[PATCH 0/2 V3] Switch to poll() in cxenstored's IO loop

[PATCH 0/2 V3] Switch to poll() in cxenstored's IO loop

[Wei Liu]

This is the third version of this patch series. All issues discovered in the
switching of xenconsoled have been fixed.

[PATCH 1/2] mini-os: implement poll(2)

[Wei Liu]

It is just a wrapper around select(2). This implementation mimics Linux's
do_poll.

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Samuel Thibault <samuel.thibault@ens-lyon.org>
---
 extras/mini-os/include/posix/poll.h |    1 +
 extras/mini-os/lib/sys.c            |  117 ++++++++++++++++++++++++++++++++++-
 2 files changed, 117 insertions(+), 1 deletion(-)
 create mode 100644 extras/mini-os/include/posix/poll.h

diff --git a/extras/mini-os/include/posix/poll.h b/extras/mini-os/include/posix/poll.h
new file mode 100644
index 0000000..06fb41a
--- /dev/null
+++ b/extras/mini-os/include/posix/poll.h
@@ -0,0 +1 @@
+#include <sys/poll.h>
diff --git a/extras/mini-os/lib/sys.c b/extras/mini-os/lib/sys.c
index 3cc3340..cfbdc90 100644
--- a/extras/mini-os/lib/sys.c
+++ b/extras/mini-os/lib/sys.c
@@ -31,6 +31,7 @@
 #include <tpm_tis.h>
 #include <xenbus.h>
 #include <xenstore.h>
+#include <poll.h>
 
 #include <sys/types.h>
 #include <sys/unistd.h>
@@ -678,6 +679,29 @@ static void dump_set(int nfds, fd_set *readfds, fd_set *writefds, fd_set *except
 #define dump_set(nfds, readfds, writefds, exceptfds, timeout)
 #endif
 
+#ifdef LIBC_DEBUG
+static void dump_pollfds(struct pollfd *pfd, int nfds, int timeout)
+{
+    int i, comma, fd;
+
+    printk("[");
+    comma = 0;
+    for (i = 0; i < nfds; i++) {
+        fd = pfd[i].fd;
+        if (comma)
+            printk(", ");
+        printk("%d(%c)/%02x", fd, file_types[files[fd].type],
+            pfd[i].events);
+            comma = 1;
+    }
+    printk("]");
+
+    printk(", %d, %d", nfds, timeout);
+}
+#else
+#define dump_pollfds(pfds, nfds, timeout)
+#endif
+
 /* Just poll without blocking */
 static int select_poll(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds)
 {
@@ -983,6 +1007,98 @@ out:
     return ret;
 }
 
