[Top] [All Lists]

Re: zerocopy changes in 3c59x.c

To: Andrew Morton <andrewm@xxxxxxxxxx>
Subject: Re: zerocopy changes in 3c59x.c
From: "David S. Miller" <davem@xxxxxxxxxx>
Date: Fri, 26 Jan 2001 05:48:45 -0800 (PST)
Cc: Alexey Kuznetsov <kuznet@xxxxxxxxxxxxx>, "netdev@xxxxxxxxxxx" <netdev@xxxxxxxxxxx>
In-reply-to: <3A718110.5C540AE@xxxxxxxxxx>
References: <3A717B44.82C66152@xxxxxxxxxx> <3A716447.CF6E8BB0@xxxxxxxxxx> <3A714788.82C064BD@xxxxxxxxxx> <14961.24733.869800.77633@xxxxxxxxxxxxxxx> <14961.30181.671982.174763@xxxxxxxxxxxxxxx> <14961.31482.977020.162767@xxxxxxxxxxxxxxx> <3A718110.5C540AE@xxxxxxxxxx>
Sender: owner-netdev@xxxxxxxxxxx
Andrew Morton writes:
 > And just to clarify: it is currently the case that we
 > do support scatter/gather on devices which don't have 
 > hardware checksums on transmit.

Yes, but I won't allow this for ipv4/ipv6 in the final zerocopy patch
I send to Linus.  The reason (did you attend my talk at UNSW last
week??? :-))) is that if SG-only is allowed it is possible for the
paged data to get modified between when we calculate the checksum
in software and the device transmits the packet.  We don't get
exclusive access to the pages, we just grab a reference to them.
Even though a TCP retransmit would fixup the checksum this is horrible
from a quality of implementation standpoint.

However, as I mentioned, DecNET and friends will be able to make use
of SG-only devices.

David S. Miller

<Prev in Thread] Current Thread [Next in Thread>