Whamcloud - gitweb
- landing b_fid.
[fs/lustre-release.git] / lustre / liblustre / super.c
index 6971d87..7b5fc95 100644 (file)
@@ -3,7 +3,7 @@
  *
  * Lustre Light Super operations
  *
- *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ *  Copyright (c) 2002-2004 Cluster File Systems, Inc.
  *
  *   This file is part of Lustre, http://www.lustre.org.
  *
 #include <sys/stat.h>
 #include <sys/fcntl.h>
 #include <sys/queue.h>
+
 #ifndef __CYGWIN__
 # include <sys/statvfs.h>
 #else
 # include <sys/statfs.h>
 #endif
 
-#include <sysio.h>
 #include <fs.h>
+#include <sysio.h>
 #include <mount.h>
 #include <inode.h>
 #include <file.h>
 
 #undef LIST_HEAD
-
 #include "llite_lib.h"
 
+#ifndef MAY_EXEC
+#  define MAY_EXEC        1
+#  define MAY_WRITE       2
+#  define MAY_READ        4
+#endif
+
+#define S_IXUGO (S_IXUSR|S_IXGRP|S_IXOTH)
+
+static int ll_permission(struct inode *inode, int mask)
+{
+        struct llu_inode_info *lli = llu_i2info(inode);
+        mode_t mode = lli->lli_st_mode;
+
+        if (current->fsuid == lli->lli_st_uid)
+                mode >>= 6;
+        else if (in_group_p(lli->lli_st_gid))
+                mode >>= 3;
+
+        if ((mode & mask & (MAY_READ|MAY_WRITE|MAY_EXEC)) == mask)
+                return 0;
+
+        if ((mask & (MAY_READ|MAY_WRITE)) ||
+            (lli->lli_st_mode & S_IXUGO))
+                if (capable(CAP_DAC_OVERRIDE))
+                        return 0;
+
+        if (mask == MAY_READ ||
+            (S_ISDIR(lli->lli_st_mode) && !(mask & MAY_WRITE))) {
+                if (capable(CAP_DAC_READ_SEARCH))
+                        return 0;
+        }
+
+        return -EACCES;
+}
+
 static void llu_fsop_gone(struct filesys *fs)
 {
         struct llu_sb_info *sbi = (struct llu_sb_info *) fs->fs_private;
-        struct obd_device *obd = class_exp2obd(sbi->ll_mdc_exp);
-        struct ll_fid rootfid;
+        struct obd_device *obd = class_exp2obd(sbi->ll_lmv_exp);
+        struct lustre_cfg lcfg;
+        int next = 0;
         ENTRY;
 
         list_del(&sbi->ll_conn_chain);
-        obd_disconnect(sbi->ll_osc_exp, 0);
+        obd_disconnect(sbi->ll_lov_exp, 0);
+        obd_disconnect(sbi->ll_lmv_exp, 0);
 
-        /* NULL request to force sync on the MDS, and get the last_committed
-         * value to flush remaining RPCs from the sending queue on client.
-         *
-         * XXX This should be an mdc_sync() call to sync the whole MDS fs,
-         *     which we can call for other reasons as well.
-         */
-        if (!obd->obd_no_recov)
-                mdc_getstatus(sbi->ll_mdc_exp, &rootfid);
+        while ((obd = class_devices_in_group(&sbi->ll_sb_uuid, &next)) != NULL)
+        {
+                int err;
 
-        obd_disconnect(sbi->ll_mdc_exp, 0);
+                LCFG_INIT(lcfg, LCFG_CLEANUP, obd->obd_name);
+                err = class_process_config(&lcfg);
+                if (err) {
+                        CERROR("cleanup failed: %s\n", obd->obd_name);
+                }
 
-        OBD_FREE(sbi, sizeof(*sbi));
+                LCFG_INIT(lcfg, LCFG_DETACH, obd->obd_name);
+                err = class_process_config(&lcfg);
+                if (err) {
+                        CERROR("detach failed: %s\n", obd->obd_name);
+                }
+        }
 
+        obd_disconnect(sbi->ll_lmv_exp, 0);
+        OBD_FREE(sbi, sizeof(*sbi));
         EXIT;
 }
 
-static struct inode_ops llu_inode_ops;
+struct inode_ops llu_inode_ops;
 
 void llu_update_inode(struct inode *inode, struct mds_body *body,
                       struct lov_stripe_md *lsm)
@@ -96,8 +139,19 @@ void llu_update_inode(struct inode *inode, struct mds_body *body,
                 }
         }
 
+        /*
+         * updating inode lustre id. We try optimize things a little bit here,
+         * thus do not do it all the time.
+         */
+        if (body->valid & OBD_MD_FID)
+                id_assign_fid(&lli->lli_id, &body->id1);
+        if ((body->valid & OBD_MD_FLID) || (body->valid & OBD_MD_FLGENER))
+                id_assign_stc(&lli->lli_id, &body->id1);
         if (body->valid & OBD_MD_FLID)
-                lli->lli_st_ino = body->ino;
+                lli->lli_st_ino = id_ino(&body->id1);
+        if (body->valid & OBD_MD_FLGENER)
+                lli->lli_st_generation = id_gen(&body->id1);
+
         if (body->valid & OBD_MD_FLATIME)
                 LTIME_S(lli->lli_st_atime) = body->atime;
         if (body->valid & OBD_MD_FLMTIME)
@@ -116,22 +170,12 @@ void llu_update_inode(struct inode *inode, struct mds_body *body,
                 lli->lli_st_flags = body->flags;
         if (body->valid & OBD_MD_FLNLINK)
                 lli->lli_st_nlink = body->nlink;
-        if (body->valid & OBD_MD_FLGENER)
-                lli->lli_st_generation = body->generation;
         if (body->valid & OBD_MD_FLRDEV)
                 lli->lli_st_rdev = body->rdev;
         if (body->valid & OBD_MD_FLSIZE)
                 lli->lli_st_size = body->size;
         if (body->valid & OBD_MD_FLBLOCKS)
                 lli->lli_st_blocks = body->blocks;
-
-        /* fillin fid */
-        if (body->valid & OBD_MD_FLID)
-                lli->lli_fid.id = body->ino;
-        if (body->valid & OBD_MD_FLGENER)
-                lli->lli_fid.generation = body->generation;
-        if (body->valid & OBD_MD_FLTYPE)
-                lli->lli_fid.f_type = body->mode & S_IFMT;
 }
 
 void obdo_to_inode(struct inode *dst, struct obdo *src, obd_flag valid)
@@ -299,7 +343,7 @@ int llu_inode_getattr(struct inode *inode, struct lov_stripe_md *lsm)
 }
 
 static struct inode* llu_new_inode(struct filesys *fs,
-                                   struct ll_fid *fid)
+                                   struct lustre_id *id)
 {
        struct inode *inode;
         struct llu_inode_info *lli;
@@ -316,17 +360,17 @@ static struct inode* llu_new_inode(struct filesys *fs,
         lli->lli_maxbytes = (__u64)(~0UL);
         lli->lli_file_data = NULL;
 
-        lli->lli_sysio_fid.fid_data = &lli->lli_fid;
-        lli->lli_sysio_fid.fid_len = sizeof(lli->lli_fid);
+        lli->lli_sysio_fid.fid_data = &lli->lli_id;
+        lli->lli_sysio_fid.fid_len = sizeof(lli->lli_id);
 
-        memcpy(&lli->lli_fid, fid, sizeof(*fid));
+        memcpy(&lli->lli_id, id, sizeof(*id));
 
         /* file identifier is needed by functions like _sysio_i_find() */
        inode = _sysio_i_new(fs, &lli->lli_sysio_fid,
 #ifndef AUTOMOUNT_FILE_NAME
-                            fid->f_type & S_IFMT,
+                            id->li_stc.u.e3s.l3s_type & S_IFMT,
 #else
-                            fid->f_type, /* all of the bits! */
+                            id->li_stc.u.e3s.l3s_type, /* all of the bits! */
 #endif
                              0, 0,
                             &llu_inode_ops, lli);
@@ -337,34 +381,35 @@ static struct inode* llu_new_inode(struct filesys *fs,
         return inode;
 }
 
-static int llu_have_md_lock(struct inode *inode)
+static int llu_have_md_lock(struct inode *inode, __u64 lockpart)
 {
         struct llu_sb_info *sbi = llu_i2sbi(inode);
         struct llu_inode_info *lli = llu_i2info(inode);
         struct lustre_handle lockh;
         struct ldlm_res_id res_id = { .name = {0} };
         struct obd_device *obddev;
+        ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
         int flags;
         ENTRY;
 
         LASSERT(inode);
 
-        obddev = sbi->ll_mdc_exp->exp_obd;
-        res_id.name[0] = lli->lli_st_ino;
-        res_id.name[1] = lli->lli_st_generation;
+        obddev = sbi->ll_lmv_exp->exp_obd;
+        res_id.name[0] = id_fid(&lli->lli_id);
+        res_id.name[1] = id_group(&lli->lli_id);
 
         CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
 
         /* FIXME use LDLM_FL_TEST_LOCK instead */
         flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
-        if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
-                            NULL, LCK_PR, &lockh)) {
+        if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_IBITS,
+                            &policy, LCK_PR, &lockh)) {
                 ldlm_lock_decref(&lockh, LCK_PR);
                 RETURN(1);
         }
 
-        if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
-                            NULL, LCK_PW, &lockh)) {
+        if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_IBITS,
+                            &policy, LCK_PW, &lockh)) {
                 ldlm_lock_decref(&lockh, LCK_PW);
                 RETURN(1);
         }
@@ -382,27 +427,28 @@ static int llu_inode_revalidate(struct inode *inode)
                 RETURN(0);
         }
 
-        if (!llu_have_md_lock(inode)) {
+        if (!llu_have_md_lock(inode, MDS_INODELOCK_UPDATE)) {
                 struct lustre_md md;
                 struct ptlrpc_request *req = NULL;
                 struct llu_sb_info *sbi = llu_i2sbi(inode);
-                struct ll_fid fid;
+                struct lustre_id id;
                 unsigned long valid = 0;
                 int rc, ealen = 0;
 
-                /* Why don't we update all valid MDS fields here, if we're
-                 * doing an RPC anyways?  -phil */
+                /* Why don't we update all valid MDS fields here, if we're doing
+                 * an RPC anyways?  -phil */
                 if (S_ISREG(lli->lli_st_mode)) {
-                        ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
+                        ealen = obd_size_diskmd(sbi->ll_lov_exp, NULL);
                         valid |= OBD_MD_FLEASIZE;
                 }
-                ll_inode2fid(&fid, inode);
-                rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
+                ll_inode2id(&id, inode);
+                rc = mdc_getattr(sbi->ll_lmv_exp, &id, valid, ealen, &req);
                 if (rc) {
                         CERROR("failure %d inode %lu\n", rc, lli->lli_st_ino);
                         RETURN(-abs(rc));
                 }
-                rc = mdc_req2lustre_md(req, 0, sbi->ll_osc_exp, &md);
+                rc = mdc_req2lustre_md(sbi->ll_lmv_exp, req, 0, 
+                                       sbi->ll_lov_exp, &md);
 
                 /* XXX Too paranoid? */
                 if (((md.body->valid ^ valid) & OBD_MD_FLEASIZE) &&
@@ -421,7 +467,7 @@ static int llu_inode_revalidate(struct inode *inode)
 
                 llu_update_inode(inode, md.body, md.lsm);
                 if (md.lsm != NULL && llu_i2info(inode)->lli_smd != md.lsm)
-                        obd_free_memmd(sbi->ll_osc_exp, &md.lsm);
+                        obd_free_memmd(sbi->ll_lov_exp, &md.lsm);
 
                 if (md.body->valid & OBD_MD_FLSIZE)
                         set_bit(LLI_F_HAVE_MDS_SIZE_LOCK,
@@ -435,14 +481,7 @@ static int llu_inode_revalidate(struct inode *inode)
 
         /* ll_glimpse_size will prefer locally cached writes if they extend
          * the file */
-        {
-                struct ost_lvb lvb;
-                ldlm_error_t err;
-
-                err = llu_glimpse_size(inode, &lvb);
-                lli->lli_st_size = lvb.lvb_size;
-        }
-        RETURN(0);
+        RETURN(llu_glimpse_size(inode));
 }
 
 static void copy_stat_buf(struct inode *ino, struct intnl_stat *b)
@@ -471,6 +510,8 @@ static int llu_iop_getattr(struct pnode *pno,
         int rc;
         ENTRY;
 
+        liblustre_wait_event(0);
+
         if (!ino) {
                 LASSERT(pno);
                 LASSERT(pno->p_base->pb_ino);
@@ -493,18 +534,19 @@ static int llu_iop_getattr(struct pnode *pno,
 
 static int null_if_equal(struct ldlm_lock *lock, void *data)
 {
-        if (data == lock->l_ast_data)
+        if (data == lock->l_ast_data) {
                 lock->l_ast_data = NULL;
 
-        if (lock->l_req_mode != lock->l_granted_mode)
-                return LDLM_ITER_STOP;
+                if (lock->l_req_mode != lock->l_granted_mode)
+                        LDLM_ERROR(lock,"clearing inode with ungranted lock\n");
+        }
 
         return LDLM_ITER_CONTINUE;
 }
 
 void llu_clear_inode(struct inode *inode)
 {
-        struct ll_fid fid;
+        struct lustre_id id;
         struct llu_inode_info *lli = llu_i2info(inode);
         struct llu_sb_info *sbi = llu_i2sbi(inode);
         ENTRY;
@@ -512,16 +554,16 @@ void llu_clear_inode(struct inode *inode)
         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%lu(%p)\n", lli->lli_st_ino,
                lli->lli_st_generation, inode);
 
-        ll_inode2fid(&fid, inode);
+        ll_inode2id(&id, inode);
         clear_bit(LLI_F_HAVE_MDS_SIZE_LOCK, &(lli->lli_flags));
-        mdc_change_cbdata(sbi->ll_mdc_exp, &fid, null_if_equal, inode);
+        mdc_change_cbdata(sbi->ll_lmv_exp, &id, null_if_equal, inode);
 
         if (lli->lli_smd)
-                obd_change_cbdata(sbi->ll_osc_exp, lli->lli_smd,
+                obd_change_cbdata(sbi->ll_lov_exp, lli->lli_smd,
                                   null_if_equal, inode);
 
         if (lli->lli_smd) {
-                obd_free_memmd(sbi->ll_osc_exp, &lli->lli_smd);
+                obd_free_memmd(sbi->ll_lov_exp, &lli->lli_smd);
                 lli->lli_smd = NULL;
         }
 
@@ -539,6 +581,7 @@ void llu_iop_gone(struct inode *inode)
         struct llu_inode_info *lli = llu_i2info(inode);
         ENTRY;
 
+        liblustre_wait_event(0);
         llu_clear_inode(inode);
 
         OBD_FREE(lli, sizeof(*lli));
@@ -639,11 +682,11 @@ int llu_setattr_raw(struct inode *inode, struct iattr *attr)
          * inode ourselves so we can call obdo_from_inode() always. */
         if (ia_valid & (lsm ? ~(ATTR_SIZE | ATTR_FROM_OPEN | ATTR_RAW) : ~0)) {
                 struct lustre_md md;
-                llu_prepare_mdc_op_data(&op_data, inode, NULL, NULL, 0, 0);
-
-                rc = mdc_setattr(sbi->ll_mdc_exp, &op_data,
-                                  attr, NULL, 0, NULL, 0, &request);
+                llu_prepare_mdc_data(&op_data, inode, NULL, NULL, 0, 0);
 
+                rc = mdc_setattr(sbi->ll_lmv_exp, &op_data,
+                                 attr, NULL, 0, NULL, 0, &request);
+                
                 if (rc) {
                         ptlrpc_req_finished(request);
                         if (rc != -EPERM && rc != -EACCES)
@@ -651,11 +694,16 @@ int llu_setattr_raw(struct inode *inode, struct iattr *attr)
                         RETURN(rc);
                 }
 
-                rc = mdc_req2lustre_md(request, 0, sbi->ll_osc_exp, &md);
+                rc = mdc_req2lustre_md(sbi->ll_lmv_exp, request, 0, 
+                                       sbi->ll_lov_exp, &md);
                 if (rc) {
                         ptlrpc_req_finished(request);
                         RETURN(rc);
                 }
+
+                /* Won't invoke vmtruncate as we already cleared ATTR_SIZE,
+                 * but needed to set timestamps backwards on utime. */
+                inode_setattr(inode, attr);
                 llu_update_inode(inode, md.body, md.lsm);
                 ptlrpc_req_finished(request);
 
@@ -672,7 +720,7 @@ int llu_setattr_raw(struct inode *inode, struct iattr *attr)
                         /* from sys_utime() */
                         if (!(ia_valid & (ATTR_MTIME_SET | ATTR_ATIME_SET))) {
                                 if (current->fsuid != lli->lli_st_uid &&
-                                    (rc = ll_permission(inode, 0/*MAY_WRITE*/, NULL)) != 0)
+                                    (rc = ll_permission(inode, MAY_WRITE)) != 0)
                                         RETURN(rc);
                         } else {
                                /* from inode_change_ok() */
@@ -706,9 +754,6 @@ int llu_setattr_raw(struct inode *inode, struct iattr *attr)
                 }
 
                 rc = llu_vmtruncate(inode, attr->ia_size);
-                if (rc == 0)
-                        set_bit(LLI_F_HAVE_OST_SIZE_LOCK,
-                                &llu_i2info(inode)->lli_flags);
 
                 /* unlock now as we don't mind others file lockers racing with
                  * the mds updates below? */
@@ -727,7 +772,7 @@ int llu_setattr_raw(struct inode *inode, struct iattr *attr)
                 oa.o_valid = OBD_MD_FLID;
                 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
                                             OBD_MD_FLMTIME | OBD_MD_FLCTIME);
-                rc = obd_setattr(sbi->ll_osc_exp, &oa, lsm, NULL);
+                rc = obd_setattr(sbi->ll_lov_exp, &oa, lsm, NULL);
                 if (rc)
                         CERROR("obd_setattr fails: rc=%d\n", rc);
         }
@@ -745,6 +790,11 @@ static int llu_iop_setattr(struct pnode *pno,
         struct iattr iattr;
         ENTRY;
 
+        liblustre_wait_event(0);
+
+        LASSERT(!(mask & ~(SETATTR_MTIME | SETATTR_ATIME | 
+                           SETATTR_UID | SETATTR_GID |
+                           SETATTR_LEN | SETATTR_MODE)));
         memset(&iattr, 0, sizeof(iattr));
 
         if (mask & SETATTR_MODE) {
@@ -794,8 +844,8 @@ static int llu_iop_symlink_raw(struct pnode *pno, const char *tgt)
         if (llu_i2info(dir)->lli_st_nlink >= EXT2_LINK_MAX)
                 RETURN(err);
 
-        llu_prepare_mdc_op_data(&op_data, dir, NULL, name, len, 0);
-        err = mdc_create(sbi->ll_mdc_exp, &op_data,
+        llu_prepare_mdc_data(&op_data, dir, NULL, name, len, 0);
+        err = mdc_create(sbi->ll_lmv_exp, &op_data,
                          tgt, strlen(tgt) + 1, S_IFLNK | S_IRWXUGO,
                          current->fsuid, current->fsgid, 0, &request);
         ptlrpc_req_finished(request);
@@ -808,7 +858,7 @@ static int llu_readlink_internal(struct inode *inode,
 {
         struct llu_inode_info *lli = llu_i2info(inode);
         struct llu_sb_info *sbi = llu_i2sbi(inode);
-        struct ll_fid fid;
+        struct lustre_id id;
         struct mds_body *body;
         int rc, symlen = lli->lli_st_size + 1;
         ENTRY;
@@ -821,8 +871,8 @@ static int llu_readlink_internal(struct inode *inode,
                 RETURN(0);
         }
 
-        ll_inode2fid(&fid, inode);
-        rc = mdc_getattr(sbi->ll_mdc_exp, &fid,
+        ll_inode2id(&id, inode);
+        rc = mdc_getattr(sbi->ll_lmv_exp, &id,
                          OBD_MD_LINKNAME, symlen, request);
         if (rc) {
                 CERROR("inode %lu: rc = %d\n", lli->lli_st_ino, rc);
@@ -913,11 +963,11 @@ static int llu_iop_mknod_raw(struct pnode *pno,
         case S_IFBLK:
         case S_IFIFO:
         case S_IFSOCK:
-                llu_prepare_mdc_op_data(&op_data, dir, NULL,
-                                        pno->p_base->pb_name.name,
-                                        pno->p_base->pb_name.len,
-                                        0);
-                err = mdc_create(sbi->ll_mdc_exp, &op_data, NULL, 0, mode,
+                llu_prepare_mdc_data(&op_data, dir, NULL,
+                                     pno->p_base->pb_name.name,
+                                     pno->p_base->pb_name.len,
+                                     0);
+                err = mdc_create(sbi->ll_lmv_exp, &op_data, NULL, 0, mode,
                                  current->fsuid, current->fsgid, dev, &request);
                 ptlrpc_req_finished(request);
                 break;
@@ -944,13 +994,18 @@ static int llu_iop_link_raw(struct pnode *old, struct pnode *new)
         LASSERT(src);
         LASSERT(dir);
 
-        llu_prepare_mdc_op_data(&op_data, src, dir, name, namelen, 0);
-        rc = mdc_link(llu_i2sbi(src)->ll_mdc_exp, &op_data, &request);
+        liblustre_wait_event(0);
+        llu_prepare_mdc_data(&op_data, src, dir, name, namelen, 0);
+        rc = mdc_link(llu_i2sbi(src)->ll_lmv_exp, &op_data, &request);
         ptlrpc_req_finished(request);
+        liblustre_wait_event(0);
 
         RETURN(rc);
 }
 
+/*
+ * libsysio will clear the inode immediately after return
+ */
 static int llu_iop_unlink_raw(struct pnode *pno)
 {
         struct inode *dir = pno->p_base->pb_parent->pb_ino;
@@ -965,30 +1020,20 @@ static int llu_iop_unlink_raw(struct pnode *pno)
 
         LASSERT(target);
 
-        llu_prepare_mdc_op_data(&op_data, dir, NULL, name, len, 0);
-        rc = mdc_unlink(llu_i2sbi(dir)->ll_mdc_exp, &op_data, &request);
-        if (!rc) {
+        liblustre_wait_event(0);
+        llu_prepare_mdc_data(&op_data, dir, NULL, name, len, 0);
+        rc = mdc_unlink(llu_i2sbi(dir)->ll_lmv_exp, &op_data, &request);
+        if (!rc)
                 rc = llu_objects_destroy(request, dir);
-
-                llu_i2info(target)->lli_stale_flag = 1;
-                unhook_stale_inode(pno);
-        }
-
         ptlrpc_req_finished(request);
+        liblustre_wait_event(0);
         RETURN(rc);
 }
 
-/* FIXME
- * following cases need to be considered later:
- * - rename an opened file/dir
- * - an opened file be removed in rename
- * - rename to remove and hardlink (?opened)
- */
 static int llu_iop_rename_raw(struct pnode *old, struct pnode *new)
 {
         struct inode *src = old->p_parent->p_base->pb_ino;
         struct inode *tgt = new->p_parent->p_base->pb_ino;
-        struct inode *tgtinode = new->p_base->pb_ino;
         const char *oldname = old->p_base->pb_name.name;
         int oldnamelen = old->p_base->pb_name.len;
         const char *newname = new->p_base->pb_name.name;
@@ -1001,17 +1046,12 @@ static int llu_iop_rename_raw(struct pnode *old, struct pnode *new)
         LASSERT(src);
         LASSERT(tgt);
 
-        llu_prepare_mdc_op_data(&op_data, src, tgt, NULL, 0, 0);
-        rc = mdc_rename(llu_i2sbi(src)->ll_mdc_exp, &op_data,
+        llu_prepare_mdc_data(&op_data, src, tgt, NULL, 0, 0);
+        rc = mdc_rename(llu_i2sbi(src)->ll_lmv_exp, &op_data,
                         oldname, oldnamelen, newname, newnamelen,
                         &request);
         if (!rc) {
                 rc = llu_objects_destroy(request, src);
-
-                if (tgtinode) {
-                        llu_i2info(tgtinode)->lli_stale_flag = 1;
-                        unhook_stale_inode(new);
-                }
         }
 
         ptlrpc_req_finished(request);
@@ -1028,7 +1068,7 @@ static int llu_statfs_internal(struct llu_sb_info *sbi,
         int rc;
         ENTRY;
 
-        rc = obd_statfs(class_exp2obd(sbi->ll_mdc_exp), osfs, max_age);
+        rc = obd_statfs(class_exp2obd(sbi->ll_lmv_exp), osfs, max_age);
         if (rc) {
                 CERROR("mdc_statfs fails: rc = %d\n", rc);
                 RETURN(rc);
@@ -1037,7 +1077,7 @@ static int llu_statfs_internal(struct llu_sb_info *sbi,
         CDEBUG(D_SUPER, "MDC blocks "LPU64"/"LPU64" objects "LPU64"/"LPU64"\n",
                osfs->os_bavail, osfs->os_blocks, osfs->os_ffree,osfs->os_files);
 
-        rc = obd_statfs(class_exp2obd(sbi->ll_osc_exp), &obd_osfs, max_age);
+        rc = obd_statfs(class_exp2obd(sbi->ll_lov_exp), &obd_osfs, max_age);
         if (rc) {
                 CERROR("obd_statfs fails: rc = %d\n", rc);
                 RETURN(rc);
@@ -1105,6 +1145,8 @@ static int llu_iop_statvfs(struct pnode *pno,
         int rc;
         ENTRY;
 
+        liblustre_wait_event(0);
+
 #ifndef __CYGWIN__
         LASSERT(pno->p_base->pb_ino);
         rc = llu_statfs(llu_i2sbi(pno->p_base->pb_ino), &fs);
@@ -1120,7 +1162,7 @@ static int llu_iop_statvfs(struct pnode *pno,
         buf->f_files = fs.f_files;  /* Total number serial numbers */
         buf->f_ffree = fs.f_ffree;  /* Number free serial numbers */
         buf->f_favail = fs.f_ffree; /* Number free ser num for non-privileged*/
-        buf->f_fsid = fs.f_fsid.__val[1];
+        buf->f_fsid = fs.f_fstc.__val[1];
         buf->f_flag = 0;            /* No equiv in statfs; maybe use type? */
         buf->f_namemax = fs.f_namelen;
 #endif
@@ -1147,8 +1189,8 @@ static int llu_iop_mkdir_raw(struct pnode *pno, mode_t mode)
                 RETURN(err);
 
         mode = (mode & (S_IRWXUGO|S_ISVTX) & ~current->fs->umask) | S_IFDIR;
-        llu_prepare_mdc_op_data(&op_data, dir, NULL, name, len, 0);
-        err = mdc_create(llu_i2sbi(dir)->ll_mdc_exp, &op_data, NULL, 0, mode,
+        llu_prepare_mdc_data(&op_data, dir, NULL, name, len, 0);
+        err = mdc_create(llu_i2sbi(dir)->ll_lmv_exp, &op_data, NULL, 0, mode,
                          current->fsuid, current->fsgid, 0, &request);
         ptlrpc_req_finished(request);
         RETURN(err);
@@ -1168,23 +1210,46 @@ static int llu_iop_rmdir_raw(struct pnode *pno)
         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,dir=%lu/%lu(%p)\n",
                name, lli->lli_st_ino, lli->lli_st_generation, dir);
 
-        llu_prepare_mdc_op_data(&op_data, dir, NULL, name, len, S_IFDIR);
-        rc = mdc_unlink(llu_i2sbi(dir)->ll_mdc_exp, &op_data, &request);
+        llu_prepare_mdc_data(&op_data, dir, NULL, name, len, S_IFDIR);
+        rc = mdc_unlink(llu_i2sbi(dir)->ll_lmv_exp, &op_data, &request);
         ptlrpc_req_finished(request);
 
-        /* libsysio: remove the pnode right away */
-        if (!rc) {
-                llu_i2info(pno->p_base->pb_ino)->lli_stale_flag = 1;
-                unhook_stale_inode(pno);
-        }
-
         RETURN(rc);
 }
 
-static int llu_iop_fcntl(struct inode *ino, int cmd, va_list ap)
+#ifdef O_DIRECT
+#define FCNTL_FLMASK (O_APPEND|O_NONBLOCK|O_ASYNC|O_DIRECT)
+#else
+#define FCNTL_FLMASK (O_APPEND|O_NONBLOCK|O_ASYNC)
+#endif
+#define FCNTL_FLMASK_INVALID (O_NONBLOCK|O_ASYNC)
+
+static int llu_iop_fcntl(struct inode *ino, int cmd, va_list ap, int *rtn)
 {
-        CERROR("liblustre did not support fcntl\n");
-        return -ENOSYS;
+        struct llu_inode_info *lli = llu_i2info(ino);
+        long flags;
+
+        switch (cmd) {
+        case F_GETFL:
+                *rtn = lli->lli_open_flags;
+                return 0;
+        case F_SETFL:
+                flags = va_arg(ap, long);
+                flags &= FCNTL_FLMASK;
+                if (flags & FCNTL_FLMASK_INVALID) {
+                        CERROR("liblustre does not support O_NONBLOCK, O_ASYNC, "
+                               "and O_DIRECT on file descriptor\n");
+                        *rtn = -1;
+                        return EINVAL;
+                }
+                lli->lli_open_flags = (int) flags;
+                *rtn = 0;
+                return 0;
+        }
+
+        CERROR("unsupported fcntl cmd %x\n", cmd);
+        *rtn = -1;
+        return ENOSYS;
 }
 
 static int llu_get_grouplock(struct inode *inode, unsigned long arg)
@@ -1250,6 +1315,8 @@ static int llu_iop_ioctl(struct inode *ino, unsigned long int request,
 {
         unsigned long arg;
 
+        liblustre_wait_event(0);
+
         switch (request) {
         case LL_IOC_GROUP_LOCK:
                 arg = va_arg(ap, unsigned long);
@@ -1268,11 +1335,13 @@ static int llu_iop_ioctl(struct inode *ino, unsigned long int request,
  */
 static int llu_iop_sync(struct inode *inode)
 {
+        liblustre_wait_event(0);
         return 0;
 }
 
 static int llu_iop_datasync(struct inode *inode)
 {
+        liblustre_wait_event(0);
         return 0;
 }
 
@@ -1284,41 +1353,41 @@ struct filesys_ops llu_filesys_ops =
 struct inode *llu_iget(struct filesys *fs, struct lustre_md *md)
 {
         struct inode *inode;
-        struct ll_fid fid;
-        struct file_identifier fileid = {&fid, sizeof(fid)};
+        struct lustre_id id;
+        struct file_identifier fileid = {&id, sizeof(id)};
 
         if ((md->body->valid &
              (OBD_MD_FLGENER | OBD_MD_FLID | OBD_MD_FLTYPE)) !=
-            (OBD_MD_FLGENER | OBD_MD_FLID | OBD_MD_FLTYPE))
-                CERROR("invalide fields!\n");
+            (OBD_MD_FLGENER | OBD_MD_FLID | OBD_MD_FLTYPE)) {
+                CERROR("bad md body valid mask 0x%x\n", md->body->valid);
+                LBUG();
+                return ERR_PTR(-EPERM);
+        }
 
-        /* try to find existing inode */
-        fid.id = md->body->ino;
-        fid.generation = md->body->generation;
-        fid.f_type = md->body->mode & S_IFMT;
+        id = md->body->id1;
 
+        /* try to find existing inode */
         inode = _sysio_i_find(fs, &fileid);
         if (inode) {
                 struct llu_inode_info *lli = llu_i2info(inode);
 
-                if (lli->lli_stale_flag ||
-                    lli->lli_st_generation != md->body->generation)
+                if (inode->i_zombie ||
+                    lli->lli_st_generation != id_gen(&md->body->id1)) {
                         I_RELE(inode);
+                }
                 else {
                         llu_update_inode(inode, md->body, md->lsm);
                         return inode;
                 }
         }
 
-        inode = llu_new_inode(fs, &fid);
+        inode = llu_new_inode(fs, &id);
         if (inode)
                 llu_update_inode(inode, md->body, md->lsm);
         
         return inode;
 }
 
-extern struct list_head lustre_profile_list;
-
 static int
 llu_fsswop_mount(const char *source,
                  unsigned flags,
@@ -1330,18 +1399,19 @@ llu_fsswop_mount(const char *source,
         struct inode *root;
         struct pnode_base *rootpb;
         struct obd_device *obd;
-        struct ll_fid rootfid;
+        struct lustre_id rootid;
         struct llu_sb_info *sbi;
         struct obd_statfs osfs;
         static struct qstr noname = { NULL, 0, 0 };
         struct ptlrpc_request *request = NULL;
-        struct lustre_handle mdc_conn = {0, };
-        struct lustre_handle osc_conn = {0, };
+        struct lustre_handle lmv_conn = {0, };
+        struct lustre_handle lov_conn = {0, };
         struct lustre_md md;
         class_uuid_t uuid;
+        struct config_llog_instance cfg;
         struct lustre_profile *lprof;
-        char *osc = NULL, *mdc = NULL;
-        int err = -EINVAL;
+        char *lov = NULL, *lmv = NULL;
+        int async = 1, err = -EINVAL;
 
         ENTRY;
 
@@ -1354,67 +1424,44 @@ llu_fsswop_mount(const char *source,
         generate_random_uuid(uuid);
         class_uuid_unparse(uuid, &sbi->ll_sb_uuid);
 