+/* Wrap around select */
+int poll(struct pollfd _pfd[], nfds_t _nfds, int _timeout)
+{
+    int n, ret;
+    int i, fd;
+    struct timeval _timeo, *timeo = NULL;
+    fd_set rfds, wfds, efds;
+    int max_fd = -1;
+
+    DEBUG("poll(");
+    dump_pollfds(_pfd, _nfds, _timeout);
+    DEBUG(")\n");
+
+    FD_ZERO(&rfds);
+    FD_ZERO(&wfds);
+    FD_ZERO(&efds);
+
+    n = 0;
+
+    for (i = 0; i < _nfds; i++) {
+        fd = _pfd[i].fd;
+        _pfd[i].revents = 0;
+
+        /* fd < 0, revents = 0, which is already set */
+        if (fd < 0) continue;
+
+        /* fd is invalid, revents = POLLNVAL, increment counter */
+        if (fd >= NOFILE || files[fd].type == FTYPE_NONE) {
+            n++;
+            _pfd[i].revents |= POLLNVAL;
+            continue;
+        }
+
+        /* normal case, map POLL* into readfds and writefds:
+         * POLLIN  -> readfds
+         * POLLOUT -> writefds
+         * POLL*   -> none
+         */
+        if (_pfd[i].events & POLLIN)
+            FD_SET(fd, &rfds);
+        if (_pfd[i].events & POLLOUT)
+            FD_SET(fd, &wfds);
+        /* always set exceptfds */
+        FD_SET(fd, &efds);
+        if (fd > max_fd)
+            max_fd = fd;
+    }
+
+    /* should never sleep when we already have events */
+    if (n) {
+        _timeo.tv_sec  = 0;
+        _timeo.tv_usec = 0;
+        timeo = &_timeo;
+    } else if (_timeout >= 0) {
+        /* normal case, construct _timeout, might sleep */
+        _timeo.tv_sec  = _timeout / 1000;
+        _timeo.tv_usec = (_timeout % 1000) * 1000;
+        timeo = &_timeo;
+    } else {
+        /* _timeout < 0, block forever */
+        timeo = NULL;
+    }
+
+
+    ret = select(max_fd+1, &rfds, &wfds, &efds, timeo);
+    /* error in select, just return, errno is set by select() */
+    if (ret < 0)
+        return ret;
+
+    for (i = 0; i < _nfds; i++) {
+        fd = _pfd[i].fd;
+
+        /* the revents has already been set for all error case */
+        if (fd < 0 || fd >= NOFILE || files[fd].type == FTYPE_NONE)
+            continue;
+
+        if (FD_ISSET(fd, &rfds) || FD_ISSET(fd, &wfds) || FD_ISSET(fd, &efds))
+            n++;
+        if (FD_ISSET(fd, &efds)) {
+            /* anything bad happens we set POLLERR */
+            _pfd[i].revents |= POLLERR;
+            continue;
+        }
+        if (FD_ISSET(fd, &rfds))
+            _pfd[i].revents |= POLLIN;
+        if (FD_ISSET(fd, &wfds))
+            _pfd[i].revents |= POLLOUT;
+    }
+
+    return n;
+}
+
 #ifdef HAVE_LWIP
 int socket(int domain, int type, int protocol)
 {
@@ -1360,7 +1476,6 @@ unsupported_function(int, tcgetattr, 0);
 unsupported_function(int, grantpt, -1);
 unsupported_function(int, unlockpt, -1);
 unsupported_function(char *, ptsname, NULL);
-unsupported_function(int, poll, -1);
 
 /* net/if.h */
 unsupported_function_log(unsigned int, if_nametoindex, -1);
-- 
1.7.10.4

[PATCH 2/2] Switch to poll() in cxenstored's IO loop

[Wei Liu]

Poll() can support more file descriptors than select(). We've done this for
xenconsoled, now do this for cxenstored as well.

The code is taken from xenconsoled and modified to adapt to cxenstored.

Note that poll() semantic is a bit different from select(). In Linux, if a fd
is set in IN/OUT fd_set and error occurs inside select(), this fd is still
considered readable / writable, and it is set in the returned IN/OUT fd_set.
So in later handle_input / handle_output, the connection will eventually be
talloc_free'ed(). After switching to poll(), we should take care of any error
right away, making the code clearer.

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
---
 tools/xenstore/xenstored_core.c |  191 +++++++++++++++++++++++++++------------
 tools/xenstore/xenstored_core.h |    2 +
 2 files changed, 133 insertions(+), 60 deletions(-)

diff --git a/tools/xenstore/xenstored_core.c b/tools/xenstore/xenstored_core.c
index bd44645..a2cf2a6 100644
--- a/tools/xenstore/xenstored_core.c
+++ b/tools/xenstore/xenstored_core.c
@@ -19,7 +19,7 @@
 
 #include <sys/types.h>
 #include <sys/stat.h>
-#include <sys/select.h>
+#include <poll.h>
 #ifndef NO_SOCKETS
 #include <sys/socket.h>
 #include <sys/un.h>
@@ -55,6 +55,12 @@
 #include "hashtable.h"
 
 extern xc_evtchn *xce_handle; /* in xenstored_domain.c */
+static int xce_pollfd_idx = -1;
+static struct pollfd *fds;
+static unsigned int current_array_size;
+static unsigned int nr_fds;
+
+#define ROUNDUP(_x, _w) (((unsigned long)(_x)+(1UL<<(_w))-1) & ~((1UL<<(_w))-1))
 
 static bool verbose = false;
 LIST_HEAD(connections);
@@ -62,6 +68,7 @@ static int tracefd = -1;
 static bool recovery = true;
 static bool remove_local = true;
 static int reopen_log_pipe[2];
+static int reopen_log_pipe0_pollfd_idx = -1;
 static char *tracefile = NULL;
 static TDB_CONTEXT *tdb_ctx = NULL;
 
