static int
se_dbscheme_init(sedb *db, char *name, int size)
{
se *e = se_of(&db->o);
/* database id */
uint32_t id = sr_seq(&e->seq, SR_DSN);
sr_seq(&e->seq, SR_DSNNEXT);
/* prepare index scheme */
sischeme *scheme = db->scheme;
if (size == 0)
size = strlen(name);
scheme->name = ss_malloc(&e->a, size + 1);
if (ssunlikely(scheme->name == NULL))
goto error;
memcpy(scheme->name, name, size);
scheme->name[size] = 0;
scheme->id = id;
scheme->sync = 2;
scheme->mmap = 0;
scheme->storage = SI_SCACHE;
scheme->node_size = 64 * 1024 * 1024;
scheme->node_compact_load = 0;
scheme->node_page_size = 128 * 1024;
scheme->node_page_checksum = 1;
scheme->compression_copy = 0;
scheme->compression_cold = 0;
scheme->compression_cold_if = &ss_nonefilter;
scheme->compression_hot = 0;
scheme->compression_hot_if = &ss_nonefilter;
scheme->temperature = 0;
scheme->expire = 0;
scheme->amqf = 0;
scheme->fmt_storage = SF_RAW;
scheme->lru = 0;
scheme->lru_step = 128 * 1024;
scheme->buf_gc_wm = 1024 * 1024;
scheme->storage_sz = ss_strdup(&e->a, "cache");
if (ssunlikely(scheme->storage_sz == NULL))
goto error;
scheme->compression_cold_sz =
ss_strdup(&e->a, scheme->compression_cold_if->name);
if (ssunlikely(scheme->compression_cold_sz == NULL))
goto error;
scheme->compression_hot_sz =
ss_strdup(&e->a, scheme->compression_hot_if->name);
if (ssunlikely(scheme->compression_hot_sz == NULL))
goto error;
sf_upsertinit(&scheme->fmt_upsert);
sf_schemeinit(&scheme->scheme);
return 0;
error:
sr_oom(&e->error);
return -1;
}
sxstate sx_set_autocommit(sxmanager *m, sxindex *index, sx *x, svlog *log, svv *v)
{
if (sslikely(m->count_rw == 0)) {
sx_init(m, x, log);
svlogv lv;
lv.index_id = index->dsn;
lv.next = UINT32_MAX;
lv.v = v;
lv.ptr = NULL;
sv_logadd(x->log, index->r, &lv);
sr_seq(index->r->seq, SR_TSNNEXT);
sx_promote(x, SX_COMMIT);
return SX_COMMIT;
}
sx_begin(m, x, SX_RW, log, 0);
int rc = sx_set(x, index, v);
if (ssunlikely(rc == -1)) {
sx_rollback(x);
return SX_ROLLBACK;
}
sxstate s = sx_prepare(x, NULL, NULL);
switch (s) {
case SX_PREPARE:
s = sx_commit(x);
break;
case SX_LOCK:
s = sx_rollback(x);
break;
case SX_ROLLBACK:
break;
default:
assert(0);
}
return s;
}
sxstate sx_set_autocommit(sxmanager *m, sxindex *index, sx *x, svv *v)
{
if (sslikely(m->count_rw == 0)) {
sx_init(m, x);
svlogv lv;
lv.id = index->dsn;
lv.next = UINT32_MAX;
sv_init(&lv.v, &sv_vif, v, NULL);
sv_logadd(&x->log, m->r->a, &lv, index->ptr);
sr_seq(m->r->seq, SR_TSNNEXT);
return SXCOMMIT;
}
sx_begin(m, x, SXRW, 0);
int rc = sx_set(x, index, v);
if (ssunlikely(rc == -1)) {
sx_rollback(x);
return SXROLLBACK;
}
sxstate s = sx_prepare(x, NULL, NULL);
if (sslikely(s == SXPREPARE))
sx_commit(x);
else
if (s == SXLOCK)
sx_rollback(x);
return s;
}
int se_scheduler_backup(void *arg)
{
se *e = arg;
sescheduler *s = &e->sched;
if (ssunlikely(e->conf.backup_path == NULL)) {
sr_error(&e->error, "%s", "backup is not enabled");
return -1;
}
/* begin backup procedure
* state 0
*
* disable log garbage-collection
*/
sl_poolgc_enable(&e->lp, 0);
ss_mutexlock(&s->lock);
if (ssunlikely(s->backup > 0)) {
ss_mutexunlock(&s->lock);
sl_poolgc_enable(&e->lp, 1);
/* in progress */
return 0;
}
uint64_t bsn = sr_seq(&e->seq, SR_BSNNEXT);
s->backup = 1;
s->backup_bsn = bsn;
ss_mutexunlock(&s->lock);
return 0;
}
sxstate sx_setstmt(sxmanager *m, sxindex *index, sv *v)
{
sr_seq(m->seq, SR_TSNNEXT);
ssrbnode *n = NULL;
int rc = sx_match(&index->i, index->scheme, sv_pointer(v), sv_size(v), &n);
if (rc == 0 && n)
return SXLOCK;
return SXCOMMIT;
}
int sc_ctl_checkpoint(sc *s)
{
uint64_t lsn = sr_seq(s->r->seq, SR_LSN);
ss_mutexlock(&s->lock);
s->checkpoint_lsn = lsn;
s->checkpoint = 1;
ss_mutexunlock(&s->lock);
return 0;
}
int sc_ctl_snapshot(sc *s)
{
uint64_t ssn = sr_seq(s->r->seq, SR_SSNNEXT);
ss_mutexlock(&s->lock);
s->snapshot_ssn = ssn;
s->snapshot = 1;
ss_mutexunlock(&s->lock);
return 0;
}
int se_scheduler_checkpoint(void *arg)
{
se *o = arg;
sescheduler *s = &o->sched;
uint64_t lsn = sr_seq(&o->seq, SR_LSN);
ss_mutexlock(&s->lock);
s->checkpoint_lsn = lsn;
s->checkpoint = 1;
ss_mutexunlock(&s->lock);
return 0;
}
int se_scheduler_snapshot(void *arg)
{
se *o = arg;
sescheduler *s = &o->sched;
uint64_t ssn = sr_seq(&o->seq, SR_SSNNEXT);
ss_mutexlock(&s->lock);
s->snapshot_ssn = ssn;
s->snapshot = 1;
ss_mutexunlock(&s->lock);
return 0;
}
uint64_t sx_vlsn(sxmanager *m)
{
ss_spinlock(&m->lock);
uint64_t vlsn;
if (sx_count(m) > 0) {
ssrbnode *node = ss_rbmin(&m->i);
sx *min = sscast(node, sx, node);
vlsn = min->vlsn;
} else {
vlsn = sr_seq(m->r->seq, SR_LSN);
}
ss_spinunlock(&m->lock);
return vlsn;
}
static int
se_dbscheme_init(sedb *db, char *name, int size)
{
se *e = se_of(&db->o);
/* database id */
uint32_t id = sr_seq(&e->seq, SR_DSN);
sr_seq(&e->seq, SR_DSNNEXT);
/* prepare index scheme */
sischeme *scheme = db->scheme;
if (size == 0)
size = strlen(name);
scheme->name = ss_malloc(&e->a, size + 1);
if (ssunlikely(scheme->name == NULL))
goto error;
memcpy(scheme->name, name, size);
scheme->name[size] = 0;
scheme->id = id;
scheme->sync = 1;
scheme->mmap = 1;
scheme->direct_io = 0;
scheme->direct_io_page_size = 4096;
scheme->direct_io_buffer_size = 8 * 1024 * 1024;
scheme->compression = 0;
scheme->compression_if = &ss_nonefilter;
scheme->expire = 0;
scheme->buf_gc_wm = 1024 * 1024;
scheme->compression_sz =
ss_strdup(&e->a, scheme->compression_if->name);
if (ssunlikely(scheme->compression_sz == NULL))
goto error;
sf_upsertinit(&scheme->upsert);
sf_schemeinit(&scheme->scheme);
return 0;
error:
sr_oom(&e->error);
return -1;
}
int se_reqread(sereq *r)
{
sereqarg *arg = &r->arg;
sedb *db = (sedb*)r->db;
uint32_t keysize;
void *key;
if (sslikely(arg->v.v)) {
keysize = sv_size(&arg->v);
key = sv_pointer(&arg->v);
} else {
keysize = 0;
key = NULL;
}
char *prefix;
uint32_t prefixsize;
if (arg->vprefix.v) {
void *vptr = sv_vpointer(arg->vprefix.v);
prefix = sf_key(vptr, 0);
prefixsize = sf_keysize(vptr, 0);
} else {
prefix = NULL;
prefixsize = 0;
}
if (sslikely(arg->vlsn_generate))
arg->vlsn = sr_seq(db->r.seq, SR_LSN);
sitx x;
si_begin(&x, &db->index, 1);
siread q;
si_readopen(&q, &x, arg->cache,
arg->order,
arg->vlsn,
prefix,
prefixsize, key, keysize);
if (arg->update)
si_readupdate(&q, &arg->vup, arg->update_eq);
if (arg->cache_only)
si_readcache_only(&q);
if (arg->has)
si_readhas(&q);
r->rc = si_read(&q);
r->read_disk = q.read_disk;
r->read_cache = q.read_cache;
r->v = q.result.v;
si_readclose(&q);
si_commit(&x);
return r->rc;
}
sxstate sx_prepare(sx *x, sxpreparef prepare, void *arg)
{
uint64_t lsn = sr_seq(x->manager->seq, SR_LSN);
/* proceed read-only transactions */
if (x->type == SX_RO || sv_logcount_write(x->log) == 0)
return sx_promote(x, SX_PREPARE);
ssiter i;
ss_iterinit(ss_bufiter, &i);
ss_iteropen(ss_bufiter, &i, &x->log->buf, sizeof(svlogv));
sxstate rc;
for (; ss_iterhas(ss_bufiter, &i); ss_iternext(ss_bufiter, &i))
{
svlogv *lv = ss_iterof(ss_bufiter, &i);
sxv *v = lv->ptr;
if ((int)v->lo == x->log_read)
break;
if (sx_vaborted(v))
return sx_promote(x, SX_ROLLBACK);
if (sslikely(v->prev == NULL)) {
rc = sx_preparecb(x, lv, lsn, prepare, arg);
if (ssunlikely(rc != 0))
return sx_promote(x, SX_ROLLBACK);
continue;
}
if (sx_vcommitted(v->prev)) {
if (v->prev->csn > x->csn)
return sx_promote(x, SX_ROLLBACK);
continue;
}
/* force commit for read-only conflicts */
sxindex *i = v->prev->index;
if (sv_vflags(v->prev->v, i->r) & SVGET) {
rc = sx_preparecb(x, lv, lsn, prepare, arg);
if (ssunlikely(rc != 0))
return sx_promote(x, SX_ROLLBACK);
continue;
}
return sx_promote(x, SX_LOCK);
}
return sx_promote(x, SX_PREPARE);
}
int sc_backupstart(sc *s)
{
/* begin backup procedure
* state 0
*
* disable log garbage-collection
*/
sl_poolgc_enable(s->lp, 0);
ss_mutexlock(&s->lock);
if (ssunlikely(s->backup > 0)) {
ss_mutexunlock(&s->lock);
sl_poolgc_enable(s->lp, 1);
/* in progress */
return 1;
}
uint64_t bsn = sr_seq(s->r->seq, SR_BSNNEXT);
s->backup = 1;
s->backup_bsn = bsn;
ss_mutexunlock(&s->lock);
return 0;
}
sxstate sx_prepare(sx *x, sxpreparef prepare, void *arg)
{
uint64_t lsn = sr_seq(x->manager->r->seq, SR_LSN);
/* proceed read-only transactions */
if (x->type == SXRO || sv_logcount_write(&x->log) == 0)
return sx_promote(x, SXPREPARE);
ssiter i;
ss_iterinit(ss_bufiter, &i);
ss_iteropen(ss_bufiter, &i, &x->log.buf, sizeof(svlogv));
for (; ss_iterhas(ss_bufiter, &i); ss_iternext(ss_bufiter, &i))
