82574L hangs (with r233708 e1000 driver).

[John Baldwin]

On Tuesday, May 01, 2012 12:21:21 pm Konstantin Belousov wrote:
> On Thu, Apr 12, 2012 at 09:38:49PM +0300, Konstantin Belousov wrote:
> > On Mon, Apr 09, 2012 at 12:19:39PM -0400, John Baldwin wrote:
> > > On Sunday, April 08, 2012 1:11:25 am Konstantin Belousov wrote:
> > > > On Sat, Apr 07, 2012 at 04:22:07PM -0700, Jack Vogel wrote:
> > > > > Make sure you have any firmware up to the latest available, if that 
doesn't
> > > > > help
> > > > > let me know and I'll check internally to see if there are any 
outstanding
> > > > > issues
> > > > > in shared code,  that will be after the weekend.
> > > > 
> > > > I had BIOS rev. 151, after you hint I found rev. 154 on the site.
> > > > Now BIOS reports itself as MTCDT10N.86A.0154.2012.0323.1601,
> > > > March 23.
> > > > 
> > > > Unfortunately, upgrade did not changed anything in regard of hanging
> > > > interface.
> > > 
> > > Does reverting 233708 make any difference?  Have you tried futzing 
around with
> > > kgdb when it is hung to see what state the device is in (software state 
at
> > > least)?
> > It does, in a sense that without r233708 the interface becomes stuck
> > almost immediately. I just upgraded to the e1000 at r234154, which does not
> > change much.
> > 
> > I fiddled with the adapter state after the hang in kgdb more, and I
> > noted something interesting. Apparently, tx works. When I ping the remote
> > host from my suffering atom machine, remote host sees the packet. Also
> > remote machine sees some udp traffic originating from the tom, like
> > ntp queries.
> > 
> > And, on receive, the atom board does receive interrupts, em0:rx 0 counter
> > in vmstat -i increases. Even more fun, the sysctl dev.em.0.debug
> > shows increasing hw rdh (as I understand, this is hardware 'last
> > received' packet pointer for rx ring). So I looked at the packet
> > descriptor at hw rdt index, and there I see
> > (kgdb) p/x ((struct adapter *)0xffffff80010e4000)->rx_rings->rx_base[78]
> > $11 = {buffer_addr = 0x12a128800, length = 0x5ea, csum = 0x3c2b, status = 
0x0, 
> >   errors = 0x0, special = 0x0}
> > 
> > Apparently, the Descriptor Done bit is clear, so the em_rxeof() function
> > breaks from the loop, not consuming the current packet. Also, it returns
> > false due to DD bit clear. This prevents em_msix_rx() from scheduling
> > taskqueue for processing. So apparent cause for the hang is missing
> > DD bit in descriptor.
> > 
> > I am not sure isn't all this is obvious for anybody who knows em
> > internals, and were to go from there.
> 
> Ok, nobody cares.
> 
> Below is the workaround I use to prevent the interface wedging.
> It seems that the sole PCI register read (namely, the rx ring head read)
> and consequent recheck of the descriptor status greatly reduce the
> likelihood of the issue. Unfortunately, the read does not eliminate
> the hang completely. So it is not some PCIe coherency problem.
> 
> With the patch applied, I am able to copy around blu-ray images, while
> previously the interface hang in 20-30 seconds of 100Mbit/s traffic.
> Sometimes the messages are printed:
> em0: Workaround: head 1018 tail 1002 cur 1010
> em0: Workaround: head 976 tail 973 cur 974
> em0: Workaround: head 950 tail 939 cur 946
> em0: Workaround: head 435 tail 419 cur 426
> 
> Machine is still dead due to random memory corruption which I see, in
> particular, pmap sometimes read garbage from PTEs. I have no idea is
> it related to em0 rx descriptor missed writes, or is a different issue.

Humm, so if I'm reading this correctly, the card "skips" a receive
descriptor and stores a packet at the next descriptor?  That's just
bizarre.

-- 
John Baldwin