Re: [git pull] uaccess-related bits of vfs.git

[Linus Torvalds]

On Fri, May 12, 2017 at 11:57 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> First, some stats: there's a thousand-odd callers of __get_user().  Out of
> those, about 70% are in arch/, mostly in sigframe-related code.

Sure. And they can be trivially converted, and none of them should care at all.

> IOW, we have
>         * most of users in arch/* (heavily dominated by signal-related code,
> both loads and stores).  Those need careful massage; maybe unsafe-based
> solution, maybe something else, but it's obviously per-architecture work
> and these paths are sensitive.

Why are they sensitive?

Why not just do this:

  git grep -l '\<__\(\(get\)\|\(put\)\)_user(' -- arch/x86
:^arch/x86/include/asm/uaccess.h
        | xargs sed -i 's/__\(\(\(get\)\|\(put\)\)_user(\)/\1/g'

which converts all the x86 uses in one go.

Anybody who *relies* on not checking the address_limit is so broken as
to be not even funny. And anything that is so performance-sensitive
that anybody can even measure the effect of the above we can convert
later.

Sure, do it in pieces (eg each architecture separately, then
"drivers", then "networking", then whatever, until all done), just so
that *if* something actually depends on semantics (and that really
shouldn't be the case), we have at least some information from a
bisect.

But I don't see the excuse for not just doing it. If nobody notices,
it's an obvious improvement. And if somebody *does* notice, we know
how to do it properly with unsafe_xyz_user(), because "__xyz_user()"
most definitely isn't it.

An example of something that *should* be converted is
"csum_partial_copy_from_user()", but it really does need to use
"user_access_begin()" and friends, because right now it's using
stac/clac for each 16-bit word. That's *expensive*, and it's expensive
whether you use __get_user() or get_user().

Adding x86 people just to see how they react to the above "patch".

For me, in my fairly minimal personal config, the above cleanup patch
shrinks the text size of the resulting kernel by 1.7kB. Yes, most of
it is the out-of-line code, but still..

Interestingly, the signal handling code didn't change at all. It looks
like only the 32-bit code uses __put_user() for the frame setup. The
regular code uses put_user_try/put_user_catch, which is the
x86-specific early try at the unsafe versions, but it would actually
be improved by using "unsafe_put_user()" and my patch to make that use
"asm goto".

                           Linus

PS. That "patch" depends on modern git - with older versions of git
you need to do the path negation with ":!pattern", and then you need
to quote it too since '!' is special for shell.