static int osd_object_destroy(const struct lu_env *env,
struct dt_object *dt, struct thandle *th)
{
char *buf = osd_oti_get(env)->oti_str;
struct osd_object *obj = osd_dt_obj(dt);
struct osd_device *osd = osd_obj2dev(obj);
const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
struct osd_thandle *oh;
int rc;
uint64_t oid, zapid;
ENTRY;
LASSERT(obj->oo_db != NULL);
LASSERT(dt_object_exists(dt));
LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
oh = container_of0(th, struct osd_thandle, ot_super);
LASSERT(oh != NULL);
LASSERT(oh->ot_tx != NULL);
/* remove obj ref from index dir (it depends) */
zapid = osd_get_name_n_idx(env, osd, fid, buf);
rc = -zap_remove(osd->od_os, zapid, buf, oh->ot_tx);
if (rc) {
CERROR("%s: zap_remove(%s) failed: rc = %d\n",
osd->od_svname, buf, rc);
GOTO(out, rc);
}
rc = osd_xattrs_destroy(env, obj, oh);
if (rc) {
CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
osd->od_svname, buf, rc);
GOTO(out, rc);
}
/* Remove object from inode accounting. It is not fatal for the destroy
* operation if something goes wrong while updating accounting, but we
* still log an error message to notify the administrator */
rc = -zap_increment_int(osd->od_os, osd->od_iusr_oid,
obj->oo_attr.la_uid, -1, oh->ot_tx);
if (rc)
CERROR("%s: failed to remove "DFID" from accounting ZAP for usr"
" %d: rc = %d\n", osd->od_svname, PFID(fid),
obj->oo_attr.la_uid, rc);
rc = -zap_increment_int(osd->od_os, osd->od_igrp_oid,
obj->oo_attr.la_gid, -1, oh->ot_tx);
if (rc)
CERROR("%s: failed to remove "DFID" from accounting ZAP for grp"
" %d: rc = %d\n", osd->od_svname, PFID(fid),
obj->oo_attr.la_gid, rc);
oid = obj->oo_db->db_object;
if (obj->oo_destroy == OSD_DESTROY_SYNC) {
rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
if (rc)
CERROR("%s: failed to free %s "LPU64": rc = %d\n",
osd->od_svname, buf, oid, rc);
} else { /* asynchronous destroy */
rc = osd_object_unlinked_add(obj, oh);
if (rc)
GOTO(out, rc);
rc = -zap_add_int(osd->od_os, osd->od_unlinkedid,
oid, oh->ot_tx);
if (rc)
CERROR("%s: zap_add_int() failed %s "LPU64": rc = %d\n",
osd->od_svname, buf, oid, rc);
}
out:
/* not needed in the cache anymore */
set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
if (rc == 0)
obj->oo_destroyed = 1;
RETURN (0);
}
static int osd_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 *handle)
{
char *buf = osd_oti_get(env)->oti_str;
const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
struct osd_object *obj = osd_dt_obj(dt);
struct osd_device *osd = osd_obj2dev(obj);
struct osd_thandle *oh;
uint64_t zapid;
int rc;
ENTRY;
LASSERT(dof);
switch (dof->dof_type) {
case DFT_REGULAR:
case DFT_SYM:
case DFT_NODE:
if (obj->oo_dt.do_body_ops == NULL)
obj->oo_dt.do_body_ops = &osd_body_ops;
break;
default:
break;
}
LASSERT(handle != NULL);
oh = container_of0(handle, struct osd_thandle, ot_super);
LASSERT(oh->ot_tx != NULL);
switch (dof->dof_type) {
case DFT_DIR:
dt->do_index_ops = &osd_dir_ops;
case DFT_INDEX:
/* for zap create */
dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, 1, NULL);
break;
case DFT_REGULAR:
case DFT_SYM:
case DFT_NODE:
/* first, we'll create new object */
dmu_tx_hold_bonus(oh->ot_tx, DMU_NEW_OBJECT);
break;
default:
LBUG();
break;
}
/* and we'll add it to some mapping */
zapid = osd_get_name_n_idx(env, osd, fid, buf);
dmu_tx_hold_bonus(oh->ot_tx, zapid);
dmu_tx_hold_zap(oh->ot_tx, zapid, TRUE, buf);
/* we will also update inode accounting ZAPs */
dmu_tx_hold_bonus(oh->ot_tx, osd->od_iusr_oid);
dmu_tx_hold_zap(oh->ot_tx, osd->od_iusr_oid, TRUE, buf);
dmu_tx_hold_bonus(oh->ot_tx, osd->od_igrp_oid);
dmu_tx_hold_zap(oh->ot_tx, osd->od_igrp_oid, TRUE, buf);
dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
__osd_xattr_declare_set(env, obj, sizeof(struct lustre_mdt_attrs),
XATTR_NAME_LMA, oh);
rc = osd_declare_quota(env, osd, attr->la_uid, attr->la_gid, 1, oh,
false, NULL, false);
RETURN(rc);
}
/*
* Concurrency: @dt is write locked.
