Viewing: mdd_orphans.c
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
* Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*
* Orphan handling code
*
* Author: Mike Pershin <tappro@clusterfs.com>
* Pravin B Shelar <pravin.shelar@sun.com>
*/
#define DEBUG_SUBSYSTEM S_MDS
#include <obd.h>
#include <obd_class.h>
#include <obd_support.h>
#include <lustre_fid.h>
#include "mdd_internal.h"
static const char mdd_orphan_index_name[] = MDT_ORPHAN_DIR;
static const char dotdot[] = "..";
enum {
ORPH_OP_UNLINK,
};
/* obsolete after 2.11, needed for upgrades from older 2.x versions */
#define ORPHAN_FILE_NAME_FORMAT_20 "%016llx:%08x:%08x:%2x"
static struct dt_key *mdd_orphan_key_fill(const struct lu_env *env,
const struct lu_fid *lf)
{
char *key = mdd_env_info(env)->mdi_key;
LASSERT(key);
if (!(MDI_KEEP_KEY & mdd_env_info(env)->mdi_flags))
snprintf(key, sizeof(mdd_env_info(env)->mdi_key),
DFID_NOBRACE, PFID(lf));
return (struct dt_key *)key;
}
/* compatibility with orphan files created in versions before 2.11 */
static struct dt_key *mdd_orphan_key_fill_20(const struct lu_env *env,
const struct lu_fid *lf)
{
char *key = mdd_env_info(env)->mdi_key;
LASSERT(key);
if (!(MDI_KEEP_KEY & mdd_env_info(env)->mdi_flags))
snprintf(key, sizeof(mdd_env_info(env)->mdi_key),
ORPHAN_FILE_NAME_FORMAT_20,
fid_seq(lf), fid_oid(lf), fid_ver(lf),
ORPH_OP_UNLINK);
return (struct dt_key *)key;
}
static inline int mdd_orphan_insert_obj(const struct lu_env *env,
struct mdd_device *mdd,
struct mdd_object *obj,
struct thandle *th)
{
struct dt_insert_rec *rec = &mdd_env_info(env)->mdi_dt_rec;
struct dt_object *dor = mdd->mdd_orphans;
const struct lu_fid *lf = mdd_object_fid(obj);
struct dt_key *key = mdd_orphan_key_fill(env, lf);
rec->rec_fid = lf;
rec->rec_type = mdd_object_type(obj);
return dt_insert(env, dor, (const struct dt_rec *)rec, key, th);
}
int mdd_orphan_declare_insert(const struct lu_env *env, struct mdd_object *obj,
umode_t mode, struct thandle *th)
{
struct dt_insert_rec *rec = &mdd_env_info(env)->mdi_dt_rec;
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_key *key;
int rc;
key = mdd_orphan_key_fill(env, mdd_object_fid(obj));
rec->rec_fid = mdd_object_fid(obj);
rec->rec_type = mode;
rc = dt_declare_insert(env, mdd->mdd_orphans,
(const struct dt_rec *)rec, key, th);
if (rc != 0)
return rc;
rc = mdo_declare_ref_add(env, obj, th);
if (rc)
return rc;
if (!S_ISDIR(mode))
return 0;
rc = mdo_declare_ref_add(env, obj, th);
if (rc)
return rc;
rc = dt_declare_ref_add(env, mdd->mdd_orphans, th);
if (rc)
return rc;
rc = mdo_declare_index_delete(env, obj, dotdot, th);
if (rc)
return rc;
rc = mdo_declare_index_insert(env, obj,
lu_object_fid(&mdd->mdd_orphans->do_lu),
S_IFDIR, dotdot, th);
return rc;
}
/**
* mdd_orphan_insert() - add an orphan @obj to the orphan index.
* @env: current lustre environment
* @obj: file or directory.
* @th: transaction for index insert.
*
* pre requirement: obj nlink == 0 && obj->mod_count != 0 (link count is
* zero && At-least is still opened by one or more process)
*
* Return:
* * %0 on success
* * %negative index operation error.
*/
int mdd_orphan_insert(const struct lu_env *env, struct mdd_object *obj,
struct thandle *th)
{
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_object *dor = mdd->mdd_orphans;
const struct lu_fid *lf_dor = lu_object_fid(&dor->do_lu);
struct dt_object *next = mdd_object_child(obj);
struct dt_insert_rec *rec = &mdd_env_info(env)->mdi_dt_rec;
int rc;
ENTRY;
LASSERT(mdd_write_locked(env, obj) != 0);
LASSERT(!(obj->mod_flags & ORPHAN_OBJ));
dt_write_lock(env, mdd->mdd_orphans, DT_TGT_ORPHAN);
rc = mdd_orphan_insert_obj(env, mdd, obj, th);
if (rc)
GOTO(out, rc);
mdo_ref_add(env, obj, th);
if (!S_ISDIR(mdd_object_type(obj)))
GOTO(out, rc = 0);
mdo_ref_add(env, obj, th);
dt_ref_add(env, mdd->mdd_orphans, th);
/* try best to fixup directory, do not return errors from here */
if (!dt_try_as_dir(env, next, true))
GOTO(out, rc = 0);
dt_delete(env, next, (const struct dt_key *)dotdot, th);
rec->rec_fid = lf_dor;
rec->rec_type = S_IFDIR;
dt_insert(env, next, (const struct dt_rec *)rec,
(const struct dt_key *)dotdot, th);
out:
if (rc == 0)
obj->mod_flags |= ORPHAN_OBJ;
dt_write_unlock(env, mdd->mdd_orphans);
RETURN(rc);
}
int mdd_orphan_declare_delete(const struct lu_env *env, struct mdd_object *obj,
struct thandle *th)
{
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_key *key;
int rc;
key = mdd_orphan_key_fill(env, mdd_object_fid(obj));
rc = dt_declare_delete(env, mdd->mdd_orphans, key, th);
if (rc)
return rc;
if (!mdd_object_exists(obj))
return -ENOENT;
rc = mdo_declare_ref_del(env, obj, th);
if (rc)
return rc;
if (S_ISDIR(mdd_object_type(obj))) {
rc = mdo_declare_ref_del(env, obj, th);
if (rc)
return rc;
rc = dt_declare_ref_del(env, mdd->mdd_orphans, th);
}
return rc;
}
/**
* mdd_orphan_delete() - delete an orphan @obj from orphan index.
* @env: current lustre environment
* @obj: file or directory.
* @th: transaction for index deletion and object destruction.
*
* pre: obj->mod_count == 0 && ORPHAN_OBJ is set for obj.
*
* Return:
* * %0 on success
* * %negative index operation error.
*/
int mdd_orphan_delete(const struct lu_env *env, struct mdd_object *obj,
struct thandle *th)
{
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_object *dor = mdd->mdd_orphans;
struct dt_key *key;
int rc = 0;
ENTRY;
LASSERT(mdd_write_locked(env, obj) != 0);
LASSERT(obj->mod_flags & ORPHAN_OBJ);
LASSERT(obj->mod_count == 0);
LASSERT(dor);
key = mdd_orphan_key_fill(env, mdd_object_fid(obj));
dt_write_lock(env, mdd->mdd_orphans, DT_TGT_ORPHAN);
if (CFS_FAIL_CHECK(OBD_FAIL_MDS_ORPHAN_DELETE))
goto ref_del;
rc = dt_delete(env, mdd->mdd_orphans, key, th);
if (rc == -ENOENT) {
key = mdd_orphan_key_fill_20(env, mdd_object_fid(obj));
rc = dt_delete(env, mdd->mdd_orphans, key, th);
}
ref_del:
if (!rc) {
/* lov objects will be destroyed by caller */
mdo_ref_del(env, obj, th);
if (S_ISDIR(mdd_object_type(obj))) {
mdo_ref_del(env, obj, th);
dt_ref_del(env, mdd->mdd_orphans, th);
}
obj->mod_flags &= ~ORPHAN_OBJ;
} else {
CERROR("%s: could not delete orphan object "DFID": rc = %d\n",
mdd_obj_dev_name(obj), PFID(mdd_object_fid(obj)), rc);
}
dt_write_unlock(env, mdd->mdd_orphans);
RETURN(rc);
}
static int mdd_orphan_destroy(const struct lu_env *env, struct mdd_object *obj,
struct dt_key *key)
{
struct thandle *th = NULL;
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
int rc = 0, rc1 = 0;
ENTRY;
th = mdd_trans_create(env, mdd);
if (IS_ERR(th)) {
rc = PTR_ERR(th);
if (rc != -EINPROGRESS)
CERROR("%s: cannot get orphan thandle: rc = %d\n",
mdd2obd_dev(mdd)->obd_name, rc);
RETURN(rc);
}
rc = mdd_orphan_declare_delete(env, obj, th);
if (rc && rc != -ENOENT)
GOTO(stop, rc);
if (rc != -ENOENT) {
rc = mdo_declare_destroy(env, obj, th);
if (rc && rc != -ENOENT)
GOTO(stop, rc);
}
rc = mdd_trans_start(env, mdd, th);
if (rc)
GOTO(stop, rc);
mdd_write_lock(env, obj, DT_TGT_CHILD);
if (likely(obj->mod_count == 0)) {
dt_write_lock(env, mdd->mdd_orphans, DT_TGT_ORPHAN);
rc = dt_delete(env, mdd->mdd_orphans, key, th);
