X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdd%2Fmdd_object.c;h=21eae5d93e2f575e2b9b35cf03ab1ef6cef46af0;hp=6504da2f786501e7abbafc280d99346bebeb2333;hb=38e6ee6fb779c34c7ef3eb4a010332385b3b84c1;hpb=b6a3222a0b35b939f40bb93087d8993e81885b26 diff --git a/lustre/mdd/mdd_object.c b/lustre/mdd/mdd_object.c index 6504da2..21eae5d 100644 --- a/lustre/mdd/mdd_object.c +++ b/lustre/mdd/mdd_object.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2013, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -50,38 +46,204 @@ /* fid_be_cpu(), fid_cpu_to_be(). */ #include #include -#include +#include #include -#include #include "mdd_internal.h" static const struct lu_object_operations mdd_lu_obj_ops; +struct mdd_object_user { + struct list_head mou_list; /**< linked off mod_users */ + u64 mou_open_flags; /**< open mode by client */ + __u64 mou_uidgid; /**< uid_gid on client */ + int mou_opencount; /**< # opened */ + ktime_t mou_deniednext; /**< time of next access denied + * notfication + */ +}; + static int mdd_xattr_get(const struct lu_env *env, struct md_object *obj, struct lu_buf *buf, const char *name); +static int mdd_changelog_data_store_by_fid(const struct lu_env *env, + struct mdd_device *mdd, + enum changelog_rec_type type, + enum changelog_rec_flags clf_flags, + const struct lu_fid *fid, + const char *xattr_name, + struct thandle *handle); + +static inline bool has_prefix(const char *str, const char *prefix); + + +static u32 flags_helper(u64 open_flags) +{ + u32 open_mode = 0; + + if (open_flags & MDS_FMODE_EXEC) { + open_mode = MDS_FMODE_EXEC; + } else { + if (open_flags & MDS_FMODE_READ) + open_mode = MDS_FMODE_READ; + if (open_flags & + (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND)) + open_mode |= MDS_FMODE_WRITE; + } + + return open_mode; +} + +/** Allocate/init a user and its sub-structures. + * + * \param flags [IN] + * \param uid [IN] + * \param gid [IN] + * \retval mou [OUT] success valid structure + * \retval mou [OUT] + */ +static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags, + uid_t uid, gid_t gid) +{ + struct mdd_object_user *mou; + + ENTRY; + + OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem); + if (mou == NULL) + RETURN(ERR_PTR(-ENOMEM)); + + mou->mou_open_flags = open_flags; + mou->mou_uidgid = ((__u64)uid << 32) | gid; + mou->mou_opencount = 0; + mou->mou_deniednext = ktime_set(0, 0); + + RETURN(mou); +} + +/** + * Free a user and its sub-structures. + * + * \param mou [IN] user to be freed. + */ +static void mdd_obj_user_free(struct mdd_object_user *mou) +{ + OBD_SLAB_FREE_PTR(mou, mdd_object_kmem); +} + +/** + * Find if UID/GID already has this file open + * + * Caller should have write-locked \param mdd_obj. + * \param mdd_obj [IN] mdd_obj + * \param uid [IN] client uid + * \param gid [IN] client gid + * \retval user pointer or NULL if not found + */ +static +struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj, + uid_t uid, gid_t gid, + u64 open_flags) +{ + struct mdd_object_user *mou; + __u64 uidgid; + + ENTRY; + + uidgid = ((__u64)uid << 32) | gid; + list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) { + if (mou->mou_uidgid == uidgid && + flags_helper(mou->mou_open_flags) == + flags_helper(open_flags)) + RETURN(mou); + } + RETURN(NULL); +} + +/** + * Add a user to the list of openers for this file + * + * Caller should have write-locked \param mdd_obj. + * \param mdd_obj [IN] mdd_obj + * \param mou [IN] user + * \retval 0 success + * \retval -ve failure + */ +static int mdd_obj_user_add(struct mdd_object *mdd_obj, + struct mdd_object_user *mou, + bool denied) +{ + struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj); + struct mdd_object_user *tmp; + __u32 uid = mou->mou_uidgid >> 32; + __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1); + + ENTRY; + tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags); + if (tmp != NULL) + RETURN(-EEXIST); + + list_add_tail(&mou->mou_list, &mdd_obj->mod_users); + + if (denied) + /* next 'access denied' notification cannot happen before + * mou_deniednext + */ + mou->mou_deniednext = + ktime_add(ktime_get(), + ktime_set(mdd->mdd_cl.mc_deniednext, 0)); + else + mou->mou_opencount++; + + RETURN(0); +} +/** + * Remove UID from the list + * + * Caller should have write-locked \param mdd_obj. + * \param mdd_obj [IN] mdd_obj + * \param uid [IN] user + * \retval -ve failure + */ +static int mdd_obj_user_remove(struct mdd_object *mdd_obj, + struct mdd_object_user *mou) +{ + ENTRY; + + if (mou == NULL) + RETURN(-ENOENT); + + list_del_init(&mou->mou_list); + + mdd_obj_user_free(mou); + + RETURN(0); +} int mdd_la_get(const struct lu_env *env, struct mdd_object *obj, - struct lu_attr *la, struct lustre_capa *capa) + struct lu_attr *la) { - if (mdd_object_exists(obj) == 0) { - CERROR("%s: object "DFID" not found: rc = -2\n", - mdd_obj_dev_name(obj), PFID(mdd_object_fid(obj))); - return -ENOENT; - } + int rc; + + if (mdd_object_exists(obj) == 0) + return -ENOENT; + + rc = mdo_attr_get(env, obj, la); + if (unlikely(rc != 0)) { + if (rc == -ENOENT) + obj->mod_flags |= DEAD_OBJ; + return rc; + } - return mdo_attr_get(env, obj, la, capa); + if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL) + obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ; + + return 0; } struct mdd_thread_info *mdd_env_info(const struct lu_env *env) { - struct mdd_thread_info *info; - - lu_env_refill((struct lu_env *)env); - info = lu_context_key_get(&env->le_ctx, &mdd_thread_key); - LASSERT(info != NULL); - return info; + return lu_env_info(env, &mdd_thread_key); } struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len) @@ -110,21 +272,21 @@ struct lu_object *mdd_object_alloc(const struct lu_env *env, struct lu_device *d) { struct mdd_object *mdd_obj; + struct lu_object *o; OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS); - if (mdd_obj != NULL) { - struct lu_object *o; - - o = mdd2lu_obj(mdd_obj); - lu_object_init(o, NULL, d); - mdd_obj->mod_obj.mo_ops = &mdd_obj_ops; - mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops; - mdd_obj->mod_count = 0; - o->lo_ops = &mdd_lu_obj_ops; - return o; - } else { + if (!mdd_obj) return NULL; - } + + o = mdd2lu_obj(mdd_obj); + lu_object_init(o, NULL, d); + mdd_obj->mod_obj.mo_ops = &mdd_obj_ops; + mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops; + mdd_obj->mod_count = 0; + o->lo_ops = &mdd_lu_obj_ops; + INIT_LIST_HEAD(&mdd_obj->mod_users); + + return o; } static int mdd_object_init(const struct lu_env *env, struct lu_object *o, @@ -136,7 +298,7 @@ static int mdd_object_init(const struct lu_env *env, struct lu_object *o, struct lu_device *under; ENTRY; - mdd_obj->mod_cltime = 0; + mdd_obj->mod_cltime = ktime_set(0, 0); under = &d->mdd_child->dd_lu_dev; below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under); if (IS_ERR(below)) @@ -155,7 +317,7 @@ static int mdd_object_start(const struct lu_env *env, struct lu_object *o) struct mdd_object *mdd_obj = lu2mdd_obj(o); struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start); - rc = mdd_la_get(env, mdd_obj, attr, BYPASS_CAPA); + rc = mdd_la_get(env, mdd_obj, attr); } return rc; @@ -163,20 +325,28 @@ static int mdd_object_start(const struct lu_env *env, struct lu_object *o) static void mdd_object_free(const struct lu_env *env, struct lu_object *o) { - struct mdd_object *mdd = lu2mdd_obj(o); + struct mdd_object *mdd = lu2mdd_obj(o); + struct mdd_object_user *mou, *tmp2; + + /* free user list */ + list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) { + list_del(&mou->mou_list); + mdd_obj_user_free(mou); + } - lu_object_fini(o); + lu_object_fini(o); OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem); } static int mdd_object_print(const struct lu_env *env, void *cookie, lu_printer_t p, const struct lu_object *o) { - struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o); - return (*p)(env, cookie, LUSTRE_MDD_NAME"-object@%p(open_count=%d, " - "valid=%x, cltime="LPU64", flags=%lx)", + struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o); + + return (*p)(env, cookie, + LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)", mdd, mdd->mod_count, mdd->mod_valid, - mdd->mod_cltime, mdd->mod_flags); + ktime_to_ns(mdd->mod_cltime), mdd->mod_flags); } static const struct lu_object_operations mdd_lu_obj_ops = { @@ -204,8 +374,7 @@ int mdd_attr_get(const struct lu_env *env, struct md_object *obj, ENTRY; - rc = mdd_la_get(env, mdd_obj, &ma->ma_attr, - mdd_object_capa(env, md2mdd_obj(obj))); + rc = mdd_la_get(env, mdd_obj, &ma->ma_attr); if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj)) ma->ma_attr.la_nlink = 0; @@ -219,16 +388,27 @@ static int mdd_xattr_get(const struct lu_env *env, struct md_object *obj, struct lu_buf *buf, const char *name) { - struct mdd_object *mdd_obj = md2mdd_obj(obj); - int rc; + struct mdd_object *mdd_obj = md2mdd_obj(obj); + struct mdd_device *mdd; + int rc; - ENTRY; + ENTRY; - if (mdd_object_exists(mdd_obj) == 0) { - CERROR("%s: object "DFID" not found: rc = -2\n", - mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj))); - return -ENOENT; - } + if (mdd_object_exists(mdd_obj) == 0) { + CERROR("%s: object "DFID" not found: rc = -2\n", + mdd_obj_dev_name(mdd_obj), + PFID(mdd_object_fid(mdd_obj))); + return -ENOENT; + } + + /* If the object has been destroyed, then do not get LMVEA, because + * it needs to load stripes from the iteration of the master object, + * and it will cause problem if master object has been destroyed, see + * LU-6427 */ + if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) && + !(mdd_obj->mod_flags & ORPHAN_OBJ) && + strcmp(name, XATTR_NAME_LMV) == 0)) + RETURN(-ENOENT); /* If the object has been delete from the namespace, then * get linkEA should return -ENOENT as well */ @@ -236,12 +416,47 @@ static int mdd_xattr_get(const struct lu_env *env, strcmp(name, XATTR_NAME_LINK) == 0)) RETURN(-ENOENT); - mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD); - rc = mdo_xattr_get(env, mdd_obj, buf, name, - mdd_object_capa(env, mdd_obj)); - mdd_read_unlock(env, mdd_obj); + mdd_read_lock(env, mdd_obj, DT_TGT_CHILD); + rc = mdo_xattr_get(env, mdd_obj, buf, name); + mdd_read_unlock(env, mdd_obj); - RETURN(rc); + mdd = mdo2mdd(obj); + + /* record only getting user xattrs and acls */ + if (rc >= 0 && buf->lb_buf && + mdd_changelog_enabled(env, mdd, CL_GETXATTR) && + (has_prefix(name, XATTR_USER_PREFIX) || + has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) || + has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) { + struct thandle *handle; + int rc2; + + LASSERT(mdo2fid(mdd_obj) != NULL); + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); + + rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL, + NULL, handle); + if (rc2) + GOTO(stop, rc2); + + rc2 = mdd_trans_start(env, mdd, handle); + if (rc2) + GOTO(stop, rc2); + + rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0, + mdo2fid(mdd_obj), name, + handle); + +stop: + rc2 = mdd_trans_stop(env, mdd, rc2, handle); + if (rc2) + rc = rc2; + } + + RETURN(rc); } /* @@ -267,9 +482,8 @@ int mdd_readlink(const struct lu_env *env, struct md_object *obj, LASSERT(next != NULL); LASSERT(next->do_body_ops != NULL); LASSERT(next->do_body_ops->dbo_read != NULL); - mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD); - rc = next->do_body_ops->dbo_read(env, next, buf, &pos, - mdd_object_capa(env, mdd_obj)); + mdd_read_lock(env, mdd_obj, DT_TGT_CHILD); + rc = dt_read(env, next, buf, &pos); mdd_read_unlock(env, mdd_obj); RETURN(rc); } @@ -285,14 +499,49 @@ static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj, ENTRY; - mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD); - rc = mdo_xattr_list(env, mdd_obj, buf, mdd_object_capa(env, mdd_obj)); + mdd_read_lock(env, mdd_obj, DT_TGT_CHILD); + rc = mdo_xattr_list(env, mdd_obj, buf); mdd_read_unlock(env, mdd_obj); - RETURN(rc); + /* If the buffer is NULL then we are only here to get the + * length of the xattr name list. */ + if (rc < 0 || buf->lb_buf == NULL) + RETURN(rc); + + /* + * Filter out XATTR_NAME_LINK if this is an orphan object. See + * mdd_xattr_get(). + */ + if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) { + char *end = (char *)buf->lb_buf + rc; + char *p = buf->lb_buf; + + while (p < end) { + char *next = p + strlen(p) + 1; + + if (strcmp(p, XATTR_NAME_LINK) == 0) { + if (end - next > 0) + memmove(p, next, end - next); + rc -= next - p; + CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK + " of orphan "DFID"\n", + PFID(mdd_object_fid(mdd_obj))); + break; + } + + p = next; + } + } + + RETURN(rc); } -int mdd_declare_object_create_internal(const struct lu_env *env, +int mdd_invalidate(const struct lu_env *env, struct md_object *obj) +{ + return mdo_invalidate(env, md2mdd_obj(obj)); +} + +int mdd_declare_create_object_internal(const struct lu_env *env, struct mdd_object *p, struct mdd_object *c, struct lu_attr *attr, @@ -300,15 +549,14 @@ int mdd_declare_object_create_internal(const struct lu_env *env, const struct md_op_spec *spec, struct dt_allocation_hint *hint) { - struct dt_object_format *dof = &mdd_env_info(env)->mti_dof; - const struct dt_index_features *feat = spec->sp_feat; - int rc; - ENTRY; + struct dt_object_format *dof = &mdd_env_info(env)->mti_dof; + const struct dt_index_features *feat = spec->sp_feat; + int rc; + ENTRY; if (feat != &dt_directory_features && feat != NULL) { - dof->dof_type = DFT_INDEX; + dof->dof_type = DFT_INDEX; dof->u.dof_idx.di_feat = feat; - } else { dof->dof_type = dt_mode_to_dft(attr->la_mode); if (dof->dof_type == DFT_REGULAR) { @@ -322,26 +570,24 @@ int mdd_declare_object_create_internal(const struct lu_env *env, } } - rc = mdo_declare_create_obj(env, c, attr, hint, dof, handle); + rc = mdo_declare_create_object(env, c, attr, hint, dof, handle); - RETURN(rc); + RETURN(rc); } -int mdd_object_create_internal(const struct lu_env *env, struct mdd_object *p, +int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p, struct mdd_object *c, struct lu_attr *attr, struct thandle *handle, const struct md_op_spec *spec, struct dt_allocation_hint *hint) { - struct dt_object_format *dof = &mdd_env_info(env)->mti_dof; - int rc; - ENTRY; + struct dt_object_format *dof = &mdd_env_info(env)->mti_dof; + int rc; + ENTRY; LASSERT(!mdd_object_exists(c)); - rc = mdo_create_obj(env, c, attr, hint, dof, handle); - - LASSERT(ergo(rc == 0, mdd_object_exists(c))); + rc = mdo_create_object(env, c, attr, hint, dof, handle); RETURN(rc); } @@ -350,15 +596,15 @@ int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj, const struct lu_attr *attr, struct thandle *handle, int needacl) { - int rc; - ENTRY; + int rc; + ENTRY; - rc = mdo_attr_set(env, obj, attr, handle, mdd_object_capa(env, obj)); -#ifdef CONFIG_FS_POSIX_ACL - if (!rc && (attr->la_valid & LA_MODE) && needacl) - rc = mdd_acl_chmod(env, obj, attr->la_mode, handle); + rc = mdo_attr_set(env, obj, attr, handle); +#ifdef CONFIG_LUSTRE_FS_POSIX_ACL + if (!rc && (attr->la_valid & LA_MODE) && needacl) + rc = mdd_acl_chmod(env, obj, attr->la_mode, handle); #endif - RETURN(rc); + RETURN(rc); } int mdd_update_time(const struct lu_env *env, struct mdd_object *obj, @@ -386,6 +632,22 @@ int mdd_update_time(const struct lu_env *env, struct mdd_object *obj, RETURN(rc); } + +static bool is_project_state_change(const struct lu_attr *oattr, + struct lu_attr *la) +{ + if (la->la_valid & LA_PROJID && + oattr->la_projid != la->la_projid) + return true; + + if ((la->la_valid & LA_FLAGS) && + (la->la_flags & LUSTRE_PROJINHERIT_FL) != + (oattr->la_flags & LUSTRE_PROJINHERIT_FL)) + return true; + + return false; +} + /* * This gives the same functionality as the code between * sys_chmod and inode_setattr @@ -395,10 +657,12 @@ int mdd_update_time(const struct lu_env *env, struct mdd_object *obj, */ static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj, const struct lu_attr *oattr, struct lu_attr *la, - const unsigned long flags) + const struct md_attr *ma) { struct lu_ucred *uc; int rc = 0; + const unsigned long flags = ma->ma_attr_flags; + ENTRY; if (!la->la_valid) @@ -414,12 +678,18 @@ static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj, LASSERT(oattr != NULL); - /* export destroy does not have ->le_ses, but we may want - * to drop LUSTRE_SOM_FL. */ uc = lu_ucred_check(env); if (uc == NULL) RETURN(0); + if (is_project_state_change(oattr, la)) { + if (!md_capable(uc, CFS_CAP_SYS_RESOURCE) && + !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) && + !(ma->ma_enable_chprojid_gid == -1 && + mdd_permission_internal(env, obj, oattr, MAY_WRITE))) + RETURN(-EPERM); + } + if (la->la_valid == LA_CTIME) { if (!(flags & MDS_PERM_BYPASS)) /* This is only for set ctime when rename's source is @@ -431,12 +701,19 @@ static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj, RETURN(rc); } - if (la->la_valid == LA_ATIME) { - /* This is atime only set for read atime update on close. */ - if (la->la_atime >= oattr->la_atime && - la->la_atime < (oattr->la_atime + + if (flags & MDS_CLOSE_UPDATE_TIMES && + la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) { + /* This is an atime/mtime/ctime attribute update for + * close RPCs. + */ + if (la->la_valid & LA_ATIME && + la->la_atime <= (oattr->la_atime + mdd_obj2mdd_dev(obj)->mdd_atime_diff)) la->la_valid &= ~LA_ATIME; + if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime) + la->la_valid &= ~LA_CTIME; + if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime) + la->la_valid &= ~LA_MTIME; RETURN(0); } @@ -458,7 +735,7 @@ static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj, RETURN(-EPERM); if (!S_ISDIR(oattr->la_mode)) - la->la_flags &= ~LUSTRE_DIRSYNC_FL; + la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL); } if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) && @@ -571,135 +848,205 @@ static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj, } } - /* For both Size-on-MDS case and truncate case, - * "la->la_valid & (LA_SIZE | LA_BLOCKS)" are ture. - * We distinguish them by "flags & MDS_SOM". - * For SOM case, it is true, the MAY_WRITE perm has been checked - * when open, no need check again. For truncate case, it is false, - * the MAY_WRITE perm should be checked here. */ - if (flags & MDS_SOM) { - /* For the "Size-on-MDS" setattr update, merge coming - * attributes with the set in the inode. BUG 10641 */ - if ((la->la_valid & LA_ATIME) && - (la->la_atime <= oattr->la_atime)) - la->la_valid &= ~LA_ATIME; - - /* OST attributes do not have a priority over MDS attributes, - * so drop times if ctime is equal. */ - if ((la->la_valid & LA_CTIME) && - (la->la_ctime <= oattr->la_ctime)) - la->la_valid &= ~(LA_MTIME | LA_CTIME); - } else { - if (la->la_valid & (LA_SIZE | LA_BLOCKS)) { - if (!((flags & MDS_OWNEROVERRIDE) && - (uc->uc_fsuid == oattr->la_uid)) && - !(flags & MDS_PERM_BYPASS)) { - rc = mdd_permission_internal(env, obj, - oattr, MAY_WRITE); - if (rc != 0) - RETURN(rc); - } - } - if (la->la_valid & LA_CTIME) { - /* The pure setattr, it has the priority over what is - * already set, do not drop it if ctime is equal. */ - if (la->la_ctime < oattr->la_ctime) - la->la_valid &= ~(LA_ATIME | LA_MTIME | - LA_CTIME); + if (la->la_valid & (LA_SIZE | LA_BLOCKS)) { + if (!((flags & MDS_OWNEROVERRIDE) && + (uc->uc_fsuid == oattr->la_uid)) && + !(flags & MDS_PERM_BYPASS)) { + rc = mdd_permission_internal(env, obj, oattr, + MAY_WRITE); + if (rc != 0) + RETURN(rc); } } + if (la->la_valid & LA_CTIME) { + /* The pure setattr, it has the priority over what is + * already set, do not drop it if ctime is equal. */ + if (la->la_ctime < oattr->la_ctime) + la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME); + } + RETURN(0); } +static int mdd_changelog_data_store_by_fid(const struct lu_env *env, + struct mdd_device *mdd, + enum changelog_rec_type type, + enum changelog_rec_flags clf_flags, + const struct lu_fid *fid, + const char *xattr_name, + struct thandle *handle) +{ + const struct lu_ucred *uc = lu_ucred(env); + enum changelog_rec_extra_flags xflags = CLFE_INVALID; + struct llog_changelog_rec *rec; + struct lu_buf *buf; + int reclen; + int rc; + + clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS; + + if (uc) { + if (uc->uc_jobid[0] != '\0') + clf_flags |= CLF_JOBID; + xflags |= CLFE_UIDGID; + xflags |= CLFE_NID; + } + if (type == CL_OPEN || type == CL_DN_OPEN) + xflags |= CLFE_OPEN; + if (type == CL_SETXATTR || type == CL_GETXATTR) + xflags |= CLFE_XATTR; + + reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ + + changelog_rec_offset(clf_flags & CLF_SUPPORTED, + xflags & CLFE_SUPPORTED)); + buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen); + if (buf->lb_buf == NULL) + RETURN(-ENOMEM); + rec = buf->lb_buf; + + rec->cr_hdr.lrh_len = reclen; + rec->cr.cr_flags = clf_flags; + rec->cr.cr_type = (__u32)type; + rec->cr.cr_tfid = *fid; + rec->cr.cr_namelen = 0; + + if (clf_flags & CLF_JOBID) + mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid); + + if (clf_flags & CLF_EXTRA_FLAGS) { + mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags); + if (xflags & CLFE_UIDGID) + mdd_changelog_rec_extra_uidgid(&rec->cr, + uc->uc_uid, uc->uc_gid); + if (xflags & CLFE_NID) + mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid); + if (xflags & CLFE_OPEN) + mdd_changelog_rec_extra_omode(&rec->cr, clf_flags); + if (xflags & CLFE_XATTR) { + if (xattr_name == NULL) + RETURN(-EINVAL); + mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name); + } + } + + rc = mdd_changelog_store(env, mdd, rec, handle); + RETURN(rc); +} + + /** Store a data change changelog record * If this fails, we must fail the whole transaction; we don't * want the change to commit without the log entry. * \param mdd_obj - mdd_object of change - * \param handle - transacion handle + * \param handle - transaction handle */ int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd, - enum changelog_rec_type type, int flags, + enum changelog_rec_type type, + enum changelog_rec_flags clf_flags, struct mdd_object *mdd_obj, struct thandle *handle) { - const struct lu_ucred *uc = lu_ucred(env); - const struct lu_fid *tfid; - struct llog_changelog_rec *rec; - struct lu_buf *buf; - int reclen; int rc; - /* Not recording */ - if (!(mdd->mdd_cl.mc_flags & CLM_ON)) - RETURN(0); - if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0) - RETURN(0); + LASSERT(mdd_obj != NULL); + LASSERT(handle != NULL); + + if (!mdd_changelog_enabled(env, mdd, type)) + RETURN(0); + + if (mdd_is_volatile_obj(mdd_obj)) + RETURN(0); - LASSERT(mdd_obj != NULL); - LASSERT(handle != NULL); + if ((type >= CL_MTIME) && (type <= CL_ATIME) && + ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) { + /* Don't need multiple updates in this log */ + /* Don't check under lock - no big deal if we get an extra + entry */ + RETURN(0); + } - tfid = mdo2fid(mdd_obj); + rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags, + mdo2fid(mdd_obj), NULL, handle); + if (rc == 0) + mdd_obj->mod_cltime = ktime_get(); - if ((type >= CL_MTIME) && (type <= CL_ATIME) && - cfs_time_before_64(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) { - /* Don't need multiple updates in this log */ - /* Don't check under lock - no big deal if we get an extra - entry */ - RETURN(0); - } + RETURN(rc); +} - flags = (flags & CLF_FLAGMASK) | CLF_VERSION; - if (uc->uc_jobid[0] != '\0') - flags |= CLF_JOBID; +int mdd_changelog_data_store_xattr(const struct lu_env *env, + struct mdd_device *mdd, + enum changelog_rec_type type, + enum changelog_rec_flags clf_flags, + struct mdd_object *mdd_obj, + const char *xattr_name, + struct thandle *handle) +{ + int rc; - reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED)); - buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen); - if (buf->lb_buf == NULL) - RETURN(-ENOMEM); - rec = buf->lb_buf; + LASSERT(mdd_obj != NULL); + LASSERT(handle != NULL); - rec->cr.cr_flags = flags; - rec->cr.cr_type = (__u32)type; - rec->cr.cr_tfid = *tfid; - rec->cr.cr_namelen = 0; - mdd_obj->mod_cltime = cfs_time_current_64(); + if (!mdd_changelog_enabled(env, mdd, type)) + RETURN(0); - if (flags & CLF_JOBID) - mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid); + if (mdd_is_volatile_obj(mdd_obj)) + RETURN(0); - rc = mdd_changelog_store(env, mdd, rec, handle); + if ((type >= CL_MTIME) && (type <= CL_ATIME) && + ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) { + /* Don't need multiple updates in this log */ + /* Don't check under lock - no big deal if we get an extra + * entry + */ + RETURN(0); + } + + rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags, + mdo2fid(mdd_obj), xattr_name, + handle); + if (rc == 0) + mdd_obj->mod_cltime = ktime_get(); RETURN(rc); } +/* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record, + * except for open flags have a dedicated record to store 32 bits of flags */ static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type, - int flags, struct md_object *obj) + enum changelog_rec_flags clf_flags, + struct md_device *m, const struct lu_fid *fid) { - struct thandle *handle; - struct mdd_object *mdd_obj = md2mdd_obj(obj); - struct mdd_device *mdd = mdo2mdd(obj); - int rc; - ENTRY; + struct thandle *handle; + struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev); + int rc; + ENTRY; + + LASSERT(fid != NULL); - handle = mdd_trans_create(env, mdd); - if (IS_ERR(handle)) + /* We'll check this again below, but we check now before we + * start a transaction. */ + if (!mdd_changelog_enabled(env, mdd, type)) + RETURN(0); + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) RETURN(PTR_ERR(handle)); - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); + rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle); if (rc) GOTO(stop, rc); - rc = mdd_trans_start(env, mdd, handle); - if (rc) - GOTO(stop, rc); + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(stop, rc); - rc = mdd_changelog_data_store(env, mdd, type, flags, mdd_obj, - handle); + rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags, + fid, NULL, handle); stop: - mdd_trans_stop(env, mdd, rc, handle); + rc = mdd_trans_stop(env, mdd, rc, handle); - RETURN(rc); + RETURN(rc); } /** @@ -717,26 +1064,26 @@ static int mdd_attr_set_changelog(const struct lu_env *env, struct md_object *obj, struct thandle *handle, __u64 valid) { - struct mdd_device *mdd = mdo2mdd(obj); - int bits, type = 0; + struct mdd_device *mdd = mdo2mdd(obj); + int bits, type = 0; bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0; bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0; - bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0; - bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0; - bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0; - bits = bits & mdd->mdd_cl.mc_mask; + bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0; + bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0; + bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0; + bits = bits & mdd->mdd_cl.mc_mask; /* This is an implementation limit rather than a protocol limit */ - CLASSERT(CL_LAST <= sizeof(int) * 8); - if (bits == 0) - return 0; + BUILD_BUG_ON(CL_LAST > sizeof(int) * 8); + if (bits == 0) + return 0; - /* The record type is the lowest non-masked set bit */ + /* The record type is the lowest non-masked set bit */ type = __ffs(bits); - /* FYI we only store the first CLF_FLAGMASK bits of la_valid */ - return mdd_changelog_data_store(env, mdd, type, (int)valid, - md2mdd_obj(obj), handle); + /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */ + return mdd_changelog_data_store(env, mdd, type, valid, + md2mdd_obj(obj), handle); } static int mdd_declare_attr_set(const struct lu_env *env, @@ -751,11 +1098,11 @@ static int mdd_declare_attr_set(const struct lu_env *env, if (rc) return rc; -#ifdef CONFIG_FS_POSIX_ACL +#ifdef CONFIG_LUSTRE_FS_POSIX_ACL if (attr->la_valid & LA_MODE) { - mdd_read_lock(env, obj, MOR_TGT_CHILD); + mdd_read_lock(env, obj, DT_TGT_CHILD); rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, - XATTR_NAME_ACL_ACCESS, BYPASS_CAPA); + XATTR_NAME_ACL_ACCESS); mdd_read_unlock(env, obj); if (rc == -EOPNOTSUPP || rc == -ENODATA) rc = 0; @@ -773,12 +1120,14 @@ static int mdd_declare_attr_set(const struct lu_env *env, } #endif - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); + rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL, + handle); return rc; } /* * LU-3671 + * LU-7239 * * permission changes may require sync operation, to mitigate performance * impact, only do this for dir and when permission is reduced. @@ -793,91 +1142,186 @@ static inline bool permission_needs_sync(const struct lu_attr *old, if (!S_ISDIR(old->la_mode)) return false; - if (new->la_valid & (LA_UID | LA_GID)) + if (new->la_valid & LA_UID && old->la_uid != new->la_uid) return true; - if (new->la_valid & LA_MODE && - new->la_mode & (S_ISUID | S_ISGID | S_ISVTX)) + if (new->la_valid & LA_GID && old->la_gid != new->la_gid) return true; - if ((new->la_valid & LA_MODE) && - ((new->la_mode & old->la_mode) & S_IRWXUGO) != - (old->la_mode & S_IRWXUGO)) - return true; + if (new->la_valid & LA_MODE) { + /* turned on sticky bit */ + if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX)) + return true; + + /* set-GID has no impact on what is allowed, not checked */ + + /* turned off setuid bit, or one of rwx for someone */ + if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) != + (old->la_mode & (0777 | S_ISUID))) + return true; + } return false; } +static inline __u64 mdd_lmm_dom_size(void *buf) +{ + struct lov_mds_md *lmm = buf; + struct lov_comp_md_v1 *comp_v1; + struct lov_mds_md *v1; + __u32 off; + + if (lmm == NULL) + return 0; + + if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1) + return 0; + + comp_v1 = (struct lov_comp_md_v1 *)lmm; + off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset); + v1 = (struct lov_mds_md *)((char *)comp_v1 + off); + + /* DoM entry is the first entry always */ + if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT) + return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end); + + return 0; +} + /* set attr and LOV EA at once, return updated attr */ int mdd_attr_set(const struct lu_env *env, struct md_object *obj, const struct md_attr *ma) { struct mdd_object *mdd_obj = md2mdd_obj(obj); struct mdd_device *mdd = mdo2mdd(obj); - struct thandle *handle; + struct thandle *handle = NULL; struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix; struct lu_attr *attr = MDD_ENV_VAR(env, cattr); const struct lu_attr *la = &ma->ma_attr; + struct lu_ucred *uc; + bool chrgrp_by_unprivileged_user = false; int rc; ENTRY; - /* we do not use ->attr_set() for LOV/SOM/HSM EA any more */ + /* we do not use ->attr_set() for LOV/HSM EA any more */ LASSERT((ma->ma_valid & MA_LOV) == 0); LASSERT((ma->ma_valid & MA_HSM) == 0); - LASSERT((ma->ma_valid & MA_SOM) == 0); - rc = mdd_la_get(env, mdd_obj, attr, BYPASS_CAPA); + rc = mdd_la_get(env, mdd_obj, attr); if (rc) RETURN(rc); *la_copy = ma->ma_attr; - rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags); + rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma); if (rc) RETURN(rc); - /* setattr on "close" only change atime, or do nothing */ - if (la->la_valid == LA_ATIME && la_copy->la_valid == 0) - RETURN(0); - - handle = mdd_trans_create(env, mdd); - if (IS_ERR(handle)) - RETURN(PTR_ERR(handle)); - - rc = mdd_declare_attr_set(env, mdd, mdd_obj, la, handle); - if (rc) - GOTO(stop, rc); + /* no need to setattr anymore */ + if (la_copy->la_valid == 0) { + CDEBUG(D_INODE, + "%s: no valid attribute on "DFID", previous valid is %#llx\n", + mdd2obd_dev(mdd)->obd_name, + PFID(mdo2fid(mdd_obj)), la->la_valid); - rc = mdd_trans_start(env, mdd, handle); - if (rc) - GOTO(stop, rc); + RETURN(0); + } - if (mdd->mdd_sync_permission && permission_needs_sync(attr, la)) - handle->th_sync = 1; + /* If an unprivileged user changes group of some file, + * the setattr operation will be processed synchronously to + * honor the quota limit of the corresponding group. see LU-5152 */ + uc = lu_ucred_check(env); + if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID && + la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) { + /* LU-10048: disable synchronous chgrp operation for it will + * cause deadlock between MDT and OST. + la_copy->la_valid |= LA_FLAGS; + la_copy->la_flags |= LUSTRE_SET_SYNC_FL; + */ + chrgrp_by_unprivileged_user = true; + + /* Flush the possible existing client setattr requests to OSTs + * to keep the order with the current setattr operation that + * will be sent directly to OSTs. see LU-5152 */ + /* LU-11303 disable sync as this is too heavyweight. + * This should be replaced with a sync only for the object + * being modified here, not the whole filesystem. + rc = dt_sync(env, mdd->mdd_child); + if (rc) + GOTO(out, rc); + */ + } + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) { + rc = PTR_ERR(handle); + handle = NULL; + + GOTO(out, rc); + } + + rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(out, rc); + + if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission && + permission_needs_sync(attr, la)) + handle->th_sync = 1; if (la->la_valid & (LA_MTIME | LA_CTIME)) - CDEBUG(D_INODE, "setting mtime "LPU64", ctime "LPU64"\n", + CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n", la->la_mtime, la->la_ctime); - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); - if (la_copy->la_valid & LA_FLAGS) - rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1); - else if (la_copy->la_valid) /* setattr */ + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); + + /* LU-10509: setattr of LA_SIZE should be skipped case of DOM, + * otherwise following truncate will do nothing and truncated + * data may be read again. This is a quick fix until LU-11033 + * will be resolved. + */ + if (la_copy->la_valid & LA_SIZE) { + struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0); + + rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV); + if (rc) { + rc = 0; + } else { + if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0) + la_copy->la_valid &= ~LA_SIZE; + lu_buf_free(lov_buf); + } + } + + if (la_copy->la_valid) { rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1); + + if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) { + /* rollback to the original gid */ + la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL; + la_copy->la_gid = attr->la_gid; + mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1); + } + } mdd_write_unlock(env, mdd_obj); +out: if (rc == 0) - rc = mdd_attr_set_changelog(env, obj, handle, la->la_valid); + rc = mdd_attr_set_changelog(env, obj, handle, + la_copy->la_valid); - GOTO(stop, rc); + if (handle != NULL) + rc = mdd_trans_stop(env, mdd, rc, handle); -stop: - mdd_trans_stop(env, mdd, rc, handle); return rc; } static int mdd_xattr_sanity_check(const struct lu_env *env, struct mdd_object *obj, - const struct lu_attr *attr) + const struct lu_attr *attr, + const char *name) { struct lu_ucred *uc = lu_ucred_assert(env); ENTRY; @@ -885,12 +1329,70 @@ static int mdd_xattr_sanity_check(const struct lu_env *env, if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL)) RETURN(-EPERM); - if ((uc->uc_fsuid != attr->la_uid) && !md_capable(uc, CFS_CAP_FOWNER)) + if (strncmp(XATTR_USER_PREFIX, name, + sizeof(XATTR_USER_PREFIX) - 1) == 0) { + /* For sticky directories, only the owner and privileged user + * can write attributes. */ + if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) && + (uc->uc_fsuid != attr->la_uid) && + !md_capable(uc, CFS_CAP_FOWNER)) + RETURN(-EPERM); + } else if (strcmp(name, XATTR_NAME_SOM) != 0 && + (uc->uc_fsuid != attr->la_uid) && + !md_capable(uc, CFS_CAP_FOWNER)) { RETURN(-EPERM); + } RETURN(0); } +/** + * Check if a string begins with a given prefix. + * + * \param str String to check + * \param prefix Substring to check at the beginning of \a str + * \return true/false whether the condition is verified. + */ +static inline bool has_prefix(const char *str, const char *prefix) +{ + return strncmp(prefix, str, strlen(prefix)) == 0; +} + +/** + * Indicate the kind of changelog to store (if any) for a xattr set/del. + * + * \param[in] xattr_name Full extended attribute name. + * + * \return type of changelog to use, or CL_NONE if no changelog is to be emitted + */ +static enum changelog_rec_type +mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd, + const char *xattr_name) +{ + /* Layout changes systematically recorded */ + if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 || + strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 || + allowed_lustre_lov(xattr_name)) + return CL_LAYOUT; + + /* HSM information changes systematically recorded */ + if (strcmp(XATTR_NAME_HSM, xattr_name) == 0) + return CL_HSM; + + /* Avoid logging SOM xattr for every file */ + if (strcmp(XATTR_NAME_SOM, xattr_name) == 0) + return CL_NONE; + + if (has_prefix(xattr_name, XATTR_USER_PREFIX) || + has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) || + has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) || + has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) || + has_prefix(xattr_name, XATTR_SECURITY_PREFIX)) + return CL_SETXATTR; + + return CL_NONE; +} + static int mdd_declare_xattr_set(const struct lu_env *env, struct mdd_device *mdd, struct mdd_object *obj, @@ -898,30 +1400,18 @@ static int mdd_declare_xattr_set(const struct lu_env *env, const char *name, int fl, struct thandle *handle) { - int rc; + enum changelog_rec_type type; + int rc; rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle); if (rc) return rc; - /* Only record user and layout xattr changes */ - if (strncmp(XATTR_USER_PREFIX, name, - sizeof(XATTR_USER_PREFIX) - 1) == 0 || - strcmp(XATTR_NAME_LOV, name) == 0) { - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); - if (rc) - return rc; - } + type = mdd_xattr_changelog_type(env, mdd, name); + if (type < 0) + return 0; /* no changelog to store */ - /* If HSM data is modified, this could add a changelog */ - if (strcmp(XATTR_NAME_HSM, name) == 0) { - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); - if (rc) - return rc; - } - - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); - return rc; + return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle); } /* @@ -936,27 +1426,26 @@ static int mdd_declare_xattr_set(const struct lu_env *env, static int mdd_hsm_update_locked(const struct lu_env *env, struct md_object *obj, const struct lu_buf *buf, - struct thandle *handle) + struct thandle *handle, + enum changelog_rec_flags *clf_flags) { struct mdd_thread_info *info = mdd_env_info(env); - struct mdd_device *mdd = mdo2mdd(obj); - struct mdd_object *mdd_obj = md2mdd_obj(obj); - struct lu_buf *current_buf; - struct md_hsm *current_mh; - struct md_hsm *new_mh; - int rc; - ENTRY; + struct mdd_object *mdd_obj = md2mdd_obj(obj); + struct lu_buf *current_buf; + struct md_hsm *current_mh; + struct md_hsm *new_mh; + int rc; + ENTRY; OBD_ALLOC_PTR(current_mh); if (current_mh == NULL) RETURN(-ENOMEM); /* Read HSM attrs from disk */ - CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf)); - current_buf = mdd_buf_get(env, info->mti_xattr_buf, - sizeof(info->mti_xattr_buf)); - rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM, - mdd_object_capa(env, mdd_obj)); + current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf, + MIN(mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size, + XATTR_SIZE_MAX)); + rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM); rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh); if (rc < 0 && rc != -ENODATA) GOTO(free, rc); @@ -969,27 +1458,424 @@ static int mdd_hsm_update_locked(const struct lu_env *env, GOTO(free, rc = -ENOMEM); lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh); - /* If HSM flags are different, add a changelog */ rc = 0; + + /* Flags differ, set flags for the changelog that will be added */ if (current_mh->mh_flags != new_mh->mh_flags) { - int flags = 0; - hsm_set_cl_event(&flags, HE_STATE); + hsm_set_cl_event(clf_flags, HE_STATE); if (new_mh->mh_flags & HS_DIRTY) - hsm_set_cl_flags(&flags, CLF_HSM_DIRTY); - - rc = mdd_changelog_data_store(env, mdd, CL_HSM, flags, mdd_obj, - handle); + hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY); } OBD_FREE_PTR(new_mh); + EXIT; free: OBD_FREE_PTR(current_mh); - return(rc); + return rc; } +static int mdd_object_pfid_replace(const struct lu_env *env, + struct mdd_object *o) +{ + struct mdd_device *mdd = mdo2mdd(&o->mod_obj); + struct thandle *handle; + int rc; + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); + + handle->th_complex = 1; + + /* it doesn't need to track the PFID update via llog, because LFSCK + * will repair it even it goes wrong */ + rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID, + 0, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc != 0) + GOTO(out, rc); + + rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle); + if (rc) + GOTO(out, rc); + +out: + mdd_trans_stop(env, mdd, rc, handle); + return rc; +} + + +static int mdd_declare_xattr_del(const struct lu_env *env, + struct mdd_device *mdd, + struct mdd_object *obj, + const char *name, + struct thandle *handle); + static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj, const char *name); +static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj, + struct md_object *md_vic) +{ + struct mdd_device *mdd = mdo2mdd(md_obj); + struct mdd_object *obj = md2mdd_obj(md_obj); + struct mdd_object *vic = md2mdd_obj(md_vic); + struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0]; + struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1]; + struct lov_mds_md *lmm; + struct thandle *handle; + int rc; + ENTRY; + + rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic)); + if (rc == 0) /* same fid */ + RETURN(-EPERM); + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); + + if (rc > 0) { + mdd_write_lock(env, obj, DT_TGT_CHILD); + mdd_write_lock(env, vic, DT_TGT_CHILD); + } else { + mdd_write_lock(env, vic, DT_TGT_CHILD); + mdd_write_lock(env, obj, DT_TGT_CHILD); + } + + /* get EA of victim file */ + memset(buf_vic, 0, sizeof(*buf_vic)); + rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV); + if (rc < 0) { + if (rc == -ENODATA) + rc = 0; + GOTO(out, rc); + } + + /* parse the layout of victim file */ + lmm = buf_vic->lb_buf; + if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1) + GOTO(out, rc = -EINVAL); + + /* save EA of target file for restore */ + memset(buf, 0, sizeof(*buf)); + rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV); + if (rc < 0) + GOTO(out, rc); + + /* Get rid of the layout from victim object */ + rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV, + LU_XATTR_MERGE, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc != 0) + GOTO(out, rc); + + rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE, + handle); + if (rc) + GOTO(out, rc); + + rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle); + if (rc) /* wtf? */ + GOTO(out_restore, rc); + + (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle); + (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle); + EXIT; + +out_restore: + if (rc) { + int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, + LU_XATTR_REPLACE, handle); + if (rc2) + CERROR("%s: failed to rollback of layout of: "DFID + ": %d, file state unknown\n", + mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2); + } + +out: + mdd_trans_stop(env, mdd, rc, handle); + mdd_write_unlock(env, obj); + mdd_write_unlock(env, vic); + lu_buf_free(buf); + lu_buf_free(buf_vic); + + if (!rc) + (void) mdd_object_pfid_replace(env, obj); + + return rc; +} + +/** + * Extract the mirror with specified mirror id, and store the splitted + * mirror layout to @buf. + * + * \param[in] comp_v1 mirrored layout + * \param[in] mirror_id the mirror with mirror_id to be extracted + * \param[out] buf store the layout excluding the extracted mirror, + * caller free the buffer we allocated in this function + * \param[out] buf_vic store the extracted layout, caller free the buffer + * we allocated in this function + * + * \retval 0 on success; < 0 if error happens + */ +static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id, + struct lu_buf *buf, struct lu_buf *buf_vic) +{ + struct lov_comp_md_v1 *comp_rem; + struct lov_comp_md_v1 *comp_vic; + struct lov_comp_md_entry_v1 *entry; + struct lov_comp_md_entry_v1 *entry_rem; + struct lov_comp_md_entry_v1 *entry_vic; + __u16 mirror_cnt; + __u16 comp_cnt, count = 0; + int lmm_size, lmm_size_vic = 0; + int i, j, k; + int offset, offset_rem, offset_vic; + + mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1; + /* comp_v1 should contains more than 1 mirror */ + if (mirror_cnt <= 1) + return -EINVAL; + comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count); + lmm_size = le32_to_cpu(comp_v1->lcm_size); + + for (i = 0; i < comp_cnt; i++) { + entry = &comp_v1->lcm_entries[i]; + if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) { + count++; + lmm_size_vic += sizeof(*entry); + lmm_size_vic += le32_to_cpu(entry->lcme_size); + } else if (count > 0) { + /* find the specified mirror */ + break; + } + } + + if (count == 0) + return -EINVAL; + + lu_buf_alloc(buf, lmm_size - lmm_size_vic); + if (!buf->lb_buf) + return -ENOMEM; + + lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic); + if (!buf_vic->lb_buf) { + lu_buf_free(buf); + return -ENOMEM; + } + + comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf; + comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf; + + memcpy(comp_rem, comp_v1, sizeof(*comp_v1)); + comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2); + comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count); + comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic); + if (!comp_rem->lcm_mirror_count) + comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE); + + memset(comp_vic, 0, sizeof(*comp_v1)); + comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1); + comp_vic->lcm_mirror_count = 0; + comp_vic->lcm_entry_count = cpu_to_le32(count); + comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic)); + comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE); + comp_vic->lcm_layout_gen = 0; + + offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt; + offset_rem = sizeof(*comp_rem) + + sizeof(*entry_rem) * (comp_cnt - count); + offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count; + for (i = j = k = 0; i < comp_cnt; i++) { + struct lov_mds_md *lmm, *lmm_dst; + bool vic = false; + + entry = &comp_v1->lcm_entries[i]; + entry_vic = &comp_vic->lcm_entries[j]; + entry_rem = &comp_rem->lcm_entries[k]; + + if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) + vic = true; + + /* copy component entry */ + if (vic) { + memcpy(entry_vic, entry, sizeof(*entry)); + entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT); + entry_vic->lcme_offset = cpu_to_le32(offset_vic); + j++; + } else { + memcpy(entry_rem, entry, sizeof(*entry)); + entry_rem->lcme_offset = cpu_to_le32(offset_rem); + k++; + } + + lmm = (struct lov_mds_md *)((char *)comp_v1 + offset); + if (vic) + lmm_dst = (struct lov_mds_md *) + ((char *)comp_vic + offset_vic); + else + lmm_dst = (struct lov_mds_md *) + ((char *)comp_rem + offset_rem); + + /* copy component entry blob */ + memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size)); + + /* blob offset advance */ + offset += le32_to_cpu(entry->lcme_size); + if (vic) + offset_vic += le32_to_cpu(entry->lcme_size); + else + offset_rem += le32_to_cpu(entry->lcme_size); + } + + return 0; +} + +static int mdd_dom_data_truncate(const struct lu_env *env, + struct mdd_device *mdd, struct mdd_object *mo); + +static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj, + struct md_rejig_data *mrd) +{ + struct mdd_device *mdd = mdo2mdd(md_obj); + struct mdd_object *obj = md2mdd_obj(md_obj); + struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj); + struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0]; + struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1]; + struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2]; + struct lov_comp_md_v1 *lcm; + struct thandle *handle; + int rc; + bool dom_stripe = false; + + ENTRY; + + rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic)); + if (rc == 0) /* same fid */ + RETURN(-EPERM); + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); + + if (rc > 0) { + mdd_write_lock(env, obj, DT_TGT_CHILD); + mdd_write_lock(env, vic, DT_TGT_CHILD); + } else { + mdd_write_lock(env, vic, DT_TGT_CHILD); + mdd_write_lock(env, obj, DT_TGT_CHILD); + } + + /* get EA of mirrored file */ + memset(buf_save, 0, sizeof(*buf)); + rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV); + if (rc < 0) + GOTO(out, rc); + + lcm = buf_save->lb_buf; + if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1) + GOTO(out, rc = -EINVAL); + + /** + * Extract the mirror with specified mirror id, and store the splitted + * mirror layout to the victim file. + */ + memset(buf, 0, sizeof(*buf)); + memset(buf_vic, 0, sizeof(*buf_vic)); + rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic); + if (rc < 0) + GOTO(out, rc); + + dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0; + + rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV, + LU_XATTR_SPLIT, handle); + if (rc) + GOTO(out, rc); + rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV, + LU_XATTR_SPLIT, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(out, rc); + + rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE, + handle); + if (rc) + GOTO(out, rc); + + rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE, + handle); + if (rc) + GOTO(out_restore, rc); + + rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle); + if (rc) + GOTO(out, rc); + EXIT; + +out_restore: + if (rc) { + /* restore obj's layout */ + int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV, + LU_XATTR_REPLACE, handle); + if (rc2) + CERROR("%s: failed to rollback of layout of: "DFID + ": %d, file state unkonwn.