#include <linux/fiemap.h>
#include <sys/ioctl.h>
#include <sys/vfs.h>
+#include <getopt.h>
+#include <sys/inotify.h>
#include "missing.h"
#include "util.h"
#include "ioprio.h"
#include "readahead-common.h"
-#define MINCORE_VEC_SIZE (READAHEAD_FILE_SIZE_MAX/PAGE_SIZE)
+/* fixme:
+ *
+ * - detect ssd on btrfs/lvm...
+ * - read ahead directories
+ * - gzip?
+ * - remount rw?
+ * - handle files where nothing is in mincore
+ * - does ioprio_set work with fadvise()?
+ */
+
+static unsigned arg_files_max = 16*1024;
+static off_t arg_file_size_max = READAHEAD_FILE_SIZE_MAX;
+static usec_t arg_timeout = 2*USEC_PER_MINUTE;
+
+static ReadaheadShared *shared = NULL;
+
+/* Avoid collisions with the NULL pointer */
+#define SECTOR_TO_PTR(s) ULONG_TO_PTR((s)+1)
+#define PTR_TO_SECTOR(p) (PTR_TO_ULONG(p)-1)
static int btrfs_defrag(int fd) {
struct btrfs_ioctl_vol_args data;
static int pack_file(FILE *pack, const char *fn, bool on_btrfs) {
struct stat st;
void *start = MAP_FAILED;
- uint8_t vec[MINCORE_VEC_SIZE];
+ uint8_t *vec;
uint32_t b, c;
size_t l, pages;
bool mapped;
assert(fn);
if ((fd = open(fn, O_RDONLY|O_CLOEXEC|O_NOATIME|O_NOCTTY|O_NOFOLLOW)) < 0) {
+
+ if (errno == ENOENT)
+ return 0;
+
+ if (errno == EPERM || errno == EACCES)
+ return 0;
+
log_warning("open(%s) failed: %m", fn);
r = -errno;
goto finish;
}
- if ((k = file_verify(fd, fn, &st)) <= 0) {
+ if ((k = file_verify(fd, fn, arg_file_size_max, &st)) <= 0) {
r = k;
goto finish;
}
goto finish;
}
+ pages = l / page_size();
+
+ vec = alloca(pages);
+ memset(vec, 0, pages);
if (mincore(start, l, vec) < 0) {
log_warning("mincore(%s) failed: %m", fn);
r = -errno;
fputs(fn, pack);
fputc('\n', pack);
- pages = l / PAGE_SIZE;
mapped = false;
for (c = 0; c < pages; c++) {
- bool new_mapped = (vec[c] & 1);
+ bool new_mapped = !!(vec[c] & 1);
if (!mapped && new_mapped)
b = c;
static int collect(const char *root) {
enum {
- FD_FANOTIFY,
+ FD_FANOTIFY, /* Get the actual fs events */
FD_SIGNAL,
+ FD_INOTIFY, /* We get notifications to quit early via this fd */
_FD_MAX
};
struct pollfd pollfd[_FD_MAX];
- int fanotify_fd = -1, signal_fd = -1, r = 0;
+ int fanotify_fd = -1, signal_fd = -1, inotify_fd = -1, r = 0;
pid_t my_pid;
Hashmap *files = NULL;
Iterator i;
char *pack_fn_new = NULL, *pack_fn = NULL;
bool on_ssd, on_btrfs;
struct statfs sfs;
+ usec_t not_after;
assert(root);
+ write_one_line_file("/proc/self/oom_score_adj", "1000");
+
if (ioprio_set(IOPRIO_WHO_PROCESS, getpid(), IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0)) < 0)
log_warning("Failed to set IDLE IO priority class: %m");
goto finish;
}
- if ((fanotify_fd = fanotify_init(FAN_CLOEXEC, O_RDONLY|O_LARGEFILE|O_CLOEXEC|O_NOATIME)) < 0) {
+ if ((fanotify_fd = fanotify_init(FAN_CLOEXEC|FAN_NONBLOCK, O_RDONLY|O_LARGEFILE|O_CLOEXEC|O_NOATIME)) < 0) {
