io/socket: address review of the poll() conversion

Follow-up to the FD_SETSIZE fix, covering the points raised in review.

Negative/overflowing I/O timeouts.  set_io_timeout() could produce a negative
select_timeout (a peer-supplied MSG_IO_TIMEOUT value was applied unchecked),
and every wait now passes select_timeout * 1000 to poll(), where a negative
millisecond count means "wait forever" -- so a hostile or buggy peer could
stall the other side and bypass keepalives entirely.  select() used to reject
that with EINVAL, which kept the loop and check_timeout() running.  Clamp a
negative argument to 0, compute allowed_lull without overflowing near INT_MAX
(secs / 2 + secs % 2), ignore a non-positive MSG_IO_TIMEOUT value, and funnel
all three waits through poll_timeout_ms(), which keeps the count positive and
bounded.

The daemon accept loop had the same fd_set overflow.  start_accept_loop() still
stored listening sockets in an fd_set, so a daemon started with enough
descriptors already open got listener fds >= FD_SETSIZE and hit the same
undefined behaviour at startup -- verified: with the old code a transfer
through such a daemon yields nothing, with this change it succeeds.  Converted
it to poll() as well.

Readiness testing.  Treating any non-zero revents as ordinary readiness was
wrong: poll() reports POLLERR/POLLHUP/POLLNVAL unrequested, and an invalid fd
shows up as POLLNVAL on a successful poll() rather than -1/EBADF, which left
the EBADF branches dead and let an invalid ff_forward_fd reach
forward_filesfrom_data() (where EBADF reads as EOF).  Use role-specific masks
(POLL_RD_BITS / POLL_WR_BITS), handle POLLNVAL explicitly in all three loops,
and request POLLPRI so select()'s old exception set is not silently dropped.

A bidirectional fd is no longer entered twice.  A direct daemon connection uses
one fd for both directions; it now occupies a single pollfd row with OR-ed
events instead of two rows carrying different masks, which also avoids the
Cygwin < 3.3.6 duplicate-entry readiness bug.

poll() is now a declared requirement: configure.ac checks for poll.h and
poll(), failing with a clear message rather than leaving it implicit.

The test no longer hardcodes FD_SETSIZE (1024 on glibc but 65536 on 64-bit
Solaris, where it would have opened too few fds and passed vacuously); it asks
the C library for the real value via a small compiled probe and skips if that
is unavailable.  Its description now also covers the fortified-libc case, where
the pre-fix result is an abort rather than a hang.

configure.ac

@@ -5,7 +5,7 @@ AC_INIT([rsync],[ ],[https://rsync.samba.org/bug-tracking.html])
 AC_C_BIGENDIAN
 AC_HEADER_DIRENT
 AC_HEADER_SYS_WAIT
-AC_CHECK_HEADERS(sys/fcntl.h sys/select.h fcntl.h sys/time.h sys/unistd.h \
+AC_CHECK_HEADERS(poll.h sys/fcntl.h sys/select.h fcntl.h sys/time.h sys/unistd.h \
     unistd.h utime.h compat.h sys/param.h ctype.h sys/wait.h sys/stat.h \
     sys/ioctl.h sys/filio.h string.h stdlib.h sys/socket.h sys/mode.h grp.h \
     sys/un.h sys/attr.h arpa/inet.h arpa/nameser.h locale.h sys/types.h \
@@ -909,6 +909,12 @@ AC_HAVE_TYPE([struct stat64], [#include <stdio.h>
 
 # if we can't find strcasecmp, look in -lresolv (for Unixware at least)
 #
+dnl rsync's I/O readiness loops use poll() rather than select() so that a
+dnl file descriptor at or above FD_SETSIZE cannot overflow an fd_set (which
+dnl is undefined behaviour and could hang the transfer).  poll() is in
+dnl POSIX.1-2001; fail early and clearly if this target lacks it.
+AC_CHECK_FUNCS([poll], , [AC_MSG_ERROR([rsync requires poll(); please report the platform to the rsync developers])])
+
 AC_CHECK_FUNCS(strcasecmp)
 if test x"$ac_cv_func_strcasecmp" = x"no"; then
     AC_CHECK_LIB(resolv, strcasecmp)

io.c

@@ -33,6 +33,12 @@
 
 #include <poll.h>
 
+/* Readiness bits we act on.  poll() can report POLLERR/POLLHUP/POLLNVAL even
+ * when they were not requested, and POLLPRI stands in for select()'s old
+ * exception set. */
+#define POLL_RD_BITS (POLLIN | POLLPRI | POLLERR | POLLHUP)
+#define POLL_WR_BITS (POLLOUT | POLLERR | POLLHUP)
+
 /** If no timeout is specified then use a 60 second I/O timeout */
 #define SELECT_TIMEOUT 60
 
@@ -115,6 +121,19 @@ static xbuf ff_xb = EMPTY_XBUF;
 static xbuf iconv_buf = EMPTY_XBUF;
 #endif
 static int select_timeout = SELECT_TIMEOUT;
+
+/* Turn select_timeout (in seconds) into a poll() millisecond count, keeping it
+ * positive and bounded.  A negative count means "wait forever" to poll(), which
+ * would bypass our keepalives and timeout enforcement entirely. */
+static int poll_timeout_ms(void)
+{
+	int secs = select_timeout;
+
+	if (secs <= 0 || secs > SELECT_TIMEOUT)
+		secs = SELECT_TIMEOUT;
+	return secs * 1000;
+}
+
 static int active_filecnt = 0;
 static OFF_T active_bytecnt = 0;
 static int first_message = 1;
@@ -251,12 +270,12 @@ static size_t safe_read(int fd, char *buf, size_t len)
 		/* We use poll() rather than select() so that a high-numbered fd
 		 * (>= FD_SETSIZE) cannot overflow an fd_set bitmap. */
 		pfd.fd = fd;
-		pfd.events = POLLIN;
+		pfd.events = POLLIN | POLLPRI;
 		pfd.revents = 0;
 
-		cnt = poll(&pfd, 1, select_timeout * 1000);
+		cnt = poll(&pfd, 1, poll_timeout_ms());
 		if (cnt <= 0) {
-			if (cnt < 0 && errno == EBADF) {
+			if (cnt < 0 && errno != EINTR && errno != EAGAIN) {
 				rsyserr(FERROR, errno, "safe_read poll failed");
 				exit_cleanup(RERR_FILEIO);
 			}
@@ -264,7 +283,13 @@ static size_t safe_read(int fd, char *buf, size_t len)
 			continue;
 		}
 
-		if (pfd.revents) {
+		/* An invalid fd is reported here rather than via poll()'s return. */
+		if (pfd.revents & POLLNVAL) {
+			rsyserr(FERROR, EBADF, "safe_read poll failed");
+			exit_cleanup(RERR_FILEIO);
+		}
+
+		if (pfd.revents & POLL_RD_BITS) {
 			ssize_t n = read(fd, buf + got, len - got);
 			if (DEBUG_GTE(IO, 2)) {
 				rprintf(FINFO, "[%s] safe_read(%d)=%" SIZE_T_FMT_MOD "d\n",
@@ -337,9 +362,9 @@ static void safe_write(int fd, const char *buf, size_t len)
 		pfd.events = POLLOUT;
 		pfd.revents = 0;
 
-		cnt = poll(&pfd, 1, select_timeout * 1000);
+		cnt = poll(&pfd, 1, poll_timeout_ms());
 		if (cnt <= 0) {
-			if (cnt < 0 && errno == EBADF) {
+			if (cnt < 0 && errno != EINTR && errno != EAGAIN) {
 				rsyserr(FERROR, errno, "safe_write poll failed on %s", what_fd_is(fd));
 				exit_cleanup(RERR_FILEIO);
 			}
@@ -348,7 +373,12 @@ static void safe_write(int fd, const char *buf, size_t len)
 			continue;
 		}
 
