static void osp_statfs_timer_cb(unsigned long _d)
{
struct osp_device *d = (struct osp_device *) _d;
LASSERT(d);
if (d->opd_pre != NULL && osp_precreate_running(d))
wake_up(&d->opd_pre_waitq);
}
int osp_init_precreate(struct osp_device *d)
{
struct l_wait_info lwi = { 0 };
struct task_struct *task;
ENTRY;
OBD_ALLOC_PTR(d->opd_pre);
if (d->opd_pre == NULL)
RETURN(-ENOMEM);
/* initially precreation isn't ready */
d->opd_pre_status = -EAGAIN;
fid_zero(&d->opd_pre_used_fid);
d->opd_pre_used_fid.f_oid = 1;
fid_zero(&d->opd_pre_last_created_fid);
d->opd_pre_last_created_fid.f_oid = 1;
d->opd_pre_reserved = 0;
d->opd_got_disconnected = 1;
d->opd_pre_grow_slow = 0;
d->opd_pre_grow_count = OST_MIN_PRECREATE;
d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
spin_lock_init(&d->opd_pre_lock);
init_waitqueue_head(&d->opd_pre_waitq);
init_waitqueue_head(&d->opd_pre_user_waitq);
init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
/*
* Initialize statfs-related things
*/
d->opd_statfs_maxage = 5; /* default update interval */
d->opd_statfs_fresh_till = cfs_time_shift(-1000);
CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
(unsigned long long)cfs_time_current(),
(unsigned long long)d->opd_statfs_fresh_till);
cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
/*
* start thread handling precreation and statfs updates
*/
task = kthread_run(osp_precreate_thread, d,
"osp-pre-%u-%u", d->opd_index, d->opd_group);
if (IS_ERR(task)) {
CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
RETURN(PTR_ERR(task));
}
l_wait_event(d->opd_pre_thread.t_ctl_waitq,
osp_precreate_running(d) || osp_precreate_stopped(d),
&lwi);
RETURN(0);
}
int osp_init_precreate(struct osp_device *d)
{
struct l_wait_info lwi = { 0 };
int rc;
ENTRY;
/* initially precreation isn't ready */
d->opd_pre_status = -EAGAIN;
fid_zero(&d->opd_pre_used_fid);
d->opd_pre_used_fid.f_oid = 1;
fid_zero(&d->opd_pre_last_created_fid);
d->opd_pre_last_created_fid.f_oid = 1;
d->opd_pre_reserved = 0;
d->opd_got_disconnected = 1;
d->opd_pre_grow_slow = 0;
d->opd_pre_grow_count = OST_MIN_PRECREATE;
d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
spin_lock_init(&d->opd_pre_lock);
cfs_waitq_init(&d->opd_pre_waitq);
cfs_waitq_init(&d->opd_pre_user_waitq);
cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
/*
* Initialize statfs-related things
*/
d->opd_statfs_maxage = 5; /* default update interval */
d->opd_statfs_fresh_till = cfs_time_shift(-1000);
CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
(unsigned long long)cfs_time_current(),
(unsigned long long)d->opd_statfs_fresh_till);
cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
/*
* start thread handling precreation and statfs updates
*/
rc = cfs_create_thread(osp_precreate_thread, d, 0);
if (rc < 0) {
CERROR("can't start precreate thread %d\n", rc);
RETURN(rc);
}
l_wait_event(d->opd_pre_thread.t_ctl_waitq,
osp_precreate_running(d) || osp_precreate_stopped(d),
&lwi);
RETURN(0);
}
static int osp_statfs_interpret(const struct lu_env *env,
struct ptlrpc_request *req,
union ptlrpc_async_args *aa, int rc)
{
struct obd_import *imp = req->rq_import;
struct obd_statfs *msfs;
struct osp_device *d;
ENTRY;
aa = ptlrpc_req_async_args(req);
d = aa->pointer_arg[0];
LASSERT(d);
if (rc != 0)
GOTO(out, rc);
msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
if (msfs == NULL)
GOTO(out, rc = -EPROTO);
d->opd_statfs = *msfs;
osp_pre_update_status(d, rc);
/* schedule next update */
d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
d->opd_statfs_update_in_progress = 0;
CDEBUG(D_CACHE, "updated statfs %p\n", d);
RETURN(0);
out:
/* couldn't update statfs, try again as soon as possible */
if (d->opd_pre != NULL && osp_precreate_running(d))
wake_up(&d->opd_pre_waitq);