log_error("Failed to create fanotify object: %m");
r = -errno;
goto finish;
goto finish;
}
+ if ((inotify_fd = open_inotify()) < 0) {
+ r = inotify_fd;
+ goto finish;
+ }
+
+ not_after = now(CLOCK_MONOTONIC) + arg_timeout;
+
my_pid = getpid();
zero(pollfd);
pollfd[FD_FANOTIFY].events = POLLIN;
pollfd[FD_SIGNAL].fd = signal_fd;
pollfd[FD_SIGNAL].events = POLLIN;
+ pollfd[FD_INOTIFY].fd = inotify_fd;
+ pollfd[FD_INOTIFY].events = POLLIN;
sd_notify(0,
"READY=1\n"
log_debug("Collecting...");
+ if (access("/run/systemd/readahead/cancel", F_OK) >= 0) {
+ log_debug("Collection canceled");
+ r = -ECANCELED;
+ goto finish;
+ }
+
+ if (access("/run/systemd/readahead/done", F_OK) >= 0) {
+ log_debug("Got termination request");
+ goto done;
+ }
+
for (;;) {
union {
struct fanotify_event_metadata metadata;
} data;
ssize_t n;
struct fanotify_event_metadata *m;
+ usec_t t;
+ int h;
+
+ if (hashmap_size(files) > arg_files_max) {
+ log_debug("Reached maximum number of read ahead files, ending collection.");
+ break;
+ }
- if (hashmap_size(files) > READAHEAD_FILES_MAX)
+ t = now(CLOCK_MONOTONIC);
+ if (t >= not_after) {
+ log_debug("Reached maximum collection time, ending collection.");
break;
+ }
- if (poll(pollfd, _FD_MAX, -1) < 0) {
+ if ((h = poll(pollfd, _FD_MAX, (int) ((not_after - t) / USEC_PER_MSEC))) < 0) {
if (errno == EINTR)
continue;
goto finish;
}
- if (pollfd[FD_SIGNAL].revents != 0)
+ if (h == 0) {
+ log_debug("Reached maximum collection time, ending collection.");
+ break;
+ }
+
+ if (pollfd[FD_SIGNAL].revents) {
+ log_debug("Got signal.");
break;
+ }
+
+ if (pollfd[FD_INOTIFY].revents) {
+ uint8_t inotify_buffer[sizeof(struct inotify_event) + FILENAME_MAX];
+ struct inotify_event *e;
+
+ if ((n = read(inotify_fd, &inotify_buffer, sizeof(inotify_buffer))) < 0) {
+ if (errno == EINTR || errno == EAGAIN)
+ continue;
+
+ log_error("Failed to read inotify event: %m");
+ r = -errno;
+ goto finish;
+ }
+
+ e = (struct inotify_event*) inotify_buffer;
+ while (n > 0) {
+ size_t step;
+
+ if ((e->mask & IN_CREATE) && streq(e->name, "cancel")) {
+ log_debug("Collection canceled");
+ r = -ECANCELED;
+ goto finish;
+ }
+
+ if ((e->mask & IN_CREATE) && streq(e->name, "done")) {
+ log_debug("Got termination request");
+ goto done;
+ }
+
+ step = sizeof(struct inotify_event) + e->len;
+ assert(step <= (size_t) n);
+
+ e = (struct inotify_event*) ((uint8_t*) e + step);
+ n -= step;
+ }
+ }
if ((n = read(fanotify_fd, &data, sizeof(data))) < 0) {
if (errno == EINTR || errno == EAGAIN)
continue;
+ /* fanotify sometimes returns EACCES on read()
+ * where it shouldn't. For now let's just
+ * ignore it here (which is safe), but
+ * eventually this should be
+ * dropped when the kernel is fixed.
+ *
+ * https://bugzilla.redhat.com/show_bug.cgi?id=707577 */
+ if (errno == EACCES)
+ continue;
+
log_error("Failed to read event: %m");
r = -errno;
goto finish;
}
- m = &data.metadata;
- while (FAN_EVENT_OK(m, n)) {
-
- if (m->pid != my_pid && m->fd >= 0) {
- char fn[PATH_MAX];
- int k;
+ for (m = &data.metadata; FAN_EVENT_OK(m, n); m = FAN_EVENT_NEXT(m, n)) {
+ char fn[PATH_MAX];
+ int k;
- snprintf(fn, sizeof(fn), "/proc/self/fd/%i", m->fd);
- char_array_0(fn);
+ if (m->fd < 0)
+ goto next_iteration;
- if ((k = readlink_malloc(fn, &p)) >= 0) {
+ if (m->pid == my_pid)
+ goto next_iteration;
- if (hashmap_get(files, p))
- /* Already read */
- free(p);
- else {
- unsigned long ul;
+ __sync_synchronize();
+ if (m->pid == shared->replay)
+ goto next_iteration;
- ul = fd_first_block(m->fd);
+ snprintf(fn, sizeof(fn), "/proc/self/fd/%i", m->fd);
+ char_array_0(fn);
- if ((k = hashmap_put(files, p, ULONG_TO_PTR(ul))) < 0) {
+ if ((k = readlink_malloc(fn, &p)) >= 0) {
+ if (startswith(p, "/tmp") ||
+ endswith(p, " (deleted)") ||
+ hashmap_get(files, p))
+ /* Not interesting, or
+ * already read */
+ free(p);
+ else {
+ unsigned long ul;
- if (k != -EEXIST)
- log_warning("set_put() failed: %s", strerror(-k));
+ ul = fd_first_block(m->fd);
- free(p);
- }
+ if ((k = hashmap_put(files, p, SECTOR_TO_PTR(ul))) < 0) {
+ log_warning("set_put() failed: %s", strerror(-k));
+ free(p);
}
+ }
- } else
- log_warning("readlink(%s) failed: %s", fn, strerror(-k));
- }
+ } else
+ log_warning("readlink(%s) failed: %s", fn, strerror(-k));
+ next_iteration:
if (m->fd)
close_nointr_nofail(m->fd);
-
- m = FAN_EVENT_NEXT(m, n);
}
-
}
+done:
if (fanotify_fd >= 0) {
close_nointr_nofail(fanotify_fd);
fanotify_fd = -1;
log_debug("Writing Pack File...");
- on_ssd = fs_on_ssd(root);
+ on_ssd = fs_on_ssd(root) > 0;
log_debug("On SSD: %s", yes_no(on_ssd));
- on_btrfs = statfs(root, &sfs) >= 0 && sfs.f_type == BTRFS_SUPER_MAGIC;
+ on_btrfs = statfs(root, &sfs) >= 0 && (long) sfs.f_type == (long) BTRFS_SUPER_MAGIC;
log_debug("On btrfs: %s", yes_no(on_btrfs));
asprintf(&pack_fn, "%s/.readahead", root);
if (on_ssd || on_btrfs) {
/* On SSD or on btrfs, just write things out in the
- * order the files where accessed. */
+ * order the files were accessed. */
HASHMAP_FOREACH_KEY(q, p, files, i)
pack_file(pack, p, on_btrfs);
j = ordered;
HASHMAP_FOREACH_KEY(q, p, files, i) {
j->path = p;
- j->block = PTR_TO_ULONG(q);
+ j->block = PTR_TO_SECTOR(q);
j++;
}
if (signal_fd >= 0)
close_nointr_nofail(signal_fd);
+ if (inotify_fd >= 0)
+ close_nointr_nofail(inotify_fd);
+
if (pack) {
fclose(pack);
unlink(pack_fn_new);
free(pack_fn);
while ((p = hashmap_steal_first_key(files)))
- free(q);
+ free(p);
hashmap_free(files);
return r;
}
+static int help(void) {
+
+ printf("%s [OPTIONS...] [DIRECTORY]\n\n"
+ "Collect read-ahead data on early boot.\n\n"
+ " -h --help Show this help\n"
+ " --max-files=INT Maximum number of files to read ahead\n"
+ " --max-file-size=BYTES Maximum size of files to read ahead\n"
+ " --timeout=USEC Maximum time to spend collecting data\n",
+ program_invocation_short_name);
+
+ return 0;
+}
+
+static int parse_argv(int argc, char *argv[]) {
+
+ enum {
+ ARG_FILES_MAX = 0x100,
+ ARG_FILE_SIZE_MAX,
+ ARG_TIMEOUT
+ };
+
+ static const struct option options[] = {
+ { "help", no_argument, NULL, 'h' },
+ { "files-max", required_argument, NULL, ARG_FILES_MAX },
+ { "file-size-max", required_argument, NULL, ARG_FILE_SIZE_MAX },
+ { "timeout", required_argument, NULL, ARG_TIMEOUT },
+ { NULL, 0, NULL, 0 }
+ };
+
+ int c;
+
+ assert(argc >= 0);
+ assert(argv);
+
+ while ((c = getopt_long(argc, argv, "h", options, NULL)) >= 0) {
+
+ switch (c) {
+
+ case 'h':
+ help();
+ return 0;
+
+ case ARG_FILES_MAX:
+ if (safe_atou(optarg, &arg_files_max) < 0 || arg_files_max <= 0) {
+ log_error("Failed to parse maximum number of files %s.", optarg);
+ return -EINVAL;
+ }
+ break;
+
+ case ARG_FILE_SIZE_MAX: {
+ unsigned long long ull;
+
+ if (safe_atollu(optarg, &ull) < 0 || ull <= 0) {
+ log_error("Failed to parse maximum file size %s.", optarg);
+ return -EINVAL;
+ }
+
+ arg_file_size_max = (off_t) ull;
+ break;
+ }
+
+ case ARG_TIMEOUT:
+ if (parse_usec(optarg, &arg_timeout) < 0 || arg_timeout <= 0) {
+ log_error("Failed to parse timeout %s.", optarg);
+ return -EINVAL;
+ }
+
+ break;
+
+ case '?':
+ return -EINVAL;
+
+ default:
+ log_error("Unknown option code %c", c);
+ return -EINVAL;
+ }
+ }
+
+ if (optind != argc &&
+ optind != argc-1) {
+ help();
+ return -EINVAL;
+ }
+
+ return 1;
+}
+
int main(int argc, char *argv[]) {
- /* log_set_target(LOG_TARGET_SYSLOG_OR_KMSG); */
+ int r;
+ const char *root;
+
+ log_set_target(LOG_TARGET_SYSLOG_OR_KMSG);
log_parse_environment();
log_open();
- log_set_max_level(LOG_DEBUG);
+ umask(0022);
+
+ if ((r = parse_argv(argc, argv)) <= 0)
+ return r < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
+
+ root = optind < argc ? argv[optind] : "/";
+
+ if (fs_on_read_only(root) > 0) {
+ log_info("Disabling readahead collector due to read-only media.");
+ return 0;
+ }
+
+ if (!enough_ram()) {
+ log_info("Disabling readahead collector due to low memory.");
+ return 0;
+ }
+
+ if (detect_virtualization(NULL) > 0) {
+ log_info("Disabling readahead collector due to execution in virtualized environment.");
+ return 0;
+ }
+
+ if (!(shared = shared_get()))
+ return 1;
+
+ shared->collect = getpid();
+ __sync_synchronize();
- if (collect("/") < 0)
+ if (collect(root) < 0)
return 1;
return 0;