int si_plannertrace(siplan *p, sstrace *t)
{
char *plan = NULL;
switch (p->plan) {
case SI_BRANCH: plan = "branch";
break;
case SI_AGE: plan = "age";
break;
case SI_COMPACT: plan = "compact";
break;
case SI_CHECKPOINT: plan = "checkpoint";
break;
case SI_GC: plan = "gc";
break;
case SI_TEMP: plan = "temperature";
break;
case SI_BACKUP:
case SI_BACKUPEND: plan = "backup";
break;
case SI_SHUTDOWN: plan = "database shutdown";
break;
case SI_DROP: plan = "database drop";
break;
}
char *explain = NULL;
switch (p->explain) {
case SI_ENONE:
explain = "none";
break;
case SI_ERETRY:
explain = "retry expected";
break;
case SI_EINDEX_SIZE:
explain = "index size";
break;
case SI_EINDEX_AGE:
explain = "index age";
break;
case SI_EBRANCH_COUNT:
explain = "branch count";
break;
case SI_ETEMP:
explain = "temperature";
break;
}
if (p->node) {
ss_trace(t, "%s <#%" PRIu32 " explain: %s>",
plan,
p->node->self.id.id, explain);
} else {
ss_trace(t, "%s <explain: %s>", plan, explain);
}
return 0;
}
static inline int
se_backupcomplete(sescheduler *s, seworker *w)
{
/*
* a. rotate log file
* b. copy log files
* c. enable log gc
* d. rename <bsn.incomplete> into <bsn>
* e. set last backup, set COMPLETE
*/
se *e = (se*)s->env;
/* force log rotation */
ss_trace(&w->trace, "%s", "log rotation for backup");
int rc = sl_poolrotate(&e->lp);
if (ssunlikely(rc == -1))
return -1;
/* copy log files */
ss_trace(&w->trace, "%s", "log files backup");
char path[1024];
snprintf(path, sizeof(path), "%s/%" PRIu32 ".incomplete/log",
e->conf.backup_path, s->backup_bsn);
rc = sl_poolcopy(&e->lp, path, &w->dc.c);
if (ssunlikely(rc == -1)) {
sr_errorrecover(&e->error);
return -1;
}
/* enable log gc */
sl_poolgc_enable(&e->lp, 1);
/* complete backup */
snprintf(path, sizeof(path), "%s/%" PRIu32 ".incomplete",
e->conf.backup_path, s->backup_bsn);
char newpath[1024];
snprintf(newpath, sizeof(newpath), "%s/%" PRIu32,
e->conf.backup_path, s->backup_bsn);
rc = rename(path, newpath);
if (ssunlikely(rc == -1)) {
sr_error(&e->error, "backup directory '%s' rename error: %s",
path, strerror(errno));
return -1;
}
/* complete */
s->backup_last = s->backup_bsn;
s->backup_last_complete = 1;
s->backup = 0;
s->backup_bsn = 0;
return 0;
}
int sc_backupend(sc *s, scworker *w)
{
/*
* a. rotate log file
* b. copy log files
* c. enable log gc
* d. rename <bsn.incomplete> into <bsn>
* e. set last backup, set COMPLETE
*/
/* force log rotation */
ss_trace(&w->trace, "%s", "log rotation for backup");
int rc = sl_poolrotate(s->lp);
if (ssunlikely(rc == -1))
return -1;
/* copy log files */
ss_trace(&w->trace, "%s", "log files backup");
char path[1024];
snprintf(path, sizeof(path), "%s/%" PRIu32 ".incomplete/log",
s->backup_path, s->backup_bsn);
rc = sl_poolcopy(s->lp, path, &w->dc.c);
if (ssunlikely(rc == -1))
return -1;
/* complete backup */
snprintf(path, sizeof(path), "%s/%" PRIu32 ".incomplete",
s->backup_path, s->backup_bsn);
char newpath[1024];
