Whamcloud - gitweb
LU-3467 mdt: call MDT handlers via unified request handler
[fs/lustre-release.git] / lustre / mdt / mdt_reint.c
index 9576882..617ac43 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2012, Intel Corporation.
+ * Copyright (c) 2011, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -66,7 +66,7 @@ static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc,
         if (repbody->valid & OBD_MD_FLMDSCAPA)
                 RETURN(rc);
 
-       if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
+       if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa &&
            exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
                 struct lustre_capa *capa;
 
@@ -318,8 +318,7 @@ static int mdt_md_create(struct mdt_thread_info *info)
                                               " permitted for administrator or"
                                               " set mdt_enable_remote_dir_gid:"
                                               " rc = %d\n",
-                                               mdt2obd_dev(mdt)->obd_name,
-                                               -EPERM);
+                                               mdt_obd_name(mdt), -EPERM);
                                        GOTO(out_put_child, rc = -EPERM);
                                }
                        }
@@ -330,7 +329,7 @@ static int mdt_md_create(struct mdt_thread_info *info)
                                CERROR("%s: remote dir is only permitted on"
                                       " MDT0 or set_param"
                                       " mdt.*.enable_remote_dir=1\n",
-                                      mdt2obd_dev(mdt)->obd_name);
+                                      mdt_obd_name(mdt));
                                GOTO(out_put_child, rc = -EPERM);
                        }
                        if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
@@ -463,21 +462,28 @@ int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
        rc = mdt_attr_get_complex(info, mo, ma);
        if (rc) {
                CERROR("file attribute read error for "DFID": %d.\n",
-                       PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
+                       PFID(mdt_object_fid(mo)), rc);
                RETURN(rc);
        }
 
        /* If an up2date copy exists in the backend, add dirty flag */
        if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
            && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
+               struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 
                ma->ma_hsm.mh_flags |= HS_DIRTY;
+
+               mdt_lock_reg_init(lh, LCK_PW);
+               rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
+                                    MDT_LOCAL_LOCK);
+               if (rc != 0)
+                       RETURN(rc);
+
                rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
-               if (rc) {
+               if (rc)
                        CERROR("file attribute change error for "DFID": %d\n",
-                               PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
-                       RETURN(rc);
-               }
+                               PFID(mdt_object_fid(mo)), rc);
+               mdt_object_unlock(info, mo, lh, rc);
        }
 
        RETURN(rc);
@@ -502,10 +508,6 @@ static int mdt_reint_setattr(struct mdt_thread_info *info,
         if (info->mti_dlm_req)
                 ldlm_request_cancel(req, info->mti_dlm_req, 0);
 
-       if ((fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) &&
-           !OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING))
-               RETURN(-EPERM);
-
        repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
         if (IS_ERR(mo))
@@ -526,7 +528,7 @@ static int mdt_reint_setattr(struct mdt_thread_info *info,
                 if (rc)
                         GOTO(out_put, rc);
 
-                mfd = mdt_mfd_new();
+               mfd = mdt_mfd_new(med);
                 if (mfd == NULL) {
                         mdt_write_put(mo);
                         GOTO(out_put, rc = -ENOMEM);
@@ -554,7 +556,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info,
                 LASSERT(info->mti_ioepoch);
 
                spin_lock(&med->med_open_lock);
-               mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
+               mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
+                                    req_is_replay(req));
                if (mfd == NULL) {
                        spin_unlock(&med->med_open_lock);
                         CDEBUG(D_INODE, "no handle for file close: "
@@ -595,14 +598,14 @@ static int mdt_reint_setattr(struct mdt_thread_info *info,
         ma->ma_need = MA_INODE;
         ma->ma_valid = 0;
        rc = mdt_attr_get_complex(info, mo, ma);
-        if (rc != 0)
-                GOTO(out_put, rc);
+       if (rc != 0)
+               GOTO(out_put, rc);
 
