On Wed, May 10, 2017 at 09:37:04AM +0200, Arnd Bergmann wrote:

quoted

How about trying to remove all of them?  If we could actually get rid
of all of them, we could drop the arch support, and we'd get faster,
simpler, shorter uaccess code throughout the kernel.

BTW, not all get_user() under KERNEL_DS are plain loads.  There is an
exception - probe_kernel_read().

quoted

The ones in kernel/compat.c are generally garbage.  They should be
using compat_alloc_user_space().  Ditto for kernel/power/user.c.

compat_alloc_user_space() has some problems too, it adds
complexity to a rarely-tested code path and can add some noticeable
overhead in cases where user space access is slow because of
extra checks.

It's clearly better than set_fs(), but the way I prefer to convert the
code is to avoid both and instead move compat handlers next to
the native code, and splitting out the common code between native
and compat mode into a helper that takes a regular kernel pointer.

I think that's what both Al has done in the past on compat_ioctl()
and select() and what Christoph does in his latest series, but
it seems worth pointing out for others that decide to help out here.

Folks, reducing the amount of places where we play with set_fs() is certainly
a good thing.  Getting rid of them completely is something entirely different;
I have tried to plot out patch series in this direction many times during the
last 5 years or so, but it's not going to be easy.  Tomorrow I can start
posting my notes in that direction (and there are tons of those, unfortunately
mixed with git grep results, highly unprintable personal comments, etc.);
just let me grab some sleep first...

BTW, slow userland access is not just due to extra checks; access_ok(),
in particular, is pretty much noise.  The real PITA comes from the things
like STAC/CLAC on recent x86.  Or hardware overhead of cross-address-space
block copy insn (e.g. on s390, where it's optimized for multi-cacheline
blocks).  Or things like uml, where it's a matter of walking the page
tables for each sodding __get_user().  It's not always just a matter of
address space limit...