations consider the device
>> DMA width as read from the device register.
>> 
>> Fixes: 7cfc4486e7ea ("stmmac: intel: Configure EHL PSE0 GbE and PSE1 GbE to 32 bits DMA addressing")
>> Signed-off-by: Jochen Henneberg <jh@xxxxxxxxxxxxxxxxxxxxxxxxxx>
>> ---
>> V2: Fixes from checkpatch.pl for commit message
>> 
>>  drivers/net/ethernet/stmicro/stmmac/common.h  |  1 +
>>  .../net/ethernet/stmicro/stmmac/dwmac-imx.c   |  2 +-
>>  .../net/ethernet/stmicro/stmmac/dwmac-intel.c |  4 +--
>>  .../ethernet/stmicro/stmmac/dwmac-mediatek.c  |  2 +-
>>  .../net/ethernet/stmicro/stmmac/stmmac_main.c | 30 ++++++++++---------
>>  include/linux/stmmac.h                        |  2 +-
>>  6 files changed, 22 insertions(+), 19 deletions(-)
>> 
>> diff --git a/drivers/net/ethernet/stmicro/stmmac/common.h b/drivers/net/ethernet/stmicro/stmmac/common.h
>> index 6b5d96bced47..55a728b1b708 100644
>> --- a/drivers/net/ethernet/stmicro/stmmac/common.h
>> +++ b/drivers/net/ethernet/stmicro/stmmac/common.h
>> @@ -418,6 +418,7 @@ struct dma_features {
>>  	unsigned int frpbs;
>>  	unsigned int frpes;
>>  	unsigned int addr64;
>> +	unsigned int host_addr;
>
> Obvious question: is host_addr an address? From the above description it
> sounds like this is more of a host address width indicator.
>
> Maybe call these "dev_addr_width" and "host_addr_width" so it's clear
> what each of these are?

You are right. I chose the name because the original field was called
addr64 which follows the naming from the chip specification. I will
switch to host_dma_width which makes it more clear that it's a DMA
address width. For both the platform field as well as the driver's
private data.