@@ -199,7 +206,7 @@ void trace_destroy(const void *data, const char *type)
 /**
  * Signal handler for SIGHUP, which requests that the trace log is reopened
  * (in the main loop).  A single byte is written to reopen_log_pipe, to awaken
- * the select() in the main loop.
+ * the poll() in the main loop.
  */
 static void trigger_reopen_log(int signal __attribute__((unused)))
 {
@@ -279,15 +286,12 @@ static int destroy_conn(void *_conn)
 
 	/* Flush outgoing if possible, but don't block. */
 	if (!conn->domain) {
-		fd_set set;
-		struct timeval none;
-
-		FD_ZERO(&set);
-		FD_SET(conn->fd, &set);
-		none.tv_sec = none.tv_usec = 0;
+		struct pollfd pfd;
+		pfd.fd = conn->fd;
+		pfd.events = POLLOUT;
 
 		while (!list_empty(&conn->out_list)
-		       && select(conn->fd+1, NULL, &set, NULL, &none) == 1)
+		       && poll(&pfd, 1, 0) == 1)
 			if (!write_messages(conn))
 				break;
 		close(conn->fd);
@@ -299,53 +303,77 @@ static int destroy_conn(void *_conn)
 	return 0;
 }
 
-
-static void set_fd(int fd, fd_set *set, int *max)
+/* This function returns index inside the array if succeed, -1 if fail */
+static int set_fd(int fd, short events)
 {
-	if (fd < 0)
-		return;
-	FD_SET(fd, set);
-	if (fd > *max)
-		*max = fd;
-}
+	int ret;
+	if (current_array_size < nr_fds + 1) {
+		struct pollfd *new_fds = NULL;
+		unsigned long newsize;
+
+		/* Round up to 2^8 boundary, in practice this just
+		 * make newsize larger than current_array_size.
+		 */
+		newsize = ROUNDUP(nr_fds + 1, 8);
+
+		new_fds = realloc(fds, sizeof(struct pollfd)*newsize);
+		if (!new_fds)
+			goto fail;
+		fds = new_fds;
+
+		memset(&fds[0] + current_array_size, 0,
+		       sizeof(struct pollfd ) * (newsize-current_array_size));
+		current_array_size = newsize;
+	}
 
+	fds[nr_fds].fd = fd;
+	fds[nr_fds].events = events;
+	ret = nr_fds;
+	nr_fds++;
 
-static int initialize_set(fd_set *inset, fd_set *outset, int sock, int ro_sock,
-			  struct timeval **ptimeout)
+	return ret;
+fail:
+	syslog(LOG_ERR, "realloc failed, ignoring fd %d\n", fd);
+	return -1;
+}
+
+static void initialize_fds(int sock, int *p_sock_pollfd_idx,
+			   int ro_sock, int *p_ro_sock_pollfd_idx,
+			   int *ptimeout)
 {
-	static struct timeval zero_timeout = { 0 };
 	struct connection *conn;
-	int max = -1;
 
-	*ptimeout = NULL;
+	if (fds)
+		memset(fds, 0, sizeof(struct pollfd) * current_array_size);
+	nr_fds = 0;
 