\n", + mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2); + } +out: + rc = mdd_trans_stop(env, mdd, rc, handle); + + /* Truncate local DOM data if all went well */ + if (!rc && dom_stripe) + mdd_dom_data_truncate(env, mdd, obj); + + mdd_write_unlock(env, obj); + mdd_write_unlock(env, vic); + lu_buf_free(buf_save); + lu_buf_free(buf); + lu_buf_free(buf_vic); + + if (!rc) + (void) mdd_object_pfid_replace(env, obj); + + return rc; +} + +static int mdd_layout_merge_allowed(const struct lu_env *env, + struct md_object *target, + struct md_object *victim) +{ + struct mdd_object *o1 = md2mdd_obj(target); + + /* cannot extend directory's LOVEA */ + if (S_ISDIR(mdd_object_type(o1))) { + CERROR("%s: Don't extend directory's LOVEA, just set it.\n", + mdd_obj_dev_name(o1)); + RETURN(-EISDIR); + } + + RETURN(0); +} + /** * The caller should guarantee to update the object ctime * after xattr_set if needed. @@ -998,21 +1884,48 @@ static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj, const struct lu_buf *buf, const char *name, int fl) { - struct mdd_object *mdd_obj = md2mdd_obj(obj); - struct lu_attr *attr = MDD_ENV_VAR(env, cattr); - struct mdd_device *mdd = mdo2mdd(obj); - struct thandle *handle; - int rc; + struct mdd_object *mdd_obj = md2mdd_obj(obj); + struct lu_attr *attr = MDD_ENV_VAR(env, cattr); + struct mdd_device *mdd = mdo2mdd(obj); + struct thandle *handle; + enum changelog_rec_type cl_type; + enum changelog_rec_flags clf_flags = 0; + int rc; ENTRY; - rc = mdd_la_get(env, mdd_obj, attr, BYPASS_CAPA); + rc = mdd_la_get(env, mdd_obj, attr); if (rc) RETURN(rc); - rc = mdd_xattr_sanity_check(env, mdd_obj, attr); + rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name); if (rc) RETURN(rc); + if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && + (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) { + struct md_rejig_data *mrd = buf->lb_buf; + struct md_object *victim = mrd->mrd_obj; + + if (buf->lb_len != sizeof(*mrd)) + RETURN(-EINVAL); + + rc = mdd_layout_merge_allowed(env, obj, victim); + if (rc) + RETURN(rc); + + if (fl == LU_XATTR_MERGE) + /* merge layout of victim as a mirror of obj's. */ + rc = mdd_xattr_merge(env, obj, victim); + else + rc = mdd_xattr_split(env, obj, mrd); + RETURN(rc); + } + + if (strcmp(name, XATTR_NAME_LMV) == 0) { + rc = mdd_dir_layout_shrink(env, obj, buf); + RETURN(rc); + } + if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 || strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) { struct posix_acl *acl; @@ -1021,6 +1934,8 @@ static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj, * ACL. */ acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf, buf->lb_len); + if (IS_ERR(acl)) + RETURN(PTR_ERR(acl)); if (acl == NULL) { rc = mdd_xattr_del(env, obj, name); RETURN(rc); @@ -1045,38 +1960,31 @@ static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj, if (rc) GOTO(stop, rc); - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); if (strcmp(XATTR_NAME_HSM, name) == 0) { - rc = mdd_hsm_update_locked(env, obj, buf, handle); + rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags); if (rc) { mdd_write_unlock(env, mdd_obj); GOTO(stop, rc); } } - rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle, - mdd_object_capa(env, mdd_obj)); + rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle); mdd_write_unlock(env, mdd_obj); if (rc) GOTO(stop, rc); - if (strcmp(XATTR_NAME_LOV, name) == 0) - rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, - handle); - else if (strncmp(XATTR_USER_PREFIX, name, - sizeof(XATTR_USER_PREFIX) - 1) == 0 || - strncmp(POSIX_ACL_XATTR_ACCESS, name, - sizeof(POSIX_ACL_XATTR_ACCESS) - 1) == 0 || - strncmp(POSIX_ACL_XATTR_DEFAULT, name, - sizeof(POSIX_ACL_XATTR_DEFAULT) - 1) == 0) - rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, - handle); + cl_type = mdd_xattr_changelog_type(env, mdd, name); + if (cl_type < 0) + GOTO(stop, rc = 0); -stop: - mdd_trans_stop(env, mdd, rc, handle); + rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags, + mdd_obj, name, handle); - RETURN(rc); + EXIT; +stop: + return mdd_trans_stop(env, mdd, rc, handle); } static int mdd_declare_xattr_del(const struct lu_env *env, @@ -1085,18 +1993,18 @@ static int mdd_declare_xattr_del(const struct lu_env *env, const char *name, struct thandle *handle) { + enum changelog_rec_type type; int rc; rc = mdo_declare_xattr_del(env, obj, name, handle); if (rc) return rc; - /* Only record user xattr changes */ - if ((strncmp(XATTR_USER_PREFIX, name, - sizeof(XATTR_USER_PREFIX) - 1) == 0)) - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); + type = mdd_xattr_changelog_type(env, mdd, name); + if (type < 0) + return 0; /* no changelog to store */ - return rc; + return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle); } /** @@ -1110,106 +2018,91 @@ static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj, struct lu_attr *attr = MDD_ENV_VAR(env, cattr); struct mdd_device *mdd = mdo2mdd(obj); struct thandle *handle; - int rc; + int rc; ENTRY; - rc = mdd_la_get(env, mdd_obj, attr, BYPASS_CAPA); + rc = mdd_la_get(env, mdd_obj, attr); if (rc) RETURN(rc); - rc = mdd_xattr_sanity_check(env, mdd_obj, attr); + rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name); if (rc) RETURN(rc); - handle = mdd_trans_create(env, mdd); - if (IS_ERR(handle)) - RETURN(PTR_ERR(handle)); + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); - rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle); - if (rc) - GOTO(stop, rc); + rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle); + if (rc) + GOTO(stop, rc); - rc = mdd_trans_start(env, mdd, handle); - if (rc) - GOTO(stop, rc); + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(stop, rc); - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); - rc = mdo_xattr_del(env, mdd_obj, name, handle, - mdd_object_capa(env, mdd_obj)); - mdd_write_unlock(env, mdd_obj); + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); + rc = mdo_xattr_del(env, mdd_obj, name, handle); + mdd_write_unlock(env, mdd_obj); if (rc) GOTO(stop, rc); - /* Only record system & user xattr changes */ - if (strncmp(XATTR_USER_PREFIX, name, - sizeof(XATTR_USER_PREFIX) - 1) == 0 || - strncmp(POSIX_ACL_XATTR_ACCESS, name, - sizeof(POSIX_ACL_XATTR_ACCESS) - 1) == 0 || - strncmp(POSIX_ACL_XATTR_DEFAULT, name, - sizeof(POSIX_ACL_XATTR_DEFAULT) - 1) == 0) - rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, - handle); + if (mdd_xattr_changelog_type(env, mdd, name) < 0) + GOTO(stop, rc = 0); -stop: - mdd_trans_stop(env, mdd, rc, handle); + rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj, + name, handle); - RETURN(rc); + EXIT; +stop: + return mdd_trans_stop(env, mdd, rc, handle); } /* - * read lov EA of an object - * return the lov EA in an allocated lu_buf + * read lov/lmv EA of an object + * return the lov/lmv EA in an allocated lu_buf */ -int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj, - struct lu_buf *lmm_buf) +int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj, + struct lu_buf *lmm_buf, const char *name) { - struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf; - int rc, bufsize; + struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf; + int rc; + ENTRY; -repeat: - rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV, - mdd_object_capa(env, obj)); + if (buf->lb_buf == NULL) { + buf = lu_buf_check_and_alloc(buf, 4096); + if (buf->lb_buf == NULL) + RETURN(-ENOMEM); + } +repeat: + rc = mdo_xattr_get(env, obj, buf, name); if (rc == -ERANGE) { /* mti_big_buf is allocated but is too small * we need to increase it */ buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, buf->lb_len * 2); if (buf->lb_buf == NULL) - GOTO(out, rc = -ENOMEM); + RETURN(-ENOMEM); goto repeat; - } - - if (rc < 0) + } else if (rc < 0) { RETURN(rc); - - if (rc == 0) + } else if (rc == 0) { RETURN(-ENODATA); - - bufsize = rc; - if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) { - /* mti_big_buf was not allocated, so we have to - * allocate it based on the ea size */ - buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, - bufsize); - if (buf->lb_buf == NULL) - GOTO(out, rc = -ENOMEM); - goto repeat; } - lu_buf_alloc(lmm_buf, bufsize); + lu_buf_alloc(lmm_buf, rc); if (lmm_buf->lb_buf == NULL) - GOTO(out, rc = -ENOMEM); + RETURN(-ENOMEM); - memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize); - rc = 0; - EXIT; + /* + * we don't use lmm_buf directly, because we don't know xattr size, so + * by using mti_big_buf we can avoid calling mdo_xattr_get() twice. + */ + memcpy(lmm_buf->lb_buf, buf->lb_buf, rc); -out: - if (rc < 0) - lu_buf_free(lmm_buf); - return rc; + RETURN(0); } static int mdd_xattr_hsm_replace(const struct lu_env *env, @@ -1217,13 +2110,13 @@ static int mdd_xattr_hsm_replace(const struct lu_env *env, struct thandle *handle) { struct hsm_attrs *attrs; - __u32 hsm_flags; - int flags = 0; + enum hsm_states hsm_flags; + enum changelog_rec_flags clf_flags = 0; int rc; ENTRY; rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE, - handle, mdd_object_capa(env, o)); + handle); if (rc != 0) RETURN(rc); @@ -1233,9 +2126,9 @@ static int mdd_xattr_hsm_replace(const struct lu_env *env, RETURN(0); /* Add a changelog record for release. */ - hsm_set_cl_event(&flags, HE_RELEASE); + hsm_set_cl_event(&clf_flags, HE_RELEASE); rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM, - flags, o, handle); + clf_flags, o, handle); RETURN(rc); } @@ -1244,13 +2137,15 @@ static int mdd_xattr_hsm_replace(const struct lu_env *env, * the rules are: * - only normal FIDs or non-system IGIFs * - same type of objects - * - same owner/group (so quotas are still valid) + * - same owner/group (so quotas are still valid) unless this is from HSM + * release. */ static int mdd_layout_swap_allowed(const struct lu_env *env, struct mdd_object *o1, const struct lu_attr *attr1, struct mdd_object *o2, - const struct lu_attr *attr2) + const struct lu_attr *attr2, + __u64 flags) { const struct lu_fid *fid1, *fid2; ENTRY; @@ -1274,6 +2169,9 @@ static int mdd_layout_swap_allowed(const struct lu_env *env, RETURN(-EBADF); } + if (flags & SWAP_LAYOUTS_MDS_HSM) + RETURN(0); + if ((attr1->la_uid != attr2->la_uid) || (attr1->la_gid != attr2->la_gid)) RETURN(-EPERM); @@ -1281,32 +2179,152 @@ static int mdd_layout_swap_allowed(const struct lu_env *env, RETURN(0); } +/* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to + * look into the layout in MDD layer. */ +static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get) +{ + struct lov_comp_md_v1 *comp_v1; + struct lov_mds_md *v1; + int i, ent_count; + __u32 off; + + if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) { + comp_v1 = (struct lov_comp_md_v1 *)lmm; + ent_count = le16_to_cpu(comp_v1->lcm_entry_count); + + if (ent_count == 0) + return -EINVAL; + + if (get) { + off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset); + v1 = (struct lov_mds_md *)((char *)comp_v1 + off); + *oi = v1->lmm_oi; + } else { + for (i = 0; i < le32_to_cpu(ent_count); i++) { + off = le32_to_cpu(comp_v1->lcm_entries[i]. + lcme_offset); + v1 = (struct lov_mds_md *)((char *)comp_v1 + + off); + v1->lmm_oi = *oi; + } + } + } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 || + le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) { + if (get) + *oi = lmm->lmm_oi; + else + lmm->lmm_oi = *oi; + } else { + return -EINVAL; + } + return 0; +} + +static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi) +{ + return mdd_lmm_oi(lmm, oi, true); +} + +static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi) +{ + return mdd_lmm_oi(lmm, oi, false); +} + +static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get) +{ + struct lov_comp_md_v1 *comp_v1; + + if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) { + comp_v1 = (struct lov_comp_md_v1 *)lmm; + if (get) + *gen = le32_to_cpu(comp_v1->lcm_layout_gen); + else + comp_v1->lcm_layout_gen = cpu_to_le32(*gen); + } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 || + le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) { + __u16 tmp_gen = *gen; + if (get) + *gen = le16_to_cpu(lmm->lmm_layout_gen); + else + lmm->lmm_layout_gen = cpu_to_le16(tmp_gen); + } else { + return -EINVAL; + } + return 0; +} + +static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen) +{ + return mdd_lmm_gen(lmm, gen, true); +} + +static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen) +{ + return mdd_lmm_gen(lmm, gen, false); +} + +static int mdd_dom_data_truncate(const struct lu_env *env, + struct mdd_device *mdd, struct mdd_object *mo) +{ + struct thandle *th; + struct dt_object *dom; + int rc; + + dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom); + if (!dom) + GOTO(out, rc = -ENODATA); + + th = dt_trans_create(env, mdd->mdd_bottom); + if (IS_ERR(th)) + GOTO(out, rc = PTR_ERR(th)); + + rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th); + if (rc) + GOTO(stop, rc); + + rc = dt_trans_start_local(env, mdd->mdd_bottom, th); + if (rc != 0) + GOTO(stop, rc); + + rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th); +stop: + dt_trans_stop(env, mdd->mdd_bottom, th); +out: + /* Ignore failure but report the error */ + if (rc) + CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n", + mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc); + return rc; +} + /** * swap layouts between 2 lustre objects */ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, struct md_object *obj2, __u64 flags) { - struct mdd_thread_info *info = mdd_env_info(env); - struct mdd_object *fst_o = md2mdd_obj(obj1); - struct mdd_object *snd_o = md2mdd_obj(obj2); - struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr); - struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr); - struct mdd_device *mdd = mdo2mdd(obj1); - struct lov_mds_md *fst_lmm, *snd_lmm; - struct lu_buf *fst_buf = &info->mti_buf[0]; - struct lu_buf *snd_buf = &info->mti_buf[1]; - struct lu_buf *fst_hsm_buf = &info->mti_buf[2]; - struct lu_buf *snd_hsm_buf = &info->mti_buf[3]; - struct ost_id *saved_oi = NULL; - struct thandle *handle; - __u16 fst_gen, snd_gen; - int fst_fl; - int rc; - int rc2; + struct mdd_thread_info *info = mdd_env_info(env); + struct mdd_object *fst_o = md2mdd_obj(obj1); + struct mdd_object *snd_o = md2mdd_obj(obj2); + struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr); + struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr); + struct mdd_device *mdd = mdo2mdd(obj1); + struct lov_mds_md *fst_lmm, *snd_lmm; + struct lu_buf *fst_buf = &info->mti_buf[0]; + struct lu_buf *snd_buf = &info->mti_buf[1]; + struct lu_buf *fst_hsm_buf = &info->mti_buf[2]; + struct lu_buf *snd_hsm_buf = &info->mti_buf[3]; + struct ost_id *saved_oi = NULL; + struct thandle *handle; + struct mdd_object *dom_o = NULL; + __u64 domsize_dom, domsize_vlt; + __u32 fst_gen, snd_gen, saved_gen; + int fst_fl; + int rc, rc2; + ENTRY; - CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4); + BUILD_BUG_ON(ARRAY_SIZE(info->mti_buf) < 4); memset(info->mti_buf, 0, sizeof(info->mti_buf)); /* we have to sort the 2 obj, so locking will always @@ -1318,16 +2336,16 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, if (rc < 0) swap(fst_o, snd_o); - rc = mdd_la_get(env, fst_o, fst_la, BYPASS_CAPA); + rc = mdd_la_get(env, fst_o, fst_la); if (rc != 0) RETURN(rc); - rc = mdd_la_get(env, snd_o, snd_la, BYPASS_CAPA); + rc = mdd_la_get(env, snd_o, snd_la); if (rc != 0) RETURN(rc); /* check if layout swapping is allowed */ - rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la); + rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags); if (rc != 0) RETURN(rc); @@ -1336,17 +2354,62 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, RETURN(PTR_ERR(handle)); /* objects are already sorted */ - mdd_write_lock(env, fst_o, MOR_TGT_CHILD); - mdd_write_lock(env, snd_o, MOR_TGT_CHILD); + mdd_write_lock(env, fst_o, DT_TGT_CHILD); + mdd_write_lock(env, snd_o, DT_TGT_CHILD); - rc = mdd_get_lov_ea(env, fst_o, fst_buf); + rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV); if (rc < 0 && rc != -ENODATA) GOTO(stop, rc); - rc = mdd_get_lov_ea(env, snd_o, snd_buf); + rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV); if (rc < 0 && rc != -ENODATA) GOTO(stop, rc); + /* check if file has DoM. DoM file can be migrated only to another + * DoM layout with the same DoM component size or to an non-DOM + * layout. After migration to OSTs layout, local MDT inode data + * should be truncated. + * Objects are sorted by FIDs, considering that original file's FID + * is always smaller the snd_o is always original file we are migrating + * from. + */ + domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf); + domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf); + + /* Only migration is supported for DoM files, not 'swap_layouts' so + * target file must be volatile and orphan. + */ + if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) { + dom_o = domsize_dom ? snd_o : NULL; + } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) { + swap(domsize_dom, domsize_vlt); + dom_o = domsize_dom ? fst_o : NULL; + } else if (domsize_dom > 0 || domsize_vlt > 0) { + /* 'lfs swap_layouts' case, neither file should have DoM */ + rc = -EOPNOTSUPP; + CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, " + "use migration instead: rc = %d\n", rc); + GOTO(stop, rc); + } + + if (domsize_vlt > 0 && domsize_dom == 0) { + rc = -EOPNOTSUPP; + CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM " + "migration is not supported: rc = %d\n", + PFID(mdo2fid(snd_o)), rc); + GOTO(stop, rc); + } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) { + rc = -EOPNOTSUPP; + CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout " + "must have the same DoM component size: rc = %d\n", + PFID(mdo2fid(fst_o)), rc); + GOTO(stop, rc); + } else if (domsize_vlt > 0) { + /* Migration with the same DOM component size, no need to + * truncate local data, it is still being used */ + dom_o = NULL; + } + /* swapping 2 non existant layouts is a success */ if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL) GOTO(stop, rc = 0); @@ -1358,10 +2421,11 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, swap(fst_buf, snd_buf); } + fst_gen = snd_gen = 0; /* lmm and generation layout initialization */ if (fst_buf->lb_buf != NULL) { fst_lmm = fst_buf->lb_buf; - fst_gen = le16_to_cpu(fst_lmm->lmm_layout_gen); + mdd_get_lmm_gen(fst_lmm, &fst_gen); fst_fl = LU_XATTR_REPLACE; } else { fst_lmm = NULL; @@ -1370,29 +2434,45 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, } snd_lmm = snd_buf->lb_buf; - snd_gen = le16_to_cpu(snd_lmm->lmm_layout_gen); + mdd_get_lmm_gen(snd_lmm, &snd_gen); + saved_gen = fst_gen; /* increase the generation layout numbers */ snd_gen++; fst_gen++; + /* + * XXX The layout generation is used to generate component IDs for + * the composite file, we have to do some special tweaks to make + * sure the layout generation is always adequate for that job. + */ + + /* Skip invalid generation number for composite layout */ + if ((snd_gen & LCME_ID_MASK) == 0) + snd_gen++; + if ((fst_gen & LCME_ID_MASK) == 0) + fst_gen++; + /* Make sure the generation is greater than all the component IDs */ + if (fst_gen < snd_gen) + fst_gen = snd_gen; + else if (fst_gen > snd_gen) + snd_gen = fst_gen; + /* set the file specific informations in lmm */ if (fst_lmm != NULL) { saved_oi = &info->mti_oa.o_oi; - - *saved_oi = fst_lmm->lmm_oi; - fst_lmm->lmm_layout_gen = cpu_to_le16(snd_gen); - fst_lmm->lmm_oi = snd_lmm->lmm_oi; - snd_lmm->lmm_oi = *saved_oi; + mdd_get_lmm_oi(fst_lmm, saved_oi); + mdd_set_lmm_gen(fst_lmm, &snd_gen); + mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi); + mdd_set_lmm_oi(snd_lmm, saved_oi); } else { - if (snd_lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) - snd_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF); - else if (snd_lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) - snd_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF); + if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) == + cpu_to_le32(LOV_MAGIC_MAGIC)) + snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED); else GOTO(stop, rc = -EPROTO); } - snd_lmm->lmm_layout_gen = cpu_to_le16(fst_gen); + mdd_set_lmm_gen(snd_lmm, &fst_gen); /* Prepare HSM attribute if it's required */ if (flags & SWAP_LAYOUTS_MDS_HSM) { @@ -1404,13 +2484,11 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, GOTO(stop, rc = -ENOMEM); /* Read HSM attribute */ - rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM, - BYPASS_CAPA); + rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM); if (rc < 0) GOTO(stop, rc); - rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM, - BYPASS_CAPA); + rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM); if (rc < 0) GOTO(stop, rc); @@ -1464,8 +2542,7 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, } } - rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle, - mdd_object_capa(env, fst_o)); + rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle); if (rc != 0) GOTO(stop, rc); @@ -1474,27 +2551,36 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, } else { if (fst_buf->lb_buf != NULL) rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV, - LU_XATTR_REPLACE, handle, - mdd_object_capa(env, snd_o)); + LU_XATTR_REPLACE, handle); else - rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle, - mdd_object_capa(env, snd_o)); + rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle); } + if (rc != 0) + GOTO(out_restore, rc); + + /* Issue one changelog record per file */ + rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle); + if (rc) + GOTO(stop, rc); + rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle); + if (rc) + GOTO(stop, rc); + EXIT; + +out_restore: if (rc != 0) { int steps = 0; /* failure on second file, but first was done, so we have * to roll back first. */ if (fst_buf->lb_buf != NULL) { - fst_lmm->lmm_oi = *saved_oi; - fst_lmm->lmm_layout_gen = cpu_to_le16(fst_gen - 1); + mdd_set_lmm_oi(fst_lmm, saved_oi); + mdd_set_lmm_gen(fst_lmm, &saved_gen); rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV, - LU_XATTR_REPLACE, handle, - mdd_object_capa(env, fst_o)); + LU_XATTR_REPLACE, handle); } else { - rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle, - mdd_object_capa(env, fst_o)); + rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle); } if (rc2 < 0) goto do_lbug; @@ -1521,21 +2607,15 @@ static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1, */ LBUG(); } - GOTO(stop, rc); } - /* Issue one changelog record per file */ - rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle); - if (rc) - GOTO(stop, rc); +stop: + rc = mdd_trans_stop(env, mdd, rc, handle); - rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle); - if (rc) - GOTO(stop, rc); - EXIT; + /* Truncate local DOM data if all went well */ + if (!rc && dom_o) + mdd_dom_data_truncate(env, mdd, dom_o); -stop: - mdd_trans_stop(env, mdd, rc, handle); mdd_write_unlock(env, snd_o); mdd_write_unlock(env, fst_o); @@ -1543,6 +2623,374 @@ stop: lu_buf_free(snd_buf); lu_buf_free(fst_hsm_buf); lu_buf_free(snd_hsm_buf); + + if (!rc) { + (void) mdd_object_pfid_replace(env, fst_o); + (void) mdd_object_pfid_replace(env, snd_o); + } + return rc; +} + +static int mdd_declare_layout_change(const struct lu_env *env, + struct mdd_device *mdd, + struct mdd_object *obj, + struct md_layout_change *mlc, + struct thandle *handle) +{ + int rc; + + rc = mdo_declare_layout_change(env, obj, mlc, handle); + if (rc) + return rc; + + return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL, + handle); +} + +/* For PFL, this is used to instantiate necessary component objects. */ +static int +mdd_layout_instantiate_component(const struct lu_env *env, + struct mdd_object *obj, struct md_layout_change *mlc, + struct thandle *handle) +{ + struct mdd_device *mdd = mdd_obj2mdd_dev(obj); + int rc; + ENTRY; + + if (mlc->mlc_opc != MD_LAYOUT_WRITE) + RETURN(-ENOTSUPP); + + rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle); + /** + * It's possible that another layout write intent has already + * instantiated our objects, so a -EALREADY returned, and we need to + * do nothing. + */ + if (rc) + RETURN(rc == -EALREADY ? 0 : rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + RETURN(rc); + + mdd_write_lock(env, obj, DT_TGT_CHILD); + rc = mdo_layout_change(env, obj, mlc, handle); + mdd_write_unlock(env, obj); + if (rc) + RETURN(rc); + + rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle); + RETURN(rc); +} + +/** + * Change the FLR layout from RDONLY to WRITE_PENDING. + * + * It picks the primary mirror, and bumps the layout version, and set + * layout version xattr to OST objects in a sync tx. In order to facilitate + * the handling of phantom writers from evicted clients, the clients carry + * layout version of the file with write RPC, so that the OSTs can verify + * if the write RPCs are legitimate, meaning not from evicted clients. + */ +static int +mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj, + struct md_layout_change *mlc, struct thandle *handle) +{ + struct mdd_device *mdd = mdd_obj2mdd_dev(obj); + struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1]; + struct lustre_som_attrs *som = &mlc->mlc_som; + int fl = 0; + int rc; + ENTRY; + + /* Verify acceptable operations */ + switch (mlc->mlc_opc) { + case MD_LAYOUT_WRITE: + case MD_LAYOUT_RESYNC: + /* these are legal operations - this represents the case that + * a few mirrors were missed in the last resync. */ + break; + case MD_LAYOUT_RESYNC_DONE: + default: + RETURN(0); + } + + som_buf->lb_buf = som; + som_buf->lb_len = sizeof(*som); + rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM); + if (rc < 0 && rc != -ENODATA) + RETURN(rc); + + if (rc > 0) { + lustre_som_swab(som); + if (som->lsa_valid & SOM_FL_STRICT) + fl = LU_XATTR_REPLACE; + } + + rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle); + if (rc) + GOTO(out, rc); + + if (fl) { + rc = mdd_declare_xattr_set(env, mdd, obj, som_buf, + XATTR_NAME_SOM, fl, handle); + if (rc) + GOTO(out, rc); + } + + /* record a changelog for data mover to consume */ + rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(out, rc); + + /* it needs a sync tx to make FLR to work properly */ + handle->th_sync = 1; + + mdd_write_lock(env, obj, DT_TGT_CHILD); + rc = mdo_layout_change(env, obj, mlc, handle); + if (!rc && fl) { + /* SOM state transition from STRICT to STALE */ + som->lsa_valid = SOM_FL_STALE; + lustre_som_swab(som); + rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM, + fl, handle); + } + mdd_write_unlock(env, obj); + if (rc) + GOTO(out, rc); + + rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle); + if (rc) + GOTO(out, rc); + + EXIT; + +out: + return rc; +} + +/** + * Handle mirrored file state transition when it's in WRITE_PENDING. + * + * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when + * the file is in WRITE_PENDING state. If everything goes fine, the file's + * layout version will be increased, and the file's state will be changed to + * SYNC_PENDING. + */ +static int +mdd_layout_update_write_pending(const struct lu_env *env, + struct mdd_object *obj, struct md_layout_change *mlc, + struct thandle *handle) +{ + struct mdd_device *mdd = mdd_obj2mdd_dev(obj); + int rc; + ENTRY; + + switch (mlc->mlc_opc) { + case MD_LAYOUT_RESYNC: + /* Upon receiving the resync request, it should + * instantiate all stale components right away to get ready + * for mirror copy. In order to avoid layout version change, + * client should avoid sending LAYOUT_WRITE request at the + * resync state. */ + break; + case MD_LAYOUT_WRITE: + /* legal race for concurrent write, the file state has been + * changed by another client. */ + break; + default: + RETURN(-EBUSY); + } + + rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle); + if (rc) + GOTO(out, rc); + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(out, rc); + + /* it needs a sync tx to make FLR to work properly */ + handle->th_sync = 1; + + mdd_write_lock(env, obj, DT_TGT_CHILD); + rc = mdo_layout_change(env, obj, mlc, handle); + mdd_write_unlock(env, obj); + if (rc) + GOTO(out, rc); + + EXIT; + +out: + return rc; +} + +/** + * Handle the requests when a FLR file's state is in SYNC_PENDING. + * + * Only concurrent write and sync complete requests are possible when the + * file is in SYNC_PENDING. For the latter request, it will pass in the + * mirrors that have been synchronized, then the stale bit will be cleared + * to make the file's state turn into RDONLY. + * For concurrent write reqeust, it just needs to change the file's state + * to WRITE_PENDING in a sync tx. It doesn't have to change the layout + * version because the version will be increased in the transition to + * SYNC_PENDING later so that it can deny the write request from potential + * evicted SYNC clients. */ +static int +mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj, + struct md_layout_change *mlc, struct thandle *handle) +{ + struct mdd_device *mdd = mdd_obj2mdd_dev(obj); + struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1]; + int fl = 0; + int rc; + ENTRY; + + /* operation validation */ + switch (mlc->mlc_opc) { + case MD_LAYOUT_RESYNC_DONE: + /* resync complete. */ + case MD_LAYOUT_WRITE: + /* concurrent write. */ + break; + case MD_LAYOUT_RESYNC: + /* resync again, most likely the previous run failed. + * no-op if it's already in SYNC_PENDING state */ + RETURN(0); + default: + RETURN(-EBUSY); + } + + if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) { + rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM); + if (rc < 0 && rc != -ENODATA) + RETURN(rc); + + fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE; + lustre_som_swab(&mlc->mlc_som); + som_buf->lb_buf = &mlc->mlc_som; + som_buf->lb_len = sizeof(mlc->mlc_som); + } + + rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle); + if (rc) + GOTO(out, rc); + + /* record a changelog for the completion of resync */ + rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL, + handle); + if (rc) + GOTO(out, rc); + + /* RESYNC_DONE has piggybacked size and blocks */ + if (fl) { + rc = mdd_declare_xattr_set(env, mdd, obj, som_buf, + XATTR_NAME_SOM, fl, handle); + if (rc) + GOTO(out, rc); + } + + rc = mdd_trans_start(env, mdd, handle); + if (rc) + GOTO(out, rc); + + /* it needs a sync tx to make FLR to work properly */ + handle->th_sync = 1; + + rc = mdo_layout_change(env, obj, mlc, handle); + if (rc) + GOTO(out, rc); + + if (fl) { + rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM, + fl, handle); + if (rc) + GOTO(out, rc); + } + + rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle); + if (rc) + GOTO(out, rc); + EXIT; +out: + return rc; +} + +/** + * Layout change callback for object. + * + * This is only used by FLR for now. In the future, it can be exteneded to + * handle all layout change. + */ +static int +mdd_layout_change(const struct lu_env *env, struct md_object *o, + struct md_layout_change *mlc) +{ + struct mdd_object *obj = md2mdd_obj(o); + struct mdd_device *mdd = mdd_obj2mdd_dev(obj); + struct lu_buf *buf = mdd_buf_get(env, NULL, 0); + struct lov_comp_md_v1 *lcm; + struct thandle *handle; + int flr_state; + int rc; + ENTRY; + + /* Verify acceptable operations */ + switch (mlc->mlc_opc) { + case MD_LAYOUT_WRITE: + case MD_LAYOUT_RESYNC: + case MD_LAYOUT_RESYNC_DONE: + break; + default: + RETURN(-ENOTSUPP); + } + + handle = mdd_trans_create(env, mdd); + if (IS_ERR(handle)) + RETURN(PTR_ERR(handle)); + + rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV); + if (rc < 0) { + if (rc == -ENODATA) + rc = -EINVAL; + GOTO(out, rc); + } + + /* analyze the layout to make sure it's a FLR file */ + lcm = buf->lb_buf; + if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1) + GOTO(out, rc = -EINVAL); + + flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK; + + /* please refer to HLD of FLR for state transition */ + switch (flr_state) { + case LCM_FL_NONE: + rc = mdd_layout_instantiate_component(env, obj, mlc, handle); + break; + case LCM_FL_WRITE_PENDING: + rc = mdd_layout_update_write_pending(env, obj, mlc, handle); + break; + case LCM_FL_RDONLY: + rc = mdd_layout_update_rdonly(env, obj, mlc, handle); + break; + case LCM_FL_SYNC_PENDING: + rc = mdd_object_update_sync_pending(env, obj, mlc, handle); + break; + default: + rc = 0; + break; + } + EXIT; + +out: + mdd_trans_stop(env, mdd, rc, handle); + lu_buf_free(buf); return rc; } @@ -1552,6 +3000,7 @@ void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent, struct dt_allocation_hint *hint) { struct dt_object *np = parent ? mdd_object_child(parent) : NULL; + struct mdd_device *mdd = mdd_obj2mdd_dev(child); struct dt_object *nc = mdd_object_child(child); memset(hint, 0, sizeof(*hint)); @@ -1565,6 +3014,19 @@ void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent, } else { hint->dah_eadata = NULL; hint->dah_eadata_len = 0; + if (spec->sp_cr_flags & MDS_OPEN_APPEND) { + if (mdd->mdd_append_stripe_count != 0 || + mdd->mdd_append_pool[0]) + CDEBUG(D_INFO, + "using O_APPEND file striping\n"); + if (mdd->mdd_append_stripe_count) + hint->dah_append_stripes = + mdd->mdd_append_stripe_count; + if (mdd->mdd_append_pool[0]) + hint->dah_append_pool = mdd->mdd_append_pool; + } else { + hint->dah_append_stripes = 0; + } } CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)), @@ -1573,54 +3035,48 @@ void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent, nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT); } -/* - * do NOT or the MAY_*'s, you'll get the weakest - */ -int accmode(const struct lu_env *env, const struct lu_attr *la, int flags) +static int accmode(const struct lu_env *env, const struct lu_attr *la, + u64 open_flags) { - int res = 0; - /* Sadly, NFSD reopens a file repeatedly during operation, so the * "acc_mode = 0" allowance for newly-created files isn't honoured. * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file * owner can write to a file even if it is marked readonly to hide * its brokenness. (bug 5781) */ - if (flags & MDS_OPEN_OWNEROVERRIDE) { + if (open_flags & MDS_OPEN_OWNEROVERRIDE) { struct lu_ucred *uc = lu_ucred_check(env); if ((uc == NULL) || (la->la_uid == uc->uc_fsuid)) return 0; } - if (flags & FMODE_READ) - res |= MAY_READ; - if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND)) - res |= MAY_WRITE; - if (flags & MDS_FMODE_EXEC) - res = MAY_EXEC; - return res; + return mds_accmode(open_flags); } static int mdd_open_sanity_check(const struct lu_env *env, - struct mdd_object *obj, - const struct lu_attr *attr, int flag) + struct mdd_object *obj, + const struct lu_attr *attr, u64 open_flags, + int is_replay) { int mode, rc; ENTRY; - /* EEXIST check */ - if (mdd_is_dead_obj(obj)) + /* EEXIST check, also opening of *open* orphans is allowed so we can + * open-by-handle unlinked files + */ + if (mdd_is_dead_obj(obj) && !is_replay && + likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0))) RETURN(-ENOENT); if (S_ISLNK(attr->la_mode)) RETURN(-ELOOP); - mode = accmode(env, attr, flag); + mode = accmode(env, attr, open_flags); if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE)) RETURN(-EISDIR); - if (!(flag & MDS_OPEN_CREATED)) { + if (!(open_flags & MDS_OPEN_CREATED)) { rc = mdd_permission_internal(env, obj, attr, mode); if (rc) RETURN(rc); @@ -1628,68 +3084,112 @@ static int mdd_open_sanity_check(const struct lu_env *env, if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) || S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode)) - flag &= ~MDS_OPEN_TRUNC; + open_flags &= ~MDS_OPEN_TRUNC; /* For writing append-only file must open it with append mode. */ if (attr->la_flags & LUSTRE_APPEND_FL) { - if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND)) + if ((open_flags & MDS_FMODE_WRITE) && + !(open_flags & MDS_OPEN_APPEND)) RETURN(-EPERM); - if (flag & MDS_OPEN_TRUNC) + if (open_flags & MDS_OPEN_TRUNC) RETURN(-EPERM); } -#if 0 - /* - * Now, flag -- O_NOATIME does not be packed by client. - */ - if (flag & O_NOATIME) { - struct lu_ucred *uc = lu_ucred(env); - - if (uc && ((uc->uc_valid == UCRED_OLD) || - (uc->uc_valid == UCRED_NEW)) && - (uc->uc_fsuid != attr->la_uid) && - !md_capable(uc, CFS_CAP_FOWNER)) - RETURN(-EPERM); - } -#endif - RETURN(0); } static int mdd_open(const struct lu_env *env, struct md_object *obj, - int flags) + u64 open_flags, struct md_op_spec *spec) { struct mdd_object *mdd_obj = md2mdd_obj(obj); + struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj))); struct lu_attr *attr = MDD_ENV_VAR(env, cattr); + struct mdd_object_user *mou = NULL; + const struct lu_ucred *uc = lu_ucred(env); + struct mdd_device *mdd = mdo2mdd(obj); + enum changelog_rec_type type = CL_OPEN; int rc = 0; + ENTRY; - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); - rc = mdd_la_get(env, mdd_obj, attr, BYPASS_CAPA); + rc = mdd_la_get(env, mdd_obj, attr); if (rc != 0) GOTO(out, rc); - rc = mdd_open_sanity_check(env, mdd_obj, attr, flags); - if (rc != 0) + rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags, + spec->no_create); + if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN))) + type = CL_DN_OPEN; + else if (rc != 0) GOTO(out, rc); + else + mdd_obj->mod_count++; + + if (!mdd_changelog_enabled(env, mdd, type)) + GOTO(out, rc); + +find: + /* look for existing opener in list under mdd_write_lock */ + mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags); + + if (!mou) { + int rc2; + + /* add user to list */ + mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid); + if (IS_ERR(mou)) { + if (rc == 0) + rc = PTR_ERR(mou); + GOTO(out, rc); + } + rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN); + if (rc2 != 0) { + mdd_obj_user_free(mou); + if (rc2 == -EEXIST) + GOTO(find, rc2); + } + } else { + if (type == CL_DN_OPEN) { + if (ktime_before(ktime_get(), mou->mou_deniednext)) + /* same user denied again same access within + * time interval: do not record + */ + GOTO(out, rc); + + /* this user already denied, but some time ago: + * update denied time + */ + mou->mou_deniednext = + ktime_add(ktime_get(), + ktime_set(mdd->mdd_cl.mc_deniednext, + 0)); + } else { + mou->mou_opencount++; + /* same user opening file again with same flags: + * don't record + */ + GOTO(out, rc); + } + } + + /* FYI, only the bottom 32 bits of open_flags are recorded */ + mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj)); - mdd_obj->mod_count++; EXIT; out: mdd_write_unlock(env, mdd_obj); return rc; } -static int mdd_declare_close(const struct lu_env *env, - struct mdd_object *obj, - struct md_attr *ma, - struct thandle *handle) +static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj, + struct md_attr *ma, struct thandle *handle) { - int rc; + int rc; - rc = orph_declare_index_delete(env, obj, handle); - if (rc) - return rc; + rc = mdd_orphan_declare_delete(env, obj, handle); + if (rc) + return rc; return mdo_declare_destroy(env, obj, handle); } @@ -1698,24 +3198,28 @@ static int mdd_declare_close(const struct lu_env *env, * No permission check is needed. */ static int mdd_close(const struct lu_env *env, struct md_object *obj, - struct md_attr *ma, int mode) + struct md_attr *ma, u64 open_flags) { - struct mdd_object *mdd_obj = md2mdd_obj(obj); - struct mdd_device *mdd = mdo2mdd(obj); - struct thandle *handle = NULL; - int rc, is_orphan = 0; - ENTRY; + struct mdd_object *mdd_obj = md2mdd_obj(obj); + struct mdd_device *mdd = mdo2mdd(obj); + struct thandle *handle = NULL; + int is_orphan = 0; + int rc; + bool blocked = false; + bool last_close_by_uid = false; + const struct lu_ucred *uc = lu_ucred(env); + ENTRY; - if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) { - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); + if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) { + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); mdd_obj->mod_count--; mdd_write_unlock(env, mdd_obj); - if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count) - CDEBUG(D_HA, "Object "DFID" is retained in orphan " - "list\n", PFID(mdd_object_fid(mdd_obj))); - RETURN(0); - } + if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count) + CDEBUG(D_HA, "Object "DFID" is retained in orphan " + "list\n", PFID(mdd_object_fid(mdd_obj))); + RETURN(0); + } /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */ @@ -1725,28 +3229,48 @@ static int mdd_close(const struct lu_env *env, struct md_object *obj, again: if (is_orphan) { - handle = mdd_trans_create(env, mdo2mdd(obj)); - if (IS_ERR(handle)) - RETURN(PTR_ERR(handle)); + /* mdd_trans_create() maybe failed because of barrier_entry(), + * under such case, the orphan MDT-object will be left in the + * orphan list, and when the MDT remount next time, the unused + * orphans will be destroyed automatically. + * + * One exception: the former mdd_finish_unlink may failed to + * add the orphan MDT-object to the orphan list, then if the + * mdd_trans_create() failed because of barrier_entry(), the + * MDT-object will become real orphan that is neither in the + * namespace nor in the orphan list. Such bad case should be + * very rare and will be handled by e2fsck/lfsck. */ + handle = mdd_trans_create(env, mdo2mdd(obj)); + if (IS_ERR(handle)) { + rc = PTR_ERR(handle); + if (rc != -EINPROGRESS) + GOTO(stop, rc); - rc = mdd_declare_close(env, mdd_obj, ma, handle); - if (rc) - GOTO(stop, rc); + handle = NULL; + blocked = true; + goto cont; + } - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle); + rc = mdd_declare_close(env, mdd_obj, ma, handle); if (rc) GOTO(stop, rc); - rc = mdd_trans_start(env, mdo2mdd(obj), handle); - if (rc) - GOTO(stop, rc); - } + rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL, + handle); + if (rc) + GOTO(stop, rc); + + rc = mdd_trans_start(env, mdo2mdd(obj), handle); + if (rc) + GOTO(stop, rc); + } - mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD); - rc = mdd_la_get(env, mdd_obj, &ma->ma_attr, - mdd_object_capa(env, mdd_obj)); +cont: + mdd_write_lock(env, mdd_obj, DT_TGT_CHILD); + rc = mdd_la_get(env, mdd_obj, &ma->ma_attr); if (rc != 0) { - CERROR("Failed to get lu_attr of "DFID": %d\n", + CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n", + lu_dev_name(mdd2lu_dev(mdd)), PFID(mdd_object_fid(mdd_obj)), rc); GOTO(out, rc); } @@ -1756,14 +3280,36 @@ again: ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 || ma->ma_attr.la_nlink == 0); - if (is_orphan && handle == NULL) { + if (is_orphan && !handle && !blocked) { mdd_write_unlock(env, mdd_obj); goto again; } mdd_obj->mod_count--; /*release open count */ - if (!is_orphan) + /* under mdd write lock */ + /* If recording, see if we need to remove UID from list. uc is not + * initialized if the client has been evicted. */ + if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) { + struct mdd_object_user *mou; + + /* look for UID in list */ + /* If mou is NULL, it probably means logging was enabled after + * the user had the file open. So the corresponding close + * will not be logged. + */ + mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, + open_flags); + if (mou) { + mou->mou_opencount--; + if (mou->mou_opencount == 0) { + mdd_obj_user_remove(mdd_obj, mou); + last_close_by_uid = true; + } + } + } + + if (!is_orphan || blocked) GOTO(out, rc = 0); /* Orphan object */ @@ -1773,7 +3319,7 @@ again: if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) { /* remove link to object from orphan index */ LASSERT(handle != NULL); - rc = __mdd_orphan_del(env, mdd_obj, handle); + rc = mdd_orphan_delete(env, mdd_obj, handle); if (rc != 0) { CERROR("%s: unable to delete "DFID" from orphan list: " "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)), @@ -1799,33 +3345,46 @@ again: EXIT; out: - mdd_write_unlock(env, mdd_obj); - - if (rc == 0 && - (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) && - !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) { - if (handle == NULL) { - handle = mdd_trans_create(env, mdo2mdd(obj)); - if (IS_ERR(handle)) + mdd_write_unlock(env, mdd_obj); + + if (rc != 0 || blocked || + !mdd_changelog_enabled(env, mdd, CL_CLOSE)) + GOTO(stop, rc); + + /* Record CL_CLOSE in changelog only if file was opened in write mode, + * or if CL_OPEN was recorded and it's last close by user. + * Changelogs mask may change between open and close operations, but + * this is not a big deal if we have a CL_CLOSE entry with no matching + * CL_OPEN. Plus Changelogs mask may not change often. + */ + if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && + (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND | + MDS_OPEN_TRUNC))) || + ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) && + !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) { + if (handle == NULL) { + handle = mdd_trans_create(env, mdo2mdd(obj)); + if (IS_ERR(handle)) GOTO(stop, rc = PTR_ERR(handle)); - rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, - handle); + rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, + NULL, NULL, handle); if (rc) GOTO(stop, rc); - rc = mdd_trans_start(env, mdo2mdd(obj), handle); - if (rc) - GOTO(stop, rc); - } + rc = mdd_trans_start(env, mdo2mdd(obj), handle); + if (rc) + GOTO(stop, rc); + } - mdd_changelog_data_store(env, mdd, CL_CLOSE, mode, - mdd_obj, handle); - } + /* FYI, only the bottom 32 bits of open_flags are recorded */ + mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags, + mdd_obj, handle); + } stop: if (handle != NULL && !IS_ERR(handle)) - mdd_trans_stop(env, mdd, rc, handle); + rc = mdd_trans_stop(env, mdd, rc, handle); return rc; } @@ -1850,7 +3409,7 @@ static int mdd_readpage_sanity_check(const struct lu_env *env, } static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp, - int nob, const struct dt_it_ops *iops, + size_t nob, const struct dt_it_ops *iops, struct dt_it *it, __u32 attr, void *arg) { struct lu_dirpage *dp = &lp->lp_dir; @@ -1862,6 +3421,9 @@ static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp, struct lu_fid fid; int first = 1; + if (nob < sizeof(*dp)) + return -EINVAL; + memset(area, 0, sizeof (*dp)); area += sizeof (*dp); nob -= sizeof (*dp); @@ -1869,7 +3431,7 @@ static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp, ent = area; do { int len; - int recsize; + size_t recsize; len = iops->key_size(env, it); @@ -1944,7 +3506,7 @@ int mdd_readpage(const struct lu_env *env, struct md_object *obj, return -ENOENT; } - mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD); + mdd_read_lock(env, mdd_obj, DT_TGT_CHILD); rc = mdd_readpage_sanity_check(env, mdd_obj); if (rc) GOTO(out_unlock, rc); @@ -2019,7 +3581,7 @@ static int mdd_object_lock(const struct lu_env *env, struct md_object *obj, struct lustre_handle *lh, struct ldlm_enqueue_info *einfo, - ldlm_policy_data_t *policy) + union ldlm_policy_data *policy) { struct mdd_object *mdd_obj = md2mdd_obj(obj); return dt_object_lock(env, mdd_object_child(mdd_obj), lh, @@ -2029,7 +3591,7 @@ static int mdd_object_lock(const struct lu_env *env, static int mdd_object_unlock(const struct lu_env *env, struct md_object *obj, struct ldlm_enqueue_info *einfo, - ldlm_policy_data_t *policy) + union ldlm_policy_data *policy) { struct mdd_object *mdd_obj = md2mdd_obj(obj); return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy); @@ -2042,6 +3604,7 @@ const struct md_object_operations mdd_obj_ops = { .moo_xattr_get = mdd_xattr_get, .moo_xattr_set = mdd_xattr_set, .moo_xattr_list = mdd_xattr_list, + .moo_invalidate = mdd_invalidate, .moo_xattr_del = mdd_xattr_del, .moo_swap_layouts = mdd_swap_layouts, .moo_open = mdd_open, @@ -2049,8 +3612,8 @@ const struct md_object_operations mdd_obj_ops = { .moo_readpage = mdd_readpage, .moo_readlink = mdd_readlink, .moo_changelog = mdd_changelog, - .moo_capa_get = mdd_capa_get, .moo_object_sync = mdd_object_sync, .moo_object_lock = mdd_object_lock, .moo_object_unlock = mdd_object_unlock, + .moo_layout_change = mdd_layout_change, };