[ECOS] pbuf_alloc failures with LwIP

[Michael O'Dowd]

Hi Elad,

I ran into a similar problem recently. I'm using a recent CVS checkout 
rather than 3.0. Also, I'm probably not using the same ethernet HW, so I 
don't know how well my reply corresponds to your case.

The eth_drv.c file is the glue between lwIP and the underlying ethernet 
driver, so the issue that you are encountering may be specific to the 
driver. In my case, when under stress, eth_drv.c generates the error 
message: "cannot allocate pbuf to receive packet". Soon after that, the 
ethernet driver stops receiving traffic permanently, but does not crash. 
In your case, if I understand correctly, your system crashes.

The issue is that when eth_drv_recv() fails to allocate a pbuf, it 
returns without calling the ethernet driver recv() function: 
(sc->funs->recv)(). In my case, the driver requires that it's recv() 
function be called, in order to complete the processing of the packet 
reception and to free up the receive buffer(s). Failing to call it, 
apparently causes the receive path to cease functioning (I'm still 
investigating the details). In your case, I gather that it crashes the 
system.

Note: I'm running on an NXP 1788 (Cortex-M3), using the 
"devs/arm/lpc2xxx/current/src/if_lpc2xxx.c" ethernet driver.

There are two aspects to this problem:

1) In my opinion, there is a bug in eth_drv_recv(). If there are no 
pbufs available, then it should at least cause the received packet to be 
discarded. Otherwise, the system may fail whenever there is a minor 
burst of traffic on the network. It doesn't take much: there are only 16 
pbufs available by default. Whether or not the system fails, depends on 
how the ethernet driver reacts to the failure to call it's recv() 
function. I hope to fix this on my platform in the near future.

2) You should also keep an eye on your pbuf usage, just to make sure 
that you don't have a pbuf memroy leak. You could also try to allocate 
more pbufs, if you have the available memory.

If you are using the default lwip configuration, the pbuf memory 
allocation is handled by memp.[hc]. It has a fixed number of pbufs 
available. The default is 16 pbufs, and can be changed in the configtool 
under: [lwIP networking stack/Memory options/Number of memp struct pbufs].

Alternatively, if you have lots of memory, you could enable the 
checkbox: [lwIP networking stack/Memory options/Use malloc for pool 
allocations]. This bypasses the memp pools and their static limitations. 
Though this will make it harder to spot a pbuf memory leak. I haven't 
tried this personally.

Finally, (when using memp) the pbuf usage can be monitored with 
lwip/stats.h. If you have access to a serial port, try calling 
stats_display(). Here is a snippet of the pbuf related output:

 >  MEM PBUF_POOL
 >          avail: 16
 >          used: 0
 >          max: 3
 >          err: 0

The "err" counter increases when pbuf_alloc() fails.

Hope that helps,

Regards,

Michael O'Dowd
Kuantic SAS

On 12/06/2012 22:40, Elad Yosef wrote:
> Hi all,
> I'm using LwIP stack on my target and experiencing crashes under stress.
>
> function eth_drv_recv) from ../io/eth/v3_0/ser/lwip/eth_drv.c
> calls pbuf_alloc() and this allocation fails.
>
> Is this result of some bad configuration?
>
> Thanks
> Elad
>


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