int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd, obd_id id, struct ofd_seq *oseq, int nr, int sync) { struct ofd_thread_info *info = ofd_info(env); struct ofd_object *fo = NULL; struct dt_object *next; struct thandle *th; struct ofd_object **batch; struct lu_fid *fid = &info->fti_fid; obd_id tmp; int rc; int i; int objects = 0; int nr_saved = nr; ENTRY; /* Don't create objects beyond the valid range for this SEQ */ if (unlikely(fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) && (id + nr) >= IDIF_MAX_OID)) { CERROR("%s:"DOSTID" hit the IDIF_MAX_OID (1<<48)!\n", ofd_name(ofd), id, ostid_seq(&oseq->os_oi)); RETURN(rc = -ENOSPC); } else if (unlikely(!fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) && (id + nr) >= OBIF_MAX_OID)) { CERROR("%s:"DOSTID" hit the OBIF_MAX_OID (1<<32)!\n", ofd_name(ofd), id, ostid_seq(&oseq->os_oi)); RETURN(rc = -ENOSPC); } OBD_ALLOC(batch, nr_saved * sizeof(struct ofd_object *)); if (batch == NULL) RETURN(-ENOMEM); info->fti_attr.la_valid = LA_TYPE | LA_MODE; /* * We mark object SUID+SGID to flag it for accepting UID+GID from * client on first write. Currently the permission bits on the OST are * never used, so this is OK. */ info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666; info->fti_dof.dof_type = dt_mode_to_dft(S_IFREG); /* Initialize a/c/m time so any client timestamp will always * be newer and update the inode. ctime = 0 is also handled * specially in osd_inode_setattr(). See LU-221, LU-1042 */ info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME; info->fti_attr.la_atime = 0; info->fti_attr.la_mtime = 0; info->fti_attr.la_ctime = 0; LASSERT(id != 0); /* prepare objects */ *fid = *lu_object_fid(&oseq->os_lastid_obj->do_lu); for (i = 0; i < nr; i++) { rc = fid_set_id(fid, id + i); if (rc != 0) { if (i == 0) GOTO(out, rc); nr = i; break; } fo = ofd_object_find(env, ofd, fid); if (IS_ERR(fo)) { if (i == 0) GOTO(out, rc = PTR_ERR(fo)); nr = i; break; } ofd_write_lock(env, fo); batch[i] = fo; } info->fti_buf.lb_buf = &tmp; info->fti_buf.lb_len = sizeof(tmp); info->fti_off = 0; th = ofd_trans_create(env, ofd); if (IS_ERR(th)) GOTO(out, rc = PTR_ERR(th)); th->th_sync |= sync; rc = dt_declare_record_write(env, oseq->os_lastid_obj, &info->fti_buf, info->fti_off, th); if (rc) GOTO(trans_stop, rc); for (i = 0; i < nr; i++) { fo = batch[i]; LASSERT(fo); if (unlikely(ofd_object_exists(fo))) { /* object may exist being re-created by write replay */ CDEBUG(D_INODE, "object "LPX64"/"LPX64" exists: " DFID"\n", ostid_seq(&oseq->os_oi), id, PFID(lu_object_fid(&fo->ofo_obj.do_lu))); continue; } next = ofd_object_child(fo); LASSERT(next != NULL); rc = dt_declare_create(env, next, &info->fti_attr, NULL, &info->fti_dof, th); if (rc) { nr = i; break; } } rc = dt_trans_start_local(env, ofd->ofd_osd, th); if (rc) GOTO(trans_stop, rc); CDEBUG(D_OTHER, "%s: create new object "DFID" nr %d\n", ofd_name(ofd), PFID(fid), nr); LASSERT(nr > 0); /* When the LFSCK scanning the whole device to verify the LAST_ID file * consistency, it will load the last_id into RAM firstly, and compare * the last_id with each OST-object's ID. If the later one is larger, * then it will regard the LAST_ID file crashed. But during the LFSCK * scanning, the OFD may continue to create new OST-objects. Those new * created OST-objects will have larger IDs than the LFSCK known ones. * So from the LFSCK view, it needs to re-load the last_id from disk * file, and if the latest last_id is still smaller than the object's * ID, then the LAST_ID file is real crashed. * * To make above mechanism to work, before OFD pre-create OST-objects, * it needs to update the LAST_ID file firstly, otherwise, the LFSCK * may cannot get latest last_id although new OST-object created. */ if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_SKIP_LASTID)) { tmp = cpu_to_le64(id + nr - 1); dt_write_lock(env, oseq->os_lastid_obj, 0); rc = dt_record_write(env, oseq->os_lastid_obj, &info->fti_buf, &info->fti_off, th); dt_write_unlock(env, oseq->os_lastid_obj); if (rc != 0) GOTO(trans_stop, rc); } for (i = 0; i < nr; i++) { fo = batch[i]; LASSERT(fo); /* Only the new created objects need to be recorded. */ if (ofd->ofd_osd->dd_record_fid_accessed) { lfsck_pack_rfa(&ofd_info(env)->fti_lr, lu_object_fid(&fo->ofo_obj.do_lu)); lfsck_in_notify(env, ofd->ofd_osd, &ofd_info(env)->fti_lr); } if (likely(!ofd_object_exists(fo) && !OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DANGLING))) { next = ofd_object_child(fo); LASSERT(next != NULL); rc = dt_create(env, next, &info->fti_attr, NULL, &info->fti_dof, th); if (rc) break; LASSERT(ofd_object_exists(fo)); } ofd_seq_last_oid_set(oseq, id + i); } objects = i; /* NOT all the wanted objects have been created, * set the LAST_ID as the real created. */ if (unlikely(objects < nr)) { int rc1; info->fti_off = 0; tmp = cpu_to_le64(ofd_seq_last_oid(oseq)); dt_write_lock(env, oseq->os_lastid_obj, 0); rc1 = dt_record_write(env, oseq->os_lastid_obj, &info->fti_buf, &info->fti_off, th); dt_write_unlock(env, oseq->os_lastid_obj); if (rc1 != 0) CERROR("%s: fail to reset the LAST_ID for seq ("LPX64 ") from "LPU64" to "LPU64"\n", ofd_name(ofd), ostid_seq(&oseq->os_oi), id + nr - 1, ofd_seq_last_oid(oseq)); } trans_stop: ofd_trans_stop(env, ofd, th, rc); out: for (i = 0; i < nr_saved; i++) { fo = batch[i]; if (fo) { ofd_write_unlock(env, fo); ofd_object_put(env, fo); } } OBD_FREE(batch, nr_saved * sizeof(struct ofd_object *)); CDEBUG((objects == 0 && rc == 0) ? D_ERROR : D_OTHER, "created %d/%d objects: %d\n", objects, nr_saved, rc); LASSERT(ergo(objects == 0, rc < 0)); RETURN(objects > 0 ? objects : rc); }
/** * Prepare buffers for write request processing. * * This function converts remote buffers from client to local buffers * and prepares the latter. If there is recovery in progress and required * object is missing then it can be re-created before write. * * \param[in] env execution environment * \param[in] exp OBD export of client * \param[in] ofd OFD device * \param[in] fid FID of object * \param[in] la object attributes * \param[in] oa OBDO structure from client * \param[in] objcount always 1 * \param[in] obj object data * \param[in] rnb remote buffers * \param[in] nr_local number of local buffers * \param[in] lnb local buffers * \param[in] jobid job ID name * * \retval 0 on successful prepare * \retval negative value on error */ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, struct ofd_device *ofd, const struct lu_fid *fid, struct lu_attr *la, struct obdo *oa, int objcount, struct obd_ioobj *obj, struct niobuf_remote *rnb, int *nr_local, struct niobuf_local *lnb, char *jobid) { struct ofd_object *fo; int i, j, k, rc = 0, tot_bytes = 0; ENTRY; LASSERT(env != NULL); LASSERT(objcount == 1); if (unlikely(exp->exp_obd->obd_recovering)) { u64 seq = fid_seq(fid); u64 oid = fid_oid(fid); struct ofd_seq *oseq; oseq = ofd_seq_load(env, ofd, seq); if (IS_ERR(oseq)) { CERROR("%s: Can't find FID Sequence "LPX64": rc = %d\n", ofd_name(ofd), seq, (int)PTR_ERR(oseq)); GOTO(out, rc = -EINVAL); } if (oid > ofd_seq_last_oid(oseq)) { int sync = 0; int diff; mutex_lock(&oseq->os_create_lock); diff = oid - ofd_seq_last_oid(oseq); /* Do sync create if the seq is about to used up */ if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) { if (unlikely(oid >= IDIF_MAX_OID - 1)) sync = 1; } else if (fid_seq_is_norm(seq)) { if (unlikely(oid >= LUSTRE_DATA_SEQ_MAX_WIDTH - 1)) sync = 1; } else { CERROR("%s : invalid o_seq "DOSTID"\n", ofd_name(ofd), POSTID(&oa->o_oi)); mutex_unlock(&oseq->os_create_lock); ofd_seq_put(env, oseq); GOTO(out, rc = -EINVAL); } while (diff > 0) { u64 next_id = ofd_seq_last_oid(oseq) + 1; int count = ofd_precreate_batch(ofd, diff); rc = ofd_precreate_objects(env, ofd, next_id, oseq, count, sync); if (rc < 0) { mutex_unlock(&oseq->os_create_lock); ofd_seq_put(env, oseq); GOTO(out, rc); } diff -= rc; } mutex_unlock(&oseq->os_create_lock); } ofd_seq_put(env, oseq); } fo = ofd_object_find(env, ofd, fid); if (IS_ERR(fo)) GOTO(out, rc = PTR_ERR(fo)); LASSERT(fo != NULL); ofd_read_lock(env, fo); if (!ofd_object_exists(fo)) { CERROR("%s: BRW to missing obj "DOSTID"\n", exp->exp_obd->obd_name, POSTID(&obj->ioo_oid)); ofd_read_unlock(env, fo); ofd_object_put(env, fo); GOTO(out, rc = -ENOENT); } if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) { rc = ofd_verify_ff(env, fo, oa); if (rc != 0) { ofd_read_unlock(env, fo); ofd_object_put(env, fo); GOTO(out, rc); } } /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some * space back if possible */ ofd_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt); /* parse remote buffers to local buffers and prepare the latter */ *nr_local = 0; for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) { rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i, lnb + j, 1); if (unlikely(rc < 0)) GOTO(err, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); /* correct index for local buffers to continue with */ for (k = 0; k < rc; k++) { lnb[j+k].lnb_flags = rnb[i].rnb_flags; if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED)) lnb[j+k].lnb_rc = -ENOSPC; } j += rc; *nr_local += rc; LASSERT(j <= PTLRPC_MAX_BRW_PAGES); tot_bytes += rnb[i].rnb_len; } LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local); if (unlikely(rc != 0)) GOTO(err, rc); ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes); RETURN(0); err: dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); ofd_read_unlock(env, fo); ofd_object_put(env, fo); /* ofd_grant_prepare_write() was called, so we must commit */ ofd_grant_commit(exp, oa->o_grant_used, rc); out: /* let's still process incoming grant information packed in the oa, * but without enforcing grant since we won't proceed with the write. * Just like a read request actually. */ ofd_grant_prepare_read(env, exp, oa); return rc; }
int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd, obd_id id, struct ofd_seq *oseq, int nr, int sync) { struct ofd_thread_info *info = ofd_info(env); struct ofd_object *fo = NULL; struct dt_object *next; struct thandle *th; struct ofd_object **batch; obd_id tmp; int rc; int i; int objects = 0; int nr_saved = nr; ENTRY; /* Don't create objects beyond the valid range for this SEQ */ if (unlikely(fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) && (id + nr) >= IDIF_MAX_OID)) { CERROR("%s:"DOSTID" hit the IDIF_MAX_OID (1<<48)!\n", ofd_name(ofd), id, ostid_seq(&oseq->os_oi)); RETURN(rc = -ENOSPC); } else if (unlikely(!fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) && (id + nr) >= OBIF_MAX_OID)) { CERROR("%s:"DOSTID" hit the OBIF_MAX_OID (1<<32)!\n", ofd_name(ofd), id, ostid_seq(&oseq->os_oi)); RETURN(rc = -ENOSPC); } OBD_ALLOC(batch, nr_saved * sizeof(struct ofd_object *)); if (batch == NULL) RETURN(-ENOMEM); info->fti_attr.la_valid = LA_TYPE | LA_MODE; /* * We mark object SUID+SGID to flag it for accepting UID+GID from * client on first write. Currently the permission bits on the OST are * never used, so this is OK. */ info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666; info->fti_dof.dof_type = dt_mode_to_dft(S_IFREG); /* Initialize a/c/m time so any client timestamp will always * be newer and update the inode. ctime = 0 is also handled * specially in osd_inode_setattr(). See LU-221, LU-1042 */ info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME; info->fti_attr.la_atime = 0; info->fti_attr.la_mtime = 0; info->fti_attr.la_ctime = 0; /* prepare objects */ ostid_set_seq(&info->fti_ostid, ostid_seq(&oseq->os_oi)); for (i = 0; i < nr; i++) { ostid_set_id(&info->fti_ostid, id + i); rc = ostid_to_fid(&info->fti_fid, &info->fti_ostid, 0); if (rc) { if (i == 0) GOTO(out, rc); nr = i; break; } fo = ofd_object_find(env, ofd, &info->fti_fid); if (IS_ERR(fo)) { if (i == 0) GOTO(out, rc = PTR_ERR(fo)); nr = i; break; } ofd_write_lock(env, fo); batch[i] = fo; } info->fti_buf.lb_buf = &tmp; info->fti_buf.lb_len = sizeof(tmp); info->fti_off = 0; th = ofd_trans_create(env, ofd); if (IS_ERR(th)) GOTO(out, rc = PTR_ERR(th)); th->th_sync |= sync; rc = dt_declare_record_write(env, oseq->os_lastid_obj, sizeof(tmp), info->fti_off, th); if (rc) GOTO(trans_stop, rc); for (i = 0; i < nr; i++) { fo = batch[i]; LASSERT(fo); if (unlikely(ofd_object_exists(fo))) { /* object may exist being re-created by write replay */ CDEBUG(D_INODE, "object "LPX64"/"LPX64" exists: " DFID"\n", ostid_seq(&oseq->os_oi), id, PFID(&info->fti_fid)); continue; } next = ofd_object_child(fo); LASSERT(next != NULL); rc = dt_declare_create(env, next, &info->fti_attr, NULL, &info->fti_dof, th); if (rc) { nr = i; break; } } rc = dt_trans_start_local(env, ofd->ofd_osd, th); if (rc) GOTO(trans_stop, rc); CDEBUG(D_OTHER, "%s: create new object "DFID" nr %d\n", ofd_name(ofd), PFID(&info->fti_fid), nr); for (i = 0; i < nr; i++) { fo = batch[i]; LASSERT(fo); if (likely(!ofd_object_exists(fo))) { next = ofd_object_child(fo); LASSERT(next != NULL); rc = dt_create(env, next, &info->fti_attr, NULL, &info->fti_dof, th); if (rc) break; LASSERT(ofd_object_exists(fo)); } ofd_seq_last_oid_set(oseq, id + i); } objects = i; if (objects > 0) { tmp = cpu_to_le64(ofd_seq_last_oid(oseq)); rc = dt_record_write(env, oseq->os_lastid_obj, &info->fti_buf, &info->fti_off, th); } trans_stop: ofd_trans_stop(env, ofd, th, rc); out: for (i = 0; i < nr_saved; i++) { fo = batch[i]; if (fo) { ofd_write_unlock(env, fo); ofd_object_put(env, fo); } } OBD_FREE(batch, nr_saved * sizeof(struct ofd_object *)); CDEBUG((objects == 0 && rc == 0) ? D_ERROR : D_OTHER, "created %d/%d objects: %d\n", objects, nr_saved, rc); LASSERT(ergo(objects == 0, rc < 0)); RETURN(objects > 0 ? objects : rc); }