Re: AW: [PATCH] amd64: Fix csum_partial_copy_generic()

[Eric Dumazet]

On Thu, Oct 19, 2023 at 8:39 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> On Thu, Oct 19, 2023 at 07:14:27AM +0100, Al Viro wrote:
> > On Thu, Oct 19, 2023 at 06:02:50AM +0100, Al Viro wrote:
> > > On Thu, Oct 19, 2023 at 04:44:04AM +0000, gus Gusenleitner Klaus wrote:
> > > > > On Wed, Oct 18, 2023 at 06:18:05AM +0000, gus Gusenleitner Klaus wrote:
> > > > > > The checksum calculation is wrong in case of an source buffer
> > > > > > containing zero bytes only. The expected return value is 0, the
> > > > > > actual return value is 0xfffffff.
> > > > >
> > > > > Expected where?  The actual checksum is defined modulo 0xffff, so
> > > > > 0 and 0xffffffff represent the same final value.
> > > > >
> > > > > The only twist is that in some situations we internally use 0 for
> > > > > "not calculated yet".
> > > > >
> > > > > > This problem occurs when a ICMP echo reply is sent that has set
> > > > > > zero identifier, sequence number and data.
> > > > >
> > > > > What problem?  Could you please either point to specific RFC or
> > > > > show that packets are rejected by some existing system, or...?
> > > >
> > > > Here's our situation:
> > > > Our device gets pinged by a third party manufacturer robot controller.
> > > > We have updated the kernel in our device to 5.15 from 4.9, the robot
> > > > controller is kept unchanged. At 4.9, our device's ping reply is accepted
> > > > by the robot controller, at 5.15 it's not.
> > > >
> > > > Wireshark shows a bad checksum warning:
> > > >  'Checksum: 0x0000 incorrect, should be 0xffff'
> > > >
> > >
> > > Lovely.  I think I see what's going on, give me a few to think about it...
> >
> > The real source of trouble was switching csum_and_copy_{to,from}_user()
> > to reporting faults as 0.  And yes, it's broken.  Bugger...
>
> I really hate the idea of bringing back the old horrors and splitting
> _nocheck and _user variants ;-/  Especially since we don't care (and
> never had, really) where in the EFAULT case had the damn thing faulted
> and what csum had it managed to accumulate prior to that point.
>
> The only callers are csum_and_copy_..._iter() and they discard
> the entire iovec segment if fault happens; all users of
> csum_and_copy_from_iter() are actually discarding everything in
> that case (reverting iterator to the point where it had been
> prior to the call).
>
> And they are all thread-synchronous.  Hell, it's tempting to steal
> a thread flag, clear it before the call of those suckers, set it in
> exception handlers in those and check in csum_and_copy_..._iter()
> after the call...  Let me see how ugly something of that sort would
> be...

I wonder if the csum_and_copy_...() helpers are really needed in modern days,
with much bigger cpu caches.

Maybe we could remove them and use more standard copy + standard
checksum over kernel buffers.