snprintf(newpath, sizeof(newpath), "%s/%" PRIu32,
s->backup_path, s->backup_bsn);
rc = ss_vfsrename(s->r->vfs, path, newpath);
if (ssunlikely(rc == -1)) {
sr_error(s->r->e, "backup directory '%s' rename error: %s",
path, strerror(errno));
return -1;
}
/* enable log gc */
sl_poolgc_enable(s->lp, 1);
/* complete */
ss_mutexlock(&s->lock);
s->backup_bsn_last = s->backup_bsn;
s->backup_bsn_last_complete = 1;
s->backup_in_progress = 0;
s->backup = 0;
s->backup_bsn = 0;
ss_mutexunlock(&s->lock);
return 0;
}
static inline int
sc_gc(sc *s, scworker *w)
{
ss_trace(&w->trace, "%s", "log gc");
int rc = sw_managergc(s->wm);
if (ssunlikely(rc == -1))
return -1;
return 0;
}
int se_scheduler(sescheduler *s, seworker *w)
{
setask task;
int rc = se_schedule(s, &task, w);
int job = rc;
if (task.rotate) {
rc = se_rotate(s, w);
if (ssunlikely(rc == -1))
goto error;
}
if (task.req) {
rc = se_dispatch(s, w, &task);
if (ssunlikely(rc == -1)) {
goto error;
}
}
se *e = (se*)s->env;
if (task.backup_complete)
se_reqonbackup(e);
if (job) {
rc = se_run(&task, w);
if (ssunlikely(rc == -1)) {
if (task.plan.plan != SI_BACKUP &&
task.plan.plan != SI_BACKUPEND) {
se_dbmalfunction(task.db);
goto error;
}
ss_mutexlock(&s->lock);
se_backuperror(s);
ss_mutexunlock(&s->lock);
}
}
if (task.gc) {
rc = se_gc(s, w);
if (ssunlikely(rc == -1))
goto error;
}
se_complete(s, &task);
ss_trace(&w->trace, "%s", "sleep");
return job;
error:
ss_trace(&w->trace, "%s", "malfunction");
return -1;
}
static int
se_gc(sescheduler *s, seworker *w)
{
ss_trace(&w->trace, "%s", "log gc");
se *e = (se*)s->env;
int rc = sl_poolgc(&e->lp);
if (ssunlikely(rc == -1))
return -1;
return 0;
}
int sc_step(sc *s, scworker *w, uint64_t vlsn)
{
sctask task;
sc_taskbegin(&task, w, vlsn);
int rc = sc_schedule(s, &task);
int rc_job = rc;
/* log rotation */
if (task.rotate) {
rc = sc_rotate(s, w);
if (ssunlikely(rc == -1))
goto error;
}
/* backup completion */
if (task.backup) {
rc = sc_backupend(s, w);
if (ssunlikely(rc == -1))
sc_backupstop(s);
}
if (rc_job == 1) {
rc = sc_execute(&task, w, vlsn);
if (ssunlikely(rc == -1)) {
if (task.plan.plan != SI_BACKUP &&
task.plan.plan != SI_BACKUPEND) {
sr_statusset(s->r->status, SR_MALFUNCTION);
goto error;
}
sc_backupstop(s);
}
}
sc_taskend(s, &task);
if (task.gc) {
rc = sc_gc(s, w);
if (ssunlikely(rc == -1))
goto error;
}
ss_trace(&w->trace, "%s", "sleep");
return rc_job;
error:
ss_trace(&w->trace, "%s", "malfunction");
return -1;
}
static inline int
sc_rotate(sc *s, scworker *w)
{
ss_trace(&w->trace, "%s", "log rotation");
int rc = sw_managerrotate_ready(s->wm);
if (rc) {
rc = sw_managerrotate(s->wm);
if (ssunlikely(rc == -1))
return -1;
}
return 0;
}
static int
se_rotate(sescheduler *s, seworker *w)
{
ss_trace(&w->trace, "%s", "log rotation");
se *e = (se*)s->env;
int rc = sl_poolrotate_ready(&e->lp, e->conf.log_rotate_wm);
if (rc) {
rc = sl_poolrotate(&e->lp);
if (ssunlikely(rc == -1))
return -1;
