Re: [PATCH/RFC G-U-P experts] IB/umem: Modernize our get_user_pages() parameters

[Roland Dreier <roland@kernel.org>]

Hi Andrea, sorry for the slow reply, had to work on other stuff for a bit.

On Mon, Jan 30, 2012 at 12:20 PM, Andrea Arcangeli <aarcange@redhat.com> wrote:
> If you map it with an mmap(PROT_READ|PROT_WRITE), force or not force
> won't change a thing in terms of cows. Just make sure you map your
> control memory right, then safely remove force=1 and you won't get the
> control page cowed by mistake. Then if you map it with MAP_SHARED it
> won't be mapped read-only by fork() (leading to either parent or child
> losing the control on the device), Hugh already suggested you to use
> MAP_SHARED instead of MAP_PRIVATE.

Actually this isn't about control memory for the RDMA adapter...
as you mentioned that typically is MMIO and mapped with remap_pfn
stuff, without using any GUP stuff.

I'm talking about the registration of other memory for reading/writing
by a remote system via RDMA.

The reason I'm talking about exporting kernel memory is that I wanted
to do a debugging trick where a kernel module exposed some state
into an mmap'able buffer.  And I wanted to be able to read that state
even if my broken module killed the whole system (in fact exactly
when things crash I want to be able to read the state to figure out
why I crashed!).

So I wrote a trivial userspace program that does nothing but mmap
the buffer, accept RDMA connections from remote systems, and
map the buffer for reading over those connections.  Then I can have
a second system that connects to that process and polls the buffer.

Because all the RDMA state is setup in advance, I can keep polling
even after the first system panics.  It's sort of like that firewire remote
debugging, except I only get access to a limited memory buffer.

The only difficulty is the problem that started this thread, ie a bogus
COW so the remote system ends up polling the wrong pages.  So with
my original patch, I'm able to debug but I guess we agree it's the
wrong fix for the general problem, and I'll write up a patch that adds
what I think is the correct fix (the new FOLL flag) soon.

 - R.