+ error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
+ if (error) {
+ fs_err(sdp, "%s lkid %x error %d\n",
+ name, lksb->sb_lkid, error);
+ return error;
+ }
+
+ wait_for_completion(&ls->ls_sync_wait);
+
+ if (lksb->sb_status != -DLM_EUNLOCK) {
+ fs_err(sdp, "%s lkid %x status %d\n",
+ name, lksb->sb_lkid, lksb->sb_status);
+ return -1;
+ }
+ return 0;
+}
+
+static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
+ unsigned int num, struct dlm_lksb *lksb, char *name)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ char strname[GDLM_STRNAME_BYTES];
+ int error, status;
+
+ memset(strname, 0, GDLM_STRNAME_BYTES);
+ snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
+
+ error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
+ strname, GDLM_STRNAME_BYTES - 1,
+ 0, sync_wait_cb, ls, NULL);
+ if (error) {
+ fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
+ name, lksb->sb_lkid, flags, mode, error);
+ return error;
+ }
+
+ wait_for_completion(&ls->ls_sync_wait);
+
+ status = lksb->sb_status;
+
+ if (status && status != -EAGAIN) {
+ fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
+ name, lksb->sb_lkid, flags, mode, status);
+ }
+
+ return status;
+}
+
+static int mounted_unlock(struct gfs2_sbd *sdp)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
+}
+
+static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
+ &ls->ls_mounted_lksb, "mounted_lock");
+}
+
+static int control_unlock(struct gfs2_sbd *sdp)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
+}
+
+static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
+ &ls->ls_control_lksb, "control_lock");
+}
+
+static void gfs2_control_func(struct work_struct *work)
+{
+ struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ char lvb_bits[GDLM_LVB_SIZE];
+ uint32_t block_gen, start_gen, lvb_gen, flags;
+ int recover_set = 0;
+ int write_lvb = 0;
+ int recover_size;
+ int i, error;
+
+ spin_lock(&ls->ls_recover_spin);
+ /*
+ * No MOUNT_DONE means we're still mounting; control_mount()
+ * will set this flag, after which this thread will take over
+ * all further clearing of BLOCK_LOCKS.
+ *
+ * FIRST_MOUNT means this node is doing first mounter recovery,
+ * for which recovery control is handled by
+ * control_mount()/control_first_done(), not this thread.
+ */
+ if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
+ test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
+ spin_unlock(&ls->ls_recover_spin);
+ return;
+ }
+ block_gen = ls->ls_recover_block;
+ start_gen = ls->ls_recover_start;
+ spin_unlock(&ls->ls_recover_spin);
+
+ /*
+ * Equal block_gen and start_gen implies we are between
+ * recover_prep and recover_done callbacks, which means
+ * dlm recovery is in progress and dlm locking is blocked.
+ * There's no point trying to do any work until recover_done.
+ */
+
+ if (block_gen == start_gen)
+ return;
+
+ /*
+ * Propagate recover_submit[] and recover_result[] to lvb:
+ * dlm_recoverd adds to recover_submit[] jids needing recovery
+ * gfs2_recover adds to recover_result[] journal recovery results
+ *
+ * set lvb bit for jids in recover_submit[] if the lvb has not
+ * yet been updated for the generation of the failure
+ *
+ * clear lvb bit for jids in recover_result[] if the result of
+ * the journal recovery is SUCCESS
+ */
+
+ error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
+ if (error) {
+ fs_err(sdp, "control lock EX error %d\n", error);
+ return;
+ }
+
+ control_lvb_read(ls, &lvb_gen, lvb_bits);
+
+ spin_lock(&ls->ls_recover_spin);
+ if (block_gen != ls->ls_recover_block ||
+ start_gen != ls->ls_recover_start) {
+ fs_info(sdp, "recover generation %u block1 %u %u\n",
+ start_gen, block_gen, ls->ls_recover_block);
+ spin_unlock(&ls->ls_recover_spin);
+ control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
+ return;
+ }
+
+ recover_size = ls->ls_recover_size;
+
+ if (lvb_gen <= start_gen) {
+ /*
+ * Clear lvb bits for jids we've successfully recovered.
