Here is version 6 of my patchset to add vlan support to the emulated rtl8139
nic.
Changes since v5:
* moved all receive changes to "add vlan tag extraction"
* fixed checkpatch.pl style issues
* fixed bugs in receive case related to small buffers and loopback
mode. Moved "too small buffer" code back where it used to be, though
it is changed in content.
Changes since v4:
* removed alloca(), for real. Thanks to the reviewers for their
patience. This patchset now has more versions than the vlan header
has bytes!
* corrected the unlikely, debug printf and long lines, as per comments
* cleaned out ifdef's pertaining to ethernet checksum calculation.
According to a comment since removed they were related to an
"optimization":
> RTL8139 provides frame CRC with received packet, this feature
> seems to be ignored by most drivers, disabled by default
see commit ccf1d14
I've tested v5 using x86_64 host/guest with the usual procedure. I've also ran
the clang analyzer on the qemu code base, just for fun.
Changes since v3:
* removed alloca() and #include<net/ethernet.h> as per comments
* reordered patches to put extraction before insertion. Extraction
touches only the receive path but insertion touches both. The two
patches are now needed to have vlan functionnality.
I've tested v4 with x86_64 host/guest. I used the same testing procedure as
before. I've tested a plain configuration as well as one with tso + vlan
offload, successfully.
I had to hack around the Linux 8139cp driver to be able to enable tso on vlan
which leads me to wonder, can someone with access to the C+ spec or a real
card confirm that it can do tso and vlan offload at the same time? The patch
I used for the kernel is at https://gist.github.com/851895.
Changes since v2:
insertion:
* moved insertion later in the process, to handle tso
* use qemu_sendv_packet() to insert the tag for us
* added dot1q_buf parameter to rtl8139_do_receive() to avoid some
memcpy() in loopback mode. Note that the code path through that
function is unchanged when dot1q_buf is NULL.
extraction:
* reduced the amount of copying by moving the "frame too short" logic
after the removal of the vlan tag (as is done in e1000.c for
example). Unfortunately, that logic can no longer be shared betwen
C+ and C mode.
I've posted v2 of these patches back in November
http://article.gmane.org/gmane.comp.emulators.qemu/84252
I've tested v3 on the following combinations of guest and hosts:
host: x86_64, guest: x86_64
host: x86_64, guest: ppc32
host: ppc32, guest: ppc32
Testing on the x86_64 host used '-net tap' and consisted of:
* making an http transfert on the untagged interface.
* ping -s 0-1472 to another host on a vlan.
* making an scp upload to another host on a vlan.
Testing on the ppc32 host used '-net socket' connected to an x86_64 qemu-kvm
running the virtio nic and consisted of:
* establishing an ssh connection between the two using an untagged interface.
* ping -s 0-1472 between the two using a vlan.
* making an scp transfer in both directions using a vlan.
All that was successful. Nevertheless, it doesn't exercise all code paths so
care is in order.
Please note that the lack of vlan support in rtl8139 has taken a few people
aback:
https://bugzilla.redhat.com/show_bug.cgi?id=516587
http://article.gmane.org/gmane.linux.network.general/14266
Thanks,
-Ben