}
return 0;
}
static int
se_dispatch(sescheduler *s, seworker *w, setask *t)
{
ss_trace(&w->trace, "%s", "dispatcher");
se *e = (se*)s->env;
int block = t->req == 1;
do {
int rc = se_active(e);
if (ssunlikely(rc == 0))
break;
sereq *req = se_reqdispatch(e, block);
if (req) {
se_execute(req);
se_reqready(req);
}
} while (block);
return 0;
}
static int
se_dispatch(sescheduler *s, seworker *w, setask *t)
{
ss_trace(&w->trace, "%s", "dispatcher");
se *e = (se*)s->env;
int block = t->req == 1;
do {
int rc = se_active(e);
if (ssunlikely(rc == 0))
break;
sereq *req = se_reqdispatch(e, block);
if (req) {
switch (req->op) {
case SE_REQREAD:
se_execute_read(req);
break;
default: assert(0);
}
se_reqready(req);
}
} while (block);
return 0;
}
static int
se_schedule(sescheduler *s, setask *task, seworker *w)
{
ss_trace(&w->trace, "%s", "schedule");
si_planinit(&task->plan);
uint64_t now = ss_utime();
se *e = (se*)s->env;
sedb *db;
srzone *zone = se_zoneof(e);
assert(zone != NULL);
task->checkpoint_complete = 0;
task->backup_complete = 0;
task->rotate = 0;
task->req = 0;
task->gc = 0;
task->db = NULL;
ss_mutexlock(&s->lock);
/* asynchronous reqs dispatcher */
if (s->req == 0) {
switch (zone->async) {
case 2:
if (se_reqqueue(e) == 0)
break;
case 1:
s->req = 1;
task->req = zone->async;
ss_mutexunlock(&s->lock);
return 0;
}
}
/* log gc and rotation */
if (s->rotate == 0)
{
task->rotate = 1;
s->rotate = 1;
}
/* checkpoint */
int in_progress = 0;
int rc;
checkpoint:
if (s->checkpoint) {
task->plan.plan = SI_CHECKPOINT;
task->plan.a = s->checkpoint_lsn;
rc = se_schedule_plan(s, &task->plan, &db);
switch (rc) {
case 1:
s->workers_branch++;
se_dbref(db, 1);
task->db = db;
task->gc = 1;
ss_mutexunlock(&s->lock);
return 1;
case 2: /* work in progress */
in_progress = 1;
break;
case 0: /* complete checkpoint */
s->checkpoint = 0;
s->checkpoint_lsn_last = s->checkpoint_lsn;
s->checkpoint_lsn = 0;
task->checkpoint_complete = 1;
break;
}
}
/* apply zone policy */
switch (zone->mode) {
case 0: /* compact_index */
case 1: /* compact_index + branch_count prio */
assert(0);
break;
case 2: /* checkpoint */
{
if (in_progress) {
ss_mutexunlock(&s->lock);
return 0;
}
uint64_t lsn = sr_seq(&e->seq, SR_LSN);
s->checkpoint_lsn = lsn;
s->checkpoint = 1;
goto checkpoint;
}
default: /* branch + compact */
assert(zone->mode == 3);
}
/* database shutdown-drop */
if (s->workers_gc_db < zone->gc_db_prio) {
ss_spinlock(&e->dblock);
db = NULL;
if (ssunlikely(e->db_shutdown.n > 0)) {
db = (sedb*)so_listfirst(&e->db_shutdown);
if (se_dbgarbage(db)) {
so_listdel(&e->db_shutdown, &db->o);
} else {
db = NULL;
}
}
ss_spinunlock(&e->dblock);
if (ssunlikely(db)) {
if (db->dropped)
task->plan.plan = SI_DROP;
else
task->plan.plan = SI_SHUTDOWN;
s->workers_gc_db++;
se_dbref(db, 1);
task->db = db;
ss_mutexunlock(&s->lock);
return 1;
}
}
/* backup */
if (s->backup && (s->workers_backup < zone->backup_prio))
{
/* backup procedure.
