Re: Re: [PATCH] rdma/siw: avoid smp_store_mb() on a u64

[Jason Gunthorpe]

On Fri, Jul 12, 2019 at 03:22:35PM +0200, Arnd Bergmann wrote:
> On Fri, Jul 12, 2019 at 3:05 PM Bernard Metzler <BMT@xxxxxxxxxxxxxx> wrote:
> 
> >
> > We share CQ (completion queue) notification flags between application
> > (which may be user land) and producer (kernel QP's (queue pairs)).
> > Those flags can be written by both application and QP's. The application
> > writes those flags to let the driver know if it shall inform about new
> > work completions. It can write those flags at any time.
> > Only a kernel producer reads those flags to decide if
> > the CQ notification handler shall be kicked, if a new CQ element gets
> > added to the CQ. When kicking the completion handler, the driver resets the
> > notification flag, which must get re-armed by the application.
> >
> > We use READ_ONCE() and WRITE_ONCE(), since the flags are potentially
> > shared (mmap'd) between user and kernel land.
> >
> > siw_req_notify_cq() is being called only by kernel consumers to change
> > (write) the CQ notification state. We use smp_store_mb() to make sure
> > the new value becomes visible to all kernel producers (QP's) asap.
> >
> >
> > From cfb861a09dcfb24a98ba0f1e26bdaa1529d1b006 Mon Sep 17 00:00:00 2001
> > From: Bernard Metzler <bmt@xxxxxxxxxxxxxx>
> > Date: Fri, 12 Jul 2019 13:19:27 +0200
> > Subject: [PATCH] Make shared CQ notification flags 32bit to respect 32bit
> >  architectures
> >
> > Signed-off-by: Bernard Metzler <bmt@xxxxxxxxxxxxxx>
> 
> This fixes the build for me, thanks!
> 
> Tested-by: Arnd Bergmann <arnd@xxxxxxxx>

Since this is coming up so late in the merge window, I'm inclined to
take the simple path while Bernard makes a complete solution
here. What do you think Arnd?