Re: [PATCH 3/3] cnic,bnx2,bnx2x: use UIO_MEM_DMA_COHERENT

[Jerry Snitselaar]

On Sun, Oct 01, 2023 at 01:57:25PM +0200, Greg Kroah-Hartman wrote:
> On Sun, Oct 01, 2023 at 12:44:05PM +0200, Hannes Reinecke wrote:
> > On 9/30/23 20:28, Greg Kroah-Hartman wrote:
> > > On Sat, Sep 30, 2023 at 11:19:20AM -0700, Chris Leech wrote:
> > > > On Sat, Sep 30, 2023 at 09:06:51AM +0200, Greg Kroah-Hartman wrote:
> > > > > On Fri, Sep 29, 2023 at 10:00:23AM -0700, Chris Leech wrote:
> > > > > > Make use of the new UIO_MEM_DMA_COHERENT type to properly handle mmap
> > > > > > for dma_alloc_coherent buffers.
> > > > > 
> > > > > Why are ethernet drivers messing around with UIO devices?  That's not
> > > > > what UIO is for, unless you are trying to do kernel bypass for these
> > > > > devices without anyone noticing?
> > > > > 
> > > > > confused,
> > > > 
> > > > It's confusing. The bnx2 driver stack included a cnic (converged nic?)
> > > > module that sits between the ethernet drivers (bnx2, bnx2x) and protocol
> > > > offload drivers (iscsi, fcoe, rdma).
> > > > 
> > > > The iscsi module (bnx2i) uses a passthrough interface from cnic to
> > > > handle some network configuration that the device firmware doesn't do.
> > > > It uses a uio device and a userspace component called iscsiuio to do
> > > > that.
> > > 
> > > That's horrible, and not what the UIO api is for at all.  Configure the
> > > device like any other normal kernel device, don't poke at raw memory
> > > values directly, that way lies madness.
> > > 
> > > Have a pointer to the userspace tool anywhere?  All I found looks like a
> > > full IP stack in userspace under that name, and surely that's not what
> > > this api is for...
> > > 
> > But that's how the interface is used, in particular for the bnx2i driver.
> > Problem is that the bnx2i iSCSI offload is just that, an iSCSI offload. Not
> > a TCP offload. So if the iSCSI interface is configured to
> > acquire the IP address via DHCP, someone has to run the DHCP protocol.
> > But the iSCSI offload can't, and the bnx2i PCI device is not a network
> > device so that the normal network stack can't be used.
> > And so the architects of the bnx2i card decided to use UIO to pass
> > the network traffic to userspace, and used the userspace 'iscsiuio'
> > application to run DHCP in userspace.
> > 
> > But's been that way for several years now; so long, in fact, that
> > the card itself has been out of support from Marvell (since quite some
> > years, too, IIRC). And even the successor of that card (the qedi driver)
> > is nearing EOL. Mind you, the qedi driver is using the same interface (by
> > using UIO to run DHCP in userspace), so singling out the bnx2i for bad
> > design can be construed as being unfair :-)
> 
> Ok, let's say they are all horrible! :)
> 
> > I agree, though, that the design is a mess.
> 
> Ok, so why are we papering over it and continuing to allow it to exist?
>
> What "broke" to suddenly require this UIO change?  If this has been
> around for a very long time, what has caused this to now require the UIO
> layer to change?
> 
> thanks,
> 
> greg k-h

Changes last year to the dma-mapping api to no longer allow __GFP_COMP,
in particular these two (from the e529d3507a93 dma-mapping pull for
6.2):

ffcb75458460 dma-mapping: reject __GFP_COMP in dma_alloc_attrs | 2022-11-21 | (Christoph Hellwig)
bb73955c0b1d cnic: don't pass bogus GFP_ flags to dma_alloc_coherent | 2022-11-21 | (Christoph Hellwig)


Regards,
Jerry