* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2012, Intel, Inc.
+ * Copyright (c) 2012, 2013, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*
* Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
* Author: Mikhail Pershin <mike.pershin@intel.com>
+ * Author: Di Wang <di.wang@intel.com>
*/
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
#define DEBUG_SUBSYSTEM S_MDS
#include "osp_internal.h"
d->opd_statfs_fresh_till);
}
+/*
+ * OSP tries to maintain pool of available objects so that calls to create
+ * objects don't block most of time
+ *
+ * each time OSP gets connected to OST, we should start from precreation cleanup
+ */
+static inline bool osp_precreate_running(struct osp_device *d)
+{
+ return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
+}
+
+static inline bool osp_precreate_stopped(struct osp_device *d)
+{
+ return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
+}
+
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);
+ if (d->opd_pre != NULL && osp_precreate_running(d))
+ wake_up(&d->opd_pre_waitq);
}
static int osp_statfs_interpret(const struct lu_env *env,
RETURN(0);
out:
/* couldn't update statfs, try again as soon as possible */
- cfs_waitq_signal(&d->opd_pre_waitq);
+ if (d->opd_pre != NULL && osp_precreate_running(d))
+ wake_up(&d->opd_pre_waitq);
+
if (req->rq_import_generation == imp->imp_generation)
- CERROR("%s: couldn't update statfs: rc = %d\n",
+ CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
d->opd_obd->obd_name, rc);
RETURN(rc);
}
*/
d->opd_statfs_fresh_till = cfs_time_shift(-1);
cfs_timer_disarm(&d->opd_statfs_timer);
- cfs_waitq_signal(&d->opd_pre_waitq);
+ wake_up(&d->opd_pre_waitq);
}
}
-
-/*
- * OSP tries to maintain pool of available objects so that calls to create
- * objects don't block most of time
- *
- * each time OSP gets connected to OST, we should start from precreation cleanup
- */
-static inline int osp_precreate_running(struct osp_device *d)
+static inline int osp_objs_precreated(const struct lu_env *env,
+ struct osp_device *osp)
{
- return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
+ return osp_fid_diff(&osp->opd_pre_last_created_fid,
+ &osp->opd_pre_used_fid);
}
-static inline int osp_precreate_stopped(struct osp_device *d)
+static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
+ struct osp_device *d)
{
- return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
-}
-
-static inline int osp_precreate_near_empty_nolock(struct osp_device *d)
-{
- int window = d->opd_pre_last_created - d->opd_pre_used_id;
+ int window = osp_objs_precreated(env, d);
/* don't consider new precreation till OST is healty and
* has free space */
(d->opd_pre_status == 0));
}
-static inline int osp_precreate_near_empty(struct osp_device *d)
+static inline int osp_precreate_near_empty(const struct lu_env *env,
+ struct osp_device *d)
{
int rc;
/* XXX: do we really need locking here? */
- cfs_spin_lock(&d->opd_pre_lock);
- rc = osp_precreate_near_empty_nolock(d);
- cfs_spin_unlock(&d->opd_pre_lock);
+ spin_lock(&d->opd_pre_lock);
+ rc = osp_precreate_near_empty_nolock(env, d);
+ spin_unlock(&d->opd_pre_lock);
+ return rc;
+}
+
+static inline int osp_create_end_seq(const struct lu_env *env,
+ struct osp_device *osp)
+{
+ struct lu_fid *fid = &osp->opd_pre_used_fid;
+ int rc;
+
+ spin_lock(&osp->opd_pre_lock);
+ rc = osp_fid_end_seq(env, fid);
+ spin_unlock(&osp->opd_pre_lock);
return rc;
}
-static int osp_precreate_send(struct osp_device *d)
+/**
+ * Write fid into last_oid/last_seq file.
+ **/
+int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
+ struct lu_fid *fid, int sync)
+{
+ struct osp_thread_info *oti = osp_env_info(env);
+ struct lu_buf *lb_oid = &oti->osi_lb;
+ struct lu_buf *lb_oseq = &oti->osi_lb2;
+ loff_t oid_off;
+ loff_t oseq_off;
+ struct thandle *th;
+ int rc;
+ ENTRY;
+
+ /* Note: through f_oid is only 32bits, it will also write
+ * 64 bits for oid to keep compatiblity with the previous
+ * version. */
+ lb_oid->lb_buf = &fid->f_oid;
+ lb_oid->lb_len = sizeof(obd_id);
+ oid_off = sizeof(obd_id) * osp->opd_index;
+
+ lb_oseq->lb_buf = &fid->f_seq;
+ lb_oseq->lb_len = sizeof(obd_id);
+ oseq_off = sizeof(obd_id) * osp->opd_index;
+
+ th = dt_trans_create(env, osp->opd_storage);
+ if (IS_ERR(th))
+ RETURN(PTR_ERR(th));
+
+ th->th_sync |= sync;
+ rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
+ lb_oid, oid_off, th);
+ if (rc != 0)
+ GOTO(out, rc);
+
+ rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
+ lb_oseq, oseq_off, th);
+ if (rc != 0)
+ GOTO(out, rc);
+
+ rc = dt_trans_start_local(env, osp->opd_storage, th);
+ if (rc != 0)
+ GOTO(out, rc);
+
+ rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
+ &oid_off, th);
+ if (rc != 0) {
+ CERROR("%s: can not write to last seq file: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ GOTO(out, rc);
+ }
+ rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
+ &oseq_off, th);
+ if (rc) {
+ CERROR("%s: can not write to last seq file: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ GOTO(out, rc);
+ }
+out:
+ dt_trans_stop(env, osp->opd_storage, th);
+ RETURN(rc);
+}
+
+int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
{
+ struct lu_fid *fid = &osp_env_info(env)->osi_fid;
+ struct lu_fid *last_fid = &osp->opd_last_used_fid;
+ int rc;
+ ENTRY;
+
+ rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
+ if (rc != 0) {
+ CERROR("%s: alloc fid error: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ RETURN(rc);
+ }
+
+ fid->f_oid = 1;
+ fid->f_ver = 0;
+ LASSERTF(fid_seq(fid) != fid_seq(last_fid),
+ "fid "DFID", last_fid "DFID"\n", PFID(fid),
+ PFID(last_fid));
+
+ rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
+ if (rc != 0) {
+ CERROR("%s: Can not update oid/seq file: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ RETURN(rc);
+ }
+
+ LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
+ osp->opd_obd->obd_name, fid_seq(last_fid),
+ fid_seq(fid));
+ /* Update last_xxx to the new seq */
+ spin_lock(&osp->opd_pre_lock);
+ osp->opd_last_used_fid = *fid;
+ osp->opd_gap_start_fid = *fid;
+ osp->opd_pre_used_fid = *fid;
+ osp->opd_pre_last_created_fid = *fid;
+ spin_unlock(&osp->opd_pre_lock);
+
+ RETURN(rc);
+}
+
+/**
+ * alloc fids for precreation.
+ * rc = 0 Success, @grow is the count of real allocation.
+ * rc = 1 Current seq is used up.
+ * rc < 0 Other error.
+ **/
+static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
+ struct lu_fid *fid, int *grow)
+{
+ struct osp_thread_info *osi = osp_env_info(env);
+ __u64 end;
+ int i = 0;
+
+ if (fid_is_idif(fid)) {
+ struct lu_fid *last_fid;
+ struct ost_id *oi = &osi->osi_oi;
+
+ spin_lock(&osp->opd_pre_lock);
+ last_fid = &osp->opd_pre_last_created_fid;
+ fid_to_ostid(last_fid, oi);
+ end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
+ *grow = end - ostid_id(oi);
+ ostid_set_id(oi, ostid_id(oi) + *grow);
+ spin_unlock(&osp->opd_pre_lock);
+
+ if (*grow == 0)
+ return 1;
+
+ ostid_to_fid(fid, oi, osp->opd_index);
+ return 0;
+ }
+
+ spin_lock(&osp->opd_pre_lock);
+ *fid = osp->opd_pre_last_created_fid;
+ end = fid->f_oid;
+ end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
+ *grow = end - fid->f_oid;
+ fid->f_oid += end - fid->f_oid;
+ spin_unlock(&osp->opd_pre_lock);
+
+ CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
+ *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
+
+ return *grow > 0 ? 0 : 1;
+}
+
+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 */
RETURN(rc);
}
- cfs_spin_lock(&d->opd_pre_lock);
+ 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;
- cfs_spin_unlock(&d->opd_pre_lock);
+ spin_unlock(&d->opd_pre_lock);
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
LASSERT(body);
- body->oa.o_id = d->opd_pre_last_created + grow;
- body->oa.o_seq = FID_SEQ_OST_MDT0; /* XXX: support for CMD? */
+
+ *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;
+ }
+ wake_up(&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);
+ CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
+ rc);
GOTO(out_req, rc);
}
LASSERT(req->rq_transno == 0);
if (body == NULL)
GOTO(out_req, rc = -EPROTO);
- CDEBUG(D_HA, "new last_created %lu\n", (unsigned long) body->oa.o_id);
- LASSERT(body->oa.o_id > d->opd_pre_used_id);
+ ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
+ if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
+ CERROR("%s: precreate fid "DFID" < local used fid "DFID
+ ": rc = %d\n", d->opd_obd->obd_name,
+ PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
+ GOTO(out_req, rc = -ESTALE);
+ }
- diff = body->oa.o_id - d->opd_pre_last_created;
+ diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
- cfs_spin_lock(&d->opd_pre_lock);
+ spin_lock(&d->opd_pre_lock);
if (diff < grow) {
/* the OST has not managed to create all the
* objects we asked for */
* next time if needed */
d->opd_pre_grow_slow = 0;
}
- d->opd_pre_last_created = body->oa.o_id;
- cfs_spin_unlock(&d->opd_pre_lock);
- CDEBUG(D_OTHER, "current precreated pool: %llu-%llu\n",
- d->opd_pre_used_id, d->opd_pre_last_created);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
+ fid_to_ostid(fid, &body->oa.o_oi);
+
+ 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);
+ wake_up(&d->opd_pre_user_waitq);
ptlrpc_req_finished(req);
RETURN(rc);
}
-
-static int osp_get_lastid_from_ost(struct osp_device *d)
+static int osp_get_lastfid_from_ost(const struct lu_env *env,
+ struct osp_device *d)
{
- struct ptlrpc_request *req;
+ struct ptlrpc_request *req = NULL;
struct obd_import *imp;
- obd_id *reply;
+ struct lu_fid *last_fid;
char *tmp;
- int rc;
+ int rc;
+ ENTRY;
imp = d->opd_obd->u.cli.cl_import;
LASSERT(imp);
- req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_ID);
+ req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
if (req == NULL)
RETURN(-ENOMEM);
- req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
- RCL_CLIENT, sizeof(KEY_LAST_ID));
+ req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
+ sizeof(KEY_LAST_FID));
+
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
if (rc) {
ptlrpc_request_free(req);
RETURN(rc);
}
- tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
- memcpy(tmp, KEY_LAST_ID, sizeof(KEY_LAST_ID));
+ tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
+ memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
req->rq_no_delay = req->rq_no_resend = 1;
+ last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
+ fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
+
ptlrpc_request_set_replen(req);
+
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);
GOTO(out, rc);
}
- reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
- if (reply == NULL)
+ last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
+ if (last_fid == NULL) {
+ CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
GOTO(out, rc = -EPROTO);
+ }
- d->opd_last_used_id = *reply;
- CDEBUG(D_HA, "%s: got last_id "LPU64" from OST\n",
- d->opd_obd->obd_name, d->opd_last_used_id);
+ if (!fid_is_sane(last_fid)) {
+ CERROR("%s: Got insane last_fid "DFID"\n",
+ d->opd_obd->obd_name, PFID(last_fid));
+ GOTO(out, rc = -EPROTO);
+ }
+
+ /* 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)
+ d->opd_last_used_fid = *last_fid;
+
+ CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
+ PFID(last_fid));
out:
ptlrpc_req_finished(req);
RETURN(rc);
-
}
/**
* asks OST to clean precreate orphans
* and gets next id for new objects
*/
-static int osp_precreate_cleanup_orphans(struct osp_device *d)
+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;
- LASSERT(d->opd_recovery_completed);
- LASSERT(d->opd_pre_reserved == 0);
-
- CDEBUG(D_HA, "%s: going to cleanup orphans since "LPU64"\n",
- d->opd_obd->obd_name, d->opd_last_used_id);
-
- if (d->opd_last_used_id < 2) {
- /* lastid looks strange... ask OST */
- rc = osp_get_lastid_from_ost(d);
+ /*
+ * 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);
}
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
if (rc) {
ptlrpc_request_free(req);
+ req = NULL;
GOTO(out, rc);
}
body->oa.o_flags = OBD_FL_DELORPHAN;
body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
- body->oa.o_seq = FID_SEQ_OST_MDT0;
- /* remove from NEXT after used one */
- body->oa.o_id = d->opd_last_used_id + 1;
+ fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
ptlrpc_request_set_replen(req);
req->rq_no_resend = req->rq_no_delay = 1;
rc = ptlrpc_queue_wait(req);
- if (rc)
+ if (rc) {
+ update_status = 1;
GOTO(out, rc);
+ }
body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
/*
* OST provides us with id new pool starts from in body->oa.o_id
*/
- cfs_spin_lock(&d->opd_pre_lock);
- if (le64_to_cpu(d->opd_last_used_id) > body->oa.o_id) {
- d->opd_pre_grow_count = OST_MIN_PRECREATE +
- le64_to_cpu(d->opd_last_used_id) -
- body->oa.o_id;
- d->opd_pre_last_created = le64_to_cpu(d->opd_last_used_id) + 1;
+ ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
+
+ spin_lock(&d->opd_pre_lock);
+ diff = osp_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 = body->oa.o_id + 1;
+ d->opd_pre_last_created_fid = *last_fid;
}
- d->opd_pre_used_id = d->opd_pre_last_created - 1;
+ /*
+ * 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;
- cfs_spin_unlock(&d->opd_pre_lock);
-
- CDEBUG(D_HA, "Got last_id "LPU64" from OST, last_used is "LPU64
- ", next "LPU64"\n", body->oa.o_id,
- le64_to_cpu(d->opd_last_used_id), d->opd_pre_used_id);
+ 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);
}
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);
d->opd_syn_changes, d->opd_syn_rpc_in_progress);
} else if (old == -ENOSPC) {
d->opd_pre_status = 0;
+ spin_lock(&d->opd_pre_lock);
d->opd_pre_grow_slow = 0;
d->opd_pre_grow_count = OST_MIN_PRECREATE;
- cfs_waitq_signal(&d->opd_pre_waitq);
+ spin_unlock(&d->opd_pre_lock);
+ wake_up(&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,
}
out:
- cfs_waitq_signal(&d->opd_pre_user_waitq);
+ wake_up(&d->opd_pre_user_waitq);
+}
+
+int osp_init_pre_fid(struct osp_device *osp)
+{
+ struct lu_env env;
+ struct osp_thread_info *osi;
+ struct lu_client_seq *cli_seq;
+ struct lu_fid *last_fid;
+ int rc;
+ ENTRY;
+
+ LASSERT(osp->opd_pre != NULL);
+
+ /* Return if last_used fid has been initialized */
+ if (!fid_is_zero(&osp->opd_last_used_fid))
+ RETURN(0);
+
+ rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
+ if (rc) {
+ CERROR("%s: init env error: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ RETURN(rc);
+ }
+
+ osi = osp_env_info(&env);
+ last_fid = &osi->osi_fid;
+ fid_zero(last_fid);
+ /* For a freshed fs, it will allocate a new sequence first */
+ if (osp_is_fid_client(osp) && osp->opd_group != 0) {
+ cli_seq = osp->opd_obd->u.cli.cl_seq;
+ rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
+ if (rc != 0) {
+ CERROR("%s: alloc fid error: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ GOTO(out, rc);
+ }
+ } else {
+ last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
+ }
+ last_fid->f_oid = 1;
+ last_fid->f_ver = 0;
+
+ spin_lock(&osp->opd_pre_lock);
+ osp->opd_last_used_fid = *last_fid;
+ osp->opd_pre_used_fid = *last_fid;
+ osp->opd_pre_last_created_fid = *last_fid;
+ spin_unlock(&osp->opd_pre_lock);
+ rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
+ if (rc != 0) {
+ CERROR("%s: write fid error: rc = %d\n",
+ osp->opd_obd->obd_name, rc);
+ GOTO(out, rc);
+ }
+out:
+ lu_env_fini(&env);
+ 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 };
- char pname[16];
+ struct lu_env env;
int rc;
ENTRY;
- sprintf(pname, "osp-pre-%u\n", d->opd_index);
- cfs_daemonize(pname);
+ 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);
+ }
- cfs_spin_lock(&d->opd_pre_lock);
+ spin_lock(&d->opd_pre_lock);
thread->t_flags = SVC_RUNNING;
- cfs_spin_unlock(&d->opd_pre_lock);
- cfs_waitq_signal(&thread->t_ctl_waitq);
+ spin_unlock(&d->opd_pre_lock);
+ wake_up(&thread->t_ctl_waitq);
while (osp_precreate_running(d)) {
/*
while (osp_precreate_running(d)) {
l_wait_event(d->opd_pre_waitq,
!osp_precreate_running(d) ||
- d->opd_new_connection, &lwi);
-
- if (!osp_precreate_running(d))
- break;
+ d->opd_new_connection,
+ &lwi);
if (!d->opd_new_connection)
continue;
- /* got connected */
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);
+ /* Sigh, fid client is not ready yet */
+ if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
+ continue;
+
+ /* Init fid for osp_precreate if necessary */
+ 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);
/*
- * 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. so we block new reservations and wait till all reserved
- * objects either user or released.
+ * Clean up orphans or recreate missing objects.
*/
- 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) {
- rc = osp_precreate_cleanup_orphans(d);
- if (rc) {
- 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);
- cfs_waitq_signal(&d->opd_pre_user_waitq);
- l_wait_event(d->opd_pre_waitq,
- !osp_precreate_running(d) ||
- d->opd_new_connection, &lwi);
- continue;
-
- }
- }
-
+ 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(d) ||
+ osp_precreate_near_empty(&env, d) ||
osp_statfs_need_update(d) ||
d->opd_got_disconnected, &lwi);
if (osp_statfs_need_update(d))
osp_statfs_update(d);
- if (osp_precreate_near_empty(d)) {
- rc = osp_precreate_send(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 &&
+ if (rc < 0 && rc != -ENOSPC &&
rc != -ETIMEDOUT && rc != -ENOTCONN)
CERROR("%s: cannot precreate objects:"
" rc = %d\n",
}
thread->t_flags = SVC_STOPPED;
- cfs_waitq_signal(&thread->t_ctl_waitq);
+ lu_env_fini(&env);
+ wake_up(&thread->t_ctl_waitq);
RETURN(0);
}
-static int osp_precreate_ready_condition(struct osp_device *d)
+static int osp_precreate_ready_condition(const struct lu_env *env,
+ struct osp_device *d)
{
- __u64 next;
+ if (d->opd_pre_recovering)
+ return 0;
/* ready if got enough precreated objects */
/* we need to wait for others (opd_pre_reserved) and our object (+1) */
- next = d->opd_pre_used_id + d->opd_pre_reserved + 1;
- if (next <= d->opd_pre_last_created)
+ if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
return 1;
- /* ready if OST reported no space and no destoys in progress */
+ /* ready if OST reported no space and no destroys in progress */
if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
- d->opd_pre_status != 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 != -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;
}
{
struct osp_device *d = data;
- LCONSOLE_WARN("%s: slow creates, last="LPU64", next="LPU64", "
- "reserved="LPU64", syn_changes=%lu, "
- "syn_rpc_in_progress=%d, status=%d\n",
- d->opd_obd->obd_name, d->opd_pre_last_created,
- d->opd_pre_used_id, d->opd_pre_reserved,
- d->opd_syn_changes, d->opd_syn_rpc_in_progress,
- d->opd_pre_status);
+ CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
+ "reserved="LPU64", syn_changes=%lu, "
+ "syn_rpc_in_progress=%d, status=%d\n",
+ d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
+ PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
+ d->opd_syn_changes, d->opd_syn_rpc_in_progress,
+ d->opd_pre_status);
- return 0;
+ return 1;
}
/*
struct l_wait_info lwi;
cfs_time_t expire = cfs_time_shift(obd_timeout);
int precreated, rc;
- int count = 0;
ENTRY;
- LASSERT(d->opd_pre_last_created >= d->opd_pre_used_id);
-
- lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
- osp_precreate_timeout_condition, d);
+ 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) {
- if (unlikely(rc == -ENODEV)) {
- if (cfs_time_aftereq(cfs_time_current(), expire))
- break;
- }
-
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
- /*
- * to address Andreas's concern on possible busy-loop
- * between this thread and osp_precreate_send()
- */
- if (unlikely(count++ == 1000)) {
- osp_precreate_timeout_condition(d);
- LBUG();
- }
-#endif
+ rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
/*
* 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 &&
- (d->opd_pre_last_created - d->opd_pre_used_id <=
- d->opd_pre_grow_count / 4 + 1)) {
- cfs_spin_lock(&d->opd_pre_lock);
+ 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;
- cfs_spin_unlock(&d->opd_pre_lock);
+ spin_unlock(&d->opd_pre_lock);
}
- /*
- * we never use the last object in the window
- */
- cfs_spin_lock(&d->opd_pre_lock);
- precreated = d->opd_pre_last_created - d->opd_pre_used_id;
- if (precreated > d->opd_pre_reserved) {
+ 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++;
- cfs_spin_unlock(&d->opd_pre_lock);
+ spin_unlock(&d->opd_pre_lock);
rc = 0;
/* XXX: don't wake up if precreation is in progress */
- if (osp_precreate_near_empty_nolock(d))
- cfs_waitq_signal(&d->opd_pre_waitq);
+ if (osp_precreate_near_empty_nolock(env, d) &&
+ !osp_precreate_end_seq_nolock(env, d))
+ wake_up(&d->opd_pre_waitq);
break;
}
- cfs_spin_unlock(&d->opd_pre_lock);
+ spin_unlock(&d->opd_pre_lock);
/*
* all precreated objects have been used and no-space
}
/* XXX: don't wake up if precreation is in progress */
- cfs_waitq_signal(&d->opd_pre_waitq);
+ wake_up(&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(d), &lwi);
+ osp_precreate_ready_condition(env, d), &lwi);
}
RETURN(rc);
/*
* this function relies on reservation made before
*/
-__u64 osp_precreate_get_id(struct osp_device *d)
+int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
+ struct lu_fid *fid)
{
- obd_id objid;
-
/* grab next id from the pool */
- cfs_spin_lock(&d->opd_pre_lock);
- LASSERT(d->opd_pre_used_id < d->opd_pre_last_created);
- objid = ++d->opd_pre_used_id;
+ spin_lock(&d->opd_pre_lock);
+
+ LASSERTF(osp_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_id(d, objid);
- cfs_spin_unlock(&d->opd_pre_lock);
+ osp_update_last_fid(d, fid);
+ spin_unlock(&d->opd_pre_lock);
/*
* probably main thread suspended orphan cleanup till
* 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);
+ wake_up(&d->opd_pre_waitq);
- return objid;
+ return 0;
}
/*
if (oa == NULL)
GOTO(out, rc = -ENOMEM);
- rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
+ rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
LASSERT(rc == 0);
oa->o_size = size;
oa->o_blocks = OBD_OBJECT_EOF;
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
LASSERT(body);
- lustre_set_wire_obdo(&body->oa, oa);
+ lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
/* XXX: capa support? */
/* osc_pack_capa(req, body, capa); */
int osp_init_precreate(struct osp_device *d)
{
struct l_wait_info lwi = { 0 };
- int rc;
+ 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;
- d->opd_pre_used_id = 0;
- d->opd_pre_last_created = 0;
+ 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_min_grow_count = OST_MIN_PRECREATE;
d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
- cfs_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);
+ 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
/*
* 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);
+ 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,
void osp_precreate_fini(struct osp_device *d)
{
- struct ptlrpc_thread *thread = &d->opd_pre_thread;
+ struct ptlrpc_thread *thread;
ENTRY;
cfs_timer_disarm(&d->opd_statfs_timer);
+ if (d->opd_pre == NULL)
+ RETURN_EXIT;
+
+ thread = &d->opd_pre_thread;
+
thread->t_flags = SVC_STOPPING;
- cfs_waitq_signal(&d->opd_pre_waitq);
+ wake_up(&d->opd_pre_waitq);
+
+ wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
- cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
+ OBD_FREE_PTR(d->opd_pre);
+ d->opd_pre = NULL;
EXIT;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff