static void osp_statfs_timer_cb(unsigned long _d)
{
struct osp_device *d = (struct osp_device *) _d;
LASSERT(d);
cfs_waitq_signal(&d->opd_pre_waitq);
}
/*
* this function relies on reservation made before
*/
int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
struct lu_fid *fid)
{
/* grab next id from the pool */
spin_lock(&d->opd_pre_lock);
LASSERTF(lu_fid_diff(&d->opd_pre_used_fid,
&d->opd_pre_last_created_fid) < 0,
"next fid "DFID" last created fid "DFID"\n",
PFID(&d->opd_pre_used_fid),
PFID(&d->opd_pre_last_created_fid));
d->opd_pre_used_fid.f_oid++;
memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
d->opd_pre_reserved--;
/*
* last_used_id must be changed along with getting new id otherwise
* we might miscalculate gap causing object loss or leak
*/
osp_update_last_fid(d, fid);
spin_unlock(&d->opd_pre_lock);
/*
* probably main thread suspended orphan cleanup till
* all reservations are released, see comment in
* osp_precreate_thread() just before orphan cleanup
*/
if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
cfs_waitq_signal(&d->opd_pre_waitq);
return 0;
}
static void seq_fid_alloc_fini(struct lu_client_seq *seq)
{
LASSERT(seq->lcs_update == 1);
mutex_lock(&seq->lcs_mutex);
--seq->lcs_update;
cfs_waitq_signal(&seq->lcs_waitq);
}
/*
* Workitem scheduled with (serial == 1) is strictly serialised not only with
* itself, but also with others scheduled this way.
*
* Now there's only one static serialised queue, but in the future more might
* be added, and even dynamic creation of serialised queues might be supported.
*/
void
cfs_wi_schedule(cfs_workitem_t *wi)
{
cfs_wi_sched_t *sched = cfs_wi_to_sched(wi);
LASSERT (!cfs_in_interrupt()); /* because we use plain spinlock */
LASSERT (!sched->ws_shuttingdown);
cfs_wi_sched_lock(sched);
if (!wi->wi_scheduled) {
LASSERT (cfs_list_empty(&wi->wi_list));
wi->wi_scheduled = 1;
if (!wi->wi_running) {
cfs_list_add_tail(&wi->wi_list, &sched->ws_runq);
#ifdef __KERNEL__
cfs_waitq_signal(&sched->ws_waitq);
#endif
} else {
cfs_list_add(&wi->wi_list, &sched->ws_rerunq);
}
}
LASSERT (!cfs_list_empty(&wi->wi_list));
cfs_wi_sched_unlock(sched);
return;
}
void mdt_ck_timer_callback(unsigned long castmeharder)
{
struct mdt_device *mdt = (struct mdt_device *)castmeharder;
struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
ENTRY;
thread_add_flags(thread, SVC_EVENT);
cfs_waitq_signal(&thread->t_ctl_waitq);
EXIT;
}
int LL_PROC_PROTO(proc_fail_loc)
{
int rc;
long old_fail_loc = cfs_fail_loc;
rc = ll_proc_dolongvec(table, write, filp, buffer, lenp, ppos);
if (old_fail_loc != cfs_fail_loc)
cfs_waitq_signal(&cfs_race_waitq);
return rc;
}
void mdt_ck_thread_stop(struct mdt_device *mdt)
{
struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
if (!thread_is_running(thread))
return;
thread_set_flags(thread, SVC_STOPPING);
cfs_waitq_signal(&thread->t_ctl_waitq);
l_wait_condition(thread->t_ctl_waitq, thread_is_stopped(thread));
}
/*
* XXX: there might be a case where removed object(s) do not add free
* space (empty object). if the number of such deletions is high, then
* we can start to update statfs too often - a rpc storm
* TODO: some throttling is needed
*/
void osp_statfs_need_now(struct osp_device *d)
{
if (!d->opd_statfs_update_in_progress) {
/*
* if current status is -ENOSPC (lack of free space on OST)
* then we should poll OST immediately once object destroy
* is replied
*/
d->opd_statfs_fresh_till = cfs_time_shift(-1);
cfs_timer_disarm(&d->opd_statfs_timer);
cfs_waitq_signal(&d->opd_pre_waitq);
}
}
/*
* the function updates current precreation status used: functional or not
*
* rc is a last code from the transport, rc == 0 meaning transport works
* well and users of lod can use objects from this OSP
*
* the status depends on current usage of OST
*/
void osp_pre_update_status(struct osp_device *d, int rc)
{
struct obd_statfs *msfs = &d->opd_statfs;
int old = d->opd_pre_status;
__u64 used;
d->opd_pre_status = rc;
if (rc)
goto out;
/* Add a bit of hysteresis so this flag isn't continually flapping,
* and ensure that new files don't get extremely fragmented due to
* only a small amount of available space in the filesystem.
* We want to set the NOSPC flag when there is less than ~0.1% free
* and clear it when there is at least ~0.2% free space, so:
* avail < ~0.1% max max = avail + used
* 1025 * avail < avail + used used = blocks - free
* 1024 * avail < used
* 1024 * avail < blocks - free
* avail < ((blocks - free) >> 10)
*
* On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
* lose that amount of space so in those cases we report no space left
* if their is less than 1 GB left. */
if (likely(msfs->os_type)) {
used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
1 << 30);
if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
d->opd_pre_status = -ENOSPC;
if (old != -ENOSPC)
CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
LPU64" free, "LPU64" used, "LPU64" "
"avail -> %d: rc = %d\n",
d->opd_obd->obd_name, msfs->os_blocks,
msfs->os_bfree, used, msfs->os_bavail,
d->opd_pre_status, rc);
CDEBUG(D_INFO,
"non-commited changes: %lu, in progress: %u\n",
d->opd_syn_changes, d->opd_syn_rpc_in_progress);
} else if (old == -ENOSPC) {
d->opd_pre_status = 0;
d->opd_pre_grow_slow = 0;
d->opd_pre_grow_count = OST_MIN_PRECREATE;
cfs_waitq_signal(&d->opd_pre_waitq);
CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
" free, "LPU64" used, "LPU64" avail -> %d: "
"rc = %d\n", d->opd_obd->obd_name,
msfs->os_blocks, msfs->os_bfree, used,
msfs->os_bavail, d->opd_pre_status, rc);
}
}
void osp_precreate_fini(struct osp_device *d)
{
struct ptlrpc_thread *thread = &d->opd_pre_thread;
ENTRY;
cfs_timer_disarm(&d->opd_statfs_timer);
thread->t_flags = SVC_STOPPING;
cfs_waitq_signal(&d->opd_pre_waitq);
cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
EXIT;
}
static void lcw_dispatch_stop(void)
{
ENTRY;
LASSERT(lcw_refcount == 0);
CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
cfs_set_bit(LCW_FLAG_STOP, &lcw_flags);
cfs_waitq_signal(&lcw_event_waitq);
cfs_wait_for_completion(&lcw_stop_completion);
CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
EXIT;
}
/* return a page that has 'len' bytes left at the end */
static struct cfs_trace_page *
cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
{
struct cfs_trace_page *tage;
if (tcd->tcd_cur_pages > 0) {
__LASSERT(!cfs_list_empty(&tcd->tcd_pages));
tage = cfs_tage_from_list(tcd->tcd_pages.prev);
if (tage->used + len <= CFS_PAGE_SIZE)
return tage;
}
if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
if (tcd->tcd_cur_stock_pages > 0) {
tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
-- tcd->tcd_cur_stock_pages;
cfs_list_del_init(&tage->linkage);
} else {
tage = cfs_tage_alloc(CFS_ALLOC_ATOMIC);
if (tage == NULL) {
if (printk_ratelimit())
printk(CFS_KERN_WARNING
"cannot allocate a tage (%ld)\n",
tcd->tcd_cur_pages);
return NULL;
}
}
tage->used = 0;
tage->cpu = cfs_smp_processor_id();
tage->type = tcd->tcd_type;
cfs_list_add_tail(&tage->linkage, &tcd->tcd_pages);
tcd->tcd_cur_pages++;
if (tcd->tcd_cur_pages > 8 && thread_running) {
struct tracefiled_ctl *tctl = &trace_tctl;
/*
* wake up tracefiled to process some pages.
*/
cfs_waitq_signal(&tctl->tctl_waitq);
}
return tage;
}
return NULL;
}
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 */
cfs_waitq_signal(&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);
}
/** Queues DONE_WRITING if
* - done writing is allowed;
* - inode has no no dirty pages; */
void ll_queue_done_writing(struct inode *inode, unsigned long flags)
{
struct ll_inode_info *lli = ll_i2info(inode);
struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
ENTRY;
spin_lock(&lli->lli_lock);
lli->lli_flags |= flags;
if ((lli->lli_flags & LLIF_DONE_WRITING) &&
cfs_list_empty(&club->cob_pending_list)) {
struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
CWARN("ino %lu/%u(flags %u) som valid it just after "
"recovery\n",
inode->i_ino, inode->i_generation,
lli->lli_flags);
/* DONE_WRITING is allowed and inode has no dirty page. */
spin_lock(&lcq->lcq_lock);
LASSERT(cfs_list_empty(&lli->lli_close_list));
CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
inode->i_ino, inode->i_generation);
cfs_list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
/* Avoid a concurrent insertion into the close thread queue:
* an inode is already in the close thread, open(), write(),
* close() happen, epoch is closed as the inode is marked as
* LLIF_EPOCH_PENDING. When pages are written inode should not
* be inserted into the queue again, clear this flag to avoid
* it. */
lli->lli_flags &= ~LLIF_DONE_WRITING;
cfs_waitq_signal(&lcq->lcq_waitq);
spin_unlock(&lcq->lcq_lock);
}
spin_unlock(&lli->lli_lock);
EXIT;
}
static int filter_quota_setinfo(struct obd_device *obd, void *data)
{
struct obd_export *exp = data;
struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
struct obd_import *imp = exp->exp_imp_reverse;
ENTRY;
LASSERT(imp != NULL);
/* setup the quota context import */
cfs_spin_lock(&qctxt->lqc_lock);
if (qctxt->lqc_import != NULL) {
cfs_spin_unlock(&qctxt->lqc_lock);
if (qctxt->lqc_import == imp)
CDEBUG(D_WARNING, "%s: lqc_import(%p) of obd(%p) was "
"activated already.\n", obd->obd_name, imp, obd);
else
CERROR("%s: lqc_import(%p:%p) of obd(%p) was "
"activated by others.\n", obd->obd_name,
qctxt->lqc_import, imp, obd);
} else {
qctxt->lqc_import = imp;
/* make imp's connect flags equal relative exp's connect flags
* adding it to avoid the scan export list */
imp->imp_connect_data.ocd_connect_flags |=
(exp->exp_connect_flags &
(OBD_CONNECT_QUOTA64 | OBD_CONNECT_CHANGE_QS));
cfs_spin_unlock(&qctxt->lqc_lock);
CDEBUG(D_QUOTA, "%s: lqc_import(%p) of obd(%p) is reactivated "
"now.\n", obd->obd_name, imp, obd);
cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
/* start quota slave recovery thread. (release high limits) */
qslave_start_recovery(obd, qctxt);
}
RETURN(0);
}
static void lcw_cb(ulong_ptr_t data)
{
struct lc_watchdog *lcw = (struct lc_watchdog *)data;
ENTRY;
if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
EXIT;
return;
}
lcw->lcw_state = LC_WATCHDOG_EXPIRED;
cfs_spin_lock_bh(&lcw->lcw_lock);
LASSERT(cfs_list_empty(&lcw->lcw_list));
cfs_spin_lock_bh(&lcw_pending_timers_lock);
lcw->lcw_refcount++; /* +1 for pending list */
cfs_list_add(&lcw->lcw_list, &lcw_pending_timers);
cfs_waitq_signal(&lcw_event_waitq);
cfs_spin_unlock_bh(&lcw_pending_timers_lock);
cfs_spin_unlock_bh(&lcw->lcw_lock);
EXIT;
}
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);
cfs_waitq_signal(&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);
cfs_waitq_signal(&thread->t_ctl_waitq);
RETURN(0);
}
static int mdt_ck_thread_main(void *args)
{
struct mdt_device *mdt = args;
struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
struct lustre_capa_key *bkey = &mdt->mdt_capa_keys[0],
*rkey = &mdt->mdt_capa_keys[1];
