X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Flod%2Flod_object.c;h=02c14c21af2db221a9e1ab8760cbb35a0c10daf2;hb=e88992a3b5b9d9ba0a69883671f1b5888514e05d;hp=8fa06a6567fea51cc67d7129609ac20c74cbb740;hpb=040179b84330689b1b2cd128dcdd711bfc8f0b4e;p=fs%2Flustre-release.git diff --git a/lustre/lod/lod_object.c b/lustre/lod/lod_object.c index 8fa06a6..02c14c2 100644 --- a/lustre/lod/lod_object.c +++ b/lustre/lod/lod_object.c @@ -23,7 +23,7 @@ * Copyright 2009 Sun Microsystems, Inc. All rights reserved * Use is subject to license terms. * - * Copyright (c) 2012, Intel Corporation. + * Copyright (c) 2012, 2013, Intel Corporation. */ /* * lustre/lod/lod_object.c @@ -31,9 +31,6 @@ * Author: Alex Zhuravlev */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS #include @@ -45,11 +42,16 @@ #include #include #include +#include #include +#include #include "lod_internal.h" -extern cfs_mem_cache_t *lod_object_kmem; +static const char dot[] = "."; +static const char dotdot[] = ".."; + +extern struct kmem_cache *lod_object_kmem; static const struct dt_body_operations lod_body_lnk_ops; static int lod_index_lookup(const struct lu_env *env, struct dt_object *dt, @@ -298,9 +300,17 @@ static int lod_declare_attr_set(const struct lu_env *env, * Therefore we need not load striping unless ownership is * changing. This should save memory and (we hope) speed up * rename(). */ - if (!(attr->la_valid & (LA_UID | LA_GID))) - RETURN(rc); + if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) { + if (!(attr->la_valid & (LA_UID | LA_GID))) + RETURN(rc); + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_OWNER)) + RETURN(0); + } else { + if (!(attr->la_valid & (LA_UID | LA_GID | LA_MODE | + LA_ATIME | LA_MTIME | LA_CTIME))) + RETURN(rc); + } /* * load striping information, notice we don't do this when object * is being initialized as we don't need this information till @@ -310,12 +320,38 @@ static int lod_declare_attr_set(const struct lu_env *env, if (rc) RETURN(rc); + if (lo->ldo_stripenr == 0) + RETURN(0); + + if (!(attr->la_valid & ~(LA_ATIME | LA_MTIME | LA_CTIME))) { + struct lu_attr *la = &lod_env_info(env)->lti_attr; + bool setattr_time = false; + + rc = dt_attr_get(env, dt_object_child(dt), la, + BYPASS_CAPA); + if (rc != 0) + RETURN(rc); + + /* If it will only setattr time, it will only set + * time < current_time */ + if ((attr->la_valid & LA_ATIME && + attr->la_atime < la->la_atime) || + (attr->la_valid & LA_CTIME && + attr->la_ctime < la->la_ctime) || + (attr->la_valid & LA_MTIME && + attr->la_mtime < la->la_mtime)) + setattr_time = true; + + if (!setattr_time) + RETURN(0); + } /* * if object is striped declare changes on the stripes */ - LASSERT(lo->ldo_stripe || lo->ldo_stripenr == 0); + LASSERT(lo->ldo_stripe); for (i = 0; i < lo->ldo_stripenr; i++) { LASSERT(lo->ldo_stripe[i]); + rc = dt_declare_attr_set(env, lo->ldo_stripe[i], attr, handle); if (rc) { CERROR("failed declaration: %d\n", rc); @@ -323,6 +359,23 @@ static int lod_declare_attr_set(const struct lu_env *env, } } + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_STRIPE) && + dt_object_exists(next) != 0 && + dt_object_remote(next) == 0) + dt_declare_xattr_del(env, next, XATTR_NAME_LOV, handle); + + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_CHANGE_STRIPE) && + dt_object_exists(next) && + dt_object_remote(next) == 0 && S_ISREG(attr->la_mode)) { + struct lod_thread_info *info = lod_env_info(env); + struct lu_buf *buf = &info->lti_buf; + + buf->lb_buf = info->lti_ea_store; + buf->lb_len = info->lti_ea_store_size; + dt_declare_xattr_set(env, next, buf, XATTR_NAME_LOV, + LU_XATTR_REPLACE, handle); + } + RETURN(rc); } @@ -344,13 +397,48 @@ static int lod_attr_set(const struct lu_env *env, if (rc) RETURN(rc); - if (!(attr->la_valid & (LA_UID | LA_GID))) - RETURN(rc); + if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) { + if (!(attr->la_valid & (LA_UID | LA_GID))) + RETURN(rc); + + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_OWNER)) + RETURN(0); + } else { + if (!(attr->la_valid & (LA_UID | LA_GID | LA_MODE | + LA_ATIME | LA_MTIME | LA_CTIME))) + RETURN(rc); + } + + if (lo->ldo_stripenr == 0) + RETURN(0); + + if (!(attr->la_valid & ~(LA_ATIME | LA_MTIME | LA_CTIME))) { + struct lu_attr *la = &lod_env_info(env)->lti_attr; + bool setattr_time = false; + + rc = dt_attr_get(env, dt_object_child(dt), la, + BYPASS_CAPA); + if (rc != 0) + RETURN(rc); + + /* If it will only setattr time, it will only set + * time < current_time */ + if ((attr->la_valid & LA_ATIME && + attr->la_atime < la->la_atime) || + (attr->la_valid & LA_CTIME && + attr->la_ctime < la->la_ctime) || + (attr->la_valid & LA_MTIME && + attr->la_mtime < la->la_mtime)) + setattr_time = true; + + if (!setattr_time) + RETURN(0); + } /* * if object is striped, apply changes to all the stripes */ - LASSERT(lo->ldo_stripe || lo->ldo_stripenr == 0); + LASSERT(lo->ldo_stripe); for (i = 0; i < lo->ldo_stripenr; i++) { LASSERT(lo->ldo_stripe[i]); rc = dt_attr_set(env, lo->ldo_stripe[i], attr, handle, capa); @@ -360,6 +448,44 @@ static int lod_attr_set(const struct lu_env *env, } } + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_STRIPE) && + dt_object_exists(next) != 0 && + dt_object_remote(next) == 0) + dt_xattr_del(env, next, XATTR_NAME_LOV, handle, BYPASS_CAPA); + + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_CHANGE_STRIPE) && + dt_object_exists(next) && + dt_object_remote(next) == 0 && S_ISREG(attr->la_mode)) { + struct lod_thread_info *info = lod_env_info(env); + struct lu_buf *buf = &info->lti_buf; + struct ost_id *oi = &info->lti_ostid; + struct lu_fid *fid = &info->lti_fid; + struct lov_mds_md_v1 *lmm; + struct lov_ost_data_v1 *objs; + __u32 magic; + int rc1; + + rc1 = lod_get_lov_ea(env, lo); + if (rc1 <= 0) + RETURN(rc); + + buf->lb_buf = info->lti_ea_store; + buf->lb_len = info->lti_ea_store_size; + lmm = info->lti_ea_store; + magic = le32_to_cpu(lmm->lmm_magic); + if (magic == LOV_MAGIC_V1) + objs = &(lmm->lmm_objects[0]); + else + objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0]; + ostid_le_to_cpu(&objs->l_ost_oi, oi); + ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx)); + fid->f_oid--; + fid_to_ostid(fid, oi); + ostid_cpu_to_le(oi, &objs->l_ost_oi); + dt_xattr_set(env, next, buf, XATTR_NAME_LOV, + LU_XATTR_REPLACE, handle, BYPASS_CAPA); + } + RETURN(rc); } @@ -390,15 +516,20 @@ static int lod_xattr_get(const struct lu_env *env, struct dt_object *dt, struct lov_desc *desc = &dev->lod_desc; if (buf->lb_buf == NULL) { - rc = sizeof(struct lov_user_md_v1); - } else if (buf->lb_len >= sizeof(struct lov_user_md_v1)) { - lum->lmm_magic = LOV_USER_MAGIC_V1; - ostid_set_seq(&lum->lmm_oi, FID_SEQ_LOV_DEFAULT); - lum->lmm_pattern = desc->ld_pattern; - lum->lmm_stripe_size = desc->ld_default_stripe_size; - lum->lmm_stripe_count = desc->ld_default_stripe_count; - lum->lmm_stripe_offset = desc->ld_default_stripe_offset; - rc = sizeof(struct lov_user_md_v1); + rc = sizeof(*lum); + } else if (buf->lb_len >= sizeof(*lum)) { + lum->lmm_magic = cpu_to_le32(LOV_USER_MAGIC_V1); + lmm_oi_set_seq(&lum->lmm_oi, FID_SEQ_LOV_DEFAULT); + lmm_oi_set_id(&lum->lmm_oi, 0); + lmm_oi_cpu_to_le(&lum->lmm_oi, &lum->lmm_oi); + lum->lmm_pattern = cpu_to_le32(desc->ld_pattern); + lum->lmm_stripe_size = cpu_to_le32( + desc->ld_default_stripe_size); + lum->lmm_stripe_count = cpu_to_le16( + desc->ld_default_stripe_count); + lum->lmm_stripe_offset = cpu_to_le16( + desc->ld_default_stripe_offset); + rc = sizeof(*lum); } else { rc = -ERANGE; } @@ -407,6 +538,393 @@ static int lod_xattr_get(const struct lu_env *env, struct dt_object *dt, RETURN(rc); } +static int lod_verify_md_striping(struct lod_device *lod, + const struct lmv_user_md_v1 *lum) +{ + int rc = 0; + ENTRY; + + if (unlikely(le32_to_cpu(lum->lum_magic) != LMV_USER_MAGIC)) + GOTO(out, rc = -EINVAL); + + if (unlikely(le32_to_cpu(lum->lum_stripe_count) == 0)) + GOTO(out, rc = -EINVAL); + + if (unlikely(le32_to_cpu(lum->lum_stripe_count) > + lod->lod_remote_mdt_count + 1)) + GOTO(out, rc = -EINVAL); +out: + if (rc != 0) + CERROR("%s: invalid lmv_user_md: magic = %x, " + "stripe_offset = %d, stripe_count = %u: rc = %d\n", + lod2obd(lod)->obd_name, le32_to_cpu(lum->lum_magic), + (int)le32_to_cpu(lum->lum_stripe_offset), + le32_to_cpu(lum->lum_stripe_count), rc); + return rc; +} + +int lod_prep_lmv_md(const struct lu_env *env, struct dt_object *dt, + struct lu_buf *lmv_buf) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lod_device *lod = lu2lod_dev(dt->do_lu.lo_dev); + struct lod_object *lo = lod_dt_obj(dt); + struct lmv_mds_md_v1 *lmm1; + int stripe_count; + int lmm_size; + int type = LU_SEQ_RANGE_ANY; + int i; + int rc; + __u32 mdtidx; + ENTRY; + + LASSERT(lo->ldo_dir_striped != 0); + LASSERT(lo->ldo_stripenr > 0); + stripe_count = lo->ldo_stripenr + 1; + lmm_size = lmv_mds_md_size(stripe_count, LMV_MAGIC); + if (info->lti_ea_store_size < lmm_size) { + rc = lod_ea_store_resize(info, lmm_size); + if (rc != 0) + RETURN(rc); + } + + lmm1 = (struct lmv_mds_md_v1 *)info->lti_ea_store; + lmm1->lmv_magic = cpu_to_le32(LMV_MAGIC); + lmm1->lmv_stripe_count = cpu_to_le32(stripe_count); + lmm1->lmv_hash_type = cpu_to_le32(lo->ldo_dir_hash_type); + rc = lod_fld_lookup(env, lod, lu_object_fid(&dt->do_lu), + &mdtidx, &type); + if (rc != 0) + RETURN(rc); + + lmm1->lmv_master_mdt_index = cpu_to_le32(mdtidx); + fid_cpu_to_le(&lmm1->lmv_stripe_fids[0], lu_object_fid(&dt->do_lu)); + for (i = 0; i < lo->ldo_stripenr; i++) { + struct dt_object *dto; + + dto = lo->ldo_stripe[i]; + LASSERT(dto != NULL); + fid_cpu_to_le(&lmm1->lmv_stripe_fids[i + 1], + lu_object_fid(&dto->do_lu)); + } + + lmv_buf->lb_buf = info->lti_ea_store; + lmv_buf->lb_len = lmm_size; + lo->ldo_dir_striping_cached = 1; + + RETURN(rc); +} + +int lod_parse_dir_striping(const struct lu_env *env, struct lod_object *lo, + const struct lu_buf *buf) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_tgt_descs *ltd = &lod->lod_mdt_descs; + struct dt_object **stripe; + union lmv_mds_md *lmm = buf->lb_buf; + struct lmv_mds_md_v1 *lmv1 = &lmm->lmv_md_v1; + struct lu_fid *fid = &info->lti_fid; + int i; + int rc = 0; + ENTRY; + + if (le32_to_cpu(lmv1->lmv_magic) == LMV_MAGIC_MIGRATE) + RETURN(0); + + if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1) + RETURN(-EINVAL); + + if (le32_to_cpu(lmv1->lmv_stripe_count) <= 1) + RETURN(0); + + fid_le_to_cpu(fid, &lmv1->lmv_stripe_fids[0]); + /* Do not load striping information for slave inode */ + if (!lu_fid_eq(fid, lu_object_fid(&lo->ldo_obj.do_lu))) { + lo->ldo_dir_slave_stripe = 1; + RETURN(0); + } + + LASSERT(lo->ldo_stripe == NULL); + OBD_ALLOC(stripe, sizeof(stripe[0]) * + (le32_to_cpu(lmv1->lmv_stripe_count) - 1)); + if (stripe == NULL) + RETURN(-ENOMEM); + + /* skip master stripe */ + for (i = 1; i < le32_to_cpu(lmv1->lmv_stripe_count); i++) { + struct lod_tgt_desc *tgt; + int idx; + int type = LU_SEQ_RANGE_ANY; + struct dt_object *dto; + + fid_le_to_cpu(fid, &lmv1->lmv_stripe_fids[i]); + rc = lod_fld_lookup(env, lod, fid, &idx, &type); + if (rc != 0) + GOTO(out, rc); + + tgt = LTD_TGT(ltd, idx); + if (tgt == NULL) + GOTO(out, rc = -ESTALE); + + dto = dt_locate_at(env, tgt->ltd_tgt, fid, + lo->ldo_obj.do_lu.lo_dev->ld_site->ls_top_dev, + NULL); + if (IS_ERR(dto)) + GOTO(out, rc = PTR_ERR(dto)); + + stripe[i - 1] = dto; + } +out: + lo->ldo_stripe = stripe; + lo->ldo_stripenr = le32_to_cpu(lmv1->lmv_stripe_count) - 1; + lo->ldo_stripes_allocated = le32_to_cpu(lmv1->lmv_stripe_count) - 1; + if (rc != 0) + lod_object_free_striping(env, lo); + + RETURN(rc); +} + +static int lod_prep_md_striped_create(const struct lu_env *env, + struct dt_object *dt, + struct lu_attr *attr, + const struct lmv_user_md_v1 *lum, + struct thandle *th) +{ + struct lod_device *lod = lu2lod_dev(dt->do_lu.lo_dev); + struct lod_tgt_descs *ltd = &lod->lod_mdt_descs; + struct lod_object *lo = lod_dt_obj(dt); + struct dt_object **stripe; + struct lu_buf lmv_buf; + int stripe_count; + int *idx_array; + int rc = 0; + int i; + int j; + ENTRY; + + /* The lum has been verifed in lod_verify_md_striping */ + LASSERT(le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC); + LASSERT(le32_to_cpu(lum->lum_stripe_count) > 0); + + /* Do not need allocated master stripe */ + stripe_count = le32_to_cpu(lum->lum_stripe_count); + OBD_ALLOC(stripe, sizeof(stripe[0]) * (stripe_count - 1)); + if (stripe == NULL) + RETURN(-ENOMEM); + + OBD_ALLOC(idx_array, sizeof(idx_array[0]) * stripe_count); + if (idx_array == NULL) + GOTO(out_free, rc = -ENOMEM); + + idx_array[0] = le32_to_cpu(lum->lum_stripe_offset); + for (i = 1; i < stripe_count; i++) { + struct lod_tgt_desc *tgt; + struct