*/
static int osd_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 zpl_direntry *zde = &osd_oti_get(env)->oti_zde.lzd_reg;
const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
struct osd_object *obj = osd_dt_obj(dt);
struct osd_device *osd = osd_obj2dev(obj);
char *buf = osd_oti_get(env)->oti_str;
struct osd_thandle *oh;
dmu_buf_t *db;
uint64_t zapid;
int rc;
ENTRY;
/* concurrent create declarations should not see
* the object inconsistent (db, attr, etc).
* in regular cases acquisition should be cheap */
down(&obj->oo_guard);
LASSERT(osd_invariant(obj));
LASSERT(!dt_object_exists(dt));
LASSERT(dof != NULL);
LASSERT(th != NULL);
oh = container_of0(th, struct osd_thandle, ot_super);
/*
* XXX missing: Quote handling.
*/
LASSERT(obj->oo_db == NULL);
/* to follow ZFS on-disk format we need
* to initialize parent dnode properly */
zapid = 0;
if (hint != NULL && hint->dah_parent != NULL &&
!dt_object_remote(hint->dah_parent))
zapid = osd_dt_obj(hint->dah_parent)->oo_db->db_object;
db = osd_create_type_f(dof->dof_type)(env, obj, attr, zapid, oh);
if (IS_ERR(db))
GOTO(out, rc = PTR_ERR(db));
zde->zde_pad = 0;
zde->zde_dnode = db->db_object;
zde->zde_type = IFTODT(attr->la_mode & S_IFMT);
zapid = osd_get_name_n_idx(env, osd, fid, buf);
rc = -zap_add(osd->od_os, zapid, buf, 8, 1, zde, oh->ot_tx);
if (rc)
GOTO(out, rc);
/* Add new object to inode accounting.
* Errors are not considered as fatal */
rc = -zap_increment_int(osd->od_os, osd->od_iusr_oid,
(attr->la_valid & LA_UID) ? attr->la_uid : 0, 1,
oh->ot_tx);
if (rc)
CERROR("%s: failed to add "DFID" to accounting ZAP for usr %d "
"(%d)\n", osd->od_svname, PFID(fid), attr->la_uid, rc);
rc = -zap_increment_int(osd->od_os, osd->od_igrp_oid,
(attr->la_valid & LA_GID) ? attr->la_gid : 0, 1,
oh->ot_tx);
if (rc)
CERROR("%s: failed to add "DFID" to accounting ZAP for grp %d "
"(%d)\n", osd->od_svname, PFID(fid), attr->la_gid, rc);
/* configure new osd object */
obj->oo_db = db;
rc = osd_object_init0(env, obj);
LASSERT(ergo(rc == 0, dt_object_exists(dt)));
LASSERT(osd_invariant(obj));
rc = osd_init_lma(env, obj, fid, oh);
if (rc) {
CERROR("%s: can not set LMA on "DFID": rc = %d\n",
osd->od_svname, PFID(fid), rc);
/* ignore errors during LMA initialization */
rc = 0;
}
out:
up(&obj->oo_guard);
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);
}
int cl_glimpse_lock(const struct lu_env *env, struct cl_io *io,
struct inode *inode, struct cl_object *clob, int agl)
{
struct cl_lock_descr *descr = &ccc_env_info(env)->cti_descr;
struct cl_inode_info *lli = cl_i2info(inode);
const struct lu_fid *fid = lu_object_fid(&clob->co_lu);
struct ccc_io *cio = ccc_env_io(env);
struct cl_lock *lock;
int result;
result = 0;
if (!(lli->lli_flags & LLIF_MDS_SIZE_LOCK)) {
CDEBUG(D_DLMTRACE, "Glimpsing inode "DFID"\n", PFID(fid));
if (lli->lli_has_smd) {
/* NOTE: this looks like DLM lock request, but it may
* not be one. Due to CEF_ASYNC flag (translated
* to LDLM_FL_HAS_INTENT by osc), this is
* glimpse request, that won't revoke any
* conflicting DLM locks held. Instead,
* ll_glimpse_callback() will be called on each
* client holding a DLM lock against this file,
* and resulting size will be returned for each
* stripe. DLM lock on [0, EOF] is acquired only
* if there were no conflicting locks. If there
* were conflicting locks, enqueuing or waiting
* fails with -ENAVAIL, but valid inode
* attributes are returned anyway. */
*descr = whole_file;
descr->cld_obj = clob;
descr->cld_mode = CLM_PHANTOM;
descr->cld_enq_flags = CEF_ASYNC | CEF_MUST;
if (agl)
descr->cld_enq_flags |= CEF_AGL;
cio->cui_glimpse = 1;
/*
* CEF_ASYNC is used because glimpse sub-locks cannot
* deadlock (because they never conflict with other
* locks) and, hence, can be enqueued out-of-order.
*
* CEF_MUST protects glimpse lock from conversion into
* a lockless mode.
*/
lock = cl_lock_request(env, io, descr, "glimpse",
current);
cio->cui_glimpse = 0;
if (lock == NULL)
return 0;
if (IS_ERR(lock))
return PTR_ERR(lock);
LASSERT(agl == 0);
result = cl_wait(env, lock);
if (result == 0) {
cl_merge_lvb(env, inode);
if (cl_isize_read(inode) > 0 &&
inode->i_blocks == 0) {
/*
* LU-417: Add dirty pages block count
* lest i_blocks reports 0, some "cp" or
* "tar" may think it's a completely
* sparse file and skip it.
*/
inode->i_blocks = dirty_cnt(inode);
}
cl_unuse(env, lock);
}
cl_lock_release(env, lock, "glimpse", current);
} else {
CDEBUG(D_DLMTRACE, "No objects for inode\n");
cl_merge_lvb(env, inode);
}
}
return result;
}
static int llog_osd_prev_block(const struct lu_env *env,
struct llog_handle *loghandle,
int prev_idx, void *buf, int len)
{
struct llog_thread_info *lgi = llog_info(env);
struct dt_object *o;
struct dt_device *dt;
loff_t cur_offset;
int rc;
if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
return -EINVAL;
CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
LASSERT(loghandle);
LASSERT(loghandle->lgh_ctxt);
o = loghandle->lgh_obj;
LASSERT(o);
LASSERT(dt_object_exists(o));
dt = lu2dt_dev(o->do_lu.lo_dev);
LASSERT(dt);
cur_offset = LLOG_CHUNK_SIZE;
llog_skip_over(&cur_offset, 0, prev_idx);
rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
if (rc)
GOTO(out, rc);
while (cur_offset < lgi->lgi_attr.la_size) {
struct llog_rec_hdr *rec, *last_rec;
struct llog_rec_tail *tail;
lgi->lgi_buf.lb_len = len;
lgi->lgi_buf.lb_buf = buf;
/* It is OK to have locking around dt_read() only, see
* comment in llog_osd_next_block for details
*/
dt_read_lock(env, o, 0);
rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
dt_read_unlock(env, o);
if (rc < 0) {
CERROR("%s: can't read llog block from log "DFID
" offset "LPU64": rc = %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
GOTO(out, rc);
}
if (rc == 0) /* end of file, nothing to do */
GOTO(out, rc);
if (rc < sizeof(*tail)) {
CERROR("%s: invalid llog block at log id "DOSTID"/%u "
"offset "LPU64"\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, cur_offset);
GOTO(out, rc = -EINVAL);
}
rec = buf;
if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
lustre_swab_llog_rec(rec);
tail = (struct llog_rec_tail *)((char *)buf + rc -
sizeof(struct llog_rec_tail));
/* get the last record in block */
last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
le32_to_cpu(tail->lrt_len));
if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
lustre_swab_llog_rec(last_rec);
LASSERT(last_rec->lrh_index == tail->lrt_index);
/* this shouldn't happen */
if (tail->lrt_index == 0) {
CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
"offset "LPU64"\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, cur_offset);
GOTO(out, rc = -EINVAL);
}
if (tail->lrt_index < prev_idx)
continue;
/* sanity check that the start of the new buffer is no farther
* than the record that we wanted. This shouldn't happen. */
if (rec->lrh_index > prev_idx) {
CERROR("%s: missed desired record? %u > %u\n",
o->do_lu.lo_dev->ld_obd->obd_name,
rec->lrh_index, prev_idx);
GOTO(out, rc = -ENOENT);
}
GOTO(out, rc = 0);
}
GOTO(out, rc = -EIO);
out:
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);
}
/* appends if idx == -1, otherwise overwrites record idx. */
static int llog_osd_write_rec(const struct lu_env *env,
struct llog_handle *loghandle,
struct llog_rec_hdr *rec,
struct llog_cookie *reccookie, int cookiecount,
void *buf, int idx, struct thandle *th)
{
struct llog_thread_info *lgi = llog_info(env);
struct llog_log_hdr *llh;
int reclen = rec->lrh_len;
int index, rc, old_tail_idx;
struct llog_rec_tail *lrt;
struct dt_object *o;
size_t left;
LASSERT(env);
llh = loghandle->lgh_hdr;
LASSERT(llh);
o = loghandle->lgh_obj;
LASSERT(o);
LASSERT(th);
CDEBUG(D_OTHER, "new record %x to "DFID"\n",
rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
/* record length should not bigger than LLOG_CHUNK_SIZE */
if (buf)
rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
else
rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
if (rc)
return rc;
rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
if (rc)
return rc;
if (buf)
/* write_blob adds header and tail to lrh_len. */
reclen = sizeof(*rec) + rec->lrh_len +
sizeof(struct llog_rec_tail);
if (idx != -1) {
/* no header: only allowed to insert record 1 */
if (idx != 1 && lgi->lgi_attr.la_size == 0)
LBUG();
if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
return -EINVAL;
if (!ext2_test_bit(idx, llh->llh_bitmap))
CERROR("%s: modify unset record %u\n",
o->do_lu.lo_dev->ld_obd->obd_name, idx);
if (idx != rec->lrh_index)
CERROR("%s: index mismatch %d %u\n",
o->do_lu.lo_dev->ld_obd->obd_name, idx,
rec->lrh_index);
lgi->lgi_off = 0;
rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
&lgi->lgi_off, th);
/* we are done if we only write the header or on error */
if (rc || idx == 0)
return rc;
if (buf) {
/* We assume that caller has set lgh_cur_* */
lgi->lgi_off = loghandle->lgh_cur_offset;
CDEBUG(D_OTHER,
"modify record "DOSTID": idx:%d/%u/%d, len:%u "
"offset %llu\n",
POSTID(&loghandle->lgh_id.lgl_oi), idx,
rec->lrh_index,
loghandle->lgh_cur_idx, rec->lrh_len,
(long long)(lgi->lgi_off - sizeof(*llh)));
if (rec->lrh_index != loghandle->lgh_cur_idx) {
CERROR("%s: modify idx mismatch %u/%d\n",
o->do_lu.lo_dev->ld_obd->obd_name, idx,
loghandle->lgh_cur_idx);
return -EFAULT;
}
} else {
/* Assumes constant lrh_len */
lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
}
rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
if (rc == 0 && reccookie) {
reccookie->lgc_lgl = loghandle->lgh_id;
reccookie->lgc_index = idx;
rc = 1;
}
return rc;
}
/* Make sure that records don't cross a chunk boundary, so we can
* process them page-at-a-time if needed. If it will cross a chunk
* boundary, write in a fake (but referenced) entry to pad the chunk.
*
* We know that llog_current_log() will return a loghandle that is
* big enough to hold reclen, so all we care about is padding here.
*/
LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
lgi->lgi_off = lgi->lgi_attr.la_size;
left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
/* NOTE: padding is a record, but no bit is set */
if (left != 0 && left != reclen &&
left < (reclen + LLOG_MIN_REC_SIZE)) {
index = loghandle->lgh_last_idx + 1;
rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
if (rc)
return rc;
loghandle->lgh_last_idx++; /*for pad rec*/
}
/* if it's the last idx in log file, then return -ENOSPC */
if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
return -ENOSPC;
loghandle->lgh_last_idx++;
index = loghandle->lgh_last_idx;
LASSERT(index < LLOG_BITMAP_SIZE(llh));
rec->lrh_index = index;
if (buf == NULL) {
lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
sizeof(*lrt));
lrt->lrt_len = rec->lrh_len;
lrt->lrt_index = rec->lrh_index;
}
/* The caller should make sure only 1 process access the lgh_last_idx,
* Otherwise it might hit the assert.*/
LASSERT(index < LLOG_BITMAP_SIZE(llh));
spin_lock(&loghandle->lgh_hdr_lock);
if (ext2_set_bit(index, llh->llh_bitmap)) {
CERROR("%s: index %u already set in log bitmap\n",
o->do_lu.lo_dev->ld_obd->obd_name, index);
spin_unlock(&loghandle->lgh_hdr_lock);
LBUG(); /* should never happen */
}
llh->llh_count++;
spin_unlock(&loghandle->lgh_hdr_lock);
old_tail_idx = llh->llh_tail.lrt_index;
llh->llh_tail.lrt_index = index;
lgi->lgi_off = 0;
rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
th);
if (rc)
GOTO(out, rc);
rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
if (rc)
GOTO(out, rc);
LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
lgi->lgi_off = lgi->lgi_attr.la_size;
rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
out:
/* cleanup llog for error case */
if (rc) {
spin_lock(&loghandle->lgh_hdr_lock);
ext2_clear_bit(index, llh->llh_bitmap);
llh->llh_count--;
spin_unlock(&loghandle->lgh_hdr_lock);
/* restore the header */
loghandle->lgh_last_idx--;
llh->llh_tail.lrt_index = old_tail_idx;
lgi->lgi_off = 0;
llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
&lgi->lgi_off, th);
}
CDEBUG(D_RPCTRACE, "added record "DOSTID": idx: %u, %u\n",
POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
if (rc == 0 && reccookie) {
reccookie->lgc_lgl = loghandle->lgh_id;
reccookie->lgc_index = index;
if ((rec->lrh_type == MDS_UNLINK_REC) ||
(rec->lrh_type == MDS_SETATTR64_REC))
reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
else if (rec->lrh_type == OST_SZ_REC)
reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
else
reccookie->lgc_subsys = -1;
rc = 1;
}
return rc;
}
/* sets:
* - cur_offset to the furthest point read in the log file
* - cur_idx to the log index preceding cur_offset
* returns -EIO/-EINVAL on error
*/
static int llog_osd_next_block(const struct lu_env *env,
struct llog_handle *loghandle, int *cur_idx,
int next_idx, __u64 *cur_offset, void *buf,
int len)
{
struct llog_thread_info *lgi = llog_info(env);
struct dt_object *o;
struct dt_device *dt;
int rc;
LASSERT(env);
LASSERT(lgi);
if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
return -EINVAL;
CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
next_idx, *cur_idx, *cur_offset);
LASSERT(loghandle);
LASSERT(loghandle->lgh_ctxt);
o = loghandle->lgh_obj;
LASSERT(o);
LASSERT(dt_object_exists(o));
dt = lu2dt_dev(o->do_lu.lo_dev);
LASSERT(dt);
rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
if (rc)
GOTO(out, rc);
while (*cur_offset < lgi->lgi_attr.la_size) {
struct llog_rec_hdr *rec, *last_rec;
struct llog_rec_tail *tail;
llog_skip_over(cur_offset, *cur_idx, next_idx);
/* read up to next LLOG_CHUNK_SIZE block */
lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
(*cur_offset & (LLOG_CHUNK_SIZE - 1));
lgi->lgi_buf.lb_buf = buf;
/* Note: read lock is not needed around la_size get above at
* the time of dt_attr_get(). There are only two cases that
* matter. Either la_size == cur_offset, in which case the
* entire read is skipped, or la_size > cur_offset and the loop
* is entered and this thread is blocked at dt_read_lock()
* until the write is completed. When the write completes, then
* the dt_read() will be done with the full length, and will
* get the full data.
*/
dt_read_lock(env, o, 0);
rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
dt_read_unlock(env, o);
if (rc < 0) {
CERROR("%s: can't read llog block from log "DFID
" offset "LPU64": rc = %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
PFID(lu_object_fid(&o->do_lu)), *cur_offset,
rc);
GOTO(out, rc);
}
if (rc < len) {
/* signal the end of the valid buffer to
* llog_process */
memset(buf + rc, 0, len - rc);
}
if (rc == 0) /* end of file, nothing to do */
GOTO(out, rc);
if (rc < sizeof(*tail)) {
CERROR("%s: invalid llog block at log id "DOSTID"/%u "
"offset "LPU64"\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, *cur_offset);
GOTO(out, rc = -EINVAL);
}
rec = buf;
if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
lustre_swab_llog_rec(rec);
tail = (struct llog_rec_tail *)((char *)buf + rc -
sizeof(struct llog_rec_tail));
/* get the last record in block */
last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
le32_to_cpu(tail->lrt_len));
if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
lustre_swab_llog_rec(last_rec);
LASSERT(last_rec->lrh_index == tail->lrt_index);
*cur_idx = tail->lrt_index;
/* this shouldn't happen */
if (tail->lrt_index == 0) {
CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
"offset "LPU64"\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, *cur_offset);
GOTO(out, rc = -EINVAL);
}
if (tail->lrt_index < next_idx)
continue;
/* sanity check that the start of the new buffer is no farther
* than the record that we wanted. This shouldn't happen. */
if (rec->lrh_index > next_idx) {
CERROR("%s: missed desired record? %u > %u\n",
o->do_lu.lo_dev->ld_obd->obd_name,
rec->lrh_index, next_idx);
GOTO(out, rc = -ENOENT);
}
GOTO(out, rc = 0);
}
GOTO(out, rc = -EIO);
out:
return rc;
}
static int llog_osd_read_header(const struct lu_env *env,
struct llog_handle *handle)
{
struct llog_rec_hdr *llh_hdr;
struct dt_object *o;
struct llog_thread_info *lgi;
int rc;
LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
o = handle->lgh_obj;
LASSERT(o);
lgi = llog_info(env);
rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
if (rc)
return rc;
LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
if (lgi->lgi_attr.la_size == 0) {
CDEBUG(D_HA, "not reading header from 0-byte log\n");
return LLOG_EEMPTY;
}
lgi->lgi_off = 0;
lgi->lgi_buf.lb_buf = handle->lgh_hdr;
lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
if (rc) {
CERROR("%s: error reading log header from "DFID": rc = %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
PFID(lu_object_fid(&o->do_lu)), rc);
return rc;
}
llh_hdr = &handle->lgh_hdr->llh_hdr;
if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
lustre_swab_llog_hdr(handle->lgh_hdr);
if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
CERROR("%s: bad log %s "DFID" header magic: %#x "
"(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
handle->lgh_name ? handle->lgh_name : "",
PFID(lu_object_fid(&o->do_lu)),
llh_hdr->lrh_type, LLOG_HDR_MAGIC);
return -EIO;
} else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
CERROR("%s: incorrectly sized log %s "DFID" header: "
"%#x (expected %#x)\n"
"you may need to re-run lconf --write_conf.\n",
o->do_lu.lo_dev->ld_obd->obd_name,
handle->lgh_name ? handle->lgh_name : "",
PFID(lu_object_fid(&o->do_lu)),
llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
return -EIO;
}
handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
return 0;
}
/**
* 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 = NULL;
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))
GOTO(out, rc = 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, rc);
}
rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import, update,
&req);
if (rc != 0)
GOTO(out, 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:
if (req != NULL)
ptlrpc_req_finished(req);
if (update != NULL)
osp_update_request_destroy(update);
RETURN(rc);
}
/*
* Update a record in a quota index file.
*
* \param env - is the environment passed by the caller
* \param th - is the transaction to use for disk writes
* \param obj - is the on-disk index to be updated.
* \param id - is the key to be updated
* \param rec - is the input record containing the new quota settings.
* \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
* \param ver - is the new version of the index if LQUOTA_SET_VER is set or is
* used to return the new version of the index when
* LQUOTA_BUMP_VER is set.
*
* \retval - 0 on success, appropriate error on failure
*/
int lquota_disk_write(const struct lu_env *env, struct thandle *th,
struct dt_object *obj, union lquota_id *id,
struct dt_rec *rec, __u32 flags, __u64 *ver)
{
struct lquota_thread_info *qti = lquota_info(env);
struct dt_key *key = (struct dt_key *)&id->qid_uid;
int rc;
ENTRY;
LASSERT(dt_object_exists(obj));
LASSERT(obj->do_index_ops != NULL);
/* lock index */
dt_write_lock(env, obj, 0);
/* check whether there is already an existing record for this ID */
rc = dt_lookup(env, obj, (struct dt_rec *)&qti->qti_rec, key);
if (rc == 0) {
/* delete existing record in order to replace it */
rc = dt_delete(env, obj, key, th);
if (rc)
GOTO(out, rc);
} else if (rc == -ENOENT) {
/* probably first insert */
rc = 0;
} else {
GOTO(out, rc);
}
if (rec != NULL) {
/* insert record with updated quota settings */
rc = dt_insert(env, obj, rec, key, th, 1);
if (rc) {
/* try to insert the old one */
rc = dt_insert(env, obj, (struct dt_rec *)&qti->qti_rec,
key, th, 1);
LASSERTF(rc == 0, "failed to insert record in quota "
"index "DFID"\n",
PFID(lu_object_fid(&obj->do_lu)));
GOTO(out, rc);
}
}
if (flags != 0) {
LASSERT(ver);
if (flags & LQUOTA_BUMP_VER) {
/* caller wants to bump the version, let's first read
* it */
*ver = dt_version_get(env, obj);
(*ver)++;
} else {
LASSERT(flags & LQUOTA_SET_VER);
}
dt_version_set(env, obj, *ver, th);
}
EXIT;
out:
dt_write_unlock(env, obj);
return rc;
}
/*
* Look-up a slave index file. If the slave index isn't found:
* - if local is set to false, we allocate a FID from FID_SEQ_QUOTA sequence and
* create the index.
* - otherwise, we create the index file with a local reserved FID (see
* lquota_local_oid)
*
* \param env - is the environment passed by the caller
* \param dev - is the backend dt_device where to look-up/create the slave index
* \param parent - is the parent directory where to create the slave index if
* it does not exist already
* \param glb_fid - is the fid of the global index file associated with this
* slave index.
* \param uuid - is the uuid of slave which is (re)connecting to the master
* target
* \param local - indicate whether to use local reserved FID (LQUOTA_USR_OID
* & LQUOTA_GRP_OID) for the slave index creation or to
* allocate a new fid from sequence FID_SEQ_QUOTA
*
* \retval - pointer to the dt_object of the slave index on success,
* appropriate error on failure
*/
struct dt_object *lquota_disk_slv_find_create(const struct lu_env *env,
struct dt_device *dev,
struct dt_object *parent,
struct lu_fid *glb_fid,
struct obd_uuid *uuid,
bool local)
{
struct lquota_thread_info *qti = lquota_info(env);
struct dt_object *slv_idx;
int rc;
ENTRY;
LASSERT(uuid != NULL);
CDEBUG(D_QUOTA, "lookup/create slave index file for %s\n",
obd_uuid2str(uuid));
/* generate filename associated with the slave */
rc = lquota_disk_slv_filename(glb_fid, uuid, qti->qti_buf);
if (rc)
RETURN(ERR_PTR(rc));
/* Slave indexes uses the FID_SEQ_QUOTA sequence since they can be read
* through the network */
qti->qti_fid.f_seq = FID_SEQ_QUOTA;
qti->qti_fid.f_ver = 0;
if (local) {
int type;
rc = lquota_extract_fid(glb_fid, NULL, NULL, &type);
if (rc)
RETURN(ERR_PTR(rc));
/* use predefined fid in the reserved oid list */
qti->qti_fid.f_oid = qtype2slv_oid(type);
slv_idx = local_index_find_or_create_with_fid(env, dev,
&qti->qti_fid,
parent,
qti->qti_buf,
LQUOTA_MODE,
&dt_quota_slv_features);
} else {
/* allocate fid dynamically if index does not exist already */
qti->qti_fid.f_oid = LQUOTA_GENERATED_OID;
/* lookup/create slave index file */
slv_idx = lquota_disk_find_create(env, dev, parent,
&qti->qti_fid,
&dt_quota_slv_features,
qti->qti_buf);
}
if (IS_ERR(slv_idx))
RETURN(slv_idx);
/* install index operation vector */
if (slv_idx->do_index_ops == NULL) {
rc = slv_idx->do_ops->do_index_try(env, slv_idx,
&dt_quota_slv_features);
if (rc) {
CERROR("%s: failed to setup index operations for "DFID
" rc:%d\n", dev->dd_lu_dev.ld_obd->obd_name,
PFID(lu_object_fid(&slv_idx->do_lu)), rc);
dt_object_put(env, slv_idx);
slv_idx = ERR_PTR(rc);
}
}
RETURN(slv_idx);
}