*
* state 0 (start)
* -------
*
* a. disable log gc
* b. mark to start backup (state 1)
*
* state 1 (background, delayed start)
* -------
*
* a. create backup_path/<bsn.incomplete> directory
* b. create database directories
* c. create log directory
* d. state 2
*
* state 2 (background, copy)
* -------
*
* a. schedule and execute node backup which bsn < backup_bsn
* b. state 3
*
* state 3 (background, completion)
* -------
*
* a. rotate log file
* b. copy log files
* c. enable log gc, schedule gc
* d. rename <bsn.incomplete> into <bsn>
* e. set last backup, set COMPLETE
*
*/
if (s->backup == 1) {
/* state 1 */
rc = se_backupstart(s);
if (ssunlikely(rc == -1)) {
se_backuperror(s);
goto backup_error;
}
s->backup = 2;
}
/* state 2 */
task->plan.plan = SI_BACKUP;
task->plan.a = s->backup_bsn;
rc = se_schedule_plan(s, &task->plan, &db);
switch (rc) {
case 1:
s->workers_backup++;
se_dbref(db, 1);
task->db = db;
ss_mutexunlock(&s->lock);
return 1;
case 2: /* work in progress */
break;
case 0: /* state 3 */
rc = se_backupcomplete(s, w);
if (ssunlikely(rc == -1)) {
se_backuperror(s);
goto backup_error;
}
s->backup_events++;
task->gc = 1;
task->backup_complete = 1;
break;
}
backup_error:;
}
/* garbage-collection */
if (s->gc) {
if (s->workers_gc < zone->gc_prio) {
task->plan.plan = SI_GC;
task->plan.a = sx_vlsn(&e->xm);
task->plan.b = zone->gc_wm;
rc = se_schedule_plan(s, &task->plan, &db);
switch (rc) {
case 1:
s->workers_gc++;
se_dbref(db, 1);
task->db = db;
ss_mutexunlock(&s->lock);
return 1;
case 2: /* work in progress */
break;
case 0: /* state 3 */
s->gc = 0;
s->gc_last = now;
break;
}
}
} else {
if (zone->gc_prio && zone->gc_period) {
if ( (now - s->gc_last) >= ((uint64_t)zone->gc_period * 1000000) ) {
s->gc = 1;
}
}
}
/* index aging */
if (s->age) {
if (s->workers_branch < zone->branch_prio) {
task->plan.plan = SI_AGE;
task->plan.a = zone->branch_age * 1000000; /* ms */
task->plan.b = zone->branch_age_wm;
rc = se_schedule_plan(s, &task->plan, &db);
switch (rc) {
case 1:
s->workers_branch++;
se_dbref(db, 1);
task->db = db;
ss_mutexunlock(&s->lock);
return 1;
case 0:
s->age = 0;
s->age_last = now;
break;
}
}
} else {
if (zone->branch_prio && zone->branch_age_period) {
if ( (now - s->age_last) >= ((uint64_t)zone->branch_age_period * 1000000) ) {
s->age = 1;
}
}
}
/* branching */
if (s->workers_branch < zone->branch_prio)
{
/* schedule branch task using following
* priority:
*
* a. peek node with the largest in-memory index
* which is equal or greater then branch
* watermark.
* If nothing is found, stick to b.
*
* b. peek node with the largest in-memory index,
* which has oldest update time.
*
* c. if no branch work is needed, schedule a
* compaction job
*
*/
task->plan.plan = SI_BRANCH;
task->plan.a = zone->branch_wm;
rc = se_schedule_plan(s, &task->plan, &db);
if (rc == 1) {
s->workers_branch++;
se_dbref(db, 1);
task->db = db;
task->gc = 1;
ss_mutexunlock(&s->lock);
return 1;
}
}
/* compaction */
task->plan.plan = SI_COMPACT;
task->plan.a = zone->compact_wm;
task->plan.b = zone->compact_mode;
rc = se_schedule_plan(s, &task->plan, &db);
if (rc == 1) {
se_dbref(db, 1);
task->db = db;
ss_mutexunlock(&s->lock);
return 1;
}
ss_mutexunlock(&s->lock);
return 0;
}
static inline siplannerrc
sc_do(sc *s, sctask *task)
{
siplannerrc rc;
scdb *db = task->db;
sicompaction *c = &db->index->scheme.compaction;
ss_trace(&task->w->trace, "%s", "schedule");
/* checkpoint */
if (db->checkpoint) {
task->plan.plan = SI_CHECKPOINT;
task->plan.a = db->checkpoint_vlsn;
rc = si_plan(db->index, &task->plan);
switch (rc) {
case SI_PMATCH:
return rc;
case SI_PNONE:
sc_task_checkpoint_done(db, task->time);
break;
case SI_PRETRY:
break;
}
}
/* node delayed gc */
task->plan.plan = SI_NODEGC;
rc = si_plan(db->index, &task->plan);
if (rc == SI_PMATCH)
return rc;
/* backup */
if (db->backup)
{
/* backup procedure.
*
* state 0 (start)
* -------
*
* a. disable log gc
* b. mark to start backup (state 1)
*
* state 1 (background, delayed start)
* -------
*
* a. create backup_path/<bsn.incomplete> directory
* b. create database directories
* c. create log directory
* d. state 2
*
* state 2 (background, copy)
* -------
*
* a. schedule and execute node backup which bsn < backup_bsn
* b. state 3
*
* state 3 (background, completion)
* -------
*
* a. rotate log file
* b. copy log files
* c. enable log gc, schedule gc
* d. rename <bsn.incomplete> into <bsn>
* e. set last backup, set COMPLETE
*
*/
/* state 2 */
task->plan.plan = SI_BACKUP;
task->plan.a = s->backup_bsn;
rc = sc_plan(s, task, SC_QBACKUP);
switch (rc) {
case SI_PMATCH:
db->workers[SC_QBACKUP]++;
task->db = db;
return SI_PMATCH;
case SI_PNONE:
sc_task_backup_done(task->db);
assert(s->backup_in_progress > 0);
s->backup_in_progress--;
/* state 3 */
if (s->backup_in_progress == 0)
task->backup = 1;
break;
case SI_PRETRY:
break;
}
}
/* expire */
if (db->expire) {
task->plan.plan = SI_EXPIRE;
task->plan.a = db->index->scheme.expire;
rc = sc_plan(s, task, SC_QEXPIRE);
switch (rc) {
case SI_PMATCH:
db->workers[SC_QEXPIRE]++;
return SI_PMATCH;
case SI_PNONE:
sc_task_expire_done(db, task->time);
break;
case SI_PRETRY:
break;
}
}
/* garbage-collection */
if (db->gc) {
task->plan.plan = SI_GC;
task->plan.a = task->vlsn;
task->plan.b = c->gc_wm;
rc = sc_plan(s, task, SC_QGC);
switch (rc) {
case SI_PMATCH:
db->workers[SC_QGC]++;
return SI_PMATCH;
case SI_PNONE:
sc_task_gc_done(db, task->time);
break;
case SI_PRETRY:
break;
}
}
/* compaction */
task->plan.plan = SI_COMPACTION;
rc = si_plan(db->index, &task->plan);
if (rc == SI_PMATCH)
return SI_PMATCH;
si_planinit(&task->plan);
return SI_PNONE;
}
