chiark / gitweb /
move /dev/.udev/ to /dev/.run/udev/ and convert old udev database at udevd startup
[elogind.git] / udev / udevd.c
index b48a91391de87b7c6a2eaed5c9593a1dcf0eb3fc..fe0e3c39295724776a71f248eaf13ef4df5f3fab 100644 (file)
@@ -34,6 +34,7 @@
 #include <sys/time.h>
 #include <sys/prctl.h>
 #include <sys/socket.h>
+#include <sys/un.h>
 #include <sys/signalfd.h>
 #include <sys/select.h>
 #include <sys/poll.h>
@@ -44,6 +45,7 @@
 #include <sys/utsname.h>
 
 #include "udev.h"
+#include "sd-daemon.h"
 
 #define UDEVD_PRIORITY                 -4
 #define UDEV_PRIORITY                  -2
@@ -69,7 +71,6 @@ static void log_fn(struct udev *udev, int priority,
        }
 }
 
-static bool debug_trace;
 static struct udev_rules *rules;
 static struct udev_queue_export *udev_queue_export;
 static struct udev_ctrl *udev_ctrl;
@@ -78,8 +79,9 @@ static int worker_watch[2];
 static pid_t settle_pid;
 static bool stop_exec_queue;
 static bool reload_config;
-static int max_childs;
-static int childs;
+static int children;
+static int children_max;
+static int exec_delay;
 static sigset_t orig_sigmask;
 static struct udev_list_node event_list;
 static struct udev_list_node worker_list;
@@ -121,6 +123,7 @@ struct event {
        const char *devpath_old;
        dev_t devnum;
        bool is_block;
+       int ifindex;
 };
 
 static struct event *node_to_event(struct udev_list_node *node)
@@ -205,7 +208,7 @@ static void worker_unref(struct worker *worker)
 
        udev_list_node_remove(&worker->node);
        udev_monitor_unref(worker->monitor);
-       childs--;
+       children--;
        info(worker->udev, "worker [%u] cleaned up\n", worker->pid);
        free(worker);
 }
@@ -226,8 +229,10 @@ static void worker_new(struct event *event)
        udev_monitor_enable_receiving(worker_monitor);
 
        worker = calloc(1, sizeof(struct worker));
-       if (worker == NULL)
+       if (worker == NULL) {
+               udev_monitor_unref(worker_monitor);
                return;
+       }
        /* worker + event reference */
        worker->refcount = 2;
        worker->udev = event->udev;
@@ -286,6 +291,9 @@ static void worker_new(struct event *event)
                        /* set timeout to prevent hanging processes */
                        alarm(UDEV_EVENT_TIMEOUT);
 
+                       if (exec_delay > 0)
+                               udev_event->exec_delay = exec_delay;
+
                        /* apply rules, create node, symlinks */
                        err = udev_event_execute_rules(udev_event, rules);
 
@@ -355,7 +363,7 @@ static void worker_new(struct event *event)
                worker->event = event;
                event->state = EVENT_RUNNING;
                udev_list_node_append(&worker->node, &worker_list);
-               childs++;
+               children++;
                info(event->udev, "seq %llu forked new worker [%u]\n", udev_device_get_seqnum(event->dev), pid);
                break;
        }
@@ -386,8 +394,9 @@ static void event_run(struct event *event, bool force)
                return;
        }
 
-       if (!force && childs >= max_childs) {
-               info(event->udev, "maximum number (%i) of childs reached\n", childs);
+       if (!force && children >= children_max) {
+               if (children_max > 1)
+                       info(event->udev, "maximum number (%i) of children reached\n", children);
                return;
        }
 
@@ -395,13 +404,13 @@ static void event_run(struct event *event, bool force)
        worker_new(event);
 }
 
-static void event_queue_insert(struct udev_device *dev)
+static int event_queue_insert(struct udev_device *dev)
 {
        struct event *event;
 
        event = calloc(1, sizeof(struct event));
        if (event == NULL)
-               return;
+               return -1;
 
        event->udev = udev_device_get_udev(dev);
        event->dev = dev;
@@ -411,6 +420,7 @@ static void event_queue_insert(struct udev_device *dev)
        event->devpath_old = udev_device_get_devpath_old(dev);
        event->devnum = udev_device_get_devnum(dev);
        event->is_block = (strcmp("block", udev_device_get_subsystem(dev)) == 0);
+       event->ifindex = udev_device_get_ifindex(dev);
 
        udev_queue_export_device_queued(udev_queue_export, dev);
        info(event->udev, "seq %llu queued, '%s' '%s'\n", udev_device_get_seqnum(dev),
@@ -422,8 +432,10 @@ static void event_queue_insert(struct udev_device *dev)
        /* run all events with a timeout set immediately */
        if (udev_device_get_timeout(dev) > 0) {
                event_run(event, true);
-               return;
+               return 0;
        }
+
+       return 0;
 }
 
 static void worker_kill(struct udev *udev, int retain)
@@ -431,10 +443,10 @@ static void worker_kill(struct udev *udev, int retain)
        struct udev_list_node *loop;
        int max;
 
-       if (childs <= retain)
+       if (children <= retain)
                return;
 
-       max = childs - retain;
+       max = children - retain;
 
        udev_list_node_foreach(loop, &worker_list) {
                struct worker *worker = node_to_worker(loop);
@@ -476,6 +488,10 @@ static bool is_devpath_busy(struct event *event)
                if (major(event->devnum) != 0 && event->devnum == loop_event->devnum && event->is_block == loop_event->is_block)
                        return true;
 
+               /* check network device ifindex */
+               if (event->ifindex != 0 && event->ifindex == loop_event->ifindex)
+                       return true;
+
                /* check our old name */
                if (event->devpath_old != NULL && strcmp(loop_event->devpath, event->devpath_old) == 0) {
                        event->delaying_seqnum = loop_event->seqnum;
@@ -491,6 +507,11 @@ static bool is_devpath_busy(struct event *event)
 
                /* identical device event found */
                if (loop_event->devpath_len == event->devpath_len) {
+                       /* devices names might have changed/swapped in the meantime */
+                       if (major(event->devnum) != 0 && (event->devnum != loop_event->devnum || event->is_block != loop_event->is_block))
+                               continue;
+                       if (event->ifindex != 0 && event->ifindex != loop_event->ifindex)
+                               continue;
                        event->delaying_seqnum = loop_event->seqnum;
                        return true;
                }
@@ -625,10 +646,10 @@ static void handle_ctrl_msg(struct udev_ctrl *uctrl)
                worker_kill(udev, 0);
        }
 
-       i = udev_ctrl_get_set_max_childs(ctrl_msg);
+       i = udev_ctrl_get_set_children_max(ctrl_msg);
        if (i >= 0) {
-               info(udev, "udevd message (SET_MAX_CHILDS) received, max_childs=%i\n", i);
-               max_childs = i;
+               info(udev, "udevd message (SET_MAX_CHILDREN) received, children_max=%i\n", i);
+               children_max = i;
        }
 
        settle_pid = udev_ctrl_get_settle(ctrl_msg);
@@ -891,10 +912,14 @@ static void static_dev_create_links(struct udev *udev, DIR *dir)
        unsigned int i;
 
        for (i = 0; i < ARRAY_SIZE(stdlinks); i++) {
-               udev_selinux_setfscreateconat(udev, dirfd(dir), stdlinks[i].link, S_IFLNK);
-               if (symlinkat(stdlinks[i].target, dirfd(dir), stdlinks[i].link) < 0 && errno == EEXIST)
-                       utimensat(dirfd(dir), stdlinks[i].link, NULL, AT_SYMLINK_NOFOLLOW);
-               udev_selinux_resetfscreatecon(udev);
+               struct stat sb;
+
+               if (stat(stdlinks[i].target, &sb) == 0) {
+                       udev_selinux_setfscreateconat(udev, dirfd(dir), stdlinks[i].link, S_IFLNK);
+                       if (symlinkat(stdlinks[i].target, dirfd(dir), stdlinks[i].link) < 0 && errno == EEXIST)
+                               utimensat(dirfd(dir), stdlinks[i].link, NULL, AT_SYMLINK_NOFOLLOW);
+                       udev_selinux_resetfscreatecon(udev);
+               }
        }
 }
 