/* We should remove object even dt_delete failed */
if (mdd_object_exists(obj) &&
!lu_object_is_dying(obj->mod_obj.mo_lu.lo_header)) {
mdo_ref_del(env, obj, th);
if (S_ISDIR(mdd_object_type(obj))) {
mdo_ref_del(env, obj, th);
dt_ref_del(env, mdd->mdd_orphans, th);
}
rc1 = mdo_destroy(env, obj, th);
}
dt_write_unlock(env, mdd->mdd_orphans);
}
mdd_write_unlock(env, obj);
stop:
mdd_trans_stop(env, mdd, 0, th);
RETURN(rc ? rc : rc1);
}
/**
* mdd_orphan_key_test_and_delete() - Delete unused orphan with FID @lf from
* PENDING directory
* @env: current lustre environment
* @mdd: MDD device finishing recovery
* @lf: FID of file or directory to delete
* @key: cookie for this entry in index iterator
*
* Return:
* * %0 success
* * %negative error
*/
static int mdd_orphan_key_test_and_delete(const struct lu_env *env,
struct mdd_device *mdd,
struct lu_fid *lf, struct dt_key *key)
{
struct mdd_object *mdo;
int rc;
mdo = mdd_object_find(env, mdd, lf);
if (IS_ERR(mdo))
return PTR_ERR(mdo);
rc = -EBUSY;
if (mdo->mod_count == 0) {
CDEBUG(D_HA, "Found orphan "DFID", delete it\n", PFID(lf));
rc = mdd_orphan_destroy(env, mdo, key);
if (rc) /* below message checked in replay-single.sh test_37 */
CERROR("%s: error unlinking orphan "DFID": rc = %d\n",
mdd2obd_dev(mdd)->obd_name, PFID(lf), rc);
} else {
mdd_write_lock(env, mdo, DT_TGT_CHILD);
if (likely(mdo->mod_count > 0)) {
CDEBUG(D_HA, "Found orphan "DFID" count %d, skip it\n",
PFID(lf), mdo->mod_count);
mdo->mod_flags |= ORPHAN_OBJ;
}
mdd_write_unlock(env, mdo);
}
mdd_object_put(env, mdo);
return rc;
}
/**
* mdd_orphan_index_iterate() - delete unreferenced files and directories in the
* PENDING directory
* @env: current lustre environment
* @thread: info about orphan cleanup thread
*
* Files that remain in PENDING after client->MDS recovery has completed
* have to be referenced (opened) by some client during recovery, or they
* will be deleted here (for clients that did not complete recovery).
*
* Return:
* * %0 success
* * %negative error
*/
static int mdd_orphan_index_iterate(const struct lu_env *env,
struct mdd_generic_thread *thread)
{
struct mdd_device *mdd = (struct mdd_device *)thread->mgt_data;
struct dt_object *dor = mdd->mdd_orphans;
struct lu_dirent *ent = &mdd_env_info(env)->mdi_ent;
const struct dt_it_ops *iops;
struct dt_it *it;
struct lu_fid fid;
int key_sz = 0;
int rc;
ENTRY;
iops = &dor->do_index_ops->dio_it;
it = iops->init(env, dor, LUDA_64BITHASH);
if (IS_ERR(it)) {
rc = PTR_ERR(it);
CERROR("%s: cannot clean '%s': rc = %d\n",
mdd2obd_dev(mdd)->obd_name, mdd_orphan_index_name, rc);
GOTO(out, rc);
}
rc = iops->load(env, it, 0);
if (rc < 0)
GOTO(out_put, rc);
if (rc == 0) {
rc = -EIO;
CERROR("%s: error loading iterator to clean '%s': rc = %d\n",
mdd2obd_dev(mdd)->obd_name, mdd_orphan_index_name, rc);
/* Index contains no zero key? */
GOTO(out_put, rc);
}
mdd_env_info(env)->mdi_flags |= MDI_KEEP_KEY;
do {
if (thread->mgt_abort)
break;
key_sz = iops->key_size(env, it);
/* filter out "." and ".." entries from PENDING dir. */
if (key_sz < 8)
goto next;
rc = iops->rec(env, it, (struct dt_rec *)ent, LUDA_64BITHASH);
if (rc != 0) {
CERROR("%s: fail to get FID for orphan it: rc = %d\n",
mdd2obd_dev(mdd)->obd_name, rc);
goto next;
}
fid_le_to_cpu(&fid, &ent->lde_fid);
if (!fid_is_sane(&fid)) {
CERROR("%s: bad FID "DFID" cleaning '%s'\n",
mdd2obd_dev(mdd)->obd_name, PFID(&fid),
mdd_orphan_index_name);
goto next;
}
/* kill orphan object */
iops->put(env, it);
rc = mdd_orphan_key_test_and_delete(env, mdd, &fid,
(struct dt_key *)ent->lde_name);
/* after index delete reset iterator */
if (rc == 0)
rc = iops->get(env, it, (const void *)"");
next:
rc = iops->next(env, it);
} while (rc == 0);
GOTO(out_put, rc = 0);
out_put:
iops->put(env, it);
iops->fini(env, it);
out:
return rc;
}
/**
* mdd_orphan_index_init() - open the PENDING directory for device @mdd
* @env: current lustre environment
* @mdd: mdd device being started.
*
* The PENDING directory persistently tracks files and directories that were
* unlinked from the namespace (nlink == 0) but are still held open by clients.
* Those inodes shouldn't be deleted if the MDS crashes, because the clients
* would not be able to recover and reopen those files. Instead, these inodes
* are linked into the PENDING directory on disk, and only deleted if all
* clients close them, or the MDS finishes client recovery without any client
* reopening them (i.e. former clients didn't join recovery).
*
* Return:
* * %0 success
* * %negative index operation error.
*/
int mdd_orphan_index_init(const struct lu_env *env, struct mdd_device *mdd)
{
struct lu_fid fid;
struct dt_object *d;
int rc = 0;
ENTRY;
/* create PENDING dir */
fid_zero(&fid);
rc = mdd_local_file_create(env, mdd, &mdd->mdd_local_root_fid,
mdd_orphan_index_name, S_IFDIR | 0755, &fid);
if (rc < 0)
RETURN(rc);
d = dt_locate(env, mdd->mdd_child, &fid);
if (IS_ERR(d))
RETURN(PTR_ERR(d));
LASSERT(lu_object_exists(&d->do_lu));
if (!dt_try_as_dir(env, d, true)) {
CERROR("%s: orphan dir '%s' is not an index: rc = %d\n",
mdd2obd_dev(mdd)->obd_name, mdd_orphan_index_name, rc);
dt_object_put(env, d);
RETURN(-ENOTDIR);
}
mdd->mdd_orphans = d;
RETURN(0);
}
void mdd_orphan_index_fini(const struct lu_env *env, struct mdd_device *mdd)
{
ENTRY;
if (mdd->mdd_orphans != NULL) {
dt_object_put(env, mdd->mdd_orphans);
mdd->mdd_orphans = NULL;
}
EXIT;
}
static int mdd_orphan_cleanup_thread(void *args)
{
struct mdd_generic_thread *thread = (struct mdd_generic_thread *)args;
struct lu_env *env = NULL;
int rc;
ENTRY;
complete(&thread->mgt_started);
OBD_ALLOC_PTR(env);
if (env == NULL)
GOTO(out, rc = -ENOMEM);
rc = lu_env_init(env, LCT_MD_THREAD);
if (rc)
GOTO(out, rc);
rc = mdd_orphan_index_iterate(env, thread);
lu_env_fini(env);
GOTO(out, rc);
out:
OBD_FREE_PTR(env);
complete(&thread->mgt_finished);
return rc;
}
/**
* mdd_orphan_cleanup() - Iterate orphan index to cleanup orphan objects after
* recovery is done.
* @env: current lustre environment
* @d: mdd device in recovery.
*
* Return:
* * %0 success
* * %negative error
*/
int mdd_orphan_cleanup(const struct lu_env *env, struct mdd_device *d)
{
int rc = -ENOMEM;
char *name = NULL;
OBD_ALLOC(name, MTI_NAME_MAXLEN);
if (name == NULL)
goto out;
snprintf(name, MTI_NAME_MAXLEN, "orph_%s", mdd2obd_dev(d)->obd_name);
rc = mdd_generic_thread_start(&d->mdd_orphan_cleanup_thread,
mdd_orphan_cleanup_thread, d, name);
out:
if (rc)
CERROR("%s: start orphan cleanup thread failed: rc = %d\n",
mdd2obd_dev(d)->obd_name, rc);
OBD_FREE(name, MTI_NAME_MAXLEN);
return rc;
}
