X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ianmdlvl/git?p=elogind.git;a=blobdiff_plain;f=src%2Fjournal%2Fjournald.c;h=5dc5c95cfa9b59e7591e4a1d56a7810c83f21441;hp=c29d75c9c4045a22af163f3fbe5a689d87eb30f2;hb=671e021c92c835c6c701dc61463149d05b6f31af;hpb=32fca2d5b746b86bbbbd4f1b57064076cfc2ae4b

diff --git a/src/journal/journald.c b/src/journal/journald.c
index c29d75c9c..5dc5c95cf 100644
--- a/src/journal/journald.c
+++ b/src/journal/journald.c
@@ -29,7 +29,9 @@
 #include <sys/ioctl.h>
 #include <linux/sockios.h>
 #include <sys/statvfs.h>
+#include <sys/mman.h>
 
+#include <libudev.h>
 #include <systemd/sd-journal.h>
 #include <systemd/sd-messages.h>
 #include <systemd/sd-daemon.h>
@@ -46,6 +48,8 @@
 #include "list.h"
 #include "journal-rate-limit.h"
 #include "journal-internal.h"
+#include "journal-vacuum.h"
+#include "journal-authenticate.h"
 #include "conf-parser.h"
 #include "journald.h"
 #include "virt.h"
@@ -70,6 +74,8 @@
 #define RECHECK_AVAILABLE_SPACE_USEC (30*USEC_PER_SEC)
 
 #define N_IOVEC_META_FIELDS 17
+#define N_IOVEC_KERNEL_FIELDS 64
+#define N_IOVEC_UDEV_FIELDS 32
 
 #define ENTRY_SIZE_MAX (1024*1024*32)
 
@@ -279,7 +285,6 @@ static JournalFile* find_journal(Server *s, uid_t uid) {
         char *p;
         int r;
         JournalFile *f;
-        char ids[33];
         sd_id128_t machine;
 
         assert(s);
@@ -303,7 +308,8 @@ static JournalFile* find_journal(Server *s, uid_t uid) {
         if (f)
                 return f;
 
-        if (asprintf(&p, "/var/log/journal/%s/user-%lu.journal", sd_id128_to_string(machine, ids), (unsigned long) uid) < 0)
+        if (asprintf(&p, "/var/log/journal/" SD_ID128_FORMAT_STR "/user-%lu.journal",
+                     SD_ID128_FORMAT_VAL(machine), (unsigned long) uid) < 0)
                 return s->system_journal;
 
         while (hashmap_size(s->user_journals) >= USER_JOURNALS_MAX) {
@@ -313,7 +319,7 @@ static JournalFile* find_journal(Server *s, uid_t uid) {
                 journal_file_close(f);
         }
 
-        r = journal_file_open_reliably(p, O_RDWR|O_CREAT, 0640, &s->system_metrics, s->system_journal, &f);
+        r = journal_file_open_reliably(p, O_RDWR|O_CREAT, 0640, s->compress, s->seal, &s->system_metrics, s->mmap, s->system_journal, &f);
         free(p);
 
         if (r < 0)
@@ -339,7 +345,7 @@ static void server_rotate(Server *s) {
         log_info("Rotating...");
 
         if (s->runtime_journal) {
-                r = journal_file_rotate(&s->runtime_journal);
+                r = journal_file_rotate(&s->runtime_journal, s->compress, false);
                 if (r < 0)
                         if (s->runtime_journal)
                                 log_error("Failed to rotate %s: %s", s->runtime_journal->path, strerror(-r));
@@ -350,7 +356,7 @@ static void server_rotate(Server *s) {
         }
 
         if (s->system_journal) {
-                r = journal_file_rotate(&s->system_journal);
+                r = journal_file_rotate(&s->system_journal, s->compress, s->seal);
                 if (r < 0)
                         if (s->system_journal)
                                 log_error("Failed to rotate %s: %s", s->system_journal->path, strerror(-r));
@@ -362,7 +368,7 @@ static void server_rotate(Server *s) {
         }
 
         HASHMAP_FOREACH_KEY(f, k, s->user_journals, i) {
-                r = journal_file_rotate(&f);
+                r = journal_file_rotate(&f, s->compress, s->seal);
                 if (r < 0)
                         if (f->path)
                                 log_error("Failed to rotate %s: %s", f->path, strerror(-r));
@@ -393,7 +399,7 @@ static void server_vacuum(Server *s) {
 
         if (s->system_journal) {
                 if (asprintf(&p, "/var/log/journal/%s", ids) < 0) {
-                        log_error("Out of memory.");
+                        log_oom();
                         return;
                 }
 
@@ -405,7 +411,7 @@ static void server_vacuum(Server *s) {
 
         if (s->runtime_journal) {
                 if (asprintf(&p, "/run/log/journal/%s", ids) < 0) {
-                        log_error("Out of memory.");
+                        log_oom();
                         return;
                 }
 
@@ -478,6 +484,10 @@ static void write_to_journal(Server *s, uid_t uid, struct iovec *iovec, unsigned
                 server_rotate(s);
                 server_vacuum(s);
                 vacuumed = true;
+
+                f = find_journal(s, uid);
+                if (!f)
+                        return;
         }
 
         for (;;) {
@@ -717,7 +727,7 @@ static void driver_message(Server *s, sd_id128_t message_id, const char *format,
         assert(s);
         assert(format);
 
-        IOVEC_SET_STRING(iovec[n++], "PRIORITY=5");
+        IOVEC_SET_STRING(iovec[n++], "PRIORITY=6");
         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=driver");
 
         memcpy(buffer, "MESSAGE=", 8);
@@ -864,7 +874,8 @@ static void forward_syslog_iovec(Server *s, const struct iovec *iovec, unsigned
                         return;
         }
 
-        log_debug("Failed to forward syslog message: %m");
+        if (errno != ENOENT)
+                log_debug("Failed to forward syslog message: %m");
 }
 
 static void forward_syslog_raw(Server *s, int priority, const char *buffer, struct ucred *ucred, struct timeval *tv) {
@@ -950,14 +961,16 @@ static void forward_kmsg(Server *s, int priority, const char *identifier, const
         char header_priority[6], header_pid[16];
         int n = 0;
         char *ident_buf = NULL;
-        int fd;
 
         assert(s);
         assert(priority >= 0);
         assert(priority <= 999);
         assert(message);
 
-        if (LOG_PRI(priority) > s->max_level_kmsg)
+        if (_unlikely_(LOG_PRI(priority) > s->max_level_kmsg))
+                return;
+
+        if (_unlikely_(s->dev_kmsg_fd < 0))
                 return;
 
         /* Never allow messages with kernel facility to be written to
@@ -992,18 +1005,9 @@ static void forward_kmsg(Server *s, int priority, const char *identifier, const
         IOVEC_SET_STRING(iovec[n++], message);
         IOVEC_SET_STRING(iovec[n++], "\n");
 
-        fd = open("/dev/kmsg", O_WRONLY|O_NOCTTY|O_CLOEXEC);
-        if (fd < 0) {
-                log_debug("Failed to open /dev/kmsg for logging: %s", strerror(errno));
-                goto finish;
-        }
-
-        if (writev(fd, iovec, n) < 0)
+        if (writev(s->dev_kmsg_fd, iovec, n) < 0)
                 log_debug("Failed to write to /dev/kmsg for logging: %s", strerror(errno));
 
-        close_nointr_nofail(fd);
-
-finish:
         free(ident_buf);
 }
 
@@ -1169,6 +1173,7 @@ static void process_syslog_message(Server *s, const char *buf, struct ucred *ucr
         free(syslog_priority);
         free(syslog_facility);
         free(syslog_identifier);
+        free(syslog_pid);
 }
 
 static bool valid_user_field(const char *p, size_t l) {
@@ -1265,7 +1270,7 @@ static void process_native_message(
                         u = MAX((n+N_IOVEC_META_FIELDS+1) * 2U, 4U);
                         c = realloc(iovec, u * sizeof(struct iovec));
                         if (!c) {
-                                log_error("Out of memory");
+                                log_oom();
                                 break;
                         }
 
@@ -1352,7 +1357,7 @@ static void process_native_message(
 
                         k = malloc((e - p) + 1 + l);
                         if (!k) {
-                                log_error("Out of memory");
+                                log_oom();
                                 break;
                         }
 
@@ -1445,7 +1450,7 @@ static void process_native_file(
 
         p = malloc(st.st_size);
         if (!p) {
-                log_error("Out of memory");
+                log_oom();
                 return;
         }
 
@@ -1537,10 +1542,8 @@ static int stdout_stream_line(StdoutStream *s, char *p) {
                         s->identifier = NULL;
                 else  {
                         s->identifier = strdup(p);
-                        if (!s->identifier) {
-                                log_error("Out of memory");
-                                return -ENOMEM;
-                        }
+                        if (!s->identifier)
+                                return log_oom();
                 }
 
                 s->state = STDOUT_STREAM_UNIT_ID;
@@ -1552,10 +1555,8 @@ static int stdout_stream_line(StdoutStream *s, char *p) {
                                 s->unit_id = NULL;
                         else  {
                                 s->unit_id = strdup(p);
-                                if (!s->unit_id) {
-                                        log_error("Out of memory");
-                                        return -ENOMEM;
-                                }
+                                if (!s->unit_id)
+                                        return log_oom();
                         }
                 }
 
@@ -1756,9 +1757,8 @@ static int stdout_stream_new(Server *s) {
 
         stream = new0(StdoutStream, 1);
         if (!stream) {
-                log_error("Out of memory.");
                 close_nointr_nofail(fd);
-                return -ENOMEM;
+                return log_oom();
         }
 
         stream->fd = fd;
@@ -1801,110 +1801,195 @@ fail:
         return r;
 }
 
-static int parse_kernel_timestamp(char **_p, usec_t *t) {
-        usec_t r;
-        int k, i;
-        char *p;
+static bool is_us(const char *pid) {
+        pid_t t;
 
-        assert(_p);
-        assert(*_p);
-        assert(t);
+        assert(pid);
 
-        p = *_p;
+        if (parse_pid(pid, &t) < 0)
+                return false;
 
-        if (strlen(p) < 14 || p[0] != '[' || p[13] != ']' || p[6] != '.')
-                return 0;
+        return t == getpid();
+}
 
-        r = 0;
+static void dev_kmsg_record(Server *s, char *p, size_t l) {
+        struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
+        char *message = NULL, *syslog_priority = NULL, *syslog_pid = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *source_time = NULL;
+        int priority, r;
+        unsigned n = 0, z = 0, j;
+        usec_t usec;
+        char *identifier = NULL, *pid = NULL, *e, *f, *k;
+        uint64_t serial;
+        size_t pl;
+        char *kernel_device = NULL;
 
-        for (i = 1; i <= 5; i++) {
-                r *= 10;
+        assert(s);
+        assert(p);
 
-                if (p[i] == ' ')
-                        continue;
+        if (l <= 0)
+                return;
 
-                k = undecchar(p[i]);
-                if (k < 0)
-                        return 0;
+        e = memchr(p, ',', l);
+        if (!e)
+                return;
+        *e = 0;
 
-                r += k;
-        }
+        r = safe_atoi(p, &priority);
+        if (r < 0 || priority < 0 || priority > 999)
+                return;
 
-        for (i = 7; i <= 12; i++) {
-                r *= 10;
+        if (s->forward_to_kmsg && (priority & LOG_FACMASK) != LOG_KERN)
+                return;
 
-                k = undecchar(p[i]);
-                if (k < 0)
-                        return 0;
+        l -= (e - p) + 1;
+        p = e + 1;
+        e = memchr(p, ',', l);
+        if (!e)
+                return;
+        *e = 0;
 
-                r += k;
+        r = safe_atou64(p, &serial);
+        if (r < 0)
+                return;
+
+        if (s->kernel_seqnum) {
+                /* We already read this one? */
+                if (serial < *s->kernel_seqnum)
+                        return;
+
+                /* Did we lose any? */
+                if (serial > *s->kernel_seqnum)
+                        driver_message(s, SD_MESSAGE_JOURNAL_MISSED, "Missed %llu kernel messages", (unsigned long long) serial - *s->kernel_seqnum - 1);
+
+                /* Make sure we never read this one again. Note that
+                 * we always store the next message serial we expect
+                 * here, simply because this makes handling the first
+                 * message with serial 0 easy. */
+                *s->kernel_seqnum = serial + 1;
         }
 
-        *t = r;
-        *_p += 14;
-        *_p += strspn(*_p, WHITESPACE);
+        l -= (e - p) + 1;
+        p = e + 1;
+        f = memchr(p, ';', l);
+        if (!f)
+                return;
+        /* Kernel 3.6 has the flags field, kernel 3.5 lacks that */
+        e = memchr(p, ',', l);
+        if (!e || f < e)
+                e = f;
+        *e = 0;
+
+        r = parse_usec(p, &usec);
+        if (r < 0)
+                return;
 
-        return 1;
-}
+        l -= (f - p) + 1;
+        p = f + 1;
+        e = memchr(p, '\n', l);
+        if (!e)
+                return;
+        *e = 0;
 
-static bool is_us(const char *pid) {
-        pid_t t;
+        pl = e - p;
+        l -= (e - p) + 1;
+        k = e + 1;
 
-        assert(pid);
+        for (j = 0; l > 0 && j < N_IOVEC_KERNEL_FIELDS; j++) {
+                char *m;
+                /* Meta data fields attached */
 
-        if (parse_pid(pid, &t) < 0)
-                return false;
+                if (*k != ' ')
+                        break;
 
-        return t == getpid();
-}
+                k ++, l --;
 
-static void proc_kmsg_line(Server *s, const char *p) {
-        struct iovec iovec[N_IOVEC_META_FIELDS + 7];
-        char *message = NULL, *syslog_priority = NULL, *syslog_pid = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *source_time = NULL;
-        int priority = LOG_KERN | LOG_INFO;
-        unsigned n = 0;
-        usec_t usec;
-        char *identifier = NULL, *pid = NULL;
+                e = memchr(k, '\n', l);
+                if (!e)
+                        return;
 
-        assert(s);
-        assert(p);
+                *e = 0;
 
-        if (isempty(p))
-                return;
+                m = cunescape_length_with_prefix(k, e - k, "_KERNEL_");
+                if (!m)
+                        break;
 
-        parse_syslog_priority((char **) &p, &priority);
+                if (startswith(m, "_KERNEL_DEVICE="))
+                        kernel_device = m + 15;
 
-        if (s->forward_to_kmsg && (priority & LOG_FACMASK) != LOG_KERN)
-                return;
+                IOVEC_SET_STRING(iovec[n++], m);
+                z++;
 
-        if (parse_kernel_timestamp((char **) &p, &usec) > 0) {
-                if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu",
-                             (unsigned long long) usec) >= 0)
-                        IOVEC_SET_STRING(iovec[n++], source_time);
+                l -= (e - k) + 1;
+                k = e + 1;
+        }
+
+        if (kernel_device) {
+                struct udev_device *ud;
+
+                ud = udev_device_new_from_device_id(s->udev, kernel_device);
+                if (ud) {
+                        const char *g;
+                        struct udev_list_entry *ll;
+                        char *b;
+
+                        g = udev_device_get_devnode(ud);
+                        if (g) {
+                                b = strappend("_UDEV_DEVNODE=", g);
+                                if (b) {
+                                        IOVEC_SET_STRING(iovec[n++], b);
+                                        z++;
+                                }
+                        }
+
+                        g = udev_device_get_sysname(ud);
+                        if (g) {
+                                b = strappend("_UDEV_SYSNAME=", g);
+                                if (b) {
+                                        IOVEC_SET_STRING(iovec[n++], b);
+                                        z++;
+                                }
+                        }
+
+                        j = 0;
+                        ll = udev_device_get_devlinks_list_entry(ud);
+                        udev_list_entry_foreach(ll, ll) {
+
+                                if (j > N_IOVEC_UDEV_FIELDS)
+                                        break;
+
+                                g = udev_list_entry_get_name(ll);
+                                b = strappend("_UDEV_DEVLINK=", g);
+                                if (g) {
+                                        IOVEC_SET_STRING(iovec[n++], b);
+                                        z++;
+                                }
+
+                                j++;
+                        }
+
+                        udev_device_unref(ud);
+                }
         }
 
+        if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu",
+                     (unsigned long long) usec) >= 0)
+                IOVEC_SET_STRING(iovec[n++], source_time);
+
         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=kernel");
 