struct lustre_capa_key *tmp;
struct lu_env env;
struct mdt_thread_info *info;
struct md_device *next;
struct l_wait_info lwi = { 0 };
mdsno_t mdsnum;
int rc;
ENTRY;
cfs_daemonize_ctxt("mdt_ck");
cfs_block_allsigs();
thread_set_flags(thread, SVC_RUNNING);
cfs_waitq_signal(&thread->t_ctl_waitq);
rc = lu_env_init(&env, LCT_MD_THREAD|LCT_REMEMBER|LCT_NOREF);
if (rc)
RETURN(rc);
thread->t_env = &env;
env.le_ctx.lc_thread = thread;
env.le_ctx.lc_cookie = 0x1;
info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
LASSERT(info != NULL);
tmp = &info->mti_capa_key;
mdsnum = mdt_md_site(mdt)->ms_node_id;
while (1) {
l_wait_event(thread->t_ctl_waitq,
thread_is_stopping(thread) ||
thread_is_event(thread),
&lwi);
if (thread_is_stopping(thread))
break;
thread_clear_flags(thread, SVC_EVENT);
if (cfs_time_before(cfs_time_current(), mdt->mdt_ck_expiry))
break;
*tmp = *rkey;
make_capa_key(tmp, mdsnum, rkey->lk_keyid);
next = mdt->mdt_child;
rc = next->md_ops->mdo_update_capa_key(&env, next, tmp);
if (!rc) {
cfs_spin_lock(&capa_lock);
*bkey = *rkey;
*rkey = *tmp;
cfs_spin_unlock(&capa_lock);
rc = write_capa_keys(&env, mdt, mdt->mdt_capa_keys);
if (rc) {
cfs_spin_lock(&capa_lock);
*rkey = *bkey;
memset(bkey, 0, sizeof(*bkey));
cfs_spin_unlock(&capa_lock);
} else {
set_capa_key_expiry(mdt);
DEBUG_CAPA_KEY(D_SEC, rkey, "new");
}
}
if (rc) {
DEBUG_CAPA_KEY(D_ERROR, rkey, "update failed for");
/* next retry is in 300 sec */
mdt->mdt_ck_expiry = jiffies + 300 * CFS_HZ;
}
cfs_timer_arm(&mdt->mdt_ck_timer, mdt->mdt_ck_expiry);
CDEBUG(D_SEC, "mdt_ck_timer %lu\n", mdt->mdt_ck_expiry);
}
lu_env_fini(&env);
thread_set_flags(thread, SVC_STOPPED);
cfs_waitq_signal(&thread->t_ctl_waitq);
RETURN(0);
}
int libcfs_debug_dumplog_thread(void *arg)
{
libcfs_debug_dumplog_internal(arg);
cfs_waitq_signal(&debug_ctlwq);
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);
d->opd_pre_recovering = 0;
/*
* 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 {
cfs_waitq_signal(&d->opd_pre_user_waitq);
}
}
RETURN(rc);
}
static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
{
struct osp_thread_info *oti = osp_env_info(env);
struct ptlrpc_request *req;
struct obd_import *imp;
struct ost_body *body;
int rc, grow, diff;
struct lu_fid *fid = &oti->osi_fid;
ENTRY;
/* don't precreate new objects till OST healthy and has free space */
if (unlikely(d->opd_pre_status)) {
CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
d->opd_obd->obd_name, d->opd_pre_status);
RETURN(0);
}
/*
* if not connection/initialization is compeleted, ignore
*/
imp = d->opd_obd->u.cli.cl_import;
LASSERT(imp);
req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
if (req == NULL)
RETURN(-ENOMEM);
req->rq_request_portal = OST_CREATE_PORTAL;
/* we should not resend create request - anyway we will have delorphan
* and kill these objects */
req->rq_no_delay = req->rq_no_resend = 1;
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
if (rc) {
ptlrpc_request_free(req);
RETURN(rc);
}
spin_lock(&d->opd_pre_lock);
if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
grow = d->opd_pre_grow_count;
spin_unlock(&d->opd_pre_lock);
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
LASSERT(body);
*fid = d->opd_pre_last_created_fid;
rc = osp_precreate_fids(env, d, fid, &grow);
if (rc == 1) {
/* Current seq has been used up*/
if (!osp_is_fid_client(d)) {
osp_pre_update_status(d, -ENOSPC);
rc = -ENOSPC;
}
cfs_waitq_signal(&d->opd_pre_waitq);
GOTO(out_req, rc);
}
if (!osp_is_fid_client(d)) {
/* Non-FID client will always send seq 0 because of
* compatiblity */
LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
fid->f_seq = 0;
}
fid_to_ostid(fid, &body->oa.o_oi);
