On Wed 02-10-13 09:20:09, Christoph Hellwig wrote:
On Wed, Oct 02, 2013 at 04:27:41PM +0200, Jan Kara wrote:
Hello,
In my quest for changing locking around page faults to make things easier for filesystems I found out get_user_pages() users could use a cleanup. The knowledge about necessary locking for get_user_pages() is in tons of places in drivers and quite a few of them actually get it wrong (don't have mmap_sem when calling get_user_pages() or hold mmap_sem when calling copy_from_user() in the surrounding code). Rather often this actually doesn't seem necessary. This patch series converts lots of places to use either get_user_pages_fast() or a new simple wrapper get_user_pages_unlocked() to remove the knowledge of mmap_sem from the drivers. I'm still looking into converting a few remaining drivers (most notably v4l2) which are more complex.
Even looking over the kerneldoc comment next to it I still fail to understand when you'd want to use get_user_pages_fast and when not.
AFAIU get_user_pages_fast() should be used 1) if you don't need any special get_user_pages() arguments (like calling it for mm of a different process, forcing COW, or similar). 2) you don't expect pages to be unmapped (then get_user_pages_fast() is actually somewhat slower because it walks page tables twice).
Honza