[Intel-wired-lan] [PATCH 0/7] [RESEND] [net] intel: Use smp_rmb rather than read_barrier_depends

[Duyck, Alexander H <alexander.h.duyck@intel.com>]

On Fri, 2017-11-17 at 10:16 -0600, Brian King wrote:
> On 11/16/2017 04:57 PM, Michael Ellerman wrote:
> > Brian King <brking@linux.vnet.ibm.com> writes:
> > 
> > > On 11/16/2017 01:33 PM, Jesse Brandeburg wrote:
> > > > On Thu, 16 Nov 2017 09:37:48 -0600
> > > > Brian King <brking@linux.vnet.ibm.com> wrote:
> > > > 
> > > > > Resending as the first attempt is not showing up in the list archive.
> > > > > 
> > > > > This patch converts several network drivers to use smp_rmb
> > > > > rather than read_barrier_depends. The initial issue was
> > > > > discovered with ixgbe on a Power machine which resulted
> > > > > in skb list corruption due to fetching a stale skb pointer.
> > > > > More details can be found in the ixgbe patch description.
> > > > 
> > > > Thanks for the fix Brian, I bet it was a tough debug.
> > > > 
> > > > The only users in the entire kernel of read_barrier_depends() (not
> > > > smp_read_barrier_depends) are the Intel network drivers.
> > > > 
> > > > Wouldn't it be better for power to just fix read_barrier_depends to do
> > > > the right thing on power? The question I'm not sure of the answer to is:
> > > > Is it really the wrong barrier to be using or is the implementation in
> > > > the kernel powerpc wrong?
> > > > 
> > > > So I think the right thing might actually to be to:
> > > > Fix arch powerpc read_barrier_depends to not be a noop, as the
> > > > semantics of the read_barrier_depends seems to be sufficient to solve
> > > > this problem, but it seems not to work for powerpc?
> > > 
> > > Jesse,
> > > 
> > > Thanks for the quick response.
> > > 
> > > Cc'ing linuxppc-dev as well. 
> > > 
> > > I did think about changing the powerpc definition of read_barrier_depends,
> > > but after reading up on that barrier, decided it was not the correct barrier
> > > to be used in this context. Here is some good historical background on
> > > read_barrier_depends that I found, along with an example.
> > > 
> > > https://lwn.net/Articles/5159/
> > > 
> > > Since there is no data-dependency in the code in question here, I think
> > > the smp_rmb is the proper barrier to use.
> > 
> > Yes I agree.
> > 
> > The read_barrier_depends() is correct to order the load of eop_desc and
> > then the dependent load of eop_desc->wb.status, but it's only required
> > or does anything on Alpha.
> > 
> > > For background, the code in question looks like this:
> > > 
> > > CPU 1                                   CPU2
> > > ============================            ============================
> > > 1: ixgbe_xmit_frame_ring                ixgbe_clean_tx_irq
> > > 2:  first->skb = skb                     eop_desc = tx_buffer->next_to_watch
> > >                                          if (!eop_desc)
> > >                                              break;
> > > 3:  ixgbe_tx_map                         read_barrier_depends()
> > >                                          if (!(eop_desc->wb.status) ... )
> > >                                              break;
> > > 4:   wmb                                 
> > > 5:   first->next_to_watch = tx_desc      napi_consume_skb(tx_buffer->skb ..);
> > > 6:   writel(i, tx_ring->tail);
> > > 
> > > What we see on powerpc is that tx_buffer->skb on CPU2 is getting loaded
> > > prior to tx_buffer->next_to_watch. Changing the read_barrier_depends
> > > to a smp_rmb solves this and prevents us from dereferencing old pointer.
> > 
> > Right. Given that read_barrier_depends() is a nop, there's nothing there
> > to order the load of tx_buffer->skb vs anything else.
> > 
> > If it's actually the load of tx_buffer->skb that's the issue then the
> > smp_rmb() should really be immediately prior to that, rather than where
> > the read_barrier_depends() currently is.
> 
> Alex,
> 
> How would you like to proceed? read_barrier_depends is a noop on all archs
> except alpha and blackfin. On those two archs, read_barrier_depends and
> smp_rmb end up resulting in the same code. So, I can either:
> 
> 1. Remove the setting of tx_buffer->skb to NULL to address your concern and proceed
> with the rest of the patch set unchanged.

I am good with this option. We just need to be certain that it solves
the original issue you saw.

> 2. Leave the read_barrier_depends, as it is the right barrier to order the load
> of eop_desc with respect to eop_desc->wb.status, and then *add* an smp_rmb in
> the same code path to address the speculative load of the skb that I was running into.
> This is arguably more pure from the perspective of the use of the different
> barriers, but has the downside of additional overhead on alpha and blackfin.
> 
> Do you have a preference? 

If you have the smp_rmb there is no need for the read_barrier_depends
as having both barriers would be redundant anyway. It was there as more
of a mental place holder than anything else since I suspect these
drivers would never be run on an alpha architecture anyway.

> Thanks,
> 
> Brian

Thanks for finding this issue and taking the time to resolve it.

- Alex