-		if (pfd.revents) {
+		if (pfd.revents & POLLNVAL) {
+			rsyserr(FERROR, EBADF, "safe_write poll failed on %s", what_fd_is(fd));
+			exit_cleanup(RERR_FILEIO);
+		}
+
+		if (pfd.revents & POLL_WR_BITS) {
 			n = write(fd, buf, len);
 			if (n < 0) {
 				if (errno == EINTR)
@@ -657,7 +687,7 @@ static char *perform_io(size_t needed, int flags)
 		if (iobuf.in_fd >= 0 && iobuf.in.size - iobuf.in.len) {
 			if (!read_batch || batch_fd >= 0) {
 				pfds[npfds].fd = iobuf.in_fd;
-				pfds[npfds].events = POLLIN;
+				pfds[npfds].events = POLLIN | POLLPRI;
 				pfds[npfds].revents = 0;
 				in_pollpos = npfds++;
 			}
@@ -714,10 +744,18 @@ static char *perform_io(size_t needed, int flags)
 			} else
 				out = NULL;
 			if (out) {
-				pfds[npfds].fd = iobuf.out_fd;
-				pfds[npfds].events = POLLOUT;
-				pfds[npfds].revents = 0;
-				out_pollpos = npfds++;
+				/* A direct daemon connection uses one fd for both
+				 * directions; give it a single row with both events
+				 * rather than two rows carrying different masks. */
+				if (in_pollpos >= 0 && iobuf.out_fd == iobuf.in_fd) {
+					pfds[in_pollpos].events |= POLLOUT;
+					out_pollpos = in_pollpos;
+				} else {
+					pfds[npfds].fd = iobuf.out_fd;
+					pfds[npfds].events = POLLOUT;
+					pfds[npfds].revents = 0;
+					out_pollpos = npfds++;
+				}
 				if (iobuf.out_fd > max_fd)
 					max_fd = iobuf.out_fd;
 			}
@@ -757,10 +795,10 @@ static char *perform_io(size_t needed, int flags)
 				poll_timeout = 0;
 			else {
 				extra_flist_sending_enabled = False;
-				poll_timeout = select_timeout * 1000;
+				poll_timeout = poll_timeout_ms();
 			}
 		} else
-			poll_timeout = select_timeout * 1000;
+			poll_timeout = poll_timeout_ms();
 
 		cnt = poll(pfds, npfds, poll_timeout);
 
@@ -784,7 +822,20 @@ static char *perform_io(size_t needed, int flags)
 				pfds[out_pollpos].revents = 0;
 		}
 
-		if (iobuf.in_fd >= 0 && in_pollpos >= 0 && pfds[in_pollpos].revents) {
+		if (cnt > 0) {
+			/* poll() reports a bad fd here, not via its return value. */
+			int p;
+			for (p = 0; p < npfds; p++) {
+				if (pfds[p].revents & POLLNVAL) {
+					msgs2stderr = 1;
+					rsyserr(FERROR, EBADF, "perform_io: poll reported an invalid fd");
+					exit_cleanup(RERR_SOCKETIO);
+				}
+			}
+		}
+
+		if (iobuf.in_fd >= 0 && in_pollpos >= 0
+		 && pfds[in_pollpos].revents & POLL_RD_BITS) {
 			size_t len, pos = iobuf.in.pos + iobuf.in.len;
 			ssize_t n;
 			if (pos >= iobuf.in.size) {
@@ -833,7 +884,7 @@ static char *perform_io(size_t needed, int flags)
 			exit_cleanup(RERR_TIMEOUT);
 		}
 
-		if (out && out_pollpos >= 0 && pfds[out_pollpos].revents) {
+		if (out && out_pollpos >= 0 && pfds[out_pollpos].revents & POLL_WR_BITS) {
 			size_t len = iobuf.raw_flushing_ends_before ? iobuf.raw_flushing_ends_before - out->pos : out->len;
 			ssize_t n;
 
@@ -894,7 +945,8 @@ static char *perform_io(size_t needed, int flags)
 				wait_for_receiver(); /* generator only */
 		}
 
