[Xen-devel] Double gnttab_end_access in mini-os netfront

[Marek Marczykowski-GÃrecki <marmarek@xxxxxxxxxxxxxxxxxxxxxx>]

Hi,

When running HVM on Xen 4.6 with qemu in stubdom, I've found that
something goes wrong with disk frontend (at least that was visible
problem - a lot of read and write errors in stubdom log). But further
debugging (including -DGNT_DEBUG) leads to double gnttab_end_access in
netfront. 

The stack trace is:
ASSERTION FAILED: !(!inuse[ref]) at gnttab.c:42.
Do_exit called!
0x00000000000f3ffb: get_time_values_from_xen at 
xen-4.6.0/extras/mini-os/arch/x86/time.c:134 (discriminator 1)
0x00000000000d74a8: HYPERVISOR_sched_op at 
xen-4.6.0/extras/mini-os/include/x86/x86_64/hypercall-x86_64.h:180
0x00000000000d6a2e: put_free_entry at xen-4.6.0/extras/mini-os/gnttab.c:43
0x00000000000d6bff: gnttab_end_access at xen-4.6.0/extras/mini-os/gnttab.c:115
0x00000000000d8a50: network_rx at xen-4.6.0/extras/mini-os/netfront.c:134
0x00000000000d9ec4: netfront_receive at xen-4.6.0/extras/mini-os/netfront.c:671
0x00000000000dd302: get_current at 
xen-4.6.0/extras/mini-os/include/x86/arch_sched.h:16
0x0000000000079f72: tap_send at 
xen-4.6.0/stubdom/../tools/qemu-xen-traditional/net.c:756
0x00000000000069f9: main_loop_wait at 
xen-4.6.0/stubdom/../tools/qemu-xen-traditional/vl.c:3826
0x0000000000021f27: main_loop at 
xen-4.6.0/stubdom/../tools/qemu-xen-traditional/i386-dm/helper2.c:612 
(discriminator 1)
0x000000000000950d: quit_timers at 
xen-4.6.0/stubdom/../tools/qemu-xen-traditional/vl.c:1866
0x00000000000d7f57: call_main at xen-4.6.0/extras/mini-os/main.c:163
0x0000000000003423: thread_starter at :?
0x0000000000000000: _start at ??:?

It was working in Xen 4.4. The only commit touching xenfront.c (in any
meaningful way) from that time is this one:
http://xenbits.xen.org/gitweb/?p=mini-os.git;a=commit;h=7c8f348390652a67e9356eec9cd2b0f76a9c7c72

With that commit reverted, issue vanishes.

I guess it's because before this commit, there was "if (rx->status ==
NETIF_RSP_NULL) continue" before "gnttab_end_access(buf->gref)", but now
that case is handled after gnttab_end_access (using "if (rx->status >
NETIF_RSP_NULL)"). I think the fix would be to restore that "continue"
line.

PS What is the correct place for such reports? minios-devel? xen-devel?
both?

-- 
Best Regards,
Marek Marczykowski-GÃrecki
Invisible Things Lab
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?

Attachment: pgpaxexcotO8L.pgp
Description: PGP signature

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel