> -----Original Message----- > From: Francois Romieu <romieu@xxxxxxxxxxxxx> > Sent: Monday, March 20, 2023 7:33 AM > To: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx> > Cc: linux-hyperv@xxxxxxxxxxxxxxx; netdev@xxxxxxxxxxxxxxx; Dexuan Cui > <decui@xxxxxxxxxxxxx>; KY Srinivasan <kys@xxxxxxxxxxxxx>; Paul Rosswurm > <paulros@xxxxxxxxxxxxx>; olaf@xxxxxxxxx; vkuznets@xxxxxxxxxx; > davem@xxxxxxxxxxxxx; wei.liu@xxxxxxxxxx; edumazet@xxxxxxxxxx; > kuba@xxxxxxxxxx; pabeni@xxxxxxxxxx; leon@xxxxxxxxxx; Long Li > <longli@xxxxxxxxxxxxx>; ssengar@xxxxxxxxxxxxxxxxxxx; linux- > kernel@xxxxxxxxxxxxxxx > Subject: Re: [PATCH net-next] net: mana: Add support for jumbo frame > > [You don't often get email from romieu@xxxxxxxxxxxxx. Learn why this is > important at https://aka.ms/LearnAboutSenderIdentification ] > > Haiyang Zhang <haiyangz@xxxxxxxxxxxxx> : > > > From: Francois Romieu <romieu@xxxxxxxxxxxxx> > [...] > > > I do not see where the driver could depend on the MTU. Even if it fails, > > > a single call to mana_change_mtu should thus never wreck the old working > > > state/configuration. > > > > > > Stated differently, the detach/attach implementation is simple but > > > it makes the driver less reliable than it could be. > > > > > > No ? > > > > No, it doesn't make the driver less reliable. To safely remove and reallocate > > DMA buffers with different size, we have to stop the traffic. So, > mana_detach() > > is called. We also call mana_detach() in mana_close(). So the process in > > mana_change_mtu() is no more risky than ifdown/ifup of the NIC. > > > > In some rare cases, if the system memory is running really low, the bigger > > buffer allocation may fail, so we re-try with the previous MTU. I don't expect > > it to fail again. But we still check & log the error code for completeness and > > debugging. > > In a ideal world, I would expect change_mtu() to allocate the new resources, > bail out if some allocation fails, stop the traffic, swap the old and new > resources, then restart the traffic and release the old resources. > This way the device is never left in a failed state. Thanks for the idea -- I will look into that. But the real world implementation may be less than ideal :) - Haiyang
