huge memory leak

[Andy Fleming <afleming@motorola.com>]

I'm working on an ethernet driver for an integrated processor (the 
MPC8540/8560).  The source is currently available in 
drivers/net/gianfar.c in Linux 2.4.  Currently, when I set the number 
of transmit buffer descriptors to 16 or lower (using ethtool -G eth1 tx 
16), and run a sufficiently packet intensive test (ping -f does not 
suffice, but ftp does), memory starts leaking like crazy (about 1-2 MB 
per second).  If I don't stop the transfer within seconds, all 128 MB 
of my memory is taken.

I'm having a great deal of trouble tracking this down.  It appears that 
the descriptor ring is filling up, and I am telling the queue to halt 
by calling netif_stop_queue().  I have scattered printks about, and it 
does not appear that the driver is having to drop any packets 
(currently, it detects that the next hard_start_xmit() call would fail, 
and stops the queue if that is the case).  However, the counter I 
increment when the ring fills up only increments when this memory leak 
manifests, so I suspect that the ring filling up is connected somehow 
to the problem.  I'm not seeing anywhere in the kernel that could be 
losing so many packets, and I can't see anywhere in the driver that 
could, so I'm at a loss.  I'm hoping someone on this list might have a 
better intuition as to what is causing this bug.

Andy Fleming
Software Enablement
Motorola SPS

-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html