TCP issue with large packet under RTEMS

[Joel Sherrill]

On 02/09/2011 11:10 AM, João Rasta wrote:
> Hi Joel, Eric,
>
> I followed Eric's suggestion and it worked. I thought that read() 
> should be automatically blocking untill all numelements passed to the 
> read() function were read. I wonder why read() exits without this 
> condition to fulfill..
>
> Joel, how can i increase the TCP/IP stack size in order to get more 
> performance out of the read routine? I mean, using only one system 
> call to get the whole data would be perfect. Previously the 
> application did not lock, it simply returned the read with less bytes 
> than the buffer size and continued executing.
>
The number of reads required is not tuneable.  The issue is how much
data can the stack hold while the application hasn't read it yet.

It is in the networkconfig.h.  This is for the one for pc386/qemu:

#if   defined(MEMORY_CUSTOM)
     MEMORY_CUSTOM,
#elif defined(MEMORY_SCARCE)
     100*1024,                 /* MBUF space */
     200*1024,                 /* MBUF cluster space */
#elif defined(MEMORY_HUGE)
     2*1024*1024,              /* MBUF space */
     5*1024*1024,              /* MBUF cluster space */
#else
     180*1024,                 /* MBUF space */
     350*1024,                 /* MBUF cluster space */
#endif

It has to be enough where the application can get behind while
the driver is pumping data in.  I am guessing you would be
something like at least 512K/1M.
>
> Thanks a lot.
>
> Best,
> JM
>
> On Wed, Feb 9, 2011 at 4:55 PM, Eric Norum <wenorum at lbl.gov 
> <mailto:wenorum at lbl.gov>> wrote:
>
>     TCP (SOCK_STREAM) receivers always need to expect to have to
>     complete their reads in multiple operations -- you need to put
>     your read call in a loop:
>
>            char *buf;
>            ssize_t nleft, nread;
>            .
>            .
>            .
>            buf = .....
>            nleft = .....
>            while (nleft) {
>                    nread = read(sd, buf, nleft);
>                    if (nread < 0)
>                            error......
>                    if (nread == 0)
>                            end-of-file......
>                    buf += nread;
>                    nleft -= nread;
>            }
>
>     On Feb 9, 2011, at 8:36 AM, João Rasta wrote:
>
>     > Hi,
>     >
>     > I'm using RTEMS 4.10 and GRETH MAC in order to exchange data
>     with a client PC via TCP. I have a LEON-3 based processor
>     architecture. I can successfully attach the GRETH ethernet driver
>     using the RTEMS driver manager and exchange small ammounts of
>     data. However, when i need to transfer an image array
>     (436160Bytes) from the PC to the RTEMS system, the read() call
>     only returns 17508Bytes read instead of 436160Bytes .
>     >
>     > The send() directive in the PC side returns the correct number
>     of bytes sent so the problem must be on the RTEMS side. The read
>     buffer has the correct size.
>     >
>     > 1) Is there any programmable limitation related to the number of
>     bytes to receive? Since i'm using a TCP socket, there is no
>     apparent reason for byte loss i guess..
>     >
>     > 2) In what circumstances should read() from a TCP socket return
>     less bytes than the ones sent?
>     >
>     >
>     > Best,
>     > JM
>     >
>
>     --
>     Eric Norum
>     wenorum at lbl.gov <mailto:wenorum at lbl.gov>
>
>


-- 
Joel Sherrill, Ph.D.             Director of Research&  Development
joel.sherrill at OARcorp.com        On-Line Applications Research
Ask me about RTEMS: a free RTOS  Huntsville AL 35805
    Support Available             (256) 722-9985