chiark / gitweb /
[PATCH] udevd: throttle the forking of processes
authorkay.sievers@vrfy.org <kay.sievers@vrfy.org>
Sun, 16 Jan 2005 23:53:08 +0000 (00:53 +0100)
committerGreg KH <gregkh@suse.de>
Wed, 27 Apr 2005 06:21:58 +0000 (23:21 -0700)
If the system reaches a defined limit of processes in running state, udevd
starts to count its own processes in running state from its session (all
forked hotplug child processes, subprocesses and callouts) and throttles
further process forking if the limit is reached.

This should help setups with hundreds of events emitted hotplug events
in parallel with hundreds of processes in "R" state. which makes the machine
unresponsible.

I placed a 100% cpu time consuming program in /etc/hotplug.d/ which runs for 5
seconds. With this patch I can load "scsi_debug add_host=100" without any major
problem. Without the patch the box is completly unresponsible for many minutes.

udevd.c
udevd.h

diff --git a/udevd.c b/udevd.c
index a8328bbb81a60d2a093089d6490ddb8f7e58577b..53c7bf3df1bd0055b46c24d1caaeb02f042f57f3 100644 (file)
--- a/udevd.c
+++ b/udevd.c
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <ctype.h>
+#include <dirent.h>
+#include <fcntl.h>
 #include <sys/time.h>
 #include <sys/types.h>
 #include <sys/socket.h>
 #include <sys/un.h>
 #include <sys/time.h>
 #include <sys/types.h>
 #include <sys/socket.h>
 #include <sys/un.h>
-#include <fcntl.h>
 #include <sys/sysinfo.h>
 #include <sys/stat.h>
 
 #include <sys/sysinfo.h>
 #include <sys/stat.h>
 
@@ -45,6 +47,7 @@
 
 /* global variables*/
 static int udevsendsock;
 
 /* global variables*/
 static int udevsendsock;
+static pid_t sid;
 
 static int pipefds[2];
 static long startup_time;
 
 static int pipefds[2];
 static long startup_time;
@@ -144,6 +147,8 @@ static void udev_run(struct hotplug_msg *msg)
                /* child */
                close(udevsendsock);
                logging_close();
                /* child */
                close(udevsendsock);
                logging_close();
+
+               setpriority(PRIO_PROCESS, 0, UDEV_PRIORITY);
                execve(udev_bin, argv, msg->envp);
                dbg("exec of child failed");
                _exit(1);
                execve(udev_bin, argv, msg->envp);
                dbg("exec of child failed");
                _exit(1);
@@ -151,9 +156,6 @@ static void udev_run(struct hotplug_msg *msg)
        case -1:
                dbg("fork of child failed");
                run_queue_delete(msg);
        case -1:
                dbg("fork of child failed");
                run_queue_delete(msg);
-               /* note: we never managed to run, so we had no impact on 
-                * running_with_devpath(), so don't bother setting run_exec_q
-                */
                break;
        default:
                /* get SIGCHLD in main loop */
                break;
        default:
                /* get SIGCHLD in main loop */
@@ -162,6 +164,100 @@ static void udev_run(struct hotplug_msg *msg)
        }
 }
 
        }
 }
 
+static int running_processes(void)
+{
+       int f;
+       static char buf[4096];
+       int len;
+       int running;
+       const char *pos;
+
+       f = open("/proc/stat", O_RDONLY);
+       if (f == -1)
+               return -1;
+
+       len = read(f, buf, sizeof(buf));
+       close(f);
+
+       if (len <= 0)
+               return -1;
+       else
+               buf[len] = '\0';
+
+       pos = strstr(buf, "procs_running ");
+       if (pos == NULL)
+               return -1;
+
+       if (sscanf(pos, "procs_running %u", &running) != 1)
+               return -1;
+
+       return running;
+}
+
+/* return the number of process es in our session, count only until limit */
+static int running_processes_in_session(pid_t session, int limit)
+{
+       DIR *dir;
+       struct dirent *dent;
+       int running = 0;
+
+       dir = opendir("/proc");
+       if (!dir)
+               return -1;
+
+       /* read process info from /proc */
+       for (dent = readdir(dir); dent != NULL; dent = readdir(dir)) {
+               int f;
+               char procdir[64];
+               char line[256];
+               const char *pos;
+               char state;
+               pid_t ppid, pgrp, sess;
+               int len;
+
+               if (!isdigit(dent->d_name[0]))
+                       continue;
+
+               snprintf(procdir, sizeof(procdir), "/proc/%s/stat", dent->d_name);
+               procdir[sizeof(procdir)-1] = '\0';
+
+               f = open(procdir, O_RDONLY);
+               if (f == -1)
+                       continue;
+
+               len = read(f, line, sizeof(line));
+               close(f);
+
+               if (len <= 0)
+                       continue;
+               else
+                       line[len] = '\0';
+
+               /* skip ugly program name */
+               pos = strrchr(line, ')') + 2;
+               if (pos == NULL)
+                       continue;
+
+               if (sscanf(pos, "%c %d %d %d ", &state, &ppid, &pgrp, &sess) != 4)
+                       continue;
+
+               /* count only processes in our session */
+               if (sess != session)
+                       continue;
+
+               /* count only running, no sleeping processes */
+               if (state != 'R')
+                       continue;
+
+               running++;
+               if (limit > 0 && running >= limit)
+                       break;
+       }
+       closedir(dir);
+
+       return running;
+}
+
 static int compare_devpath(const char *running, const char *waiting)
 {
        int i;
 static int compare_devpath(const char *running, const char *waiting)
 {
        int i;
@@ -218,13 +314,30 @@ static void exec_queue_manager(void)
        struct hotplug_msg *loop_msg;
        struct hotplug_msg *tmp_msg;
        struct hotplug_msg *msg;
        struct hotplug_msg *loop_msg;
        struct hotplug_msg *tmp_msg;
        struct hotplug_msg *msg;
+       int running;
+
+       running = running_processes();
+       dbg("%d processes runnning on system", running);
+       if (running < 0)
+               running = THROTTLE_MAX_RUNNING_CHILDS;
 
        list_for_each_entry_safe(loop_msg, tmp_msg, &exec_list, list) {
 
        list_for_each_entry_safe(loop_msg, tmp_msg, &exec_list, list) {
+               /* check running processes in our session and possibly throttle */
+               if (running >= THROTTLE_MAX_RUNNING_CHILDS) {
+                       running = running_processes_in_session(sid, THROTTLE_MAX_RUNNING_CHILDS+10);
+                       dbg("%d processes running in session", running);
+                       if (running >= THROTTLE_MAX_RUNNING_CHILDS) {
+                               dbg("delay seq %llu, cause too many processes already running", loop_msg->seqnum);
+                               return;
+                       }
+               }
+
                msg = running_with_devpath(loop_msg);
                if (!msg) {
                        /* move event to run list */
                        list_move_tail(&loop_msg->list, &running_list);
                        udev_run(loop_msg);
                msg = running_with_devpath(loop_msg);
                if (!msg) {
                        /* move event to run list */
                        list_move_tail(&loop_msg->list, &running_list);
                        udev_run(loop_msg);
+                       running++;
                        dbg("moved seq %llu to running list", loop_msg->seqnum);
                } else {
                        dbg("delay seq %llu (%s), cause seq %llu (%s) is still running",
                        dbg("moved seq %llu to running list", loop_msg->seqnum);
                } else {
                        dbg("delay seq %llu (%s), cause seq %llu (%s) is still running",
@@ -529,10 +642,17 @@ int main(int argc, char *argv[], char *envp[])
                }
        }
 
                }
        }
 
+       /* become session leader */
+       sid = setsid();
+       dbg("our session is %d", sid);
+
        /* make sure we don't lock any path */
        chdir("/");
        umask(umask(077) | 022);
 
        /* make sure we don't lock any path */
        chdir("/");
        umask(umask(077) | 022);
 
+       /*set a reasonable scheduling priority for the daemon */
+       setpriority(PRIO_PROCESS, 0, UDEVD_PRIORITY);
+
        /* Set fds to dev/null */
        fd = open( "/dev/null", O_RDWR );
        if (fd >= 0)  {
        /* Set fds to dev/null */
        fd = open( "/dev/null", O_RDWR );
        if (fd >= 0)  {
@@ -544,9 +664,6 @@ int main(int argc, char *argv[], char *envp[])
        } else
                dbg("error opening /dev/null %s", strerror(errno));
 
        } else
                dbg("error opening /dev/null %s", strerror(errno));
 
-       /* become session leader */
-       setsid();
-
        /* setup signal handler pipe */
        retval = pipe(pipefds);
        if (retval < 0) {
        /* setup signal handler pipe */
        retval = pipe(pipefds);
        if (retval < 0) {
@@ -597,14 +714,14 @@ int main(int argc, char *argv[], char *envp[])
        else
                udev_bin = UDEV_BIN;
 
        else
                udev_bin = UDEV_BIN;
 
-       /* possible set of expected_seqnum number */
+       /* possible init of expected_seqnum value */
        udevd_expected_seqnum = getenv("UDEVD_EXPECTED_SEQNUM");
        if (udevd_expected_seqnum != NULL) {
                expected_seqnum = strtoull(udevd_expected_seqnum, NULL, 10);
                dbg("initialize expected_seqnum to %llu", expected_seqnum);
        }
 
        udevd_expected_seqnum = getenv("UDEVD_EXPECTED_SEQNUM");
        if (udevd_expected_seqnum != NULL) {
                expected_seqnum = strtoull(udevd_expected_seqnum, NULL, 10);
                dbg("initialize expected_seqnum to %llu", expected_seqnum);
        }
 
-       /* get current time to provide shorter startup timeout */
+       /* get current time to provide shorter timeout on startup */
        sysinfo(&info);
        startup_time = info.uptime;
 
        sysinfo(&info);
        startup_time = info.uptime;
 
diff --git a/udevd.h b/udevd.h
index 28f62cf528ffe66c26b18a6cd220c91504afad10..ae1d20e5ecfd309f6679ae6610b67c9872159177 100644 (file)
--- a/udevd.h
+++ b/udevd.h
 #define UDEVD_SOCK_PATH                        "udevd"
 #define SEND_WAIT_MAX_SECONDS          3
 #define SEND_WAIT_LOOP_PER_SECOND      10
 #define UDEVD_SOCK_PATH                        "udevd"
 #define SEND_WAIT_MAX_SECONDS          3
 #define SEND_WAIT_LOOP_PER_SECOND      10
+
+#define UDEVD_PRIORITY                 -4
+#define UDEV_PRIORITY                  -2
+
+/* duration of initialization phase with shorter timeout */
 #define INIT_TIME_SEC                  5
 #define EVENT_INIT_TIMEOUT_SEC         2
 #define INIT_TIME_SEC                  5
 #define EVENT_INIT_TIMEOUT_SEC         2
+
+/* timeout to wait for missing events */
 #define EVENT_TIMEOUT_SEC              10
 
 #define EVENT_TIMEOUT_SEC              10
 
+/* start to throttle forking if maximum number of running childs in our session is reached */
+#define THROTTLE_MAX_RUNNING_CHILDS    10
+
 /* environment buffer, should match the kernel's size in lib/kobject_uevent.h */
 #define HOTPLUG_BUFFER_SIZE            1024
 #define HOTPLUG_NUM_ENVP               32
 /* environment buffer, should match the kernel's size in lib/kobject_uevent.h */
 #define HOTPLUG_BUFFER_SIZE            1024
 #define HOTPLUG_NUM_ENVP               32