[ECOS] eth_drv_send() called recursively

[ECOS] eth_drv_send() called recursively

[Xiaochen Zhou]

Hi,everyone:

Here is ethernet packet sending flow in our ecos application:

1. eth_drv_send() call my_eth_driver_send()
2. my_eth_driver_send() call eth_drv_tx_done()
3. eth_drv_tx_done() call eth_drv_send()

eth_drv_send() and eth_drv_tx_done() are implemented in
io/eth/current/src/net/eth_drv.c

eth_drv_tx_done(struct eth_drv_sc *sc, CYG_ADDRESS key, int status)
{
    ....
    eth_drv_send(ifp);
    ....

eth_drv_send(struct ifnet *ifp)
{
    ....
    while ((sc->funs->can_send)(sc) > 0) {
        IF_DEQUEUE(&ifp->if_snd, m0);
        if (m0 == 0) {
            break;
        }
    ....
}

If sending queue has many mbufs (we using queue for QOS application),
sending thread will call eth_drv_send() recursively, the stack of packet
sending thread will overflow.

Can anyone tell me why call eth_drv_send() in eth_drv_tx_done()? In
eth_drv_send(), there are a loop to call IF_DEQUEUE(), so I think it is
useless and harmful. Can I delete "eth_drv_send(ifp)" line in
eth_drv_tx_done()?

Thanks!

Xiaochen Zhou

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] eth_drv_send() called recursively

[Nick Garnett]

Xiaochen Zhou <zhouxiaochen@h3c.com> writes:

> Hi,everyone:
> 
> Here is ethernet packet sending flow in our ecos application:
> 
> 1. eth_drv_send() call my_eth_driver_send()
> 2. my_eth_driver_send() call eth_drv_tx_done()
> 3. eth_drv_tx_done() call eth_drv_send()
> 
> eth_drv_send() and eth_drv_tx_done() are implemented in
> io/eth/current/src/net/eth_drv.c
> 
> eth_drv_tx_done(struct eth_drv_sc *sc, CYG_ADDRESS key, int status)
> {
>     ....
>     eth_drv_send(ifp);
>     ....
> 
> eth_drv_send(struct ifnet *ifp)
> {
>     ....
>     while ((sc->funs->can_send)(sc) > 0) {
>         IF_DEQUEUE(&ifp->if_snd, m0);
>         if (m0 == 0) {
>             break;
>         }
>     ....
> }
> 
> If sending queue has many mbufs (we using queue for QOS application),
> sending thread will call eth_drv_send() recursively, the stack of packet
> sending thread will overflow.
> 
> Can anyone tell me why call eth_drv_send() in eth_drv_tx_done()? In
> eth_drv_send(), there are a loop to call IF_DEQUEUE(), so I think it is
> useless and harmful. Can I delete "eth_drv_send(ifp)" line in
> eth_drv_tx_done()?

No!

In general eth_drv_tx_done() is not called from the driver's send
routine, it is usually called from the deliver/poll routine once the
packet has actually been transmitted. Usually in response to an
interrupt.

The only situation in which eth_drv_tx_done() may be called from the
send routine is if you have copied the packet out to an internal
buffer, in the driver or hardware. In this case the can_send routine
will fend off susequent send attempts until the packet is sent and a
buffer is available.

So I suspect that either your driver is calling eth_drv_tx_done() in
the wrong place, or your can_send routine is not working correctly.

Whatever the problem, deleting lines from generic code that has been
working corretly for many years is not the correct solution.


-- 
Nick Garnett                                     eCos Kernel Architect
eCosCentric Limited     http://www.eCosCentric.com/   The eCos experts
Barnwell House, Barnwell Drive, Cambridge, UK.    Tel: +44 1223 245571
Registered in England and Wales: Reg No 4422071.


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss