Whamcloud - gitweb
LU-14182 lov: cancel layout lock on replay deadlock
[fs/lustre-release.git] / lustre / llite / namei.c
index ed6eb70..6814416 100644 (file)
@@ -155,26 +155,44 @@ struct inode *ll_iget(struct super_block *sb, ino_t hash,
         RETURN(inode);
 }
 
-static void ll_invalidate_negative_children(struct inode *dir)
+/* mark negative sub file dentries invalid and prune unused dentries */
+static void ll_prune_negative_children(struct inode *dir)
 {
-       struct dentry *dentry, *tmp_subdir;
+       struct dentry *dentry;
+       struct dentry *child;
 
+       ENTRY;
+
+restart:
        spin_lock(&dir->i_lock);
        hlist_for_each_entry(dentry, &dir->i_dentry, d_alias) {
                spin_lock(&dentry->d_lock);
-               if (!list_empty(&dentry->d_subdirs)) {
-                       struct dentry *child;
-
-                       list_for_each_entry_safe(child, tmp_subdir,
-                                                &dentry->d_subdirs,
-                                                d_child) {
-                               if (child->d_inode == NULL)
-                                       d_lustre_invalidate(child, 1);
+               list_for_each_entry(child, &dentry->d_subdirs, d_child) {
+                       if (child->d_inode)
+                               continue;
+
+                       spin_lock_nested(&child->d_lock, DENTRY_D_LOCK_NESTED);
+                       __d_lustre_invalidate(child);
+                       if (!ll_d_count(child)) {
+                               dget_dlock(child);
+                               __d_drop(child);
+                               spin_unlock(&child->d_lock);
+                               spin_unlock(&dentry->d_lock);
+                               spin_unlock(&dir->i_lock);
+
+                               CDEBUG(D_DENTRY, "prune negative dentry %pd\n",
+                                      child);
+
+                               dput(child);
+                               goto restart;
                        }
+                       spin_unlock(&child->d_lock);
                }
                spin_unlock(&dentry->d_lock);
        }
        spin_unlock(&dir->i_lock);
+
+       EXIT;
 }
 
 int ll_test_inode_by_fid(struct inode *inode, void *opaque)
@@ -190,17 +208,12 @@ static int ll_dom_lock_cancel(struct inode *inode, struct ldlm_lock *lock)
        int rc;
        ENTRY;
 
-       if (!lli->lli_clob) {
-               /* due to DoM read on open, there may exist pages for Lustre
-                * regular file even though cl_object is not set up yet. */
-               truncate_inode_pages(inode->i_mapping, 0);
-               RETURN(0);
-       }
-
        env = cl_env_get(&refcheck);
        if (IS_ERR(env))
                RETURN(PTR_ERR(env));
 
+       OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_REPLAY_PAUSE, cfs_fail_val);
+
        /* reach MDC layer to flush data under  the DoM ldlm lock */
        rc = cl_object_flush(env, lli->lli_clob, lock);
        if (rc == -ENODATA) {
@@ -305,7 +318,7 @@ static void ll_lock_cancel_bits(struct ldlm_lock *lock, __u64 to_cancel)
        lli = ll_i2info(inode);
 
        if (bits & MDS_INODELOCK_UPDATE)
-               ll_file_set_flag(lli, LLIF_UPDATE_ATIME);
+               set_bit(LLIF_UPDATE_ATIME, &lli->lli_flags);
 
        if ((bits & MDS_INODELOCK_UPDATE) && S_ISDIR(inode->i_mode)) {
                CDEBUG(D_INODE, "invalidating inode "DFID" lli = %p, "
@@ -350,18 +363,17 @@ static void ll_lock_cancel_bits(struct ldlm_lock *lock, __u64 to_cancel)
                                                        ll_test_inode_by_fid,
                                                        (void *)&lli->lli_pfid);
                        if (master_inode) {
-                               ll_invalidate_negative_children(master_inode);
+                               ll_prune_negative_children(master_inode);
                                iput(master_inode);
                        }
                } else {
-                       ll_invalidate_negative_children(inode);
+                       ll_prune_negative_children(inode);
                }
        }
 
        if ((bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM)) &&
-           inode->i_sb->s_root != NULL &&
-           inode != inode->i_sb->s_root->d_inode)
-               ll_invalidate_aliases(inode);
+           !is_root_inode(inode))
+               ll_prune_aliases(inode);
 
        if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM))
                forget_all_cached_acls(inode);
@@ -603,6 +615,27 @@ struct dentry *ll_splice_alias(struct inode *inode, struct dentry *de)
        if (rc < 0)
                return ERR_PTR(rc);
        d_add(de, inode);
+
+       /* this needs only to be done for foreign symlink dirs as
+        * DCACHE_SYMLINK_TYPE is already set by d_flags_for_inode()
+        * kernel routine for files with symlink ops (ie, real symlink)
+        */
+       if (inode && S_ISDIR(inode->i_mode) &&
+           ll_sbi_has_foreign_symlink(ll_i2sbi(inode)) &&
+#ifdef HAVE_IOP_GET_LINK
+           inode->i_op->get_link) {
+#else
+           inode->i_op->follow_link) {
+#endif
+               CDEBUG(D_INFO, "%s: inode "DFID": faking foreign dir as a symlink\n",
+                      ll_i2sbi(inode)->ll_fsname, PFID(ll_inode2fid(inode)));
+               spin_lock(&de->d_lock);
+               /* like d_flags_for_inode() already does for files */
+               de->d_flags = (de->d_flags & ~DCACHE_ENTRY_TYPE) |
+                             DCACHE_SYMLINK_TYPE;
+               spin_unlock(&de->d_lock);
+       }
+
        CDEBUG(D_DENTRY, "Add dentry %p inode %p refc %d flags %#x\n",
               de, de->d_inode, ll_d_count(de), de->d_flags);
         return de;
@@ -634,10 +667,41 @@ static int ll_lookup_it_finish(struct ptlrpc_request *request,
                if (rc)
                        RETURN(rc);
 
-               if (it->it_op & IT_OPEN)
-                       ll_dom_finish_open(inode, request, it);
+               /* If encryption context was returned by MDT, put it in
+                * inode now to save an extra getxattr and avoid deadlock.
+                */
+               if (body->mbo_valid & OBD_MD_ENCCTX) {
+                       encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
+                       encctxlen = req_capsule_get_size(pill,
+                                                        &RMF_FILE_ENCCTX,
+                                                        RCL_SERVER);
+
+                       if (encctxlen) {
+                               CDEBUG(D_SEC,
+                                      "server returned encryption ctx for "DFID"\n",
+                                      PFID(ll_inode2fid(inode)));
+                               rc = ll_xattr_cache_insert(inode,
+                                              LL_XATTR_NAME_ENCRYPTION_CONTEXT,
+                                                          encctx, encctxlen);
+                               if (rc)
+                                       CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
+                                             ll_i2sbi(inode)->ll_fsname,
+                                             PFID(ll_inode2fid(inode)), rc);
+                               else if (encrypt) {
+                                       rc = llcrypt_get_encryption_info(inode);
+                                       if (rc)
+                                               CDEBUG(D_SEC,
+                                                "cannot get enc info for "DFID": rc = %d\n",
+                                                PFID(ll_inode2fid(inode)), rc);
+                               }
+                       }
+               }
 
                ll_set_lock_data(ll_i2sbi(parent)->ll_md_exp, inode, it, &bits);
+               /* OPEN can return data if lock has DoM+LAYOUT bits set */
+               if (it->it_op & IT_OPEN &&
+                   bits & MDS_INODELOCK_DOM && bits & MDS_INODELOCK_LAYOUT)
+                       ll_dom_finish_open(inode, request);
 
                /* We used to query real size from OSTs here, but actually
                 * this is not needed. For stat() calls size would be updated
@@ -678,29 +742,6 @@ static int ll_lookup_it_finish(struct ptlrpc_request *request,
                                      PFID(ll_inode2fid(inode)),
                                      rc);
                }
-
-               /* If encryption context was returned by MDT, put it in
-                * inode now to save an extra getxattr and avoid deadlock.
-                */
-               if (body->mbo_valid & OBD_MD_ENCCTX) {
-                       encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
-                       encctxlen = req_capsule_get_size(pill,
-                                                        &RMF_FILE_ENCCTX,
-                                                        RCL_SERVER);
-
-                       if (encctxlen) {
-                               CDEBUG(D_SEC,
-                                      "server returned encryption ctx for "DFID"\n",
-                                      PFID(ll_inode2fid(inode)));
-                               rc = ll_xattr_cache_insert(inode,
-                                              LL_XATTR_NAME_ENCRYPTION_CONTEXT,
-                                              encctx, encctxlen);
-                               if (rc)
-                                       CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
-                                             ll_i2sbi(inode)->ll_fsname,
-                                             PFID(ll_inode2fid(inode)), rc);
-                       }
-               }
        }
 
        /* Only hash *de if it is unhashed (new dentry).
@@ -1103,18 +1144,21 @@ static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
        it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
 
        if (ll_sbi_has_encrypt(ll_i2sbi(dir)) && IS_ENCRYPTED(dir)) {
-               /* we know that we are going to create a regular file because
+               /* in case of create, this is going to be a regular file because
                 * we set S_IFREG bit on it->it_create_mode above
                 */
                rc = llcrypt_get_encryption_info(dir);
                if (rc)
                        GOTO(out_release, rc);
-               if (!llcrypt_has_encryption_key(dir))
-                       GOTO(out_release, rc = -ENOKEY);
-               encrypt = true;
-               rc = 0;
+               if (open_flags & O_CREAT) {
+                       if (!llcrypt_has_encryption_key(dir))
+                               GOTO(out_release, rc = -ENOKEY);
+                       encrypt = true;
+               }
        }
 
+       OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE2, cfs_fail_val);
+
        /* Dentry added to dcache tree in ll_lookup_it */
        de = ll_lookup_it(dir, dentry, it, &secctx, &secctxlen, &pca, encrypt,
                          &encctx, &encctxlen);
@@ -1173,7 +1217,9 @@ static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
                        }
                }
 
-               if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN)) {
+               /* check also if a foreign file is openable */
+               if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN) &&
+                   ll_foreign_is_openable(dentry, open_flags)) {
                        /* Open dentry. */
                        if (S_ISFIFO(dentry->d_inode->i_mode)) {
                                /* We cannot call open here as it might
@@ -1659,6 +1705,10 @@ static int ll_rmdir(struct inode *dir, struct dentry *dchild)
        if (unlikely(d_mountpoint(dchild)))
                 RETURN(-EBUSY);
 
+       /* some foreign dir may not be allowed to be removed */
+       if (!ll_foreign_is_removable(dchild, false))
+               RETURN(-EPERM);
+
        op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len,
                                     S_IFDIR, LUSTRE_OPC_ANY, NULL);
        if (IS_ERR(op_data))
@@ -1670,13 +1720,25 @@ static int ll_rmdir(struct inode *dir, struct dentry *dchild)
        op_data->op_fid2 = op_data->op_fid3;
        rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
        ll_finish_md_op_data(op_data);
-       if (!rc)
-               ll_update_times(request, dir);
+       if (!rc) {
+               struct mdt_body *body;
 
-       ptlrpc_req_finished(request);
-       if (!rc)
+               ll_update_times(request, dir);
                ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR,
                                   ktime_us_delta(ktime_get(), kstart));
+
+               /*
+                * The server puts attributes in on the last unlink, use them
+                * to update the link count so the inode can be freed
+                * immediately.
+                */
+               body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
+               if (body->mbo_valid & OBD_MD_FLNLINK)
+                       set_nlink(dchild->d_inode, body->mbo_nlink);
+       }
+
+       ptlrpc_req_finished(request);
+
        RETURN(rc);
 }
 
@@ -1732,6 +1794,10 @@ static int ll_unlink(struct inode *dir, struct dentry *dchild)
        if (unlikely(d_mountpoint(dchild)))
                RETURN(-EBUSY);
 
+       /* some foreign file/dir may not be allowed to be unlinked */
+       if (!ll_foreign_is_removable(dchild, false))
+               RETURN(-EPERM);
+
        op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len, 0,
                                     LUSTRE_OPC_ANY, NULL);
        if (IS_ERR(op_data))
@@ -1780,6 +1846,7 @@ static int ll_rename(struct inode *src, struct dentry *src_dchild,
        struct ll_sb_info *sbi = ll_i2sbi(src);
        struct md_op_data *op_data;
        ktime_t kstart = ktime_get();
+       umode_t mode = 0;
        int err;
        ENTRY;
 
@@ -1804,15 +1871,21 @@ static int ll_rename(struct inode *src, struct dentry *src_dchild,
        if (err)
                RETURN(err);
 
-       op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, 0,
+       if (src_dchild->d_inode)
+               mode = src_dchild->d_inode->i_mode;
+
+       if (tgt_dchild->d_inode)
+               mode = tgt_dchild->d_inode->i_mode;
+
+       op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, mode,
                                     LUSTRE_OPC_ANY, NULL);
        if (IS_ERR(op_data))
                RETURN(PTR_ERR(op_data));
 
-       if (src_dchild->d_inode != NULL)
+       if (src_dchild->d_inode)
                op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
 
-       if (tgt_dchild->d_inode != NULL)
+       if (tgt_dchild->d_inode)
                op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
 
        err = md_rename(sbi->ll_md_exp, op_data,