Re: pcap_read_linux_mmap is always blocking

[Dragos Ilie <dragos.ilie () gmail com>]

Guy Harris wrote:
> On Jul 6, 2009, at 9:39 AM, Dragos Ilie wrote:
>
> > I traced the issue to pcap-linux.c:pcap_read_linux_mmap(), which is the
> > read handler selected when HAVE_PACKET_RING is defined. The handler
> > calls poll() with a negative value, which means an infinite timeout.
> > This occurs when the libpcap variable md.timeout is equal to zero
> > (default value). If the caller specifies a positive md.timeout then
> > pcap_setnonblock_mmap() re-computes the timeout as shown below:
> >
> > p->md.timeout = p->md.timeout*-1 - 1;
> >
> > When the user-specified timeout is negative the expression above
> > becomes:
> >
> > p->md.timeout = (p->md.timeout+1)*-1;
>
> The user-specified timeout isn't supposed to be negative; the behavior
> when that is done has never been specified.  If your application is
> setting the timeout to a negative value, it should stop doing so.
>
> Internally, libpcap uses a negative timeout value in the pcap_t data
> structure to indicate non-blocking mode; if it's negative, poll()
> isn't called.  If, when pcap_setnonblock() is called, the
> p->md.timeout is:
>
>     zero, p->md.timeout becomes -1;
>
>     1 through INT_MAX, p->md.timeout becomes -2 through -INT_MAX-1;
>
> and that should be negative in all cases, so poll() shouldn't be called.
> -
> This is the tcpdump-workers list.
> Visit https://cod.sandelman.ca/ to unsubscribe.

My application is not setting the timeout.  This corresponds to the case
where p->md.timeout == 0. The function pcap_setnonblock_mmap first
checks the nonblock flag and if it is set *and* if p->md.timeout > 0 it
maps the value to p->md.timeout*-1 - 1

...
static int
pcap_setnonblock_mmap(pcap_t *p, int nonblock, char *errbuf)
{
        /* map each value to the corresponding 2's complement, to
         * preserve the timeout value provided with pcap_set_timeout */
        if (nonblock) {
                if (p->md.timeout > 0)
                        p->md.timeout = p->md.timeout*-1 - 1;
        } else
...


pcap_read_linux_mmap calls poll() unconditionally:

 ret = poll(&pollinfo, 1, (handle->md.timeout > 0)?
                                                handle->md.timeout: -1);

If md.timeout==0 (which is my case) the result is that the poll-timeout
is set to -1 (infinite). Am I missing something?

On a related note, if HAVE_PACKET_RING is not defined I get non-blocking
behavior, without any changes to my application.

-
This is the tcpdump-workers list.
Visit https://cod.sandelman.ca/ to unsubscribe.