dt_object *dto; + struct lu_fid fid; + int idx; + struct lu_object_conf conf = { 0 }; + + idx = (idx_array[i - 1] + 1) % (lod->lod_remote_mdt_count + 1); + + for (j = 0; j < lod->lod_remote_mdt_count; + j++, idx = (idx + 1) % (lod->lod_remote_mdt_count + 1)) { + bool already_allocated = false; + int k; + + CDEBUG(D_INFO, "try idx %d, mdt cnt %d," + " allocated %d, last allocated %d\n", idx, + lod->lod_remote_mdt_count, i, idx_array[i - 1]); + + /* Find next avaible target */ + if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx)) + continue; + + /* check whether the idx already exists + * in current allocated array */ + for (k = 0; k < i; k++) { + if (idx_array[k] == idx) { + already_allocated = true; + break; + } + } + + if (already_allocated) + continue; + + break; + } + + /* Can not allocate more stripes */ + if (j == lod->lod_remote_mdt_count) { + CDEBUG(D_INFO, "%s: require stripes %d only get %d\n", + lod2obd(lod)->obd_name, stripe_count, i - 1); + break; + } + + CDEBUG(D_INFO, "idx %d, mdt cnt %d," + " allocated %d, last allocated %d\n", idx, + lod->lod_remote_mdt_count, i, idx_array[i - 1]); + + tgt = LTD_TGT(ltd, idx); + LASSERT(tgt != NULL); + + rc = obd_fid_alloc(tgt->ltd_exp, &fid, NULL); + if (rc < 0) + GOTO(out_put, rc); + rc = 0; + + conf.loc_flags = LOC_F_NEW; + dto = dt_locate_at(env, tgt->ltd_tgt, &fid, + dt->do_lu.lo_dev->ld_site->ls_top_dev, &conf); + if (IS_ERR(dto)) + GOTO(out_put, rc = PTR_ERR(dto)); + stripe[i - 1] = dto; + idx_array[i] = idx; + } + + lo->ldo_dir_striped = 1; + lo->ldo_stripe = stripe; + lo->ldo_stripenr = i - 1; + lo->ldo_stripes_allocated = stripe_count - 1; + + if (lo->ldo_stripenr == 0) + GOTO(out_put, rc = -ENOSPC); + + rc = lod_prep_lmv_md(env, dt, &lmv_buf); + if (rc != 0) + GOTO(out_put, rc); + + for (i = 0; i < lo->ldo_stripenr; i++) { + struct dt_object *dto; + + dto = stripe[i]; + /* only create slave striped object */ + rc = dt_declare_create(env, dto, attr, NULL, NULL, th); + if (rc != 0) + GOTO(out_put, rc); + + if (!dt_try_as_dir(env, dto)) + GOTO(out_put, rc = -EINVAL); + + rc = dt_declare_insert(env, dto, + (const struct dt_rec *)lu_object_fid(&dto->do_lu), + (const struct dt_key *)dot, th); + if (rc != 0) + GOTO(out_put, rc); + + /* master stripe FID will be put to .. */ + rc = dt_declare_insert(env, dto, + (const struct dt_rec *)lu_object_fid(&dt->do_lu), + (const struct dt_key *)dotdot, th); + if (rc != 0) + GOTO(out_put, rc); + + /* probably nothing to inherite */ + if (lo->ldo_striping_cached && + !LOVEA_DELETE_VALUES(lo->ldo_def_stripe_size, + lo->ldo_def_stripenr, + lo->ldo_def_stripe_offset)) { + struct lod_thread_info *info; + struct lov_user_md_v3 *v3; + + /* sigh, lti_ea_store has been used for lmv_buf, + * so we have to allocate buffer for default + * stripe EA */ + OBD_ALLOC_PTR(v3); + if (v3 == NULL) + GOTO(out_put, rc = -ENOMEM); + + memset(v3, 0, sizeof(*v3)); + v3->lmm_magic = cpu_to_le32(LOV_USER_MAGIC_V3); + v3->lmm_stripe_count = + cpu_to_le32(lo->ldo_def_stripenr); + v3->lmm_stripe_offset = + cpu_to_le32(lo->ldo_def_stripe_offset); + v3->lmm_stripe_size = + cpu_to_le32(lo->ldo_def_stripe_size); + if (lo->ldo_pool) + strncpy(v3->lmm_pool_name, lo->ldo_pool, + LOV_MAXPOOLNAME); + + info = lod_env_info(env); + info->lti_buf.lb_buf = v3; + info->lti_buf.lb_len = sizeof(*v3); + rc = dt_declare_xattr_set(env, dto, + &info->lti_buf, + XATTR_NAME_LOV, + 0, th); + OBD_FREE_PTR(v3); + if (rc != 0) + GOTO(out_put, rc); + } + rc = dt_declare_xattr_set(env, dto, &lmv_buf, XATTR_NAME_LMV, 0, + th); + if (rc != 0) + GOTO(out_put, rc); + } + + rc = dt_declare_xattr_set(env, dt_object_child(dt), &lmv_buf, + XATTR_NAME_LMV, 0, th); + if (rc != 0) + GOTO(out_put, rc); + +out_put: + if (rc < 0) { + for (i = 0; i < stripe_count - 1; i++) + if (stripe[i] != NULL) + lu_object_put(env, &stripe[i]->do_lu); + OBD_FREE(stripe, sizeof(stripe[0]) * (stripe_count - 1)); + } + +out_free: + if (idx_array != NULL) + OBD_FREE(idx_array, sizeof(idx_array[0]) * stripe_count); + + RETURN(rc); +} + +/** + * Declare create striped md object. + */ +static int lod_declare_xattr_set_lmv(const struct lu_env *env, + struct dt_object *dt, + struct lu_attr *attr, + const struct lu_buf *lum_buf, + struct thandle *th) +{ + struct lod_object *lo = lod_dt_obj(dt); + struct lod_device *lod = lu2lod_dev(dt->do_lu.lo_dev); + struct lmv_user_md_v1 *lum; + int rc; + ENTRY; + + lum = lum_buf->lb_buf; + LASSERT(lum != NULL); + + CDEBUG(D_INFO, "lum magic = %x count = %u offset = %d\n", + le32_to_cpu(lum->lum_magic), le32_to_cpu(lum->lum_stripe_count), + (int)le32_to_cpu(lum->lum_stripe_offset)); + + if (le32_to_cpu(lum->lum_stripe_count) <= 1) + GOTO(out, rc = 0); + + rc = lod_verify_md_striping(lod, lum); + if (rc != 0) + GOTO(out, rc); + + /* prepare dir striped objects */ + rc = lod_prep_md_striped_create(env, dt, attr, lum, th); + if (rc != 0) { + /* failed to create striping, let's reset + * config so that others don't get confused */ + lod_object_free_striping(env, lo); + GOTO(out, rc); + } +out: + RETURN(rc); +} + /* * LOV xattr is a storage for striping, and LOD owns this xattr. * but LOD allows others to control striping to some extent @@ -436,7 +954,7 @@ static int lod_declare_xattr_set(const struct lu_env *env, * LU_XATTR_REPLACE is set to indicate a layout swap */ mode = dt->do_lu.lo_header->loh_attr & S_IFMT; - if ((S_ISREG(mode) || !mode) && !strcmp(name, XATTR_NAME_LOV) && + if ((S_ISREG(mode) || mode == 0) && strcmp(name, XATTR_NAME_LOV) == 0 && !