On Mon, Aug 19, 2019 at 02:52:13PM +0000, Bernard Metzler wrote: > > >To: "Bernard Metzler" <[email protected]> > >From: "Jason Gunthorpe" <[email protected]> > >Date: 08/19/2019 04:19PM > >Cc: "Geert Uytterhoeven" <[email protected]>, "Doug Ledford" > ><[email protected]>, [email protected], > >[email protected] > >Subject: [EXTERNAL] Re: Re: Re: [PATCH] RDMA/siw: Fix compiler > >warnings on 32-bit due to u64/pointer abuse > > > >On Mon, Aug 19, 2019 at 02:15:36PM +0000, Bernard Metzler wrote: > >> > >> >To: "Bernard Metzler" <[email protected]> > >> >From: "Jason Gunthorpe" <[email protected]> > >> >Date: 08/19/2019 03:52PM > >> >Cc: "Geert Uytterhoeven" <[email protected]>, "Doug Ledford" > >> ><[email protected]>, [email protected], > >> >[email protected] > >> >Subject: [EXTERNAL] Re: Re: [PATCH] RDMA/siw: Fix compiler > >warnings > >> >on 32-bit due to u64/pointer abuse > >> > > >> >On Mon, Aug 19, 2019 at 01:36:11PM +0000, Bernard Metzler wrote: > >> >> >If the value is really a kernel pointer, then it ought to be > >> >printed > >> >> >with %p. We have been getting demanding on this point lately in > >> >RDMA > >> >> >to enforce the ability to keep kernel pointers secret. > >> >> > > >> >> >> - wqe->sqe.sge[0].laddr = (u64)&wqe->sqe.sge[1]; > >> >> >> + wqe->sqe.sge[0].laddr = > >> >> >> (uintptr_t)&wqe->sqe.sge[1]; > >> >> > > >> >> >[..] > >> >> > > >> >> >> rv = siw_rx_kva(srx, > >> >> >> - (void *)(sge->laddr + > >> >> >> frx->sge_off), > >> >> >> + (void *)(uintptr_t)(sge->laddr > >> >> >> + frx->sge_off), > >> >> >> sge_bytes); > >> >> > > >> >> >Bernard, this is nonsense, what is going on here with > >sge->laddr > >> >that > >> >> >it can't be a void *? > >> >> > > >> >> siw_sge is defined in siw-abi.h. We make the address u64 to keep > >> >the ABI > >> >> arch independent. > >> > > >> >Eh? How does the siw-abi.h store a kernel pointer? Sounds like > >kernel > >> >and user types are being mixed. > >> > > >> > >> siw-abi.h defines the work queue elements of a siw send queue. > >> For user land, the send queue is mmapped. Kernel or user land > >> clients write to its send queue when posting work > >> (SGE: buffer address, length, local key). > > > >Should have different types.. Don't want to accidently mix a laddr > >under user control with one under kernel control. > > > Well we have an unsigned 64bit for both user and kernel > application buffer addresses throughout the rdma stack,
We do not. Kernel addresses are consistenyly void * or dma_addr_t Most places that consume a data address are using lkeys anyhow, which does have a lkey & u64, but that u64 is not a application buffer address, but the IOVA of the lkey, which is very different. I really have no idea why siw needs to mix kernel VAs with user pointers, particularly in wqes... Jason
