static int control_first_done(struct gfs2_sbd *sdp)
{
struct lm_lockstruct *ls = &sdp->sd_lockstruct;
char lvb_bits[GDLM_LVB_SIZE];
uint32_t start_gen, block_gen;
int error;
restart:
spin_lock(&ls->ls_recover_spin);
start_gen = ls->ls_recover_start;
block_gen = ls->ls_recover_block;
if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
/* sanity check, should not happen */
fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
start_gen, block_gen, ls->ls_recover_flags);
spin_unlock(&ls->ls_recover_spin);
control_unlock(sdp);
return -1;
}
if (start_gen == block_gen) {
/*
* Wait for the end of a dlm recovery cycle to switch from
* first mounter recovery. We can ignore any recover_slot
* callbacks between the recover_prep and next recover_done
* because we are still the first mounter and any failed nodes
* have not fully mounted, so they don't need recovery.
*/
spin_unlock(&ls->ls_recover_spin);
fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
dlm_recovery_wait, TASK_UNINTERRUPTIBLE);
goto restart;
}
clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
spin_unlock(&ls->ls_recover_spin);
memset(lvb_bits, 0, sizeof(lvb_bits));
control_lvb_write(ls, start_gen, lvb_bits);
error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
if (error)
fs_err(sdp, "control_first_done mounted PR error %d\n", error);
error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
if (error)
fs_err(sdp, "control_first_done control NL error %d\n", error);
return error;
}
static int control_first_done(struct gfs2_sbd *sdp)
{
struct lm_lockstruct *ls = &sdp->sd_lockstruct;
char lvb_bits[GDLM_LVB_SIZE];
uint32_t start_gen, block_gen;
int error;
restart:
spin_lock(&ls->ls_recover_spin);
start_gen = ls->ls_recover_start;
block_gen = ls->ls_recover_block;
if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
/* */
fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
start_gen, block_gen, ls->ls_recover_flags);
spin_unlock(&ls->ls_recover_spin);
control_unlock(sdp);
return -1;
}
if (start_gen == block_gen) {
/*
*/
spin_unlock(&ls->ls_recover_spin);
fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
dlm_recovery_wait, TASK_UNINTERRUPTIBLE);
goto restart;
}
clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
spin_unlock(&ls->ls_recover_spin);
memset(lvb_bits, 0, sizeof(lvb_bits));
control_lvb_write(ls, start_gen, lvb_bits);
error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
if (error)
fs_err(sdp, "control_first_done mounted PR error %d\n", error);
error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
if (error)
fs_err(sdp, "control_first_done control NL error %d\n", error);
return error;
}
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;
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, ls->ls_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, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
continue;
__clear_bit_le(i, ls->ls_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, ls->ls_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, ls->ls_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, ls->ls_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 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);
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);
if (block_gen == start_gen)
return;
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) {
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) {
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) {
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);
}
}
write_lvb = 1;
} else {
}
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;
}
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;
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);
}
}