-	FD_ZERO(inset);
-	FD_ZERO(outset);
+	*ptimeout = -1;
 
 	if (sock != -1)
-		set_fd(sock, inset, &max);
+		*p_sock_pollfd_idx = set_fd(sock, POLLIN|POLLPRI);
 	if (ro_sock != -1)
-		set_fd(ro_sock, inset, &max);
+		*p_ro_sock_pollfd_idx = set_fd(ro_sock, POLLIN|POLLPRI);
 	if (reopen_log_pipe[0] != -1)
-		set_fd(reopen_log_pipe[0], inset, &max);
+		reopen_log_pipe0_pollfd_idx =
+			set_fd(reopen_log_pipe[0], POLLIN|POLLPRI);
 
 	if (xce_handle != NULL)
-		set_fd(xc_evtchn_fd(xce_handle), inset, &max);
+		xce_pollfd_idx = set_fd(xc_evtchn_fd(xce_handle),
+					POLLIN|POLLPRI);
 
 	list_for_each_entry(conn, &connections, list) {
 		if (conn->domain) {
 			if (domain_can_read(conn) ||
 			    (domain_can_write(conn) &&
 			     !list_empty(&conn->out_list)))
-				*ptimeout = &zero_timeout;
+				*ptimeout = 0;
 		} else {
-			set_fd(conn->fd, inset, &max);
+			short events = POLLIN|POLLPRI;
 			if (!list_empty(&conn->out_list))
-				FD_SET(conn->fd, outset);
+				events |= POLLOUT;
+			conn->pollfd_idx = set_fd(conn->fd, events);
 		}
 	}
-
-	return max;
 }
 
 /* Is child a subnode of parent, or equal? */
@@ -1330,6 +1358,7 @@ struct connection *new_connection(connwritefn_t *write, connreadfn_t *read)
 		return NULL;
 
 	new->fd = -1;
+	new->pollfd_idx = -1;
 	new->write = write;
 	new->read = read;
 	new->can_write = true;
@@ -1770,14 +1799,13 @@ int priv_domid = 0;
 
 int main(int argc, char *argv[])
 {
-	int opt, *sock, *ro_sock, max;
-	fd_set inset, outset;
+	int opt, *sock, *ro_sock;
+	int sock_pollfd_idx = -1, ro_sock_pollfd_idx = -1;
 	bool dofork = true;
 	bool outputpid = false;
 	bool no_domain_init = false;
 	const char *pidfile = NULL;
-	int evtchn_fd = -1;
-	struct timeval *timeout;
+	int timeout;
 
 	while ((opt = getopt_long(argc, argv, "DE:F:HNPS:t:T:RLVW:", options,
 				  NULL)) != -1) {
@@ -1880,11 +1908,9 @@ int main(int argc, char *argv[])
 
 	signal(SIGHUP, trigger_reopen_log);
 
-	if (xce_handle != NULL)
-		evtchn_fd = xc_evtchn_fd(xce_handle);
-
 	/* Get ready to listen to the tools. */
-	max = initialize_set(&inset, &outset, *sock, *ro_sock, &timeout);
+	initialize_fds(*sock, &sock_pollfd_idx, *ro_sock, &ro_sock_pollfd_idx,
+		       &timeout);
 
 	/* Tell the kernel we're up and running. */
 	xenbus_notify_running();
@@ -1893,27 +1919,57 @@ int main(int argc, char *argv[])
 	for (;;) {
 		struct connection *conn, *next;
 
-		if (select(max+1, &inset, &outset, NULL, timeout) < 0) {
+		if (poll(fds, nr_fds, timeout) < 0) {
 			if (errno == EINTR)
 				continue;
-			barf_perror("Select failed");
+			barf_perror("Poll failed");
 		}
 
-		if (reopen_log_pipe[0] != -1 && FD_ISSET(reopen_log_pipe[0], &inset)) {
-			char c;
-			if (read(reopen_log_pipe[0], &c, 1) != 1)
-				barf_perror("read failed");
-			reopen_log();
+		if (reopen_log_pipe0_pollfd_idx != -1) {
+			if (fds[reopen_log_pipe0_pollfd_idx].revents
+			    & ~POLLIN) {
+				close(reopen_log_pipe[0]);
+				close(reopen_log_pipe[1]);
+				init_pipe(reopen_log_pipe);
+			} else if (fds[reopen_log_pipe0_pollfd_idx].revents
+				   & POLLIN) {
+				char c;
+				if (read(reopen_log_pipe[0], &c, 1) != 1)
+					barf_perror("read failed");
+				reopen_log();
+			}
+			reopen_log_pipe0_pollfd_idx = -1;
 		}
 
-		if (*sock != -1 && FD_ISSET(*sock, &inset))
-			accept_connection(*sock, true);
+		if (sock_pollfd_idx != -1) {
+			if (fds[sock_pollfd_idx].revents & ~POLLIN) {
+				barf_perror("sock poll failed");
+				break;
+			} else if (fds[sock_pollfd_idx].revents & POLLIN) {
+				accept_connection(*sock, true);
+				sock_pollfd_idx = -1;
+			}
+		}
 
-		if (*ro_sock != -1 && FD_ISSET(*ro_sock, &inset))
-			accept_connection(*ro_sock, false);
+		if (ro_sock_pollfd_idx != -1) {
+			if (fds[ro_sock_pollfd_idx].revents & ~POLLIN) {
+				barf_perror("ro sock poll failed");
+				break;
+			} else if (fds[ro_sock_pollfd_idx].revents & POLLIN) {
+				accept_connection(*ro_sock, false);
+				ro_sock_pollfd_idx = -1;
+			}
+		}
 
-		if (evtchn_fd != -1 && FD_ISSET(evtchn_fd, &inset))
-			handle_event();
+		if (xce_pollfd_idx != -1) {
+			if (fds[xce_pollfd_idx].revents & ~POLLIN) {
+				barf_perror("xce_handle poll failed");
+				break;
+			} else if (fds[xce_pollfd_idx].revents & POLLIN) {
+				handle_event();
+				xce_pollfd_idx = -1;
+			}
+		}
 
 		next = list_entry(connections.next, typeof(*conn), list);
 		if (&next->list != &connections)
@@ -1939,21 +1995,36 @@ int main(int argc, char *argv[])
 				if (talloc_free(conn) == 0)
 					continue;
 			} else {
-				if (FD_ISSET(conn->fd, &inset))
-					handle_input(conn);
+				if (conn->pollfd_idx != -1) {
+					if (fds[conn->pollfd_idx].revents
+					    & ~(POLLIN|POLLOUT))
+						talloc_free(conn);
+					else if (fds[conn->pollfd_idx].revents
+						 & POLLIN)
+						handle_input(conn);
+				}
 				if (talloc_free(conn) == 0)
 					continue;
 
 				talloc_increase_ref_count(conn);
-				if (FD_ISSET(conn->fd, &outset))
-					handle_output(conn);
+
+				if (conn->pollfd_idx != -1) {
+					if (fds[conn->pollfd_idx].revents
+					    & ~(POLLIN|POLLOUT))
+						talloc_free(conn);
+					else if (fds[conn->pollfd_idx].revents
+						 & POLLOUT)
+						handle_output(conn);
+				}
 				if (talloc_free(conn) == 0)
 					continue;
+
+				conn->pollfd_idx = -1;
 			}
 		}
 
-		max = initialize_set(&inset, &outset, *sock, *ro_sock,
-				     &timeout);
+		initialize_fds(*sock, &sock_pollfd_idx, *ro_sock,
+			       &ro_sock_pollfd_idx, &timeout);
 	}
 }
 
diff --git a/tools/xenstore/xenstored_core.h b/tools/xenstore/xenstored_core.h
index 492ca0d..cfbcf6f 100644
--- a/tools/xenstore/xenstored_core.h
+++ b/tools/xenstore/xenstored_core.h
@@ -60,6 +60,8 @@ struct connection
 
 	/* The file descriptor we came in on. */
 	int fd;
+	/* The index of pollfd in global pollfd array */
+	int pollfd_idx;
 
 	/* Who am I? 0 for socket connections. */
 	unsigned int id;
-- 
1.7.10.4

Re: [PATCH 0/2 V3] Switch to poll() in cxenstored's IO loop

[Ian Campbell]

On Mon, 2013-03-25 at 11:17 +0000, Wei Liu wrote:
> This is the third version of this patch series. All issues discovered in the
> switching of xenconsoled have been fixed.

Looks good, I acked the second one (which was a lot like the previous
iteration + using idx instead of ptr) and applied both.