/**
* 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->lb_len, oid_off, th);
if (rc != 0)
GOTO(out, rc);
rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
lb_oseq->lb_len, 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_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 void osp_object_assign_fid(const struct lu_env *env,
struct osp_device *d, struct osp_object *o)
{
struct osp_thread_info *osi = osp_env_info(env);
LASSERT(fid_is_zero(lu_object_fid(&o->opo_obj.do_lu)));
LASSERT(o->opo_reserved);
o->opo_reserved = 0;
osp_precreate_get_fid(env, d, &osi->osi_fid);
lu_object_assign_fid(env, &o->opo_obj.do_lu, &osi->osi_fid);
}
static inline int osp_objs_precreated(const struct lu_env *env,
struct osp_device *osp)
{
struct lu_fid *fid1 = &osp->opd_pre_last_created_fid;
struct lu_fid *fid2 = &osp->opd_pre_used_fid;
LASSERTF(fid_seq(fid1) == fid_seq(fid2),
"Created fid"DFID" Next fid "DFID"\n", PFID(fid1), PFID(fid2));
if (fid_is_idif(fid1)) {
struct ost_id *oi1 = &osp_env_info(env)->osi_oi;
struct ost_id *oi2 = &osp_env_info(env)->osi_oi2;
LASSERT(fid_is_idif(fid1) && fid_is_idif(fid2));
fid_to_ostid(fid1, oi1);
fid_to_ostid(fid2, oi2);
LASSERT(ostid_id(oi1) >= ostid_id(oi2));
return ostid_id(oi1) - ostid_id(oi2);
}
return fid_oid(fid1) - fid_oid(fid2);
}
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_object_init(const struct lu_env *env, struct lu_object *o,
const struct lu_object_conf *conf)
{
struct osp_object *po = lu2osp_obj(o);
int rc = 0;
ENTRY;
if (is_ost_obj(o)) {
po->opo_obj.do_ops = &osp_obj_ops;
} else {
struct lu_attr *la = &osp_env_info(env)->osi_attr;
po->opo_obj.do_ops = &osp_md_obj_ops;
o->lo_header->loh_attr |= LOHA_REMOTE;
rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o),
la, NULL);
if (rc == 0)
o->lo_header->loh_attr |=
LOHA_EXISTS | (la->la_mode & S_IFMT);
if (rc == -ENOENT)
rc = 0;
}
RETURN(rc);
}
/**
* 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);
}
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;
}
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);
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);
wake_up(&d->opd_pre_user_waitq);
ptlrpc_req_finished(req);
RETURN(rc);
}
/**
* Implementation of dt_index_operations::dio_lookup
*
* Look up record by key under a remote index object. It packs lookup update
* into RPC, sends to the remote OUT and waits for the lookup result.
*
* \param[in] env execution environment
* \param[in] dt index object to lookup
* \param[out] rec record in which to return lookup result
* \param[in] key key of index which will be looked up
*
* \retval 1 if the lookup succeeds.
* \retval negative errno if the lookup fails.
*/
static int osp_md_index_lookup(const struct lu_env *env, struct dt_object *dt,
struct dt_rec *rec, const struct dt_key *key)
{
struct lu_buf *lbuf = &osp_env_info(env)->osi_lb2;
struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
struct dt_device *dt_dev = &osp->opd_dt_dev;
struct osp_update_request *update;
struct object_update_reply *reply;
struct ptlrpc_request *req = NULL;
struct lu_fid *fid;
int rc;
ENTRY;
/* Because it needs send the update buffer right away,
* just create an update buffer, instead of attaching the
* update_remote list of the thandle.