if (req->rq_import_generation == imp->imp_generation)
CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
d->opd_obd->obd_name, rc);
RETURN(rc);
}
static int osp_precreate_thread(void *_arg)
{
struct osp_device *d = _arg;
struct ptlrpc_thread *thread = &d->opd_pre_thread;
struct l_wait_info lwi = { 0 };
struct lu_env env;
int rc;
ENTRY;
rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
if (rc) {
CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
rc);
RETURN(rc);
}
spin_lock(&d->opd_pre_lock);
thread->t_flags = SVC_RUNNING;
spin_unlock(&d->opd_pre_lock);
wake_up(&thread->t_ctl_waitq);
while (osp_precreate_running(d)) {
/*
* need to be connected to OST
*/
while (osp_precreate_running(d)) {
l_wait_event(d->opd_pre_waitq,
!osp_precreate_running(d) ||
d->opd_new_connection,
&lwi);
if (!d->opd_new_connection)
continue;
d->opd_new_connection = 0;
d->opd_got_disconnected = 0;
break;
}
if (!osp_precreate_running(d))
break;
LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) {
/* Get new sequence for client first */
LASSERT(d->opd_exp != NULL);
d->opd_obd->u.cli.cl_seq->lcs_exp =
class_export_get(d->opd_exp);
rc = osp_init_pre_fid(d);
if (rc != 0) {
class_export_put(d->opd_exp);
d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
CERROR("%s: init pre fid error: rc = %d\n",
d->opd_obd->obd_name, rc);
continue;
}
}
osp_statfs_update(d);
/*
* Clean up orphans or recreate missing objects.
*/
rc = osp_precreate_cleanup_orphans(&env, d);
if (rc != 0)
continue;
/*
* connected, can handle precreates now
*/
while (osp_precreate_running(d)) {
l_wait_event(d->opd_pre_waitq,
!osp_precreate_running(d) ||
osp_precreate_near_empty(&env, d) ||
osp_statfs_need_update(d) ||
d->opd_got_disconnected, &lwi);
if (!osp_precreate_running(d))
break;
/* something happened to the connection
* have to start from the beginning */
if (d->opd_got_disconnected)
break;
if (osp_statfs_need_update(d))
osp_statfs_update(d);
/* To avoid handling different seq in precreate/orphan
* cleanup, it will hold precreate until current seq is
* used up. */
if (unlikely(osp_precreate_end_seq(&env, d) &&
!osp_create_end_seq(&env, d)))
continue;
if (unlikely(osp_precreate_end_seq(&env, d) &&
osp_create_end_seq(&env, d))) {
LCONSOLE_INFO("%s:"LPX64" is used up."
" Update to new seq\n",
d->opd_obd->obd_name,
fid_seq(&d->opd_pre_last_created_fid));
rc = osp_precreate_rollover_new_seq(&env, d);
if (rc)
continue;
}
if (osp_precreate_near_empty(&env, d)) {
rc = osp_precreate_send(&env, d);
/* osp_precreate_send() sets opd_pre_status
* in case of error, that prevent the using of
* failed device. */
if (rc < 0 && rc != -ENOSPC &&
rc != -ETIMEDOUT && rc != -ENOTCONN)
CERROR("%s: cannot precreate objects:"
" rc = %d\n",
d->opd_obd->obd_name, rc);
}
}
}
thread->t_flags = SVC_STOPPED;
lu_env_fini(&env);
wake_up(&thread->t_ctl_waitq);
RETURN(0);
}
/**
* asks OST to clean precreate orphans
* and gets next id for new objects
*/
static int osp_precreate_cleanup_orphans(struct lu_env *env,
struct osp_device *d)
{
struct osp_thread_info *osi = osp_env_info(env);
struct lu_fid *last_fid = &osi->osi_fid;
struct ptlrpc_request *req = NULL;
struct obd_import *imp;
struct ost_body *body;
struct l_wait_info lwi = { 0 };
int update_status = 0;
int rc;
int diff;
ENTRY;
/*
* wait for local recovery to finish, so we can cleanup orphans
* orphans are all objects since "last used" (assigned), but
* there might be objects reserved and in some cases they won't
* be used. we can't cleanup them till we're sure they won't be
* used. also can't we allow new reservations because they may
* end up getting orphans being cleaned up below. so we block
* new reservations and wait till all reserved objects either
* user or released.
