Re: [PATCH 03/11] fs: add new read_uptr and write_uptr file operations

[Christoph Hellwig <hch@lst.de>]

On Mon, Jun 29, 2020 at 10:02:48AM -0700, Linus Torvalds wrote:
> That said, is there no practical limit on how big "optlen" can be?

There are some pretty huge ones, like the sctp one that can take
a basically unlimited list of sockaddr structures.

> Sure, I realize that a lot of setsockopt users may not use all of the
> data, but let's say that "optlen" is 128, but the actual low-level
> setsockopt operation only uses the first 16 bytes, maybe we could
> always just copy the 128 bytes from user space into kernel space, and
> just say "setsockopt() always gets a kernel pointer".

One issue is that a lot setsockopt calls are in the fast path, and
even have micro-optimizations like putting an int on stack for the
fast path to avoid the memory allocation.  While I don't know for
sure I fear that always doing a large allocation could end up having
a performance impact.  But otherwise I like that idea, and did in
fact start some prep work until I realized what I did was futile.

> Then the bpf use is even simpler. It would just pass the kernel
> pointer natively.
> 
> Because that seems to be what the BPF code really wants to do: it
> takes the user optval, and munges it into a kernel optval, and then
> (if that has been done) runs the low-level sock_setsockopt() under
> KERNEL_DS.
> 
> Couldn't we switch things around instead, and just *always* copy
> things from user space, and sock_setsockopt (and
> sock->ops->setsockopt) _always_ get a kernel buffer?
> 
> And avoid the set_fs(KERNEL_DS) games entirely that way?

I'd love to be able to do that.  And now that we want through this
whole mess than Nth time I have another idea:

 - we assume optlen is correct, which should cover about 90% of
   the protocols
 - but to override that a new setsockopt_len method is added that
   returns the correct length for all the messy ones.

Let me try if that works out.