[openib-general] Re: [PATCH][RFC][0/4] InfiniBand userspace verbs implementation

[Andrew Morton]

Timur Tabi <timur.tabi at ammasso.com> wrote:
>
> Andrew Morton wrote:
> 
> > This is because there is no file descriptor or anything else associated
> > with the pages which permits the kernel to clean stuff up on unclean
> > application exit.  Also there are the obvious issues with permitting
> > pinning of unbounded amounts of memory.
> 
> Then that might explain the "bug" that we're seeing with get_user_pages().  We've been 
> assuming that get_user_pages() mappings are permanent.

They are permanent until someone runs put_page() against all the pages. 
What I'm saying is that all current callers of get_user_pages() _do_ run
put_page() within the same syscall or upon I/O termination.

> Well, I was just about to re-implement get_user_pages() support in our driver to 
> demonstrate the bug.  I guess I'll hold off on that.
> 
> If you look at the Infiniband code that was recently submitted, I think you'll see it does 
> exactly that: after calling mlock(), the driver calls get_user_pages(), and it stores the 
> page mappings for future use.

Where?

bix:/usr/src/linux-2.6.12-rc3> grep -rl get_user_pages .
./arch/i386/lib/usercopy.c
./arch/sparc64/kernel/ptrace.c
./drivers/video/pvr2fb.c
./drivers/media/video/video-buf.c
./drivers/scsi/sg.c
./drivers/scsi/st.c
./include/asm-ia64/pgtable.h
./include/linux/mm.h
./include/asm-um/archparam-i386.h
./include/asm-i386/fixmap.h
./fs/nfs/direct.c
./fs/aio.c
./fs/binfmt_elf.c
./fs/bio.c
./fs/direct-io.c
./kernel/futex.c
./kernel/ptrace.c
./mm/memory.c
./mm/nommu.c
./mm/rmap.c
./mm/mempolicy.c