*/
spin_lock(&d->opd_pre_lock);
d->opd_pre_recovering = 1;
spin_unlock(&d->opd_pre_lock);
/*
* The locking above makes sure the opd_pre_reserved check below will
* catch all osp_precreate_reserve() calls who find
* "!opd_pre_recovering".
*/
l_wait_event(d->opd_pre_waitq,
(!d->opd_pre_reserved && d->opd_recovery_completed) ||
!osp_precreate_running(d) || d->opd_got_disconnected,
&lwi);
if (!osp_precreate_running(d) || d->opd_got_disconnected)
GOTO(out, rc = -EAGAIN);
CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
*last_fid = d->opd_last_used_fid;
/* The OSP should already get the valid seq now */
LASSERT(!fid_is_zero(last_fid));
if (fid_oid(&d->opd_last_used_fid) < 2) {
/* lastfid looks strange... ask OST */
rc = osp_get_lastfid_from_ost(env, d);
if (rc)
GOTO(out, rc);
}
imp = d->opd_obd->u.cli.cl_import;
LASSERT(imp);
req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
if (req == NULL)
GOTO(out, rc = -ENOMEM);
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
if (rc) {
ptlrpc_request_free(req);
req = NULL;
GOTO(out, rc);
}
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
GOTO(out, rc = -EPROTO);
body->oa.o_flags = OBD_FL_DELORPHAN;
body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
ptlrpc_request_set_replen(req);
/* Don't resend the delorphan req */
req->rq_no_resend = req->rq_no_delay = 1;
rc = ptlrpc_queue_wait(req);
if (rc) {
update_status = 1;
GOTO(out, rc);
}
body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
GOTO(out, rc = -EPROTO);
/*
* OST provides us with id new pool starts from in body->oa.o_id
*/
ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
spin_lock(&d->opd_pre_lock);
diff = lu_fid_diff(&d->opd_last_used_fid, last_fid);
if (diff > 0) {
d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
d->opd_pre_last_created_fid = d->opd_last_used_fid;
} else {
d->opd_pre_grow_count = OST_MIN_PRECREATE;
d->opd_pre_last_created_fid = *last_fid;
}
/*
* This empties the pre-creation pool and effectively blocks any new
* reservations.
*/
LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
LUSTRE_DATA_SEQ_MAX_WIDTH);
d->opd_pre_used_fid = d->opd_pre_last_created_fid;
d->opd_pre_grow_slow = 0;
spin_unlock(&d->opd_pre_lock);
CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
"last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
out:
if (req)
ptlrpc_req_finished(req);
spin_lock(&d->opd_pre_lock);
d->opd_pre_recovering = 0;
spin_unlock(&d->opd_pre_lock);
/*
* If rc is zero, the pre-creation window should have been emptied.
* Since waking up the herd would be useless without pre-created
* objects, we defer the signal to osp_precreate_send() in that case.
*/
if (rc != 0) {
if (update_status) {
CERROR("%s: cannot cleanup orphans: rc = %d\n",
d->opd_obd->obd_name, rc);
/* we can't proceed from here, OST seem to
* be in a bad shape, better to wait for
* a new instance of the server and repeat
* from the beginning. notify possible waiters
* this OSP isn't quite functional yet */
osp_pre_update_status(d, rc);
} else {
wake_up(&d->opd_pre_user_waitq);
}
}
RETURN(rc);
}
