WIP store parent dir for subdirs in inode

src/nfs_conn.cpp

@@ -389,7 +389,7 @@ static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
         auto direntry = json11::Json::parse(value, err);
         if (err != "")
         {
-            fprintf(stderr, "Invalid direntry %s: %s\n", kv_direntry_key(dir_ino, filename).c_str(), err.c_str());
+            fprintf(stderr, "Invalid JSON in direntry %s = %s: %s\n", kv_direntry_key(dir_ino, filename).c_str(), value.c_str(), err.c_str());
             *reply = (LOOKUP3res){ .status = NFS3ERR_IO };
             rpc_queue_reply(rop);
             return;
@@ -832,9 +832,13 @@ static void kv_do_create(kv_create_state *st)
             return;
         }
         st->new_id = new_id;
-        json11::Json direntry = json11::Json::object{ { "ino", st->new_id } };
+        auto direntry = json11::Json::object{ { "ino", st->new_id } };
+        if (st->attrobj["type"].uint64_value() == NF3DIR)
+        {
+            direntry["type"] = (uint64_t)NF3DIR;
+        }
         st->attrs = std::move(st->attrobj);
-        st->direntry_text = direntry.dump().c_str();
+        st->direntry_text = json11::Json(direntry).dump().c_str();
         // Set direntry
         st->self->parent->db->set(kv_direntry_key(st->dir_ino, st->filename), st->direntry_text, [st](int res)
         {
@@ -898,8 +902,8 @@ static void kv_do_create(kv_create_state *st)
                 auto direntry = json11::Json::parse(value, err);
                 if (err != "")
                 {
-                    fprintf(stderr, "direntry %s contains invalid JSON: %s, overwriting\n",
-                        kv_direntry_key(st->dir_ino, st->filename).c_str(), value.c_str());
+                    fprintf(stderr, "Invalid JSON in direntry %s = %s: %s, overwriting\n",
+                        kv_direntry_key(st->dir_ino, st->filename).c_str(), value.c_str(), err.c_str());
                     return true;
                 }
                 if (st->exclusive && direntry["verf"].uint64_value() == st->verf)
@@ -990,6 +994,7 @@ static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
     st->dir_ino = kv_fh_inode(args->where.dir);
     st->filename = args->where.name;
     st->attrobj["type"] = NF3DIR;
+    st->attrobj["parent_ino"] = st->dir_ino;
     kv_create_setattr(st->attrobj, args->attributes);
     st->cb = [st](int res) { kv_create_reply<MKDIR3RES>(st, res); };
     kv_do_create(st);
@@ -1091,8 +1096,8 @@ resume_1:
     st->direntry = json11::Json::parse(st->direntry_text, err);
     if (err != "")
     {
-        fprintf(stderr, "direntry %s contains invalid JSON: %s, deleting\n",
-            kv_direntry_key(st->dir_ino, st->filename).c_str(), st->direntry_text.c_str());
+        fprintf(stderr, "Invalid JSON in direntry %s = %s: %s, deleting\n",
+            kv_direntry_key(st->dir_ino, st->filename).c_str(), st->direntry_text.c_str(), err.c_str());
         // Just delete direntry and skip inode
     }
     else
@@ -1119,8 +1124,8 @@ resume_2:
         st->ientry = json11::Json::parse(st->ientry_text, err);
         if (err != "")
         {
-            fprintf(stderr, "inode %s contains invalid JSON: %s, treating as a regular file\n",
-                kv_inode_key(st->ino).c_str(), st->ientry_text.c_str());
+            fprintf(stderr, "Invalid JSON in inode %s = %s: %s, treating as a regular file\n",
+                kv_inode_key(st->ino).c_str(), st->ientry_text.c_str(), err.c_str());
         }
     // (1-2) Check type
     st->type = st->ientry["type"].uint64_value();
@@ -1346,6 +1351,8 @@ static void nfs_kv_continue_rename(nfs_kv_rename_state *st, int state)
     else if (state == 2) goto resume_2;
     else if (state == 3) goto resume_3;
     else if (state == 4) goto resume_4;
+    else if (state == 5) goto resume_5;
+    else if (state == 6) goto resume_6;
     else
     {
         fprintf("BUG: invalid state in nfs_kv_continue_delete()");
@@ -1380,19 +1387,71 @@ resume_2:
     if (st->res < 0)
     {
         auto cb = std::move(st->cb);
-        cb(res);
+        cb(st->res);
         return;
     }
+    {
+        std::string err;
+        st->direntry = json11::Json::parse(st->old_direntry_text, err);
+        if (err != "")
+        {
+            fprintf(stderr, "Invalid JSON in direntry %s = %s: %s\n",
+                kv_direntry_key(st->old_dir_ino, st->old_name).c_str(),
+                st->old_direntry_text.c_str(), err.c_str());
+        }
+    }
+    if (st->direntry["type"].uint64_value() == NF3DIR &&
+        st->direntry["ino"].uint64_value() != 0 &&
+        st->new_dir_ino != st->old_dir_ino)
+    {
+        // Read & check inode
+        kv_read_inode(st->self, st->direntry["ino"].uint64_value(), [st](int res, uint64_t ino, const std::string & value, json11::Json ientry)
+        {
+            st->res = res;
+            st->ientry_text = value;
+            st->ientry = ientry;
+            nfs_kv_continue_rename(3);
+        });
+        return;
+resume_3:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        // Change parent reference
+        {
+            auto ientry_new = st->ientry.object_items();
+            ientry_new["parent_ino"] = st->new_dir_ino;
+            st->self->parent->db->set(kv_inode_key(st->direntry["ino"].uint64_value()), json11::Json(ientry_new).dump(), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_rename(st, 4);
+            }, [st](int res, const std::string & old_value)
+            {
+                return old_value == st->ientry_text;
+            });
+        }
+        return;
+resume_4:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+    }
     st->self->parent->db->set(kv_direntry_key(st->new_dir_ino, st->new_name), [st](int res)
     {
         st->res = res;
-        nfs_kv_continue_rename(st, 3);
+        nfs_kv_continue_rename(st, 5);
     }, [st](int res, const std::string & old_value)
     {
         return res == -ENOENT;
     });
     return;
-resume_3:
+resume_5:
     if (st->res < 0)
     {
         if (st->res == -EAGAIN)
@@ -1401,13 +1460,13 @@ resume_3:
         st->self->parent->db->set(kv_direntry_key(st->old_dir_ino, st->old_name), st->old_direntry_text, [st](int res)
         {
             st->res = res;
-            nfs_kv_continue_rename(st, 4);
+            nfs_kv_continue_rename(st, 6);
         }, [st](int res, const std::string & old_value)
         {
             return res == -ENOENT;
         });
         return;
-resume_4:
+resume_6:
         if (st->res < 0)
         {
             if (st->res == -EAGAIN)
@@ -1427,6 +1486,7 @@ static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
 {
     auto st = new nfs_kv_rename_state;
     st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
     RENAME3args *args = (RENAME3args*)rop->request;
     st->old_dir_ino = kv_fh_inode(args->from.dir);
     st->new_dir_ino = kv_fh_inode(args->to.dir);
@@ -1462,6 +1522,11 @@ static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
 
 static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
 {
+    // 1) Find the source file
+    // 2) If it's a directory - fail with -EISDIR
+    // 3) Update the inode entry with refcount++
+    // 5) Create the new direntry with the same inode reference
+    //    Fail and rollback refcount if it already exists
     //nfs_client_t *self = (nfs_client_t*)opaque;
     //LINK3args *args = (LINK3args*)rop->request;
     LINK3res *reply = (LINK3res*)rop->reply;
@@ -1471,54 +1536,129 @@ static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static void fill_dir_entry(nfs_client_t *self, rpc_op_t *rop,
-    std::map<std::string, nfs_dir_t>::iterator dir_id_it, struct entryplus3 *entry, bool is_plus)
+struct nfs_kv_readdir_state
 {
-    if (dir_id_it == self->parent->dir_info.end())
-    {
-        return;
-    }
-    entry->fileid = dir_id_it->second.id;
-    if (is_plus)
-    {
-        entry->name_attributes = (post_op_attr){
-            .attributes_follow = 1,
-            .attributes = get_dir_attributes(self, dir_id_it->first),
-        };
-        entry->name_handle = (post_op_fh3){
-            .handle_follows = 1,
-            .handle = xdr_copy_string(rop->xdrs, "S"+base64_encode(sha256(dir_id_it->first))),
-        };
-    }
-}
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    bool is_plus = false;
+    READDIRPLUS3args args;
+    uint64_t dir_ino = 0;
+    std::string prefix;
+    void *list_handle;
+    bool eof = false;
+    int reply_size = 0;
+    std::vector<entryplus3> entries;
+};
 
 static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
 {
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    READDIRPLUS3args plus_args;
-    READDIRPLUS3args *args = NULL;
+    auto st = new nfs_kv_rename_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
     if (is_plus)
-        args = ((READDIRPLUS3args*)rop->request);
+        st->args = *((READDIRPLUS3args*)rop->request);
     else
     {
-        args = &plus_args;
         READDIR3args *in_args = ((READDIR3args*)rop->request);
-        args->dir = in_args->dir;
-        args->cookie = in_args->cookie;
-        *((uint64_t*)args->cookieverf) = *((uint64_t*)in_args->cookieverf);
-        args->dircount = 512;
-        args->maxcount = in_args->count;
+        st->args.dir = in_args->dir;
+        st->args.cookie = in_args->cookie;
+        *((uint64_t*)st->args.cookieverf) = *((uint64_t*)in_args->cookieverf);
+        st->args.dircount = 512;
+        st->args.maxcount = in_args->count;
     }
-    std::string dirhash = args->dir;
-    std::string dir;
-    if (dirhash != "roothandle")
+    st->dir_ino = kv_fh_inode(st->args.dir);
+    st->prefix = kv_direntry_key(dir_ino, "");
+    st->list_handle = self->parent->db->list_start(prefix);
+    st->eof = true;
+    // Limit results based on maximum reply size
+    // Sadly we have to calculate reply size by hand
+    // reply without entries is 4+4+(dir_attributes ? sizeof(fattr3) : 0)+8+4 bytes
+    st->reply_size = 20;
+    if (st->reply_size > st->args.maxcount)
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
-            dir = dir_it->second;
+        // Error, too small max reply size
+        if (is_plus)
+        {
+            READDIRPLUS3res *reply = (READDIRPLUS3res*)rop->reply;
+            *reply = (READDIRPLUS3res){ .status = NFS3ERR_TOOSMALL };
+            rpc_queue_reply(rop);
         }
-    std::string prefix = dir.size() ? dir+"/" : self->parent->name_prefix;
-    std::map<std::string, struct entryplus3> entries;
+        else
+        {
+            READDIR3res *reply = (READDIR3res*)rop->reply;
+            *reply = (READDIR3res){ .status = NFS3ERR_TOOSMALL };
+            rpc_queue_reply(rop);
+        }
+        return;
+    }
+    // Add . and ..
+    if (st->args.cookie == 0)
+    {
+        kv_read_inode(st->self, kv_fh_inode(st->args.dir), [st](int res, const std::string & value, json11::Json ientry)
+        {
+            
+        });
+        auto dir_id_it = self->parent->dir_info.find(dir);
+        fill_dir_entry(self, rop, dir_id_it, &entries["."], is_plus);
+        auto sl = dir.rfind("/");
+        if (sl != std::string::npos)
+        {
+            auto dir_id_it = self->parent->dir_info.find(dir.substr(0, sl));
+            fill_dir_entry(self, rop, dir_id_it, &entries[".."], is_plus);
+        }
+    }
+    st->self->parent->db->list_next(st->list_handle, [=](int res, const std::string & key, const std::string & value)
+    {
+        if (res == -ENOENT || key.size() > prefix.size() || key.substr(0, prefix.size()) != prefix)
+        {
+            self->parent->db->list_close(list_handle);
+            return;
+        }
+        auto direntry = json11::Json::parse(value, err);
+        if (err != "")
+        {
+            fprintf(stderr, "readdir: direntry %s contains invalid JSON: %s, skipping\n",
+                key.c_str(), value.c_str());
+            self->parent->db->list_next(list_handle);
+            return;
+        }
+        auto ino = direntry["ino"].uint64_value();
+        auto name = kv_direntry_filename(key);
+        auto fh = kv_fh(ino);
+        // 1 entry3 is (8+4+(filename_len+3)/4*4+8) bytes
+        // 1 entryplus3 is (8+4+(filename_len+3)/4*4+8
+        //   + 4+(name_attributes ? (sizeof(fattr3) = 84) : 0)
+        //   + 4+(name_handle ? 4+(handle_len+3)/4*4 : 0)) bytes
+        auto entry_size = 20 + len_pad4(name) + (is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+        if (st->reply_size + entry_size > st->args.maxcount)
+        {
+            st->eof = false;
+            self->parent->db->list_close(list_handle);
+            return;
+        }
+        auto idx = st->entries.size();
+        st->entries.push_back((entryplus3){});
+        auto entry = &st->entries[idx];
+        entry->name = xdr_copy_string(rop->xdrs, name);
+        entry->fileid = ino;
+        if (is_plus)
+        {
+            st->getattr_queue.push_back(idx);
+            if (st->getattr_running < self->parent->readdir_getattr_parallel)
+            {
+                kv_readdir_getattr_next(st);
+            }
+            entry->name_handle = (post_op_fh3){
+                .handle_follows = 1,
+                .handle = xdr_copy_string(fh),
+            };
+        }
+        self->parent->db->list_next(list_handle);
+    });
+    
+
+
+
     for (auto & ic: self->parent->cli->st_cli.inode_config)
     {
         auto & inode_cfg = ic.second;
@@ -1540,18 +1680,6 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
             // fileid will change when the user creates snapshots
             // however, we hope that clients tolerate it well
             // Linux does, even though it complains about "fileid changed" in dmesg
-            entries[subname].fileid = ic.first;
-            if (is_plus)
-            {
-                entries[subname].name_attributes = (post_op_attr){
-                    .attributes_follow = 1,
-                    .attributes = get_file_attributes(self, ic.first),
-                };
-                entries[subname].name_handle = (post_op_fh3){
-                    .handle_follows = 1,
-                    .handle = xdr_copy_string(rop->xdrs, "S"+base64_encode(sha256(inode_cfg.name))),
-                };
-            }
         }
         else
         {