blob: 42557b17a7fb9bb19324579fd0b67a7b0d67dc50 [file] [log] [blame]
From 9e407e0be01695e7b927f5820ade87ee9602c248 Mon Sep 17 00:00:00 2001
From: Alexander Kanavin <alex.kanavin@gmail.com>
Date: Fri, 15 Sep 2017 17:00:14 +0300
Subject: [PATCH] Use epoll API on Linux
Also a couple of other modifications due to epoll having
a different approach to how the working set of fds is defined
and used:
1) open_client() returns an index into the array of clients
2) close_client() has a protection against being called twice
with the same client (which would mess up the active_clients
counter)
Upstream-Status: Submitted [Seebs CC'd by email]
Signed-off-by: Alexander Kanavin <alex.kanavin@gmail.com>
---
enums/exit_status.in | 3 +
pseudo_server.c | 189 ++++++++++++++++++++++++++++++++++++++++++++++++++-
2 files changed, 190 insertions(+), 2 deletions(-)
diff --git a/enums/exit_status.in b/enums/exit_status.in
index 6be44d3..88f94cd 100644
--- a/enums/exit_status.in
+++ b/enums/exit_status.in
@@ -18,3 +18,6 @@ listen_fd, "server loop had no valid listen fd"
pseudo_loaded, "server couldn't get out of pseudo environment"
pseudo_prefix, "couldn't get valid pseudo prefix"
pseudo_invocation, "invalid server command arguments"
+epoll_create, "epoll_create() failed"
+epoll_ctl, "epoll_ctl() failed"
+
diff --git a/pseudo_server.c b/pseudo_server.c
index ff16efd..14d34de 100644
--- a/pseudo_server.c
+++ b/pseudo_server.c
@@ -40,6 +40,12 @@
#include "pseudo_client.h"
#include "pseudo_db.h"
+// This has to come after pseudo includes, as that's where PSEUDO_PORT defines are
+#ifdef PSEUDO_PORT_LINUX
+#include <sys/epoll.h>
+#endif
+
+
static int listen_fd = -1;
typedef struct {
@@ -59,6 +65,7 @@ static int active_clients = 0, highest_client = 0, max_clients = 0;
#define LOOP_DELAY 2
#define DEFAULT_PSEUDO_SERVER_TIMEOUT 30
+#define EPOLL_MAX_EVENTS 10
int pseudo_server_timeout = DEFAULT_PSEUDO_SERVER_TIMEOUT;
static int die_peacefully = 0;
static int die_forcefully = 0;
@@ -80,6 +87,9 @@ quit_now(int signal) {
static int messages = 0, responses = 0;
static struct timeval message_time = { .tv_sec = 0 };
+#ifdef PSEUDO_PORT_LINUX
+static void pseudo_server_loop_epoll(void);
+#endif
static void pseudo_server_loop(void);
/* helper function to make a directory, just like mkdir -p.
@@ -369,12 +379,16 @@ pseudo_server_start(int daemonize) {
kill(ppid, SIGUSR1);
}
}
+#ifdef PSEUDO_PORT_LINUX
+ pseudo_server_loop_epoll();
+#else
pseudo_server_loop();
+#endif
return 0;
}
/* mess with internal tables as needed */
-static void
+static unsigned int
open_client(int fd) {
pseudo_client_t *new_clients;
int i;
@@ -390,7 +404,7 @@ open_client(int fd) {
++active_clients;
if (i > highest_client)
highest_client = i;
- return;
+ return i;
}
}
@@ -414,9 +428,11 @@ open_client(int fd) {
max_clients += 16;
++active_clients;
+ return max_clients - 16;
} else {
pseudo_diag("error allocating new client, fd %d\n", fd);
close(fd);
+ return 0;
}
}
@@ -433,6 +449,10 @@ close_client(int client) {
client, highest_client);
return;
}
+ if (clients[client].fd == -1) {
+ pseudo_debug(PDBGF_SERVER, "client %d already closed\n", client);
+ return;
+ }
close(clients[client].fd);
clients[client].fd = -1;
free(clients[client].tag);
@@ -566,6 +586,171 @@ serve_client(int i) {
}
}
+#ifdef PSEUDO_PORT_LINUX
+static void pseudo_server_loop_epoll(void)
+{
+ struct sockaddr_un client;
+ socklen_t len;
+ int i;
+ int rc;
+ int fd;
+ int timeout;
+ struct epoll_event ev, events[EPOLL_MAX_EVENTS];
+ int loop_timeout = pseudo_server_timeout;
+
+ clients = malloc(16 * sizeof(*clients));
+
+ clients[0].fd = listen_fd;
+ clients[0].pid = getpid();
+
+ for (i = 1; i < 16; ++i) {
+ clients[i].fd = -1;
+ clients[i].pid = 0;
+ clients[i].tag = NULL;
+ clients[i].program = NULL;
+ }
+
+ active_clients = 1;
+ max_clients = 16;
+ highest_client = 0;
+
+ pseudo_debug(PDBGF_SERVER, "server loop started.\n");
+ if (listen_fd < 0) {
+ pseudo_diag("got into loop with no valid listen fd.\n");
+ exit(PSEUDO_EXIT_LISTEN_FD);
+ }
+
+ timeout = LOOP_DELAY * 1000;
+
+ int epollfd = epoll_create1(0);
+ if (epollfd == -1) {
+ pseudo_diag("epoll_create1() failed.\n");
+ exit(PSEUDO_EXIT_EPOLL_CREATE);
+ }
+ ev.events = EPOLLIN;
+ ev.data.u64 = 0;
+ if (epoll_ctl(epollfd, EPOLL_CTL_ADD, clients[0].fd, &ev) == -1) {
+ pseudo_diag("epoll_ctl() failed with listening socket.\n");
+ exit(PSEUDO_EXIT_EPOLL_CTL);
+ }
+
+ pdb_log_msg(SEVERITY_INFO, NULL, NULL, NULL, "server started (pid %d)", getpid());
+
+ for (;;) {
+ rc = epoll_wait(epollfd, events, EPOLL_MAX_EVENTS, timeout);
+ if (rc == 0 || (rc == -1 && errno == EINTR)) {
+ /* If there's no clients, start timing out. If there
+ * are active clients, never time out.
+ */
+ if (active_clients == 1) {
+ loop_timeout -= LOOP_DELAY;
+ /* maybe flush database to disk */
+ pdb_maybe_backup();
+ if (loop_timeout <= 0) {
+ pseudo_debug(PDBGF_SERVER, "no more clients, got bored.\n");
+ die_peacefully = 1;
+ } else {
+ /* display this if not exiting */
+ pseudo_debug(PDBGF_SERVER | PDBGF_BENCHMARK, "%d messages handled in %.4f seconds, %d responses\n",
+ messages,
+ (double) message_time.tv_sec +
+ (double) message_time.tv_usec / 1000000.0,
+ responses);
+ }
+ }
+ } else if (rc > 0) {
+ loop_timeout = pseudo_server_timeout;
+ for (i = 0; i < rc; ++i) {
+ if (clients[events[i].data.u64].fd == listen_fd) {
+ if (!die_forcefully) {
+ len = sizeof(client);
+ if ((fd = accept(listen_fd, (struct sockaddr *) &client, &len)) != -1) {
+ /* Don't allow clients to end up on fd 2, because glibc's
+ * malloc debug uses that fd unconditionally.
+ */
+ if (fd == 2) {
+ int newfd = fcntl(fd, F_DUPFD, 3);
+ close(fd);
+ fd = newfd;
+ }
+ pseudo_debug(PDBGF_SERVER, "new client fd %d\n", fd);
+ /* A new client implicitly cancels any
+ * previous shutdown request, or a
+ * shutdown for lack of clients.
+ */
+ pseudo_server_timeout = DEFAULT_PSEUDO_SERVER_TIMEOUT;
+ die_peacefully = 0;
+
+ ev.events = EPOLLIN;
+ ev.data.u64 = open_client(fd);
+ if (ev.data.u64 != 0 && epoll_ctl(epollfd, EPOLL_CTL_ADD, clients[ev.data.u64].fd, &ev) == -1) {
+ pseudo_diag("epoll_ctl() failed with accepted socket.\n");
+ exit(PSEUDO_EXIT_EPOLL_CTL);
+ }
+ } else if (errno == EMFILE) {
+ pseudo_debug(PDBGF_SERVER, "Hit max open files, dropping a client.\n");
+ /* In theory there is a potential race here where if we close a client,
+ it may have sent us a fastop message which we don't act upon.
+ If we don't close a filehandle we'll loop indefinitely thought.
+ Only close one per loop iteration in the interests of caution */
+ for (int j = 1; j <= highest_client; ++j) {
+ if (clients[j].fd != -1) {
+ close_client(j);
+ break;
+ }
+ }
+ }
+ }
+ } else {
+ struct timeval tv1, tv2;
+ int rc;
+ gettimeofday(&tv1, NULL);
+ rc = serve_client(events[i].data.u64);
+ gettimeofday(&tv2, NULL);
+ ++messages;
+ if (rc == 0)
+ ++responses;
+ message_time.tv_sec += (tv2.tv_sec - tv1.tv_sec);
+ message_time.tv_usec += (tv2.tv_usec - tv1.tv_usec);
+ if (message_time.tv_usec < 0) {
+ message_time.tv_usec += 1000000;
+ --message_time.tv_sec;
+ } else while (message_time.tv_usec > 1000000) {
+ message_time.tv_usec -= 1000000;
+ ++message_time.tv_sec;
+ }
+ }
+ if (die_forcefully)
+ break;
+ }
+ pseudo_debug(PDBGF_SERVER, "server loop complete [%d clients left]\n", active_clients);
+ } else {
+ pseudo_diag("epoll_wait failed: %s\n", strerror(errno));
+ break;
+ }
+ if (die_peacefully || die_forcefully) {
+ pseudo_debug(PDBGF_SERVER, "quitting.\n");
+ pseudo_debug(PDBGF_SERVER | PDBGF_BENCHMARK, "server %d exiting: handled %d messages in %.4f seconds\n",
+ getpid(), messages,
+ (double) message_time.tv_sec +
+ (double) message_time.tv_usec / 1000000.0);
+ pdb_log_msg(SEVERITY_INFO, NULL, NULL, NULL, "server %d exiting: handled %d messages in %.4f seconds",
+ getpid(), messages,
+ (double) message_time.tv_sec +
+ (double) message_time.tv_usec / 1000000.0);
+ /* and at this point, we'll start refusing connections */
+ close(clients[0].fd);
+ /* This is a good place to insert a delay for
+ * debugging race conditions during startup. */
+ /* usleep(300000); */
+ exit(0);
+ }
+ }
+
+}
+
+#endif
+
/* get clients, handle messages, shut down.
* This doesn't actually do any work, it just calls a ton of things which
* do work.
--
2.14.1