* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2016, Intel Corporation.
+ * Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
static const struct lu_object_operations mdd_lu_obj_ops;
+struct mdd_object_user {
+ struct list_head mou_list; /**< linked off mod_users */
+ int mou_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,
+ int 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 unsigned int flags_helper(int flags)
+{
+ unsigned int mflags = 0;
+
+ if (flags & MDS_FMODE_EXEC) {
+ mflags = MDS_FMODE_EXEC;
+ } else {
+ if (flags & MDS_FMODE_READ)
+ mflags = MDS_FMODE_READ;
+ if (flags &
+ (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
+ mflags |= MDS_FMODE_WRITE;
+ }
+
+ return mflags;
+}
+
+/** 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(int 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_flags = 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, int 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_flags) == flags_helper(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_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 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,
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);
}
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,
strcmp(name, XATTR_NAME_LINK) == 0))
RETURN(-ENOENT);
- mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
+ mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
rc = mdo_xattr_get(env, mdd_obj, buf, name);
- mdd_read_unlock(env, mdd_obj);
+ mdd_read_unlock(env, mdd_obj);
- RETURN(rc);
+ mdd = mdo2mdd(obj);
+
+ /* record only getting user xattrs and acls */
+ if (rc >= 0 &&
+ 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);
}
/*
}
static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
- struct mdd_device *mdd,
- enum changelog_rec_type type, int flags,
- const struct lu_fid *fid,
- struct thandle *handle)
-{
- const struct lu_ucred *uc = lu_ucred(env);
- struct llog_changelog_rec *rec;
- struct lu_buf *buf;
- int reclen;
- int rc;
+ struct mdd_device *mdd,
+ enum changelog_rec_type type,
+ int flags, const struct lu_fid *fid,
+ const char *xattr_name,
+ struct thandle *handle)
+{
+ const struct lu_ucred *uc = lu_ucred(env);
+ struct llog_changelog_rec *rec;
+ struct lu_buf *buf;
+ int reclen;
+ int xflags = CLFE_INVALID;
+ int rc;
- flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
- if (uc != NULL && uc->uc_jobid[0] != '\0')
- flags |= CLF_JOBID;
+ flags = (flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
- reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED));
+ if (uc) {
+ if (uc->uc_jobid[0] != '\0')
+ 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(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 = flags;
rec->cr.cr_type = (__u32)type;
rec->cr.cr_tfid = *fid;
if (flags & CLF_JOBID)
mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
- rc = mdd_changelog_store(env, mdd, rec, handle);
+ if (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, 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);
}
LASSERT(mdd_obj != NULL);
LASSERT(handle != NULL);
- /* Not recording */
- if (!(mdd->mdd_cl.mc_flags & CLM_ON))
- RETURN(0);
- if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
+ if (!mdd_changelog_enabled(env, mdd, type))
RETURN(0);
if (mdd_is_volatile_obj(mdd_obj))
}
rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
- mdo2fid(mdd_obj), handle);
+ mdo2fid(mdd_obj), NULL, handle);
+ if (rc == 0)
+ mdd_obj->mod_cltime = ktime_get();
+
+ RETURN(rc);
+}
+
+int mdd_changelog_data_store_xattr(const struct lu_env *env,
+ struct mdd_device *mdd,
+ enum changelog_rec_type type,
+ int flags, struct mdd_object *mdd_obj,
+ const char *xattr_name,
+ struct thandle *handle)
+{
+ int rc;
+
+ 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);
+
+ 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, flags,
+ mdo2fid(mdd_obj), xattr_name,
+ handle);
if (rc == 0)
mdd_obj->mod_cltime = ktime_get();
int rc;
ENTRY;
- /* Not recording */
- if (!(mdd->mdd_cl.mc_flags & CLM_ON))
- RETURN(0);
- if (!(mdd->mdd_cl.mc_mask & (1 << type)))
- RETURN(0);
-
LASSERT(fid != NULL);
+ /* 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);
GOTO(stop, rc);
rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
- fid, handle);
+ fid, NULL, handle);
stop:
rc = mdd_trans_stop(env, mdd, rc, handle);
}
#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;
}
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;
int rc;
ENTRY;
RETURN(0);
}
- handle = mdd_trans_create(env, mdd);
- if (IS_ERR(handle))
- RETURN(PTR_ERR(handle));
+ /* 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) {
+ la_copy->la_valid |= LA_FLAGS;