body->oa.o_valid = OBD_MD_FLGROUP;
ptlrpc_request_set_replen(req);
rc = ptlrpc_queue_wait(req);
if (rc) {
CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
rc);
GOTO(out_req, rc);
}
LASSERT(req->rq_transno == 0);
body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
GOTO(out_req, rc = -EPROTO);
ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0,
"reply fid "DFID" pre used fid "DFID"\n", PFID(fid),
PFID(&d->opd_pre_used_fid));
diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid);
spin_lock(&d->opd_pre_lock);
if (diff < grow) {
/* the OST has not managed to create all the
* objects we asked for */
d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
d->opd_pre_grow_slow = 1;
} else {
/* the OST is able to keep up with the work,
* we could consider increasing grow_count
* next time if needed */
d->opd_pre_grow_slow = 0;
}
d->opd_pre_last_created_fid = *fid;
spin_unlock(&d->opd_pre_lock);
CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
PFID(&d->opd_pre_last_created_fid));
out_req:
/* now we can wakeup all users awaiting for objects */
osp_pre_update_status(d, rc);
cfs_waitq_signal(&d->opd_pre_user_waitq);
ptlrpc_req_finished(req);
RETURN(rc);
}
/*
* called to reserve object in the pool
* return codes:
* ENOSPC - no space on corresponded OST
* EAGAIN - precreation is in progress, try later
* EIO - no access to OST
*/
int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
{
struct l_wait_info lwi;
cfs_time_t expire = cfs_time_shift(obd_timeout);
int precreated, rc;
ENTRY;
LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
"Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
PFID(&d->opd_pre_used_fid));
/*
* wait till:
* - preallocation is done
* - no free space expected soon
* - can't connect to OST for too long (obd_timeout)
* - OST can allocate fid sequence.
*/
while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
rc == -ENODEV || rc == -EAGAIN) {
/*
* increase number of precreations
*/
precreated = osp_objs_precreated(env, d);
if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
d->opd_pre_grow_slow == 0 &&
precreated <= (d->opd_pre_grow_count / 4 + 1)) {
spin_lock(&d->opd_pre_lock);
d->opd_pre_grow_slow = 1;
d->opd_pre_grow_count *= 2;
spin_unlock(&d->opd_pre_lock);
}
spin_lock(&d->opd_pre_lock);
precreated = osp_objs_precreated(env, d);
if (precreated > d->opd_pre_reserved &&
!d->opd_pre_recovering) {
d->opd_pre_reserved++;
spin_unlock(&d->opd_pre_lock);
rc = 0;
/* XXX: don't wake up if precreation is in progress */
if (osp_precreate_near_empty_nolock(env, d) &&
!osp_precreate_end_seq_nolock(env, d))
cfs_waitq_signal(&d->opd_pre_waitq);
break;
}
spin_unlock(&d->opd_pre_lock);
/*
* all precreated objects have been used and no-space
* status leave us no chance to succeed very soon
* but if there is destroy in progress, then we should
* wait till that is done - some space might be released
*/
if (unlikely(rc == -ENOSPC)) {
if (d->opd_syn_changes) {
/* force local commit to release space */
dt_commit_async(env, d->opd_storage);
}
if (d->opd_syn_rpc_in_progress) {
/* just wait till destroys are done */
/* see l_wait_even() few lines below */
}
if (d->opd_syn_changes +
d->opd_syn_rpc_in_progress == 0) {
/* no hope for free space */
break;
}
}
/* XXX: don't wake up if precreation is in progress */
cfs_waitq_signal(&d->opd_pre_waitq);
lwi = LWI_TIMEOUT(expire - cfs_time_current(),
osp_precreate_timeout_condition, d);
if (cfs_time_aftereq(cfs_time_current(), expire)) {
rc = -ETIMEDOUT;
break;
}
l_wait_event(d->opd_pre_user_waitq,
osp_precreate_ready_condition(env, d), &lwi);
}
RETURN(rc);
}
void cfs_complete(cfs_completion_t *c)
{
LASSERT(c != NULL);
c->done = 1;
cfs_waitq_signal(&c->wait);
}