@@ -946,22 +971,115 @@ static int mem_size_mb(void)
        return memsize;
 }
 
+static int convert_db(struct udev *udev)
+{
+       char filename[UTIL_PATH_SIZE];
+       FILE *f;
+       struct udev_enumerate *udev_enumerate;
+       struct udev_list_entry *list_entry;
+
+       /* current database */
+       util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.run/udev/db3", NULL);
+       if (access(filename, F_OK) >= 0)
+               return 0;
+
+       /* make sure we do not get here again */
+       util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.run/udev/db3/", NULL);
+       util_create_path(udev, filename);
+
+       /* old database */
+       util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.udev/db", NULL);
+       if (access(filename, F_OK) < 0)
+               return 0;
+
+       f = fopen("/dev/kmsg", "w");
+       if (f != NULL) {
+               fprintf(f, "<6>udev[%u]: converting old udev database\n", getpid());
+               fclose(f);
+       }
+
+       udev_enumerate = udev_enumerate_new(udev);
+       if (udev_enumerate == NULL)
+               return -1;
+       udev_enumerate_scan_devices(udev_enumerate);
+       udev_list_entry_foreach(list_entry, udev_enumerate_get_list_entry(udev_enumerate)) {
+               struct udev_device *device;
+
+               device = udev_device_new_from_syspath(udev, udev_list_entry_get_name(list_entry));
+               if (device == NULL)
+                       continue;
+
+               /* try to find the old database for devices without a current one */
+               if (udev_device_read_db(device, NULL) < 0) {
+                       bool have_db;
+                       const char *id;
+                       struct stat stats;
+                       char devpath[UTIL_PATH_SIZE];
+                       char from[UTIL_PATH_SIZE];
+
+                       have_db = false;
+
+                       /* find database in old location */
+                       id = udev_device_get_id_filename(device);
+                       util_strscpyl(from, sizeof(from), udev_get_dev_path(udev), "/.udev/db/", id, NULL);
+                       if (lstat(from, &stats) == 0) {
+                               if (!have_db) {
+                                       udev_device_read_db(device, from);
+                                       have_db = true;
+                               }
+                               unlink(from);
+                       }
+
+                       /* find old database with $subsys:$sysname name */
+                       util_strscpyl(from, sizeof(from), udev_get_dev_path(udev),
+                                    "/.udev/db/", udev_device_get_subsystem(device), ":",
+                                    udev_device_get_sysname(device), NULL);
+                       if (lstat(from, &stats) == 0) {
+                               if (!have_db) {
+                                       udev_device_read_db(device, from);
+                                       have_db = true;
+                               }
+                               unlink(from);
+                       }
+
+                       /* find old database with the encoded devpath name */
+                       util_path_encode(udev_device_get_devpath(device), devpath, sizeof(devpath));
+                       util_strscpyl(from, sizeof(from), udev_get_dev_path(udev),
+                                     "/.udev/db/", devpath, NULL);
+                       if (lstat(from, &stats) == 0) {
+                               if (!have_db) {
+                                       udev_device_read_db(device, from);
+                                       have_db = true;
+                               }
+                               unlink(from);
+                       }
+
+                       /* write out new database */
+                       if (have_db)
+                               udev_device_update_db(device);
+               }
+               udev_device_unref(device);
+       }
+       udev_enumerate_unref(udev_enumerate);
+       return 0;
+}
+
 int main(int argc, char *argv[])
 {
        struct udev *udev;
        int fd;
        FILE *f;
        sigset_t mask;
-       const char *value;
        int daemonize = false;
        int resolve_names = 1;
        static const struct option options[] = {
                { "daemon", no_argument, NULL, 'd' },
-               { "debug-trace", no_argument, NULL, 't' },
                { "debug", no_argument, NULL, 'D' },
+               { "children-max", required_argument, NULL, 'c' },
+               { "exec-delay", required_argument, NULL, 'e' },
+               { "resolve-names", required_argument, NULL, 'N' },
                { "help", no_argument, NULL, 'h' },
                { "version", no_argument, NULL, 'V' },
-               { "resolve-names", required_argument, NULL, 'N' },
                {}
        };
        int rc = 1;
@@ -978,7 +1096,7 @@ int main(int argc, char *argv[])
        for (;;) {
                int option;
 
