> > int ehca_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
> > {
>
> Can you split this monster routine into individual functions for
> each type of mmap please? With two helpers to get and verify the cq/qp
> shared by the individual sub-variants, that would also help to get rid
> of all those magic offsets.
>
> Actually, this routine directly comes from ib_device.mmap - Roland,
> can you shed some light on what's going on here?
Each userspace-accessible IB device gets a single device node like
/dev/infiniband/uverbsX. Opening that gives userspace a "context".
One of the things userspace can do with that fd is mmap() on it --
that was originally envisioned as a way to map a page of hardware
registers directly in to the userspace process.
It seems ehca needs to allocate lots of different things in the kernel
via mmap(). What you're saying I guess is that ideally each of these
would be mmap() on a different fd rather than using different
offsets. It's a little awkward to open multiple device nodes to get
multiple fds, since there's not a good way to attach them all to the
same context. I guess we could create some hack to return more file
handles, but I think that cure is worse than the disease of using
magic offsets...
Maybe longer term we need to look at a scheme like cell's spufs but
I'm still not confident we have the RDMA interface quite ready to
freeze at the system call level.
- R.
Roland Dreier wrote on 11.01.2007 20:54:58:
> > int ehca_mmap(struct ib_ucontext *context, struct vm_area_struct
*vma)
> > {
>
> Can you split this monster routine into individual functions for
> each type of mmap please? With two helpers to get and verify the
cq/qp
> shared by the individual sub-variants, that would also help to get rid
> of all those magic offsets.
>
> Actually, this routine directly comes from ib_device.mmap - Roland,
> can you shed some light on what's going on here?
Each userspace-accessible IB device gets a single device node like
/dev/infiniband/uverbsX. Opening that gives userspace a "context".
One of the things userspace can do with that fd is mmap() on it --
that was originally envisioned as a way to map a page of hardware
registers directly in to the userspace process.
It seems ehca needs to allocate lots of different things in the kernel
via mmap(). What you're saying I guess is that ideally each of these
would be mmap() on a different fd rather than using different
offsets. It's a little awkward to open multiple device nodes to get
multiple fds, since there's not a good way to attach them all to the
same context. I guess we could create some hack to return more file
handles, but I think that cure is worse than the disease of using
magic offsets...
Maybe longer term we need to look at a scheme like cell's spufs but
I'm still not confident we have the RDMA interface quite ready to
freeze at the system call level.
- R.
...as Roland mentions, we're not completely free to change the filehandle
usage,
it's shared by ~5 drivers now.
I'd say lets investigate the direction of an own filesystem unless
there's no other clean solution.
We can polish the current version a bit, but that won't change the "magic
offsets".
Roland, could you take this patchset into your tree?
We hope it adresses the major security concern and vm_insert_page.
We're preparing the next patch for the yield deadlock topic with this
patchset as prereq.
Gruss / Regards . . . Christoph Raisch