socket.socket.settimeout implementation

[jnair]

When using socket.socket.settimeout we normally only guard against
"socket.timeout" exception.Now the implementation of "settimeout" in
"Python-2.4.3/Modules/socketmodule.c" sets the socket fd to
nonblocking and uses "select()" to timeout
as seen below in line 1487 and 1386 :

static PyObject *
1470 sock_settimeout(PySocketSockObject *s, PyObject *arg)
1471 {
1472 double timeout;
..............................
1485
1486 s->sock_timeout = timeout;
1487 internal_setblocking(s, timeout < 0.0);
1488
1489 Py_INCREF(Py_None);
1490 return Py_None;
1491 }
1492

1362 sock_accept(PySocketSockObject *s)
1363 {
...

1385 Py_BEGIN_ALLOW_THREADS
1386 timeout = internal_select(s, 0);
1387 if (!timeout)
1388 newfd = accept(s->sock_fd, (struct sockaddr *)&addrbuf,
1389 &addrlen);
1390 Py_END_ALLOW_THREADS
1391


Now internal_select() returns "1" on timemout else "0" as seen below
:

673 n = select(s->sock_fd+1, &fds, NULL, NULL,&tv);
674 if (n == 0)
675 return 1;
676 return 0;

back to the sock_accept() fuction on line 1387

1387 if (!timeout)
1388 newfd = accept(s->sock_fd, (struct sockaddr *) &addrbuf,
1389 &addrlen)


if "select()" returned before timeout we call "accept()"

Now my observation on FC4 ( kernel version 2.6.11-1.1369_FC4)
"select()" can return before timeout with a
"Interrupted system call" (EINTR) error,

my question is , is it possible that our internal_select() fuction as
seen on line 676 above returns "0" and we go on to do a "accept()" call
which returns with a "Resource temporarily unavailable" error if the
sockfd is not readable. So user expects only "socket.timeout" but gets
"Resource temporarily unavailable" which stops the program
from proceeding.

Our Case : We are using CherryPy , which sets a timeout on socket using
"settimeout" and whenever our code uses "os.seteuid" cherrpy dies
giving
out the "Resource temporarily unavailable".

I did write a "C" program using pthreads , one thread doing
non blocking accept() and select() and other thread doing a seteuid()
..
select() did bail out giving Interrupted system call Error.

Thanks
Jitendra Nair

 

[Lawrence D'Oliveiro]

Now my observation on FC4 ( kernel version 2.6.11-1.1369_FC4)
"select()" can return before timeout with a
"Interrupted system call" (EINTR) error,

Nothing Red-Hat-specific, or even Linux-specific, about this. It's a
standard *nix thing. _All_ potentially time-consuming system calls (such
as select(2), read(2), write(2)) can return EINTR, or some other
indication that they didn't do everything they were asked to do.

my question is , is it possible that our internal_select() fuction as
seen on line 676 above returns "0" and we go on to do a "accept()" call
which returns with a "Resource temporarily unavailable" error if the
sockfd is not readable.

The semantics of select(2) is that you're supposed to check the
returrned bitmasks and only take the appropriate actions if the
corresponding read-ready, write-ready and error bits are set.

See the select(2) man page for more on this.