-               option = getopt_long(argc, argv, "dDthV", options, NULL);
+               option = getopt_long(argc, argv, "c:deDthV", options, NULL);
                if (option == -1)
                        break;
 
@@ -986,8 +1104,11 @@ int main(int argc, char *argv[])
                case 'd':
                        daemonize = true;
                        break;
-               case 't':
-                       debug_trace = true;
+               case 'c':
+                       children_max = strtoul(optarg, NULL, 0);
+                       break;
+               case 'e':
+                       exec_delay = strtoul(optarg, NULL, 0);
                        break;
                case 'D':
                        debug = true;
@@ -1008,8 +1129,15 @@ int main(int argc, char *argv[])
                        }
                        break;
                case 'h':
-                       printf("Usage: udevd [--help] [--daemon] [--debug-trace] [--debug] "
-                              "[--resolve-names=early|late|never] [--version]\n");
+                       printf("Usage: udevd OPTIONS\n"
+                              "  --daemon\n"
+                              "  --debug\n"
+                              "  --children-max=<maximum number of workers>\n"
+                              "  --exec-delay=<seconds to wait before executing RUN=>\n"
+                              "  --resolve-names=early|late|never\n" 
+                              "  --version\n"
+                              "  --help\n"
+                              "\n");
                        goto exit;
                case 'V':
                        printf("%s\n", VERSION);
@@ -1019,13 +1147,55 @@ int main(int argc, char *argv[])
                }
        }
 
+       /*
+        * read the kernel commandline, in case we need to get into debug mode
+        *   udev.log-priority=<level>              syslog priority
+        *   udev.children-max=<number of workers>  events are fully serialized if set to 1
+        *
+        */
+       f = fopen("/proc/cmdline", "r");
+       if (f != NULL) {
+               char cmdline[4096];
+
+               if (fgets(cmdline, sizeof(cmdline), f) != NULL) {
+                       char *pos;
+
+                       pos = strstr(cmdline, "udev.log-priority=");
+                       if (pos != NULL) {
+                               pos += strlen("udev.log-priority=");
+                               udev_set_log_priority(udev, util_log_priority(pos));
+                       }
+
+                       pos = strstr(cmdline, "udev.children-max=");
+                       if (pos != NULL) {
+                               pos += strlen("udev.children-max=");
+                               children_max = strtoul(pos, NULL, 0);
+                       }
+
+                       pos = strstr(cmdline, "udev.exec-delay=");
+                       if (pos != NULL) {
+                               pos += strlen("udev.exec-delay=");
+                               exec_delay = strtoul(pos, NULL, 0);
+                       }
+               }
+               fclose(f);
+       }
+
        if (getuid() != 0) {
                fprintf(stderr, "root privileges required\n");
                err(udev, "root privileges required\n");
                goto exit;
        }
 
-       /* make sure std{in,out,err} fd's are in a sane state */
+       /* set umask before creating any file/directory */
+       chdir("/");
+       umask(022);
+
+       /* create standard links, copy static nodes, create nodes from modules */
+       static_dev_create(udev);
+       static_dev_create_from_modules(udev);
+
+       /* before opening new files, make sure std{in,out,err} fds are in a sane state */
        fd = open("/dev/null", O_RDWR);
        if (fd < 0) {
                fprintf(stderr, "cannot open /dev/null\n");
@@ -1036,7 +1206,6 @@ int main(int argc, char *argv[])
        if (write(STDERR_FILENO, 0, 0) < 0)
                dup2(fd, STDERR_FILENO);
 
-       /* init control socket, bind() ensures, that only one udevd instance is running */
        udev_ctrl = udev_ctrl_new_from_socket(udev, UDEV_CTRL_SOCK_PATH);
        if (udev_ctrl == NULL) {
                fprintf(stderr, "error initializing control socket");
@@ -1083,7 +1252,7 @@ int main(int argc, char *argv[])
                                  IN_DELETE | IN_MOVE | IN_CLOSE_WRITE);
 
                /* watch dynamic rules directory */
-               util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.udev/rules.d", NULL);
+               util_strscpyl(filename, sizeof(filename), udev_get_dev_path(udev), "/.run/udev/rules.d", NULL);
                if (stat(filename, &statbuf) != 0) {
                        util_create_path(udev, filename);
                        udev_selinux_setfscreatecon(udev, filename, S_IFDIR|0755);
@@ -1127,6 +1296,17 @@ int main(int argc, char *argv[])
                goto exit;
        }
 
+       /* if needed, convert old database from earlier udev version */
+       convert_db(udev);
+
+       if (!debug) {
+               dup2(fd, STDIN_FILENO);
+               dup2(fd, STDOUT_FILENO);
+               dup2(fd, STDERR_FILENO);
+       }
+       if (fd > STDERR_FILENO)
+               close(fd);
+
        if (daemonize) {
                pid_t pid;
 
@@ -1142,59 +1322,48 @@ int main(int argc, char *argv[])
                        rc = 0;
                        goto exit;
                }
+       } else {
+               sd_notify(1, "READY=1");
        }
 
+       /* set scheduling priority for the main daemon process */
+       setpriority(PRIO_PROCESS, 0, UDEVD_PRIORITY);
+
+       setsid();
+
        f = fopen("/dev/kmsg", "w");
        if (f != NULL) {
-               fprintf(f, "<6>udev: starting version " VERSION "\n");
+               fprintf(f, "<6>udev[%u]: starting version " VERSION "\n", getpid());
                fclose(f);
        }
 
-       /* redirect std{out,err} */
-       if (!debug && !debug_trace) {
-               dup2(fd, STDIN_FILENO);
-               dup2(fd, STDOUT_FILENO);
-               dup2(fd, STDERR_FILENO);
-       }
-       if (fd > STDERR_FILENO)
-               close(fd);
-
-       /* set scheduling priority for the daemon */
-       setpriority(PRIO_PROCESS, 0, UDEVD_PRIORITY);
-
-       chdir("/");
-       umask(022);
-       setsid();
-
-       /* OOM_DISABLE == -17 */
-       fd = open("/proc/self/oom_adj", O_RDWR);
+       fd = open("/proc/self/oom_score_adj", O_RDWR);
        if (fd < 0) {
-               err(udev, "error disabling OOM: %m\n");
+               /* Fallback to old interface */
+               fd = open("/proc/self/oom_adj", O_RDWR);
+               if (fd < 0) {
+                       err(udev, "error disabling OOM: %m\n");
+               } else {
+                       /* OOM_DISABLE == -17 */
+                       write(fd, "-17", 3);
+                       close(fd);
+               }
        } else {
-               write(fd, "-17", 3);
+               write(fd, "-1000", 5);
                close(fd);
        }
 
-       /* in trace mode run one event after the other */
-       if (debug_trace) {
-               max_childs = 1;
-       } else {
+       if (children_max <= 0) {
                int memsize = mem_size_mb();
 
+               /* set value depending on the amount of RAM */
                if (memsize > 0)
-                       max_childs = 128 + (memsize / 8);
+                       children_max = 128 + (memsize / 8);
                else
-                       max_childs = 128;
+                       children_max = 128;
        }
+       info(udev, "set children_max to %u\n", children_max);
 
-       /* possibly overwrite maximum limit of executed events */
-       value = getenv("UDEVD_MAX_CHILDS");
-       if (value)
-               max_childs = strtoul(value, NULL, 10);
-       info(udev, "initialize max_childs to %u\n", max_childs);
-
-       static_dev_create(udev);
-       static_dev_create_from_modules(udev);
        udev_rules_apply_static_dev_perms(rules);
 
        udev_list_init(&event_list);
@@ -1205,7 +1374,7 @@ int main(int argc, char *argv[])
                int timeout;
 
                /* set timeout to kill idle workers */
-               if (udev_list_is_empty(&event_list) && childs > 2)
+               if (udev_list_is_empty(&event_list) && children > 2)
                        timeout = 3 * 1000;
                else
                        timeout = -1;
@@ -1228,9 +1397,8 @@ int main(int argc, char *argv[])
 
                        dev = udev_monitor_receive_device(monitor);
                        if (dev != NULL)
-                               event_queue_insert(dev);
-                       else
-                               udev_device_unref(dev);
+                               if (event_queue_insert(dev) < 0)
+                                       udev_device_unref(dev);
                }
 
                /* start new events */