-        /* zeroconf */
-        if (g_zconf) {
-                struct config_llog_instance cfg;
-                int len;
-
-                if (!g_zconf_mdsname) {
-                        CERROR("no mds name\n");
-                        GOTO(out_free, err = -EINVAL);
-                }
-
-                /* generate a string unique to this super, let's try
-                 the address of the super itself.*/
-                len = (sizeof(sbi) * 2) + 1; 
-                OBD_ALLOC(sbi->ll_instance, len);
-                if (sbi->ll_instance == NULL) 
-                        GOTO(out_free, err = -ENOMEM);
-                sprintf(sbi->ll_instance, "%p", sbi);
-
-                cfg.cfg_instance = sbi->ll_instance;
-                cfg.cfg_uuid = sbi->ll_sb_uuid;
-                err = liblustre_process_log(&cfg, 1);
-                if (err < 0) {
-                        CERROR("Unable to process log: %s\n", g_zconf_profile);
-
-                        GOTO(out_free, err);
-                }
-
-                lprof = class_get_profile(g_zconf_profile);
-                if (lprof == NULL) {
-                        CERROR("No profile found: %s\n", g_zconf_profile);
-                        GOTO(out_free, err = -EINVAL);
-                }
-                if (osc)
-                        OBD_FREE(osc, strlen(osc) + 1);
-                OBD_ALLOC(osc, strlen(lprof->lp_osc) + 
-                          strlen(sbi->ll_instance) + 2);
-                sprintf(osc, "%s-%s", lprof->lp_osc, sbi->ll_instance);
-
-                if (mdc)
-                        OBD_FREE(mdc, strlen(mdc) + 1);
-                OBD_ALLOC(mdc, strlen(lprof->lp_mdc) + 
-                          strlen(sbi->ll_instance) + 2);
-                sprintf(mdc, "%s-%s", lprof->lp_mdc, sbi->ll_instance);
-        } else {
-                /* setup from dump_file */
-                if (list_empty(&lustre_profile_list)) {
-                        CERROR("no profile\n");
-                        GOTO(out_free, err = -EINVAL);
-                }
+        /* generate a string unique to this super, let's try
+         the address of the super itself.*/
+        OBD_ALLOC(sbi->ll_instance, sizeof(sbi) * 2 + 1);
+        if (sbi->ll_instance == NULL) 
+                GOTO(out_free, err = -ENOMEM);
+        sprintf(sbi->ll_instance, "%p", sbi);
+
+        /* retrive & parse config log */
+        cfg.cfg_instance = sbi->ll_instance;
+        cfg.cfg_uuid = sbi->ll_sb_uuid;
+        err = liblustre_process_log(&cfg, 1);
+        if (err < 0) {
+                CERROR("Unable to process log: %s\n", g_zconf_profile);
+                GOTO(out_free, err);
+        }
 
-                lprof = list_entry(lustre_profile_list.next,
-                                   struct lustre_profile, lp_list);
-                osc = lprof->lp_osc;
-                mdc = lprof->lp_mdc;
+        lprof = class_get_profile(g_zconf_profile);
+        if (lprof == NULL) {
+                CERROR("No profile found: %s\n", g_zconf_profile);
+                GOTO(out_free, err = -EINVAL);
         }
+        if (lov)
+                OBD_FREE(lov, strlen(lov) + 1);
+        OBD_ALLOC(lov, strlen(lprof->lp_lov) + 
+                  strlen(sbi->ll_instance) + 2);
+        sprintf(lov, "%s-%s", lprof->lp_lov, sbi->ll_instance);
 
-        if (!osc) {
+        if (lmv)
+                OBD_FREE(lmv, strlen(lmv) + 1);
+        OBD_ALLOC(lmv, strlen(lprof->lp_lmv) + 
+                  strlen(sbi->ll_instance) + 2);
+        sprintf(lmv, "%s-%s", lprof->lp_lmv, sbi->ll_instance);
+
+        if (!lov) {
                 CERROR("no osc\n");
                 GOTO(out_free, err = -EINVAL);
         }
-        if (!mdc) {
+        if (!lmv) {
                 CERROR("no mdc\n");
                 GOTO(out_free, err = -EINVAL);
         }
@@ -1425,62 +1472,68 @@ llu_fsswop_mount(const char *source,
                 goto out_free;
         }
 
-        obd = class_name2obd(mdc);
+        obd = class_name2obd(lmv);
         if (!obd) {
-                CERROR("MDC %s: not setup or attached\n", mdc);
+                CERROR("MDC %s: not setup or attached\n", lmv);
                 GOTO(out_free, err = -EINVAL);
         }
-
-        if (mdc_init_ea_size(obd, osc))
+        obd_set_info(obd->obd_self_export, strlen("async"), "async",
+                     sizeof(async), &async);
+#warning "FIXME ASAP!"
+#if 0
+        if (mdc_init_ea_size(obd, lov))
                 GOTO(out_free, err = -EINVAL);
-
+#endif
         /* setup mdc */
-        err = obd_connect(&mdc_conn, obd, &sbi->ll_sb_uuid);
+        err = obd_connect(&lmv_conn, obd, &sbi->ll_sb_uuid, 0);
         if (err) {
-                CERROR("cannot connect to %s: rc = %d\n", mdc, err);
+                CERROR("cannot connect to %s: rc = %d\n", lmv, err);
                 GOTO(out_free, err);
         }
-        sbi->ll_mdc_exp = class_conn2export(&mdc_conn);
+        sbi->ll_lmv_exp = class_conn2export(&lmv_conn);
 
         err = obd_statfs(obd, &osfs, 100000000);
         if (err)
-                GOTO(out_mdc, err);
+                GOTO(out_lmv, err);
 
         /*
          * FIXME fill fs stat data into sbi here!!! FIXME
          */
 
-        /* setup osc */
-        obd = class_name2obd(osc);
+        /* setup lov */
+        obd = class_name2obd(lov);
         if (!obd) {
-                CERROR("OSC %s: not setup or attached\n", osc);
-                GOTO(out_mdc, err = -EINVAL);
+                CERROR("OSC %s: not setup or attached\n", lov);
+                GOTO(out_lmv, err = -EINVAL);
         }
+        obd_set_info(obd->obd_self_export, strlen("async"), "async",
+                     sizeof(async), &async);
 
-        err = obd_connect(&osc_conn, obd, &sbi->ll_sb_uuid);
+        err = obd_connect(&lov_conn, obd, &sbi->ll_sb_uuid, 0);
         if (err) {
-                CERROR("cannot connect to %s: rc = %d\n", osc, err);
-                GOTO(out_mdc, err);
+                CERROR("cannot connect to %s: rc = %d\n", lov, err);
+                GOTO(out_lmv, err);
         }
-        sbi->ll_osc_exp = class_conn2export(&osc_conn);
+        sbi->ll_lov_exp = class_conn2export(&lov_conn);
 
-        err = mdc_getstatus(sbi->ll_mdc_exp, &rootfid);
+        err = mdc_getstatus(sbi->ll_lmv_exp, &rootid);
         if (err) {
                 CERROR("cannot mds_connect: rc = %d\n", err);
-                GOTO(out_osc, err);
+                GOTO(out_lov, err);
         }
-        CDEBUG(D_SUPER, "rootfid "LPU64"\n", rootfid.id);
-        sbi->ll_rootino = rootfid.id;
+        CDEBUG(D_SUPER, "rootid "LPU64"\n", rootid.li_stc.u.e3s.l3s_ino);
+        sbi->ll_rootino = rootid.li_stc.u.e3s.l3s_ino;
 
         /* fetch attr of root inode */
-        err = mdc_getattr(sbi->ll_mdc_exp, &rootfid,
+        err = mdc_getattr(sbi->ll_lmv_exp, &rootid,
                           OBD_MD_FLNOTOBD|OBD_MD_FLBLOCKS, 0, &request);
         if (err) {
                 CERROR("mdc_getattr failed for root: rc = %d\n", err);
-                GOTO(out_osc, err);
+                GOTO(out_lov, err);
         }
 
-        err = mdc_req2lustre_md(request, 0, sbi->ll_osc_exp, &md);
+        err = mdc_req2lustre_md(sbi->ll_lmv_exp, request, 0, 
+                                sbi->ll_lov_exp, &md);
         if (err) {
                 CERROR("failed to understand root inode md: rc = %d\n",err);
                 GOTO(out_request, err);
@@ -1489,7 +1542,7 @@ llu_fsswop_mount(const char *source,
         LASSERT(sbi->ll_rootino != 0);
 
         root = llu_iget(fs, &md);
-        if (root == NULL) {
+        if (!root || IS_ERR(root)) {
                 CERROR("fail to generate root inode\n");
                 GOTO(out_request, err = -EBADF);
         }
@@ -1519,10 +1572,10 @@ out_inode:
         _sysio_i_gone(root);
 out_request:
         ptlrpc_req_finished(request);
-out_osc:
-        obd_disconnect(sbi->ll_osc_exp, 0);
-out_mdc:
-        obd_disconnect(sbi->ll_mdc_exp, 0);
+out_lov:
+        obd_disconnect(sbi->ll_lov_exp, 0);
+out_lmv:
+        obd_disconnect(sbi->ll_lmv_exp, 0);
 out_free:
         OBD_FREE(sbi, sizeof(*sbi));
         return err;
@@ -1532,32 +1585,30 @@ struct fssw_ops llu_fssw_ops = {
         llu_fsswop_mount
 };
 
-static struct inode_ops llu_inode_ops = {
-        inop_lookup:    llu_iop_lookup,
-        inop_getattr:   llu_iop_getattr,
-        inop_setattr:   llu_iop_setattr,
-        inop_getdirentries:     llu_iop_getdirentries,
-        inop_mkdir:     llu_iop_mkdir_raw,
-        inop_rmdir:     llu_iop_rmdir_raw,
-        inop_symlink:   llu_iop_symlink_raw,
-        inop_readlink:  llu_iop_readlink,
-        inop_open:      llu_iop_open,
-        inop_close:     llu_iop_close,
-        inop_link:      llu_iop_link_raw,
-        inop_unlink:    llu_iop_unlink_raw,
-        inop_rename:    llu_iop_rename_raw,
-        inop_ipreadv:   llu_iop_ipreadv,
-        inop_ipwritev:  llu_iop_ipwritev,
-        inop_iodone:    llu_iop_iodone,
-        inop_fcntl:     llu_iop_fcntl,
-        inop_sync:      llu_iop_sync,
-        inop_datasync:  llu_iop_datasync,
-        inop_ioctl:     llu_iop_ioctl,
-        inop_mknod:     llu_iop_mknod_raw,
+struct inode_ops llu_inode_ops = {
+        .inop_lookup         = llu_iop_lookup,
+        .inop_getattr        = llu_iop_getattr,
+        .inop_setattr        = llu_iop_setattr,
+        .inop_getdirentries  = llu_iop_getdirentries,
+        .inop_mkdir          = llu_iop_mkdir_raw,
+        .inop_rmdir          = llu_iop_rmdir_raw,
+        .inop_symlink        = llu_iop_symlink_raw,
+        .inop_readlink       = llu_iop_readlink,
+        .inop_open           = llu_iop_open,
+        .inop_close          = llu_iop_close,
+        .inop_link           = llu_iop_link_raw,
+        .inop_unlink         = llu_iop_unlink_raw,
+        .inop_rename         = llu_iop_rename_raw,
+        .inop_iodone         = llu_iop_iodone,
+        .inop_fcntl          = llu_iop_fcntl,
+        .inop_sync           = llu_iop_sync,
+        .inop_read           = llu_iop_read,
+        .inop_write          = llu_iop_write,
+        .inop_datasync       = llu_iop_datasync,
+        .inop_ioctl          = llu_iop_ioctl,
+        .inop_mknod          = llu_iop_mknod_raw,
 #ifdef _HAVE_STATVFS
-        inop_statvfs:   llu_iop_statvfs,
+        .inop_statvfs        = llu_iop_statvfs,
 #endif
-        inop_gone:      llu_iop_gone,
+        .inop_gone           = llu_iop_gone,
 };
-
-#warning "time_after() defined in liblustre.h need to be rewrite in userspace"