[PATCH v12 01/12] lib: introduce copy_struct_{to,from}_user helpers

[Peter Zijlstra]

On Thu, Sep 05, 2019 at 07:26:22PM +1000, Aleksa Sarai wrote:
> On 2019-09-05, Peter Zijlstra <peterz at infradead.org> wrote:
> > On Thu, Sep 05, 2019 at 06:19:22AM +1000, Aleksa Sarai wrote:
> > > +/**
> > > + * copy_struct_to_user: copy a struct to user space
> > > + * @dst:   Destination address, in user space.
> > > + * @usize: Size of @dst struct.
> > > + * @src:   Source address, in kernel space.
> > > + * @ksize: Size of @src struct.
> > > + *
> > > + * Copies a struct from kernel space to user space, in a way that guarantees
> > > + * backwards-compatibility for struct syscall arguments (as long as future
> > > + * struct extensions are made such that all new fields are *appended* to the
> > > + * old struct, and zeroed-out new fields have the same meaning as the old
> > > + * struct).
> > > + *
> > > + * @ksize is just sizeof(*dst), and @usize should've been passed by user space.
> > > + * The recommended usage is something like the following:
> > > + *
> > > + *   SYSCALL_DEFINE2(foobar, struct foo __user *, uarg, size_t, usize)
> > > + *   {
> > > + *      int err;
> > > + *      struct foo karg = {};
> > > + *
> > > + *      // do something with karg
> > > + *
> > > + *      err = copy_struct_to_user(uarg, usize, &karg, sizeof(karg));
> > > + *      if (err)
> > > + *        return err;
> > > + *
> > > + *      // ...
> > > + *   }
> > > + *
> > > + * There are three cases to consider:
> > > + *  * If @usize == @ksize, then it's copied verbatim.
> > > + *  * If @usize < @ksize, then kernel space is "returning" a newer struct to an
> > > + *    older user space. In order to avoid user space getting incomplete
> > > + *    information (new fields might be important), all trailing bytes in @src
> > > + *    (@ksize - @usize) must be zerored
> > 
> > s/zerored/zero/, right?
> 
> It should've been "zeroed".

That reads wrong to me; that way it reads like this function must take
that action and zero out the 'rest'; which is just wrong.

This function must verify those bytes are zero, not make them zero.

> > >                                          , otherwise -EFBIG is returned.
> > 
> > 'Funny' that, copy_struct_from_user() below seems to use E2BIG.
> 
> This is a copy of the semantics that sched_[sg]etattr(2) uses -- E2BIG for
> a "too big" struct passed to the kernel, and EFBIG for a "too big"
> struct passed to user-space. I would personally have preferred EMSGSIZE
> instead of EFBIG, but felt using the existing error codes would be less
> confusing.

Sadly a recent commit:

  1251201c0d34 ("sched/core: Fix uclamp ABI bug, clean up and robustify sched_read_attr() ABI logic and code")

Made the situation even 'worse'.


> > > +	if (unlikely(!access_ok(src, usize)))
> > > +		return -EFAULT;
> > > +
> > > +	/* Deal with trailing bytes. */
> > > +	if (usize < ksize)
> > > +		memset(dst + size, 0, rest);
> > > +	else if (usize > ksize) {
> > > +		const void __user *addr = src + size;
> > > +		char buffer[BUFFER_SIZE] = {};
> > 
> > Isn't that too big for on-stack?
> 
> Is a 64-byte buffer too big? I picked the number "at random" to be the
> size of a cache line, but I could shrink it down to 32 bytes if the size
> is an issue (I wanted to avoid needless allocations -- hence it being
> on-stack).

Ah, my ctags gave me a definition of BUFFER_SIZE that was 512. I suppose
64 should be OK.

> > > +
> > > +		while (rest > 0) {
> > > +			size_t bufsize = min(rest, sizeof(buffer));
> > > +
> > > +			if (__copy_from_user(buffer, addr, bufsize))
> > > +				return -EFAULT;
> > > +			if (memchr_inv(buffer, 0, bufsize))
> > > +				return -E2BIG;
> > > +
> > > +			addr += bufsize;
> > > +			rest -= bufsize;
> > > +		}
> > 
> > The perf implementation uses get_user(); but if that is too slow, surely
> > we can do something with uaccess_try() here?
> 
> Is there a non-x86-specific way to do that (unless I'm mistaken only x86
> has uaccess_try() or the other *_try() wrappers)? The main "performance
> improvement" (if you can even call it that) is that we use memchr_inv()
> which finds non-matching characters more efficiently than just doing a
> loop.

Oh, you're right, that's x86 only :/