/** * Implementation of the llog_operations::lop_write * * This function writes the new record in the llog or modify the existed one. * * \param[in] env execution environment * \param[in] loghandle llog handle of the current llog * \param[in] rec llog record header. This is a real header of * the full llog record to write. This is * the beginning of buffer to write, the length * of buffer is stored in \a rec::lrh_len * \param[out] reccookie pointer to the cookie to return back if needed. * It is used for further cancel of this llog * record. * \param[in] idx index of the llog record. If \a idx == -1 then * this is append case, otherwise \a idx is * the index of record to modify * \param[in] th current transaction handle * * \retval 0 on successful write && \a reccookie == NULL * 1 on successful write && \a reccookie != NULL * \retval negative error if write failed */ 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 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; struct llog_rec_tail *lrt; struct dt_object *o; size_t left; bool header_is_updated = false; ENTRY; 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 (reclen > LLOG_CHUNK_SIZE) RETURN(-E2BIG); rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL); if (rc) RETURN(rc); /** * The modification case. * If idx set then the record with that index must be modified. * There are three cases possible: * 1) the common case is the llog header update (idx == 0) * 2) the llog record modification during llog process. * This is indicated by the \a loghandle::lgh_cur_idx > 0. * In that case the \a loghandle::lgh_cur_offset * 3) otherwise this is assumed that llog consist of records of * fixed size, i.e. catalog. The llog header must has llh_size * field equal to record size. The record offset is calculated * just by /a idx value * * During modification we don't need extra header update because * the bitmap and record count are not changed. The record header * and tail remains the same too. */ if (idx != LLOG_NEXT_IDX) { /* llog can be empty only when first record is being written */ LASSERT(ergo(idx > 0, lgi->lgi_attr.la_size > 0)); 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); RETURN(-ENOENT); } if (idx != rec->lrh_index) { CERROR("%s: modify index mismatch %d %u\n", o->do_lu.lo_dev->ld_obd->obd_name, idx, rec->lrh_index); RETURN(-EFAULT); } if (idx == LLOG_HEADER_IDX) { /* llog header update */ LASSERT(reclen == sizeof(struct llog_log_hdr)); LASSERT(rec == &llh->llh_hdr); lgi->lgi_off = 0; lgi->lgi_buf.lb_len = reclen; lgi->lgi_buf.lb_buf = rec; rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th); RETURN(rc); } else if (loghandle->lgh_cur_idx > 0) { /** * The lgh_cur_offset can be used only if index is * the same. */ if (idx != loghandle->lgh_cur_idx) { CERROR("%s: modify index mismatch %d %d\n", o->do_lu.lo_dev->ld_obd->obd_name, idx, loghandle->lgh_cur_idx); RETURN(-EFAULT); } lgi->lgi_off = loghandle->lgh_cur_offset; CDEBUG(D_OTHER, "modify record "DOSTID": idx:%d, " "len:%u offset %llu\n", POSTID(&loghandle->lgh_id.lgl_oi), idx, rec->lrh_len, (long long)lgi->lgi_off); } else if (llh->llh_size > 0) { if (llh->llh_size != rec->lrh_len) { CERROR("%s: wrong record size, llh_size is %u" " but record size is %u\n", o->do_lu.lo_dev->ld_obd->obd_name, llh->llh_size, rec->lrh_len); RETURN(-EINVAL); } lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen; } else { /* This can be result of lgh_cur_idx is not set during * llog processing or llh_size is not set to proper * record size for fixed records llog. Therefore it is * impossible to get record offset. */ CERROR("%s: can't get record offset, idx:%d, " "len:%u.\n", o->do_lu.lo_dev->ld_obd->obd_name, idx, rec->lrh_len); RETURN(-EFAULT); } /* update only data, header and tail remain the same */ lgi->lgi_off += sizeof(struct llog_rec_hdr); lgi->lgi_buf.lb_len = REC_DATA_LEN(rec); lgi->lgi_buf.lb_buf = REC_DATA(rec); rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th); if (rc == 0 && reccookie) { reccookie->lgc_lgl = loghandle->lgh_id; reccookie->lgc_index = idx; rc = 1; } RETURN(rc); } /** * The append case. * The most common case of using llog. The new index is assigned to * the new record, new bit is set in llog bitmap and llog count is * incremented. * * 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. */ 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); /* increment the last_idx along with llh_tail index, they should * be equal for a llog lifetime */ loghandle->lgh_last_idx++; index = loghandle->lgh_last_idx; llh->llh_tail.lrt_index = index; /** * NB: the caller should make sure only 1 process access * the lgh_last_idx, e.g. append should be exclusive. * Otherwise it might hit the assert. */ LASSERT(index < LLOG_BITMAP_SIZE(llh)); rec->lrh_index = index; lrt = rec_tail(rec); lrt->lrt_len = rec->lrh_len; lrt->lrt_index = rec->lrh_index; /* the lgh_hdr_lock protects llog header data from concurrent * update/cancel, the llh_count and llh_bitmap are protected */ 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); lgi->lgi_off = 0; lgi->lgi_buf.lb_len = llh->llh_hdr.lrh_len; lgi->lgi_buf.lb_buf = &llh->llh_hdr; rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th); if (rc) GOTO(out, rc); header_is_updated = true; 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; lgi->lgi_buf.lb_len = reclen; lgi->lgi_buf.lb_buf = rec; rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th); if (rc < 0) GOTO(out, rc); CDEBUG(D_OTHER, "added record "DOSTID": idx: %u, %u\n", POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len); if (reccookie != NULL) { 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); out: /* cleanup llog for error case */ spin_lock(&loghandle->lgh_hdr_lock); ext2_clear_bit(index, llh->llh_bitmap); llh->llh_count--; spin_unlock(&loghandle->lgh_hdr_lock); /* restore llog last_idx */ loghandle->lgh_last_idx--; llh->llh_tail.lrt_index = loghandle->lgh_last_idx; /* restore the header on disk if it was written */ if (header_is_updated) { lgi->lgi_off = 0; lgi->lgi_buf.lb_len = llh->llh_hdr.lrh_len; lgi->lgi_buf.lb_buf = &llh->llh_hdr; dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th); } 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; struct llog_rec_tail *lrt; struct dt_object *o; size_t left; ENTRY; 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 "LPX64": idx:%d/%u/%d, len:%u " "offset %llu\n", loghandle->lgh_id.lgl_oid, 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); 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) RETURN(rc); rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL); if (rc) RETURN(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); if (rc) RETURN(rc); CDEBUG(D_RPCTRACE, "added record "LPX64": idx: %u, %u\n", loghandle->lgh_id.lgl_oid, 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); }