Kill off use of etherboot.h outside drivers/net.

safe dropping of the netdev ref by the driver while other refs still
exist.

Add netdev_irq() method.  Net device open()/close() methods should no
longer enable or disable IRQs.

Remove rx_quota; it wasn't used anywhere and added too much complexity
to implementing correct interrupt-masking behaviour in pxe_undi.c.

This reverts commit 3487640397.

This reverts commit 68add6e814.

This version uses the gPXE driver API rather than the legacy wrapper API.

Convert printf() to DBG(); printf() is not allowed in drivers.

structures.

DHCP still broken and #if 0'd out.

Achieved via Perl using:

perl -pi -e 's/pk_buff/io_buffer/g; s/Packet buffer/I\/O buffer/ig; ' \
	-e 's/pkbuff\.h/iobuf.h/g; s/pkb_/iob_/g; s/_pkb/_iob/g; ' \
	-e 's/pkb/iobuf/g; s/PKB/IOB/g;'

ISA 3c509 is currently non-functional, although the EISA (3c509-eisa) and
MCA (3c529) variants should build OK.

None of this code is yet tested.

quota is 1.

names.

Add "dev" pointer in struct net_device to tie network interfaces back to a
hardware device.

Force natural alignment of data types in __table() macros.  This seems to
prevent gcc from taking the unilateral decision to occasionally increase
their alignment (which screws up the table packing).

of alloc_pkb() exhaustion when e.g. an iSCSI-booted DOS session is left
idle for a long time at the C:\ prompt and builds up a huge packet
backlog.

of the packet buffer.  Although pkbuffs are guaranteed to be at least
PKB_ZLEN bytes long, there's no guarantee that the data starts at the
start of the packet buffer.

Since we have to align data for rtl8139 anyway, and since the start of
the packet buffer is guaranteed to be aligned to PKB_ALIGN, we may as
well just shuffle the packet down so that it starts at the start of
the packet buffer.  This simultaneously aligns the packet and ensures
that there is enough room for the zero-padding.