-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Pravin B Shelar <pravin.shelar@sun.com>
*/
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
#define DEBUG_SUBSYSTEM S_MDS
#include <obd.h>
#include "mdd_internal.h"
const char orph_index_name[] = "PENDING";
+const char *dotdot = "..";
enum {
ORPH_OP_UNLINK,
static struct dt_key* orph_key_fill(const struct lu_env *env,
const struct lu_fid *lf, __u32 op)
{
- char *key = mdd_env_info(env)->mti_orph_key;
+ char *key = mdd_env_info(env)->mti_key;
int rc;
LASSERT(key);
static struct dt_key* orph_key_fill_18(const struct lu_env *env,
const struct lu_fid *lf)
{
- char *key = mdd_env_info(env)->mti_orph_key;
+ char *key = mdd_env_info(env)->mti_key;
int rc;
LASSERT(key);
return ERR_PTR(rc);
}
-static int orphan_key_to_fid(char *key, struct lu_fid *lf)
-{
- int rc = 0;
- unsigned int op;
-
- rc = sscanf(key, ORPHAN_FILE_NAME_FORMAT,
- (long long unsigned int *)&lf->f_seq, &lf->f_oid,
- &lf->f_ver, &op);
- if (rc == 4)
- return 0;
-
- /* build igif */
- rc = sscanf(key, ORPHAN_FILE_NAME_FORMAT_18,
- (long long unsigned int *)&lf->f_seq, &lf->f_oid);
- if (rc == 2) {
- lf->f_ver = 0;
- return 0;
- }
-
- CERROR("can not parse orphan file name %s\n",key);
- return -EINVAL;
-}
-
static inline void mdd_orphan_write_lock(const struct lu_env *env,
struct mdd_device *mdd)
{
}
+int orph_declare_index_insert(const struct lu_env *env,
+ struct mdd_object *obj,
+ cfs_umode_t mode, struct thandle *th)
+{
+ struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
+ struct dt_key *key;
+ int rc;
+
+ key = orph_key_fill(env, mdo2fid(obj), ORPH_OP_UNLINK);
+
+ rc = dt_declare_insert(env, mdd->mdd_orphans, NULL, key, th);
+ if (rc)
+ 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, NULL, dotdot, th);
+
+ return rc;
+}
+
static int orph_index_insert(const struct lu_env *env,
struct mdd_object *obj,
__u32 op,
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);
- const struct dt_key *dotdot = (const struct dt_key *) "..";
int rc;
ENTRY;
LASSERT(mdd_write_locked(env, obj) != 0);
LASSERT(!(obj->mod_flags & ORPHAN_OBJ));
- LASSERT(obj->mod_count > 0);
mdd_orphan_write_lock(env, mdd);
if (!dt_try_as_dir(env, next))
goto out;
next->do_index_ops->dio_delete(env, next,
- dotdot, th, BYPASS_CAPA);
+ (const struct dt_key *)dotdot,
+ th, BYPASS_CAPA);
next->do_index_ops->dio_insert(env, next,
- (struct dt_rec *) lf_dor,
- dotdot, th, BYPASS_CAPA, 1);
+ (struct dt_rec *)lf_dor,
+ (const struct dt_key *)dotdot,
+ th, BYPASS_CAPA, 1);
out:
if (rc == 0)
RETURN(rc);
}
-/**
- * destroy osd object on mdd and associated ost objects.
- *
- * \param obj orphan object
- * \param mdd used for sending llog msg to osts
- *
- * \retval 0 success
- * \retval -ve error
- */
-static int orphan_object_kill(const struct lu_env *env,
+int orph_declare_index_delete(const struct lu_env *env,
struct mdd_object *obj,
- struct mdd_device *mdd,
+
struct thandle *th)
{
- struct lu_attr *la = &mdd_env_info(env)->mti_la;
- int rc = 0;
- ENTRY;
+ struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
+ struct dt_key *key;
+ int rc;
- /* No need to lock this object as its recovery phase, and
- * no other thread can access it. But we need to lock it
- * as its precondition for osd api we using. */
+ key = orph_key_fill(env, mdo2fid(obj), ORPH_OP_UNLINK);
+
+ rc = dt_declare_delete(env, mdd->mdd_orphans, key, th);
+ if (rc)
+ return rc;
+
+ rc = mdo_declare_ref_del(env, obj, th);
+ if (rc)
+ return rc;
- mdo_ref_del(env, obj, th);
if (S_ISDIR(mdd_object_type(obj))) {
- mdo_ref_del(env, obj, th);
- mdd_orphan_ref_del(env, mdd, th);
- } else {
- /* regular file , cleanup linked ost objects */
- rc = mdd_la_get(env, obj, la, BYPASS_CAPA);
- if (rc == 0)
- rc = mdd_lov_destroy(env, mdd, obj, la);
+ rc = mdo_declare_ref_del(env, obj, th);
+ if (rc)
+ return rc;
+
+ rc = dt_declare_ref_del(env, mdd->mdd_orphans, th);
}
- RETURN(rc);
+
+ return rc;
}
static int orph_index_delete(const struct lu_env *env,
int rc = 0;
ENTRY;
- mdd_txn_param_build(env, mdd, MDD_TXN_UNLINK_OP);
- th = mdd_trans_start(env, mdd);
+ th = mdd_trans_create(env, mdd);
if (IS_ERR(th)) {
CERROR("Cannot get thandle\n");
RETURN(-ENOMEM);
}
+ rc = orph_declare_index_delete(env, obj, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = mdo_declare_destroy(env, obj, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = mdd_trans_start(env, mdd, th);
+ if (rc)
+ GOTO(stop, rc);
mdd_write_lock(env, obj, MOR_TGT_CHILD);
if (likely(obj->mod_count == 0)) {
mdd_orphan_write_lock(env, mdd);
rc = mdd_orphan_delete_obj(env, mdd, key, th);
- if (!rc)
- orphan_object_kill(env, obj, mdd, th);
- else
+ if (rc == 0) {
+ mdo_ref_del(env, obj, th);
+ if (S_ISDIR(mdd_object_type(obj))) {
+ mdo_ref_del(env, obj, th);
+ mdd_orphan_ref_del(env, mdd, th);
+ }
+ rc = mdo_destroy(env, obj, th);
+ } else
CERROR("could not delete object: rc = %d\n",rc);
mdd_orphan_write_unlock(env, mdd);
}
mdd_write_unlock(env, obj);
+
+stop:
mdd_trans_stop(env, mdd, 0, th);
RETURN(rc);
}
+/**
+ * Delete unused orphan with FID \a lf from PENDING directory
+ *
+ * \param mdd MDD device finishing recovery
+ * \param lf FID of file or directory to delete
+ * \param key cookie for this entry in index iterator
+ *
+ * \retval 0 success
+ * \retval -ve error
+ */
static int orph_key_test_and_del(const struct lu_env *env,
struct mdd_device *mdd,
struct lu_fid *lf,
rc = -EBUSY;
if (mdo->mod_count == 0) {
- CWARN("Found orphan! Delete it\n");
+ CDEBUG(D_HA, "Found orphan "DFID", delete it\n", PFID(lf));
rc = orphan_object_destroy(env, mdo, key);
+ if (rc) /* so replay-single.sh test_37 works */
+ CERROR("%s: error unlinking orphan "DFID" from "
+ "PENDING: rc = %d\n",
+ mdd2obd_dev(mdd)->obd_name, PFID(lf), rc);
} else {
mdd_write_lock(env, mdo, MOR_TGT_CHILD);
if (likely(mdo->mod_count > 0)) {
- CDEBUG(D_HA, "Found orphan, open count = %d\n",
- mdo->mod_count);
+ 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);
return rc;
}
+/**
+ * delete unreferenced files and directories in the PENDING directory
+ *
+ * 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).