{
svlogv *lv = ss_iterof(ss_bufiter, &i);
sxv *v = lv->v.v;
if ((int)v->lo == x->log_read)
break;
if (sx_vaborted(v))
return sx_promote(x, SXROLLBACK);
if (sslikely(v->prev == NULL)) {
if (prepare && lsn != x->vlsn) {
sxindex *i = v->index;
if (prepare(x, &lv->v, arg, i->ptr))
return sx_promote(x, SXROLLBACK);
}
continue;
}
if (sx_vcommitted(v->prev)) {
if (v->prev->csn > x->csn)
return sx_promote(x, SXROLLBACK);
continue;
}
/* force commit for read-only conflicts */
if (v->prev->v->flags & SVGET)
continue;
return sx_promote(x, SXLOCK);
}
return sx_promote(x, SXPREPARE);
}
sxstate sx_get_autocommit(sxmanager *m, sxindex *index)
{
(void)m;
sr_seq(index->r->seq, SR_TSNNEXT);
return SX_COMMIT;
}
int se_scheduler_branch(void *arg)
{
sedb *db = arg;
se *e = se_of(&db->o);
srzone *z = se_zoneof(e);
seworker *w = se_workerpool_pop(&e->sched.workers, &e->r);
if (ssunlikely(w == NULL))
return -1;
int rc;
while (1) {
uint64_t vlsn = sx_vlsn(&e->xm);
uint64_t vlsn_lru = si_lru_vlsn(&db->index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_BRANCH,
.a = z->branch_wm,
.b = 0,
.c = 0,
.node = NULL
};
rc = si_plan(&db->index, &plan);
if (rc == 0)
break;
rc = si_execute(&db->index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
se_workerpool_push(&e->sched.workers, w);
return rc;
}
int se_scheduler_compact(void *arg)
{
sedb *db = arg;
se *e = se_of(&db->o);
srzone *z = se_zoneof(e);
seworker *w = se_workerpool_pop(&e->sched.workers, &e->r);
if (ssunlikely(w == NULL))
return -1;
int rc;
while (1) {
uint64_t vlsn = sx_vlsn(&e->xm);
uint64_t vlsn_lru = si_lru_vlsn(&db->index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_COMPACT,
.a = z->compact_wm,
.b = z->compact_mode,
.c = 0,
.node = NULL
};
rc = si_plan(&db->index, &plan);
if (rc == 0)
break;
rc = si_execute(&db->index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
se_workerpool_push(&e->sched.workers, w);
return rc;
}
int se_scheduler_compact_index(void *arg)
{
sedb *db = arg;
se *e = se_of(&db->o);
srzone *z = se_zoneof(e);
seworker *w = se_workerpool_pop(&e->sched.workers, &e->r);
if (ssunlikely(w == NULL))
return -1;
int rc;
while (1) {
uint64_t vlsn = sx_vlsn(&e->xm);
uint64_t vlsn_lru = si_lru_vlsn(&db->index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_COMPACT_INDEX,
.a = z->branch_wm,
.b = 0,
.c = 0,
.node = NULL
};
rc = si_plan(&db->index, &plan);
if (rc == 0)
break;
rc = si_execute(&db->index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
se_workerpool_push(&e->sched.workers, w);
return rc;
}
int se_scheduler_anticache(void *arg)
{
se *o = arg;
sescheduler *s = &o->sched;
uint64_t asn = sr_seq(&o->seq, SR_ASNNEXT);
ss_mutexlock(&s->lock);
s->anticache_asn = asn;
s->anticache_storage = o->conf.anticache;
s->anticache = 1;
ss_mutexunlock(&s->lock);
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;
}
int se_scheduler_branch(void *arg)
{
sedb *db = arg;
se *e = se_of(&db->o);
srzone *z = se_zoneof(e);
seworker stub;
se_workerstub_init(&stub);
int rc;
while (1) {
uint64_t vlsn = sx_vlsn(&e->xm);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_BRANCH,
.a = z->branch_wm,
.b = 0,
.c = 0,
.node = NULL
};
rc = si_plan(&db->index, &plan);
if (rc == 0)
break;
rc = si_execute(&db->index, &stub.dc, &plan, vlsn);
if (ssunlikely(rc == -1))
break;
}
se_workerstub_free(&stub, &db->r);
return rc;
}
int se_scheduler_compact(void *arg)
{
sedb *db = arg;
se *e = se_of(&db->o);
srzone *z = se_zoneof(e);
seworker stub;
se_workerstub_init(&stub);
int rc;
while (1) {
uint64_t vlsn = sx_vlsn(&e->xm);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_COMPACT,
.a = z->compact_wm,
.b = z->compact_mode,
.c = 0,
.node = NULL
};
rc = si_plan(&db->index, &plan);
if (rc == 0)
break;
rc = si_execute(&db->index, &stub.dc, &plan, vlsn);
if (ssunlikely(rc == -1))
break;
}
se_workerstub_free(&stub, &db->r);
return rc;
}
int se_scheduler_checkpoint(void *arg)
{
se *o = arg;
sescheduler *s = &o->sched;
uint64_t lsn = sr_seq(&o->seq, SR_LSN);
ss_mutexlock(&s->lock);
s->checkpoint_lsn = lsn;
s->checkpoint = 1;
ss_mutexunlock(&s->lock);
return 0;
}
int sc_ctl_branch(sc *s, uint64_t vlsn, si *index)
{
sr *r = s->r;
int rc = sr_statusactive(r->status);
if (ssunlikely(rc == 0))
return 0;
srzone *z = sr_zoneof(r);
scworker *w = sc_workerpool_pop(&s->wp, r);
if (ssunlikely(w == NULL))
return -1;
while (1) {
uint64_t vlsn_lru = si_lru_vlsn(index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_BRANCH,
.a = z->branch_wm,
.b = 0,
.c = 0,
.node = NULL
};
rc = si_plan(index, &plan);
if (rc == 0)
break;
rc = si_execute(index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
sc_workerpool_push(&s->wp, w);
return rc;
}
int sc_ctl_compact(sc *s, uint64_t vlsn, si *index)
{
sr *r = s->r;
int rc = sr_statusactive(r->status);
if (ssunlikely(rc == 0))
return 0;
srzone *z = sr_zoneof(r);
scworker *w = sc_workerpool_pop(&s->wp, r);
if (ssunlikely(w == NULL))
return -1;
while (1) {
uint64_t vlsn_lru = si_lru_vlsn(index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_COMPACT,
.a = z->compact_wm,
.b = z->compact_mode,
.c = 0,
.node = NULL
};
rc = si_plan(index, &plan);
if (rc == 0)
break;
rc = si_execute(index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
sc_workerpool_push(&s->wp, w);
return rc;
}
int sc_ctl_compact_index(sc *s, uint64_t vlsn, si *index)
{
sr *r = s->r;
int rc = sr_statusactive(r->status);
if (ssunlikely(rc == 0))
return 0;
srzone *z = sr_zoneof(r);
scworker *w = sc_workerpool_pop(&s->wp, r);
if (ssunlikely(w == NULL))
return -1;
while (1) {
uint64_t vlsn_lru = si_lru_vlsn(index);
