*/
/*
* This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
*
* lustre/osp/osp_precreate.c
*
struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
LASSERT(d);
+ /* invalidate statfs data so osp_precreate_thread() can refresh */
+ d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC);
if (d->opd_pre_task)
wake_up(&d->opd_pre_waitq);
}
static void osp_pre_update_status_msfs(struct osp_device *d,
struct obd_statfs *msfs, int rc)
{
+ CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc);
if (rc)
d->opd_pre_status = rc;
else
osp_pre_update_msfs(d, msfs);
- wake_up(&d->opd_pre_user_waitq);
+ wake_up_all(&d->opd_pre_user_waitq);
}
/* Pass in the old statfs data in case the limits have changed */
union ptlrpc_async_args *aa = args;
struct obd_import *imp = req->rq_import;
struct obd_statfs *msfs;
+ struct obd_statfs *sfs;
struct osp_device *d;
u64 maxage_ns;
jiffies + cfs_time_seconds(d->opd_statfs_maxage));
d->opd_statfs_update_in_progress = 0;
- CDEBUG(D_CACHE, "updated statfs %p\n", d);
+ sfs = &d->opd_statfs;
+ CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, "
+ "%u bsize, %u reserved mb low, %u reserved mb high,"
+ "%llu files, %llu free files\n", d->opd_obd->obd_name, d,
+ sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize,
+ d->opd_reserved_mb_low, d->opd_reserved_mb_high,
+ sfs->os_files, sfs->os_ffree);
RETURN(0);
out:
d->opd_obd->obd_name,
atomic_read(&d->opd_sync_changes));
osp_sync_add_commit_cb_1s(env, d, th);
- dt_trans_stop(env, d->opd_storage, th);
}
+ dt_trans_stop(env, d->opd_storage, th);
}
out:
/* don't consider new precreation till OST is healty and
* has free space */
- return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
- (d->opd_pre_status == 0));
+ return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 ||
+ d->opd_force_creation) && (d->opd_pre_status == 0));
}
/**
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;
+
+ /* Delorphan happens only with a first MDT-OST connect. resend/replay
+ * handles objects creation on reconnects, no need to do delorhpan
+ * in this case.
+ */
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
if (rc) {
}
spin_lock(&d->opd_pre_lock);
- if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
+ if (d->opd_force_creation)
+ d->opd_pre_create_count = OST_MIN_PRECREATE;
+ else if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
grow = d->opd_pre_create_count;
spin_unlock(&d->opd_pre_lock);
if (rc) {
CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
rc);
+ if (req->rq_net_err)
+ /* have osp_precreate_reserve() to wait for repeat */
+ rc = -ENOTCONN;
GOTO(out_req, rc);
}
- LASSERT(req->rq_transno == 0);
body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
fid_to_ostid(fid, &body->oa.o_oi);
d->opd_pre_last_created_fid = *fid;
+ d->opd_force_creation = false;
spin_unlock(&d->opd_pre_lock);
CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
out_req:
/* now we can wakeup all users awaiting for objects */
osp_pre_update_status(d, rc);
- wake_up(&d->opd_pre_user_waitq);
+
+ /* pause to let osp_precreate_reserve to go first */
+ CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2);
ptlrpc_req_finished(req);
RETURN(rc);
*
* \param[in] env LU environment provided by the caller
* \param[in] d OSP device
+ * \param[in] update update or not update last used fid
*
* \retval 0 on success
* \retval negative negated errno on error
**/
static int osp_get_lastfid_from_ost(const struct lu_env *env,
- struct osp_device *d)
+ struct osp_device *d, bool update)
{
struct ptlrpc_request *req = NULL;
struct obd_import *imp;
rc = ptlrpc_queue_wait(req);
if (rc) {
- /* bad-bad OST.. let sysadm sort this out */
- if (rc == -ENOTSUPP) {
- CERROR("%s: server does not support FID: rc = %d\n",
- d->opd_obd->obd_name, -ENOTSUPP);
- }
- ptlrpc_set_import_active(imp, 0);
+ /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld),
+ * let sysadm sort this * out.
+ */
+ if (rc == -EFAULT)
+ ptlrpc_set_import_active(imp, 0);
GOTO(out, rc);
}
/* Only update the last used fid, if the OST has objects for
* this sequence, i.e. fid_oid > 0 */
- if (fid_oid(last_fid) > 0)
+ if (fid_oid(last_fid) > 0 && update)
d->opd_last_used_fid = *last_fid;
+ if (fid_oid(last_fid) == 0 &&
+ fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) {
+ /* reformatted OST, it requires creation request
+ * to recreate objects
+ */
+ d->opd_force_creation = true;
+ }
CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
PFID(last_fid));
ENTRY;
/*
+ * Do cleanup orphans only with a first connection, after that
+ * all precreate requests uses resend/replay flags to support OST
+ * failover/reconnect.
+ */
+ if (d->opd_cleanup_orphans_done) {
+ rc = osp_get_lastfid_from_ost(env, d, false);
+ RETURN(0);
+ }
+ /*
* 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
CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
+ OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val);
+
*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) {
+ if (fid_oid(&d->opd_last_used_fid) < 2 ||
+ OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
/* lastfid looks strange... ask OST */
- rc = osp_get_lastfid_from_ost(env, d);
+ rc = osp_get_lastfid_from_ost(env, d, true);
if (rc)
GOTO(out, rc);
}
if (req)
ptlrpc_req_finished(req);
+
/*
* If rc is zero, the pre-creation window should have been emptied.
* Since waking up the herd would be useless without pre-created
* this OSP isn't quite functional yet */
osp_pre_update_status(d, rc);
} else {
- wake_up(&d->opd_pre_user_waitq);
+ wake_up_all(&d->opd_pre_user_waitq);
}
} else {
spin_lock(&d->opd_pre_lock);
d->opd_pre_recovering = 0;
spin_unlock(&d->opd_pre_lock);
+ d->opd_cleanup_orphans_done = true;
}
RETURN(rc);
if (!d->opd_new_connection)
continue;
+ OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY,
+ cfs_fail_val);
d->opd_new_connection = 0;
d->opd_got_disconnected = 0;
break;
if (d->opd_pre == NULL)
continue;
+ if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) {
+ d->opd_pre_recovering = 1;
+ break;
+ }
+
/* To avoid handling different seq in precreate/orphan
* cleanup, it will hold precreate until current seq is
* used up. */
static int osp_precreate_ready_condition(const struct lu_env *env,
struct osp_device *d)
{
+ /* Bail out I/O fails to OST */
+ if (d->opd_pre_status != 0 &&
+ d->opd_pre_status != -EAGAIN &&
+ d->opd_pre_status != -ENODEV &&
+ d->opd_pre_status != -ENOTCONN &&
+ d->opd_pre_status != -ENOSPC) {
+ /* DEBUG LU-3230 */
+ if (d->opd_pre_status != -EIO)
+ CERROR("%s: precreate failed opd_pre_status %d\n",
+ d->opd_obd->obd_name, d->opd_pre_status);
+ return 1;
+ }
+
if (d->opd_pre_recovering)
return 0;
d->opd_pre_status == -ENOSPC)
return 1;
- /* Bail out I/O fails to OST */
- if (d->opd_pre_status != 0 &&
- d->opd_pre_status != -EAGAIN &&
- d->opd_pre_status != -ENODEV &&
- d->opd_pre_status != -ENOTCONN &&
- d->opd_pre_status != -ENOSPC) {
- /* DEBUG LU-3230 */
- if (d->opd_pre_status != -EIO)
- CERROR("%s: precreate failed opd_pre_status %d\n",
- d->opd_obd->obd_name, d->opd_pre_status);
- return 1;
- }
-
return 0;
}
* \retval -EAGAIN try later, slow precreation in progress
* \retval -EIO when no access to OST
*/
-int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
+int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d,
+ bool can_block)
{
time64_t expire = ktime_get_seconds() + obd_timeout;
int precreated, rc, synced = 0;
break;
}
+ if (!can_block) {
+ LASSERT(d->opd_pre);
+ rc = -ENOBUFS;
+ break;
+ }
+
+ CDEBUG(D_INFO, "%s: Sleeping on objects\n",
+ d->opd_obd->obd_name);
if (wait_event_idle_timeout(
d->opd_pre_user_waitq,
osp_precreate_ready_condition(env, d),
atomic_read(&d->opd_sync_changes),
atomic_read(&d->opd_sync_rpcs_in_progress),
d->opd_pre_status);
+ } else {
+ CDEBUG(D_INFO, "%s: Waked up, status=%d\n",
+ d->opd_obd->obd_name, d->opd_pre_status);
}
}
* XXX: decide how do we do here with resend
* if we don't resend, then client may see wrong file size
* if we do resend, then MDS thread can get stuck for quite long
- * and if we don't resend, then client will also get -EWOULDBLOCK !!
+ * and if we don't resend, then client will also get -EAGAIN !!
* (see LU-7975 and sanity/test_27F use cases)
* but let's decide not to resend/delay this truncate request to OST
* and allow Client to decide to resend, in a less agressive way from
* after_reply(), by returning -EINPROGRESS instead of
- * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
+ * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the
* end of this routine
*/
req->rq_no_resend = req->rq_no_delay = 1;
rc = ptlrpc_queue_wait(req);
if (rc) {
- /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
+ /* -EAGAIN/-EWOULDBLOCK means OST is unreachable at the moment
* since we have decided not to resend/delay, but this could
* lead to wrong size to be seen at Client side and even process
* trying to open to exit/fail if not itself handling -EAGAIN.
* So it should be better to return -EINPROGRESS instead and
* leave the decision to resend at Client side in after_reply()
*/
- if (rc == -EWOULDBLOCK) {
+ if (rc == -EAGAIN) {
rc = -EINPROGRESS;
CDEBUG(D_HA, "returning -EINPROGRESS instead of "
"-EWOULDBLOCK/-EAGAIN to allow Client to "
d->opd_pre_max_create_count = OST_MAX_PRECREATE;
d->opd_reserved_mb_high = 0;
d->opd_reserved_mb_low = 0;
+ d->opd_cleanup_orphans_done = false;
+ d->opd_force_creation = false;
RETURN(0);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff