read(2) hangs waiting for data from a closed socket

[Alex Riesen]

Hi,

I noticed that conky (http://conky.sourceforge.net/) freezes while reading
from hddtemp socket (it's an IPv4 TCP socket, loopback).
I didn't notice when exactly this started happening (I started running
full desktop on 2.6.35 only since -rc3) and haven't bisected yet (I'll try,
time and our daughter permitting).

Little what I have (2.6.35 crashes on me yet):

- hddtemp has closed this socket already (lsof -p `pidof hddtemp`
shows only its listening socket, no established connections). I can even
kill hddtemp altogether, I does not matter for conky's read

- conky waits in read(2) with this trace:
 conky         S ffff880033550348     0 19324      1 0x00000001
  ffff880074a85b08 0000000000000082 0000000000004000 ffff880074a85fd8
  0000000000013640 ffff880074a85fd8 ffff880033550000 0000000000013640
  0000000000013640 0000000000013640 0000000000013640 ffff880033550000
 Call Trace:
  [<ffffffff81465d81>] schedule_timeout+0x28/0x243
  [<ffffffff81028a12>] ? scale_rt_power+0x23/0x69
  [<ffffffff81034683>] ? load_balance+0x40a/0xf99
  [<ffffffff814676cb>] ? _raw_spin_unlock_bh+0xf/0x11
  [<ffffffff8136463e>] sk_wait_data+0x87/0xd1
  [<ffffffff81050e32>] ? autoremove_wake_function+0x0/0x38
  [<ffffffff813a793a>] tcp_recvmsg+0x431/0x88f
  [<ffffffff813c2b0b>] inet_recvmsg+0x6b/0x80
  [<ffffffff81360fe9>] sock_aio_read+0x118/0x12c
  [<ffffffff810cc722>] do_sync_read+0xc7/0x10d
  [<ffffffff8104930e>] ? ptrace_notify+0x95/0xb1
  [<ffffffff81161aea>] ? security_file_permission+0x11/0x13
  [<ffffffff810cd1a7>] vfs_read+0xbe/0x147
  [<ffffffff810cd2f4>] sys_read+0x47/0x6f
  [<ffffffff81002ba9>] tracesys+0xd9/0xde

(I have modified conky to use read, original code used recv,
doesn't matter in this case, though).

- It doesn't happen every time conky connects to hddtemp,
but surely happens if leave it running for five minutes. I could
reproduce the problem by running a simple program which
just connects to hddtemp, polls(2), reads and exits, in a tight
loop, so it is not something specific to conky. An while the conky's
code is not pretty it is certainly nothing unusual: basically
connect, select for incoming, blocking read if select returns
a ready socket, and close (see below).

- the system is a Dell M1330 laptop (Core 2 Duo, 64bit kernel
and userspace). Config attached.

The conky's code (conky_1.6.1-0ubuntu4):

char *get_hddtemp_info(char *dev, char *hostaddr, int port, char *unit)
{
	int sockfd = 0;
	struct hostent he, *he_res = 0;
	int he_errno;
	char hostbuff[2048];
	struct sockaddr_in addr;
	struct timeval tv;
	fd_set rfds;
	int len, i, devlen = strlen(dev);
	char sep;
	char *p, *out, *r = NULL;

	if ((sockfd = socket(AF_INET, SOCK_STREAM, 0)) == -1) {
		perror("socket");
		return NULL;
	}

	do {
#ifdef HAVE_GETHOSTBYNAME_R
		if (gethostbyname_r(hostaddr, &he, hostbuff, sizeof(hostbuff),
&he_res, &he_errno)) {	// get the host info
			ERR("hddtemp gethostbyname_r: %s", hstrerror(h_errno));
			break;
		}
#else /* HAVE_GETHOSTBYNAME_R */
		he_res = gethostbyname(hostaddr);
		if (!he_res) {
			perror("gethostbyname");
			break;
		}
#endif /* HAVE_GETHOSTBYNAME_R */

		addr.sin_family = AF_INET;
		addr.sin_port = htons(port);
		addr.sin_addr = *((struct in_addr *) he_res->h_addr);
		memset(&(addr.sin_zero), 0, 8);

		if (connect(sockfd, (struct sockaddr *) &addr,
					sizeof(struct sockaddr)) == -1) {
			perror("connect");
			break;
		}

		FD_ZERO(&rfds);
		FD_SET(sockfd, &rfds);

		/* We're going to wait up to a quarter a second to see whether there's
		 * any data available. Polling with timeout set to 0 doesn't seem to work
		 * with hddtemp. */
		tv.tv_sec = 0;
		tv.tv_usec = 250000;

		i = select(sockfd + 1, &rfds, NULL, NULL, &tv);
		if (i == -1) {
			if (errno == EINTR) {	/* silently ignore interrupted system call */
				break;
			} else {
				perror("select");
			}
		}

		/* No data available */
		if (i <= 0) {
			break;
		}

		p = buf;
		len = 0;
		do {
			i = read(sockfd, p, BUFLEN - (p - buf));
			if (i < 0) {
				perror("read");
				break;
			}
			len += i;
			p += i;
		} while (i > 0 && p < buf + BUFLEN - 1);

		if (len < 2) {
			break;
		}

		buf[len] = 0;

		/* The first character read is the separator. */
		sep = buf[0];
		p = buf + 1;

		while (*p) {
			if (!strncmp(p, dev, devlen)) {
				p += devlen + 1;
				p = strchr(p, sep);
				if (!p) {
					break;
				}
				p++;
				out = p;
				p = strchr(p, sep);
				if (!p) {
					break;
				}
				*p = '\0';
				p++;
				*unit = *p;
				if (!strncmp(out, "NA", 2)) {
					strncpy(buf, "N/A", BUFLEN);
					r = buf;
				} else {
					r = out;
				}
				break;
			} else {
				for (i = 0; i < 5; i++) {
					p = strchr(p, sep);
					if (!p) {
						break;
					}
					p++;
				}
				if (!p && i < 5) {
					break;
				}
			}
		}
	} while (0);
	close(sockfd);
	return r;
}

Attachment: 2.6.35-rc3.config
Description: Binary data