-        mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
+       mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 
-       if (info->mti_mdt->mdt_opts.mo_oss_capa &&
+       if (info->mti_mdt->mdt_lut.lut_oss_capa &&
            exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
-           S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
+           S_ISREG(lu_object_attr(&mo->mot_obj)) &&
            (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
                 struct lustre_capa *capa;
 
@@ -658,7 +661,7 @@ static int mdt_reint_create(struct mdt_thread_info *info,
                break;
        default:
                CERROR("%s: Unsupported mode %o\n",
-                      mdt2obd_dev(info->mti_mdt)->obd_name,
+                      mdt_obd_name(info->mti_mdt),
                       info->mti_attr.ma_attr.la_mode);
                RETURN(err_serious(-EOPNOTSUPP));
        }
@@ -776,12 +779,12 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
 
                if (!fid_is_zero(rr->rr_fid2)) {
                        CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
-                              mdt2obd_dev(info->mti_mdt)->obd_name,
+                              mdt_obd_name(info->mti_mdt),
                               (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
                        GOTO(put_child, rc = -ENOENT);
                }
                CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
-                      mdt2obd_dev(info->mti_mdt)->obd_name,
+                      mdt_obd_name(info->mti_mdt),
                       (char *)rr->rr_name, PFID(mdt_object_fid(mc)));
 
                if (!mdt_is_dne_client(req->rq_export))
@@ -794,7 +797,7 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
                        if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
                                CERROR("%s: unlink remote entry is only "
                                       "permitted for administrator: rc = %d\n",
-                                       mdt2obd_dev(info->mti_mdt)->obd_name,
+                                       mdt_obd_name(info->mti_mdt),
                                        -EPERM);
                                GOTO(put_child, rc = -EPERM);
                        }
@@ -819,17 +822,20 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
                repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
                GOTO(unlock_child, rc = -EREMOTE);
        } else if (info->mti_spec.sp_rm_entry) {
-               CERROR("%s: lfs rmdir should not be used on local dir %s\n",
-                      mdt2obd_dev(info->mti_mdt)->obd_name,
-                      (char *)rr->rr_name);
-               GOTO(put_child, rc = -EPERM);
+               rc = -EPERM;
+               CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n",
+                      mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc);
+               GOTO(put_child, rc);
        }
 
-       rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
-                             MDT_CROSS_LOCK);
-       if (rc != 0) {
+       /* We used to acquire MDS_INODELOCK_FULL here but we can't do
+        * this now because a running HSM restore on the child (unlink
+        * victim) will hold the layout lock. See LU-4002. */
+       rc = mdt_object_lock(info, mc, child_lh,
+                            MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
+                            MDT_CROSS_LOCK);
+       if (rc != 0)
                GOTO(put_child, rc);
-       }
 
         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
@@ -941,13 +947,14 @@ static int mdt_reint_link(struct mdt_thread_info *info,
 
        if (mdt_object_remote(ms)) {
                mdt_object_put(info->mti_env, ms);
-               CERROR("Target directory "DFID" is on another MDT\n",
-                       PFID(rr->rr_fid1));
+               CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
+                      mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
+                      PFID(rr->rr_fid2));
                GOTO(out_unlock_parent, rc = -EXDEV);
        }
 
-        rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
-                            MDT_CROSS_LOCK);
+       rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
+                            MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
         if (rc != 0) {
                 mdt_object_put(info->mti_env, ms);
                 GOTO(out_unlock_parent, rc);
@@ -969,6 +976,11 @@ static int mdt_reint_link(struct mdt_thread_info *info,
                 GOTO(out_unlock_child, rc);
         /* save version of file name for replay, it must be ENOENT here */
         if (!req_is_replay(mdt_info_req(info))) {
+               if (rc != -ENOENT) {
+                       CDEBUG(D_INFO, "link target %.*s existed!\n",
+                              rr->rr_namelen, (char *)rr->rr_name);
+                       GOTO(out_unlock_child, rc = -EEXIST);
+               }
                 info->mti_ver[2] = ENOENT_VERSION;
                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
         }
@@ -991,31 +1003,31 @@ out_unlock_parent:
  * (lh->mlh_pdo_hash) in parallel directory lock.
  */
 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
-                       struct mdt_lock_handle *lh,
-                       struct mdt_object *obj, __u64 ibits)
+                             struct mdt_lock_handle *lh,
+                             struct mdt_object *obj, __u64 ibits)
 {
-        struct ldlm_res_id *res_id = &info->mti_res_id;
-        struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
-        ldlm_policy_data_t *policy = &info->mti_policy;
-        int rc;
+       struct ldlm_res_id *res = &info->mti_res_id;
+       struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
+       ldlm_policy_data_t *policy = &info->mti_policy;
+       int rc;
 
-        /*
-         * Finish res_id initializing by name hash marking part of
-         * directory which is taking modification.
-         */
-        LASSERT(lh->mlh_pdo_hash != 0);
-        fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
-        memset(policy, 0, sizeof(*policy));
-        policy->l_inodebits.bits = ibits;
-        /*
-         * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
-         * going to be sent to client. If it is - mdt_intent_policy() path will
-         * fix it up and turn FL_LOCAL flag off.
-         */
-        rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
-                          res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
-                          &info->mti_exp->exp_handle.h_cookie);
-        return rc;
+       /*
+        * Finish res_id initializing by name hash marking part of
+        * directory which is taking modification.
+        */
+       LASSERT(lh->mlh_pdo_hash != 0);
+       fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
+       memset(policy, 0, sizeof(*policy));
+       policy->l_inodebits.bits = ibits;
+       /*
+        * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
+        * going to be sent to client. If it is - mdt_intent_policy() path will
+        * fix it up and turn FL_LOCAL flag off.
+        */
+       rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
+                         res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
+                         &info->mti_exp->exp_handle.h_cookie);
+       return rc;
 }
 
 static int mdt_rename_lock(struct mdt_thread_info *info,
@@ -1032,7 +1044,7 @@ static int mdt_rename_lock(struct mdt_thread_info *info,
 
        memset(policy, 0, sizeof *policy);
        policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 4, 53, 0)
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 5, 53, 0)
        /* In phase I, we will not do cross-rename, so local BFL lock would
         * be enough
         */
@@ -1233,8 +1245,8 @@ static int mdt_reint_rename(struct mdt_thread_info *info,
 
         lh_oldp = &info->mti_lh[MDT_LH_OLD];
         mdt_lock_reg_init(lh_oldp, LCK_EX);
-        rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
-                             MDT_CROSS_LOCK);
+       rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
+                            MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
         if (rc != 0) {
                 mdt_object_put(info->mti_env, mold);
                 GOTO(out_unlock_target, rc);
@@ -1281,8 +1293,14 @@ static int mdt_reint_rename(struct mdt_thread_info *info,
                        GOTO(out_unlock_old, rc = -EXDEV);
                }
 
-                rc = mdt_object_lock(info, mnew, lh_newp,
-                                     MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
+               /* We used to acquire MDS_INODELOCK_FULL here but we
+                * can't do this now because a running HSM restore on
+                * the rename onto victim will hold the layout
+                * lock. See LU-4002. */
+               rc = mdt_object_lock(info, mnew, lh_newp,
+                                    MDS_INODELOCK_LOOKUP |
+                                    MDS_INODELOCK_UPDATE,
+                                    MDT_CROSS_LOCK);
                 if (rc != 0) {
                         mdt_object_put(info->mti_env, mnew);
                         GOTO(out_unlock_old, rc);
@@ -1293,7 +1311,14 @@ static int mdt_reint_rename(struct mdt_thread_info *info,
         } else if (rc != -EREMOTE && rc != -ENOENT) {
                 GOTO(out_unlock_old, rc);
         } else {
-                mdt_enoent_version_save(info, 3);
+               /* If mnew does not exist and mold are remote directory,
+                * it only allows rename if they are under same directory */
+               if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
+                       CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
+                              PFID(old_fid));
+                       GOTO(out_unlock_old, rc = -EXDEV);
+               }
+               mdt_enoent_version_save(info, 3);
         }
 
         /* step 5: rename it */