[ofa-general] ib_umem_get always wants write access

[Pete Wyckoff]

rdreier at cisco.com wrote on Wed, 21 Mar 2007 11:23 -0700:
>  > I've wondered about this for a while.  In ib_umem_get, there is a
>  > call to get_user_pages that does the work of virtual to physical
>  > translation and increasing the ref counts.  It is always invoked
>  > with write == 1, even if cmd.access_flags == 0 (read only
>  > registration).
>  > 
>  > This is fine for anonymous private memory, or writeable shared
>  > memory.  But consider pinning a read-only section of memory, such as
>  > shared read-only data or text segment, or a file mapping of a file
>  > that was opened O_RDONLY.  Having write == 1 there forces a full
>  > copy of all these pages.
>  > 
>  > The force argument is explicitly set to 1 only when access_flags
>  > does not specify write access, giving gup permission to do the
>  > copy-on-write, essentially.  That seems correct, but always setting
>  > write to 1 has me confused.
>  > 
>  > Is there some IB semantic reason for forcing the registered pages to
>  > be writable?
> 
> I'm having a hard time remembering the exact reasoning, but the basic
> idea is that we need to allow read-only memory to be registered but we
> also need to force allocated but not touched memory to be faulted in.

Thanks, I'll try to setup a scenario where read-only memory is not
present, then gup with write = 0 and see if it does not do the
faulting properly.  It's not clear to me now.

The performance degradation of COW-ing read-only pages is noticable.

		-- Pete