siplan plan = {
.explain = SI_ENONE,
.plan = SI_COMPACT_INDEX,
.a = z->branch_wm,
.b = 0,
.c = 0,
.node = NULL
};
rc = si_plan(index, &plan);
if (rc == 0)
break;
rc = si_execute(index, &w->dc, &plan, vlsn, vlsn_lru);
if (ssunlikely(rc == -1))
break;
}
sc_workerpool_push(&s->wp, w);
return rc;
}
int sc_ctl_anticache(sc *s)
{
uint64_t asn = sr_seq(s->r->seq, SR_ASNNEXT);
ss_mutexlock(&s->lock);
s->anticache_asn = asn;
s->anticache_storage = s->anticache_limit;
s->anticache = 1;
ss_mutexunlock(&s->lock);
return 0;
}
static inline int
si_split(si *index, sdc *c, ssbuf *result,
sinode *parent,
ssiter *i,
uint64_t size_node,
uint64_t size_stream,
uint32_t stream,
uint64_t vlsn)
{
sr *r = &index->r;
uint32_t timestamp = ss_timestamp();
int rc;
sdmergeconf mergeconf = {
.stream = stream,
.size_stream = size_stream,
.size_node = size_node,
.size_page = index->scheme.compaction.node_page_size,
.checksum = index->scheme.compaction.node_page_checksum,
.expire = index->scheme.expire,
.timestamp = timestamp,
.compression = index->scheme.compression,
.compression_if = index->scheme.compression_if,
.direct_io = index->scheme.direct_io,
.direct_io_page_size = index->scheme.direct_io_page_size,
.vlsn = vlsn
};
sinode *n = NULL;
sdmerge merge;
rc = sd_mergeinit(&merge, r, i, &c->build, &c->build_index,
&c->upsert, &mergeconf);
if (ssunlikely(rc == -1))
return -1;
while ((rc = sd_merge(&merge)) > 0)
{
/* create new node */
uint64_t id = sr_seq(index->r.seq, SR_NSNNEXT);
n = si_nodenew(r, id, parent->id);
if (ssunlikely(n == NULL))
goto error;
rc = si_nodecreate(n, r, &index->scheme);
if (ssunlikely(rc == -1))
goto error;
/* write pages */
uint64_t offset;
offset = sd_iosize(&c->io, &n->file);
while ((rc = sd_mergepage(&merge, offset)) == 1) {
rc = sd_writepage(r, &n->file, &c->io, merge.build);
if (ssunlikely(rc == -1))
goto error;
offset = sd_iosize(&c->io, &n->file);
}
if (ssunlikely(rc == -1))
goto error;
offset = sd_iosize(&c->io, &n->file);
rc = sd_mergeend(&merge, offset);
if (ssunlikely(rc == -1))
goto error;
/* write index */
rc = sd_writeindex(r, &n->file, &c->io, &merge.index);
if (ssunlikely(rc == -1))
goto error;
/* mmap mode */
if (index->scheme.mmap) {
rc = si_nodemap(n, r);
if (ssunlikely(rc == -1))
goto error;
}
/* add node to the list */
rc = ss_bufadd(result, index->r.a, &n, sizeof(sinode*));
if (ssunlikely(rc == -1)) {
sr_oom_malfunction(index->r.e);
goto error;
}
n->index = merge.index;
}
if (ssunlikely(rc == -1))
goto error;
return 0;
error:
if (n)
si_nodefree(n, r, 0);
sd_mergefree(&merge);
si_splitfree(result, r);
return -1;
}
static int
si_merge(si *index, sdc *c, sinode *node,
uint64_t vlsn,
ssiter *stream,
uint64_t size_stream,
uint32_t n_stream)
{
sr *r = &index->r;
ssbuf *result = &c->a;
ssiter i;
/* begin compaction.
