On Thu, Aug 08, 2019 at 02:17:57PM +0000, Bernard Metzler wrote: > > >To: "Bernard Metzler" <BMT@xxxxxxxxxxxxxx>, "Jason Gunthorpe" > ><jgg@xxxxxxxx> > >From: "Doug Ledford" <dledford@xxxxxxxxxx> > >Date: 08/07/2019 08:53PM > >Cc: linux-rdma@xxxxxxxxxxxxxxx > >Subject: [EXTERNAL] Re: Re: [PATCH 1/1] Make user mmapped CQ arming > >flags field 32 bit size to remove 64 bit architecture dependency of > >siw. > > > >On Wed, 2019-08-07 at 17:49 +0000, Bernard Metzler wrote: > >> > >> It hurts, but I did finally setup qemu with a ppc image to check, > >> and so you are right! > >> > >> ... > >> > >> #ifdef __BIG_ENDIAN__ > >> > >> seem to be available in both kernel and user land... > >> > >> But, general question: siw in its current shape isn't out > >> for long, older versions from github are already broken with > >> the abi. So, silently changing the abi at this stage of siw > >> deployment is no option? It's a hassle to see an old mistake > >> carried along forever with that #ifdef statement...no? > > > >I was thinking about that myself. > > > >This really hasn't been out long enough to completely tie our hands > >here. A point update to rdma-core will resolve any user side issues. > > > > What we are aiming at is ensuring backward compatibility > for 64bit-BE architectures, which are using siw since this year. > The community is likely of size zero. > I found it hard to find a machine, even in the ppc world, which > let me test that BE thing. I ended up with an emulator. So I > assume it is no real world issue to now just change the 64bits > flag into 32bits and add a 32bit pad for ABI compliance. This seems reasonable to me, document all these details in the commit message. Jason