+ * Because all nodes attempt to recover failed journals,
+ * a journal can be recovered multiple times successfully
+ * in succession. Only the first will really do recovery,
+ * the others find it clean, but still report a successful
+ * recovery. So, another node may have already recovered
+ * the jid and cleared the lvb bit for it.
+ */
+ for (i = 0; i < recover_size; i++) {
+ if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
+ continue;
+
+ ls->ls_recover_result[i] = 0;
+
+ if (!test_bit_le(i, lvb_bits + JID_BITMAP_OFFSET))
+ continue;
+
+ __clear_bit_le(i, lvb_bits + JID_BITMAP_OFFSET);
+ write_lvb = 1;
+ }
+ }
+
+ if (lvb_gen == start_gen) {
+ /*
+ * Failed slots before start_gen are already set in lvb.
+ */
+ for (i = 0; i < recover_size; i++) {
+ if (!ls->ls_recover_submit[i])
+ continue;
+ if (ls->ls_recover_submit[i] < lvb_gen)
+ ls->ls_recover_submit[i] = 0;
+ }
+ } else if (lvb_gen < start_gen) {
+ /*
+ * Failed slots before start_gen are not yet set in lvb.
+ */
+ for (i = 0; i < recover_size; i++) {
+ if (!ls->ls_recover_submit[i])
+ continue;
+ if (ls->ls_recover_submit[i] < start_gen) {
+ ls->ls_recover_submit[i] = 0;
+ __set_bit_le(i, lvb_bits + JID_BITMAP_OFFSET);
+ }
+ }
+ /* even if there are no bits to set, we need to write the
+ latest generation to the lvb */
+ write_lvb = 1;
+ } else {
+ /*
+ * we should be getting a recover_done() for lvb_gen soon
+ */
+ }
+ spin_unlock(&ls->ls_recover_spin);
+
+ if (write_lvb) {
+ control_lvb_write(ls, start_gen, lvb_bits);
+ flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
+ } else {
+ flags = DLM_LKF_CONVERT;
+ }
+
+ error = control_lock(sdp, DLM_LOCK_NL, flags);
+ if (error) {
+ fs_err(sdp, "control lock NL error %d\n", error);
+ return;
+ }
+
+ /*
+ * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
+ * and clear a jid bit in the lvb if the recovery is a success.
+ * Eventually all journals will be recovered, all jid bits will
+ * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
+ */
+
+ for (i = 0; i < recover_size; i++) {
+ if (test_bit_le(i, lvb_bits + JID_BITMAP_OFFSET)) {
+ fs_info(sdp, "recover generation %u jid %d\n",
+ start_gen, i);
+ gfs2_recover_set(sdp, i);
+ recover_set++;
+ }
+ }
+ if (recover_set)
+ return;
+
+ /*
+ * No more jid bits set in lvb, all recovery is done, unblock locks
+ * (unless a new recover_prep callback has occured blocking locks
+ * again while working above)
+ */
+
+ spin_lock(&ls->ls_recover_spin);
+ if (ls->ls_recover_block == block_gen &&
+ ls->ls_recover_start == start_gen) {
+ clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
+ spin_unlock(&ls->ls_recover_spin);
+ fs_info(sdp, "recover generation %u done\n", start_gen);
+ gfs2_glock_thaw(sdp);
+ } else {
+ fs_info(sdp, "recover generation %u block2 %u %u\n",
+ start_gen, block_gen, ls->ls_recover_block);
+ spin_unlock(&ls->ls_recover_spin);
+ }
+}
+
+static int control_mount(struct gfs2_sbd *sdp)
+{
+ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
+ char lvb_bits[GDLM_LVB_SIZE];
+ uint32_t start_gen, block_gen, mount_gen, lvb_gen;
+ int mounted_mode;
+ int retries = 0;
+ int error;
+
+ memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
+ memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
+ memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
+ ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
+ init_completion(&ls->ls_sync_wait);
+
+ set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
+
+ error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
+ if (error) {
+ fs_err(sdp, "control_mount control_lock NL error %d\n", error);
+ return error;
+ }
+
+ error = mounted_lock(sdp, DLM_LOCK_NL, 0);
+ if (error) {
+ fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
+ control_unlock(sdp);
+ return error;
+ }
+ mounted_mode = DLM_LOCK_NL;
+
+restart:
+ if (retries++ && signal_pending(current)) {
+ error = -EINTR;
+ goto fail;
+ }
+
+ /*
+ * We always start with both locks in NL. control_lock is
+ * demoted to NL below so we don't need to do it here.
+ */
+
+ if (mounted_mode != DLM_LOCK_NL) {
+ error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
+ if (error)
+ goto fail;
+ mounted_mode = DLM_LOCK_NL;
+ }
+
+ /*
+ * Other nodes need to do some work in dlm recovery and gfs2_control
+ * before the recover_done and control_lock will be ready for us below.
+ * A delay here is not required but often avoids having to retry.
+ */
+
+ msleep_interruptible(500);
+
+ /*
+ * Acquire control_lock in EX and mounted_lock in either EX or PR.
+ * control_lock lvb keeps track of any pending journal recoveries.
+ * mounted_lock indicates if any other nodes have the fs mounted.
+ */
+
+ error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
+ if (error == -EAGAIN) {
+ goto restart;
+ } else if (error) {
+ fs_err(sdp, "control_mount control_lock EX error %d\n", error);
+ goto fail;
+ }
+
+ error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
+ if (!error) {
+ mounted_mode = DLM_LOCK_EX;
+ goto locks_done;
+ } else if (error != -EAGAIN) {
+ fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
+ goto fail;
+ }
+
+ error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
+ if (!error) {
+ mounted_mode = DLM_LOCK_PR;
+ goto locks_done;
+ } else {
+ /* not even -EAGAIN should happen here */
+ fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
+ goto fail;
+ }
+
+locks_done:
+ /*
+ * If we got both locks above in EX, then we're the first mounter.
+ * If not, then we need to wait for the control_lock lvb to be
+ * updated by other mounted nodes to reflect our mount generation.
+ *
+ * In simple first mounter cases, first mounter will see zero lvb_gen,
+ * but in cases where all existing nodes leave/fail before mounting
+ * nodes finish control_mount, then all nodes will be mounting and
+ * lvb_gen will be non-zero.
+ */
+
+ control_lvb_read(ls, &lvb_gen, lvb_bits);
+
+ if (lvb_gen == 0xFFFFFFFF) {
+ /* special value to force mount attempts to fail */
+ fs_err(sdp, "control_mount control_lock disabled\n");
+ error = -EINVAL;
+ goto fail;
+ }
+
+ if (mounted_mode == DLM_LOCK_EX) {
+ /* first mounter, keep both EX while doing first recovery */
+ spin_lock(&ls->ls_recover_spin);
+ clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
+ set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
+ set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
+ spin_unlock(&ls->ls_recover_spin);
+ fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
+ return 0;
+ }
+
+ error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
+ if (error)
+ goto fail;
+
+ /*
+ * We are not first mounter, now we need to wait for the control_lock
+ * lvb generation to be >= the generation from our first recover_done
+ * and all lvb bits to be clear (no pending journal recoveries.)
+ */
+
+ if (!all_jid_bits_clear(lvb_bits)) {
+ /* journals need recovery, wait until all are clear */
+ fs_info(sdp, "control_mount wait for journal recovery\n");
+ goto restart;
+ }
+
+ spin_lock(&ls->ls_recover_spin);
+ block_gen = ls->ls_recover_block;
+ start_gen = ls->ls_recover_start;
+ mount_gen = ls->ls_recover_mount;
+
+ if (lvb_gen < mount_gen) {
+ /* wait for mounted nodes to update control_lock lvb to our
+ generation, which might include new recovery bits set */
+ fs_info(sdp, "control_mount wait1 block %u start %u mount %u "
+ "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
+ lvb_gen, ls->ls_recover_flags);
+ spin_unlock(&ls->ls_recover_spin);
+ goto restart;
+ }
+
+ if (lvb_gen != start_gen) {
+ /* wait for mounted nodes to update control_lock lvb to the
+ latest recovery generation */
+ fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
+ "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
+ lvb_gen, ls->ls_recover_flags);
+ spin_unlock(&ls->ls_recover_spin);
+ goto restart;
+ }
+
+ if (block_gen == start_gen) {
+ /* dlm recovery in progress, wait for it to finish */
+ fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
+ "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
+ lvb_gen, ls->ls_recover_flags);
+ spin_unlock(&ls->ls_recover_spin);
+ goto restart;