From mboxrd@z Thu Jan  1 00:00:00 1970
Message-ID: <2e44aa9ac8b270827a30c09349cc4a35@collyer.net>
Date: Tue, 17 Aug 2004 01:04:13 -0700
From: geoff@collyer.net
To: 9fans@cse.psu.edu, geoff@collyer.net
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Cc:
Subject: [9fans] booting & igbe
Topicbox-Message-UUID: d4acca6a-eacd-11e9-9e20-41e7f4b1d025

(9fans is constipated tonight, so my messages will probably all arrive
at once.)

The builtin-PXE booting on the Intel gigabit Ethernet cards is nice.
Unfortunately, it currently has to load 9pxeload in order to get
plan9.ini parsed and then load the kernel itself, which requires a
bootstrap igbe driver, which I think I have *this close* to working.
I know jmk's away until the end of the month; anybody else who's
familiar with the card (82540EM Intel PRO/1000 MT) who thinks this
sounds familiar, please let me know:

	9(pxe)load sends broadcast BOOTP requests, get a BOOTP reply,
	prints the IP & server name, sends a TFTP request, which is
	never seen by the server (verified with snoopy and prints in
	tftpd).  It's going at 1Gb/s, full-duplex through an SMC 8508T
	gigabit switch, which otherwise has been working fine.  The
	cpu server is also connected via igbe to the SMC switch at
	full duplex, 1Gb/s.  Sending multiple BOOTP requests works as
	expected, drawing multiple replies.  prints in the boot driver
	suggest that the packets are getting into the transmit ring
	buffer and that the card is draining the buffer.  The bits are
	set to pad short packets, which shouldn't be necessary anyway
	since we're running full duplex.  openbsd with an i82557 card
	can tftp from the same cpu server through the same SMC switch.
	Ether & IP addresses all look right.  I've run out of theories
	of what could be going wrong, short of nasty bugs or
	deliberate malice somewhere.

I'm trying to load fs kernels (i.e., small kernels that PXE can
handle) currently, so it's tempting to just stick 1K of (possibly
preprocessed plan9.ini) on the front of the kernel being loaded and
let PXE load the whole mess and avoid 9pxeload completely.