X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flod%2Flod_object.c;h=13a7e31aac9faa2fb1cb5d48be0bebd85b45496a;hp=dc896ef533f892367978916e1a1d66b2386c4dda;hb=e2ac6e1eaa108eef3493837e9bd881629582ea1d;hpb=0a83d948f37bec7fca6e9aa30f59f26354273b23 diff --git a/lustre/lod/lod_object.c b/lustre/lod/lod_object.c index dc896ef..13a7e31 100644 --- a/lustre/lod/lod_object.c +++ b/lustre/lod/lod_object.c @@ -98,9 +98,9 @@ static int lod_declare_insert(const struct lu_env *env, struct dt_object *dt, */ static int lod_insert(const struct lu_env *env, struct dt_object *dt, const struct dt_rec *rec, const struct dt_key *key, - struct thandle *th, int ign) + struct thandle *th) { - return lod_sub_insert(env, dt_object_child(dt), rec, key, th, ign); + return lod_sub_insert(env, dt_object_child(dt), rec, key, th); } /** @@ -377,15 +377,24 @@ static struct dt_index_operations lod_index_ops = { static struct dt_it *lod_striped_it_init(const struct lu_env *env, struct dt_object *dt, __u32 attr) { - struct lod_object *lo = lod_dt_obj(dt); - struct dt_object *next; - struct lod_it *it = &lod_env_info(env)->lti_it; - struct dt_it *it_next; - ENTRY; + struct lod_object *lo = lod_dt_obj(dt); + struct dt_object *next; + struct lod_it *it = &lod_env_info(env)->lti_it; + struct dt_it *it_next; + __u16 index = 0; LASSERT(lo->ldo_dir_stripe_count > 0); - next = lo->ldo_stripe[0]; - LASSERT(next != NULL); + + do { + next = lo->ldo_stripe[index]; + if (next && dt_object_exists(next)) + break; + } while (++index < lo->ldo_dir_stripe_count); + + /* no valid stripe */ + if (!next || !dt_object_exists(next)) + return ERR_PTR(-ENODEV); + LASSERT(next->do_index_ops != NULL); it_next = next->do_index_ops->dio_it.init(env, next, attr); @@ -398,7 +407,7 @@ static struct dt_it *lod_striped_it_init(const struct lu_env *env, * additional ones */ LASSERT(it->lit_obj == NULL); - it->lit_stripe_index = 0; + it->lit_stripe_index = index; it->lit_attr = attr; it->lit_it = it_next; it->lit_obj = dt; @@ -433,10 +442,10 @@ static void lod_striped_it_fini(const struct lu_env *env, struct dt_it *di) LOD_CHECK_STRIPED_IT(env, it, lo); next = lo->ldo_stripe[it->lit_stripe_index]; - LASSERT(next != NULL); - LASSERT(next->do_index_ops != NULL); - - next->do_index_ops->dio_it.fini(env, it->lit_it); + if (next) { + LASSERT(next->do_index_ops != NULL); + next->do_index_ops->dio_it.fini(env, it->lit_it); + } } /* the iterator not in use any more */ @@ -457,15 +466,15 @@ static void lod_striped_it_fini(const struct lu_env *env, struct dt_it *di) static int lod_striped_it_get(const struct lu_env *env, struct dt_it *di, const struct dt_key *key) { - const struct lod_it *it = (const struct lod_it *)di; - struct lod_object *lo = lod_dt_obj(it->lit_obj); - struct dt_object *next; - ENTRY; + const struct lod_it *it = (const struct lod_it *)di; + struct lod_object *lo = lod_dt_obj(it->lit_obj); + struct dt_object *next; LOD_CHECK_STRIPED_IT(env, it, lo); next = lo->ldo_stripe[it->lit_stripe_index]; LASSERT(next != NULL); + LASSERT(dt_object_exists(next)); LASSERT(next->do_index_ops != NULL); return next->do_index_ops->dio_it.get(env, it->lit_it, key); @@ -480,9 +489,16 @@ static int lod_striped_it_get(const struct lu_env *env, struct dt_it *di, */ static void lod_striped_it_put(const struct lu_env *env, struct dt_it *di) { - struct lod_it *it = (struct lod_it *)di; - struct lod_object *lo = lod_dt_obj(it->lit_obj); - struct dt_object *next; + struct lod_it *it = (struct lod_it *)di; + struct lod_object *lo = lod_dt_obj(it->lit_obj); + struct dt_object *next; + + /* + * If lit_it == NULL, then it means the sub_it has been finished, + * which only happens in failure cases, see lod_striped_it_next() + */ + if (!it->lit_it) + return; LOD_CHECK_STRIPED_IT(env, it, lo); @@ -503,17 +519,20 @@ static void lod_striped_it_put(const struct lu_env *env, struct dt_it *di) */ static int lod_striped_it_next(const struct lu_env *env, struct dt_it *di) { - struct lod_it *it = (struct lod_it *)di; - struct lod_object *lo = lod_dt_obj(it->lit_obj); - struct dt_object *next; - struct dt_it *it_next; - int rc; + struct lod_it *it = (struct lod_it *)di; + struct lod_object *lo = lod_dt_obj(it->lit_obj); + struct dt_object *next; + struct dt_it *it_next; + __u32 index; + int rc; + ENTRY; LOD_CHECK_STRIPED_IT(env, it, lo); next = lo->ldo_stripe[it->lit_stripe_index]; LASSERT(next != NULL); + LASSERT(dt_object_exists(next)); LASSERT(next->do_index_ops != NULL); again: rc = next->do_index_ops->dio_it.next(env, it->lit_it); @@ -546,33 +565,44 @@ again: RETURN(rc); } - /* go to next stripe */ - if (it->lit_stripe_index + 1 >= lo->ldo_dir_stripe_count) - RETURN(1); - - it->lit_stripe_index++; - next->do_index_ops->dio_it.put(env, it->lit_it); next->do_index_ops->dio_it.fini(env, it->lit_it); it->lit_it = NULL; - next = lo->ldo_stripe[it->lit_stripe_index]; - LASSERT(next != NULL); - rc = next->do_ops->do_index_try(env, next, &dt_directory_features); - if (rc != 0) - RETURN(rc); + /* go to next stripe */ + index = it->lit_stripe_index; + while (++index < lo->ldo_dir_stripe_count) { + next = lo->ldo_stripe[index]; + if (!next) + continue; - LASSERT(next->do_index_ops != NULL); + if (!dt_object_exists(next)) + continue; + + rc = next->do_ops->do_index_try(env, next, + &dt_directory_features); + if (rc != 0) + RETURN(rc); + + LASSERT(next->do_index_ops != NULL); + + it_next = next->do_index_ops->dio_it.init(env, next, + it->lit_attr); + if (IS_ERR(it_next)) + RETURN(PTR_ERR(it_next)); + + rc = next->do_index_ops->dio_it.get(env, it_next, + (const struct dt_key *)""); + if (rc <= 0) + RETURN(rc == 0 ? -EIO : rc); - it_next = next->do_index_ops->dio_it.init(env, next, it->lit_attr); - if (!IS_ERR(it_next)) { it->lit_it = it_next; + it->lit_stripe_index = index; goto again; - } else { - rc = PTR_ERR(it_next); + } - RETURN(rc); + RETURN(1); } /** @@ -956,7 +986,9 @@ static int lod_index_try(const struct lu_env *env, struct dt_object *dt, int i; for (i = 0; i < lo->ldo_dir_stripe_count; i++) { - if (dt_object_exists(lo->ldo_stripe[i]) == 0) + if (!lo->ldo_stripe[i]) + continue; + if (!dt_object_exists(lo->ldo_stripe[i])) continue; rc = lo->ldo_stripe[i]->do_ops->do_index_try(env, lo->ldo_stripe[i], feat); @@ -1044,12 +1076,16 @@ static int lod_attr_get(const struct lu_env *env, } static inline void lod_adjust_stripe_info(struct lod_layout_component *comp, - struct lov_desc *desc) + struct lov_desc *desc, + int append_stripes) { if (comp->llc_pattern != LOV_PATTERN_MDT) { - if (!comp->llc_stripe_count) + if (append_stripes) { + comp->llc_stripe_count = append_stripes; + } else if (!comp->llc_stripe_count) { comp->llc_stripe_count = desc->ld_default_stripe_count; + } } if (comp->llc_stripe_size <= 0) comp->llc_stripe_size = desc->ld_default_stripe_size; @@ -1241,6 +1277,8 @@ static int lod_declare_attr_set(const struct lu_env *env, for (i = 0; i < lo->ldo_dir_stripe_count; i++) { if (lo->ldo_stripe[i] == NULL) continue; + if (!dt_object_exists(lo->ldo_stripe[i])) + continue; rc = lod_sub_declare_attr_set(env, lo->ldo_stripe[i], attr, th); if (rc != 0) @@ -1385,7 +1423,7 @@ static int lod_attr_set(const struct lu_env *env, buf->lb_len = info->lti_ea_store_size; lmm = info->lti_ea_store; magic = le32_to_cpu(lmm->lmm_magic); - if (magic == LOV_MAGIC_COMP_V1) { + if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) { struct lov_comp_md_v1 *lcm = buf->lb_buf; struct lov_comp_md_entry_v1 *lcme = &lcm->lcm_entries[0]; @@ -1419,7 +1457,8 @@ static int lod_attr_set(const struct lu_env *env, buf->lb_buf = info->lti_ea_store; buf->lb_len = info->lti_ea_store_size; lcm = buf->lb_buf; - if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1) + if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1 && + le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_SEL) RETURN(-EINVAL); le32_add_cpu(&lcm->lcm_layout_gen, 1); @@ -1454,21 +1493,35 @@ static int lod_xattr_get(const struct lu_env *env, struct dt_object *dt, rc = dt_xattr_get(env, dt_object_child(dt), buf, name); if (strcmp(name, XATTR_NAME_LMV) == 0) { struct lmv_mds_md_v1 *lmv1; + struct lmv_foreign_md *lfm; int rc1 = 0; if (rc > (typeof(rc))sizeof(*lmv1)) RETURN(rc); - if (rc < (typeof(rc))sizeof(*lmv1)) + /* short (<= sizeof(struct lmv_mds_md_v1)) foreign LMV case */ + /* XXX empty foreign LMV is not allowed */ + if (rc <= offsetof(typeof(*lfm), lfm_value)) RETURN(rc = rc > 0 ? -EINVAL : rc); if (buf->lb_buf == NULL || buf->lb_len == 0) { CLASSERT(sizeof(*lmv1) <= sizeof(info->lti_key)); + /* lti_buf is large enough for *lmv1 or a short + * (<= sizeof(struct lmv_mds_md_v1)) foreign LMV + */ info->lti_buf.lb_buf = info->lti_key; info->lti_buf.lb_len = sizeof(*lmv1); rc = dt_xattr_get(env, dt_object_child(dt), &info->lti_buf, name); + if (unlikely(rc <= offsetof(typeof(*lfm), + lfm_value))) + RETURN(rc = rc > 0 ? -EINVAL : rc); + + lfm = info->lti_buf.lb_buf; + if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) + RETURN(rc); + if (unlikely(rc != sizeof(*lmv1))) RETURN(rc = rc > 0 ? -EINVAL : rc); @@ -1481,6 +1534,13 @@ static int lod_xattr_get(const struct lu_env *env, struct dt_object *dt, le32_to_cpu(lmv1->lmv_stripe_count), LMV_MAGIC_V1); } else { + lfm = buf->lb_buf; + if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) + RETURN(rc); + + if (rc != sizeof(*lmv1)) + RETURN(rc = rc > 0 ? -EINVAL : rc); + rc1 = lod_load_lmv_shards(env, lod_dt_obj(dt), buf, false); } @@ -1488,6 +1548,13 @@ static int lod_xattr_get(const struct lu_env *env, struct dt_object *dt, RETURN(rc = rc1 != 0 ? rc1 : rc); } + if ((rc > 0) && buf->lb_buf && strcmp(name, XATTR_NAME_LOV) == 0) { + struct lov_comp_md_v1 *lcm = buf->lb_buf; + + if (lcm->lcm_magic == cpu_to_le32(LOV_MAGIC_SEL)) + lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1); + } + if (rc != -ENODATA || !S_ISDIR(dt->do_lu.lo_header->loh_attr & S_IFMT)) RETURN(rc); @@ -1662,7 +1729,8 @@ int lod_parse_dir_striping(const struct lu_env *env, struct lod_object *lo, LASSERT(mutex_is_locked(&lo->ldo_layout_mutex)); - if (le32_to_cpu(lmv1->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION) + /* XXX may be useless as not called for foreign LMV ?? */ + if (le32_to_cpu(lmv1->lmv_magic) == LMV_MAGIC_FOREIGN) RETURN(0); if (le32_to_cpu(lmv1->lmv_magic) == LMV_MAGIC_STRIPE) { @@ -1689,8 +1757,10 @@ int lod_parse_dir_striping(const struct lu_env *env, struct lod_object *lo, __u32 idx; fid_le_to_cpu(fid, &lmv1->lmv_stripe_fids[i]); - if (!fid_is_sane(fid)) - GOTO(out, rc = -ESTALE); + if (!fid_is_sane(fid)) { + stripe[i] = NULL; + continue; + } rc = lod_fld_lookup(env, lod, fid, &idx, &type); if (rc != 0) @@ -1785,6 +1855,10 @@ static int lod_dir_declare_create_stripes(const struct lu_env *env, struct linkea_data ldata = { NULL }; struct lu_buf linkea_buf; + /* OBD_FAIL_MDS_STRIPE_FID may leave stripe uninitialized */ + if (!dto) + continue; + rc = lod_sub_declare_create(env, dto, attr, NULL, dof, th); if (rc != 0) GOTO(out, rc); @@ -1895,9 +1969,8 @@ static int lod_prep_md_striped_create(const struct lu_env *env, /* The lum has been verifed in lod_verify_md_striping */ LASSERT(le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC || le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC); - LASSERT(le32_to_cpu(lum->lum_stripe_count) > 0); - stripe_count = le32_to_cpu(lum->lum_stripe_count); + stripe_count = lo->ldo_dir_stripe_count; OBD_ALLOC(idx_array, sizeof(idx_array[0]) * stripe_count); if (idx_array == NULL) @@ -2013,6 +2086,11 @@ static int lod_prep_md_striped_create(const struct lu_env *env, * in the above loop */ LASSERT(tgt_dt != NULL); LASSERT(fid_is_sane(&fid)); + + /* fail