+ la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
+
+ /* Flush the possible existing sync requests to OSTs to
+ * keep the order of sync for the current setattr operation
+ * will be sent directly to OSTs. see LU-5152 */
+ 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(stop, rc);
+ if (rc)
+ GOTO(out, rc);
- rc = mdd_trans_start(env, mdd, handle);
- if (rc)
- GOTO(stop, rc);
+ rc = mdd_trans_start(env, mdd, handle);
+ if (rc)
+ GOTO(out, rc);
if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
handle->th_sync = 1;
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 */
+
+ /* 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_copy->la_valid);
- GOTO(stop, rc);
-
-stop:
- rc = mdd_trans_stop(env, mdd, rc, handle);
+ if (handle != NULL)
+ rc = mdd_trans_stop(env, mdd, rc, handle);
return rc;
}
(uc->uc_fsuid != attr->la_uid) &&
!md_capable(uc, CFS_CAP_FOWNER))
RETURN(-EPERM);
- } else {
- if ((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);
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))
- return CL_XATTR;
+ 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 -1;
+ return CL_NONE;
}
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;
- if (mdd_xattr_changelog_type(env, mdd, name) < 0)
+ type = mdd_xattr_changelog_type(env, mdd, name);
+ if (type < 0)
return 0; /* no changelog to store */
- return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
+ return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
}
/*
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,
/* get EA of victim file */
memset(buf_vic, 0, sizeof(*buf_vic));
- rc = mdd_get_lov_ea(env, vic, buf_vic);
+ rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
if (rc < 0) {
if (rc == -ENODATA)
rc = 0;
/* save EA of target file for restore */
memset(buf, 0, sizeof(*buf));
- rc = mdd_get_lov_ea(env, obj, buf);
+ rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
if (rc < 0)
GOTO(out, rc);
GOTO(out, rc);
rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
- if (rc) { /* wtf? */
- int rc2;
+ 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;
- rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
- LU_XATTR_REPLACE, handle);
+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);
- GOTO(out, 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:
mdd_trans_stop(env, mdd, rc, handle);
mdd_write_unlock(env, obj);
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)
+/**
+ * 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 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);
+ 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_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;
+ 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, MOR_TGT_CHILD);
+ mdd_write_lock(env, vic, MOR_TGT_CHILD);
+ } else {
+ mdd_write_lock(env, vic, MOR_TGT_CHILD);
+ mdd_write_lock(env, obj, MOR_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);
+
+ 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:
+ mdd_trans_stop(env, mdd, rc, handle);
+ 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);
if (rc)
RETURN(rc);
- if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && fl == LU_XATTR_MERGE) {
- struct md_object *victim = buf->lb_buf;
+ 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(victim))
+ if (buf->lb_len != sizeof(*mrd))
RETURN(-EINVAL);
rc = mdd_layout_merge_allowed(env, obj, victim);
if (rc)
RETURN(rc);
- /* merge layout of victim as a mirror of obj's. */
- rc = mdd_xattr_merge(env, obj, victim);
+ 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 (cl_type < 0)
GOTO(stop, rc = 0);
- rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
- handle);
+ rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, cl_flags,
+ mdd_obj, name, handle);
EXIT;
stop:
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;
- if (mdd_xattr_changelog_type(env, mdd, name) < 0)
+ type = mdd_xattr_changelog_type(env, mdd, name);
+ if (type < 0)
return 0; /* no changelog to store */
- return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
+ return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
}
/**
if (mdd_xattr_changelog_type(env, mdd, name) < 0)
GOTO(stop, rc = 0);
- rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
+ rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
+ name, handle);
EXIT;
stop:
}
/*
- * 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);
+ 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,
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;
- __u32 fst_gen, snd_gen, saved_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);
mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
mdd_write_lock(env, snd_o, MOR_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);
else
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;
*/
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);
-
- rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
- if (rc)
- GOTO(stop, rc);
- EXIT;
-
stop:
rc = mdd_trans_stop(env, mdd, rc, handle);
+ /* Truncate local DOM data if all went well */
+ if (!rc && dom_o)
+ mdd_dom_data_truncate(env, mdd, dom_o);
+
mdd_write_unlock(env, snd_o);
mdd_write_unlock(env, fst_o);
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;
}
if (rc)
return rc;
- return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
+ return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
+ handle);
}
/* For PFL, this is used to instantiate necessary component objects. */
-int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
- struct md_layout_change *mlc)
+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_object *mdd_obj = md2mdd_obj(obj);
- struct mdd_device *mdd = mdo2mdd(obj);
- struct thandle *handle;
+ struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
int rc;
ENTRY;
- handle = mdd_trans_create(env, mdd);
- if (IS_ERR(handle))
- RETURN(PTR_ERR(handle));
+ if (mlc->mlc_opc != MD_LAYOUT_WRITE)
+ RETURN(-ENOTSUPP);
- rc = mdd_declare_layout_change(env, mdd, mdd_obj, mlc, handle);
+ 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)
- GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
+ RETURN(rc == -EALREADY ? 0 : rc);
rc = mdd_trans_start(env, mdd, handle);
if (rc)
- GOTO(stop, rc);
+ RETURN(rc);
- mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
- rc = mdo_layout_change(env, mdd_obj, mlc, handle);
- mdd_write_unlock(env, mdd_obj);
+ mdd_write_lock(env, obj, MOR_TGT_CHILD);
+ rc = mdo_layout_change(env, obj, mlc, handle);
+ mdd_write_unlock(env, obj);
if (rc)
- GOTO(stop, rc);
+ RETURN(rc);
- rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
-stop:
- RETURN(mdd_trans_stop(env, mdd, rc, handle));
+ 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, MOR_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, MOR_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;
}
void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
return 0;
}
- if (flags & FMODE_READ)
+ if (flags & MDS_FMODE_READ)
res |= MAY_READ;
- if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
+ if (flags & (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
res |= MAY_WRITE;
if (flags & MDS_FMODE_EXEC)
res = MAY_EXEC;
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) &&
+ likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
RETURN(-ENOENT);
if (S_ISLNK(attr->la_mode))
/* 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 ((flag & MDS_FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
RETURN(-EPERM);
if (flag & MDS_OPEN_TRUNC)
RETURN(-EPERM);
int flags)
{
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;
mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
GOTO(out, rc);
rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
- if (rc != 0)
+ 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);
- mdd_obj->mod_count++;
+find:
+ /* look for existing opener in list under mdd_write_lock */
+ mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, flags);
+
+ if (!mou) {
+ int rc2;
+
+ /* add user to list */
+ mou = mdd_obj_user_alloc(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);
+ }
+ }
+
+ mdd_changelog(env, type, flags, md_dev, mdo2fid(mdd_obj));
+
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);
}
int is_orphan = 0;
int rc;
bool blocked = false;
+ int last_close_by_uid = 0;
+ const struct lu_ucred *uc = lu_ucred(env);
ENTRY;
if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
if (rc)
GOTO(stop, rc);
- 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);
mdd_write_lock(env, mdd_obj, MOR_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);
}
mdd_obj->mod_count--; /*release open count */
+ /* under mdd write lock */
+ /* If recording, see if we need to remove UID from list */
+ if (mdd_changelog_enabled(env, mdd, CL_OPEN)) {
+ 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, mode);
+ if (mou) {
+ mou->mou_opencount--;
+ if (mou->mou_opencount == 0) {
+ mdd_obj_user_remove(mdd_obj, mou);
+ last_close_by_uid = 1;
+ }
+ }
+ }
+
if (!is_orphan || blocked)
GOTO(out, rc = 0);
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)),
out:
mdd_write_unlock(env, mdd_obj);
- if (!rc && !blocked &&
- (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
+ 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)) &&
+ (mode & (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);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff