When the virtio net feature that allows merging of RX buffers
is enabled, buffers belonging to a merged set must be processed in
strict sequential order; to achieve this, the seqno field of the
vnet_rx structure is used to prevent different CPUS from processing
RX buffers simultaneously. However, part of the code that processes
the first buffer of a merged set is being executed outside the
critical section, which can cause corruption of the relevant pbuf
struct. This issue can manifest itself as failure to process
incoming network traffic.
Fix the above issue by introducing the vnet_rx_hdr_remove() helper
function, which is called to process both the first buffer of a
merged set (inside the critical section), and the only buffer of a
received packet when merging is not being done.

Closes #2099.

These macros are used when memory read/write ordering needs to be
enforced in order for different CPUs to have a consistent view of
shared memory. They can expand to an actual memory fence (on
architectures with a weakly ordered memory model), or to a simple
compiler barrier (on architectures with strong memory ordering).
The use of these macros instead of read_barrier() and
write_barrier() (which always insert a memory fence) creates more
efficient code for architectures (such as x86) with strong memory
ordering.
Replace calls to read/write_barrier() with calls to
smp_read/write_barrier() when memory is not used for device I/O;
remove barriers when not needed; in spin_unlock(), replace compiler
barrier with SMP write barrier, otherwise code is not safe with
weakly ordered memory model.

When the virtio net feature that allows merging of RX buffers is
enabled, the seqno field of the vnet_rx structure is used to
prevent different CPUS from processing simultaneously RX buffers
belonging to the same merged set.
Add memory barriers at the beginning and end of the critical
section, in order to prevent memory read/write reordering from
causing different CPUs to run parts of the critical section at the
same time.