a remote stripe FID allocation */ + if (i && OBD_FAIL_CHECK(OBD_FAIL_MDS_STRIPE_FID)) + continue; + conf.loc_flags = LOC_F_NEW; dto = dt_locate_at(env, tgt_dt, &fid, dt->do_lu.lo_dev->ld_site->ls_top_dev, @@ -2054,6 +2132,27 @@ out_free: } /** + * + * Alloc cached foreign LMV + * + * \param[in] lo object + * \param[in] size size of foreign LMV + * + * \retval 0 on success + * \retval negative if failed + */ +int lod_alloc_foreign_lmv(struct lod_object *lo, size_t size) +{ + OBD_ALLOC_LARGE(lo->ldo_foreign_lmv, size); + if (lo->ldo_foreign_lmv == NULL) + return -ENOMEM; + lo->ldo_foreign_lmv_size = size; + lo->ldo_dir_is_foreign = 1; + + return 0; +} + +/** * Declare create striped md object. * * The function declares intention to create a striped directory. This is a @@ -2090,8 +2189,16 @@ static int lod_declare_xattr_set_lmv(const struct lu_env *env, le32_to_cpu(lum->lum_magic), le32_to_cpu(lum->lum_stripe_count), (int)le32_to_cpu(lum->lum_stripe_offset)); - if (lo->ldo_dir_stripe_count == 0) + if (lo->ldo_dir_stripe_count == 0) { + if (lo->ldo_dir_is_foreign) { + rc = lod_alloc_foreign_lmv(lo, lum_buf->lb_len); + if (rc != 0) + GOTO(out, rc); + memcpy(lo->ldo_foreign_lmv, lum, lum_buf->lb_len); + lo->ldo_dir_stripe_loaded = 1; + } GOTO(out, rc = 0); + } /* prepare dir striped objects */ rc = lod_prep_md_striped_create(env, dt, attr, lum, dof, th); @@ -2167,7 +2274,7 @@ static int lod_dir_declare_layout_add(const struct lu_env *env, fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]); if (!fid_is_sane(fid)) - GOTO(out, rc = -ESTALE); + continue; rc = lod_fld_lookup(env, lod, fid, &idx, &type); if (rc) @@ -2288,7 +2395,8 @@ static int lod_dir_declare_layout_delete(const struct lu_env *env, for (i = final_stripe_count; i < lo->ldo_dir_stripe_count; i++) { dto = lo->ldo_stripe[i]; - LASSERT(dto); + if (!dto) + continue; if (!dt_try_as_dir(env, dto)) return -ENOTDIR; @@ -2358,7 +2466,8 @@ static int lod_dir_layout_delete(const struct lu_env *env, for (i = final_stripe_count; i < lo->ldo_dir_stripe_count; i++) { dto = lo->ldo_stripe[i]; - LASSERT(dto); + if (!dto) + continue; rc = lod_sub_delete(env, dto, (const struct dt_key *)dotdot, th); @@ -2440,7 +2549,11 @@ static int lod_dir_declare_xattr_set(const struct lu_env *env, RETURN(0); for (i = 0; i < lo->ldo_dir_stripe_count; i++) { - LASSERT(lo->ldo_stripe[i]); + if (!lo->ldo_stripe[i]) + continue; + + if (!dt_object_exists(lo->ldo_stripe[i])) + continue; rc = lod_sub_declare_xattr_set(env, lo->ldo_stripe[i], buf, name, fl, th); @@ -2473,20 +2586,25 @@ lod_obj_stripe_replace_parent_fid_cb(const struct lu_env *env, return rc; } - filter_fid_le_to_cpu(ff, ff, sizeof(*ff)); - /* - * mdd_declare_migrate_create() declares this via source object because - * target is not ready yet, so declare anyway. + * locd_buf is set if it's called by dir migration, which doesn't check + * pfid and comp id. */ - if (!data->locd_declare && - lu_fid_eq(lu_object_fid(&lo->ldo_obj.do_lu), &ff->ff_parent) && - ff->ff_layout.ol_comp_id == comp->llc_id) - return 0; + if (data->locd_buf) { + memset(ff, 0, sizeof(*ff)); + ff->ff_parent = *(struct lu_fid *)data->locd_buf->lb_buf; + } else { + filter_fid_le_to_cpu(ff, ff, sizeof(*ff)); + + if (lu_fid_eq(lod_object_fid(lo), &ff->ff_parent) && + ff->ff_layout.ol_comp_id == comp->llc_id) + return 0; + + memset(ff, 0, sizeof(*ff)); + ff->ff_parent = *lu_object_fid(&lo->ldo_obj.do_lu); + } /* rewrite filter_fid */ - memset(ff, 0, sizeof(*ff)); - ff->ff_parent = *lu_object_fid(&lo->ldo_obj.do_lu); ff->ff_parent.f_ver = stripe_idx; ff->ff_layout.ol_stripe_size = comp->llc_stripe_size; ff->ff_layout.ol_stripe_count = comp->llc_stripe_count; @@ -2522,11 +2640,11 @@ lod_obj_stripe_replace_parent_fid_cb(const struct lu_env *env, */ static int lod_replace_parent_fid(const struct lu_env *env, struct dt_object *dt, + const struct lu_buf *buf, struct thandle *th, bool declare) { struct lod_object *lo = lod_dt_obj(dt); struct lod_thread_info *info = lod_env_info(env); - struct lu_buf *buf = &info->lti_buf; struct filter_fid *ff; struct lod_obj_stripe_cb_data data = { { 0 } }; int rc; @@ -2548,11 +2666,9 @@ static int lod_replace_parent_fid(const struct lu_env *env, RETURN(rc); } - buf->lb_buf = info->lti_ea_store; - buf->lb_len = info->lti_ea_store_size; - data.locd_declare = declare; data.locd_stripe_cb = lod_obj_stripe_replace_parent_fid_cb; + data.locd_buf = buf; rc = lod_obj_for_each_stripe(env, lo, th, &data); RETURN(rc); @@ -2571,7 +2687,8 @@ inline __u16 lod_comp_entry_stripe_count(struct lod_object *lo, else if ((__u16)-1 == entry->llc_stripe_count) return lod->lod_desc.ld_tgt_count; else - return lod_get_stripe_count(lod, lo, entry->llc_stripe_count); + return lod_get_stripe_count(lod, lo, + entry->llc_stripe_count, false); } static int lod_comp_md_size(struct lod_object *lo, bool is_dir) @@ -2579,7 +2696,7 @@ static int lod_comp_md_size(struct lod_object *lo, bool is_dir) int magic, size = 0, i; struct lod_layout_component *comp_entries; __u16 comp_cnt; - bool is_composite; + bool is_composite, is_foreign = false; if (is_dir) { comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt; @@ -2590,8 +2707,11 @@ static int lod_comp_md_size(struct lod_object *lo, bool is_dir) comp_cnt = lo->ldo_comp_cnt; comp_entries = lo->ldo_comp_entries; is_composite = lo->ldo_is_composite; + is_foreign = lo->ldo_is_foreign; } + if (is_foreign) + return lo->ldo_foreign_lov_size; LASSERT(comp_cnt != 0 && comp_entries != NULL); if (is_composite) { @@ -2688,7 +2808,7 @@ static int lod_declare_layout_add(const struct lu_env *env, lod_comp->llc_stripe_count = v1->lmm_stripe_count; lod_comp->llc_stripe_size = v1->lmm_stripe_size; - lod_adjust_stripe_info(lod_comp, desc); + lod_adjust_stripe_info(lod_comp, desc, 0); if (v1->lmm_magic == LOV_USER_MAGIC_V3) { v3 = (struct lov_user_md_v3 *) v1; @@ -2793,6 +2913,8 @@ static int lod_declare_layout_set(const struct lu_env *env, for (i = 0; i < comp_v1->lcm_entry_count; i++) { __u32 id = comp_v1->lcm_entries[i].lcme_id; __u32 flags = comp_v1->lcm_entries[i].lcme_flags; + __u32 mirror_flag = flags & LCME_MIRROR_FLAGS; + bool neg = flags & LCME_FL_NEG; if (flags & LCME_FL_INIT) { if (changed) @@ -2800,16 +2922,30 @@ static int lod_declare_layout_set(const struct lu_env *env, RETURN(-EINVAL); } + flags &= ~(LCME_MIRROR_FLAGS | LCME_FL_NEG); for (j = 0; j < lo->ldo_comp_cnt; j++) { lod_comp = &lo->ldo_comp_entries[j]; - if (id != lod_comp->llc_id) + + /* lfs only put one flag in each entry */ + if ((flags && id != lod_comp->llc_id) || + (mirror_flag && mirror_id_of(id) != + mirror_id_of(lod_comp->llc_id))) continue; - if (flags & LCME_FL_NEG) { - flags &= ~LCME_FL_NEG; - lod_comp->llc_flags &= ~flags; + if (neg) { + if (flags) + lod_comp->llc_flags &= ~flags; + if (mirror_flag) + lod_comp->llc_flags &= ~mirror_flag; } else { - lod_comp->llc_flags |= flags; + if (flags) + lod_comp->llc_flags |= flags; + if (mirror_flag) { + lod_comp->llc_flags |= mirror_flag; + if (mirror_flag & LCME_FL_NOSYNC) + lod_comp->llc_timestamp = + ktime_get_real_seconds(); + } } changed = true; } @@ -3104,6 +3240,7 @@ static int lod_declare_layout_merge(const struct lu_env *env, struct lov_comp_md_v1 *cur_lcm; struct lov_comp_md_v1 *merge_lcm; struct lov_comp_md_entry_v1 *lcme; + struct lov_mds_md_v1 *lmm; size_t size = 0; size_t offset; __u16 cur_entry_count; @@ -3112,6 +3249,8 @@ static int lod_declare_layout_merge(const struct lu_env *env, __u16 mirror_id = 0; __u32 mirror_count; int rc, i; + bool merge_has_dom; + ENTRY; merge_lcm = mbuf->lb_buf; @@ -3140,6 +3279,7 @@ static int lod_declare_layout_merge(const struct lu_env *env, rc = lod_layout_convert(info); break; case LOV_MAGIC_COMP_V1: + case LOV_MAGIC_SEL: rc = 0; break; default: @@ -3198,6 +3338,13 @@ static int lod_declare_layout_merge(const struct lu_env *env, } mirror_id = mirror_id_of(id) + 1; + + /* check if first entry in new layout is DOM */ + lmm = (struct lov_mds_md_v1 *)((char *)merge_lcm + + merge_lcm->lcm_entries[0].lcme_offset); + merge_has_dom = lov_pattern(le32_to_cpu(lmm->lmm_pattern)) == + LOV_PATTERN_MDT; + for (i = 0; i < merge_entry_count; i++) { struct lov_comp_md_entry_v1 *merge_lcme; @@ -3206,6 +3353,8 @@ static int lod_declare_layout_merge(const struct lu_env *env, *lcme = *merge_lcme; lcme->lcme_offset = cpu_to_le32(offset); + if (merge_has_dom && i == 0) + lcme->lcme_flags |= cpu_to_le32(LCME_FL_STALE); id = pflr_id(mirror_id, i + 1); lcme->lcme_id = cpu_to_le32(id); @@ -3345,7 +3494,7 @@ static int lod_declare_xattr_set(const struct lu_env *env, } else if (S_ISDIR(mode)) { rc = lod_dir_declare_xattr_set(env, dt, buf, name, fl, th); } else if (strcmp(name, XATTR_NAME_FID) == 0) { - rc = lod_replace_parent_fid(env, dt, th, true); + rc = lod_replace_parent_fid(env, dt, buf, th, true); } else { rc = lod_sub_declare_xattr_set(env, next, buf, name, fl, th); } @@ -3392,7 +3541,11 @@ static int lod_xattr_set_internal(const struct lu_env *env, RETURN(0); for (i = 0; i < lo->ldo_dir_stripe_count; i++) { - LASSERT(lo->ldo_stripe[i]); + if (!lo->ldo_stripe[i]) + continue; + + if (!dt_object_exists(lo->ldo_stripe[i])) + continue; rc = lod_sub_xattr_set(env, lo->ldo_stripe[i], buf, name, fl, th); @@ -3502,8 +3655,23 @@ static int lod_xattr_set_lov_on_dir(const struct lu_env *env, pool_name); break; case LOV_USER_MAGIC_COMP_V1: + { + struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lum; + struct lov_comp_md_entry_v1 *lcme; + int i, comp_cnt; + + comp_cnt = le16_to_cpu(lcm->lcm_entry_count); + for (i = 0; i < comp_cnt; i++) { + lcme = &lcm->lcm_entries[i]; + if (lcme->lcme_flags & cpu_to_le32(LCME_FL_EXTENSION)) { + lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_SEL); + break; + } + } + is_del = false; break; + } default: CERROR("Invalid magic %x\n", lum->lmm_magic); RETURN(-EINVAL); @@ -3544,20 +3712,24 @@ static int lod_xattr_set_default_lmv_on_dir(const struct lu_env *env, const char *name, int fl, struct thandle *th) { - struct lmv_user_md_v1 *lum; - int rc; + struct lmv_user_md_v1 *lum; + int rc; + ENTRY; LASSERT(buf != NULL && buf->lb_buf != NULL); lum = buf->lb_buf; - CDEBUG(D_OTHER, "set default stripe_count # %u stripe_offset %d\n", + CDEBUG(D_INFO, + "set default stripe_count # %u stripe_offset %d hash %u\n", le32_to_cpu(lum->lum_stripe_count), - (int)le32_to_cpu(lum->lum_stripe_offset)); + (int)le32_to_cpu(lum->lum_stripe_offset), + le32_to_cpu(lum->lum_hash_type)); if (LMVEA_DELETE_VALUES((le32_to_cpu(lum->lum_stripe_count)), le32_to_cpu(lum->lum_stripe_offset)) && - le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC) { + le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC && + !(le32_to_cpu(lum->lum_hash_type) & LMV_HASH_FLAG_SPACE)) { rc = lod_xattr_del_internal(env, dt, name, th); if (rc == -ENODATA) rc = 0; @@ -3611,8 +3783,15 @@ static int lod_xattr_set_lmv(const struct lu_env *env, struct dt_object *dt, /* The stripes are supposed to be allocated in declare phase, * if there are no stripes being allocated, it will skip */ - if (lo->ldo_dir_stripe_count == 0) + if (lo->ldo_dir_stripe_count == 0) { + if (lo->ldo_dir_is_foreign) { + rc = lod_sub_xattr_set(env, dt_object_child(dt), buf, + XATTR_NAME_LMV, fl, th); + if (rc != 0) + RETURN(rc); + } RETURN(0); + } rc = dt_attr_get(env, dt_object_child(dt), attr); if (rc != 0) @@ -3643,10 +3822,18 @@ static int lod_xattr_set_lmv(const struct lu_env *env, struct dt_object *dt, struct linkea_data ldata = { NULL }; struct lu_buf linkea_buf; + /* OBD_FAIL_MDS_STRIPE_FID may leave stripe uninitialized */ + if (!dto) + continue; + + /* fail a remote stripe creation */ + if (i && OBD_FAIL_CHECK(OBD_FAIL_MDS_STRIPE_CREATE)) + continue; + /* if it's source stripe of migrating directory, don't create */ if (!((lo->ldo_dir_hash_type & LMV_HASH_FLAG_MIGRATION) && i >= lo->ldo_dir_migrate_offset)) { - dt_write_lock(env, dto, MOR_TGT_CHILD); + dt_write_lock(env, dto, DT_TGT_CHILD); rc = lod_sub_create(env, dto, attr, NULL, dof, th); if (rc != 0) { dt_write_unlock(env, dto); @@ -3661,14 +3848,14 @@ static int lod_xattr_set_lmv(const struct lu_env *env, struct dt_object *dt, rec->rec_fid = lu_object_fid(&dto->do_lu); rc = lod_sub_insert(env, dto, (const struct dt_rec *)rec, - (const struct dt_key *)dot, th, 0); + (const struct dt_key *)dot, th); if (rc != 0) GOTO(out, rc); } rec->rec_fid = lu_object_fid(&dt->do_lu); rc = lod_sub_insert(env, dto, (struct dt_rec *)rec, - (const struct dt_key *)dotdot, th, 0); + (const struct dt_key *)dotdot, th); if (rc != 0) GOTO(out, rc); @@ -3712,7 +3899,7 @@ static int lod_xattr_set_lmv(const struct lu_env *env, struct dt_object *dt, rec->rec_fid = lu_object_fid(&dto->do_lu); rc = lod_sub_insert(env, dt_object_child(dt), (const struct dt_rec *)rec, - (const struct dt_key *)stripe_name, th, 0); + (const struct dt_key *)stripe_name, th); if (rc != 0) GOTO(out, rc); @@ -3804,12 +3991,34 @@ static int lod_dir_striping_create_internal(const struct lu_env *env, th); if (rc != 0) RETURN(rc); + } else { + /* foreign LMV EA case */ + if (lmu) { + struct lmv_foreign_md *lfm = lmu->lb_buf; + + if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) { + rc = lod_declare_xattr_set_lmv(env, dt, attr, + lmu, dof, th); + } + } else { + if (lo->ldo_dir_is_foreign) { + LASSERT(lo->ldo_foreign_lmv != NULL && + lo->ldo_foreign_lmv_size > 0); + info->lti_buf.lb_buf = lo->ldo_foreign_lmv; + info->lti_buf.lb_len = lo->ldo_foreign_lmv_size; + lmu = &info->lti_buf; + rc = lod_xattr_set_lmv(env, dt, lmu, + XATTR_NAME_LMV, 0, th); + } + } } /* Transfer default LMV striping from the parent */ if (lds != NULL && lds->lds_dir_def_striping_set && - !LMVEA_DELETE_VALUES(lds->lds_dir_def_stripe_count, - lds->lds_dir_def_stripe_offset)) { + !(LMVEA_DELETE_VALUES(lds->lds_dir_def_stripe_count, + lds->lds_dir_def_stripe_offset) && + le32_to_cpu(lds->lds_dir_def_hash_type) != + LMV_HASH_TYPE_UNKNOWN)) { struct lmv_user_md_v1 *v1 = info->lti_ea_store; if (info->lti_ea_store_size < sizeof(*v1)) { @@ -3925,7 +4134,7 @@ static int lod_generate_and_set_lovea(const struct lu_env *env, LASSERT(lo); - if (lo->ldo_comp_cnt == 0) { + if (lo->ldo_comp_cnt == 0 && !lo->ldo_is_foreign) { lod_striping_free(env, lo); rc = lod_sub_xattr_del(env, next, XATTR_NAME_LOV, th); RETURN(rc); @@ -3950,129 +4159,328 @@ static int lod_generate_and_set_lovea(const struct lu_env *env, RETURN(rc); } +static __u32 lod_gen_component_id(struct lod_object *lo, + int mirror_id, int comp_idx); + /** - * Delete layout component(s) + * Repeat an existing component + * + * Creates a new layout by replicating an existing component. Uses striping + * policy from previous component as a template for the striping for the new + * new component. + * + * New component starts with zero length, will be extended (or removed) before + * returning layout to client. + * + * NB: Reallocates layout components array (lo->ldo_comp_entries), invalidating + * any pre-existing pointers to components. Handle with care. * * \param[in] env execution environment for this thread - * \param[in] dt object - * \param[in] th transaction handle + * \param[in,out] lo object to update the layout of + * \param[in] index index of component to copy * * \retval 0 on success - * \retval negative error number on failure + * \retval negative errno on error */ -static int lod_layout_del(const struct lu_env *env, struct dt_object *dt, - struct thandle *th) +static int lod_layout_repeat_comp(const struct lu_env *env, + struct lod_object *lo, int index) { - struct lod_layout_component *lod_comp; - struct lod_object *lo = lod_dt_obj(dt); - struct dt_object *next = dt_object_child(dt); - struct lu_attr *attr = &lod_env_info(env)->lti_attr; - int rc, i, j, left; - - LASSERT(lo->ldo_is_composite); - LASSERT(lo->ldo_comp_cnt > 0 && lo->ldo_comp_entries != NULL); - - left = lo->ldo_comp_cnt; - for (i = (lo->ldo_comp_cnt - 1); i >= 0; i--) { - lod_comp = &lo->ldo_comp_entries[i]; - - if (lod_comp->llc_id != LCME_ID_INVAL) - break; - left--; + struct lod_layout_component *lod_comp; + struct lod_layout_component *new_comp = NULL; + struct lod_layout_component *comp_array; + int rc = 0, i, new_cnt = lo->ldo_comp_cnt + 1; + __u16 mirror_id; + int offset = 0; + ENTRY; - /* Not instantiated component */ - if (lod_comp->llc_stripe == NULL) - continue; + lod_comp = &lo->ldo_comp_entries[index]; + LASSERT(lod_comp_inited(lod_comp) && lod_comp->llc_id != LCME_ID_INVAL); - LASSERT(lod_comp->llc_stripe_count > 0); - for (j = 0; j < lod_comp->llc_stripe_count; j++) { - struct dt_object *obj = lod_comp->llc_stripe[j]; + CDEBUG(D_LAYOUT, "repeating component %d\n", index); - if (obj == NULL) - continue; - rc = lod_sub_destroy(env, obj, th); - if (rc) - GOTO(out, rc); + OBD_ALLOC(comp_array, sizeof(*comp_array) * new_cnt); + if (comp_array == NULL) + GOTO(out, rc = -ENOMEM); - lu_object_put(env, &obj->do_lu); - lod_comp->llc_stripe[j] = NULL; - } - OBD_FREE(lod_comp->llc_stripe, sizeof(struct dt_object *) * - lod_comp->llc_stripes_allocated); - lod_comp->llc_stripe = NULL; - OBD_FREE(lod_comp->llc_ost_indices, - sizeof(__u32) * lod_comp->llc_stripes_allocated); - lod_comp->llc_ost_indices = NULL; - lod_comp->llc_stripes_allocated = 0; - lod_obj_set_pool(lo, i, NULL); - if (lod_comp->llc_ostlist.op_array) { - OBD_FREE(lod_comp->llc_ostlist.op_array, - lod_comp->llc_ostlist.op_size); - lod_comp->llc_ostlist.op_array = NULL; - lod_comp->llc_ostlist.op_size = 0; + for (i = 0; i < lo->ldo_comp_cnt; i++) { + memcpy(&comp_array[i + offset], &lo->ldo_comp_entries[i], + sizeof(*comp_array)); + + /* Duplicate this component in to the next slot */ + if (i == index) { + new_comp = &comp_array[i + 1]; + memcpy(&comp_array[i + 1], &lo->ldo_comp_entries[i], + sizeof(*comp_array)); + /* We must now skip this new component when copying */ + offset = 1; } } - LASSERTF(left >= 0 && left < lo->ldo_comp_cnt, "left = %d\n", left); - if (left > 0) { - struct lod_layout_component *comp_array; - - OBD_ALLOC(comp_array, sizeof(*comp_array) * left); - if (comp_array == NULL) - GOTO(out, rc = -ENOMEM); - - memcpy(&comp_array[0], &lo->ldo_comp_entries[0], - sizeof(*comp_array) * left); + /* Set up copied component */ + new_comp->llc_flags &= ~LCME_FL_INIT; + new_comp->llc_stripe = NULL; + new_comp->llc_stripes_allocated = 0; + new_comp->llc_stripe_offset = LOV_OFFSET_DEFAULT; + /* for uninstantiated components, layout gen stores default stripe + * offset */ + new_comp->llc_layout_gen = lod_comp->llc_stripe_offset; + /* This makes the repeated component zero-length, placed at the end of + * the preceding component */ + new_comp->llc_extent.e_start = new_comp->llc_extent.e_end; + new_comp->llc_timestamp = lod_comp->llc_timestamp; + new_comp->llc_pool = NULL; + + rc = lod_set_pool(&new_comp->llc_pool, lod_comp->llc_pool); + if (rc) + GOTO(out, rc); - OBD_FREE(lo->ldo_comp_entries, - sizeof(*comp_array) * lo->ldo_comp_cnt); - lo->ldo_comp_entries = comp_array; - lo->ldo_comp_cnt = left; + if (new_comp->llc_ostlist.op_array) { + __u32 *op_array = NULL; - LASSERT(lo->ldo_mirror_count == 1); - lo->ldo_mirrors[0].lme_end = left - 1; - lod_obj_inc_layout_gen(lo); - } else { - lod_free_comp_entries(lo); + OBD_ALLOC(op_array, new_comp->llc_ostlist.op_size); + if (!op_array) + GOTO(out, rc = -ENOMEM); + memcpy(op_array, &new_comp->llc_ostlist.op_array, + new_comp->llc_ostlist.op_size); + new_comp->llc_ostlist.op_array = op_array; } - LASSERT(dt_object_exists(dt)); - rc = dt_attr_get(env, next, attr); - if (rc) - GOTO(out, rc); + OBD_FREE(lo->ldo_comp_entries, + sizeof(*comp_array) * lo->ldo_comp_cnt); + lo->ldo_comp_entries = comp_array; + lo->ldo_comp_cnt = new_cnt; - if (attr->la_size > 0) { - attr->la_size = 0; - attr->la_valid = LA_SIZE; - rc = lod_sub_attr_set(env, next, attr, th); - if (rc) - GOTO(out, rc); - } + /* Generate an id for the new component */ + mirror_id = mirror_id_of(new_comp->llc_id); + new_comp->llc_id = LCME_ID_INVAL; + new_comp->llc_id = lod_gen_component_id(lo, mirror_id, index + 1); + if (new_comp->llc_id == LCME_ID_INVAL) + GOTO(out, rc = -ERANGE); - rc = lod_generate_and_set_lovea(env, lo, th); EXIT; out: if (rc) - lod_striping_free(env, lo); + OBD_FREE(comp_array, sizeof(*comp_array) * new_cnt); + return rc; } +static int lod_layout_data_init(struct lod_thread_info *info, __u32 comp_cnt) +{ + ENTRY; -static int lod_get_default_lov_striping(const struct lu_env *env, - struct lod_object *lo, - struct lod_default_striping *lds); -/** - * Implementation of dt_object_operations::do_xattr_set. - * - * Sets specified extended attribute on the object. Three types of EAs are - * special: - * LOV EA - stores striping for a regular file or default striping (when set - * on a directory) - * LMV EA - stores a marker for the striped directories - * DMV EA - stores default directory striping - * - * When striping is applied to a non-striped existing object (this is called - * late striping), then LOD notices the caller wants to turn the object into a + /* clear memory region that will be used for layout change */ + memset(&info->lti_layout_attr, 0, sizeof(struct lu_attr)); + info->lti_count = 0; + + if (info->lti_comp_size >= comp_cnt) + RETURN(0); + + if (info->lti_comp_size > 0) { + OBD_FREE(info->lti_comp_idx, + info->lti_comp_size * sizeof(__u32)); + info->lti_comp_size = 0; + } + + OBD_ALLOC(info->lti_comp_idx, comp_cnt * sizeof(__u32)); + if (!info->lti_comp_idx) + RETURN(-ENOMEM); + + info->lti_comp_size = comp_cnt; + RETURN(0); +} + +/** + * Prepare new layout minus deleted components + * + * Removes components marked for deletion (LCME_ID_INVAL) by copying to a new + * layout and skipping those components. Removes stripe objects if any exist. + * + * NB: + * Reallocates layout components array (lo->ldo_comp_entries), invalidating + * any pre-existing pointers to components. + * + * Caller is responsible for updating mirror end (ldo_mirror[].lme_end). + * + * \param[in] env execution environment for this thread + * \param[in,out] lo object to update the layout of + * \param[in] th transaction handle for this operation + * + * \retval # of components deleted + * \retval negative errno on error + */ +static int lod_layout_del_prep_layout(const struct lu_env *env, + struct lod_object *lo, + struct thandle *th) +{ + struct lod_layout_component *lod_comp; + struct lod_thread_info *info = lod_env_info(env); + int rc = 0, i, j, deleted = 0; + + ENTRY; + + LASSERT(lo->ldo_is_composite); + LASSERT(lo->ldo_comp_cnt > 0 && lo->ldo_comp_entries != NULL); + + rc = lod_layout_data_init(info, lo->ldo_comp_cnt); + if (rc) + RETURN(rc); + + for (i = 0; i < lo->ldo_comp_cnt; i++) { + lod_comp = &lo->ldo_comp_entries[i]; + + if (lod_comp->llc_id != LCME_ID_INVAL) { + /* Build array of things to keep */ + info->lti_comp_idx[info->lti_count++] = i; + continue; + } + + lod_obj_set_pool(lo, i, NULL); + if (lod_comp->llc_ostlist.op_array) { + OBD_FREE(lod_comp->llc_ostlist.op_array, + lod_comp->llc_ostlist.op_size); + lod_comp->llc_ostlist.op_array = NULL; + lod_comp->llc_ostlist.op_size = 0; + } + + deleted++; + CDEBUG(D_LAYOUT, "deleting comp %d, left %d\n", i, + lo->ldo_comp_cnt - deleted); + + /* No striping info for this component */ + if (lod_comp->llc_stripe == NULL) + continue; + + LASSERT(lod_comp->llc_stripe_count > 0); + for (j = 0; j < lod_comp->llc_stripe_count; j++) { + struct dt_object *obj = lod_comp->llc_stripe[j]; + + if (obj == NULL) + continue; + + /* components which are not init have no sub objects + * to destroy */ + if (lod_comp_inited(lod_comp)) { + rc = lod_sub_destroy(env, obj, th); + if (rc) + GOTO(out, rc); + } + + lu_object_put(env, &obj->do_lu); + lod_comp->llc_stripe[j] = NULL; + } + OBD_FREE(lod_comp->llc_stripe, sizeof(struct dt_object *) * + lod_comp->llc_stripes_allocated); + lod_comp->llc_stripe = NULL; + OBD_FREE(lod_comp->llc_ost_indices, + sizeof(__u32) * lod_comp->llc_stripes_allocated); + lod_comp->llc_ost_indices = NULL; + lod_comp->llc_stripes_allocated = 0; + } + + /* info->lti_count has the amount of left components */ + LASSERTF(info->lti_count >= 0 && info->lti_count < lo->ldo_comp_cnt, + "left = %d, lo->ldo_comp_cnt %d\n", (int)info->lti_count, + (int)lo->ldo_comp_cnt); + + if (info->lti_count > 0) { + struct lod_layout_component *comp_array; + + OBD_ALLOC(comp_array, sizeof(*comp_array) * info->lti_count); + if (comp_array == NULL) + GOTO(out, rc = -ENOMEM); + + for (i = 0; i < info->lti_count; i++) { + memcpy(&comp_array[i], + &lo->ldo_comp_entries[info->lti_comp_idx[i]], + sizeof(*comp_array)); + } + + OBD_FREE(lo->ldo_comp_entries, + sizeof(*comp_array) * lo->ldo_comp_cnt); + lo->ldo_comp_entries = comp_array; + lo->ldo_comp_cnt = info->lti_count; + } else { + lod_free_comp_entries(lo); + } + + EXIT; +out: + return rc ? rc : deleted; +} + +/** + * Delete layout component(s) + * + * This function sets up the layout data in the env and does the setattrs + * required to write out the new layout. The layout itself is modified in + * lod_layout_del_prep_layout. + * + * \param[in] env execution environment for this thread + * \param[in] dt object + * \param[in] th transaction handle + * + * \retval 0 on success + * \retval negative error number on failure + */ +static int lod_layout_del(const struct lu_env *env, struct dt_object *dt, + struct thandle *th) +{ + struct lod_object *lo = lod_dt_obj(dt); + struct dt_object *next = dt_object_child(dt); + struct lu_attr *attr = &lod_env_info(env)->lti_attr; + int rc; + + LASSERT(lo->ldo_mirror_count == 1); + + rc = lod_layout_del_prep_layout(env, lo, th); + if (rc < 0) + GOTO(out, rc); + + /* Only do this if we didn't delete all components */ + if (lo->ldo_comp_cnt > 0) { + lo->ldo_mirrors[0].lme_end = lo->ldo_comp_cnt - 1; + lod_obj_inc_layout_gen(lo); + } + + LASSERT(dt_object_exists(dt)); + rc = dt_attr_get(env, next, attr); + if (rc) + GOTO(out, rc); + + if (attr->la_size > 0) { + attr->la_size = 0; + attr->la_valid = LA_SIZE; + rc = lod_sub_attr_set(env, next, attr, th); + if (rc) + GOTO(out, rc); + } + + rc = lod_generate_and_set_lovea(env, lo, th); + EXIT; +out: + if (rc) + lod_striping_free(env, lo); + return rc; +} + + +static int lod_get_default_lov_striping(const struct lu_env *env, + struct lod_object *lo, + struct lod_default_striping *lds, + struct dt_allocation_hint *ah); +/** + * Implementation of dt_object_operations::do_xattr_set. + * + * Sets specified extended attribute on the object. Three types of EAs are + * special: + * LOV EA - stores striping for a regular file or default striping (when set + * on a directory) + * LMV EA - stores a marker for the striped directories + * DMV EA - stores default directory striping + * + * When striping is applied to a non-striped existing object (this is called + * late striping), then LOD notices the caller wants to turn the object into a * striped one. The stripe objects are created and appropriate EA is set: * LOV EA storing all the stripes directly or LMV EA storing just a small header * with striping configuration. @@ -4110,14 +4518,14 @@ static int lod_xattr_set(const struct lu_env *env, RETURN(rc); } else if (S_ISDIR(dt->do_lu.lo_header->loh_attr) && strcmp(name, XATTR_NAME_LOV) == 0) { - struct lod_thread_info *info = lod_env_info(env); - struct lod_default_striping *lds = &info->lti_def_striping; + struct lod_default_striping *lds = lod_lds_buf_get(env); struct lov_user_md_v1 *v1 = buf->lb_buf; char pool[LOV_MAXPOOLNAME + 1]; bool is_del; /* get existing striping config */ - rc = lod_get_default_lov_striping(env, lod_dt_obj(dt), lds); + rc = lod_get_default_lov_striping(env, lod_dt_obj(dt), lds, + NULL); if (rc) RETURN(rc); @@ -4214,7 +4622,7 @@ static int lod_xattr_set(const struct lu_env *env, } RETURN(rc); } else if (strcmp(name, XATTR_NAME_FID) == 0) { - rc = lod_replace_parent_fid(env, dt, th, false); + rc = lod_replace_parent_fid(env, dt, buf, th, false); RETURN(rc); } @@ -4248,7 +4656,13 @@ static int lod_declare_xattr_del(const struct lu_env *env, if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) RETURN(0); - /* set xattr to each stripes, if needed */ + /* NB: don't delete stripe LMV, because when we do this, normally we + * will remove stripes, besides, if directory LMV is corrupt, this will + * prevent deleting its LMV and fixing it (via LFSCK). + */ + if (!strcmp(name, XATTR_NAME_LMV)) + RETURN(0); + rc = lod_striping_load(env, lo); if (rc != 0) RETURN(rc); @@ -4259,7 +4673,9 @@ static int lod_declare_xattr_del(const struct lu_env *env, for (i = 0; i < lo->ldo_dir_stripe_count; i++) { struct dt_object *dto = lo->ldo_stripe[i]; - LASSERT(dto); + if (!dto) + continue; + rc = lod_sub_declare_xattr_del(env, dto, name, th); if (rc != 0) break; @@ -4292,13 +4708,17 @@ static int lod_xattr_del(const struct lu_env *env, struct dt_object *dt, if (rc != 0 || !S_ISDIR(dt->do_lu.lo_header->loh_attr)) RETURN(rc); + if (!strcmp(name, XATTR_NAME_LMV)) + RETURN(0); + if (lo->ldo_dir_stripe_count == 0) RETURN(0); for (i = 0; i < lo->ldo_dir_stripe_count; i++) { struct dt_object *dto = lo->ldo_stripe[i]; - LASSERT(dto); + if (!dto) + continue; rc = lod_sub_xattr_del(env, dto, name, th); if (rc != 0) @@ -4384,19 +4804,19 @@ skip: */ static int lod_get_default_lov_striping(const struct lu_env *env, struct lod_object *lo, - struct lod_default_striping *lds) + struct lod_default_striping *lds, + struct dt_allocation_hint *ah) { struct lod_thread_info *info = lod_env_info(env); struct lov_user_md_v1 *v1 = NULL; struct lov_user_md_v3 *v3 = NULL; struct lov_comp_md_v1 *comp_v1 = NULL; - __u16 comp_cnt; - __u16 mirror_cnt; - bool composite; - int rc, i, j; - ENTRY; + __u16 comp_cnt; + __u16 mirror_cnt; + bool composite; + int rc, i, j; - lds->lds_def_striping_set = 0; + ENTRY; rc = lod_get_lov_ea(env, lo); if (rc < 0) @@ -4416,17 +4836,21 @@ static int lod_get_default_lov_striping(const struct lu_env *env, lustre_swab_lov_user_md_v3(v3); lustre_swab_lov_user_md_objects(v3->lmm_objects, v3->lmm_stripe_count); - } else if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_COMP_V1)) { + } else if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_COMP_V1) || + v1->lmm_magic == __swab32(LOV_USER_MAGIC_SEL)) { comp_v1 = (struct lov_comp_md_v1 *)v1; lustre_swab_lov_comp_md_v1(comp_v1); } if (v1->lmm_magic != LOV_MAGIC_V3 && v1->lmm_magic != LOV_MAGIC_V1 && v1->lmm_magic != LOV_MAGIC_COMP_V1 && + v1->lmm_magic != LOV_MAGIC_SEL && v1->lmm_magic != LOV_USER_MAGIC_SPECIFIC) RETURN(-ENOTSUPP); - if (v1->lmm_magic == LOV_MAGIC_COMP_V1) { + if ((v1->lmm_magic == LOV_MAGIC_COMP_V1 || + v1->lmm_magic == LOV_MAGIC_SEL) && + !(ah && ah->dah_append_stripes)) { comp_v1 = (struct lov_comp_md_v1 *)v1; comp_cnt = comp_v1->lcm_entry_count; if (comp_cnt == 0) @@ -4450,7 +4874,6 @@ static int lod_get_default_lov_striping(const struct lu_env *env, for (i = 0; i < comp_cnt; i++) { struct lod_layout_component *lod_comp; - struct lu_extent *ext; char *pool; lod_comp = &lds->lds_def_comp_entries[i]; @@ -4464,30 +4887,38 @@ static int lod_get_default_lov_striping(const struct lu_env *env, if (composite) { v1 = (struct lov_user_md *)((char *)comp_v1 + comp_v1->lcm_entries[i].lcme_offset); - ext = &comp_v1->lcm_entries[i].lcme_extent; - lod_comp->llc_extent = *ext; + lod_comp->llc_extent = + comp_v1->lcm_entries[i].lcme_extent; + /* We only inherit certain flags from the layout */ + lod_comp->llc_flags = + comp_v1->lcm_entries[i].lcme_flags & + LCME_TEMPLATE_FLAGS; } - if (v1->lmm_pattern != LOV_PATTERN_RAID0 && - v1->lmm_pattern != LOV_PATTERN_MDT && - v1->lmm_pattern != 0) { + if (!lov_pattern_supported(v1->lmm_pattern) && + !(v1->lmm_pattern & LOV_PATTERN_F_RELEASED)) { lod_free_def_comp_entries(lds); RETURN(-EINVAL); } - CDEBUG(D_LAYOUT, DFID" stripe_count=%d stripe_size=%d " - "stripe_offset=%d\n", + CDEBUG(D_LAYOUT, DFID" stripe_count=%d stripe_size=%d stripe_offset=%d append_stripes=%d\n", PFID(lu_object_fid(&lo->ldo_obj.do_lu)), (int)v1->lmm_stripe_count, (int)v1->lmm_stripe_size, - (int)v1->lmm_stripe_offset); + (int)v1->lmm_stripe_offset, + ah ? ah->dah_append_stripes : 0); - lod_comp->llc_stripe_count = v1->lmm_stripe_count; + if (ah && ah->dah_append_stripes) + lod_comp->llc_stripe_count = ah->dah_append_stripes; + else + lod_comp->llc_stripe_count = v1->lmm_stripe_count; lod_comp->llc_stripe_size = v1->lmm_stripe_size; lod_comp->llc_stripe_offset = v1->lmm_stripe_offset; lod_comp->llc_pattern = v1->lmm_pattern; pool = NULL; - if (v1->lmm_magic == LOV_USER_MAGIC_V3) { + if (ah && ah->dah_append_pool && ah->dah_append_pool[0]) { + pool = ah->dah_append_pool; + } else if (v1->lmm_magic == LOV_USER_MAGIC_V3) { /* XXX: sanity check here */ v3 = (struct lov_user_md_v3 *) v1; if (v3->lmm_pool_name[0] != '\0') @@ -4525,27 +4956,30 @@ static int lod_get_default_lmv_striping(const struct lu_env *env, struct lod_object *lo, struct lod_default_striping *lds) { - struct lod_thread_info *info = lod_env_info(env); - struct lmv_user_md_v1 *v1 = NULL; - int rc; - ENTRY; + struct lmv_user_md *lmu; + int rc; lds->lds_dir_def_striping_set = 0; + rc = lod_get_default_lmv_ea(env, lo); if (rc < 0) - RETURN(rc); + return rc; - if (rc < (typeof(rc))sizeof(struct lmv_user_md)) - RETURN(0); + if (rc >= (int)sizeof(*lmu)) { + struct lod_thread_info *info = lod_env_info(env); - v1 = info->lti_ea_store; + lmu = info->lti_ea_store; - lds->lds_dir_def_stripe_count = le32_to_cpu(v1->lum_stripe_count); - lds->lds_dir_def_stripe_offset = le32_to_cpu(v1->lum_stripe_offset); - lds->lds_dir_def_hash_type = le32_to_cpu(v1->lum_hash_type); - lds->lds_dir_def_striping_set = 1; + lds->lds_dir_def_stripe_count = + le32_to_cpu(lmu->lum_stripe_count); + lds->lds_dir_def_stripe_offset = + le32_to_cpu(lmu->lum_stripe_offset); + lds->lds_dir_def_hash_type = + le32_to_cpu(lmu->lum_hash_type); + lds->lds_dir_def_striping_set = 1; + } - RETURN(0); + return 0; } /** @@ -4566,7 +5000,7 @@ static int lod_get_default_striping(const struct lu_env *env, { int rc, rc1; - rc = lod_get_default_lov_striping(env, lo, lds); + rc = lod_get_default_lov_striping(env, lo, lds, NULL); rc1 = lod_get_default_lmv_striping(env, lo, lds); if (rc == 0 && rc1 < 0) rc = rc1; @@ -4608,8 +5042,9 @@ static void lod_striping_from_default(struct lod_object *lo, struct lod_layout_component *def_comp = &lds->lds_def_comp_entries[i]; - CDEBUG(D_LAYOUT, "Inherite from default: size:%hu " - "nr:%u offset:%u pattern %#x %s\n", + CDEBUG(D_LAYOUT, "Inherit from default: flags=%#x " + "size=%hu nr=%u offset=%u pattern=%#x pool=%s\n", + def_comp->llc_flags, def_comp->llc_stripe_size, def_comp->llc_stripe_count, def_comp->llc_stripe_offset, @@ -4648,7 +5083,7 @@ static void lod_striping_from_default(struct lod_object *lo, if (!lo->ldo_is_composite) continue; - lod_adjust_stripe_info(obj_comp, desc); + lod_adjust_stripe_info(obj_comp, desc, 0); } } else if (lds->lds_dir_def_striping_set && S_ISDIR(mode)) { if (lo->ldo_dir_stripe_count == 0) @@ -4658,7 +5093,8 @@ static void lod_striping_from_default(struct lod_object *lo, lo->ldo_dir_stripe_offset = lds->lds_dir_def_stripe_offset; if (lo->ldo_dir_hash_type == 0) - lo->ldo_dir_hash_type = lds->lds_dir_def_hash_type; + lo->ldo_dir_hash_type = lds->lds_dir_def_hash_type & + ~LMV_HASH_FLAG_SPACE; CDEBUG(D_LAYOUT, "striping from default dir: count:%hu, " "offset:%u, hash_type:%u\n", @@ -4667,7 +5103,8 @@ static void lod_striping_from_default(struct lod_object *lo, } } -static inline bool lod_need_inherit_more(struct lod_object *lo, bool from_root) +static inline bool lod_need_inherit_more(struct lod_object *lo, bool from_root, + char *append_pool) { struct lod_layout_component *lod_comp; @@ -4687,6 +5124,9 @@ static inline bool lod_need_inherit_more(struct lod_object *lo, bool from_root) lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT)) return true; + if (append_pool && append_pool[0]) + return true; + return false; } @@ -4709,7 +5149,7 @@ static void lod_ah_init(const struct lu_env *env, { struct lod_device *d = lu2lod_dev(child->do_lu.lo_dev); struct lod_thread_info *info = lod_env_info(env); - struct lod_default_striping *lds = &info->lti_def_striping; + struct lod_default_striping *lds = lod_lds_buf_get(env); struct dt_object *nextp = NULL; struct dt_object *nextc; struct lod_object *lp = NULL; @@ -4721,6 +5161,9 @@ static void lod_ah_init(const struct lu_env *env, LASSERT(child); + if (ah->dah_append_stripes == -1) + ah->dah_append_stripes = d->lod_desc.ld_tgt_count; + if (likely(parent)) { nextp = dt_object_child(parent); lp = lod_dt_obj(parent); @@ -4744,8 +5187,22 @@ static void lod_ah_init(const struct lu_env *env, /* other default values are 0 */ lc->ldo_dir_stripe_offset = -1; - /* get default striping from parent object */ - if (likely(lp != NULL)) + /* no default striping configuration is needed for + * foreign dirs + */ + if (ah->dah_eadata != NULL && ah->dah_eadata_len != 0 && + le32_to_cpu(lum1->lum_magic) == LMV_MAGIC_FOREIGN) { + lc->ldo_dir_is_foreign = true; + /* keep stripe_count 0 and stripe_offset -1 */ + CDEBUG(D_INFO, "no default striping for foreign dir\n"); + RETURN_EXIT; + } + + /* + * If parent object is not root directory, + * then get default striping from parent object. + */ + if (likely(lp != NULL) && !fid_is_root(lod_object_fid(lp))) lod_get_default_striping(env, lp, lds); /* set child default striping info, default value is NULL */ @@ -4808,7 +5265,7 @@ static void lod_ah_init(const struct lu_env *env, * Try from the parent first. */ if (likely(lp != NULL)) { - rc = lod_get_default_lov_striping(env, lp, lds); + rc = lod_get_default_lov_striping(env, lp, lds, ah); if (rc == 0) lod_striping_from_default(lc, lds, child_mode); } @@ -4836,8 +5293,10 @@ static void lod_ah_init(const struct lu_env *env, * - parent has plain(v1/v3) default layout, and some attributes * are not specified in the default layout; */ - if (d->lod_md_root != NULL && lod_need_inherit_more(lc, true)) { - rc = lod_get_default_lov_striping(env, d->lod_md_root, lds); + if (d->lod_md_root != NULL && + lod_need_inherit_more(lc, true, ah->dah_append_pool)) { + rc = lod_get_default_lov_striping(env, d->lod_md_root, lds, + ah); if (rc) goto out; if (lc->ldo_comp_cnt == 0) { @@ -4855,7 +5314,8 @@ static void lod_ah_init(const struct lu_env *env, if (lod_comp->llc_stripe_size <= 0) lod_comp->llc_stripe_size = def_comp->llc_stripe_size; - if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) + if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT && + (!lod_comp->llc_pool || !lod_comp->llc_pool[0])) lod_comp->llc_stripe_offset = def_comp->llc_stripe_offset; if (lod_comp->llc_pool == NULL) @@ -4867,7 +5327,7 @@ out: * fs default striping may not be explicitly set, or historically set * in config log, use them. */ - if (lod_need_inherit_more(lc, false)) { + if (lod_need_inherit_more(lc, false, ah->dah_append_pool)) { if (lc->ldo_comp_cnt == 0) { rc = lod_alloc_comp_entries(lc, 0, 1); if (rc) @@ -4881,7 +5341,9 @@ out: LASSERT(!lc->ldo_is_composite); lod_comp = &lc->ldo_comp_entries[0]; desc = &d->lod_desc; - lod_adjust_stripe_info(lod_comp, desc); + lod_adjust_stripe_info(lod_comp, desc, ah->dah_append_stripes); + if (ah->dah_append_pool && ah->dah_append_pool[0]) + lod_obj_set_pool(lc, 0, ah->dah_append_pool); } EXIT; @@ -5045,6 +5507,48 @@ out: RETURN(rc); } +/* + * Whether subdirectories under \a dt should be created on MDTs by space QoS + * + * If LMV_HASH_FLAG_SPACE is set on directory default layout, its subdirectories + * should be created on MDT by space QoS. + * + * \param[in] env execution environment + * \param[in] dev lu device + * \param[in] dt object + * + * \retval 1 if directory should create subdir by space usage + * \retval 0 if not + * \retval -ev if failed + */ +static inline int dt_object_qos_mkdir(const struct lu_env *env, + struct lu_device *dev, + struct dt_object *dt) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lu_object *obj; + struct lod_object *lo; + struct lmv_user_md *lmu; + int rc; + + obj = lu_object_find_slice(env, dev, lu_object_fid(&dt->do_lu), NULL); + if (IS_ERR(obj)) + return PTR_ERR(obj); + + lo = lu2lod_obj(obj); + + rc = lod_get_default_lmv_ea(env, lo); + dt_object_put(env, dt); + if (rc <= 0) + return rc; + + if (rc < (int)sizeof(*lmu)) + return -EINVAL; + + lmu = info->lti_ea_store; + return !!(le32_to_cpu(lmu->lum_hash_type) & LMV_HASH_FLAG_SPACE); +} + /** * Implementation of dt_object_operations::do_declare_create. * @@ -5109,22 +5613,28 @@ static int lod_declare_create(const struct lu_env *env, struct dt_object *dt, GOTO(out, rc = -EREMOTE); if (lo->ldo_dir_stripe_offset == -1) { - /* child and parent should be in the same MDT */ - if (hint->dah_parent != NULL && - dt_object_remote(hint->dah_parent)) - GOTO(out, rc = -EREMOTE); + /* + * child and parent should be in the same MDT, + * but if parent has plain layout, it's allowed. + */ + if (hint->dah_parent && + dt_object_remote(hint->dah_parent)) { + rc = dt_object_qos_mkdir(env, + lo->ldo_obj.do_lu.lo_dev, + hint->dah_parent); + if (rc <= 0) + GOTO(out, rc ? rc : -EREMOTE); + } } else if (lo->ldo_dir_stripe_offset != ss->ss_node_id) { struct lod_device *lod; struct lod_tgt_descs *ltd; struct lod_tgt_desc *tgt = NULL; bool found_mdt = false; - int i; lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); ltd = &lod->lod_mdt_descs; - cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) { - tgt = LTD_TGT(ltd, i); + ltd_foreach_tgt(ltd, tgt) { if (tgt->ltd_index == lo->ldo_dir_stripe_offset) { found_mdt = true; @@ -5234,7 +5744,8 @@ int lod_striped_create(const struct lu_env *env, struct dt_object *dt, int rc = 0, i, j; ENTRY; - LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL); + LASSERT((lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL) || + lo->ldo_is_foreign); mirror_id = 0; /* non-flr file's mirror_id is 0 */ if (lo->ldo_mirror_count > 1) { @@ -5303,6 +5814,23 @@ out: RETURN(rc); } +static inline bool lod_obj_is_dom(struct dt_object *dt) +{ + struct lod_object *lo = lod_dt_obj(dt); + + if (!dt_object_exists(dt_object_child(dt))) + return false; + + if (S_ISDIR(dt->do_lu.lo_header->loh_attr)) + return false; + + if (!lo->ldo_comp_cnt) + return false; + + return (lov_pattern(lo->ldo_comp_entries[0].llc_pattern) == + LOV_PATTERN_MDT); +} + /** * Implementation of dt_object_operations::do_create. * @@ -5325,7 +5853,8 @@ static int lod_create(const struct lu_env *env, struct dt_object *dt, RETURN(rc); if (S_ISREG(dt->do_lu.lo_header->loh_attr) && - lod_obj_is_striped(dt) && dof->u.dof_reg.striped != 0) { + (lod_obj_is_striped(dt) || lod_obj_is_dom(dt)) && + dof->u.dof_reg.striped != 0) { LASSERT(lod_dt_obj(dt)->ldo_comp_cached == 0); rc = lod_striped_create(env, dt, attr, dof, th); } @@ -5362,11 +5891,13 @@ lod_obj_stripe_destroy_cb(const struct lu_env *env, struct lod_object *lo, static int lod_declare_destroy(const struct lu_env *env, struct dt_object *dt, struct thandle *th) { - struct dt_object *next = dt_object_child(dt); - struct lod_object *lo = lod_dt_obj(dt); + struct dt_object *next = dt_object_child(dt); + struct lod_object *lo = lod_dt_obj(dt); struct lod_thread_info *info = lod_env_info(env); - char *stripe_name = info->lti_key; - int rc, i; + struct dt_object *stripe; + char *stripe_name = info->lti_key; + int rc, i; + ENTRY; /* @@ -5386,13 +5917,17 @@ static int lod_declare_destroy(const struct lu_env *env, struct dt_object *dt, RETURN(rc); for (i = 0; i < lo->ldo_dir_stripe_count; i++) { + stripe = lo->ldo_stripe[i]; + if (!stripe) + continue; + rc = lod_sub_declare_ref_del(env, next, th); if (rc != 0) RETURN(rc); - snprintf(stripe_name, sizeof(info->lti_key), DFID":%d", - PFID(lu_object_fid(&lo->ldo_stripe[i]->do_lu)), - i); + snprintf(stripe_name, sizeof(info->lti_key), + DFID":%d", + PFID(lu_object_fid(&stripe->do_lu)), i); rc = lod_sub_declare_delete(env, next, (const struct dt_key *)stripe_name, th); if (rc != 0) @@ -5417,14 +5952,18 @@ static int lod_declare_destroy(const struct lu_env *env, struct dt_object *dt, /* declare destroy all striped objects */ if (S_ISDIR(dt->do_lu.lo_header->loh_attr)) { for (i = 0; i < lo->ldo_dir_stripe_count; i++) { - if (lo->ldo_stripe[i] == NULL) + stripe = lo->ldo_stripe[i]; + if (!stripe) continue; - rc = lod_sub_declare_ref_del(env, lo->ldo_stripe[i], - th); + if (!dt_object_exists(stripe)) + continue; - rc = lod_sub_declare_destroy(env, lo->ldo_stripe[i], - th); + rc = lod_sub_declare_ref_del(env, stripe, th); + if (rc != 0) + break; + + rc = lod_sub_declare_destroy(env, stripe, th); if (rc != 0) break; } @@ -5454,11 +5993,13 @@ static int lod_destroy(const struct lu_env *env, struct dt_object *dt, struct dt_object *next = dt_object_child(dt); struct lod_object *lo = lod_dt_obj(dt); struct lod_thread_info *info = lod_env_info(env); - char *stripe_name = info->lti_key; - unsigned int i; - int rc; - ENTRY; - + char *stripe_name = info->lti_key; + struct dt_object *stripe; + unsigned int i; + int rc; + + ENTRY; + /* destroy sub-stripe of master object */ if (S_ISDIR(dt->do_lu.lo_header->loh_attr)) { rc = next->do_ops->do_index_try(env, next, @@ -5467,17 +6008,20 @@ static int lod_destroy(const struct lu_env *env, struct dt_object *dt, RETURN(rc); for (i = 0; i < lo->ldo_dir_stripe_count; i++) { + stripe = lo->ldo_stripe[i]; + if (!stripe) + continue; + rc = lod_sub_ref_del(env, next, th); if (rc != 0) RETURN(rc); snprintf(stripe_name, sizeof(info->lti_key), DFID":%d", - PFID(lu_object_fid(&lo->ldo_stripe[i]->do_lu)), - i); + PFID(lu_object_fid(&stripe->do_lu)), i); CDEBUG(D_INFO, DFID" delete stripe %s "DFID"\n", PFID(lu_object_fid(&dt->do_lu)), stripe_name, - PFID(lu_object_fid(&lo->ldo_stripe[i]->do_lu))); + PFID(lu_object_fid(&stripe->do_lu))); rc = lod_sub_delete(env, next, (const struct dt_key *)stripe_name, th); @@ -5500,20 +6044,22 @@ static int lod_destroy(const struct lu_env *env, struct dt_object *dt, /* destroy all striped objects */ if (S_ISDIR(dt->do_lu.lo_header->loh_attr)) { for (i = 0; i < lo->ldo_dir_stripe_count; i++) { - if (lo->ldo_stripe[i] == NULL) + stripe = lo->ldo_stripe[i]; + if (!stripe) + continue; + + if (!dt_object_exists(stripe)) continue; + if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_SPEOBJ) || i == cfs_fail_val) { - dt_write_lock(env, lo->ldo_stripe[i], - MOR_TGT_CHILD); - rc = lod_sub_ref_del(env, lo->ldo_stripe[i], - th); - dt_write_unlock(env, lo->ldo_stripe[i]); + dt_write_lock(env, stripe, DT_TGT_CHILD); + rc = lod_sub_ref_del(env, stripe, th); + dt_write_unlock(env, stripe); if (rc != 0) break; - rc = lod_sub_destroy(env, lo->ldo_stripe[i], - th); + rc = lod_sub_destroy(env, stripe, th); if (rc != 0) break; } @@ -5621,8 +6167,10 @@ static int lod_object_unlock(const struct lu_env *env, struct dt_object *dt, * NB, ha_count may not equal to ldo_dir_stripe_count, because dir * layout may change, e.g., shrink dir layout after migration. */ - for (i = 0; i < lo->ldo_dir_stripe_count; i++) - dt_invalidate(env, lo->ldo_stripe[i]); + for (i = 0; i < lo->ldo_dir_stripe_count; i++) { + if (lo->ldo_stripe[i]) + dt_invalidate(env, lo->ldo_stripe[i]); + } slave_locks_size = offsetof(typeof(*slave_locks), ha_handles[slave_locks->ha_count]); @@ -5683,17 +6231,19 @@ static int lod_object_lock(const struct lu_env *env, for (i = 0; i < lo->ldo_dir_stripe_count; i++) { struct lustre_handle lockh; struct ldlm_res_id *res_id; + struct dt_object *stripe; + + stripe = lo->ldo_stripe[i]; + if (!stripe) + continue; res_id = &lod_env_info(env)->lti_res_id; - fid_build_reg_res_name(lu_object_fid(&lo->ldo_stripe[i]->do_lu), - res_id); + fid_build_reg_res_name(lu_object_fid(&stripe->do_lu), res_id); einfo->ei_res_id = res_id; - LASSERT(lo->ldo_stripe[i] != NULL); - if (dt_object_remote(lo->ldo_stripe[i])) { + if (dt_object_remote(stripe)) { set_bit(i, (void *)slave_locks->ha_map); - rc = dt_object_lock(env, lo->ldo_stripe[i], &lockh, - einfo, policy); + rc = dt_object_lock(env, stripe, &lockh, einfo, policy); } else { struct ldlm_namespace *ns = einfo->ei_namespace; ldlm_blocking_callback blocking = einfo->ei_cb_local_bl; @@ -5705,7 +6255,7 @@ static int lod_object_lock(const struct lu_env *env, dlmflags |= LDLM_FL_COS_INCOMPAT; LASSERT(ns != NULL); - rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, + rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_IBITS, policy, einfo->ei_mode, &dlmflags, blocking, completion, NULL, @@ -5737,31 +6287,6 @@ static int lod_invalidate(const struct lu_env *env, struct dt_object *dt) return dt_invalidate(env, dt_object_child(dt)); } -static int lod_layout_data_init(struct lod_thread_info *info, __u32 comp_cnt) -{ - ENTRY; - - /* clear memory region that will be used for layout change */ - memset(&info->lti_layout_attr, 0, sizeof(struct lu_attr)); - info->lti_count = 0; - - if (info->lti_comp_size >= comp_cnt) - RETURN(0); - - if (info->lti_comp_size > 0) { - OBD_FREE(info->lti_comp_idx, - info->lti_comp_size * sizeof(__u32)); - info->lti_comp_size = 0; - } - - OBD_ALLOC(info->lti_comp_idx, comp_cnt * sizeof(__u32)); - if (!info->lti_comp_idx) - RETURN(-ENOMEM); - - info->lti_comp_size = comp_cnt; - RETURN(0); -} - static int lod_declare_instantiate_components(const struct lu_env *env, struct lod_object *lo, struct thandle *th) { @@ -5788,6 +6313,553 @@ static int lod_declare_instantiate_components(const struct lu_env *env, RETURN(rc); } +/** + * Check OSTs for an existing component for further extension + * + * Checks if OSTs are still healthy and not out of space. Gets free space + * on OSTs (relative to allocation watermark rmb_low) and compares to + * the proposed new_end for this component. + * + * Decides whether or not to extend a component on its current OSTs. + * + * \param[in] env execution environment for this thread + * \param[in] lo object we're checking + * \param[in] index index of this component + * \param[in] extension_size extension size for this component + * \param[in] extent layout extent for requested operation + * \param[in] comp_extent extension component extent + * \param[in] write if this is write operation + * + * \retval true - OK to extend on current OSTs + * \retval false - do not extend on current OSTs + */ +static bool lod_sel_osts_allowed(const struct lu_env *env, + struct lod_object *lo, + int index, __u64 extension_size, + struct lu_extent *extent, + struct lu_extent *comp_extent, int write) +{ + struct lod_layout_component *lod_comp = &lo->ldo_comp_entries[index]; + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; + __u64 available = 0; + __u64 size; + bool ret = true; + int i, rc; + + ENTRY; + + LASSERT(lod_comp->llc_stripe_count != 0); + + if (write == 0 || + (extent->e_start == 0 && extent->e_end == OBD_OBJECT_EOF)) { + /* truncate or append */ + size = extension_size; + } else { + /* In case of write op, check the real write extent, + * it may be larger than the extension_size */ + size = roundup(min(extent->e_end, comp_extent->e_end) - + max(extent->e_start, comp_extent->e_start), + extension_size); + } + /* extension_size is file level, so we must divide by stripe count to + * compare it to available space on a single OST */ + size /= lod_comp->llc_stripe_count; + + lod_getref(&lod->lod_ost_descs); + for (i = 0; i < lod_comp->llc_stripe_count; i++) { + int index = lod_comp->llc_ost_indices[i]; + struct lod_tgt_desc *ost = OST_TGT(lod, index); + struct obd_statfs_info info = { 0 }; + int j, repeated = 0; + + LASSERT(ost); + + /* Get the number of times this OST repeats in this component. + * Note: inter-component repeats are not counted as this is + * considered as a rare case: we try to not repeat OST in other + * components if possible. */ + for (j = 0; j < lod_comp->llc_stripe_count; j++) { + if (index != lod_comp->llc_ost_indices[j]) + continue; + + /* already handled */ + if (j < i) + break; + + repeated++; + } + if (j < lod_comp->llc_stripe_count) + continue; + + if (!cfs_bitmap_check(lod->lod_ost_bitmap, index)) { + CDEBUG(D_LAYOUT, "ost %d no longer present\n", index); + ret = false; + break; + } + + rc = dt_statfs_info(env, ost->ltd_ost, sfs, &info); + if (rc) { + CDEBUG(D_LAYOUT, "statfs failed for ost %d, error %d\n", + index, rc); + ret = false; + break; + } + + if (sfs->os_state & OS_STATE_ENOSPC || + sfs->os_state & OS_STATE_READONLY || + sfs->os_state & OS_STATE_DEGRADED) { + CDEBUG(D_LAYOUT, "ost %d is not availble for SEL " + "extension, state %u\n", index, sfs->os_state); + ret = false; + break; + } + + /* In bytes */ + available = sfs->os_bavail * sfs->os_bsize; + /* 'available' is relative to the allocation threshold */ + available -= (__u64) info.os_reserved_mb_low << 20; + + CDEBUG(D_LAYOUT, "ost %d lowwm: %d highwm: %d, " + "%llu %% blocks available, %llu %% blocks free\n", + index, info.os_reserved_mb_low, info.os_reserved_mb_high, + (100ull * sfs->os_bavail) / sfs->os_blocks, + (100ull * sfs->os_bfree) / sfs->os_blocks); + + if (size * repeated > available) { + ret = false; + CDEBUG(D_LAYOUT, "low space on ost %d, available %llu " + "< extension size %llu\n", index, available, + extension_size); + break; + } + } + lod_putref(lod, &lod->lod_ost_descs); + + RETURN(ret); +} + +/** + * Adjust extents after component removal + * + * When we remove an extension component, we move the start of the next + * component to match the start of the extension component, so no space is left + * without layout. + * + * \param[in] env execution environment for this thread + * \param[in] lo object + * \param[in] max_comp layout component + * \param[in] index index of this component + * + * \retval 0 on success + * \retval negative errno on error + */ +static void lod_sel_adjust_extents(const struct lu_env *env, + struct lod_object *lo, + int max_comp, int index) +{ + struct lod_layout_component *lod_comp = NULL; + struct lod_layout_component *next = NULL; + struct lod_layout_component *prev = NULL; + __u64 new_start = 0; + __u64 start; + int i; + + /* Extension space component */ + lod_comp = &lo->ldo_comp_entries[index]; + next = &lo->ldo_comp_entries[index + 1]; + prev = &lo->ldo_comp_entries[index - 1]; + + LASSERT(lod_comp != NULL && prev != NULL && next != NULL); + LASSERT(lod_comp->llc_flags & LCME_FL_EXTENSION); + + /* Previous is being removed */ + if (prev && prev->llc_id == LCME_ID_INVAL) + new_start = prev->llc_extent.e_start; + else + new_start = lod_comp->llc_extent.e_start; + + for (i = index + 1; i < max_comp; i++) { + lod_comp = &lo->ldo_comp_entries[i]; + + start = lod_comp->llc_extent.e_start; + lod_comp->llc_extent.e_start = new_start; + + /* We only move zero length extendable components */ + if (!(start == lod_comp->llc_extent.e_end)) + break; + + LASSERT(!(lod_comp->llc_flags & LCME_FL_INIT)); + + lod_comp->llc_extent.e_end = new_start; + } +} + +/* Calculate the proposed 'new end' for a component we're extending */ +static __u64 lod_extension_new_end(__u64 extension_size, __u64 extent_end, + __u32 stripe_size, __u64 component_end, + __u64 extension_end) +{ + __u64 new_end; + + LASSERT(extension_size != 0 && stripe_size != 0); + + /* Round up to extension size */ + if (extent_end == OBD_OBJECT_EOF) { + new_end = OBD_OBJECT_EOF; + } else { + /* Add at least extension_size to the previous component_end, + * covering the req layout extent */ + new_end = max(extent_end - component_end, extension_size); + new_end = roundup(new_end, extension_size); + new_end += component_end; + + /* Component end must be min stripe size aligned */ + if (new_end % stripe_size) { + CDEBUG(D_LAYOUT, "new component end is not aligned " + "by the stripe size %u: [%llu, %llu) ext size " + "%llu new end %llu, aligning\n", + stripe_size, component_end, extent_end, + extension_size, new_end); + new_end = roundup(new_end, stripe_size); + } + + /* Overflow */ + if (new_end < extent_end) + new_end = OBD_OBJECT_EOF; + } + + /* Don't extend past the end of the extension component */ + if (new_end > extension_end) + new_end = extension_end; + + return new_end; +} + +/* As lod_sel_handler() could be re-entered for the same component several + * times, this is the data for the next call. Fields could be changed to + * component indexes when needed, (e.g. if there is no need to instantiate + * all the previous components up to the current position) to tell the caller + * where to start over from. */ +struct sel_data { + int sd_force; + int sd_repeat; +}; + +/** + * Process extent updates for a particular layout component + * + * Handle layout updates for a particular extension space component touched by + * a layout update operation. Core function of self-extending PFL feature. + * + * In general, this function processes exactly *one* stage of an extension + * operation, modifying the layout accordingly, then returns to the caller. + * The caller is responsible for restarting processing with the new layout, + * which may repeatedly return to this function until the extension updates + * are complete. + * + * This function does one of a few things to the layout: + * 1. Extends the component before the current extension space component to + * allow it to accomodate the requested operation (if space/policy permit that + * component to continue on its current OSTs) + * + * 2. If extension of the existing component fails, we do one of two things: + * a. If there is a component after the extension space, we remove the + * extension space component, move the start of the next component down + * accordingly, then notify the caller to restart processing w/the new + * layout. + * b. If there is no following component, we try repeating the current + * component, creating a new component using the current one as a + * template (keeping its stripe properties but not specific striping), + * and try assigning striping for this component. If there is sufficient + * free space on the OSTs chosen for this component, it is instantiated + * and i/o continues there. + * + * If there is not sufficient space on the new OSTs, we remove this new + * component & extend the current component. + * + * Note further that uninited components followed by extension space can be zero + * length meaning that we will try to extend them before initializing them, and + * if that fails, they will be removed without initialization. + * + * 3. If we extend to/beyond the end of an extension space component, that + * component is exhausted (all of its range has been given to real components), + * so we remove it and restart processing. + * + * \param[in] env execution environment for this thread + * \param[in,out] lo object to update the layout of + * \param[in] extent layout extent for requested operation, update + * layout to fit this operation + * \param[in] th transaction handle for this operation + * \param[in,out] max_comp the highest comp for the portion of the layout + * we are operating on (For FLR, the chosen + * replica). Updated because we may remove + * components. + * \param[in] index index of the extension space component we're + * working on + * \param[in] write if this is write op + * \param[in,out] force if the extension is to be forced; set here + to force it on the 2nd call for the same + extension component + * + * \retval 0 on success + * \retval negative errno on error + */ +static int lod_sel_handler(const struct lu_env *env, + struct lod_object *lo, + struct lu_extent *extent, + struct thandle *th, int *max_comp, + int index, int write, + struct sel_data *sd) +{ + struct lod_device *d = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_thread_info *info = lod_env_info(env); + struct lod_layout_component *lod_comp; + struct lod_layout_component *prev; + struct lod_layout_component *next = NULL; + __u64 extension_size; + __u64 new_end = 0; + bool repeated; + int change = 0; + int rc = 0; + ENTRY; + + /* First component cannot be extension space */ + if (index == 0) { + CERROR("%s: "DFID" first component cannot be extension space\n", + lod2obd(d)->obd_name, PFID(lod_object_fid(lo))); + RETURN(-EINVAL); + } + + lod_comp = &lo->ldo_comp_entries[index]; + prev = &lo->ldo_comp_entries[index - 1]; + if ((index + 1) < *max_comp) + next = &lo->ldo_comp_entries[index + 1]; + + /* extension size uses the stripe size field as KiB */ + extension_size = lod_comp->llc_stripe_size * SEL_UNIT_SIZE; + + CDEBUG(D_LAYOUT, "prev start %llu, extension start %llu, extension end" + " %llu, extension size %llu\n", prev->llc_extent.e_start, + lod_comp->llc_extent.e_start, lod_comp->llc_extent.e_end, + extension_size); + + /* Two extension space components cannot be adjacent & extension space + * components cannot be init */ + if ((prev->llc_flags & LCME_FL_EXTENSION) || + !(ergo(next, !(next->llc_flags & LCME_FL_EXTENSION))) || + lod_comp_inited(lod_comp)) { + CERROR("%s: "DFID" invalid extension space components\n", + lod2obd(d)->obd_name, PFID(lod_object_fid(lo))); + RETURN(-EINVAL); + } + + if (!prev->llc_stripe) { + CDEBUG(D_LAYOUT, "Previous component not inited\n"); + info->lti_count = 1; + info->lti_comp_idx[0] = index - 1; + rc = lod_declare_instantiate_components(env, lo, th); + /* ENOSPC tells us we can't use this component. If there is + * a next or we are repeating, we either spill over (next) or + * extend the original comp (repeat). Otherwise, return the + * error to the user. */ + if (rc == -ENOSPC && (next || sd->sd_repeat)) + rc = 1; + if (rc < 0) + RETURN(rc); + } + + if (sd->sd_force == 0 && rc == 0) + rc = !lod_sel_osts_allowed(env, lo, index - 1, + extension_size, extent, + &lod_comp->llc_extent, write); + + repeated = !!(sd->sd_repeat); + sd->sd_repeat = 0; + sd->sd_force = 0; + + /* Extend previous component */ + if (rc == 0) { + new_end = lod_extension_new_end(extension_size, extent->e_end, + prev->llc_stripe_size, + prev->llc_extent.e_end, + lod_comp->llc_extent.e_end); + + CDEBUG(D_LAYOUT, "new end %llu\n", new_end); + lod_comp->llc_extent.e_start = new_end; + prev->llc_extent.e_end = new_end; + + if (prev->llc_extent.e_end == lod_comp->llc_extent.e_end) { + CDEBUG(D_LAYOUT, "Extension component exhausted\n"); + lod_comp->llc_id = LCME_ID_INVAL; + change--; + } + } else { + /* rc == 1, failed to extend current component */ + LASSERT(rc == 1); + if (next) { + /* Normal 'spillover' case - Remove the extension + * space component & bring down the start of the next + * component. */ + lod_comp->llc_id = LCME_ID_INVAL; + change--; + if (!(prev->llc_flags & LCME_FL_INIT)) { + prev->llc_id = LCME_ID_INVAL; + change--; + } + lod_sel_adjust_extents(env, lo, *max_comp, index); + } else if (lod_comp_inited(prev)) { + /* If there is no next, and the previous component is + * INIT'ed, try repeating the previous component. */ + LASSERT(repeated == 0); + rc = lod_layout_repeat_comp(env, lo, index - 1); + if (rc < 0) + RETURN(rc); + change++; + /* The previous component is a repeated component. + * Record this so we don't keep trying to repeat it. */ + sd->sd_repeat = 1; + } else { + /* If the previous component is not INIT'ed, this may + * be a component we have just instantiated but failed + * to extend. Or even a repeated component we failed + * to prepare a striping for. Do not repeat but instead + * remove the repeated component & force the extention + * of the original one */ + sd->sd_force = 1; + if (repeated) { + prev->llc_id = LCME_ID_INVAL; + change--; + } + } + } + + if (change < 0) { + rc = lod_layout_del_prep_layout(env, lo, NULL); + if (rc < 0) + RETURN(rc); + LASSERTF(-rc == change, + "number deleted %d != requested %d\n", -rc, + change); + } + *max_comp = *max_comp + change; + + /* lod_del_prep_layout reallocates ldo_comp_entries, so we must + * refresh these pointers before using them */ + lod_comp = &lo->ldo_comp_entries[index]; + prev = &lo->ldo_comp_entries[index - 1]; + CDEBUG(D_LAYOUT, "After extent updates: prev start %llu, current start " + "%llu, current end %llu max_comp %d ldo_comp_cnt %d\n", + prev->llc_extent.e_start, lod_comp->llc_extent.e_start, + lod_comp->llc_extent.e_end, *max_comp, lo->ldo_comp_cnt); + + /* Layout changed successfully */ + RETURN(0); +} + +/** + * Declare layout extent updates + * + * Handles extensions. Identifies extension components touched by current + * operation and passes them to processing function. + * + * Restarts with updated layouts from the processing function until the current + * operation no longer touches an extension space component. + * + * \param[in] env execution environment for this thread + * \param[in,out] lo object to update the layout of + * \param[in] extent layout extent for requested operation, update layout to + * fit this operation + * \param[in] th transaction handle for this operation + * \param[in] pick identifies chosen mirror for FLR layouts + * \param[in] write if this is write op + * + * \retval 1 on layout changed, 0 on no change + * \retval negative errno on error + */ +static int lod_declare_update_extents(const struct lu_env *env, + struct lod_object *lo, struct lu_extent *extent, + struct thandle *th, int pick, int write) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lod_layout_component *lod_comp; + bool layout_changed = false; + struct sel_data sd = { 0 }; + int start_index; + int i = 0; + int max_comp = 0; + int rc = 0, rc2; + int change = 0; + ENTRY; + + /* This makes us work on the components of the chosen mirror */ + start_index = lo->ldo_mirrors[pick].lme_start; + max_comp = lo->ldo_mirrors[pick].lme_end + 1; + if (lo->ldo_flr_state == LCM_FL_NONE) + LASSERT(start_index == 0 && max_comp == lo->ldo_comp_cnt); + + CDEBUG(D_LAYOUT, "extent->e_start %llu, extent->e_end %llu\n", + extent->e_start, extent->e_end); + for (i = start_index; i < max_comp; i++) { + lod_comp = &lo->ldo_comp_entries[i]; + + /* We've passed all components of interest */ + if (lod_comp->llc_extent.e_start >= extent->e_end) + break; + + if (lod_comp->llc_flags & LCME_FL_EXTENSION) { + layout_changed = true; + rc = lod_sel_handler(env, lo, extent, th, &max_comp, + i, write, &sd); + if (rc < 0) + GOTO(out, rc); + + /* Nothing has changed behind the prev one */ + i -= 2; + continue; + } + } + + /* We may have added or removed components. If so, we must update the + * start & ends of all the mirrors after the current one, and the end + * of the current mirror. */ + change = max_comp - 1 - lo->ldo_mirrors[pick].lme_end; + if (change) { + lo->ldo_mirrors[pick].lme_end += change; + for (i = pick + 1; i < lo->ldo_mirror_count; i++) { + lo->ldo_mirrors[i].lme_start += change; + lo->ldo_mirrors[i].lme_end += change; + } + } + + EXIT; +out: + /* The amount of components has changed, adjust the lti_comp_idx */ + rc2 = lod_layout_data_init(info, lo->ldo_comp_cnt); + + return rc < 0 ? rc : rc2 < 0 ? rc2 : layout_changed; +} + +/* If striping is already instantiated or INIT'ed DOM? */ +static bool lod_is_instantiation_needed(struct lod_layout_component *comp) +{ + return !(((lov_pattern(comp->llc_pattern) == LOV_PATTERN_MDT) && + lod_comp_inited(comp)) || comp->llc_stripe); +} + +/** + * Declare layout update for a non-FLR layout. + * + * \param[in] env execution environment for this thread + * \param[in,out] lo object to update the layout of + * \param[in] layout layout intent for requested operation, "update" is + * a process of reacting to this + * \param[in] buf buffer containing lov ea (see comment on usage inline) + * \param[in] th transaction handle for this operation + * + * \retval 0 on success + * \retval negative errno on error + */ static int lod_declare_update_plain(const struct lu_env *env, struct lod_object *lo, struct layout_intent *layout, const struct lu_buf *buf, struct thandle *th) @@ -5796,6 +6868,7 @@ static int lod_declare_update_plain(const struct lu_env *env, struct lod_device *d = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); struct lod_layout_component *lod_comp; struct lov_comp_md_v1 *comp_v1 = NULL; + bool layout_changed = false; bool replay = false; int i, rc; ENTRY; @@ -5831,6 +6904,11 @@ static int lod_declare_update_plain(const struct lu_env *env, /* old on-disk EA is stored in info->lti_buf */ comp_v1 = (struct lov_comp_md_v1 *)info->lti_buf.lb_buf; replay = true; + layout_changed = true; + + rc = lod_layout_data_init(info, lo->ldo_comp_cnt); + if (rc) + GOTO(out, rc); } else { /* non replay path */ rc = lod_striping_load(env, lo); @@ -5843,18 +6921,27 @@ static int lod_declare_update_plain(const struct lu_env *env, if (lo->ldo_comp_cnt > 1 && lod_comp->llc_extent.e_end != OBD_OBJECT_EOF && lod_comp->llc_extent.e_end < layout->li_extent.e_end) { - CDEBUG(replay ? D_ERROR : D_LAYOUT, - "%s: the defined layout [0, %#llx) does not covers " - "the write range "DEXT"\n", - lod2obd(d)->obd_name, lod_comp->llc_extent.e_end, - PEXT(&layout->li_extent)); + CDEBUG_LIMIT(replay ? D_ERROR : D_LAYOUT, + "%s: the defined layout [0, %#llx) does not " + "covers the write range "DEXT"\n", + lod2obd(d)->obd_name, lod_comp->llc_extent.e_end, + PEXT(&layout->li_extent)); GOTO(out, rc = -EINVAL); } - CDEBUG(D_LAYOUT, "%s: "DFID": instantiate components "DEXT"\n", + CDEBUG(D_LAYOUT, "%s: "DFID": update components "DEXT"\n", lod2obd(d)->obd_name, PFID(lod_object_fid(lo)), PEXT(&layout->li_extent)); + if (!replay) { + rc = lod_declare_update_extents(env, lo, &layout->li_extent, + th, 0, layout->li_opc == LAYOUT_INTENT_WRITE); + if (rc < 0) + GOTO(out, rc); + else if (rc) + layout_changed = true; + } + /* * Iterate ld->ldo_comp_entries, find the component whose extent under * the write range and not instantianted. @@ -5866,7 +6953,8 @@ static int lod_declare_update_plain(const struct lu_env *env, break; if (!replay) { - if (lod_comp_inited(lod_comp)) + /* If striping is instantiated or INIT'ed DOM skip */ + if (!lod_is_instantiation_needed(lod_comp)) continue; } else { /** @@ -5890,17 +6978,19 @@ static int lod_declare_update_plain(const struct lu_env *env, LASSERT(info->lti_comp_idx != NULL); info->lti_comp_idx[info->lti_count++] = i; + layout_changed = true; } - if (info->lti_count == 0) + if (!layout_changed) RETURN(-EALREADY); lod_obj_inc_layout_gen(lo); rc = lod_declare_instantiate_components(env, lo, th); + EXIT; out: if (rc) lod_striping_free(env, lo); - RETURN(rc); + return rc; } static inline int lod_comp_index(struct lod_object *lo, @@ -5915,15 +7005,21 @@ static inline int lod_comp_index(struct lod_object *lo, /** * Stale other mirrors by writing extent. */ -static void lod_stale_components(struct lod_object *lo, int primary, - struct lu_extent *extent) +static int lod_stale_components(const struct lu_env *env, struct lod_object *lo, + int primary, struct lu_extent *extent, + struct thandle *th) { struct lod_layout_component *pri_comp, *lod_comp; + struct lu_extent pri_extent; + int rc = 0; int i; + ENTRY; /* The writing extent decides which components in the primary * are affected... */ CDEBUG(D_LAYOUT, "primary mirror %d, "DEXT"\n", primary, PEXT(extent)); + +restart: lod_foreach_mirror_comp(pri_comp, lo, primary) { if (!lu_extent_is_overlapped(extent, &pri_comp->llc_extent)) continue; @@ -5932,15 +7028,27 @@ static void lod_stale_components(struct lod_object *lo, int primary, lod_comp_index(lo, pri_comp), PEXT(&pri_comp->llc_extent)); + pri_extent.e_start = pri_comp->llc_extent.e_start; + pri_extent.e_end = pri_comp->llc_extent.e_end; + for (i = 0; i < lo->ldo_mirror_count; i++) { if (i == primary) continue; + rc = lod_declare_update_extents(env, lo, &pri_extent, + th, i, 0); + /* if update_extents changed the layout, it may have + * reallocated the component array, so start over to + * avoid using stale pointers */ + if (rc == 1) + goto restart; + if (rc < 0) + RETURN(rc); /* ... and then stale other components that are * overlapping with primary components */ lod_foreach_mirror_comp(lod_comp, lo, i) { if (!lu_extent_is_overlapped( - &pri_comp->llc_extent, + &pri_extent, &lod_comp->llc_extent)) continue; @@ -5952,6 +7060,8 @@ static void lod_stale_components(struct lod_object *lo, int primary, } } } + + RETURN(rc); } /** @@ -6098,6 +7208,34 @@ static int lod_primary_pick(const struct lu_env *env, struct lod_object *lo, RETURN(picked); } +static int lod_prepare_resync_mirror(const struct lu_env *env, + struct lod_object *lo, + __u16 mirror_id) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lod_layout_component *lod_comp; + bool neg = !!(MIRROR_ID_NEG & mirror_id); + int i; + + mirror_id &= ~MIRROR_ID_NEG; + + for (i = 0; i < lo->ldo_mirror_count; i++) { + if ((!neg && lo->ldo_mirrors[i].lme_id != mirror_id) || + (neg && lo->ldo_mirrors[i].lme_id == mirror_id)) + continue; + + lod_foreach_mirror_comp(lod_comp, lo, i) { + if (lod_comp_inited(lod_comp)) + continue; + + info->lti_comp_idx[info->lti_count++] = + lod_comp_index(lo, lod_comp); + } + } + + return 0; +} + /** * figure out the components should be instantiated for resync. */ @@ -6159,6 +7297,7 @@ static int lod_declare_update_rdonly(const struct lu_env *env, if (mlc->mlc_opc == MD_LAYOUT_WRITE) { struct layout_intent *layout = mlc->mlc_intent; + int write = layout->li_opc == LAYOUT_INTENT_WRITE; int picked; extent = layout->li_extent; @@ -6173,6 +7312,12 @@ static int lod_declare_update_rdonly(const struct lu_env *env, PFID(lod_object_fid(lo)), lo->ldo_mirrors[picked].lme_id); + /* Update extents of primary before staling */ + rc = lod_declare_update_extents(env, lo, &extent, th, picked, + write); + if (rc < 0) + GOTO(out, rc); + if (layout->li_opc == LAYOUT_INTENT_TRUNC) { /** * trunc transfers [0, size) in the intent extent, we'd @@ -6183,7 +7328,9 @@ static int lod_declare_update_rdonly(const struct lu_env *env, } /* stale overlapping components from other mirrors */ - lod_stale_components(lo, picked, &extent); + rc = lod_stale_components(env, lo, picked, &extent, th); + if (rc < 0) + GOTO(out, rc); /* restore truncate intent extent */ if (layout->li_opc == LAYOUT_INTENT_TRUNC) @@ -6197,7 +7344,7 @@ static int lod_declare_update_rdonly(const struct lu_env *env, &lod_comp->llc_extent)) break; - if (lod_comp_inited(lod_comp)) + if (!lod_is_instantiation_needed(lod_comp)) continue; info->lti_comp_idx[info->lti_count++] = @@ -6213,23 +7360,33 @@ static int lod_declare_update_rdonly(const struct lu_env *env, * prep uninited all components assuming any non-stale mirror * could be picked as the primary mirror. */ - for (i = 0; i < lo->ldo_mirror_count; i++) { - if (lo->ldo_mirrors[i].lme_stale) - continue; + if (mlc->mlc_mirror_id == 0) { + /* normal resync */ + for (i = 0; i < lo->ldo_mirror_count; i++) { + if (lo->ldo_mirrors[i].lme_stale) + continue; - lod_foreach_mirror_comp(lod_comp, lo, i) { - if (!lod_comp_inited(lod_comp)) - break; + lod_foreach_mirror_comp(lod_comp, lo, i) { + if (!lod_comp_inited(lod_comp)) + break; - if (extent.e_end < lod_comp->llc_extent.e_end) - extent.e_end = - lod_comp->llc_extent.e_end; + if (extent.e_end < + lod_comp->llc_extent.e_end) + extent.e_end = + lod_comp->llc_extent.e_end; + } } + rc = lod_prepare_resync(env, lo, &extent); + if (rc) + GOTO(out, rc); + } else { + /* mirror write, try to init its all components */ + rc = lod_prepare_resync_mirror(env, lo, + mlc->mlc_mirror_id); + if (rc) + GOTO(out, rc); } - rc = lod_prepare_resync(env, lo, &extent); - if (rc) - GOTO(out, rc); /* change the file state to SYNC_PENDING */ lo->ldo_flr_state = LCM_FL_SYNC_PENDING; } @@ -6241,7 +7398,7 @@ static int lod_declare_update_rdonly(const struct lu_env *env, if (lo->ldo_layout_gen > (LCME_ID_MAX >> 1)) { __u32 layout_version; - cfs_get_random_bytes(&layout_version, sizeof(layout_version)); + get_random_bytes(&layout_version, sizeof(layout_version)); lo->ldo_layout_gen = layout_version & 0xffff; } @@ -6313,6 +7470,9 @@ static int lod_declare_update_write_pending(const struct lu_env *env, * 2. transfer layout version to all objects to close write era. */ if (mlc->mlc_opc == MD_LAYOUT_WRITE) { + struct layout_intent *layout = mlc->mlc_intent; + int write = layout->li_opc == LAYOUT_INTENT_WRITE; + LASSERT(mlc->mlc_intent != NULL); extent = mlc->mlc_intent->li_extent; @@ -6320,6 +7480,12 @@ static int lod_declare_update_write_pending(const struct lu_env *env, CDEBUG(D_LAYOUT, DFID": intent to write: "DEXT"\n", PFID(lod_object_fid(lo)), PEXT(&extent)); + /* 1. Update extents of primary before staling */ + rc = lod_declare_update_extents(env, lo, &extent, th, primary, + write); + if (rc < 0) + GOTO(out, rc); + if (mlc->mlc_intent->li_opc == LAYOUT_INTENT_TRUNC) { /** * trunc transfers [0, size) in the intent extent, we'd @@ -6328,10 +7494,13 @@ static int lod_declare_update_write_pending(const struct lu_env *env, extent.e_start = extent.e_end; extent.e_end = OBD_OBJECT_EOF; } - /* 1. stale overlapping components */ - lod_stale_components(lo, primary, &extent); - /* 2. find out the components need instantiating. + /* 2. stale overlapping components */ + rc = lod_stale_components(env, lo, primary, &extent, th); + if (rc < 0) + GOTO(out, rc); + + /* 3. find the components which need instantiating. * instantiate [0, mlc->mlc_intent->e_end) */ /* restore truncate intent extent */ @@ -6344,7 +7513,7 @@ static int lod_declare_update_write_pending(const struct lu_env *env, &lod_comp->llc_extent)) break; - if (lod_comp_inited(lod_comp)) + if (!lod_is_instantiation_needed(lod_comp)) continue; CDEBUG(D_LAYOUT, "write instantiate %d / %d\n", @@ -6353,16 +7522,26 @@ static int lod_declare_update_write_pending(const struct lu_env *env, lod_comp_index(lo, lod_comp); } } else { /* MD_LAYOUT_RESYNC */ - lod_foreach_mirror_comp(lod_comp, lo, primary) { - if (!lod_comp_inited(lod_comp)) - break; + if (mlc->mlc_mirror_id == 0) { + /* normal resync */ + lod_foreach_mirror_comp(lod_comp, lo, primary) { + if (!lod_comp_inited(lod_comp)) + break; + + extent.e_end = lod_comp->llc_extent.e_end; + } - extent.e_end = lod_comp->llc_extent.e_end; + rc = lod_prepare_resync(env, lo, &extent); + if (rc) + GOTO(out, rc); + } else { + /* mirror write, try to init its all components */ + rc = lod_prepare_resync_mirror(env, lo, + mlc->mlc_mirror_id); + if (rc) + GOTO(out, rc); } - rc = lod_prepare_resync(env, lo, &extent); - if (rc) - GOTO(out, rc); /* change the file state to SYNC_PENDING */ lo->ldo_flr_state = LCM_FL_SYNC_PENDING; } @@ -6461,9 +7640,9 @@ static int lod_declare_update_sync_pending(const struct lu_env *env, GOTO(out, rc = -EINVAL); } - CDEBUG(D_LAYOUT, DFID": resynced %u/%zu components\n", + CDEBUG(D_LAYOUT, DFID": synced %u resynced %u/%zu components\n", PFID(lod_object_fid(lo)), - resync_components, mlc->mlc_resync_count); + sync_components, resync_components, mlc->mlc_resync_count); lo->ldo_flr_state = LCM_FL_RDONLY; lod_obj_inc_layout_gen(lo); @@ -6613,11 +7792,11 @@ static ssize_t lod_declare_write(const struct lu_env *env, */ static ssize_t lod_write(const struct lu_env *env, struct dt_object *dt, const struct lu_buf *buf, loff_t *pos, - struct thandle *th, int iq) + struct thandle *th) { LASSERT(S_ISREG(dt->do_lu.lo_header->loh_attr) || S_ISLNK(dt->do_lu.lo_header->loh_attr)); - return lod_sub_write(env, dt_object_child(dt), buf, pos, th, iq); + return lod_sub_write(env, dt_object_child(dt), buf, pos, th); } static int lod_declare_punch(const struct lu_env *env, struct dt_object *dt, @@ -6678,16 +7857,8 @@ static int lod_object_init(const struct lu_env *env, struct lu_object *lo, ENTRY; rc = lod_fld_lookup(env, lod, lu_object_fid(lo), &idx, &type); - if (rc != 0) { - /* Note: Sometimes, it will Return EAGAIN here, see - * ptrlpc_import_delay_req(), which might confuse - * lu_object_find_at() and make it wait there incorrectly. - * so we convert it to EIO here.*/ - if (rc == -EAGAIN) - rc = -EIO; - + if (rc != 0) RETURN(rc); - } if (type == LU_SEQ_RANGE_MDT && idx == lu_site2seq(lo->lo_dev->ld_site)->ss_node_id) { @@ -6731,6 +7902,56 @@ static int lod_object_init(const struct lu_env *env, struct lu_object *lo, /** * + * Alloc cached foreign LOV + * + * \param[in] lo object + * \param[in] size size of foreign LOV + * + * \retval 0 on success + * \retval negative if failed + */ +int lod_alloc_foreign_lov(struct lod_object *lo, size_t size) +{ + OBD_ALLOC_LARGE(lo->ldo_foreign_lov, size); + if (lo->ldo_foreign_lov == NULL) + return -ENOMEM; + lo->ldo_foreign_lov_size = size; + lo->ldo_is_foreign = 1; + return 0; +} + +/** + * + * Free cached foreign LOV + * + * \param[in] lo object + */ +void lod_free_foreign_lov(struct lod_object *lo) +{ + if (lo->ldo_foreign_lov != NULL) + OBD_FREE_LARGE(lo->ldo_foreign_lov, lo->ldo_foreign_lov_size); + lo->ldo_foreign_lov = NULL; + lo->ldo_foreign_lov_size = 0; + lo->ldo_is_foreign = 0; +} + +/** + * + * Free cached foreign LMV + * + * \param[in] lo object + */ +void lod_free_foreign_lmv(struct lod_object *lo) +{ + if (lo->ldo_foreign_lmv != NULL) + OBD_FREE_LARGE(lo->ldo_foreign_lmv, lo->ldo_foreign_lmv_size); + lo->ldo_foreign_lmv = NULL; + lo->ldo_foreign_lmv_size = 0; + lo->ldo_dir_is_foreign = 0; +} + +/** + * * Release resources associated with striping. * * If the object is striped (regular or directory), then release @@ -6744,7 +7965,13 @@ void lod_striping_free_nolock(const struct lu_env *env, struct lod_object *lo) struct lod_layout_component *lod_comp; int i, j; - if (lo->ldo_stripe != NULL) { + if (unlikely(lo->ldo_is_foreign)) { + lod_free_foreign_lov(lo); + lo->ldo_comp_cached = 0; + } else if (unlikely(lo->ldo_dir_is_foreign)) { + lod_free_foreign_lmv(lo); + lo->ldo_dir_stripe_loaded = 0; + } else if (lo->ldo_stripe != NULL) { LASSERT(lo->ldo_comp_entries == NULL); LASSERT(lo->ldo_dir_stripes_allocated > 0);