Doorbell records are currently embedded within the completion queue
and receive work queue strucures, which are allocated using zalloc()
and so have an alignment guarantee of only sizeof(void*), i.e. four
bytes.  This is sufficient for the receive work queue, but not for the
completion queue, which requires an alignment guarantee of eight
bytes.

Though not guaranteed, it so happens that zalloc() will always return
a pointer that is exactly four bytes above a sixteen-byte boundary.
The completion queue doorbell record is therefore always misaligned,
and the value passed to the hardware via SW2HW_CQ is actually always
pointing to the page_offset value within the MTT descriptor (which
directly precedes the inline doorbell record).  Provided that the page
offset is greater than 0x100, this looks to the hardware like an
update_ci value of greater than 0x010000 (taking into account
endianness differences), and so the hardware will happily deliver more
than 0x010000 completions before stopping.  Hence this problem is
rarely observable.

Fix by allocating the doorbell records separately and using the
correct alignment constraints.

Signed-off-by: Michael Brown <mcb30@ipxe.org>