         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
 
-        if ((priority & LOG_FACMASK) == LOG_KERN) {
-
-                if (s->forward_to_syslog)
-                        forward_syslog(s, priority, "kernel", p, NULL, NULL);
-
+        if ((priority & LOG_FACMASK) == LOG_KERN)
                 IOVEC_SET_STRING(iovec[n++], "SYSLOG_IDENTIFIER=kernel");
-        } else {
-                read_identifier(&p, &identifier, &pid);
+        else {
+                read_identifier((const char**) &p, &identifier, &pid);
 
                 /* Avoid any messages we generated ourselves via
                  * log_info() and friends. */
                 if (pid && is_us(pid))
                         goto finish;
 
-                if (s->forward_to_syslog)
-                        forward_syslog(s, priority, identifier, p, NULL, NULL);
-
                 if (identifier) {
                         syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
                         if (syslog_identifier)
@@ -1921,13 +2006,16 @@ static void proc_kmsg_line(Server *s, const char *p) {
                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
         }
 
-        message = strappend("MESSAGE=", p);
+        message = cunescape_length_with_prefix(p, pl, "MESSAGE=");
         if (message)
                 IOVEC_SET_STRING(iovec[n++], message);
 
         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), NULL, NULL, NULL, 0, NULL, priority);
 
 finish:
+        for (j = 0; j < z; j++)
+                free(iovec[j].iov_base);
+
         free(message);
         free(syslog_priority);
         free(syslog_identifier);
@@ -1938,41 +2026,6 @@ finish:
         free(pid);
 }
 
-static void proc_kmsg_scan(Server *s) {
-        char *p;
-        size_t remaining;
-
-        assert(s);
-
-        p = s->proc_kmsg_buffer;
-        remaining = s->proc_kmsg_length;
-        for (;;) {
-                char *end;
-                size_t skip;
-
-                end = memchr(p, '\n', remaining);
-                if (end)
-                        skip = end - p + 1;
-                else if (remaining >= sizeof(s->proc_kmsg_buffer) - 1) {
-                        end = p + sizeof(s->proc_kmsg_buffer) - 1;
-                        skip = remaining;
-                } else
-                        break;
-
-                *end = 0;
-
-                proc_kmsg_line(s, p);
-
-                remaining -= skip;
-                p += skip;
-        }
-
-        if (p > s->proc_kmsg_buffer) {
-                memmove(s->proc_kmsg_buffer, p, remaining);
-                s->proc_kmsg_length = remaining;
-        }
-}
-
 static int system_journal_open(Server *s) {
         int r;
         char *fn;
@@ -2009,14 +2062,12 @@ static int system_journal_open(Server *s) {
                 if (!fn)
                         return -ENOMEM;
 
-                r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, &s->system_metrics, NULL, &s->system_journal);
+                r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, s->compress, s->seal, &s->system_metrics, s->mmap, NULL, &s->system_journal);
                 free(fn);
 
-                if (r >= 0) {
-                        s->system_journal->compress = s->compress;
-
+                if (r >= 0)
                         server_fix_perms(s, s->system_journal, 0);
-                } else if (r < 0) {
+                else if (r < 0) {
 
                         if (r != -ENOENT && r != -EROFS)
                                 log_warning("Failed to open system journal: %s", strerror(-r));
@@ -2038,7 +2089,7 @@ static int system_journal_open(Server *s) {
                          * if it already exists, so that we can flush
                          * it into the system journal */
 
-                        r = journal_file_open(fn, O_RDWR, 0640, &s->runtime_metrics, NULL, &s->runtime_journal);
+                        r = journal_file_open(fn, O_RDWR, 0640, s->compress, false, &s->runtime_metrics, s->mmap, NULL, &s->runtime_journal);
                         free(fn);
 
                         if (r < 0) {
@@ -2054,7 +2105,7 @@ static int system_journal_open(Server *s) {
                          * it if necessary. */
 
                         (void) mkdir_parents(fn, 0755);
-                        r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, &s->runtime_metrics, NULL, &s->runtime_journal);
+                        r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, s->compress, false, &s->runtime_metrics, s->mmap, NULL, &s->runtime_journal);
                         free(fn);
 
                         if (r < 0) {
@@ -2063,11 +2114,8 @@ static int system_journal_open(Server *s) {
                         }
                 }
 
-                if (s->runtime_journal) {
-                        s->runtime_journal->compress = s->compress;
-
+                if (s->runtime_journal)
                         server_fix_perms(s, s->runtime_journal, 0);
-                }
         }
 
         return r;
@@ -2142,52 +2190,57 @@ finish:
         journal_file_close(s->runtime_journal);
         s->runtime_journal = NULL;
 
-        if (r >= 0) {
-                char path[] = "/run/log/journal/xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
-                sd_id128_to_string(machine, path + 17);
-                rm_rf(path, false, true, false);
-        }
+        if (r >= 0)
+                rm_rf("/run/log/journal", false, true, false);
 
         return r;
 }
 
-static int server_read_proc_kmsg(Server *s) {
+static int server_read_dev_kmsg(Server *s) {
+        char buffer[8192+1]; /* the kernel-side limit per record is 8K currently */
         ssize_t l;
+
         assert(s);
-        assert(s->proc_kmsg_fd >= 0);
+        assert(s->dev_kmsg_fd >= 0);
 
-        l = read(s->proc_kmsg_fd, s->proc_kmsg_buffer + s->proc_kmsg_length, sizeof(s->proc_kmsg_buffer) - 1 - s->proc_kmsg_length);
-        if (l == 0) /* the kernel is stupid and in some race
-                     * conditions returns 0 in the middle of the
-                     * stream. */
+        l = read(s->dev_kmsg_fd, buffer, sizeof(buffer) - 1);
+        if (l == 0)
                 return 0;
         if (l < 0) {
+                /* Old kernels who don't allow reading from /dev/kmsg
+                 * return EINVAL when we try. So handle this cleanly,
+                 * but don' try to ever read from it again. */
+                if (errno == EINVAL) {
+                        epoll_ctl(s->epoll_fd, EPOLL_CTL_DEL, s->dev_kmsg_fd, NULL);
+                        return 0;
+                }
 
-                if (errno == EAGAIN || errno == EINTR)
+                if (errno == EAGAIN || errno == EINTR || errno == EPIPE)
                         return 0;
 
                 log_error("Failed to read from kernel: %m");
                 return -errno;
         }
 
-        s->proc_kmsg_length += l;
-
-        proc_kmsg_scan(s);
+        dev_kmsg_record(s, buffer, l);
         return 1;
 }
 
-static int server_flush_proc_kmsg(Server *s) {
+static int server_flush_dev_kmsg(Server *s) {
         int r;
 
         assert(s);
 
-        if (s->proc_kmsg_fd < 0)
+        if (s->dev_kmsg_fd < 0)
+                return 0;
+
+        if (!s->dev_kmsg_readable)
                 return 0;
 
-        log_info("Flushing /proc/kmsg...");
+        log_info("Flushing /dev/kmsg...");
 
         for (;;) {
-                r = server_read_proc_kmsg(s);
+                r = server_read_dev_kmsg(s);
                 if (r < 0)
                         return r;
 
@@ -2223,6 +2276,8 @@ static int process_event(Server *s, struct epoll_event *ev) {
                         return -errno;
                 }
 
+                log_info("Received SIG%s", signal_to_string(sfsi.ssi_signo));
+
                 if (sfsi.ssi_signo == SIGUSR1) {
                         touch("/run/systemd/journal/flushed");
                         server_flush_to_var(s);
@@ -2235,10 +2290,9 @@ static int process_event(Server *s, struct epoll_event *ev) {
                         return 1;
                 }
 
-                log_debug("Received SIG%s", signal_to_string(sfsi.ssi_signo));
                 return 0;
 
-        } else if (ev->data.fd == s->proc_kmsg_fd) {
+        } else if (ev->data.fd == s->dev_kmsg_fd) {
                 int r;
 
                 if (ev->events != EPOLLIN) {
@@ -2246,7 +2300,7 @@ static int process_event(Server *s, struct epoll_event *ev) {
                         return -EIO;
                 }
 
-                r = server_read_proc_kmsg(s);
+                r = server_read_dev_kmsg(s);
                 if (r < 0)
                         return r;
 
@@ -2591,28 +2645,68 @@ static int open_stdout_socket(Server *s) {
         return 0;
 }
 
-static int open_proc_kmsg(Server *s) {
+static int open_dev_kmsg(Server *s) {
         struct epoll_event ev;
 
         assert(s);
 
-        if (!s->import_proc_kmsg)
-                return 0;
-
-        s->proc_kmsg_fd = open("/proc/kmsg", O_RDONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY);
-        if (s->proc_kmsg_fd < 0) {
-                log_warning("Failed to open /proc/kmsg, ignoring: %m");
+        s->dev_kmsg_fd = open("/dev/kmsg", O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY);
+        if (s->dev_kmsg_fd < 0) {
+                log_warning("Failed to open /dev/kmsg, ignoring: %m");
                 return 0;
         }
 
         zero(ev);
         ev.events = EPOLLIN;
-        ev.data.fd = s->proc_kmsg_fd;
-        if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->proc_kmsg_fd, &ev) < 0) {
-                log_error("Failed to add /proc/kmsg fd to epoll object: %m");
+        ev.data.fd = s->dev_kmsg_fd;
+        if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->dev_kmsg_fd, &ev) < 0) {
+
+                /* This will fail with EPERM on older kernels where
+                 * /dev/kmsg is not readable. */
+                if (errno == EPERM)
+                        return 0;
+
+                log_error("Failed to add /dev/kmsg fd to epoll object: %m");
                 return -errno;
         }
 
+        s->dev_kmsg_readable = true;
+
+        return 0;
+}
+
+static int open_kernel_seqnum(Server *s) {
+        int fd;
+        uint64_t *p;
+
+        assert(s);
+
+        /* We store the seqnum we last read in an mmaped file. That
+         * way we can just use it like a variable, but it is
+         * persistant and automatically flushed at reboot. */
+
+        fd = open("/run/systemd/journal/kernel-seqnum", O_RDWR|O_CREAT|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
+        if (fd < 0) {
+                log_error("Failed to open /run/systemd/journal/kernel-seqnum, ignoring: %m");
+                return 0;
+        }
+
+        if (posix_fallocate(fd, 0, sizeof(uint64_t)) < 0) {
+                log_error("Failed to allocate sequential number file, ignoring: %m");
+                close_nointr_nofail(fd);
+                return 0;
+        }
+
+        p = mmap(NULL, sizeof(uint64_t), PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
+        if (p == MAP_FAILED) {
+                log_error("Failed to map sequential number file, ignoring: %m");
+                close_nointr_nofail(fd);
+                return 0;
+        }
+
+        close_nointr_nofail(fd);
+        s->kernel_seqnum = p;
+
         return 0;
 }
 
@@ -2730,8 +2824,9 @@ static int server_init(Server *s) {
         assert(s);
 
         zero(*s);
-        s->syslog_fd = s->native_fd = s->stdout_fd = s->signal_fd = s->epoll_fd = s->proc_kmsg_fd = -1;
+        s->syslog_fd = s->native_fd = s->stdout_fd = s->signal_fd = s->epoll_fd = s->dev_kmsg_fd = -1;
         s->compress = true;
+        s->seal = true;
 
         s->rate_limit_interval = DEFAULT_RATE_LIMIT_INTERVAL;
         s->rate_limit_burst = DEFAULT_RATE_LIMIT_BURST;
@@ -2749,11 +2844,15 @@ static int server_init(Server *s) {
         server_parse_config_file(s);
         server_parse_proc_cmdline(s);
 
+        mkdir_p("/run/systemd/journal", 0755);
+
         s->user_journals = hashmap_new(trivial_hash_func, trivial_compare_func);
-        if (!s->user_journals) {
-                log_error("Out of memory.");
-                return -ENOMEM;
-        }
+        if (!s->user_journals)
+                return log_oom();
+
+        s->mmap = mmap_cache_new();
+        if (!s->mmap)
+                return log_oom();
 
         s->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
         if (s->epoll_fd < 0) {
@@ -2814,7 +2913,11 @@ static int server_init(Server *s) {
         if (r < 0)
                 return r;
 
-        r = open_proc_kmsg(s);
+        r = open_dev_kmsg(s);
+        if (r < 0)
+                return r;
+
+        r = open_kernel_seqnum(s);
         if (r < 0)
                 return r;
 
@@ -2822,6 +2925,10 @@ static int server_init(Server *s) {
         if (r < 0)
                 return r;
 
+        s->udev = udev_new();
+        if (!s->udev)
+                return -ENOMEM;
+
         s->rate_limit = journal_rate_limit_new(s->rate_limit_interval, s->rate_limit_burst);
         if (!s->rate_limit)
                 return -ENOMEM;
@@ -2833,6 +2940,22 @@ static int server_init(Server *s) {
         return 0;
 }
 
+static void maybe_append_tags(Server *s) {
+#ifdef HAVE_GCRYPT
+        JournalFile *f;
+        Iterator i;
+        usec_t n;
+
+        n = now(CLOCK_REALTIME);
+
+        if (s->system_journal)
+                journal_file_maybe_append_tag(s->system_journal, n);
+
+        HASHMAP_FOREACH(f, s->user_journals, i)
+                journal_file_maybe_append_tag(f, n);
+#endif
+}
+
 static void server_done(Server *s) {
         JournalFile *f;
         assert(s);
@@ -2866,14 +2989,23 @@ static void server_done(Server *s) {
         if (s->stdout_fd >= 0)
                 close_nointr_nofail(s->stdout_fd);
 
-        if (s->proc_kmsg_fd >= 0)
-                close_nointr_nofail(s->proc_kmsg_fd);
+        if (s->dev_kmsg_fd >= 0)
+                close_nointr_nofail(s->dev_kmsg_fd);
 
         if (s->rate_limit)
                 journal_rate_limit_free(s->rate_limit);
 
+        if (s->kernel_seqnum)
+                munmap(s->kernel_seqnum, sizeof(uint64_t));
+
         free(s->buffer);
         free(s->tty_path);
+
+        if (s->mmap)
+                mmap_cache_unref(s->mmap);
+
+        if (s->udev)
+                udev_unref(s->udev);
 }
 
 int main(int argc, char *argv[]) {
@@ -2892,6 +3024,7 @@ int main(int argc, char *argv[]) {
 
         log_set_target(LOG_TARGET_SAFE);
         log_set_facility(LOG_SYSLOG);
+        log_set_max_level(LOG_DEBUG);
         log_parse_environment();
         log_open();
 
@@ -2903,7 +3036,7 @@ int main(int argc, char *argv[]) {
 
         server_vacuum(&server);
         server_flush_to_var(&server);
-        server_flush_proc_kmsg(&server);
+        server_flush_dev_kmsg(&server);
 
         log_debug("systemd-journald running as pid %lu", (unsigned long) getpid());
         driver_message(&server, SD_MESSAGE_JOURNAL_START, "Journal started");
@@ -2914,8 +3047,26 @@ int main(int argc, char *argv[]) {
 
         for (;;) {
                 struct epoll_event event;
+                int t;
+
+#ifdef HAVE_GCRYPT
+                usec_t u;
+
+                if (server.system_journal &&
+                    journal_file_next_evolve_usec(server.system_journal, &u)) {
+                        usec_t n;
+
+                        n = now(CLOCK_REALTIME);
 
-                r = epoll_wait(server.epoll_fd, &event, 1, -1);
+                        if (n >= u)
+                                t = 0;
+                        else
+                                t = (int) ((u - n + USEC_PER_MSEC - 1) / USEC_PER_MSEC);
+                } else
+#endif
+                        t = -1;
+
+                r = epoll_wait(server.epoll_fd, &event, 1, t);
                 if (r < 0) {
 
                         if (errno == EINTR)
@@ -2924,14 +3075,17 @@ int main(int argc, char *argv[]) {
                         log_error("epoll_wait() failed: %m");
                         r = -errno;
                         goto finish;
-                } else if (r == 0)
-                        break;
+                }
 
-                r = process_event(&server, &event);
-                if (r < 0)
-                        goto finish;
-                else if (r == 0)
-                        break;
+                if (r > 0) {
+                        r = process_event(&server, &event);
+                        if (r < 0)
+                                goto finish;
+                        else if (r == 0)
+                                break;
+                }
+
+                maybe_append_tags(&server);
         }
 
         log_debug("systemd-journald stopped as pid %lu", (unsigned long) getpid());