Date: Wed, 10 Jan 2007 02:56:49 +0300 From: Oleg Bulyzhin <oleg@freebsd.org> To: John Polstra <jdp@polstra.com> Cc: Doug Barton <dougb@freebsd.org>, net@freebsd.org Subject: Re: [Fwd: Re: bge Ierr rate increase from 5.3R -> 6.1R] Message-ID: <20070109235649.GC5246@lath.rinet.ru> In-Reply-To: <XFMail.20070108163329.jdp@polstra.com> References: <20061231061903.J47883@delplex.bde.org> <XFMail.20070108163329.jdp@polstra.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Jan 08, 2007 at 04:33:29PM -0800, John Polstra wrote:
> On 30-Dec-2006 Bruce Evans wrote:
> > More debugging showed that almost any of the reads of the phy in mii
> > cause an input error,
>
> The errors appear to be caused by the code in bge_miibus_{read,write}reg
> that clears and then restores the BGE_MIMODE_AUTOPOLL bit of the
> BGE_MI_MODE register. If you remove those chunks of code, the errors
> go away even when mii_tick is called periodically, and the chip
> performs quite well under heavy traffic load in both directions.
I've seen this kind of ierrors (LINK_LOST error flag in rx buffer descriptor)
only on bcm5701 chip. Other chips i tried (5700 and 5721) seems are not
affected.
>
> I know that some of the *_MODE registers in these chips cause problems
> if they are read, so I tried putting a shadow copy of the register
> into the softc and doing only writes to the hardware register. But
> that didn't help in this case.
>
> I seem to recall having heard that Broadcom discourages the use of
> PHY autopolling for unspecified reasons. The Linux driver from
> Broadcom had code to support it, but it was never actually used for
> any devices. The current Linux driver uses autopolling only for the
> Altima chip that's present on the inexpensive Netgear GA302T card.
If i remember right, Broadcom recommends using phy interrupts for
tracing link events, though auto-polling noted as alternative method
for that.
>
> I agree with you that we shouldn't need to call mii_tick periodically,
> except perhaps for a few very old variants of the Broadcom chip.
>
> John
mii_tick periodic calls are requried - MII layer use it for link
autonegotiation. Though i guess it's possible to avoid phy access if we
have link up.
I see two different ways to solve this:
1) convert driver for using phy interrupts as default (like linux driver),
thus BGE_MI_MODE read/write can be avoided.
2) do not touch PHY inside bge_tick() if we have link up. Actually, this
check was in driver but was removed as workaround of existed link state
issues (due to some bugs in if_bge.c and brgphy.c link loss events were
detected with few seconds delay or even not detected at all).
Those bugs are fixed now, so we can revive this check (bge_link_upd() can
be simplified a bit too).
--
Oleg.
================================================================
=== Oleg Bulyzhin -- OBUL-RIPN -- OBUL-RIPE -- oleg@rinet.ru ===
================================================================
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20070109235649.GC5246>