Re: [PATCH] gup: return -EFAULT on access_ok failure

[Chris Wilson <chris@chris-wilson.co.uk>]

Quoting Michael S. Tsirkin (2018-04-05 20:34:08)
> On Thu, Apr 05, 2018 at 11:43:27AM -0700, Linus Torvalds wrote:
> > On Thu, Apr 5, 2018 at 11:28 AM, Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > to repeat what you are saying IIUC __get_user_pages_fast returns 0 if it can't
> > > pin any pages and that is by design.  Returning 0 on error isn't usual I think
> > > so I guess this behaviour should we well documented.
> > 
> > Arguably it happens elsewhere too, and not just in the kernel.
> > "read()" at past the end of a file is not an error, you'll just get 0
> > for EOF.
> > 
> > So it's not really "returning 0 on error".
> > 
> > It really is simply returning the number of pages it got. End of
> > story. That number of pages can be smaller than the requested number
> > of pages, and _that_ is due to some error, but note how it can return
> > "5" on error too - you asked for 10 pages, but the error happened in
> > the middle!
> > 
> > So the right way to check for error is to bverify that you get the
> > number of pages that you asked for. If you don't, something bad
> > happened.
> > 
> > Of course, many users don't actually care about "I didn't get
> > everything". They only care about "did I get _something_". Then that 0
> > ends up being the error case, but note how it depends on the caller.
> > 
> > > What about get_user_pages_fast though?
> > 
> > We do seem to special-case the first page there. I'm not sure it's a
> > good idea. But like the __get_user_pages_fast(), we seem to have users
> > that know about the particular semantics and depend on it.
> > 
> > It's all ugly, I agree.
> > 
> > End result: we can't just change semantics of either of them.
> > 
> > At least not without going through every single user and checking that
> > they are ok with it.
> > 
> > Which I guess I could be ok with. Maybe changing the semantics of
> > __get_user_pages_fast() is acceptable, if you just change it
> > *everywhere* (which includes not just he users, but also the couple of
> > architecture-specific versions of that same function that we have.
> > 
> >                     Linus
> 
> OK I hope I understood what you are saying here.
> 
> At least drivers/gpu/drm/i915/i915_gem_userptr.c seems to
> get it wrong:
> 
>         pinned = __get_user_pages_fast(obj->userptr.ptr,
> 
>         if (pinned < 0) {
>                 pages = ERR_PTR(pinned);
>                 pinned = 0;
>         } else if (pinned < num_pages) {
>                 pages = __i915_gem_userptr_get_pages_schedule(obj);
>                 active = pages == ERR_PTR(-EAGAIN);
>         } else {
>                 pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
>                 active = !IS_ERR(pages);
>         }
> 
> The <0 path is never taken.

Please note that it only recently lost other paths that set an error
beforehand. Not exactly wrong since the short return is expected and
handled.

> Cc maintainers - should that driver be changed to use
> get_user_pages_fast?

It's not allowed to fault. __gup_fast has that comment, gup_fast does
not.
-Chris