*
* Split merge stream into a number of
* a new nodes.
*/
int rc;
rc = si_split(index, c, result,
node, stream,
index->scheme.compaction.node_size,
size_stream,
n_stream,
vlsn);
if (ssunlikely(rc == -1))
return -1;
SS_INJECTION(r->i, SS_INJECTION_SI_COMPACTION_0,
si_splitfree(result, r);
sr_malfunction(r->e, "%s", "error injection");
return -1);
/* mask removal of a single node as a
* single node update */
int count = ss_bufused(result) / sizeof(sinode*);
int count_index;
si_lock(index);
count_index = index->n;
si_unlock(index);
sinode *n;
if (ssunlikely(count == 0 && count_index == 1))
{
n = si_bootstrap(index, node->id);
if (ssunlikely(n == NULL))
return -1;
rc = ss_bufadd(result, r->a, &n, sizeof(sinode*));
if (ssunlikely(rc == -1)) {
sr_oom_malfunction(r->e);
si_nodefree(n, r, 1);
return -1;
}
count++;
}
/* commit compaction changes */
si_lock(index);
svindex *j = si_nodeindex(node);
si_plannerremove(&index->p, node);
si_nodesplit(node);
switch (count) {
case 0: /* delete */
si_remove(index, node);
si_redistribute_index(index, r, c, node);
break;
case 1: /* self update */
n = *(sinode**)result->s;
n->i0 = *j;
n->used = j->used;
si_nodelock(n);
si_replace(index, node, n);
si_plannerupdate(&index->p, n);
break;
default: /* split */
rc = si_redistribute(index, r, c, node, result);
if (ssunlikely(rc == -1)) {
si_unlock(index);
si_splitfree(result, r);
return -1;
}
ss_iterinit(ss_bufiterref, &i);
ss_iteropen(ss_bufiterref, &i, result, sizeof(sinode*));
n = ss_iterof(ss_bufiterref, &i);
n->used = n->i0.used;
si_nodelock(n);
si_replace(index, node, n);
si_plannerupdate(&index->p, n);
for (ss_iternext(ss_bufiterref, &i); ss_iterhas(ss_bufiterref, &i);
ss_iternext(ss_bufiterref, &i)) {
n = ss_iterof(ss_bufiterref, &i);
n->used = n->i0.used;
si_nodelock(n);
si_insert(index, n);
si_plannerupdate(&index->p, n);
}
break;
}
sv_indexinit(j);
si_unlock(index);
/* compaction completion */
/* seal nodes */
ss_iterinit(ss_bufiterref, &i);
ss_iteropen(ss_bufiterref, &i, result, sizeof(sinode*));
while (ss_iterhas(ss_bufiterref, &i))
{
n = ss_iterof(ss_bufiterref, &i);
if (index->scheme.sync) {
rc = ss_filesync(&n->file);
if (ssunlikely(rc == -1)) {
sr_malfunction(r->e, "db file '%s' sync error: %s",
ss_pathof(&n->file.path),
strerror(errno));
return -1;
}
}
rc = si_noderename_seal(n, r, &index->scheme);
if (ssunlikely(rc == -1)) {
si_nodefree(node, r, 0);
return -1;
}
SS_INJECTION(r->i, SS_INJECTION_SI_COMPACTION_3,
si_nodefree(node, r, 0);
sr_malfunction(r->e, "%s", "error injection");
return -1);
ss_iternext(ss_bufiterref, &i);
}
SS_INJECTION(r->i, SS_INJECTION_SI_COMPACTION_1,
si_nodefree(node, r, 0);
sr_malfunction(r->e, "%s", "error injection");
return -1);
/* gc node */
uint16_t refs = si_noderefof(node);
if (sslikely(refs == 0)) {
rc = si_nodefree(node, r, 1);
if (ssunlikely(rc == -1))
return -1;
} else {
/* node concurrently being read, schedule for
* delayed removal */
si_nodegc(node, r, &index->scheme);
si_lock(index);
ss_listappend(&index->gc, &node->gc);
index->gc_count++;
si_unlock(index);
}
SS_INJECTION(r->i, SS_INJECTION_SI_COMPACTION_2,
sr_malfunction(r->e, "%s", "error injection");
return -1);
/* complete new nodes */
ss_iterinit(ss_bufiterref, &i);
ss_iteropen(ss_bufiterref, &i, result, sizeof(sinode*));
while (ss_iterhas(ss_bufiterref, &i))
{
n = ss_iterof(ss_bufiterref, &i);
rc = si_noderename_complete(n, r, &index->scheme);
if (ssunlikely(rc == -1))
return -1;
SS_INJECTION(r->i, SS_INJECTION_SI_COMPACTION_4,
sr_malfunction(r->e, "%s", "error injection");
return -1);
ss_iternext(ss_bufiterref, &i);
}
/* unlock */
si_lock(index);
ss_iterinit(ss_bufiterref, &i);
ss_iteropen(ss_bufiterref, &i, result, sizeof(sinode*));
while (ss_iterhas(ss_bufiterref, &i))
{
n = ss_iterof(ss_bufiterref, &i);
si_nodeunlock(n);
ss_iternext(ss_bufiterref, &i);
}
si_unlock(index);
return 0;
}
static int
se_dbscheme_init(sedb *db, char *name)
{
se *e = se_of(&db->o);
/* prepare index scheme */
sischeme *scheme = &db->scheme;
scheme->name = ss_strdup(&e->a, name);
if (ssunlikely(scheme->name == NULL))
goto e0;
scheme->id = sr_seq(&e->seq, SR_DSNNEXT);
scheme->sync = 2;
scheme->mmap = 0;
scheme->storage = SI_SCACHE;
scheme->cache_mode = 0;
scheme->cache_sz = NULL;
scheme->node_size = 64 * 1024 * 1024;
scheme->node_compact_load = 0;
scheme->node_page_size = 128 * 1024;
scheme->node_page_checksum = 1;
scheme->compression_key = 0;
scheme->compression = 0;
scheme->compression_if = &ss_nonefilter;
scheme->compression_branch = 0;
scheme->compression_branch_if = &ss_nonefilter;
scheme->amqf = 0;
scheme->fmt = SF_KV;
scheme->fmt_storage = SF_SRAW;
scheme->path_fail_on_exists = 0;
scheme->path_fail_on_drop = 1;
scheme->lru = 0;
scheme->lru_step = 128 * 1024;
scheme->buf_gc_wm = 1024 * 1024;
scheme->storage_sz = ss_strdup(&e->a, "cache");
if (ssunlikely(scheme->storage_sz == NULL))
goto e1;
scheme->compression_sz =
ss_strdup(&e->a, scheme->compression_if->name);
if (ssunlikely(scheme->compression_sz == NULL))
goto e1;
scheme->compression_branch_sz =
ss_strdup(&e->a, scheme->compression_branch_if->name);
if (ssunlikely(scheme->compression_branch_sz == NULL))
goto e1;
sf_upsertinit(&scheme->fmt_upsert);
scheme->fmt_sz = ss_strdup(&e->a, "kv");
if (ssunlikely(scheme->fmt_sz == NULL))
goto e1;
/* init single key part as string */
int rc;
sr_schemeinit(&scheme->scheme);
srkey *part = sr_schemeadd(&scheme->scheme);
rc = sr_keysetname(part, &e->a, "key");
if (ssunlikely(rc == -1))
goto e1;
rc = sr_keyset(part, &e->a, "string");
if (ssunlikely(rc == -1))
goto e1;
return 0;
e1:
si_schemefree(&db->scheme, &db->r);
e0:
sr_oom(&e->error);
return -1;
}