-		if (ff_forward_fd >= 0 && ff_pollpos >= 0 && pfds[ff_pollpos].revents) {
+		if (ff_forward_fd >= 0 && ff_pollpos >= 0
+		 && pfds[ff_pollpos].revents & POLL_RD_BITS) {
 			/* This can potentially flush all output and enable
 			 * multiplexed output, so keep this last in the loop
 			 * and be sure to not cache anything that would break
@@ -1153,8 +1205,11 @@ void io_set_sock_fds(int f_in, int f_out)
 
 void set_io_timeout(int secs)
 {
+	if (secs < 0)
+		secs = 0;
 	io_timeout = secs;
-	allowed_lull = (io_timeout + 1) / 2;
+	/* Round up without overflowing when secs is near INT_MAX. */
+	allowed_lull = secs / 2 + secs % 2;
 
 	if (!io_timeout || allowed_lull > SELECT_TIMEOUT)
 		select_timeout = SELECT_TIMEOUT;
@@ -1559,7 +1614,10 @@ static void read_a_msg(void)
 			goto invalid_msg;
 		val = raw_read_int();
 		iobuf.in_multiplexed = 1;
-		if (!io_timeout || io_timeout > val) {
+		/* Ignore a non-positive value: it would otherwise disable our
+		 * timeout entirely (and a negative one used to be rejected by
+		 * select(), but would make poll() wait forever). */
+		if (val > 0 && (!io_timeout || io_timeout > val)) {
 			if (INFO_GTE(MISC, 2))
 				rprintf(FINFO, "Setting --timeout=%d to match server\n", val);
 			set_io_timeout(val);

socket.c

@@ -25,6 +25,7 @@
  * emulate it using the KAME implementation. */
 
 #include "rsync.h"
+#include <poll.h>
 #include "itypes.h"
 #include "ifuncs.h"
 #ifdef HAVE_NETINET_IN_SYSTM_H
@@ -536,8 +537,8 @@ static void sigchld_handler(UNUSED(int val))
 
 void start_accept_loop(int port, int (*fn)(int, int))
 {
-	fd_set deffds;
-	int *sp, maxfd, i;
+	struct pollfd *pfds;
+	int *sp, nsp, i;
 
 #ifdef HAVE_SIGACTION
 	sigact.sa_flags = SA_NOCLDSTOP;
@@ -549,8 +550,12 @@ void start_accept_loop(int port, int (*fn)(int, int))
 		exit_cleanup(RERR_SOCKETIO);
 
 	/* ready to listen */
-	FD_ZERO(&deffds);
-	for (i = 0, maxfd = -1; sp[i] >= 0; i++) {
+	for (nsp = 0; sp[nsp] >= 0; nsp++) {}
+	/* poll() rather than select(): a listening fd at or above FD_SETSIZE
+	 * would overflow an fd_set, which is undefined behaviour (issue #231). */
+	if (!(pfds = new_array(struct pollfd, nsp ? nsp : 1)))
+		out_of_memory("start_accept_loop");
+	for (i = 0; sp[i] >= 0; i++) {
 		if (listen(sp[i], lp_listen_backlog()) < 0) {
 			rsyserr(FERROR, errno, "listen() on socket failed");
 #ifdef INET6
@@ -560,36 +565,32 @@ void start_accept_loop(int port, int (*fn)(int, int))
 #endif
 			exit_cleanup(RERR_SOCKETIO);
 		}
-		FD_SET(sp[i], &deffds);
-		if (maxfd < sp[i])
-			maxfd = sp[i];
+		pfds[i].fd = sp[i];
+		pfds[i].events = POLLIN;
+		pfds[i].revents = 0;
 	}
 
 	/* now accept incoming connections - forking a new process
 	 * for each incoming connection */
 	while (1) {
-		fd_set fds;
 		pid_t pid;
 		int fd;
 		struct sockaddr_storage addr;
 		socklen_t addrlen = sizeof addr;
 
-		/* close log file before the potentially very long select so
+		/* close log file before the potentially very long wait so the
 		 * file can be trimmed by another process instead of growing
 		 * forever */
 		logfile_close();
 
-#ifdef FD_COPY
-		FD_COPY(&deffds, &fds);
-#else
-		fds = deffds;
-#endif
+		for (i = 0; i < nsp; i++)
+			pfds[i].revents = 0;
 
-		if (select(maxfd + 1, &fds, NULL, NULL, NULL) < 1)
+		if (poll(pfds, nsp, -1) < 1)
 			continue;
 