*/
update = osp_update_request_create(dt_dev);
if (IS_ERR(update))
RETURN(PTR_ERR(update));
rc = osp_update_rpc_pack(env, index_lookup, update, OUT_INDEX_LOOKUP,
lu_object_fid(&dt->do_lu), rec, key);
if (rc != 0) {
CERROR("%s: Insert update error: rc = %d\n",
dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
GOTO(out, rc);
}
rc = osp_remote_sync(env, osp, update, &req);
if (rc < 0)
GOTO(out, rc);
reply = req_capsule_server_sized_get(&req->rq_pill,
&RMF_OUT_UPDATE_REPLY,
OUT_UPDATE_REPLY_SIZE);
if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
CERROR("%s: Wrong version %x expected %x: rc = %d\n",
dt_dev->dd_lu_dev.ld_obd->obd_name,
reply->ourp_magic, UPDATE_REPLY_MAGIC, -EPROTO);
GOTO(out, rc = -EPROTO);
}
rc = object_update_result_data_get(reply, lbuf, 0);
if (rc < 0)
GOTO(out, rc);
if (lbuf->lb_len != sizeof(*fid)) {
CERROR("%s: lookup "DFID" %s wrong size %d\n",
dt_dev->dd_lu_dev.ld_obd->obd_name,
PFID(lu_object_fid(&dt->do_lu)), (char *)key,
(int)lbuf->lb_len);
GOTO(out, rc = -EINVAL);
}
fid = lbuf->lb_buf;
if (ptlrpc_rep_need_swab(req))
lustre_swab_lu_fid(fid);
if (!fid_is_sane(fid)) {
CERROR("%s: lookup "DFID" %s invalid fid "DFID"\n",
dt_dev->dd_lu_dev.ld_obd->obd_name,
PFID(lu_object_fid(&dt->do_lu)), (char *)key, PFID(fid));
GOTO(out, rc = -EINVAL);
}
memcpy(rec, fid, sizeof(*fid));
GOTO(out, rc = 1);
out:
if (req != NULL)
ptlrpc_req_finished(req);
osp_update_request_destroy(update);
return rc;
}
static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
struct lu_buf *rbuf, loff_t *pos)
{
struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
struct dt_device *dt_dev = &osp->opd_dt_dev;
struct lu_buf *lbuf = &osp_env_info(env)->osi_lb2;
char *ptr = rbuf->lb_buf;
struct osp_update_request *update;
struct ptlrpc_request *req = NULL;
struct out_read_reply *orr;
struct ptlrpc_bulk_desc *desc;
struct object_update_reply *reply;
__u32 left_size;
int nbufs;
int i;
int rc;
ENTRY;
/* Because it needs send the update buffer right away,
* just create an update buffer, instead of attaching the
* update_remote list of the thandle. */
update = osp_update_request_create(dt_dev);
if (IS_ERR(update))
RETURN(PTR_ERR(update));
rc = osp_update_rpc_pack(env, read, update, OUT_READ,
lu_object_fid(&dt->do_lu),
rbuf->lb_len, *pos);
if (rc != 0) {
CERROR("%s: cannot insert update: rc = %d\n",
dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
GOTO(out_update, rc);
}
CDEBUG(D_INFO, "%s "DFID" read offset %llu size %zu\n",
dt_dev->dd_lu_dev.ld_obd->obd_name,
PFID(lu_object_fid(&dt->do_lu)), *pos, rbuf->lb_len);
rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import, update,
&req);
if (rc != 0)
GOTO(out_update, rc);
nbufs = (rbuf->lb_len + OUT_BULK_BUFFER_SIZE - 1) /
OUT_BULK_BUFFER_SIZE;
/* allocate bulk descriptor */
desc = ptlrpc_prep_bulk_imp(req, nbufs, 1,
PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KVEC,
MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
if (desc == NULL)
GOTO(out, rc = -ENOMEM);
/* split the buffer into small chunk size */
left_size = rbuf->lb_len;
for (i = 0; i < nbufs; i++) {
int read_size;
read_size = left_size > OUT_BULK_BUFFER_SIZE ?
OUT_BULK_BUFFER_SIZE : left_size;
desc->bd_frag_ops->add_iov_frag(desc, ptr, read_size);
ptr += read_size;
}
/* This will only be called with read-only update, and these updates
* might be used to retrieve update log during recovery process, so
* it will be allowed to send during recovery process */
req->rq_allow_replay = 1;
req->rq_bulk_read = 1;
/* send request to master and wait for RPC to complete */
rc = ptlrpc_queue_wait(req);
if (rc != 0)
GOTO(out, rc);
rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
req->rq_bulk->bd_nob_transferred);
if (rc < 0)
GOTO(out, rc);
reply = req_capsule_server_sized_get(&req->rq_pill,
&RMF_OUT_UPDATE_REPLY,
OUT_UPDATE_REPLY_SIZE);
if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
CERROR("%s: invalid update reply magic %x expected %x:"
" rc = %d\n", dt_dev->dd_lu_dev.ld_obd->obd_name,
reply->ourp_magic, UPDATE_REPLY_MAGIC, -EPROTO);
GOTO(out, rc = -EPROTO);
}
rc = object_update_result_data_get(reply, lbuf, 0);
if (rc < 0)
GOTO(out, rc);
if (lbuf->lb_len < sizeof(*orr))
GOTO(out, rc = -EPROTO);
orr = lbuf->lb_buf;
orr_le_to_cpu(orr, orr);
rc = orr->orr_size;
*pos = orr->orr_offset;
out:
ptlrpc_req_finished(req);
out_update:
osp_update_request_destroy(update);
RETURN(rc);
}
static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
struct lu_attr *attr,
struct dt_allocation_hint *hint,
struct dt_object_format *dof, struct thandle *th)
{
struct osp_thread_info *osi = osp_env_info(env);
struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
struct osp_object *o = dt2osp_obj(dt);
int rc = 0;
struct lu_fid *fid = &osi->osi_fid;
ENTRY;
if (o->opo_reserved) {
/* regular case, fid is assigned holding trunsaction open */
osp_object_assign_fid(env, d, o);
}
memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
LASSERTF(fid_is_sane(fid), "fid for osp_obj %p is insane"DFID"!\n",
osp_obj, PFID(fid));
if (!o->opo_reserved) {
/* special case, id was assigned outside of transaction
* see comments in osp_declare_attr_set */
spin_lock(&d->opd_pre_lock);
osp_update_last_fid(d, fid);
spin_unlock(&d->opd_pre_lock);
}
CDEBUG(D_INODE, "fid for osp_obj %p is "DFID"!\n", osp_obj, PFID(fid));
/* If the precreate ends, it means it will be ready to rollover to
* the new sequence soon, all the creation should be synchronized,
* otherwise during replay, the replay fid will be inconsistent with
* last_used/create fid */
if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
th->th_sync = 1;
/*
* it's OK if the import is inactive by this moment - id was created
* by OST earlier, we just need to maintain it consistently on the disk
* once import is reconnected, OSP will claim this and other objects
* used and OST either keep them, if they exist or recreate
*/
/* we might have lost precreated objects */
if (unlikely(d->opd_gap_count) > 0) {
spin_lock(&d->opd_pre_lock);
if (d->opd_gap_count > 0) {
int count = d->opd_gap_count;
ostid_set_id(&osi->osi_oi,
fid_oid(&d->opd_gap_start_fid));
d->opd_gap_count = 0;
spin_unlock(&d->opd_pre_lock);
CDEBUG(D_HA, "Writting gap "DFID"+%d in llog\n",
PFID(&d->opd_gap_start_fid), count);
/* real gap handling is disabled intil ORI-692 will be
* fixed, now we only report gaps */
} else {
spin_unlock(&d->opd_pre_lock);
}
}
/* new object, the very first ->attr_set()
* initializing attributes needs no logging */
o->opo_new = 1;
/* Only need update last_used oid file, seq file will only be update
* during seq rollover */
osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
&d->opd_last_used_fid.f_oid, d->opd_index);
rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
&osi->osi_off, th);
CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
RETURN(rc);
}
static int osp_declare_object_create(const struct lu_env *env,
struct dt_object *dt,
struct lu_attr *attr,
struct dt_allocation_hint *hint,
struct dt_object_format *dof,
struct thandle *th)
{
struct osp_thread_info *osi = osp_env_info(env);
struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
struct osp_object *o = dt2osp_obj(dt);
const struct lu_fid *fid;
int rc = 0;
ENTRY;
/* should happen to non-0 OSP only so that at least one object
* has been already declared in the scenario and LOD should
* cleanup that */
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
RETURN(-ENOSPC);
LASSERT(d->opd_last_used_oid_file);
fid = lu_object_fid(&dt->do_lu);
/*
* There can be gaps in precreated ids and record to unlink llog
* XXX: we do not handle gaps yet, implemented before solution
* was found to be racy, so we disabled that. there is no
* point in making useless but expensive llog declaration.
*/
/* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
if (unlikely(!fid_is_zero(fid))) {
/* replay case: caller knows fid */
osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
sizeof(osi->osi_id), osi->osi_off,
th);
RETURN(rc);
}
/*
* in declaration we need to reserve object so that we don't block
* awaiting precreation RPC to complete
*/
rc = osp_precreate_reserve(env, d);
/*
* we also need to declare update to local "last used id" file for
* recovery if object isn't used for a reason, we need to release
* reservation, this can be made in osd_object_release()
*/
if (rc == 0) {
/* mark id is reserved: in create we don't want to talk
* to OST */
LASSERT(o->opo_reserved == 0);
o->opo_reserved = 1;
/* common for all OSPs file hystorically */
osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
sizeof(osi->osi_id), osi->osi_off,
th);
} else {
/* not needed in the cache anymore */
set_bit(LU_OBJECT_HEARD_BANSHEE,
&dt->do_lu.lo_header->loh_flags);
}
RETURN(rc);
}