(fl & LU_XATTR_REPLACE)) { /* * this is a request to manipulate object's striping @@ -451,15 +969,56 @@ static int lod_declare_xattr_set(const struct lu_env *env, attr->la_mode = S_IFREG; } rc = lod_declare_striped_object(env, dt, attr, buf, th); - if (rc) + } else if (S_ISDIR(mode)) { + struct lod_device *d = lu2lod_dev(dt->do_lu.lo_dev); + struct lod_object *lo = lod_dt_obj(dt); + int i; + + if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) { + struct lmv_user_md_v1 *lum; + + LASSERT(buf != NULL && buf->lb_buf != NULL); + lum = buf->lb_buf; + rc = lod_verify_md_striping(d, lum); + if (rc != 0) + RETURN(rc); + } + + rc = dt_declare_xattr_set(env, next, buf, name, fl, th); + if (rc != 0) + RETURN(rc); + + /* set xattr to each stripes, if needed */ + rc = lod_load_striping(env, lo); + if (rc != 0) RETURN(rc); - } - rc = dt_declare_xattr_set(env, next, buf, name, fl, th); + if (lo->ldo_stripenr == 0) + RETURN(rc); + + for (i = 0; i < lo->ldo_stripenr; i++) { + LASSERT(lo->ldo_stripe[i]); + rc = dt_declare_xattr_set(env, lo->ldo_stripe[i], buf, + name, fl, th); + if (rc != 0) + break; + } + } else { + rc = dt_declare_xattr_set(env, next, buf, name, fl, th); + } RETURN(rc); } +static void lod_lov_stripe_cache_clear(struct lod_object *lo) +{ + lo->ldo_striping_cached = 0; + lo->ldo_def_striping_set = 0; + lod_object_set_pool(lo, NULL); + lo->ldo_def_stripe_size = 0; + lo->ldo_def_stripenr = 0; +} + static int lod_xattr_set_lov_on_dir(const struct lu_env *env, struct dt_object *dt, const struct lu_buf *buf, @@ -475,15 +1034,12 @@ static int lod_xattr_set_lov_on_dir(const struct lu_env *env, int rc; ENTRY; - LASSERT(l->ldo_stripe == NULL); - l->ldo_striping_cached = 0; - l->ldo_def_striping_set = 0; - lod_object_set_pool(l, NULL); - l->ldo_def_stripe_size = 0; - l->ldo_def_stripenr = 0; - - LASSERT(buf); - LASSERT(buf->lb_buf); + /* If it is striped dir, we should clear the stripe cache for + * slave stripe as well, but there are no effective way to + * notify the LOD on the slave MDT, so we do not cache stripe + * information for slave stripe for now. XXX*/ + lod_lov_stripe_cache_clear(l); + LASSERT(buf != NULL && buf->lb_buf != NULL); lum = buf->lb_buf; rc = lod_verify_striping(d, buf, 0); @@ -517,24 +1073,161 @@ static int lod_xattr_set_lov_on_dir(const struct lu_env *env, RETURN(rc); } +static int lod_xattr_set_default_lmv_on_dir(const struct lu_env *env, + struct dt_object *dt, + const struct lu_buf *buf, + const char *name, int fl, + struct thandle *th, + struct lustre_capa *capa) +{ + struct dt_object *next = dt_object_child(dt); + struct lod_object *l = lod_dt_obj(dt); + 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", + le32_to_cpu(lum->lum_stripe_count), + (int)le32_to_cpu(lum->lum_stripe_offset)); + + 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) { + rc = dt_xattr_del(env, next, name, th, capa); + if (rc == -ENODATA) + rc = 0; + } else { + rc = dt_xattr_set(env, next, buf, name, fl, th, capa); + if (rc != 0) + RETURN(rc); + + /* Update default stripe cache */ + if (l->ldo_dir_stripe == NULL) { + OBD_ALLOC_PTR(l->ldo_dir_stripe); + if (l->ldo_dir_stripe == NULL) + RETURN(-ENOMEM); + } + + l->ldo_dir_striping_cached = 0; + l->ldo_dir_def_striping_set = 1; + l->ldo_dir_def_stripenr = + le32_to_cpu(lum->lum_stripe_count) - 1; + } + + RETURN(rc); +} + +static int lod_xattr_set_lmv(const struct lu_env *env, struct dt_object *dt, + const struct lu_buf *buf, const char *name, + int fl, struct thandle *th, + struct lustre_capa *capa) +{ + struct lod_object *lo = lod_dt_obj(dt); + struct lu_buf lmv_buf; + int i; + int rc; + ENTRY; + + if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) + RETURN(-ENOTDIR); + + /* The stripes are supposed to be allocated in declare phase, + * if there are no stripes being allocated, it will skip */ + if (lo->ldo_stripenr == 0) + RETURN(0); + + rc = lod_prep_lmv_md(env, dt, &lmv_buf); + if (rc != 0) + RETURN(rc); + + for (i = 0; i < lo->ldo_stripenr; i++) { + struct dt_object *dto; + struct lu_attr *attr = &lod_env_info(env)->lti_attr; + + dto = lo->ldo_stripe[i]; + memset(attr, 0, sizeof(*attr)); + attr->la_valid = LA_TYPE | LA_MODE; + attr->la_mode = S_IFDIR; + rc = dt_create(env, dto, attr, NULL, NULL, th); + if (rc != 0) + RETURN(rc); + + rc = dt_insert(env, dto, + (const struct dt_rec *)lu_object_fid(&dto->do_lu), + (const struct dt_key *)dot, th, capa, 0); + if (rc != 0) + RETURN(rc); + + rc = dt_insert(env, dto, + (struct dt_rec *)lu_object_fid(&dt->do_lu), + (const struct dt_key *)dotdot, th, capa, 0); + if (rc != 0) + RETURN(rc); + + if (lo->ldo_striping_cached && + !LOVEA_DELETE_VALUES(lo->ldo_def_stripe_size, + lo->ldo_def_stripenr, + lo->ldo_def_stripe_offset)) { + struct lod_thread_info *info; + struct lov_user_md_v3 *v3; + + /* sigh, lti_ea_store has been used for lmv_buf, + * so we have to allocate buffer for default + * stripe EA */ + OBD_ALLOC_PTR(v3); + if (v3 == NULL) + RETURN(-ENOMEM); + + memset(v3, 0, sizeof(*v3)); + v3->lmm_magic = cpu_to_le32(LOV_USER_MAGIC_V3); + v3->lmm_stripe_count = + cpu_to_le32(lo->ldo_def_stripenr); + v3->lmm_stripe_offset = + cpu_to_le32(lo->ldo_def_stripe_offset); + v3->lmm_stripe_size = + cpu_to_le32(lo->ldo_def_stripe_size); + if (lo->ldo_pool) + strncpy(v3->lmm_pool_name, lo->ldo_pool, + LOV_MAXPOOLNAME); + + info = lod_env_info(env); + info->lti_buf.lb_buf = v3; + info->lti_buf.lb_len = sizeof(*v3); + rc = dt_xattr_set(env, dto, &info->lti_buf, + XATTR_NAME_LOV, 0, th, capa); + OBD_FREE_PTR(v3); + if (rc != 0) + RETURN(rc); + } + + rc = dt_xattr_set(env, dto, &lmv_buf, XATTR_NAME_LMV, fl, th, + capa); + } + + rc = dt_xattr_set(env, dt_object_child(dt), &lmv_buf, XATTR_NAME_LMV, + fl, th, capa); + + RETURN(rc); +} + static int lod_xattr_set(const struct lu_env *env, struct dt_object *dt, const struct lu_buf *buf, const char *name, int fl, struct thandle *th, struct lustre_capa *capa) { + struct lod_object *lo = lod_dt_obj(dt); struct dt_object *next = dt_object_child(dt); __u32 attr; int rc; + int i; ENTRY; attr = dt->do_lu.lo_header->loh_attr & S_IFMT; - if (S_ISDIR(attr)) { - if (strncmp(name, XATTR_NAME_LOV, strlen(XATTR_NAME_LOV)) == 0) - rc = lod_xattr_set_lov_on_dir(env, dt, buf, name, - fl, th, capa); - else - rc = dt_xattr_set(env, next, buf, name, fl, th, capa); - + if (S_ISDIR(attr) && strcmp(name, XATTR_NAME_LOV) == 0) { + rc = lod_xattr_set_lov_on_dir(env, dt, buf, name, fl, th, capa); } else if (S_ISREG(attr) && !strcmp(name, XATTR_NAME_LOV)) { /* in case of lov EA swap, just set it * if not, it is a replay so check striping match what we @@ -548,7 +1241,11 @@ static int lod_xattr_set(const struct lu_env *env, } else { rc = lod_striping_create(env, dt, NULL, NULL, th); } - RETURN(rc); + } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) { + if (!S_ISDIR(attr)) + RETURN(-ENOTDIR); + rc = lod_xattr_set_default_lmv_on_dir(env, dt, buf, name, fl, + th, capa); } else { /* * behave transparantly for all other EAs @@ -556,6 +1253,20 @@ static int lod_xattr_set(const struct lu_env *env, rc = dt_xattr_set(env, next, buf, name, fl, th, capa); } + if (rc != 0 || !S_ISDIR(attr)) + RETURN(rc); + + if (lo->ldo_stripenr == 0) + RETURN(rc); + + for (i = 0; i < lo->ldo_stripenr; i++) { + LASSERT(lo->ldo_stripe[i]); + rc = dt_xattr_set(env, lo->ldo_stripe[i], buf, name, fl, th, + capa); + if (rc != 0) + break; + } + RETURN(rc); } @@ -570,6 +1281,8 @@ static int lod_xattr_del(const struct lu_env *env, struct dt_object *dt, const char *name, struct thandle *th, struct lustre_capa *capa) { + if (!strcmp(name, XATTR_NAME_LOV)) + lod_object_free_striping(env, lod_dt_obj(dt)); return dt_xattr_del(env, dt_object_child(dt), name, th, capa); } @@ -604,20 +1317,19 @@ static inline int lod_object_will_be_striped(int is_reg, const struct lu_fid *fi return (is_reg && fid_seq(fid) != FID_SEQ_LOCAL_FILE); } -static int lod_cache_parent_striping(const struct lu_env *env, - struct lod_object *lp) + +static int lod_cache_parent_lov_striping(const struct lu_env *env, + struct lod_object *lp) { + struct lod_thread_info *info = lod_env_info(env); struct lov_user_md_v1 *v1 = NULL; struct lov_user_md_v3 *v3 = NULL; int rc; ENTRY; - /* dt_ah_init() is called from MDD without parent being write locked + /* called from MDD without parent being write locked, * lock it here */ dt_write_lock(env, dt_object_child(&lp->ldo_obj), 0); - if (lp->ldo_striping_cached) - GOTO(unlock, rc = 0); - rc = lod_get_lov_ea(env, lp); if (rc < 0) GOTO(unlock, rc); @@ -632,7 +1344,8 @@ static int lod_cache_parent_striping(const struct lu_env *env, GOTO(unlock, rc = 0); } - v1 = (struct lov_user_md_v1 *)lod_env_info(env)->lti_ea_store; + rc = 0; + v1 = info->lti_ea_store; if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_V1)) lustre_swab_lov_user_md_v1(v1); else if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_V3)) @@ -644,23 +1357,63 @@ static int lod_cache_parent_striping(const struct lu_env *env, if (v1->lmm_pattern != LOV_PATTERN_RAID0 && v1->lmm_pattern != 0) GOTO(unlock, rc = 0); + CDEBUG(D_INFO, DFID" stripe_count=%d stripe_size=%d stripe_offset=%d\n", + PFID(lu_object_fid(&lp->ldo_obj.do_lu)), + (int)v1->lmm_stripe_count, + (int)v1->lmm_stripe_size, (int)v1->lmm_stripe_offset); + lp->ldo_def_stripenr = v1->lmm_stripe_count; lp->ldo_def_stripe_size = v1->lmm_stripe_size; lp->ldo_def_stripe_offset = v1->lmm_stripe_offset; lp->ldo_striping_cached = 1; lp->ldo_def_striping_set = 1; - 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]) lod_object_set_pool(lp, v3->lmm_pool_name); } + EXIT; +unlock: + dt_write_unlock(env, dt_object_child(&lp->ldo_obj)); + return rc; +} + + +static int lod_cache_parent_lmv_striping(const struct lu_env *env, + struct lod_object *lp) +{ + struct lod_thread_info *info = lod_env_info(env); + struct lmv_user_md_v1 *v1 = NULL; + int rc; + ENTRY; + + /* called from MDD without parent being write locked, + * lock it here */ + dt_write_lock(env, dt_object_child(&lp->ldo_obj), 0); + rc = lod_get_default_lmv_ea(env, lp); + if (rc < 0) + GOTO(unlock, rc); + + if (rc < sizeof(struct lmv_user_md)) { + /* don't lookup for non-existing or invalid striping */ + lp->ldo_dir_def_striping_set = 0; + lp->ldo_dir_striping_cached = 1; + lp->ldo_dir_def_stripenr = 0; + lp->ldo_dir_def_stripe_offset = + (typeof(v1->lum_stripe_offset))(-1); + lp->ldo_dir_def_hash_type = LMV_HASH_TYPE_FNV_1A_64; + GOTO(unlock, rc = 0); + } + + rc = 0; + v1 = info->lti_ea_store; - CDEBUG(D_OTHER, "def. striping: # %d, sz %d, off %d %s%s on "DFID"\n", - lp->ldo_def_stripenr, lp->ldo_def_stripe_size, - lp->ldo_def_stripe_offset, v3 ? "from " : "", - v3 ? lp->ldo_pool : "", PFID(lu_object_fid(&lp->ldo_obj.do_lu))); + lp->ldo_dir_def_stripenr = le32_to_cpu(v1->lum_stripe_count) - 1; + lp->ldo_dir_def_stripe_offset = le32_to_cpu(v1->lum_stripe_offset); + lp->ldo_dir_def_hash_type = le32_to_cpu(v1->lum_hash_type); + lp->ldo_dir_def_striping_set = 1; + lp->ldo_dir_striping_cached = 1; EXIT; unlock: @@ -668,6 +1421,31 @@ unlock: return rc; } +static int lod_cache_parent_striping(const struct lu_env *env, + struct lod_object *lp, + umode_t child_mode) +{ + int rc = 0; + ENTRY; + + rc = lod_load_striping(env, lp); + if (rc != 0) + RETURN(rc); + + if (!lp->ldo_striping_cached) { + /* we haven't tried to get default striping for + * the directory yet, let's cache it in the object */ + rc = lod_cache_parent_lov_striping(env, lp); + if (rc != 0) + RETURN(rc); + } + + if (S_ISDIR(child_mode) && !lp->ldo_dir_striping_cached) + rc = lod_cache_parent_lmv_striping(env, lp); + + RETURN(rc); +} + /** * used to transfer default striping data to the object being created */ @@ -675,7 +1453,7 @@ static void lod_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah, struct dt_object *parent, struct dt_object *child, - cfs_umode_t child_mode) + umode_t child_mode) { struct lod_device *d = lu2lod_dev(child->do_lu.lo_dev); struct dt_object *nextp = NULL; @@ -704,14 +1482,28 @@ static void lod_ah_init(const struct lu_env *env, * can be called with local object existing */ if (!dt_object_exists(nextc) || dt_object_remote(nextc)) - nextc->do_ops->do_ah_init(env, ah, nextp, nextc, child_mode); + nextc->do_ops->do_ah_init(env, ah, dt_object_remote(nextp) ? + NULL : nextp, nextc, child_mode); if (S_ISDIR(child_mode)) { - if (lp->ldo_striping_cached == 0) { - /* we haven't tried to get default striping for - * the directory yet, let's cache it in the object */ - lod_cache_parent_striping(env, lp); + int rc; + + if (lc->ldo_dir_stripe == NULL) { + OBD_ALLOC_PTR(lc->ldo_dir_stripe); + if (lc->ldo_dir_stripe == NULL) + return; + } + + if (lp->ldo_dir_stripe == NULL) { + OBD_ALLOC_PTR(lp->ldo_dir_stripe); + if (lp->ldo_dir_stripe == NULL) + return; } + + rc = lod_cache_parent_striping(env, lp, child_mode); + if (rc != 0) + return; + /* transfer defaults to new directory */ if (lp->ldo_striping_cached) { if (lp->ldo_pool) @@ -722,11 +1514,66 @@ static void lod_ah_init(const struct lu_env *env, lc->ldo_striping_cached = 1; lc->ldo_def_striping_set = 1; CDEBUG(D_OTHER, "inherite EA sz:%d off:%d nr:%d\n", - (int)lc->ldo_def_stripenr, (int)lc->ldo_def_stripe_size, - (int)lc->ldo_def_stripe_offset); + (int)lc->ldo_def_stripe_offset, + (int)lc->ldo_def_stripenr); } - return; + + /* transfer dir defaults to new directory */ + if (lp->ldo_dir_striping_cached) { + lc->ldo_dir_def_stripenr = lp->ldo_dir_def_stripenr; + lc->ldo_dir_def_stripe_offset = + lp->ldo_dir_def_stripe_offset; + lc->ldo_dir_def_hash_type = + lp->ldo_dir_def_hash_type; + lc->ldo_dir_striping_cached = 1; + lc->ldo_dir_def_striping_set = 1; + CDEBUG(D_INFO, "inherit default EA nr:%d off:%d t%u\n", + (int)lc->ldo_dir_def_stripenr, + (int)lc->ldo_dir_def_stripe_offset, + lc->ldo_dir_def_hash_type); + } + + /* If the directory is specified with certain stripes */ + if (ah->dah_eadata != NULL && ah->dah_eadata_len != 0) { + const struct lmv_user_md_v1 *lum1 = ah->dah_eadata; + int rc; + + rc = lod_verify_md_striping(d, lum1); + if (rc == 0 && + le32_to_cpu(lum1->lum_stripe_count) > 1) { + /* Directory will be striped only if + * stripe_count > 1 */ + lc->ldo_stripenr = + le32_to_cpu(lum1->lum_stripe_count) - 1; + lc->ldo_dir_stripe_offset = + le32_to_cpu(lum1->lum_stripe_offset); + lc->ldo_dir_hash_type = + le32_to_cpu(lum1->lum_hash_type); + CDEBUG(D_INFO, "set stripe EA nr:%hu off:%d\n", + lc->ldo_stripenr, + (int)lc->ldo_dir_stripe_offset); + } + } else if (lp->ldo_dir_def_striping_set) { + /* If there are default dir stripe from parent */ + lc->ldo_stripenr = lp->ldo_dir_def_stripenr; + lc->ldo_dir_stripe_offset = + lp->ldo_dir_def_stripe_offset; + lc->ldo_dir_hash_type = + lp->ldo_dir_def_hash_type; + CDEBUG(D_INFO, "inherit EA nr:%hu off:%d\n", + lc->ldo_stripenr, + (int)lc->ldo_dir_stripe_offset); + } else { + /* set default stripe for this directory */ + lc->ldo_stripenr = 0; + lc->ldo_dir_stripe_offset = -1; + } + + CDEBUG(D_INFO, "final striping count:%hu, offset:%d\n", + lc->ldo_stripenr, (int)lc->ldo_dir_stripe_offset); + + goto out; } /* @@ -736,17 +1583,12 @@ static void lod_ah_init(const struct lu_env *env, */ if (!lod_object_will_be_striped(S_ISREG(child_mode), lu_object_fid(&child->do_lu))) - return; - + goto out; /* * try from the parent */ if (likely(parent)) { - if (lp->ldo_striping_cached == 0) { - /* we haven't tried to get default striping for - * the directory yet, let's cache it in the object */ - lod_cache_parent_striping(env, lp); - } + lod_cache_parent_striping(env, lp, child_mode); lc->ldo_def_stripe_offset = (__u16) -1; @@ -774,6 +1616,12 @@ static void lod_ah_init(const struct lu_env *env, lc->ldo_stripenr, lc->ldo_stripe_size, lc->ldo_pool ? lc->ldo_pool : ""); +out: + /* we do not cache stripe information for slave stripe, see + * lod_xattr_set_lov_on_dir */ + if (lp != NULL && lp->ldo_dir_slave_stripe) + lod_lov_stripe_cache_clear(lp); + EXIT; } @@ -823,7 +1671,6 @@ static int lod_declare_init_size(const struct lu_env *env, RETURN(rc); } - /** * Create declaration of striped object */ @@ -844,22 +1691,33 @@ int lod_declare_striped_object(const struct lu_env *env, struct dt_object *dt, GOTO(out, rc = -ENOMEM); } - /* choose OST and generate appropriate objects */ - rc = lod_qos_prep_create(env, lo, attr, lovea, th); - if (rc) { - /* failed to create striping, let's reset - * config so that others don't get confused */ - lod_object_free_striping(env, lo); - GOTO(out, rc); - } + if (!dt_object_remote(next)) { + /* choose OST and generate appropriate objects */ + rc = lod_qos_prep_create(env, lo, attr, lovea, th); + if (rc) { + /* failed to create striping, let's reset + * config so that others don't get confused */ + lod_object_free_striping(env, lo); + GOTO(out, rc); + } - /* - * declare storage for striping data - */ - info->lti_buf.lb_len = lov_mds_md_size(lo->ldo_stripenr, + /* + * declare storage for striping data + */ + info->lti_buf.lb_len = lov_mds_md_size(lo->ldo_stripenr, lo->ldo_pool ? LOV_MAGIC_V3 : LOV_MAGIC_V1); - rc = dt_declare_xattr_set(env, next, &info->lti_buf, XATTR_NAME_LOV, - 0, th); + } else { + /* LOD can not choose OST objects for remote objects, i.e. + * stripes must be ready before that. Right now, it can only + * happen during migrate, i.e. migrate process needs to create + * remote regular file (mdd_migrate_create), then the migrate + * process will provide stripeEA. */ + LASSERT(lovea != NULL); + info->lti_buf = *lovea; + } + + rc = dt_declare_xattr_set(env, next, &info->lti_buf, + XATTR_NAME_LOV, 0, th); if (rc) GOTO(out, rc); @@ -875,6 +1733,146 @@ out: RETURN(rc); } +int lod_dir_striping_create_internal(const struct lu_env *env, + struct dt_object *dt, + struct lu_attr *attr, + const struct dt_object_format *dof, + struct thandle *th, + bool declare) +{ + struct lod_thread_info *info = lod_env_info(env); + struct dt_object *next = dt_object_child(dt); + struct lod_object *lo = lod_dt_obj(dt); + int rc; + ENTRY; + + if (lo->ldo_dir_def_striping_set && + !LMVEA_DELETE_VALUES(lo->ldo_stripenr, + lo->ldo_dir_stripe_offset)) { + struct lmv_user_md_v1 *v1 = info->lti_ea_store; + int stripe_count = lo->ldo_stripenr + 1; + + if (info->lti_ea_store_size < sizeof(*v1)) { + rc = lod_ea_store_resize(info, sizeof(*v1)); + if (rc != 0) + RETURN(rc); + v1 = info->lti_ea_store; + } + + memset(v1, 0, sizeof(*v1)); + v1->lum_magic = cpu_to_le32(LMV_USER_MAGIC); + v1->lum_stripe_count = cpu_to_le32(stripe_count); + v1->lum_stripe_offset = + cpu_to_le32(lo->ldo_dir_stripe_offset); + + info->lti_buf.lb_buf = v1; + info->lti_buf.lb_len = sizeof(*v1); + + if (declare) + rc = lod_declare_xattr_set_lmv(env, dt, attr, + &info->lti_buf, th); + else + rc = lod_xattr_set_lmv(env, dt, &info->lti_buf, + XATTR_NAME_LMV, 0, th, + BYPASS_CAPA); + if (rc != 0) + RETURN(rc); + } + + /* Transfer default LMV striping from the parent */ + if (lo->ldo_dir_striping_cached && + !LMVEA_DELETE_VALUES(lo->ldo_dir_def_stripenr, + lo->ldo_dir_def_stripe_offset)) { + struct lmv_user_md_v1 *v1 = info->lti_ea_store; + int def_stripe_count = lo->ldo_dir_def_stripenr + 1; + + if (info->lti_ea_store_size < sizeof(*v1)) { + rc = lod_ea_store_resize(info, sizeof(*v1)); + if (rc != 0) + RETURN(rc); + v1 = info->lti_ea_store; + } + + memset(v1, 0, sizeof(*v1)); + v1->lum_magic = cpu_to_le32(LMV_USER_MAGIC); + v1->lum_stripe_count = cpu_to_le32(def_stripe_count); + v1->lum_stripe_offset = + cpu_to_le32(lo->ldo_dir_def_stripe_offset); + v1->lum_hash_type = + cpu_to_le32(lo->ldo_dir_def_hash_type); + + info->lti_buf.lb_buf = v1; + info->lti_buf.lb_len = sizeof(*v1); + if (declare) + rc = dt_declare_xattr_set(env, next, &info->lti_buf, + XATTR_NAME_DEFAULT_LMV, 0, + th); + else + rc = dt_xattr_set(env, next, &info->lti_buf, + XATTR_NAME_DEFAULT_LMV, 0, th, + BYPASS_CAPA); + if (rc != 0) + RETURN(rc); + } + + /* Transfer default LOV striping from the