+ *
+ * \param mdd MDD device finishing recovery
+ *
+ * \retval 0 success
+ * \retval -ve error
+ */
static int orph_index_iterate(const struct lu_env *env,
- struct mdd_device *mdd)
+ struct mdd_device *mdd)
{
- struct dt_object *dor = mdd->mdd_orphans;
- char *mti_key = mdd_env_info(env)->mti_orph_key;
- const struct dt_it_ops *iops;
- struct dt_it *it;
- char *key;
- struct lu_fid fid;
- int result = 0;
+ struct dt_object *dor = mdd->mdd_orphans;
+ struct lu_dirent *ent = &mdd_env_info(env)->mti_ent;
+ const struct dt_it_ops *iops;
+ struct dt_it *it;
+ struct lu_fid fid;
int key_sz = 0;
int rc;
__u64 cookie;
ENTRY;
/* In recovery phase, do not need for any lock here */
-
iops = &dor->do_index_ops->dio_it;
it = iops->init(env, dor, LUDA_64BITHASH, BYPASS_CAPA);
- if (!IS_ERR(it)) {
- result = iops->load(env, it, 0);
- if (result > 0) {
- /* main cycle */
- do {
-
- key = (void *)iops->key(env, it);
- if (IS_ERR(key)) {
- CERROR("key failed when clean pending.\n");
- goto next;
- }
- key_sz = iops->key_size(env, it);
-
- /* filter out "." and ".." entries from
- * PENDING dir. */
- if (key_sz < 8)
- goto next;
-
- memcpy(mti_key, key, key_sz);
- mti_key[key_sz] = 0;
-
- if (orphan_key_to_fid(mti_key, &fid))
- goto next;
- if (!fid_is_sane(&fid)) {
- CERROR("fid is not sane when clean pending.\n");
- goto next;
- }
-
- /* kill orphan object */
- cookie = iops->store(env, it);
- iops->put(env, it);
- rc = orph_key_test_and_del(env, mdd, &fid,
- (struct dt_key *)mti_key);
-
- /* after index delete reset iterator */
- if (!rc)
- result = iops->get(env, it,
- (const void *)"");
- else
- result = iops->load(env, it, cookie);
-next:
- result = iops->next(env, it);
- } while (result == 0);
- result = 0;
- } else if (result == 0) {
- CERROR("Input/Output for clean pending.\n");
- /* Index contains no zero key? */
- result = -EIO;
- }
- iops->put(env, it);
- iops->fini(env, it);
- } else {
- result = PTR_ERR(it);
- CERROR("Cannot clean pending (%d).\n", result);
+ if (IS_ERR(it)) {
+ rc = PTR_ERR(it);
+ CERROR("%s: cannot clean PENDING: rc = %d\n",
+ mdd2obd_dev(mdd)->obd_name, rc);
+ GOTO(out, rc);
+ }
+
+ rc = iops->load(env, it, 0);
+ if (rc < 0)
+ GOTO(out_put, rc);
+ if (rc == 0) {
+ CERROR("%s: error loading iterator to clean PENDING\n",
+ mdd2obd_dev(mdd)->obd_name);
+ /* Index contains no zero key? */
+ GOTO(out_put, rc = -EIO);
}
- RETURN(result);
+ do {
+ 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 PENDING\n",
+ mdd2obd_dev(mdd)->obd_name, PFID(&fid));
+ goto next;
+ }
+
+ /* kill orphan object */
+ cookie = iops->store(env, it);
+ iops->put(env, it);
+ rc = orph_key_test_and_del(env, mdd, &fid,
+ (struct dt_key *)ent->lde_name);
+
+ /* after index delete reset iterator */
+ if (rc == 0)
+ rc = iops->get(env, it, (const void *)"");
+ else
+ rc = iops->load(env, it, cookie);
+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;
}
+/**
+ * open the PENDING directory for device \a mdd
+ *
+ * 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).
+ * \param d mdd device being started.
+ *
+ * \retval 0 success
+ * \retval -ve index operation error.
+ *
+ */
int orph_index_init(const struct lu_env *env, struct mdd_device *mdd)
{
struct lu_fid fid;
}
/**
- * Iterate orphan index to cleanup orphan objects in case of recovery.
+ * Iterate orphan index to cleanup orphan objects after recovery is done.
* \param d mdd device in recovery.
- *
*/
-
int __mdd_orphan_cleanup(const struct lu_env *env, struct mdd_device *d)
{
return orph_index_iterate(env, d);
}
/**
- * delete an orphan \a obj from orphan index.
+ * add an orphan \a obj to the orphan index.
* \param obj file or directory.
* \param th transaction for index insert.
*
* \retval 0 success
* \retval -ve index operation error.
*/
-
int __mdd_orphan_add(const struct lu_env *env,
struct mdd_object *obj, struct thandle *th)
{
* \retval 0 success
* \retval -ve index operation error.
*/
-
int __mdd_orphan_del(const struct lu_env *env,
struct mdd_object *obj, struct thandle *th)
{