-		for (i = 0, fd = -1; sp[i] >= 0; i++) {
-			if (FD_ISSET(sp[i], &fds)) {
+		for (i = 0, fd = -1; i < nsp; i++) {
+			if (pfds[i].revents & (POLLIN | POLLERR | POLLHUP)) {
 				fd = accept(sp[i], (struct sockaddr *)&addr, &addrlen);
 				break;
 			}

testsuite/highfd-hang_test.py

@@ -1,84 +1,120 @@
 #!/usr/bin/env python3
-"""Regression test for issue #231: rsync must not hang when its I/O file
-descriptors are numbered at/above FD_SETSIZE.
+"""Regression test for issue #231: high-numbered I/O descriptors.
 
-rsync's I/O loops in io.c used to call select() with fd_set bitmaps, which can
-only represent descriptors below FD_SETSIZE (1024 with glibc).  If rsync is
-started with many descriptors already open -- e.g. inherited from a parent that
-leaked fds -- its own socket/pipe fds get allocated above 1024.  FD_SET() and
+rsync's I/O readiness loops used to call select() with fd_set bitmaps, which can
+only represent descriptors below FD_SETSIZE.  If rsync is started with many
+descriptors already open -- e.g. inherited from a parent that leaked fds -- its
+own socket and pipe fds get allocated at or above that limit.  FD_SET() and
 FD_ISSET() then index past the end of the fixed-size fd_set, which is undefined
-behavior: select() reports the fd ready, but FD_ISSET() reads the out-of-bounds
-bit as 0, so the read/write never happens and rsync spins at 100% CPU forever
-("rsync hangs, 100% of one CPU, no progress").
+behaviour: with a plain build select() reports the fd ready while FD_ISSET()
+reads the out-of-bounds bit as 0, so the read/write never happens and rsync
+spins at 100% CPU forever; with a fortified libc the process instead aborts
+("bit out of range 0 - FD_SETSIZE in fd_set").  Either way the transfer never
+completes.
 
-We reproduce that deterministically by opening enough inheritable dummy fds to
-push rsync's descriptors past FD_SETSIZE, then running an ordinary transfer
-with close_fds=False so the child inherits them.  With the poll()-based io.c
-the transfer finishes instantly; with the old select()-based code it never
-finishes and is caught by the timeout.  The cross-over is binary (instant vs.
-infinite), so the timeout is a robust signal rather than a timing race.
+We reproduce that by opening enough inheritable dummy fds to push rsync's
+descriptors past FD_SETSIZE, then running an ordinary transfer with
+close_fds=False so the child inherits them.  With the poll()-based I/O the
+transfer finishes immediately; without it, it hangs or aborts.
+
+FD_SETSIZE is *not* hardcoded: it is 1024 with glibc but 65536 on 64-bit
+Solaris, and assuming 1024 there would open too few fds to cross the limit and
+pass vacuously.  We ask the C library for the real value instead, and skip when
+we cannot (no compiler, or the fd limit cannot be raised far enough).
 """
 
 import os
 import resource
+import shutil
 import subprocess
+import tempfile
 
 from rsyncfns import (
     FROMDIR, TODIR, rmtree, rsync_argv, test_fail, test_skipped,
 )
 
-# select()'s fd_set holds FD_SETSIZE bits; glibc and most libcs use 1024.
-FD_SETSIZE = 1024
-NDUMMY = FD_SETSIZE + 80      # force rsync's fds comfortably past the limit
-TIMEOUT = 30                  # poll build: ~instant; select build: hangs forever
+TIMEOUT = 30  # poll() build: ~instant; the bug: hangs (or aborts)
+
+
+def probe_fd_setsize():
+    """Ask the C library for FD_SETSIZE rather than assuming a value."""
+    cc = os.environ.get('CC') or shutil.which('cc') or shutil.which('gcc')
+    if not cc:
+        return None
+    with tempfile.TemporaryDirectory() as td:
+        src = os.path.join(td, 'fdss.c')
+        exe = os.path.join(td, 'fdss')
+        with open(src, 'w') as f:
+            f.write('#include <stdio.h>\n'
+                    '#include <sys/select.h>\n'
+                    'int main(void){ printf("%d\\n", (int)FD_SETSIZE); return 0; }\n')
+        if subprocess.run([cc, src, '-o', exe],
+                          capture_output=True).returncode != 0:
+            return None
+        proc = subprocess.run([exe], capture_output=True, text=True)
+        if proc.returncode != 0:
+            return None
+        try:
+            return int(proc.stdout.strip())
+        except ValueError:
+            return None
+
+
+fd_setsize = probe_fd_setsize()
+if not fd_setsize:
+    test_skipped("could not determine FD_SETSIZE (no usable C compiler)")
+
+# Push rsync's descriptors comfortably past the limit.
+ndummy = fd_setsize + 80
+want = ndummy + 64
 
-# We must be able to open enough descriptors to cross FD_SETSIZE.
 soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
-want = NDUMMY + 64
 if soft < want:
     if hard != resource.RLIM_INFINITY and hard < want:
-        test_skipped(f"RLIMIT_NOFILE hard cap {hard} < {want}; cannot place "
-                     "fds above FD_SETSIZE to exercise issue #231")
+        test_skipped(f"RLIMIT_NOFILE hard cap {hard} < {want}; cannot place fds "
+                     f"above FD_SETSIZE ({fd_setsize}) to exercise issue #231")
     resource.setrlimit(resource.RLIMIT_NOFILE, (want, hard))
 
-# A small tree to transfer.
 rmtree(FROMDIR)
 rmtree(TODIR)
 FROMDIR.mkdir(parents=True, exist_ok=True)
-payload = {f"f{i}": os.urandom(1000) for i in range(20)}
+TODIR.mkdir(parents=True, exist_ok=True)
+
+payload = {f'f{i}': os.urandom(1000) for i in range(20)}
 for name, data in payload.items():
     (FROMDIR / name).write_bytes(data)
 
-# Occupy the low fd numbers with inheritable dummy descriptors so that the
-# rsync child's socket/pipe fds are forced above FD_SETSIZE.  Python opens fds
-# O_CLOEXEC by default, so mark each inheritable and use close_fds=False.
+# Occupy the low fd numbers with inheritable dummy descriptors so the rsync
+# child's socket/pipe fds land above FD_SETSIZE.  Python opens fds O_CLOEXEC by
+# default, so mark each inheritable and use close_fds=False.
 dummies = []
 try:
     while True:
         fd = os.open(os.devnull, os.O_RDONLY)
         os.set_inheritable(fd, True)
         dummies.append(fd)
-        if fd >= NDUMMY:
+        if fd >= ndummy:
             break
 
     argv = rsync_argv('-a', f'{FROMDIR}/', f'{TODIR}/')
     try:
         proc = subprocess.run(argv, timeout=TIMEOUT, close_fds=False)
     except subprocess.TimeoutExpired:
-        test_fail("rsync hung with high-numbered fds -- select()/fd_set "
-                  "overflow (issue #231 regression)")
+        test_fail(f"rsync did not finish within {TIMEOUT}s with fds above "
+                  f"FD_SETSIZE ({fd_setsize}) -- select()/fd_set overflow "
+                  "(issue #231 regression)")
 finally:
     for fd in dummies:
         os.close(fd)
 
 if proc.returncode != 0:
-    test_fail(f"rsync exited {proc.returncode}: {' '.join(argv)}")
+    test_fail(f"rsync exited {proc.returncode} with fds above FD_SETSIZE "
+              f"({fd_setsize}); a fortified libc aborts on the fd_set overflow "
+              "(issue #231 regression)")
 
-# The cap only matters if it stayed correct: verify every file transferred.
 for name, data in payload.items():
-    got = (TODIR / name).read_bytes()
-    if got != data:
+    if (TODIR / name).read_bytes() != data:
         test_fail(f"{name} differs after a high-fd transfer")
 
-print(f"issue #231: transfer with fds >= {NDUMMY} (above FD_SETSIZE) "
+print(f"issue #231: transfer with fds above FD_SETSIZE ({fd_setsize}) "
       "completed correctly")