parent */ + if (lo->ldo_striping_cached && + !LOVEA_DELETE_VALUES(lo->ldo_def_stripe_size, + lo->ldo_def_stripenr, + lo->ldo_def_stripe_offset)) { + struct lov_user_md_v3 *v3 = info->lti_ea_store; + + if (info->lti_ea_store_size < sizeof(*v3)) { + rc = lod_ea_store_resize(info, sizeof(*v3)); + if (rc != 0) + RETURN(rc); + v3 = info->lti_ea_store; + } + + memset(v3, 0, sizeof(*v3)); + v3->lmm_magic = cpu_to_le32(LOV_USER_MAGIC_V3); + v3->lmm_stripe_count = cpu_to_le16(lo->ldo_def_stripenr); + v3->lmm_stripe_offset = cpu_to_le16(lo->ldo_def_stripe_offset); + v3->lmm_stripe_size = cpu_to_le32(lo->ldo_def_stripe_size); + if (lo->ldo_pool) + strncpy(v3->lmm_pool_name, lo->ldo_pool, + LOV_MAXPOOLNAME); + + info->lti_buf.lb_buf = v3; + info->lti_buf.lb_len = sizeof(*v3); + + if (declare) + rc = dt_declare_xattr_set(env, next, &info->lti_buf, + XATTR_NAME_LOV, 0, th); + else + rc = dt_xattr_set(env, next, &info->lti_buf, + XATTR_NAME_LOV, 0, th, + BYPASS_CAPA); + if (rc != 0) + RETURN(rc); + } + + RETURN(0); +} + +static int lod_declare_dir_striping_create(const struct lu_env *env, + struct dt_object *dt, + struct lu_attr *attr, + struct dt_object_format *dof, + struct thandle *th) +{ + return lod_dir_striping_create_internal(env, dt, attr, dof, th, true); +} + +static int lod_dir_striping_create(const struct lu_env *env, + struct dt_object *dt, + struct lu_attr *attr, + struct dt_object_format *dof, + struct thandle *th) +{ + return lod_dir_striping_create_internal(env, dt, attr, dof, th, false); +} + static int lod_declare_object_create(const struct lu_env *env, struct dt_object *dt, struct lu_attr *attr, @@ -914,40 +1912,13 @@ static int lod_declare_object_create(const struct lu_env *env, if (lo->ldo_stripenr > 0) rc = lod_declare_striped_object(env, dt, attr, NULL, th); - } else if (dof->dof_type == DFT_DIR && lo->ldo_striping_cached) { - struct lod_thread_info *info = lod_env_info(env); - - struct lov_user_md_v3 *v3; - - if (LOVEA_DELETE_VALUES(lo->ldo_def_stripe_size, - lo->ldo_def_stripenr, - lo->ldo_def_stripe_offset)) - RETURN(0); - - OBD_ALLOC_PTR(v3); - if (v3 == NULL) - RETURN(-ENOMEM); - - v3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3); - v3->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0); - fid_to_ostid(lu_object_fid(&dt->do_lu), &v3->lmm_oi); - ostid_cpu_to_le(&v3->lmm_oi, &v3->lmm_oi); - v3->lmm_stripe_size = cpu_to_le32(lo->ldo_def_stripe_size); - v3->lmm_stripe_count = cpu_to_le32(lo->ldo_def_stripenr); - v3->lmm_stripe_offset = cpu_to_le16(lo->ldo_def_stripe_offset); - if (lo->ldo_pool) - strncpy(v3->lmm_pool_name, lo->ldo_pool, - LOV_MAXPOOLNAME); - - info->lti_buf.lb_buf = v3; - info->lti_buf.lb_len = sizeof(*v3); - - /* to transfer default striping from the parent */ - rc = dt_declare_xattr_set(env, next, &info->lti_buf, - XATTR_NAME_LOV, 0, th); - OBD_FREE_PTR(v3); + } else if (dof->dof_type == DFT_DIR) { + /* Orphan object (like migrating object) does not have + * lod_dir_stripe, see lod_ah_init */ + if (lo->ldo_dir_stripe != NULL) + rc = lod_declare_dir_striping_create(env, dt, attr, + dof, th); } - out: RETURN(rc); } @@ -960,8 +1931,6 @@ int lod_striping_create(const struct lu_env *env, struct dt_object *dt, int rc = 0, i; ENTRY; - LASSERT(lo->ldo_stripe); - LASSERT(lo->ldo_stripenr > 0); LASSERT(lo->ldo_striping_cached == 0); /* create all underlying objects */ @@ -992,9 +1961,10 @@ static int lod_object_create(const struct lu_env *env, struct dt_object *dt, rc = dt_create(env, next, attr, hint, dof, th); if (rc == 0) { - if (S_ISDIR(dt->do_lu.lo_header->loh_attr)) - rc = lod_store_def_striping(env, dt, th); - else if (lo->ldo_stripe) + if (S_ISDIR(dt->do_lu.lo_header->loh_attr) && + lo->ldo_dir_stripe != NULL) + rc = lod_dir_striping_create(env, dt, attr, dof, th); + else if (lo->ldo_stripe && dof->u.dof_reg.striped != 0) rc = lod_striping_create(env, dt, attr, dof, th); } @@ -1017,6 +1987,9 @@ static int lod_declare_object_destroy(const struct lu_env *env, if (rc) RETURN(rc); + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ)) + RETURN(0); + /* * load striping information, notice we don't do this when object * is being initialized as we don't need this information till @@ -1051,12 +2024,18 @@ static int lod_object_destroy(const struct lu_env *env, if (rc) RETURN(rc); + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ)) + RETURN(0); + /* destroy all underlying objects */ for (i = 0; i < lo->ldo_stripenr; i++) { LASSERT(lo->ldo_stripe[i]); - rc = dt_destroy(env, lo->ldo_stripe[i], th); - if (rc) - break; + /* for striped directory, next == ldo_stripe[0] */ + if (next != lo->ldo_stripe[i]) { + rc = dt_destroy(env, lo->ldo_stripe[i], th); + if (rc) + break; + } } RETURN(rc); @@ -1115,19 +2094,143 @@ static int lod_object_sync(const struct lu_env *env, struct dt_object *dt) return dt_object_sync(env, dt_object_child(dt)); } +struct lod_slave_locks { + int lsl_lock_count; + struct lustre_handle lsl_handle[0]; +}; + +static int lod_object_unlock_internal(const struct lu_env *env, + struct dt_object *dt, + struct ldlm_enqueue_info *einfo, + ldlm_policy_data_t *policy) +{ + struct lod_object *lo = lod_dt_obj(dt); + struct lod_slave_locks *slave_locks = einfo->ei_cbdata; + int rc = 0; + int i; + ENTRY; + + if (slave_locks == NULL) + RETURN(0); + + for (i = 0; i < slave_locks->lsl_lock_count; i++) { + if (lustre_handle_is_used(&slave_locks->lsl_handle[i])) { + int rc1; + + einfo->ei_cbdata = &slave_locks->lsl_handle[i]; + rc1 = dt_object_unlock(env, lo->ldo_stripe[i], einfo, + policy); + if (rc1 < 0) + rc = rc == 0 ? rc1 : rc; + } + } + + RETURN(rc); +} + +static int lod_object_unlock(const struct lu_env *env, struct dt_object *dt, + struct ldlm_enqueue_info *einfo, + union ldlm_policy_data *policy) +{ + struct lod_object *lo = lod_dt_obj(dt); + struct lod_slave_locks *slave_locks = einfo->ei_cbdata; + int slave_locks_size; + int rc; + ENTRY; + + if (slave_locks == NULL) + RETURN(0); + + rc = lod_load_striping(env, lo); + if (rc != 0) + RETURN(rc); + + /* Note: for remote lock for single stripe dir, MDT will cancel + * the lock by lockh directly */ + if (lo->ldo_stripenr == 0 && dt_object_remote(dt_object_child(dt))) + RETURN(0); + + if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) + RETURN(-ENOTDIR); + + /* Only cancel slave lock for striped dir */ + rc = lod_object_unlock_internal(env, dt, einfo, policy); + + slave_locks_size = sizeof(*slave_locks) + slave_locks->lsl_lock_count * + sizeof(slave_locks->lsl_handle[0]); + OBD_FREE(slave_locks, slave_locks_size); + einfo->ei_cbdata = NULL; + + RETURN(rc); +} + static int lod_object_lock(const struct lu_env *env, - struct dt_object *dt, struct lustre_handle *lh, + struct dt_object *dt, + struct lustre_handle *lh, struct ldlm_enqueue_info *einfo, - void *policy) + union ldlm_policy_data *policy) { - struct dt_object *next = dt_object_child(dt); - int rc; + struct lod_object *lo = lod_dt_obj(dt); + int rc = 0; + int i; + int slave_locks_size; + struct lod_slave_locks *slave_locks = NULL; ENTRY; - /* - * declare setattr on the local object - */ - rc = dt_object_lock(env, next, lh, einfo, policy); + /* remote object lock */ + if (!einfo->ei_enq_slave) { + LASSERT(dt_object_remote(dt)); + return dt_object_lock(env, dt_object_child(dt), lh, einfo, + policy); + } + + if (!S_ISDIR(dt->do_lu.lo_header->loh_attr)) + RETURN(-ENOTDIR); + + rc = lod_load_striping(env, lo); + if (rc != 0) + RETURN(rc); + + /* No stripes */ + if (lo->ldo_stripenr == 0) + RETURN(0); + + slave_locks_size = sizeof(*slave_locks) + lo->ldo_stripenr * + sizeof(slave_locks->lsl_handle[0]); + /* Freed in lod_object_unlock */ + OBD_ALLOC(slave_locks, slave_locks_size); + if (slave_locks == NULL) + RETURN(-ENOMEM); + slave_locks->lsl_lock_count = lo->ldo_stripenr; + + /* striped directory lock */ + for (i = 0; i < lo->ldo_stripenr; i++) { + struct lustre_handle lockh; + struct ldlm_res_id *res_id; + + 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); + einfo->ei_res_id = res_id; + + LASSERT(lo->ldo_stripe[i]); + rc = dt_object_lock(env, lo->ldo_stripe[i], &lockh, einfo, + policy); + if (rc != 0) + GOTO(out, rc); + + slave_locks->lsl_handle[i] = lockh; + } + + einfo->ei_cbdata = slave_locks; + +out: + if (rc != 0 && slave_locks != NULL) { + einfo->ei_cbdata = slave_locks; + lod_object_unlock_internal(env, dt, einfo, policy); + OBD_FREE(slave_locks, slave_locks_size); + einfo->ei_cbdata = NULL; + } RETURN(rc); } @@ -1160,6 +2263,7 @@ struct dt_object_operations lod_obj_ops = { .do_capa_get = lod_capa_get, .do_object_sync = lod_object_sync, .do_object_lock = lod_object_lock, + .do_object_unlock = lod_object_unlock, }; static ssize_t lod_read(const struct lu_env *env, struct dt_object *dt, @@ -1172,11 +2276,11 @@ static ssize_t lod_read(const struct lu_env *env, struct dt_object *dt, static ssize_t lod_declare_write(const struct lu_env *env, struct dt_object *dt, - const loff_t size, loff_t pos, + const struct lu_buf *buf, loff_t pos, struct thandle *th) { return dt_declare_record_write(env, dt_object_child(dt), - size, pos, th); + buf, pos, th); } static ssize_t lod_write(const struct lu_env *env, struct dt_object *dt, @@ -1194,23 +2298,57 @@ static const struct dt_body_operations lod_body_lnk_ops = { .dbo_write = lod_write }; -static int lod_object_init(const struct lu_env *env, struct lu_object *o, +static int lod_object_init(const struct lu_env *env, struct lu_object *lo, const struct lu_object_conf *conf) { - struct lod_device *d = lu2lod_dev(o->lo_dev); - struct lu_object *below; - struct lu_device *under; + struct lod_device *lod = lu2lod_dev(lo->lo_dev); + struct lu_device *cdev = NULL; + struct lu_object *cobj; + struct lod_tgt_descs *ltd = NULL; + struct lod_tgt_desc *tgt; + mdsno_t idx = 0; + int type = LU_SEQ_RANGE_ANY; + int rc; ENTRY; - /* - * create local object - */ - under = &d->lod_child->dd_lu_dev; - below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under); - if (below == NULL) + rc = lod_fld_lookup(env, lod, lu_object_fid(lo), &idx, &type); + if (rc != 0) + RETURN(rc); + + if (type == LU_SEQ_RANGE_MDT && + idx == lu_site2seq(lo->lo_dev->ld_site)->ss_node_id) { + cdev = &lod->lod_child->dd_lu_dev; + } else if (type == LU_SEQ_RANGE_MDT) { + ltd = &lod->lod_mdt_descs; + lod_getref(ltd); + } else if (type == LU_SEQ_RANGE_OST) { + ltd = &lod->lod_ost_descs; + lod_getref(ltd); + } else { + LBUG(); + } + + if (ltd != NULL) { + if (ltd->ltd_tgts_size > idx && + cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx)) { + tgt = LTD_TGT(ltd, idx); + + LASSERT(tgt != NULL); + LASSERT(tgt->ltd_tgt != NULL); + + cdev = &(tgt->ltd_tgt->dd_lu_dev); + } + lod_putref(lod, ltd); + } + + if (unlikely(cdev == NULL)) + RETURN(-ENOENT); + + cobj = cdev->ld_ops->ldo_object_alloc(env, lo->lo_header, cdev); + if (unlikely(cobj == NULL)) RETURN(-ENOMEM); - lu_object_add(o, below); + lu_object_add(lo, cobj); RETURN(0); } @@ -1219,6 +2357,11 @@ void lod_object_free_striping(const struct lu_env *env, struct lod_object *lo) { int i; + if (lo->ldo_dir_stripe != NULL) { + OBD_FREE_PTR(lo->ldo_dir_stripe); + lo->ldo_dir_stripe = NULL; + } + if (lo->ldo_stripe) { LASSERT(lo->ldo_stripes_allocated > 0); @@ -1233,6 +2376,7 @@ void lod_object_free_striping(const struct lu_env *env, struct lod_object *lo) lo->ldo_stripes_allocated = 0; } lo->ldo_stripenr = 0; + lo->ldo_pattern = 0; } /* @@ -1283,51 +2427,3 @@ struct lu_object_operations lod_lu_obj_ops = { .loo_object_release = lod_object_release, .loo_object_print = lod_object_print, }; - -/** - * Init remote lod object - */ -static int lod_robject_init(const struct lu_env *env, struct lu_object *lo, - const struct lu_object_conf *conf) -{ - struct lod_device *lod = lu2lod_dev(lo->lo_dev); - struct lod_tgt_descs *ltd = &lod->lod_mdt_descs; - struct lu_device *c_dev = NULL; - struct lu_object *c_obj; - int i; - ENTRY; - - lod_getref(ltd); - if (ltd->ltd_tgts_size > 0) { - cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) { - struct lod_tgt_desc *tgt; - tgt = LTD_TGT(ltd, i); - LASSERT(tgt && tgt->ltd_tgt); - if (tgt->ltd_index == - lu2lod_obj(lo)->ldo_mds_num) { - c_dev = &(tgt->ltd_tgt->dd_lu_dev); - break; - } - } - } - lod_putref(lod, ltd); - - if (unlikely(c_dev == NULL)) - RETURN(-ENOENT); - - c_obj = c_dev->ld_ops->ldo_object_alloc(env, lo->lo_header, c_dev); - if (unlikely(c_obj == NULL)) - RETURN(-ENOMEM); - - lu_object_add(lo, c_obj); - - RETURN(0); -} - -struct lu_object_operations lod_lu_robj_ops = { - .loo_object_init = lod_robject_init, - .loo_object_start = lod_object_start, - .loo_object_free = lod_object_free, - .loo_object_release = lod_object_release, - .loo_object_print = lod_object_print, -};