Return block NFS implementation back as an option too

- Do not try to allocate more DB blocks in an inode block until it's "confirmed" and "locked" by the first write
- Do not recheck for new zero DB blocks on first write into an inode block - a CAS failure means someone else is already writing into it
- Throw new allocation blocks away regardless of whether the known_version is 0 on a CAS failure

src/CMakeLists.txt

@@ -181,10 +181,48 @@ target_link_libraries(vitastor-nbd
 	vitastor_client
 )
 
+# libvitastor_kv.so
+add_library(vitastor_kv SHARED
+	kv_db.cpp
+	kv_db.h
+)
+target_link_libraries(vitastor_kv
+	vitastor_client
+)
+set_target_properties(vitastor_kv PROPERTIES VERSION ${VERSION} SOVERSION 0)
+
+# vitastor-kv
+add_executable(vitastor-kv
+	kv_cli.cpp
+)
+target_link_libraries(vitastor-kv
+	vitastor_kv
+)
+
+add_executable(vitastor-kv-stress
+	kv_stress.cpp
+)
+target_link_libraries(vitastor-kv-stress
+	vitastor_kv
+)
+
 # vitastor-nfs
 add_executable(vitastor-nfs
 	nfs_proxy.cpp
-	nfs_conn.cpp
+	nfs_block.cpp
+	nfs_kv.cpp
+	nfs_kv_create.cpp
+	nfs_kv_getattr.cpp
+	nfs_kv_link.cpp
+	nfs_kv_lookup.cpp
+	nfs_kv_read.cpp
+	nfs_kv_readdir.cpp
+	nfs_kv_remove.cpp
+	nfs_kv_rename.cpp
+	nfs_kv_setattr.cpp
+	nfs_kv_write.cpp
+	nfs_fsstat.cpp
+	nfs_mount.cpp
 	nfs_portmap.cpp
 	sha256.c
 	nfs/xdr_impl.cpp
@@ -194,6 +232,7 @@ add_executable(vitastor-nfs
 )
 target_link_libraries(vitastor-nfs
 	vitastor_client
+	vitastor_kv
 )
 
 # vitastor-cli

src/cluster_client.cpp

@@ -6,7 +6,7 @@
 #include "cluster_client_impl.h"
 #include "http_client.h" // json_is_true
 
-cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config)
+cluster_client_t::cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json config)
 {
     wb = new writeback_cache_t();
 
@@ -534,7 +534,7 @@ void cluster_client_t::execute_internal(cluster_op_t *op)
         return;
     }
     if (op->opcode == OSD_OP_WRITE && enable_writeback && !(op->flags & OP_FLUSH_BUFFER) &&
-        !op->version /* FIXME no CAS writeback */)
+        !op->version /* no CAS writeback */)
     {
         if (wb->writebacks_active >= client_max_writeback_iodepth)
         {
@@ -555,7 +555,7 @@ void cluster_client_t::execute_internal(cluster_op_t *op)
     }
     if (op->opcode == OSD_OP_WRITE && !(op->flags & OP_IMMEDIATE_COMMIT))
     {
-        if (!(op->flags & OP_FLUSH_BUFFER))
+        if (!(op->flags & OP_FLUSH_BUFFER) && !op->version /* no CAS write-repeat */)
         {
             wb->copy_write(op, CACHE_WRITTEN);
         }
@@ -1161,7 +1161,7 @@ void cluster_client_t::handle_op_part(cluster_op_part_t *part)
                 );
             }
         }
-        else
+        else if (log_level > 0)
         {
             fprintf(
                 stderr, "%s operation failed on OSD %lu: retval=%ld (expected %d)\n",

src/cluster_client.h

@@ -121,7 +121,7 @@ public:
     json11::Json::object cli_config, file_config, etcd_global_config;
     json11::Json::object config;
 
-    cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config);
+    cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json config);
     ~cluster_client_t();
     void execute(cluster_op_t *op);
     void execute_raw(osd_num_t osd_num, osd_op_t *op);

src/kv_cli.cpp

@@ -0,0 +1,401 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Vitastor shared key/value database test CLI
+
+#define _XOPEN_SOURCE
+#include <limits.h>
+
+#include <netinet/tcp.h>
+#include <sys/epoll.h>
+#include <unistd.h>
+#include <fcntl.h>
+//#include <signal.h>
+
+#include "epoll_manager.h"
+#include "str_util.h"
+#include "kv_db.h"
+
+const char *exe_name = NULL;
+
+class kv_cli_t
+{
+public:
+    kv_dbw_t *db = NULL;
+    ring_loop_t *ringloop = NULL;
+    epoll_manager_t *epmgr = NULL;
+    cluster_client_t *cli = NULL;
+    bool interactive = false;
+    int in_progress = 0;
+    char *cur_cmd = NULL;
+    int cur_cmd_size = 0, cur_cmd_alloc = 0;
+    bool finished = false, eof = false;
+    json11::Json::object cfg;
+
+    ~kv_cli_t();
+
+    static json11::Json::object parse_args(int narg, const char *args[]);
+    void run(const json11::Json::object & cfg);
+    void read_cmd();
+    void next_cmd();
+    void handle_cmd(const std::string & cmd, std::function<void()> cb);
+};
+
+kv_cli_t::~kv_cli_t()
+{
+    if (cur_cmd)
+    {
+        free(cur_cmd);
+        cur_cmd = NULL;
+    }
+    cur_cmd_alloc = 0;
+    if (db)
+        delete db;
+    if (cli)
+    {
+        cli->flush();
+        delete cli;
+    }
+    if (epmgr)
+        delete epmgr;
+    if (ringloop)
+        delete ringloop;
+}
+
+json11::Json::object kv_cli_t::parse_args(int narg, const char *args[])
+{
+    json11::Json::object cfg;
+    for (int i = 1; i < narg; i++)
+    {
+        if (!strcmp(args[i], "-h") || !strcmp(args[i], "--help"))
+        {
+            printf(
+                "Vitastor Key/Value CLI\n"
+                "(c) Vitaliy Filippov, 2023+ (VNPL-1.1)\n"
+                "\n"
+                "USAGE: %s [--etcd_address ADDR] [OTHER OPTIONS]\n",
+                exe_name
+            );
+            exit(0);
+        }
+        else if (args[i][0] == '-' && args[i][1] == '-')
+        {
+            const char *opt = args[i]+2;
+            cfg[opt] = !strcmp(opt, "json") || i == narg-1 ? "1" : args[++i];
+        }
+    }
+    return cfg;
+}
+
+void kv_cli_t::run(const json11::Json::object & cfg)
+{
+    // Create client
+    ringloop = new ring_loop_t(512);
+    epmgr = new epoll_manager_t(ringloop);
+    cli = new cluster_client_t(ringloop, epmgr->tfd, cfg);
+    db = new kv_dbw_t(cli);
+    // Load image metadata
+    while (!cli->is_ready())
+    {
+        ringloop->loop();
+        if (cli->is_ready())
+            break;
+        ringloop->wait();
+    }
+    // Run
+    fcntl(0, F_SETFL, fcntl(0, F_GETFL, 0) | O_NONBLOCK);
+    try
+    {
+        epmgr->tfd->set_fd_handler(0, false, [this](int fd, int events)
+        {
+            if (events & EPOLLIN)
+            {
+                read_cmd();
+            }
+            if (events & EPOLLRDHUP)
+            {
+                epmgr->tfd->set_fd_handler(0, false, NULL);
+                finished = true;
+            }
+        });
+        interactive = true;
+        printf("> ");
+    }
+    catch (std::exception & e)
+    {
+        // Can't add to epoll, STDIN is probably a file
+        read_cmd();
+    }
+    while (!finished)
+    {
+        ringloop->loop();
+        if (!finished)
+            ringloop->wait();
+    }
+    // Destroy the client
+    delete db;
+    db = NULL;
+    cli->flush();
+    delete cli;
+    delete epmgr;
+    delete ringloop;
+    cli = NULL;
+    epmgr = NULL;
+    ringloop = NULL;
+}
+
+void kv_cli_t::read_cmd()
+{
+    if (!cur_cmd_alloc)
+    {
+        cur_cmd_alloc = 65536;
+        cur_cmd = (char*)malloc_or_die(cur_cmd_alloc);
+    }
+    while (cur_cmd_size < cur_cmd_alloc)
+    {
+        int r = read(0, cur_cmd+cur_cmd_size, cur_cmd_alloc-cur_cmd_size);
+        if (r < 0 && errno != EAGAIN)
+            fprintf(stderr, "Error reading from stdin: %s\n", strerror(errno));
+        if (r > 0)
+            cur_cmd_size += r;
+        if (r == 0)
+            eof = true;
+        if (r <= 0)
+            break;
+    }
+    next_cmd();
+}
+
+void kv_cli_t::next_cmd()
+{
+    if (in_progress > 0)
+    {
+        return;
+    }
+    int pos = 0;
+    for (; pos < cur_cmd_size; pos++)
+    {
+        if (cur_cmd[pos] == '\n' || cur_cmd[pos] == '\r')
+        {
+            auto cmd = trim(std::string(cur_cmd, pos));
+            pos++;
+            memmove(cur_cmd, cur_cmd+pos, cur_cmd_size-pos);
+            cur_cmd_size -= pos;
+            in_progress++;
+            handle_cmd(cmd, [this]()
+            {
+                in_progress--;
+                if (interactive)
+                    printf("> ");
+                next_cmd();
+                if (!in_progress)
+                    read_cmd();
+            });
+            break;
+        }
+    }
+    if (eof && !in_progress)
+    {
+        finished = true;
+    }
+}
+
+void kv_cli_t::handle_cmd(const std::string & cmd, std::function<void()> cb)
+{
+    if (cmd == "")
+    {
+        cb();
+        return;
+    }
+    auto pos = cmd.find_first_of(" \t");
+    if (pos != std::string::npos)
+    {
+        while (pos < cmd.size()-1 && (cmd[pos+1] == ' ' || cmd[pos+1] == '\t'))
+            pos++;
+    }
+    auto opname = strtolower(pos == std::string::npos ? cmd : cmd.substr(0, pos));
+    if (opname == "open")
+    {
+        uint64_t pool_id = 0;
+        inode_t inode_id = 0;
+        uint32_t kv_block_size = 0;
+        int scanned = sscanf(cmd.c_str() + pos+1, "%lu %lu %u", &pool_id, &inode_id, &kv_block_size);
+        if (scanned == 2)
+        {
+            kv_block_size = 4096;
+        }
+        if (scanned < 2 || !pool_id || !inode_id || !kv_block_size || (kv_block_size & (kv_block_size-1)) != 0)
+        {
+            fprintf(stderr, "Usage: open <pool_id> <inode_id> [block_size]. Block size must be a power of 2. Default is 4096.\n");
+            cb();
+            return;
+        }
+        cfg["kv_block_size"] = (uint64_t)kv_block_size;
+        db->open(INODE_WITH_POOL(pool_id, inode_id), cfg, [=](int res)
+        {
+            if (res < 0)
+                fprintf(stderr, "Error opening index: %s (code %d)\n", strerror(-res), res);
+            else
+                printf("Index opened. Current size: %lu bytes\n", db->get_size());
+            cb();
+        });
+    }
+    else if (opname == "config")
+    {
+        auto pos2 = cmd.find_first_of(" \t", pos+1);
+        if (pos2 == std::string::npos)
+        {
+            fprintf(stderr, "Usage: config <property> <value>\n");
+            cb();
+            return;
+        }
+        auto key = trim(cmd.substr(pos+1, pos2-pos-1));
+        auto value = parse_size(trim(cmd.substr(pos2+1)));
+        if (key != "kv_memory_limit" &&
+            key != "kv_allocate_blocks" &&
+            key != "kv_evict_max_misses" &&
+            key != "kv_evict_attempts_per_level" &&
+            key != "kv_evict_unused_age" &&
+            key != "kv_log_level")
+        {
+            fprintf(
+                stderr, "Allowed properties: kv_memory_limit, kv_allocate_blocks,"
+                " kv_evict_max_misses, kv_evict_attempts_per_level, kv_evict_unused_age, kv_log_level\n"
+            );
+        }
+        else
+        {
+            cfg[key] = value;
+            db->set_config(cfg);
+        }
+        cb();
+    }
+    else if (opname == "get" || opname == "set" || opname == "del")
+    {
+        if (opname == "get" || opname == "del")
+        {
+            if (pos == std::string::npos)
+            {
+                fprintf(stderr, "Usage: %s <key>\n", opname.c_str());
+                cb();
+                return;
+            }
+            auto key = trim(cmd.substr(pos+1));
+            if (opname == "get")
+            {
+                db->get(key, [this, cb](int res, const std::string & value)
+                {
+                    if (res < 0)
+                        fprintf(stderr, "Error: %s (code %d)\n", strerror(-res), res);
+                    else
+                    {
+                        write(1, value.c_str(), value.size());
+                        write(1, "\n", 1);
+                    }
+                    cb();
+                });
+            }
+            else
+            {
+                db->del(key, [this, cb](int res)
+                {
+                    if (res < 0)
+                        fprintf(stderr, "Error: %s (code %d)\n", strerror(-res), res);
+                    else
+                        printf("OK\n");
+                    cb();
+                });
+            }
+        }
+        else
+        {
+            auto pos2 = cmd.find_first_of(" \t", pos+1);
+            if (pos2 == std::string::npos)
+            {
+                fprintf(stderr, "Usage: set <key> <value>\n");
+                cb();
+                return;
+            }
+            auto key = trim(cmd.substr(pos+1, pos2-pos-1));
+            auto value = trim(cmd.substr(pos2+1));
+            db->set(key, value, [this, cb](int res)
+            {
+                if (res < 0)
+                    fprintf(stderr, "Error: %s (code %d)\n", strerror(-res), res);
+                else
+                    printf("OK\n");
+                cb();
+            });
+        }
+    }
+    else if (opname == "list")
+    {
+        std::string start, end;
+        if (pos != std::string::npos)
+        {
+            auto pos2 = cmd.find_first_of(" \t", pos+1);
+            if (pos2 != std::string::npos)
+            {
+                start = trim(cmd.substr(pos+1, pos2-pos-1));
+                end = trim(cmd.substr(pos2+1));
+            }
+            else
+            {
+                start = trim(cmd.substr(pos+1));
+            }
+        }
+        void *handle = db->list_start(start);
+        db->list_next(handle, [=](int res, const std::string & key, const std::string & value)
+        {
+            if (res < 0)
+            {
+                if (res != -ENOENT)
+                {
+                    fprintf(stderr, "Error: %s (code %d)\n", strerror(-res), res);
+                }
+                db->list_close(handle);
+                cb();
+            }
+            else
+            {
+                printf("%s = %s\n", key.c_str(), value.c_str());
+                db->list_next(handle, NULL);
+            }
+        });
+    }
+    else if (opname == "close")
+    {
+        db->close([=]()
+        {
+            printf("Index closed\n");
+            cb();
+        });
+    }
+    else if (opname == "quit" || opname == "q")
+    {
+        ::close(0);
+        finished = true;
+    }
+    else
+    {
+        fprintf(
+            stderr, "Unknown operation: %s. Supported operations:\n"
+            "open <pool_id> <inode_id> [block_size]\n"
+            "config <property> <value>\n"
+            "get <key>\nset <key> <value>\ndel <key>\nlist [<start> [end]]\n"
+            "close\nquit\n", opname.c_str()
+        );
+        cb();
+    }
+}
+
+int main(int narg, const char *args[])
+{
+    setvbuf(stdout, NULL, _IONBF, 0);
+    setvbuf(stderr, NULL, _IONBF, 0);
+    exe_name = args[0];
+    kv_cli_t *p = new kv_cli_t();
+    p->run(kv_cli_t::parse_args(narg, args));
+    delete p;
+    return 0;
+}

src/kv_db.h

@@ -0,0 +1,36 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Vitastor shared key/value database
+// Parallel optimistic B-Tree O:-)
+
+#pragma once
+
+#include "cluster_client.h"
+
+struct kv_db_t;
+
+struct kv_dbw_t
+{
+    kv_dbw_t(cluster_client_t *cli);
+    ~kv_dbw_t();
+
+    void open(inode_t inode_id, json11::Json cfg, std::function<void(int)> cb);
+    void set_config(json11::Json cfg);
+    void close(std::function<void()> cb);
+
+    uint64_t get_size();
+
+    void get(const std::string & key, std::function<void(int res, const std::string & value)> cb,
+        bool allow_old_cached = false);
+    void set(const std::string & key, const std::string & value, std::function<void(int res)> cb,
+        std::function<bool(int res, const std::string & value)> cas_compare = NULL);
+    void del(const std::string & key, std::function<void(int res)> cb,
+        std::function<bool(int res, const std::string & value)> cas_compare = NULL);
+
+    void* list_start(const std::string & start);
+    void list_next(void *handle, std::function<void(int res, const std::string & key, const std::string & value)> cb);
+    void list_close(void *handle);
+
+    kv_db_t *db;
+};

src/kv_stress.cpp

@@ -0,0 +1,697 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Vitastor shared key/value database stress tester / benchmark
+
+#define _XOPEN_SOURCE
+#include <limits.h>
+
+#include <netinet/tcp.h>
+#include <sys/epoll.h>
+#include <unistd.h>
+#include <fcntl.h>
+//#include <signal.h>
+
+#include "epoll_manager.h"
+#include "str_util.h"
+#include "kv_db.h"
+
+const char *exe_name = NULL;
+
+struct kv_test_listing_t
+{
+    uint64_t count = 0, done = 0;
+    void *handle = NULL;
+    std::string next_after;
+    std::set<std::string> inflights;
+    timespec tv_begin;
+    bool error = false;
+};
+
+struct kv_test_lat_t
+{
+    const char *name = NULL;
+    uint64_t usec = 0, count = 0;
+};
+
+struct kv_test_stat_t
+{
+    kv_test_lat_t get, add, update, del, list;
+    uint64_t list_keys = 0;
+};
+
+class kv_test_t
+{
+public:
+    // Config
+    json11::Json::object kv_cfg;
+    std::string key_prefix, key_suffix;
+    uint64_t inode_id = 0;
+    uint64_t op_count = 1000000;
+    uint64_t runtime_sec = 0;
+    uint64_t parallelism = 4;
+    uint64_t reopen_prob = 1;
+    uint64_t get_prob = 30000;
+    uint64_t add_prob = 20000;
+    uint64_t update_prob = 20000;
+    uint64_t del_prob = 5000;
+    uint64_t list_prob = 300;
+    uint64_t min_key_len = 10;
+    uint64_t max_key_len = 70;
+    uint64_t min_value_len = 50;
+    uint64_t max_value_len = 300;
+    uint64_t min_list_count = 10;
+    uint64_t max_list_count = 1000;
+    uint64_t print_stats_interval = 1;
+    bool json_output = false;
+    uint64_t log_level = 1;
+    bool trace = false;
+    bool stop_on_error = false;
+    // FIXME: Multiple clients
+    kv_test_stat_t stat, prev_stat;
+    timespec prev_stat_time, start_stat_time;
+
+    // State
+    kv_dbw_t *db = NULL;
+    ring_loop_t *ringloop = NULL;
+    epoll_manager_t *epmgr = NULL;
+    cluster_client_t *cli = NULL;
+    ring_consumer_t consumer;
+    bool finished = false;
+    uint64_t total_prob = 0;
+    uint64_t ops_sent = 0, ops_done = 0;
+    int stat_timer_id = -1;
+    int in_progress = 0;
+    bool reopening = false;
+    std::set<kv_test_listing_t*> listings;
+    std::set<std::string> changing_keys;
+    std::map<std::string, std::string> values;
+
+    ~kv_test_t();
+
+    static json11::Json::object parse_args(int narg, const char *args[]);
+    void parse_config(json11::Json cfg);
+    void run(json11::Json cfg);
+    void loop();
+    void print_stats(kv_test_stat_t & prev_stat, timespec & prev_stat_time);
+    void print_total_stats();
+    void start_change(const std::string & key);
+    void stop_change(const std::string & key);
+    void add_stat(kv_test_lat_t & stat, timespec tv_begin);
+};
+
+kv_test_t::~kv_test_t()
+{
+    if (db)
+        delete db;
+    if (cli)
+    {
+        cli->flush();
+        delete cli;
+    }
+    if (epmgr)
+        delete epmgr;
+    if (ringloop)
+        delete ringloop;
+}
+
+json11::Json::object kv_test_t::parse_args(int narg, const char *args[])
+{
+    json11::Json::object cfg;
+    for (int i = 1; i < narg; i++)
+    {
+        if (!strcmp(args[i], "-h") || !strcmp(args[i], "--help"))
+        {
+            printf(
+                "Vitastor Key/Value DB stress tester / benchmark\n"
+                "(c) Vitaliy Filippov, 2023+ (VNPL-1.1)\n"
+                "\n"
+                "USAGE: %s --pool_id POOL_ID --inode_id INODE_ID [OPTIONS]\n"
+                "  --op_count 1000000\n"
+                "    Total operations to run during test. 0 means unlimited\n"
+                "  --key_prefix \"\"\n"
+                "    Prefix for all keys read or written (to avoid collisions)\n"
+                "  --key_suffix \"\"\n"
+                "    Suffix for all keys read or written (to avoid collisions, but scan all DB)\n"
+                "  --runtime 0\n"
+                "    Run for this number of seconds. 0 means unlimited\n"
+                "  --parallelism 4\n"
+                "    Run this number of operations in parallel\n"
+                "  --get_prob 30000\n"
+                "    Fraction of key retrieve operations\n"
+                "  --add_prob 20000\n"
+                "    Fraction of key addition operations\n"
+                "  --update_prob 20000\n"
+                "    Fraction of key update operations\n"
+                "  --del_prob 30000\n"
+                "    Fraction of key delete operations\n"
+                "  --list_prob 300\n"
+                "    Fraction of listing operations\n"
+                "  --min_key_len 10\n"
+                "    Minimum key size in bytes\n"
+                "  --max_key_len 70\n"
+                "    Maximum key size in bytes\n"
+                "  --min_value_len 50\n"
+                "    Minimum value size in bytes\n"
+                "  --max_value_len 300\n"
+                "    Maximum value size in bytes\n"
+                "  --min_list_count 10\n"
+                "    Minimum number of keys read in listing (0 = all keys)\n"
+                "  --max_list_count 1000\n"
+                "    Maximum number of keys read in listing\n"
+                "  --print_stats 1\n"
+                "    Print operation statistics every this number of seconds\n"
+                "  --json\n"
+                "    JSON output\n"
+                "  --stop_on_error 0\n"
+                "    Stop on first execution error, mismatch, lost key or extra key during listing\n"
+                "  --kv_memory_limit 128M\n"
+                "    Maximum memory to use for vitastor-kv index cache\n"
+                "  --kv_allocate_blocks 4\n"
+                "    Number of PG blocks used for new tree block allocation in parallel\n"
+                "  --kv_evict_max_misses 10\n"
+                "    Eviction algorithm parameter: retry eviction from another random spot\n"
+                "    if this number of keys is used currently or was used recently\n"
+                "  --kv_evict_attempts_per_level 3\n"
+                "    Retry eviction at most this number of times per tree level, starting\n"
+                "    with bottom-most levels\n"
+                "  --kv_evict_unused_age 1000\n"
+                "    Evict only keys unused during this number of last operations\n"
+                "  --kv_log_level 1\n"
+                "    Log level. 0 = errors, 1 = warnings, 10 = trace operations\n",
+                exe_name
+            );
+            exit(0);
+        }
+        else if (args[i][0] == '-' && args[i][1] == '-')
+        {
+            const char *opt = args[i]+2;
+            cfg[opt] = !strcmp(opt, "json") || i == narg-1 ? "1" : args[++i];
+        }
+    }
+    return cfg;
+}
+
+void kv_test_t::parse_config(json11::Json cfg)
+{
+    inode_id = INODE_WITH_POOL(cfg["pool_id"].uint64_value(), cfg["inode_id"].uint64_value());
+    if (cfg["op_count"].uint64_value() > 0)
+        op_count = cfg["op_count"].uint64_value();
+    key_prefix = cfg["key_prefix"].string_value();
+    key_suffix = cfg["key_suffix"].string_value();
+    if (cfg["runtime"].uint64_value() > 0)
+        runtime_sec = cfg["runtime"].uint64_value();
+    if (cfg["parallelism"].uint64_value() > 0)
+        parallelism = cfg["parallelism"].uint64_value();
+    if (!cfg["reopen_prob"].is_null())
+        reopen_prob = cfg["reopen_prob"].uint64_value();
+    if (!cfg["get_prob"].is_null())
+        get_prob = cfg["get_prob"].uint64_value();
+    if (!cfg["add_prob"].is_null())
+        add_prob = cfg["add_prob"].uint64_value();
+    if (!cfg["update_prob"].is_null())
+        update_prob = cfg["update_prob"].uint64_value();
+    if (!cfg["del_prob"].is_null())
+        del_prob = cfg["del_prob"].uint64_value();
+    if (!cfg["list_prob"].is_null())
+        list_prob = cfg["list_prob"].uint64_value();
+    if (!cfg["min_key_len"].is_null())
+        min_key_len = cfg["min_key_len"].uint64_value();
+    if (cfg["max_key_len"].uint64_value() > 0)
+        max_key_len = cfg["max_key_len"].uint64_value();
+    if (!cfg["min_value_len"].is_null())
+        min_value_len = cfg["min_value_len"].uint64_value();
+    if (cfg["max_value_len"].uint64_value() > 0)
+        max_value_len = cfg["max_value_len"].uint64_value();
+    if (!cfg["min_list_count"].is_null())
+        min_list_count = cfg["min_list_count"].uint64_value();
+    if (!cfg["max_list_count"].is_null())
+        max_list_count = cfg["max_list_count"].uint64_value();
+    if (!cfg["print_stats"].is_null())
+        print_stats_interval = cfg["print_stats"].uint64_value();
+    if (!cfg["json"].is_null())
+        json_output = true;
+    if (!cfg["stop_on_error"].is_null())
+        stop_on_error = cfg["stop_on_error"].bool_value();
+    if (!cfg["kv_memory_limit"].is_null())
+        kv_cfg["kv_memory_limit"] = cfg["kv_memory_limit"];
+    if (!cfg["kv_allocate_blocks"].is_null())
+        kv_cfg["kv_allocate_blocks"] = cfg["kv_allocate_blocks"];
+    if (!cfg["kv_evict_max_misses"].is_null())
+        kv_cfg["kv_evict_max_misses"] = cfg["kv_evict_max_misses"];
+    if (!cfg["kv_evict_attempts_per_level"].is_null())
+        kv_cfg["kv_evict_attempts_per_level"] = cfg["kv_evict_attempts_per_level"];
+    if (!cfg["kv_evict_unused_age"].is_null())
+        kv_cfg["kv_evict_unused_age"] = cfg["kv_evict_unused_age"];
+    if (!cfg["kv_log_level"].is_null())
+    {
+        log_level = cfg["kv_log_level"].uint64_value();
+        trace = log_level >= 10;
+        kv_cfg["kv_log_level"] = cfg["kv_log_level"];
+    }
+    total_prob = reopen_prob+get_prob+add_prob+update_prob+del_prob+list_prob;
+    stat.get.name = "get";
+    stat.add.name = "add";
+    stat.update.name = "update";
+    stat.del.name = "del";
+    stat.list.name = "list";
+}
+
+void kv_test_t::run(json11::Json cfg)
+{
+    srand48(time(NULL));
+    parse_config(cfg);
+    // Create client
+    ringloop = new ring_loop_t(512);
+    epmgr = new epoll_manager_t(ringloop);
+    cli = new cluster_client_t(ringloop, epmgr->tfd, cfg);
+    db = new kv_dbw_t(cli);
+    // Load image metadata
+    while (!cli->is_ready())
+    {
+        ringloop->loop();
+        if (cli->is_ready())
+            break;
+        ringloop->wait();
+    }
+    // Run
+    reopening = true;
+    db->open(inode_id, kv_cfg, [this](int res)
+    {
+        reopening = false;
+        if (res < 0)
+        {
+            fprintf(stderr, "ERROR: Open index: %d (%s)\n", res, strerror(-res));
+            exit(1);
+        }
+        if (trace)
+            printf("Index opened\n");
+        ringloop->wakeup();
+    });
+    consumer.loop = [this]() { loop(); };
+    ringloop->register_consumer(&consumer);
+    if (print_stats_interval)
+        stat_timer_id = epmgr->tfd->set_timer(print_stats_interval*1000, true, [this](int) { print_stats(prev_stat, prev_stat_time); });
+    clock_gettime(CLOCK_REALTIME, &start_stat_time);
+    prev_stat_time = start_stat_time;
+    while (!finished)
+    {
+        ringloop->loop();
+        if (!finished)
+            ringloop->wait();
+    }
+    if (stat_timer_id >= 0)
+        epmgr->tfd->clear_timer(stat_timer_id);
+    ringloop->unregister_consumer(&consumer);
+    // Print total stats
+    print_total_stats();
+    // Destroy the client
+    delete db;
+    db = NULL;
+    cli->flush();
+    delete cli;
+    delete epmgr;
+    delete ringloop;
+    cli = NULL;
+    epmgr = NULL;
+    ringloop = NULL;
+}
+
+static const char *base64_chars = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789@+/";
+
+std::string random_str(int len)
+{
+    std::string str;
+    str.resize(len);
+    for (int i = 0; i < len; i++)
+    {
+        str[i] = base64_chars[lrand48() % 64];
+    }
+    return str;
+}
+
+void kv_test_t::loop()
+{
+    if (reopening)
+    {
+        return;
+    }
+    if (ops_done >= op_count)
+    {
+        finished = true;
+    }
+    while (!finished && ops_sent < op_count && in_progress < parallelism)
+    {
+        uint64_t dice = (lrand48() % total_prob);
+        if (dice < reopen_prob)
+        {
+            reopening = true;
+            db->close([this]()
+            {
+                if (trace)
+                    printf("Index closed\n");
+                db->open(inode_id, kv_cfg, [this](int res)
+                {
+                    reopening = false;
+                    if (res < 0)
+                    {
+                        fprintf(stderr, "ERROR: Reopen index: %d (%s)\n", res, strerror(-res));
+                        finished = true;
+                        return;
+                    }
+                    if (trace)
+                        printf("Index reopened\n");
+                    ringloop->wakeup();
+                });
+            });
+            return;
+        }
+        else if (dice < reopen_prob+get_prob)
+        {
+            // get existing
+            auto key = random_str(max_key_len);
+            auto k_it = values.lower_bound(key);
+            if (k_it == values.end())
+                continue;
+            key = k_it->first;
+            if (changing_keys.find(key) != changing_keys.end())
+                continue;
+            in_progress++;
+            ops_sent++;
+            if (trace)
+                printf("get %s\n", key.c_str());
+            timespec tv_begin;
+            clock_gettime(CLOCK_REALTIME, &tv_begin);
+            db->get(key, [this, key, tv_begin](int res, const std::string & value)
+            {
+                add_stat(stat.get, tv_begin);
+                ops_done++;
+                in_progress--;
+                auto it = values.find(key);
+                if (res != (it == values.end() ? -ENOENT : 0))
+                {
+                    fprintf(stderr, "ERROR: get %s: %d (%s)\n", key.c_str(), res, strerror(-res));
+                    if (stop_on_error)
+                        exit(1);
+                }
+                else if (it != values.end() && value != it->second)
+                {
+                    fprintf(stderr, "ERROR: get %s: mismatch: %s vs %s\n", key.c_str(), value.c_str(), it->second.c_str());
+                    if (stop_on_error)
+                        exit(1);
+                }
+                ringloop->wakeup();
+            });
+        }
+        else if (dice < reopen_prob+get_prob+add_prob+update_prob)
+        {
+            bool is_add = false;
+            std::string key;
+            if (dice < reopen_prob+get_prob+add_prob)
+            {
+                // add
+                is_add = true;
+                uint64_t key_len = min_key_len + (max_key_len > min_key_len ? lrand48() % (max_key_len-min_key_len) : 0);
+                key = key_prefix + random_str(key_len) + key_suffix;
+            }
+            else
+            {
+                // update
+                key = random_str(max_key_len);
+                auto k_it = values.lower_bound(key);
+                if (k_it == values.end())
+                    continue;
+                key = k_it->first;
+            }
+            if (changing_keys.find(key) != changing_keys.end())
+                continue;
+            uint64_t value_len = min_value_len + (max_value_len > min_value_len ? lrand48() % (max_value_len-min_value_len) : 0);
+            auto value = random_str(value_len);
+            start_change(key);
+            ops_sent++;
+            in_progress++;
+            if (trace)
+                printf("set %s = %s\n", key.c_str(), value.c_str());
+            timespec tv_begin;
+            clock_gettime(CLOCK_REALTIME, &tv_begin);
+            db->set(key, value, [this, key, value, tv_begin, is_add](int res)
+            {
+                add_stat(is_add ? stat.add : stat.update, tv_begin);
+                stop_change(key);
+                ops_done++;
+                in_progress--;
+                if (res != 0)
+                {
+                    fprintf(stderr, "ERROR: set %s = %s: %d (%s)\n", key.c_str(), value.c_str(), res, strerror(-res));
+                    if (stop_on_error)
+                        exit(1);
+                }
+                else
+                {
+                    values[key] = value;
+                }
+                ringloop->wakeup();
+            }, NULL);
+        }
+        else if (dice < reopen_prob+get_prob+add_prob+update_prob+del_prob)
+        {
+            // delete
+            auto key = random_str(max_key_len);
+            auto k_it = values.lower_bound(key);
+            if (k_it == values.end())
+                continue;
+            key = k_it->first;
+            if (changing_keys.find(key) != changing_keys.end())
+                continue;
+            start_change(key);
+            ops_sent++;
+            in_progress++;
+            if (trace)
+                printf("del %s\n", key.c_str());
+            timespec tv_begin;
+            clock_gettime(CLOCK_REALTIME, &tv_begin);
+            db->del(key, [this, key, tv_begin](int res)
+            {
+                add_stat(stat.del, tv_begin);
+                stop_change(key);
+                ops_done++;
+                in_progress--;
+                if (res != 0)
+                {
+                    fprintf(stderr, "ERROR: del %s: %d (%s)\n", key.c_str(), res, strerror(-res));
+                    if (stop_on_error)
+                        exit(1);
+                }
+                else
+                {
+                    values.erase(key);
+                }
+                ringloop->wakeup();
+            }, NULL);
+        }
+        else if (dice < reopen_prob+get_prob+add_prob+update_prob+del_prob+list_prob)
+        {
+            // list
+            ops_sent++;
+            in_progress++;
+            auto key = random_str(max_key_len);
+            auto lst = new kv_test_listing_t;
+            auto k_it = values.lower_bound(key);
+            lst->count = min_list_count + (max_list_count > min_list_count ? lrand48() % (max_list_count-min_list_count) : 0);
+            lst->handle = db->list_start(k_it == values.begin() ? key_prefix : key);
+            lst->next_after = k_it == values.begin() ? key_prefix : key;
+            lst->inflights = changing_keys;
+            listings.insert(lst);
+            if (trace)
+                printf("list from %s\n", key.c_str());
+            clock_gettime(CLOCK_REALTIME, &lst->tv_begin);
+            db->list_next(lst->handle, [this, lst](int res, const std::string & key, const std::string & value)
+            {
+                if (log_level >= 11)
+                    printf("list: %s = %s\n", key.c_str(), value.c_str());
+                if (res >= 0 && key_prefix.size() && (key.size() < key_prefix.size() ||
+                    key.substr(0, key_prefix.size()) != key_prefix))
+                {
+                    // stop at this key
+                    res = -ENOENT;
+                }
+                if (res < 0 || (lst->count > 0 && lst->done >= lst->count))
+                {
+                    add_stat(stat.list, lst->tv_begin);
+                    if (res == 0)
+                    {
+                        // ok (done >= count)
+                    }
+                    else if (res != -ENOENT)
+                    {
+                        fprintf(stderr, "ERROR: list: %d (%s)\n", res, strerror(-res));
+                        lst->error = true;
+                    }
+                    else
+                    {
+                        auto k_it = lst->next_after == "" ? values.begin() : values.upper_bound(lst->next_after);
+                        while (k_it != values.end())
+                        {
+                            while (k_it != values.end() && lst->inflights.find(k_it->first) != lst->inflights.end())
+                                k_it++;
+                            if (k_it != values.end())
+                            {
+                                fprintf(stderr, "ERROR: list: missing key %s\n", (k_it++)->first.c_str());
+                                lst->error = true;
+                            }
+                        }
+                    }
+                    if (lst->error && stop_on_error)
+                        exit(1);
+                    ops_done++;
+                    in_progress--;
+                    db->list_close(lst->handle);
+                    delete lst;
+                    listings.erase(lst);
+                    ringloop->wakeup();
+                }
+                else
+                {
+                    stat.list_keys++;
+                    // Do not check modified keys in listing
+                    // Listing may return their old or new state
+                    if ((!key_suffix.size() || key.size() >= key_suffix.size() &&
+                        key.substr(key.size()-key_suffix.size()) == key_suffix) &&
+                        lst->inflights.find(key) == lst->inflights.end())
+                    {
+                        lst->done++;
+                        auto k_it = lst->next_after == "" ? values.begin() : values.upper_bound(lst->next_after);
+                        while (true)
+                        {
+                            while (k_it != values.end() && lst->inflights.find(k_it->first) != lst->inflights.end())
+                            {
+                                k_it++;
+                            }
+                            if (k_it == values.end() || k_it->first > key)
+                            {
+                                fprintf(stderr, "ERROR: list: extra key %s\n", key.c_str());
+                                lst->error = true;
+                                break;
+                            }
+                            else if (k_it->first < key)
+                            {
+                                fprintf(stderr, "ERROR: list: missing key %s\n", k_it->first.c_str());
+                                lst->error = true;
+                                lst->next_after = k_it->first;
+                                k_it++;
+                            }
+                            else
+                            {
+                                if (k_it->second != value)
+                                {
+                                    fprintf(stderr, "ERROR: list: mismatch: %s = %s but should be %s\n",
+                                        key.c_str(), value.c_str(), k_it->second.c_str());
+                                    lst->error = true;
+                                }
+                                lst->next_after = k_it->first;
+                                break;
+                            }
+                        }
+                    }
+                    db->list_next(lst->handle, NULL);
+                }
+            });
+        }
+    }
+}
+
+void kv_test_t::add_stat(kv_test_lat_t & stat, timespec tv_begin)
+{
+    timespec tv_end;
+    clock_gettime(CLOCK_REALTIME, &tv_end);
+    int64_t usec = (tv_end.tv_sec - tv_begin.tv_sec)*1000000 +
+        (tv_end.tv_nsec - tv_begin.tv_nsec)/1000;
+    if (usec > 0)
+    {
+        stat.usec += usec;
+        stat.count++;
+    }
+}
+
+void kv_test_t::print_stats(kv_test_stat_t & prev_stat, timespec & prev_stat_time)
+{
+    timespec cur_stat_time;
+    clock_gettime(CLOCK_REALTIME, &cur_stat_time);
+    int64_t usec = (cur_stat_time.tv_sec - prev_stat_time.tv_sec)*1000000 +
+        (cur_stat_time.tv_nsec - prev_stat_time.tv_nsec)/1000;
+    if (usec > 0)
+    {
+        kv_test_lat_t *lats[] = { &stat.get, &stat.add, &stat.update, &stat.del, &stat.list };
+        kv_test_lat_t *prev[] = { &prev_stat.get, &prev_stat.add, &prev_stat.update, &prev_stat.del, &prev_stat.list };
+        if (!json_output)
+        {
+            char buf[128] = { 0 };
+            for (int i = 0; i < sizeof(lats)/sizeof(lats[0]); i++)
+            {
+                snprintf(buf, sizeof(buf)-1, "%.1f %s/s (%lu us)", (lats[i]->count-prev[i]->count)*1000000.0/usec,
+                    lats[i]->name, (lats[i]->usec-prev[i]->usec)/(lats[i]->count-prev[i]->count > 0 ? lats[i]->count-prev[i]->count : 1));
+                int k;
+                for (k = strlen(buf); k < strlen(lats[i]->name)+21; k++)
+                    buf[k] = ' ';
+                buf[k] = 0;
+                printf("%s", buf);
+            }
+            printf("\n");
+        }
+        else
+        {
+            int64_t runtime = (cur_stat_time.tv_sec - start_stat_time.tv_sec)*1000000 +
+                (cur_stat_time.tv_nsec - start_stat_time.tv_nsec)/1000;
+            printf("{\"runtime\":%.1f", (double)runtime/1000000.0);
+            for (int i = 0; i < sizeof(lats)/sizeof(lats[0]); i++)
+            {
+                if (lats[i]->count > prev[i]->count)
+                {
+                    printf(
+                        ",\"%s\":{\"avg\":{\"iops\":%.1f,\"usec\":%lu},\"total\":{\"count\":%lu,\"usec\":%lu}}",
+                        lats[i]->name, (lats[i]->count-prev[i]->count)*1000000.0/usec,
+                        (lats[i]->usec-prev[i]->usec)/(lats[i]->count-prev[i]->count),
+                        lats[i]->count, lats[i]->usec
+                    );
+                }
+            }
+            printf("}\n");
+        }
+    }
+    prev_stat = stat;
+    prev_stat_time = cur_stat_time;
+}
+
+void kv_test_t::print_total_stats()
+{
+    if (!json_output)
+        printf("Total:\n");
+    kv_test_stat_t start_stats;
+    timespec start_stat_time = this->start_stat_time;
+    print_stats(start_stats, start_stat_time);
+}
+
+void kv_test_t::start_change(const std::string & key)
+{
+    changing_keys.insert(key);
+    for (auto lst: listings)
+    {
+        lst->inflights.insert(key);
+    }
+}
+
+void kv_test_t::stop_change(const std::string & key)
+{
+    changing_keys.erase(key);
+}
+
+int main(int narg, const char *args[])
+{
+    setvbuf(stdout, NULL, _IONBF, 0);
+    setvbuf(stderr, NULL, _IONBF, 0);
+    exe_name = args[0];
+    kv_test_t *p = new kv_test_t();
+    p->run(kv_test_t::parse_args(narg, args));
+    delete p;
+    return 0;
+}

src/nfs_block.cpp

@@ -1,23 +1,18 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 //
-// NFS connection handler for NFS proxy
+// NFS proxy over Vitastor block images
 
 #include <sys/time.h>
 
 #include "str_util.h"
 
 #include "nfs_proxy.h"
-
+#include "nfs_common.h"
+#include "nfs_block.h"
 #include "nfs/nfs.h"
-
 #include "cli.h"
 
-#define TRUE 1
-#define FALSE 0
-
-#define MAX_REQUEST_SIZE 128*1024*1024
-
 static unsigned len_pad4(unsigned len)
 {
     return len + (len&3 ? 4-(len&3) : 0);
@@ -28,10 +23,10 @@ static std::string get_inode_name(nfs_client_t *self, diropargs3 & what)
     // Get name
     std::string dirhash = what.dir;
     std::string dir;
-    if (dirhash != "roothandle")
+    if (dirhash != NFS_ROOT_HANDLE)
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
             dir = dir_it->second;
         else
             return "";
@@ -42,24 +37,9 @@ static std::string get_inode_name(nfs_client_t *self, diropargs3 & what)
         : self->parent->name_prefix+name);
 }
 
-static nfsstat3 vitastor_nfs_map_err(int err)
-{
-    return (err == EINVAL ? NFS3ERR_INVAL
-        : (err == ENOENT ? NFS3ERR_NOENT
-        : (err == ENOSPC ? NFS3ERR_NOSPC
-        : (err == EEXIST ? NFS3ERR_EXIST
-        : (err == EIO ? NFS3ERR_IO : (err ? NFS3ERR_IO : NFS3_OK))))));
-}
-
-static int nfs3_null_proc(void *opaque, rpc_op_t *rop)
-{
-    rpc_queue_reply(rop);
-    return 0;
-}
-
 static fattr3 get_dir_attributes(nfs_client_t *self, std::string dir)
 {
-    auto & dinf = self->parent->dir_info.at(dir);
+    auto & dinf = self->parent->blockfs->dir_info.at(dir);
     return (fattr3){
         .type = NF3DIR,
         .mode = 0755,
@@ -108,7 +88,7 @@ static fattr3 get_file_attributes(nfs_client_t *self, inode_t inode_num)
     };
 }
 
-static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     GETATTR3args *args = (GETATTR3args*)rop->request;
@@ -116,12 +96,12 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     bool is_dir = false;
     std::string dirhash = args->object;
     std::string dir;
-    if (args->object == "roothandle")
+    if (args->object == NFS_ROOT_HANDLE)
         is_dir = true;
     else
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
         {
             is_dir = true;
             dir = dir_it->second;
@@ -140,8 +120,8 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     else
     {
         uint64_t inode_num = 0;
-        auto inode_num_it = self->parent->inode_by_hash.find(dirhash);
-        if (inode_num_it != self->parent->inode_by_hash.end())
+        auto inode_num_it = self->parent->blockfs->inode_by_hash.find(dirhash);
+        if (inode_num_it != self->parent->blockfs->inode_by_hash.end())
             inode_num = inode_num_it->second;
         auto inode_it = self->parent->cli->st_cli.inode_config.find(inode_num);
         if (inode_num && inode_it != self->parent->cli->st_cli.inode_config.end())
@@ -179,16 +159,16 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     SETATTR3args *args = (SETATTR3args*)rop->request;
     SETATTR3res *reply = (SETATTR3res*)rop->reply;
     std::string handle = args->object;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
-        if (handle == "roothandle" || self->parent->dir_by_hash.find(handle) != self->parent->dir_by_hash.end())
+        if (handle == NFS_ROOT_HANDLE || self->parent->blockfs->dir_by_hash.find(handle) != self->parent->blockfs->dir_by_hash.end())
         {
             if (args->new_attributes.size.set_it)
             {
@@ -228,7 +208,7 @@ static int nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     LOOKUP3args *args = (LOOKUP3args*)rop->request;
@@ -255,8 +235,8 @@ static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
                 return 0;
             }
         }
-        auto dir_it = self->parent->dir_info.find(full_name);
-        if (dir_it != self->parent->dir_info.end())
+        auto dir_it = self->parent->blockfs->dir_info.find(full_name);
+        if (dir_it != self->parent->blockfs->dir_info.end())
         {
             *reply = (LOOKUP3res){
                 .status = NFS3_OK,
@@ -277,7 +257,7 @@ static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_access_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_access_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     ACCESS3args *args = (ACCESS3args*)rop->request;
@@ -292,7 +272,7 @@ static int nfs3_access_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     //READLINK3args *args = (READLINK3args*)rop->request;
@@ -303,14 +283,14 @@ static int nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_read_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_read_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     READ3args *args = (READ3args*)rop->request;
     READ3res *reply = (READ3res*)rop->reply;
     std::string handle = args->file;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
         *reply = (READ3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -367,14 +347,14 @@ static int nfs3_read_proc(void *opaque, rpc_op_t *rop)
 
 static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t new_size, uint64_t offset, uint64_t count, void *buf);
 
-static int nfs3_write_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_write_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     WRITE3args *args = (WRITE3args*)rop->request;
     WRITE3res *reply = (WRITE3res*)rop->reply;
     std::string handle = args->file;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
         *reply = (WRITE3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -480,8 +460,8 @@ static void complete_extend_write(nfs_client_t *self, rpc_op_t *rop, inode_t ino
 
 static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size, int err)
 {
-    auto ext_it = self->extend_writes.lower_bound((extend_size_t){ .inode = inode, .new_size = 0 });
-    while (ext_it != self->extend_writes.end() &&
+    auto ext_it = self->parent->blockfs->extend_writes.lower_bound((extend_size_t){ .inode = inode, .new_size = 0 });
+    while (ext_it != self->parent->blockfs->extend_writes.end() &&
         ext_it->first.inode == inode &&
         ext_it->first.new_size <= new_size)
     {
@@ -490,7 +470,7 @@ static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t n
         {
             complete_extend_write(self, ext_it->second.rop, inode, ext_it->second.write_res < 0
                 ? ext_it->second.write_res : ext_it->second.resize_res);
-            self->extend_writes.erase(ext_it++);
+            self->parent->blockfs->extend_writes.erase(ext_it++);
         }
         else
             ext_it++;
@@ -500,7 +480,7 @@ static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t n
 static void extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size)
 {
     // Send an extend request
-    auto & ext = self->extends[inode];
+    auto & ext = self->parent->blockfs->extends[inode];
     ext.cur_extend = new_size;
     auto inode_it = self->parent->cli->st_cli.inode_config.find(inode);
     if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
@@ -514,7 +494,7 @@ static void extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size)
             { "force_size", true },
         }), [=](const cli_result_t & r)
         {
-            auto & ext = self->extends[inode];
+            auto & ext = self->parent->blockfs->extends[inode];
             if (r.err)
             {
                 fprintf(stderr, "Error extending inode %lu to %lu bytes: %s\n", inode, new_size, r.text.c_str());
@@ -548,7 +528,7 @@ static void nfs_do_write(nfs_client_t *self, std::multimap<extend_size_t, extend
     {
         auto inode = op->inode;
         int write_res = op->retval < 0 ? op->retval : (op->retval != op->len ? -ERANGE : 0);
-        if (ewr_it == self->extend_writes.end())
+        if (ewr_it == self->parent->blockfs->extend_writes.end())
         {
             complete_extend_write(self, rop, inode, write_res);
         }
@@ -558,7 +538,7 @@ static void nfs_do_write(nfs_client_t *self, std::multimap<extend_size_t, extend
             if (ewr_it->second.resize_res <= 0)
             {
                 complete_extend_write(self, rop, inode, write_res < 0 ? write_res : ewr_it->second.resize_res);
-                self->extend_writes.erase(ewr_it);
+                self->parent->blockfs->extend_writes.erase(ewr_it);
             }
         }
     };
@@ -572,7 +552,7 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
     if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
         inode_it->second.size < new_size)
     {
-        auto ewr_it = self->extend_writes.emplace((extend_size_t){
+        auto ewr_it = self->parent->blockfs->extend_writes.emplace((extend_size_t){
             .inode = inode,
             .new_size = new_size,
         }, (extend_write_t){
@@ -580,7 +560,7 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
             .resize_res = 1,
             .write_res = 1,
         });
-        auto & ext = self->extends[inode];
+        auto & ext = self->parent->blockfs->extends[inode];
         if (ext.cur_extend > 0)
         {
             // Already resizing, just wait
@@ -595,11 +575,11 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
     }
     else
     {
-        nfs_do_write(self, self->extend_writes.end(), rop, inode, offset, count, buf);
+        nfs_do_write(self, self->parent->blockfs->extend_writes.end(), rop, inode, offset, count, buf);
     }
 }
 
-static int nfs3_create_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_create_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     CREATE3args *args = (CREATE3args*)rop->request;
@@ -650,7 +630,7 @@ static int nfs3_create_proc(void *opaque, rpc_op_t *rop)
     return 1;
 }
 
-static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     MKDIR3args *args = (MKDIR3args*)rop->request;
@@ -669,19 +649,19 @@ static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    auto dir_id_it = self->parent->dir_info.find(full_name);
-    if (dir_id_it != self->parent->dir_info.end())
+    auto dir_id_it = self->parent->blockfs->dir_info.find(full_name);
+    if (dir_id_it != self->parent->blockfs->dir_info.end())
     {
         *reply = (MKDIR3res){ .status = NFS3ERR_EXIST };
         rpc_queue_reply(rop);
         return 0;
     }
     // FIXME: Persist empty directories in some etcd keys, like /vitastor/dir/...
-    self->parent->dir_info[full_name] = (nfs_dir_t){
-        .id = self->parent->next_dir_id++,
+    self->parent->blockfs->dir_info[full_name] = (nfs_dir_t){
+        .id = self->parent->blockfs->next_dir_id++,
         .mod_rev = 0,
     };
-    self->parent->dir_by_hash["S"+base64_encode(sha256(full_name))] = full_name;
+    self->parent->blockfs->dir_by_hash["S"+base64_encode(sha256(full_name))] = full_name;
     *reply = (MKDIR3res){
         .status = NFS3_OK,
         .resok = (MKDIR3resok){
@@ -700,7 +680,7 @@ static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
 {
 //    nfs_client_t *self = (nfs_client_t*)opaque;
 //    SYMLINK3args *args = (SYMLINK3args*)rop->request;
@@ -711,7 +691,7 @@ static int nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
 {
 //    nfs_client_t *self = (nfs_client_t*)opaque;
 //    MKNOD3args *args = (MKNOD3args*)rop->request;
@@ -722,7 +702,7 @@ static int nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_remove_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_remove_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     REMOVE3res *reply = (REMOVE3res*)rop->reply;
@@ -752,7 +732,7 @@ static int nfs3_remove_proc(void *opaque, rpc_op_t *rop)
     return 1;
 }
 
-static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     RMDIR3args *args = (RMDIR3args*)rop->request;
@@ -764,8 +744,8 @@ static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    auto dir_it = self->parent->dir_info.find(full_name);
-    if (dir_it == self->parent->dir_info.end())
+    auto dir_it = self->parent->blockfs->dir_info.find(full_name);
+    if (dir_it == self->parent->blockfs->dir_info.end())
     {
         *reply = (RMDIR3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -781,8 +761,8 @@ static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
             return 0;
         }
     }
-    self->parent->dir_by_hash.erase("S"+base64_encode(sha256(full_name)));
-    self->parent->dir_info.erase(dir_it);
+    self->parent->blockfs->dir_by_hash.erase("S"+base64_encode(sha256(full_name)));
+    self->parent->blockfs->dir_info.erase(dir_it);
     *reply = (RMDIR3res){ .status = NFS3_OK };
     rpc_queue_reply(rop);
     return 0;
@@ -811,12 +791,12 @@ static int continue_dir_rename(nfs_dir_rename_state *rename_st)
     if (!rename_st->items.size())
     {
         // old dir
-        auto old_info = self->parent->dir_info.at(rename_st->old_name);
-        self->parent->dir_info.erase(rename_st->old_name);
-        self->parent->dir_by_hash.erase("S"+base64_encode(sha256(rename_st->old_name)));
+        auto old_info = self->parent->blockfs->dir_info.at(rename_st->old_name);
+        self->parent->blockfs->dir_info.erase(rename_st->old_name);
+        self->parent->blockfs->dir_by_hash.erase("S"+base64_encode(sha256(rename_st->old_name)));
         // new dir
-        self->parent->dir_info[rename_st->new_name] = old_info;
-        self->parent->dir_by_hash["S"+base64_encode(sha256(rename_st->new_name))] = rename_st->new_name;
+        self->parent->blockfs->dir_info[rename_st->new_name] = old_info;
+        self->parent->blockfs->dir_by_hash["S"+base64_encode(sha256(rename_st->new_name))] = rename_st->new_name;
         RENAME3res *reply = (RENAME3res*)rename_st->rop->reply;
         *reply = (RENAME3res){
             .status = NFS3_OK,
@@ -853,7 +833,7 @@ static int continue_dir_rename(nfs_dir_rename_state *rename_st)
 
 static void nfs_do_rename(nfs_client_t *self, rpc_op_t *rop, std::string old_name, std::string new_name);
 
-static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_rename_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     RENAME3args *args = (RENAME3args*)rop->request;
@@ -866,8 +846,8 @@ static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    bool old_is_dir = self->parent->dir_info.find(old_name) != self->parent->dir_info.end();
-    bool new_is_dir = self->parent->dir_info.find(new_name) != self->parent->dir_info.end();
+    bool old_is_dir = self->parent->blockfs->dir_info.find(old_name) != self->parent->blockfs->dir_info.end();
+    bool new_is_dir = self->parent->blockfs->dir_info.find(new_name) != self->parent->blockfs->dir_info.end();
     bool old_is_file = false, new_is_file = false;
     for (auto & ic: self->parent->cli->st_cli.inode_config)
     {
@@ -948,7 +928,7 @@ static void nfs_do_rename(nfs_client_t *self, rpc_op_t *rop, std::string old_nam
     });
 }
 
-static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_link_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     //LINK3args *args = (LINK3args*)rop->request;
@@ -962,7 +942,7 @@ static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
 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)
 {
-    if (dir_id_it == self->parent->dir_info.end())
+    if (dir_id_it == self->parent->blockfs->dir_info.end())
     {
         return;
     }
@@ -980,7 +960,7 @@ static void fill_dir_entry(nfs_client_t *self, rpc_op_t *rop,
     }
 }
 
-static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
+static void block_nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     READDIRPLUS3args plus_args;
@@ -999,10 +979,10 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     }
     std::string dirhash = args->dir;
     std::string dir;
-    if (dirhash != "roothandle")
+    if (dirhash != NFS_ROOT_HANDLE)
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
             dir = dir_it->second;
     }
     std::string prefix = dir.size() ? dir+"/" : self->parent->name_prefix;
@@ -1043,12 +1023,12 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
         }
         else
         {
-            // skip directories, they will be added from dir_info
+            // skip directories, they will be added from blockfs->dir_info
         }
     }
-    // Add directories from dir_info
-    for (auto dir_id_it = self->parent->dir_info.lower_bound(prefix);
-        dir_id_it != self->parent->dir_info.end(); dir_id_it++)
+    // Add directories from blockfs->dir_info
+    for (auto dir_id_it = self->parent->blockfs->dir_info.lower_bound(prefix);
+        dir_id_it != self->parent->blockfs->dir_info.end(); dir_id_it++)
     {
         if (prefix != "" && dir_id_it->first.substr(0, prefix.size()) != prefix)
             break;
@@ -1061,12 +1041,12 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     }
     // Add . and ..
     {
-        auto dir_id_it = self->parent->dir_info.find(dir);
+        auto dir_id_it = self->parent->blockfs->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));
+            auto dir_id_it = self->parent->blockfs->dir_info.find(dir.substr(0, sl));
             fill_dir_entry(self, rop, dir_id_it, &entries[".."], is_plus);
         }
     }
@@ -1147,7 +1127,7 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     {
         READDIRPLUS3res *reply = (READDIRPLUS3res*)rop->reply;
         *reply = { .status = NFS3_OK };
-        *(uint64_t*)(reply->resok.cookieverf) = self->parent->dir_info.at(dir).mod_rev;
+        *(uint64_t*)(reply->resok.cookieverf) = self->parent->blockfs->dir_info.at(dir).mod_rev;
         reply->resok.reply.entries = entries.size() ? &entries.begin()->second : NULL;
         reply->resok.reply.eof = eof;
     }
@@ -1155,250 +1135,123 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     {
         READDIR3res *reply = (READDIR3res*)rop->reply;
         *reply = { .status = NFS3_OK };
-        *(uint64_t*)(reply->resok.cookieverf) = self->parent->dir_info.at(dir).mod_rev;
+        *(uint64_t*)(reply->resok.cookieverf) = self->parent->blockfs->dir_info.at(dir).mod_rev;
         reply->resok.reply.entries = entries.size() ? (entry3*)&entries.begin()->second : NULL;
         reply->resok.reply.eof = eof;
     }
     rpc_queue_reply(rop);
 }
 
-static int nfs3_readdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readdir_proc(void *opaque, rpc_op_t *rop)
 {
-    nfs3_readdir_common(opaque, rop, false);
+    block_nfs3_readdir_common(opaque, rop, false);
     return 0;
 }
 
-static int nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop)
 {
-    nfs3_readdir_common(opaque, rop, true);
+    block_nfs3_readdir_common(opaque, rop, true);
     return 0;
 }
 
-// Get file system statistics
-static int nfs3_fsstat_proc(void *opaque, rpc_op_t *rop)
+void block_fs_state_t::init(nfs_proxy_t *proxy)
 {
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    //FSSTAT3args *args = (FSSTAT3args*)rop->request;
-    FSSTAT3res *reply = (FSSTAT3res*)rop->reply;
-    uint64_t tbytes = 0, fbytes = 0;
-    auto pst_it = self->parent->pool_stats.find(self->parent->default_pool_id);
-    if (pst_it != self->parent->pool_stats.end())
-    {
-        auto ttb = pst_it->second["total_raw_tb"].number_value();
-        auto ftb = (pst_it->second["total_raw_tb"].number_value() - pst_it->second["used_raw_tb"].number_value());
-        tbytes = ttb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
-        fbytes = ftb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
-    }
-    *reply = (FSSTAT3res){
-        .status = NFS3_OK,
-        .resok = (FSSTAT3resok){
-            .obj_attributes = {
-                .attributes_follow = 1,
-                .attributes = get_dir_attributes(self, ""),
-            },
-            .tbytes = tbytes, // total bytes
-            .fbytes = fbytes, // free bytes
-            .abytes = fbytes, // available bytes
-            .tfiles = (size3)(1 << 31), // maximum total files
-            .ffiles = (size3)(1 << 31), // free files
-            .afiles = (size3)(1 << 31), // available files
-            .invarsec = 0,
-        },
+    // We need inode name hashes for NFS handles to remain stateless and <= 64 bytes long
+    dir_info[""] = (nfs_dir_t){
+        .id = 1,
+        .mod_rev = 0,
     };
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_fsinfo_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    FSINFO3args *args = (FSINFO3args*)rop->request;
-    FSINFO3res *reply = (FSINFO3res*)rop->reply;
-    if (args->fsroot != "roothandle")
+    clock_gettime(CLOCK_REALTIME, &dir_info[""].mtime);
+    assert(proxy->cli->st_cli.on_inode_change_hook == NULL);
+    proxy->cli->st_cli.on_inode_change_hook = [this, proxy](inode_t changed_inode, bool removed)
     {
-        // Example error
-        *reply = (FSINFO3res){ .status = NFS3ERR_INVAL };
-    }
-    else
-    {
-        // Fill info
-        *reply = (FSINFO3res){
-            .status = NFS3_OK,
-            .resok = (FSINFO3resok){
-                .obj_attributes = {
-                    .attributes_follow = 1,
-                    .attributes = get_dir_attributes(self, ""),
-                },
-                .rtmax = 128*1024*1024,
-                .rtpref = 128*1024*1024,
-                .rtmult = 4096,
-                .wtmax = 128*1024*1024,
-                .wtpref = 128*1024*1024,
-                .wtmult = 4096,
-                .dtpref = 128,
-                .maxfilesize = 0x7fffffffffffffff,
-                .time_delta = {
-                    .seconds = 1,
-                    .nseconds = 0,
-                },
-                .properties = FSF3_SYMLINK | FSF3_HOMOGENEOUS,
-            },
-        };
-    }
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_pathconf_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    PATHCONF3args *args = (PATHCONF3args*)rop->request;
-    PATHCONF3res *reply = (PATHCONF3res*)rop->reply;
-    if (args->object != "roothandle")
-    {
-        // Example error
-        *reply = (PATHCONF3res){ .status = NFS3ERR_INVAL };
-    }
-    else
-    {
-        // Fill info
-        bool_t x = FALSE;
-        *reply = (PATHCONF3res){
-            .status = NFS3_OK,
-            .resok = (PATHCONF3resok){
-                .obj_attributes = {
-                    // Without at least one reference to a non-constant value (local variable or something else),
-                    // with gcc 8 we get "internal compiler error: side-effects element in no-side-effects CONSTRUCTOR" here
-                    // FIXME: get rid of this after raising compiler requirement
-                    .attributes_follow = x,
-                },
-                .linkmax = 0,
-                .name_max = 255,
-                .no_trunc = TRUE,
-                .chown_restricted = FALSE,
-                .case_insensitive = FALSE,
-                .case_preserving = TRUE,
-            },
-        };
-    }
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_commit_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    //COMMIT3args *args = (COMMIT3args*)rop->request;
-    cluster_op_t *op = new cluster_op_t;
-    // fsync. we don't know how to fsync a single inode, so just fsync everything
-    op->opcode = OSD_OP_SYNC;
-    op->callback = [self, rop](cluster_op_t *op)
-    {
-        COMMIT3res *reply = (COMMIT3res*)rop->reply;
-        *reply = (COMMIT3res){ .status = vitastor_nfs_map_err(op->retval) };
-        *(uint64_t*)reply->resok.verf = self->parent->server_id;
-        rpc_queue_reply(rop);
+        auto inode_cfg_it = proxy->cli->st_cli.inode_config.find(changed_inode);
+        if (inode_cfg_it == proxy->cli->st_cli.inode_config.end())
+        {
+            return;
+        }
+        auto & inode_cfg = inode_cfg_it->second;
+        std::string full_name = inode_cfg.name;
+        if (proxy->name_prefix != "" && full_name.substr(0, proxy->name_prefix.size()) != proxy->name_prefix)
+        {
+            return;
+        }
+        // Calculate directory modification time and revision (used as "cookie verifier")
+        timespec now;
+        clock_gettime(CLOCK_REALTIME, &now);
+        dir_info[""].mod_rev = dir_info[""].mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dir_info[""].mod_rev;
+        dir_info[""].mtime = now;
+        int pos = full_name.find('/', proxy->name_prefix.size());
+        while (pos >= 0)
+        {
+            std::string dir = full_name.substr(0, pos);
+            auto & dinf = dir_info[dir];
+            if (!dinf.id)
+                dinf.id = next_dir_id++;
+            dinf.mod_rev = dinf.mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dinf.mod_rev;
+            dinf.mtime = now;
+            dir_by_hash["S"+base64_encode(sha256(dir))] = dir;
+            pos = full_name.find('/', pos+1);
+        }
+        // Alter inode_by_hash
+        if (removed)
+        {
+            auto ino_it = hash_by_inode.find(changed_inode);
+            if (ino_it != hash_by_inode.end())
+            {
+                inode_by_hash.erase(ino_it->second);
+                hash_by_inode.erase(ino_it);
+            }
+        }
+        else
+        {
+            std::string hash = "S"+base64_encode(sha256(full_name));
+            auto hbi_it = hash_by_inode.find(changed_inode);
+            if (hbi_it != hash_by_inode.end() && hbi_it->second != hash)
+            {
+                // inode had a different name, remove old hash=>inode pointer
+                inode_by_hash.erase(hbi_it->second);
+            }
+            inode_by_hash[hash] = changed_inode;
+            hash_by_inode[changed_inode] = hash;
+        }
     };
-    self->parent->cli->execute(op);
-    return 1;
 }
 
-static int mount3_mnt_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    //nfs_dirpath *args = (nfs_dirpath*)rop->request;
-    nfs_mountres3 *reply = (nfs_mountres3*)rop->reply;
-    u_int flavor = RPC_AUTH_NONE;
-    reply->fhs_status = MNT3_OK;
-    reply->mountinfo.fhandle = xdr_copy_string(rop->xdrs, "roothandle");
-    reply->mountinfo.auth_flavors.auth_flavors_len = 1;
-    reply->mountinfo.auth_flavors.auth_flavors_val = (u_int*)xdr_copy_string(rop->xdrs, (char*)&flavor, sizeof(u_int)).data;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_dump_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    nfs_mountlist *reply = (nfs_mountlist*)rop->reply;
-    *reply = (struct nfs_mountbody*)malloc_or_die(sizeof(struct nfs_mountbody));
-    xdr_add_malloc(rop->xdrs, *reply);
-    (*reply)->ml_hostname = xdr_copy_string(rop->xdrs, "127.0.0.1");
-    (*reply)->ml_directory = xdr_copy_string(rop->xdrs, self->parent->export_root);
-    (*reply)->ml_next = NULL;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_umnt_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    //nfs_dirpath *arg = (nfs_dirpath*)rop->request;
-    // do nothing
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_umntall_proc(void *opaque, rpc_op_t *rop)
-{
-    // do nothing
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_export_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    nfs_exports *reply = (nfs_exports*)rop->reply;
-    *reply = (struct nfs_exportnode*)calloc_or_die(1, sizeof(struct nfs_exportnode) + sizeof(struct nfs_groupnode));
-    xdr_add_malloc(rop->xdrs, *reply);
-    (*reply)->ex_dir = xdr_copy_string(rop->xdrs, self->parent->export_root);
-    (*reply)->ex_groups = (struct nfs_groupnode*)(reply+1);
-    (*reply)->ex_groups->gr_name = xdr_copy_string(rop->xdrs, "127.0.0.1");
-    (*reply)->ex_groups->gr_next = NULL;
-    (*reply)->ex_next = NULL;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-nfs_client_t::nfs_client_t()
+void nfs_block_procs(nfs_client_t *self)
 {
     struct rpc_service_proc_t pt[] = {
-        {NFS_PROGRAM, NFS_V3, NFS3_NULL,        nfs3_null_proc,        NULL,                            0,                        NULL,                           0,                       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,     nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,     nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,      nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,      nfs3_access_proc,      (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,    nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READ,        nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        this},
-        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,       nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,      nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,       nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,     nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,       nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,      nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,       nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,      nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_LINK,        nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,     nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS, nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), this},
-        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,      nfs3_fsstat_proc,      (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,      nfs3_fsinfo_proc,      (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,    nfs3_pathconf_proc,    (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    this},
-        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,      nfs3_commit_proc,      (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,      NULL,                            0,                        NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,     (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,    NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,    (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc, NULL,                            0,                        NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,  NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       this},
+        {NFS_PROGRAM, NFS_V3, NFS3_NULL,          nfs3_null_proc,              NULL,                            0,                        NULL,                           0,                       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,       block_nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,       block_nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,        block_nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,        block_nfs3_access_proc,      (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,      block_nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READ,          block_nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,         block_nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,        block_nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,         block_nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,       block_nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,         block_nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,        block_nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,         block_nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,        block_nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LINK,          block_nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,       block_nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS,   block_nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,        nfs3_fsstat_proc,            (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,        nfs3_fsinfo_proc,            (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,      nfs3_pathconf_proc,          (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,        nfs3_commit_proc,            (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,              NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,             (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,            NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,            (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc,         NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,          NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       self},
     };
     for (int i = 0; i < sizeof(pt)/sizeof(pt[0]); i++)
     {
-        proc_table.insert(pt[i]);
+        self->proc_table.insert(pt[i]);
     }
 }
-
-nfs_client_t::~nfs_client_t()
-{
-}

src/nfs_block.h

@@ -0,0 +1,57 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over Vitastor block images - header
+
+#pragma once
+
+struct nfs_dir_t
+{
+    uint64_t id;
+    uint64_t mod_rev;
+    timespec mtime;
+};
+
+struct extend_size_t
+{
+    inode_t inode;
+    uint64_t new_size;
+};
+
+inline bool operator < (const extend_size_t &a, const extend_size_t &b)
+{
+    return a.inode < b.inode || a.inode == b.inode && a.new_size < b.new_size;
+}
+
+struct extend_write_t
+{
+    rpc_op_t *rop;
+    int resize_res, write_res; // 1 = started, 0 = completed OK, -errno = completed with error
+};
+
+struct extend_inode_t
+{
+    uint64_t cur_extend = 0, next_extend = 0;
+};
+
+struct block_fs_state_t
+{
+    // filehandle = "S"+base64(sha256(full name with prefix)) or "roothandle" for mount root)
+    uint64_t next_dir_id = 2;
+    // filehandle => dir with name_prefix
+    std::map<std::string, std::string> dir_by_hash;
+    // dir with name_prefix => dir info
+    std::map<std::string, nfs_dir_t> dir_info;
+    // filehandle => inode ID
+    std::map<std::string, inode_t> inode_by_hash;
+    // inode ID => filehandle
+    std::map<inode_t, std::string> hash_by_inode;
+
+    // inode extend requests in progress
+    std::map<inode_t, extend_inode_t> extends;
+    std::multimap<extend_size_t, extend_write_t> extend_writes;
+
+    void init(nfs_proxy_t *proxy);
+};
+
+nfsstat3 vitastor_nfs_map_err(int err);

src/nfs_common.h

@@ -0,0 +1,22 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy - common functions
+
+#pragma once
+
+#include "nfs/nfs.h"
+
+void nfs_block_procs(nfs_client_t *self);
+void nfs_kv_procs(nfs_client_t *self);
+int nfs3_fsstat_proc(void *opaque, rpc_op_t *rop);
+int nfs3_fsinfo_proc(void *opaque, rpc_op_t *rop);
+int nfs3_pathconf_proc(void *opaque, rpc_op_t *rop);
+int nfs3_access_proc(void *opaque, rpc_op_t *rop);
+int nfs3_null_proc(void *opaque, rpc_op_t *rop);
+int nfs3_commit_proc(void *opaque, rpc_op_t *rop);
+int mount3_mnt_proc(void *opaque, rpc_op_t *rop);
+int mount3_dump_proc(void *opaque, rpc_op_t *rop);
+int mount3_umnt_proc(void *opaque, rpc_op_t *rop);
+int mount3_umntall_proc(void *opaque, rpc_op_t *rop);
+int mount3_export_proc(void *opaque, rpc_op_t *rop);

src/nfs_fsstat.cpp

@@ -0,0 +1,124 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy - common FSSTAT, FSINFO, PATHCONF
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+// Get file system statistics
+int nfs3_fsstat_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    //FSSTAT3args *args = (FSSTAT3args*)rop->request;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] FSSTAT\n", self->nfs_fd);
+    FSSTAT3res *reply = (FSSTAT3res*)rop->reply;
+    uint64_t tbytes = 0, fbytes = 0;
+    auto pst_it = self->parent->pool_stats.find(self->parent->default_pool_id);
+    if (pst_it != self->parent->pool_stats.end())
+    {
+        auto ttb = pst_it->second["total_raw_tb"].number_value();
+        auto ftb = (pst_it->second["total_raw_tb"].number_value() - pst_it->second["used_raw_tb"].number_value());
+        tbytes = ttb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
+        fbytes = ftb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
+    }
+    *reply = (FSSTAT3res){
+        .status = NFS3_OK,
+        .resok = (FSSTAT3resok){
+            .obj_attributes = {
+                .attributes_follow = 0,
+                //.attributes = get_root_attributes(self),
+            },
+            .tbytes = tbytes, // total bytes
+            .fbytes = fbytes, // free bytes
+            .abytes = fbytes, // available bytes
+            .tfiles = (size3)1 << (63-POOL_ID_BITS), // maximum total files
+            .ffiles = (size3)1 << (63-POOL_ID_BITS), // free files
+            .afiles = (size3)1 << (63-POOL_ID_BITS), // available files
+            .invarsec = 0,
+        },
+    };
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int nfs3_fsinfo_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    FSINFO3args *args = (FSINFO3args*)rop->request;
+    FSINFO3res *reply = (FSINFO3res*)rop->reply;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] FSINFO %s\n", self->nfs_fd, std::string(args->fsroot).c_str());
+    if (args->fsroot != NFS_ROOT_HANDLE)
+    {
+        *reply = (FSINFO3res){ .status = NFS3ERR_INVAL };
+    }
+    else
+    {
+        // Fill info
+        *reply = (FSINFO3res){
+            .status = NFS3_OK,
+            .resok = (FSINFO3resok){
+                .obj_attributes = {
+                    .attributes_follow = 0,
+                    //.attributes = get_root_attributes(self),
+                },
+                .rtmax = 128*1024*1024,
+                .rtpref = 128*1024*1024,
+                .rtmult = 4096,
+                .wtmax = 128*1024*1024,
+                .wtpref = 128*1024*1024,
+                .wtmult = 4096,
+                .dtpref = 128,
+                .maxfilesize = 0x7fffffffffffffff,
+                .time_delta = {
+                    .seconds = 1,
+                    .nseconds = 0,
+                },
+                .properties = FSF3_SYMLINK | FSF3_HOMOGENEOUS,
+            },
+        };
+    }
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int nfs3_pathconf_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    PATHCONF3args *args = (PATHCONF3args*)rop->request;
+    PATHCONF3res *reply = (PATHCONF3res*)rop->reply;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] PATHCONF %s\n", self->nfs_fd, std::string(args->object).c_str());
+    if (args->object != NFS_ROOT_HANDLE)
+    {
+        *reply = (PATHCONF3res){ .status = NFS3ERR_INVAL };
+    }
+    else
+    {
+        // Fill info
+        *reply = (PATHCONF3res){
+            .status = NFS3_OK,
+            .resok = (PATHCONF3resok){
+                .obj_attributes = {
+                    // Without at least one reference to a non-constant value (local variable or something else),
+                    // with gcc 8 we get "internal compiler error: side-effects element in no-side-effects CONSTRUCTOR" here
+                    // FIXME: get rid of this after raising compiler requirement
+                    .attributes_follow = 0,
+                    //.attributes = get_root_attributes(self),
+                },
+                .linkmax = 0,
+                .name_max = 255,
+                .no_trunc = TRUE,
+                .chown_restricted = FALSE,
+                .case_insensitive = FALSE,
+                .case_preserving = TRUE,
+            },
+        };
+    }
+    rpc_queue_reply(rop);
+    return 0;
+}

src/nfs_kv.cpp

@@ -0,0 +1,174 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - common functions
+
+#include <sys/time.h>
+
+#include "str_util.h"
+#include "nfs_proxy.h"
+#include "nfs_common.h"
+#include "nfs_kv.h"
+
+nfstime3 nfstime_from_str(const std::string & s)
+{
+    nfstime3 t;
+    auto p = s.find(".");
+    if (p != std::string::npos)
+    {
+        t.seconds = stoull_full(s.substr(0, p), 10);
+        t.nseconds = stoull_full(s.substr(p+1), 10);
+        p = s.size()-p-1;
+        for (; p < 9; p++)
+            t.nseconds *= 10;
+        for (; p > 9; p--)
+            t.nseconds /= 10;
+    }
+    else
+        t.seconds = stoull_full(s, 10);
+    return t;
+}
+
+std::string nfstime_to_str(nfstime3 t)
+{
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%u.%09u", t.seconds, t.nseconds);
+    return buf;
+}
+
+int kv_map_type(const std::string & type)
+{
+    return (type == "" || type == "file" ? NF3REG :
+        (type == "dir" ? NF3DIR :
+        (type == "blk" ? NF3BLK :
+        (type == "chr" ? NF3CHR :
+        (type == "link" ? NF3LNK :
+        (type == "sock" ? NF3SOCK :
+        (type == "fifo" ? NF3FIFO : -1)))))));
+}
+
+fattr3 get_kv_attributes(nfs_client_t *self, uint64_t ino, json11::Json attrs)
+{
+    auto type = kv_map_type(attrs["type"].string_value());
+    auto mode = attrs["mode"].uint64_value();
+    auto nlink = attrs["nlink"].uint64_value();
+    nfstime3 mtime = nfstime_from_str(attrs["mtime"].string_value());
+    nfstime3 atime = attrs["atime"].is_null() ? mtime : nfstime_from_str(attrs["atime"].string_value());
+    // FIXME In theory we could store the binary structure itself instead of JSON
+    return (fattr3){
+        .type = (type == 0 ? NF3REG : (ftype3)type),
+        .mode = (attrs["mode"].is_null() ? (type == NF3DIR ? 0755 : 0644) : (uint32_t)mode),
+        .nlink = (nlink == 0 ? 1 : (uint32_t)nlink),
+        .uid = (uint32_t)attrs["uid"].uint64_value(),
+        .gid = (uint32_t)attrs["gid"].uint64_value(),
+        .size = (type == NF3DIR ? 4096 : attrs["size"].uint64_value()),
+        .used = (type == NF3DIR ? 4096 : attrs["alloc"].uint64_value()),
+        .rdev = (type == NF3BLK || type == NF3CHR
+            ? (specdata3){ (uint32_t)attrs["major"].uint64_value(), (uint32_t)attrs["minor"].uint64_value() }
+            : (specdata3){}),
+        .fsid = self->parent->fsid,
+        .fileid = ino,
+        .atime = atime,
+        .mtime = mtime,
+        .ctime = mtime,
+    };
+}
+
+std::string kv_direntry_key(uint64_t dir_ino, const std::string & filename)
+{
+    // encode as: d <length> <hex dir_ino> / <filename>
+    char key[24] = { 0 };
+    snprintf(key, sizeof(key), "d-%lx/", dir_ino);
+    int n = strnlen(key, sizeof(key)-1) - 3;
+    if (n < 10)
+        key[1] = '0'+n;
+    else
+        key[1] = 'A'+(n-10);
+    return (char*)key + filename;
+}
+
+std::string kv_direntry_filename(const std::string & key)
+{
+    // decode as: d <length> <hex dir_ino> / <filename>
+    auto pos = key.find("/");
+    if (pos != std::string::npos)
+        return key.substr(pos+1);
+    return key;
+}
+
+std::string kv_inode_key(uint64_t ino)
+{
+    char key[24] = { 0 };
+    snprintf(key, sizeof(key), "i-%lx", ino);
+    int n = strnlen(key, sizeof(key)-1) - 2;
+    if (n < 10)
+        key[1] = '0'+n;
+    else
+        key[1] = 'A'+(n-10);
+    return std::string(key, n+2);
+}
+
+std::string kv_fh(uint64_t ino)
+{
+    return "S"+std::string((char*)&ino, 8);
+}
+
+uint64_t kv_fh_inode(const std::string & fh)
+{
+    if (fh.size() == 1 && fh[0] == 'R')
+    {
+        return 1;
+    }
+    else if (fh.size() == 9 && fh[0] == 'S')
+    {
+        return *(uint64_t*)&fh[1];
+    }
+    else if (fh.size() > 17 && fh[0] == 'I')
+    {
+        return *(uint64_t*)&fh[fh.size()-8];
+    }
+    return 0;
+}
+
+bool kv_fh_valid(const std::string & fh)
+{
+    return fh == NFS_ROOT_HANDLE || fh.size() == 9 && fh[0] == 'S' || fh.size() > 17 && fh[0] == 'I';
+}
+
+void nfs_kv_procs(nfs_client_t *self)
+{
+    struct rpc_service_proc_t pt[] = {
+        {NFS_PROGRAM, NFS_V3, NFS3_NULL,        nfs3_null_proc,           NULL,                            0,                        NULL,                           0,                       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,     kv_nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,     kv_nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,      kv_nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,      nfs3_access_proc,         (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,    kv_nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READ,        kv_nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,       kv_nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,      kv_nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,       kv_nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,     kv_nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,       kv_nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,      kv_nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,       kv_nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,      kv_nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LINK,        kv_nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,     kv_nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS, kv_nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,      nfs3_fsstat_proc,         (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,      nfs3_fsinfo_proc,         (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,    nfs3_pathconf_proc,       (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,      nfs3_commit_proc,         (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,         NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,        (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,       NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,       (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc,    NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,     NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       self},
+    };
+    for (int i = 0; i < sizeof(pt)/sizeof(pt[0]); i++)
+    {
+        self->proc_table.insert(pt[i]);
+    }
+}

src/nfs_kv.h

@@ -0,0 +1,95 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - header
+
+#pragma once
+
+#include "nfs/nfs.h"
+
+#define KV_ROOT_INODE 1
+#define KV_NEXT_ID_KEY "id"
+#define SHARED_FILE_MAGIC_V1 0x711A5158A6EDF17E
+
+struct nfs_kv_write_state;
+
+struct list_cookie_t
+{
+    uint64_t dir_ino, cookieverf, cookie;
+};
+
+inline bool operator < (const list_cookie_t & a, const list_cookie_t & b)
+{
+    return a.dir_ino < b.dir_ino || a.dir_ino == b.dir_ino &&
+        (a.cookieverf < b.cookieverf || a.cookieverf == b.cookieverf && a.cookie < b.cookie);
+};
+
+struct list_cookie_val_t
+{
+    std::string key;
+};
+
+struct shared_alloc_queue_t
+{
+    nfs_kv_write_state *st;
+    int state;
+    uint64_t size;
+};
+
+struct kv_inode_extend_t
+{
+    int refcnt = 0;
+    uint64_t cur_extend = 0, next_extend = 0, done_extend = 0;
+    std::vector<std::function<void()>> waiters;
+};
+
+struct kv_fs_state_t
+{
+    std::map<list_cookie_t, list_cookie_val_t> list_cookies;
+    uint64_t fs_next_id = 0, fs_allocated_id = 0;
+    std::vector<uint64_t> unallocated_ids;
+    std::vector<shared_alloc_queue_t> allocating_shared;
+    uint64_t cur_shared_inode = 0, cur_shared_offset = 0;
+    std::map<inode_t, kv_inode_extend_t> extends;
+};
+
+struct shared_file_header_t
+{
+    uint64_t magic = 0;
+    uint64_t inode = 0;
+    uint64_t size = 0;
+};
+
+nfsstat3 vitastor_nfs_map_err(int err);
+nfstime3 nfstime_from_str(const std::string & s);
+std::string nfstime_to_str(nfstime3 t);
+int kv_map_type(const std::string & type);
+fattr3 get_kv_attributes(nfs_client_t *self, uint64_t ino, json11::Json attrs);
+std::string kv_direntry_key(uint64_t dir_ino, const std::string & filename);
+std::string kv_direntry_filename(const std::string & key);
+std::string kv_inode_key(uint64_t ino);
+std::string kv_fh(uint64_t ino);
+uint64_t kv_fh_inode(const std::string & fh);
+bool kv_fh_valid(const std::string & fh);
+void allocate_new_id(nfs_client_t *self, std::function<void(int res, uint64_t new_id)> cb);
+void kv_read_inode(nfs_client_t *self, uint64_t ino,
+    std::function<void(int res, const std::string & value, json11::Json ientry)> cb,
+    bool allow_cache = false);
+uint64_t align_shared_size(nfs_client_t *self, uint64_t size);
+
+int kv_nfs3_getattr_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_setattr_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_lookup_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readlink_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_read_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_write_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_create_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_mkdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_symlink_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_mknod_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_remove_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_rmdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_rename_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_link_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop);

src/nfs_kv_create.cpp

@@ -0,0 +1,324 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - CREATE, MKDIR, SYMLINK, MKNOD
+
+#include <sys/time.h>
+
+#include "str_util.h"
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+void allocate_new_id(nfs_client_t *self, std::function<void(int res, uint64_t new_id)> cb)
+{
+    if (self->parent->kvfs->fs_next_id <= self->parent->kvfs->fs_allocated_id)
+    {
+        cb(0, self->parent->kvfs->fs_next_id++);
+        return;
+    }
+    else if (self->parent->kvfs->fs_next_id > self->parent->fs_inode_count)
+    {
+        cb(-ENOSPC, 0);
+        return;
+    }
+    self->parent->db->get(KV_NEXT_ID_KEY, [=](int res, const std::string & prev_str)
+    {
+        if (res < 0 && res != -ENOENT)
+        {
+            cb(res, 0);
+            return;
+        }
+        uint64_t prev_val = stoull_full(prev_str);
+        if (prev_val >= self->parent->fs_inode_count)
+        {
+            cb(-ENOSPC, 0);
+            return;
+        }
+        if (prev_val < 1)
+        {
+            prev_val = 1;
+        }
+        uint64_t new_val = prev_val + self->parent->id_alloc_batch_size;
+        if (new_val >= self->parent->fs_inode_count)
+        {
+            new_val = self->parent->fs_inode_count;
+        }
+        self->parent->db->set(KV_NEXT_ID_KEY, std::to_string(new_val), [=](int res)
+        {
+            if (res == -EAGAIN)
+            {
+                // CAS failure - retry
+                allocate_new_id(self, cb);
+            }
+            else if (res < 0)
+            {
+                cb(res, 0);
+            }
+            else
+            {
+                self->parent->kvfs->fs_next_id = prev_val+2;
+                self->parent->kvfs->fs_allocated_id = new_val;
+                cb(0, prev_val+1);
+            }
+        }, [prev_val](int res, const std::string & value)
+        {
+            // FIXME: Allow to modify value from CAS callback? ("update" query)
+            return res < 0 || stoull_full(value) == prev_val;
+        });
+    });
+}
+
+struct kv_create_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    bool exclusive = false;
+    uint64_t verf = 0;
+    uint64_t dir_ino = 0;
+    std::string filename;
+    uint64_t new_id = 0;
+    json11::Json::object attrobj;
+    json11::Json attrs;
+    std::string direntry_text;
+    uint64_t dup_ino = 0;
+    std::function<void(int res)> cb;
+};
+
+static void kv_do_create(kv_create_state *st)
+{
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] CREATE %lu/%s ATTRS %s\n", st->self->nfs_fd, st->dir_ino, st->filename.c_str(), json11::Json(st->attrobj).dump().c_str());
+    if (st->filename == "" || st->filename.find("/") != std::string::npos)
+    {
+        auto cb = std::move(st->cb);
+        cb(-EINVAL);
+        return;
+    }
+    // Generate inode ID
+    allocate_new_id(st->self, [st](int res, uint64_t new_id)
+    {
+        if (res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(res);
+            return;
+        }
+        st->new_id = new_id;
+        auto direntry = json11::Json::object{ { "ino", st->new_id } };
+        if (st->attrobj["type"].string_value() == "dir")
+        {
+            direntry["type"] = "dir";
+        }
+        st->attrs = std::move(st->attrobj);
+        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)
+        {
+            if (res < 0)
+            {
+                st->self->parent->kvfs->unallocated_ids.push_back(st->new_id);
+                if (res == -EAGAIN)
+                {
+                    if (st->dup_ino)
+                    {
+                        st->new_id = st->dup_ino;
+                        res = 0;
+                    }
+                    else
+                        res = -EEXIST;
+                }
+                else
+                    fprintf(stderr, "create %lu/%s failed: %s (code %d)\n", st->dir_ino, st->filename.c_str(), strerror(-res), res);
+                auto cb = std::move(st->cb);
+                cb(res);
+            }
+            else
+            {
+                st->self->parent->db->set(kv_inode_key(st->new_id), st->attrs.dump().c_str(), [st](int res)
+                {
+                    if (res == -EAGAIN)
+                    {
+                        res = -EEXIST;
+                    }
+                    if (res < 0)
+                    {
+                        st->self->parent->db->del(kv_direntry_key(st->dir_ino, st->filename), [st, res](int del_res)
+                        {
+                            if (!del_res)
+                            {
+                                st->self->parent->kvfs->unallocated_ids.push_back(st->new_id);
+                            }
+                            auto cb = std::move(st->cb);
+                            cb(res);
+                        }, [st](int res, const std::string & value)
+                        {
+                            return res != -ENOENT && value == st->direntry_text;
+                        });
+                    }
+                    else
+                    {
+                        auto cb = std::move(st->cb);
+                        cb(0);
+                    }
+                }, [st](int res, const std::string & value)
+                {
+                    return res == -ENOENT;
+                });
+            }
+        }, [st](int res, const std::string & value)
+        {
+            // CAS compare - check that the key doesn't exist
+            if (res == 0)
+            {
+                std::string err;
+                auto direntry = json11::Json::parse(value, err);
+                if (err != "")
+                {
+                    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)
+                {
+                    st->dup_ino = direntry["ino"].uint64_value();
+                    return false;
+                }
+                return false;
+            }
+            return true;
+        });
+    });
+}
+
+static void kv_create_setattr(json11::Json::object & attrobj, sattr3 & sattr)
+{
+    if (sattr.mode.set_it)
+        attrobj["mode"] = (uint64_t)sattr.mode.mode;
+    if (sattr.uid.set_it)
+        attrobj["uid"] = (uint64_t)sattr.uid.uid;
+    if (sattr.gid.set_it)
+        attrobj["gid"] = (uint64_t)sattr.gid.gid;
+    if (sattr.atime.set_it)
+        attrobj["atime"] = nfstime_to_str(sattr.atime.atime);
+    if (sattr.mtime.set_it)
+        attrobj["mtime"] = nfstime_to_str(sattr.mtime.mtime);
+}
+
+template<class T, class Tok> static void kv_create_reply(kv_create_state *st, int res)
+{
+    T *reply = (T*)st->rop->reply;
+    if (res < 0)
+    {
+        *reply = (T){ .status = vitastor_nfs_map_err(-res) };
+    }
+    else
+    {
+        *reply = (T){
+            .status = NFS3_OK,
+            .resok = (Tok){
+                .obj = {
+                    .handle_follows = 1,
+                    .handle = xdr_copy_string(st->rop->xdrs, kv_fh(st->new_id)),
+                },
+                .obj_attributes = {
+                    .attributes_follow = 1,
+                    .attributes = get_kv_attributes(st->self, st->new_id, st->attrs),
+                },
+            },
+        };
+    }
+    rpc_queue_reply(st->rop);
+    delete st;
+}
+
+int kv_nfs3_create_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_create_state *st = new kv_create_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    auto args = (CREATE3args*)rop->request;
+    st->exclusive = args->how.mode == NFS_EXCLUSIVE;
+    st->verf = st->exclusive ? *(uint64_t*)&args->how.verf : 0;
+    st->dir_ino = kv_fh_inode(args->where.dir);
+    st->filename = args->where.name;
+    if (args->how.mode == NFS_EXCLUSIVE)
+    {
+        st->attrobj["verf"] = *(uint64_t*)&args->how.verf;
+    }
+    else if (args->how.mode == NFS_UNCHECKED)
+    {
+        kv_create_setattr(st->attrobj, args->how.obj_attributes);
+        if (args->how.obj_attributes.size.set_it)
+        {
+            st->attrobj["size"] = (uint64_t)args->how.obj_attributes.size.size;
+            st->attrobj["empty"] = true;
+        }
+    }
+    st->cb = [st](int res) { kv_create_reply<CREATE3res, CREATE3resok>(st, res); };
+    kv_do_create(st);
+    return 1;
+}
+
+int kv_nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_create_state *st = new kv_create_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    auto args = (MKDIR3args*)rop->request;
+    st->dir_ino = kv_fh_inode(args->where.dir);
+    st->filename = args->where.name;
+    st->attrobj["type"] = "dir";
+    st->attrobj["parent_ino"] = st->dir_ino;
+    kv_create_setattr(st->attrobj, args->attributes);
+    st->cb = [st](int res) { kv_create_reply<MKDIR3res, MKDIR3resok>(st, res); };
+    kv_do_create(st);
+    return 1;
+}
+
+int kv_nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_create_state *st = new kv_create_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    auto args = (SYMLINK3args*)rop->request;
+    st->dir_ino = kv_fh_inode(args->where.dir);
+    st->filename = args->where.name;
+    st->attrobj["type"] = "link";
+    st->attrobj["symlink"] = (std::string)args->symlink.symlink_data;
+    kv_create_setattr(st->attrobj, args->symlink.symlink_attributes);
+    st->cb = [st](int res) { kv_create_reply<SYMLINK3res, SYMLINK3resok>(st, res); };
+    kv_do_create(st);
+    return 1;
+}
+
+int kv_nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_create_state *st = new kv_create_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    auto args = (MKNOD3args*)rop->request;
+    st->dir_ino = kv_fh_inode(args->where.dir);
+    st->filename = args->where.name;
+    if (args->what.type == NF3CHR || args->what.type == NF3BLK)
+    {
+        st->attrobj["type"] = (args->what.type == NF3CHR ? "chr" : "blk");
+        st->attrobj["major"] = (uint64_t)args->what.chr_device.spec.specdata1;
+        st->attrobj["minor"] = (uint64_t)args->what.chr_device.spec.specdata2;
+        kv_create_setattr(st->attrobj, args->what.chr_device.dev_attributes);
+    }
+    else if (args->what.type == NF3SOCK || args->what.type == NF3FIFO)
+    {
+        st->attrobj["type"] = (args->what.type == NF3SOCK ? "sock" : "fifo");
+        kv_create_setattr(st->attrobj, args->what.sock_attributes);
+    }
+    else
+    {
+        *(MKNOD3res*)rop->reply = (MKNOD3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->cb = [st](int res) { kv_create_reply<MKNOD3res, MKNOD3resok>(st, res); };
+    kv_do_create(st);
+    return 1;
+}

src/nfs_kv_getattr.cpp

@@ -0,0 +1,76 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - GETATTR
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+// Attributes are always stored in the inode
+void kv_read_inode(nfs_client_t *self, uint64_t ino,
+    std::function<void(int res, const std::string & value, json11::Json ientry)> cb,
+    bool allow_cache)
+{
+    auto key = kv_inode_key(ino);
+    self->parent->db->get(key, [=](int res, const std::string & value)
+    {
+        if (ino == KV_ROOT_INODE && res == -ENOENT)
+        {
+            // Allow root inode to not exist
+            cb(0, "", json11::Json(json11::Json::object{ { "type", "dir" } }));
+            return;
+        }
+        if (res < 0)
+        {
+            if (res != -ENOENT)
+                fprintf(stderr, "Error reading inode %s: %s (code %d)\n", kv_inode_key(ino).c_str(), strerror(-res), res);
+            cb(res, "", json11::Json());
+            return;
+        }
+        std::string err;
+        auto attrs = json11::Json::parse(value, err);
+        if (err != "")
+        {
+            fprintf(stderr, "Invalid JSON in inode %s = %s: %s\n", kv_inode_key(ino).c_str(), value.c_str(), err.c_str());
+            res = -EIO;
+        }
+        cb(res, value, attrs);
+    }, allow_cache);
+}
+
+int kv_nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    GETATTR3args *args = (GETATTR3args*)rop->request;
+    GETATTR3res *reply = (GETATTR3res*)rop->reply;
+    std::string fh = args->object;
+    auto ino = kv_fh_inode(fh);
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] GETATTR %lu\n", self->nfs_fd, ino);
+    if (!kv_fh_valid(fh))
+    {
+        *reply = (GETATTR3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        return 0;
+    }
+    kv_read_inode(self, ino, [=](int res, const std::string & value, json11::Json attrs)
+    {
+        if (res < 0)
+        {
+            *reply = (GETATTR3res){ .status = vitastor_nfs_map_err(-res) };
+        }
+        else
+        {
+            *reply = (GETATTR3res){
+                .status = NFS3_OK,
+                .resok = (GETATTR3resok){
+                    .obj_attributes = get_kv_attributes(self, ino, attrs),
+                },
+            };
+        }
+        rpc_queue_reply(rop);
+    });
+    return 1;
+}

src/nfs_kv_link.cpp

@@ -0,0 +1,189 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - LINK
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_kv_link_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t dir_ino = 0;
+    std::string filename;
+    std::string ientry_text;
+    json11::Json ientry;
+    bool retrying = false;
+    int wait = 0;
+    int res = 0, res2 = 0;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_link(nfs_kv_link_state *st, int state)
+{
+    // 1) Read the source inode
+    // 2) If it's a directory - fail with -EISDIR
+    // 3) Create the new direntry with the same inode reference
+    // 4) Update the inode entry with refcount++
+    // 5) Retry update if CAS failed but the inode exists
+    // 6) Otherwise fail and remove the new direntry
+    // Yeah we may leave a bad direntry if we crash
+    // But the other option is to possibly leave an inode with too big refcount
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else if (state == 4) goto resume_4;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_link()");
+        abort();
+    }
+resume_0:
+    // Check that the source inode exists and is not a directory
+    st->wait = st->retrying ? 1 : 2;
+    st->res2 = 0;
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res == 0 ? (attrs["type"].string_value() == "dir" ? -EISDIR : 0) : res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        if (!--st->wait)
+            nfs_kv_continue_link(st, 1);
+    });
+    if (!st->retrying)
+    {
+        // Check that the new directory exists
+        kv_read_inode(st->self, st->dir_ino, [st](int res, const std::string & value, json11::Json attrs)
+        {
+            st->res2 = res == 0 ? (attrs["type"].string_value() == "dir" ? 0 : -ENOTDIR) : res;
+            if (!--st->wait)
+                nfs_kv_continue_link(st, 1);
+        });
+    }
+    return;
+resume_1:
+    if (st->res < 0 || st->res2 < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res < 0 ? st->res : st->res2);
+        return;
+    }
+    // Write the new direntry
+    if (!st->retrying)
+    {
+        st->self->parent->db->set(kv_direntry_key(st->dir_ino, st->filename),
+            json11::Json(json11::Json::object{ { "ino", st->ino } }).dump(), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_link(st, 2);
+        }, [st](int res, const std::string & old_value)
+        {
+            return res == -ENOENT;
+        });
+        return;
+resume_2:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+    }
+    // Increase inode refcount
+    {
+        auto new_ientry = st->ientry.object_items();
+        auto nlink = new_ientry["nlink"].uint64_value();
+        new_ientry["nlink"] = nlink ? nlink+1 : 2;
+        st->ientry = new_ientry;
+    }
+    st->self->parent->db->set(kv_inode_key(st->ino), st->ientry.dump(), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_link(st, 3);
+    }, [st](int res, const std::string & old_value)
+    {
+        st->res2 = res;
+        return res == 0 && old_value == st->ientry_text;
+    });
+    return;
+resume_3:
+    if (st->res2 == -ENOENT)
+    {
+        st->res = -ENOENT;
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Re-read inode and retry
+        st->retrying = true;
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        // Maybe inode was deleted in the meantime, delete our direntry
+        st->self->parent->db->del(kv_direntry_key(st->dir_ino, st->filename), [st](int res)
+        {
+            st->res2 = res;
+            nfs_kv_continue_link(st, 4);
+        });
+        return;
+resume_4:
+        if (st->res2 < 0)
+        {
+            fprintf(stderr, "Warning: failed to delete new linked direntry %lu/%s: %s (code %d)\n",
+                st->dir_ino, st->filename.c_str(), strerror(-st->res2), st->res2);
+        }
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+// FIXME: We'll need some tests for the FS
+int kv_nfs3_link_proc(void *opaque, rpc_op_t *rop)
+{
+    auto st = new nfs_kv_link_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    LINK3args *args = (LINK3args*)rop->request;
+    st->ino = kv_fh_inode(args->file);
+    st->dir_ino = kv_fh_inode(args->link.dir);
+    st->filename = args->link.name;
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] LINK %lu -> %lu/%s\n", st->self->nfs_fd, st->ino, st->dir_ino, st->filename.c_str());
+    if (!st->ino || !st->dir_ino || st->filename == "")
+    {
+        LINK3res *reply = (LINK3res*)rop->reply;
+        *reply = (LINK3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->cb = [st](int res)
+    {
+        LINK3res *reply = (LINK3res*)st->rop->reply;
+        if (res < 0)
+        {
+            *reply = (LINK3res){ .status = vitastor_nfs_map_err(res) };
+        }
+        else
+        {
+            *reply = (LINK3res){
+                .status = NFS3_OK,
+                .resok = (LINK3resok){
+                    .file_attributes = (post_op_attr){
+                        .attributes_follow = 1,
+                        .attributes = get_kv_attributes(st->self, st->ino, st->ientry),
+                    },
+                },
+            };
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_link(st, 0);
+    return 1;
+}

src/nfs_kv_lookup.cpp

@@ -0,0 +1,104 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - LOOKUP, READLINK
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+int kv_nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    LOOKUP3args *args = (LOOKUP3args*)rop->request;
+    LOOKUP3res *reply = (LOOKUP3res*)rop->reply;
+    inode_t dir_ino = kv_fh_inode(args->what.dir);
+    std::string filename = args->what.name;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] LOOKUP %lu/%s\n", self->nfs_fd, dir_ino, filename.c_str());
+    if (!dir_ino || filename == "")
+    {
+        *reply = (LOOKUP3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        return 0;
+    }
+    self->parent->db->get(kv_direntry_key(dir_ino, filename), [=](int res, const std::string & value)
+    {
+        if (res < 0)
+        {
+            *reply = (LOOKUP3res){ .status = vitastor_nfs_map_err(-res) };
+            rpc_queue_reply(rop);
+            return;
+        }
+        std::string err;
+        auto direntry = json11::Json::parse(value, err);
+        if (err != "")
+        {
+            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;
+        }
+        uint64_t ino = direntry["ino"].uint64_value();
+        kv_read_inode(self, ino, [=](int res, const std::string & value, json11::Json ientry)
+        {
+            if (res < 0)
+            {
+                *reply = (LOOKUP3res){ .status = vitastor_nfs_map_err(res == -ENOENT ? -EIO : res) };
+                rpc_queue_reply(rop);
+                return;
+            }
+            *reply = (LOOKUP3res){
+                .status = NFS3_OK,
+                .resok = (LOOKUP3resok){
+                    .object = xdr_copy_string(rop->xdrs, kv_fh(ino)),
+                    .obj_attributes = {
+                        .attributes_follow = 1,
+                        .attributes = get_kv_attributes(self, ino, ientry),
+                    },
+                },
+            };
+            rpc_queue_reply(rop);
+        });
+    });
+    return 1;
+}
+
+int kv_nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    READLINK3args *args = (READLINK3args*)rop->request;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] READLINK %lu\n", self->nfs_fd, kv_fh_inode(args->symlink));
+    READLINK3res *reply = (READLINK3res*)rop->reply;
+    if (!kv_fh_valid(args->symlink) || args->symlink == NFS_ROOT_HANDLE)
+    {
+        // Invalid filehandle or trying to read symlink from root entry
+        *reply = (READLINK3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        return 0;
+    }
+    kv_read_inode(self, kv_fh_inode(args->symlink), [=](int res, const std::string & value, json11::Json attrs)
+    {
+        if (res < 0)
+        {
+            *reply = (READLINK3res){ .status = vitastor_nfs_map_err(-res) };
+        }
+        else if (attrs["type"] != "link")
+        {
+            *reply = (READLINK3res){ .status = NFS3ERR_INVAL };
+        }
+        else
+        {
+            *reply = (READLINK3res){
+                .status = NFS3_OK,
+                .resok = (READLINK3resok){
+                    .data = xdr_copy_string(rop->xdrs, attrs["symlink"].string_value()),
+                },
+            };
+        }
+        rpc_queue_reply(rop);
+    });
+    return 1;
+}

src/nfs_kv_read.cpp

@@ -0,0 +1,162 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - READ
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_kv_read_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    bool allow_cache = true;
+    inode_t ino = 0;
+    uint64_t offset = 0, size = 0;
+    std::function<void(int)> cb;
+    // state
+    int res = 0;
+    json11::Json ientry;
+    uint64_t aligned_size = 0, aligned_offset = 0;
+    uint8_t *aligned_buf = NULL;
+    cluster_op_t *op = NULL;
+    uint8_t *buf = NULL;
+};
+
+static void nfs_kv_continue_read(nfs_kv_read_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_read()");
+        abort();
+    }
+    if (st->offset + sizeof(shared_file_header_t) < st->self->parent->shared_inode_threshold)
+    {
+        kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+        {
+            st->res = res;
+            st->ientry = attrs;
+            nfs_kv_continue_read(st, 1);
+        }, st->allow_cache);
+        return;
+resume_1:
+        if (st->res < 0 || kv_map_type(st->ientry["type"].string_value()) != NF3REG)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res < 0 ? st->res : -EINVAL);
+            return;
+        }
+        if (st->ientry["shared_ino"].uint64_value() != 0)
+        {
+            st->aligned_size = align_shared_size(st->self, st->offset+st->size);
+            st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+            st->buf = st->aligned_buf + sizeof(shared_file_header_t) + st->offset;
+            st->op = new cluster_op_t;
+            st->op->opcode = OSD_OP_READ;
+            st->op->inode = st->self->parent->fs_base_inode + st->ientry["shared_ino"].uint64_value();
+            st->op->offset = st->ientry["shared_offset"].uint64_value();
+            if (st->offset+st->size > st->ientry["size"].uint64_value())
+            {
+                st->op->len = align_shared_size(st->self, st->ientry["size"].uint64_value());
+                memset(st->aligned_buf+st->op->len, 0, st->aligned_size-st->op->len);
+            }
+            else
+                st->op->len = st->aligned_size;
+            st->op->iov.push_back(st->aligned_buf, st->op->len);
+            st->op->callback = [st, state](cluster_op_t *op)
+            {
+                st->res = op->retval == op->len ? 0 : op->retval;
+                delete op;
+                nfs_kv_continue_read(st, 2);
+            };
+            st->self->parent->cli->execute(st->op);
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+            auto hdr = ((shared_file_header_t*)st->aligned_buf);
+            if (hdr->magic != SHARED_FILE_MAGIC_V1 || hdr->inode != st->ino ||
+                align_shared_size(st->self, hdr->size) > align_shared_size(st->self, st->ientry["size"].uint64_value()))
+            {
+                // Got unrelated data - retry from the beginning
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                st->allow_cache = false;
+                nfs_kv_continue_read(st, 0);
+                return;
+            }
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+    }
+    st->aligned_offset = (st->offset & ~(st->self->parent->pool_alignment-1));
+    st->aligned_size = ((st->offset + st->size + st->self->parent->pool_alignment) &
+        ~(st->self->parent->pool_alignment-1)) - st->aligned_offset;
+    st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+    st->buf = st->aligned_buf + st->offset - st->aligned_offset;
+    st->op = new cluster_op_t;
+    st->op->opcode = OSD_OP_READ;
+    st->op->inode = st->self->parent->fs_base_inode + st->ino;
+    st->op->offset = st->aligned_offset;
+    st->op->len = st->aligned_size;
+    st->op->iov.push_back(st->aligned_buf, st->aligned_size);
+    st->op->callback = [st](cluster_op_t *op)
+    {
+        st->res = op->retval;
+        delete op;
+        nfs_kv_continue_read(st, 3);
+    };
+    st->self->parent->cli->execute(st->op);
+    return;
+resume_3:
+    auto cb = std::move(st->cb);
+    cb(st->res);
+    return;
+}
+
+int kv_nfs3_read_proc(void *opaque, rpc_op_t *rop)
+{
+    READ3args *args = (READ3args*)rop->request;
+    READ3res *reply = (READ3res*)rop->reply;
+    auto ino = kv_fh_inode(args->file);
+    if (args->count > MAX_REQUEST_SIZE || !ino)
+    {
+        *reply = (READ3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        return 0;
+    }
+    auto st = new nfs_kv_read_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    st->ino = ino;
+    st->offset = args->offset;
+    st->size = args->count;
+    st->cb = [st](int res)
+    {
+        READ3res *reply = (READ3res*)st->rop->reply;
+        *reply = (READ3res){ .status = vitastor_nfs_map_err(res) };
+        if (res == 0)
+        {
+            xdr_add_malloc(st->rop->xdrs, st->aligned_buf);
+            reply->resok.data.data = (char*)st->buf;
+            reply->resok.data.size = st->size;
+            reply->resok.count = st->size;
+            reply->resok.eof = 0;
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_read(st, 0);
+    return 1;
+}

src/nfs_kv_readdir.cpp

@@ -0,0 +1,371 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - READDIR, READDIRPLUS
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+static unsigned len_pad4(unsigned len)
+{
+    return len + (len&3 ? 4-(len&3) : 0);
+}
+
+struct nfs_kv_readdir_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    // Request:
+    bool is_plus = false;
+    uint64_t cookie = 0;
+    uint64_t cookieverf = 0;
+    uint64_t dir_ino = 0;
+    uint64_t maxcount = 0;
+    std::function<void(int)> cb;
+    // State:
+    int res = 0;
+    std::string prefix, start;
+    void *list_handle;
+    uint64_t parent_ino = 0;
+    std::string ientry_text, parent_ientry_text;
+    json11::Json ientry, parent_ientry;
+    std::string cur_key, cur_value;
+    int reply_size = 0;
+    int to_skip = 0;
+    uint64_t offset = 0;
+    int getattr_running = 0, getattr_cur = 0;
+    // Result:
+    bool eof = false;
+    //uint64_t cookieverf = 0; // same field
+    std::vector<entryplus3> entries;
+};
+
+static void nfs_kv_continue_readdir(nfs_kv_readdir_state *st, int state);
+
+static void kv_getattr_next(nfs_kv_readdir_state *st)
+{
+    while (st->is_plus && st->getattr_cur < st->entries.size() && st->getattr_running < st->self->parent->readdir_getattr_parallel)
+    {
+        auto idx = st->getattr_cur++;
+        st->getattr_running++;
+        kv_read_inode(st->self, st->entries[idx].fileid, [st, idx](int res, const std::string & value, json11::Json ientry)
+        {
+            if (res == 0)
+            {
+                st->entries[idx].name_attributes = (post_op_attr){
+                    // FIXME: maybe do not read parent attributes and leave them to a GETATTR?
+                    .attributes_follow = 1,
+                    .attributes = get_kv_attributes(st->self, st->entries[idx].fileid, ientry),
+                };
+            }
+            st->getattr_running--;
+            kv_getattr_next(st);
+            if (st->getattr_running == 0 && !st->list_handle)
+            {
+                nfs_kv_continue_readdir(st, 4);
+            }
+        });
+    }
+}
+
+static void nfs_kv_continue_readdir(nfs_kv_readdir_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else if (state == 4) goto resume_4;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_readdir()");
+        abort();
+    }
+    // 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->maxcount)
+    {
+        // Error, too small max reply size
+        auto cb = std::move(st->cb);
+        cb(-NFS3ERR_TOOSMALL);
+        return;
+    }
+    // Add . and ..
+    if (st->cookie <= 1)
+    {
+        kv_read_inode(st->self, st->dir_ino, [st](int res, const std::string & value, json11::Json ientry)
+        {
+            st->res = res;
+            st->ientry_text = value;
+            st->ientry = ientry;
+            nfs_kv_continue_readdir(st, 1);
+        });
+        return;
+resume_1:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        if (st->cookie == 0)
+        {
+            auto fh = kv_fh(st->dir_ino);
+            auto entry_size = 20 + 4/*len_pad4(".")*/ + (st->is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+            if (st->reply_size + entry_size > st->maxcount)
+            {
+                auto cb = std::move(st->cb);
+                cb(-NFS3ERR_TOOSMALL);
+                return;
+            }
+            entryplus3 dot = {};
+            dot.name = xdr_copy_string(st->rop->xdrs, ".");
+            dot.fileid = st->dir_ino;
+            dot.name_attributes = (post_op_attr){
+                .attributes_follow = 1,
+                .attributes = get_kv_attributes(st->self, st->dir_ino, st->ientry),
+            };
+            dot.name_handle = (post_op_fh3){
+                .handle_follows = 1,
+                .handle = xdr_copy_string(st->rop->xdrs, fh),
+            };
+            st->entries.push_back(dot);
+            st->reply_size += entry_size;
+        }
+        st->parent_ino = st->ientry["parent_ino"].uint64_value();
+        if (st->parent_ino)
+        {
+            kv_read_inode(st->self, st->ientry["parent_ino"].uint64_value(), [st](int res, const std::string & value, json11::Json ientry)
+            {
+                st->res = res;
+                st->parent_ientry_text = value;
+                st->parent_ientry = ientry;
+                nfs_kv_continue_readdir(st, 2);
+            });
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+        }
+        auto fh = kv_fh(st->parent_ino);
+        auto entry_size = 20 + 4/*len_pad4("..")*/ + (st->is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+        if (st->reply_size + entry_size > st->maxcount)
+        {
+            st->eof = false;
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+        entryplus3 dotdot = {};
+        dotdot.name = xdr_copy_string(st->rop->xdrs, "..");
+        dotdot.fileid = st->dir_ino;
+        dotdot.name_attributes = (post_op_attr){
+            // FIXME: maybe do not read parent attributes and leave them to a GETATTR?
+            .attributes_follow = 1,
+            .attributes = get_kv_attributes(st->self,
+                st->parent_ino ? st->parent_ino : st->dir_ino,
+                st->parent_ino ? st->parent_ientry : st->ientry),
+        };
+        dotdot.name_handle = (post_op_fh3){
+            .handle_follows = 1,
+            .handle = xdr_copy_string(st->rop->xdrs, fh),
+        };
+        st->entries.push_back(dotdot);
+        st->reply_size += entry_size;
+    }
+    st->prefix = kv_direntry_key(st->dir_ino, "");
+    st->eof = true;
+    st->start = st->prefix;
+    if (st->cookie > 1)
+    {
+        auto lc_it = st->self->parent->kvfs->list_cookies.find((list_cookie_t){ st->dir_ino, st->cookieverf, st->cookie });
+        if (lc_it != st->self->parent->kvfs->list_cookies.end())
+        {
+            st->start = lc_it->second.key;
+            st->to_skip = 1;
+            st->offset = st->cookie;
+        }
+        else
+        {
+            st->to_skip = st->cookie-2;
+            st->offset = 2;
+            st->cookieverf = ((uint64_t)lrand48() | ((uint64_t)lrand48() << 31) | ((uint64_t)lrand48() << 62));
+        }
+    }
+    else
+    {
+        st->to_skip = 0;
+        st->offset = 2;
+        st->cookieverf = ((uint64_t)lrand48() | ((uint64_t)lrand48() << 31) | ((uint64_t)lrand48() << 62));
+    }
+    {
+        auto lc_it = st->self->parent->kvfs->list_cookies.lower_bound((list_cookie_t){ st->dir_ino, st->cookieverf, 0 });
+        if (lc_it != st->self->parent->kvfs->list_cookies.end() &&
+            lc_it->first.dir_ino == st->dir_ino &&
+            lc_it->first.cookieverf == st->cookieverf &&
+            lc_it->first.cookie < st->cookie)
+        {
+            auto lc_start = lc_it;
+            while (lc_it != st->self->parent->kvfs->list_cookies.end() && lc_it->first.cookieverf == st->cookieverf)
+            {
+                lc_it++;
+            }
+            st->self->parent->kvfs->list_cookies.erase(lc_start, lc_it);
+        }
+    }
+    st->getattr_cur = st->entries.size();
+    st->list_handle = st->self->parent->db->list_start(st->start);
+    st->self->parent->db->list_next(st->list_handle, [=](int res, const std::string & key, const std::string & value)
+    {
+        st->res = res;
+        st->cur_key = key;
+        st->cur_value = value;
+        nfs_kv_continue_readdir(st, 3);
+    });
+    return;
+    while (st->list_handle)
+    {
+        st->self->parent->db->list_next(st->list_handle, NULL);
+        return;
+resume_3:
+        if (st->res == -ENOENT || st->cur_key.size() > st->prefix.size() || st->cur_key.substr(0, st->prefix.size()) != st->prefix)
+        {
+            st->self->parent->db->list_close(st->list_handle);
+            st->list_handle = NULL;
+            break;
+        }
+        if (st->to_skip > 0)
+        {
+            st->to_skip--;
+            continue;
+        }
+        std::string err;
+        auto direntry = json11::Json::parse(st->cur_value, err);
+        if (err != "")
+        {
+            fprintf(stderr, "readdir: direntry %s contains invalid JSON: %s, skipping\n",
+                st->cur_key.c_str(), st->cur_value.c_str());
+            continue;
+        }
+        auto ino = direntry["ino"].uint64_value();
+        auto name = kv_direntry_filename(st->cur_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.size()) + (st->is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+        if (st->reply_size + entry_size > st->maxcount)
+        {
+            st->eof = false;
+            st->self->parent->db->list_close(st->list_handle);
+            st->list_handle = NULL;
+            break;
+        }
+        st->reply_size += entry_size;
+        auto idx = st->entries.size();
+        st->entries.push_back((entryplus3){});
+        auto entry = &st->entries[idx];
+        entry->name = xdr_copy_string(st->rop->xdrs, name);
+        entry->fileid = ino;
+        entry->cookie = st->offset++;
+        st->self->parent->kvfs->list_cookies[(list_cookie_t){ st->dir_ino, st->cookieverf, entry->cookie }] = { .key = entry->name };
+        if (st->is_plus)
+        {
+            entry->name_handle = (post_op_fh3){
+                .handle_follows = 1,
+                .handle = xdr_copy_string(st->rop->xdrs, fh),
+            };
+            kv_getattr_next(st);
+        }
+        st->self->parent->db->list_next(st->list_handle, NULL);
+    }
+resume_4:
+    while (st->getattr_running > 0)
+    {
+        return;
+    }
+    void *prev = NULL;
+    for (int i = 0; i < st->entries.size(); i++)
+    {
+        entryplus3 *entry = &st->entries[i];
+        if (prev)
+        {
+            if (st->is_plus)
+                ((entryplus3*)prev)->nextentry = entry;
+            else
+                ((entry3*)prev)->nextentry = (entry3*)entry;
+        }
+        prev = entry;
+    }
+    // Send reply
+    auto cb = std::move(st->cb);
+    cb(0);
+}
+
+static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
+{
+    auto st = new nfs_kv_readdir_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    st->is_plus = is_plus;
+    if (st->is_plus)
+    {
+        READDIRPLUS3args *args = (READDIRPLUS3args*)rop->request;
+        st->dir_ino = kv_fh_inode(args->dir);
+        st->cookie = args->cookie;
+        st->cookieverf = *((uint64_t*)args->cookieverf);
+        st->maxcount = args->maxcount;
+    }
+    else
+    {
+        READDIR3args *args = ((READDIR3args*)rop->request);
+        st->dir_ino = kv_fh_inode(args->dir);
+        st->cookie = args->cookie;
+        st->cookieverf = *((uint64_t*)args->cookieverf);
+        st->maxcount = args->count;
+    }
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] READDIR %lu VERF %lx OFFSET %lu LIMIT %lu\n", st->self->nfs_fd, st->dir_ino, st->cookieverf, st->cookie, st->maxcount);
+    st->cb = [st](int res)
+    {
+        if (st->is_plus)
+        {
+            READDIRPLUS3res *reply = (READDIRPLUS3res*)st->rop->reply;
+            *reply = (READDIRPLUS3res){ .status = vitastor_nfs_map_err(res) };
+            *(uint64_t*)(reply->resok.cookieverf) = st->cookieverf;
+            reply->resok.reply.entries = st->entries.size() ? &st->entries[0] : NULL;
+            reply->resok.reply.eof = st->eof;
+        }
+        else
+        {
+            READDIR3res *reply = (READDIR3res*)st->rop->reply;
+            *reply = (READDIR3res){ .status = vitastor_nfs_map_err(res) };
+            *(uint64_t*)(reply->resok.cookieverf) = st->cookieverf;
+            reply->resok.reply.entries = st->entries.size() ? (entry3*)&st->entries[0] : NULL;
+            reply->resok.reply.eof = st->eof;
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_readdir(st, 0);
+}
+
+int kv_nfs3_readdir_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs3_readdir_common(opaque, rop, false);
+    return 0;
+}
+
+int kv_nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs3_readdir_common(opaque, rop, true);
+    return 0;
+}

src/nfs_kv_remove.cpp

@@ -0,0 +1,313 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - REMOVE, RMDIR
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+#include "cli.h"
+
+struct kv_del_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t dir_ino = 0;
+    std::string filename;
+    uint64_t ino = 0;
+    void *list_handle = NULL;
+    std::string prefix, list_key, direntry_text, ientry_text;
+    json11::Json direntry, ientry;
+    int type = 0;
+    bool is_rmdir = false;
+    bool rm_data = false;
+    int res = 0, res2 = 0;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_delete(kv_del_state *st, int state)
+{
+    // Overall algorithm:
+    // 1) Get inode attributes and check that it's not a directory (REMOVE)
+    // 2) Get inode attributes and check that it is a directory (RMDIR)
+    // 3) Delete direntry with CAS
+    // 4) Check that the directory didn't contain files (RMDIR) and restore it if it did
+    // 5) Reduce inode refcount by 1 or delete inode
+    // 6) If regular file and inode is deleted: delete data
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    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 if (state == 7) goto resume_7;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_delete()");
+        abort();
+    }
+    st->self->parent->db->get(kv_direntry_key(st->dir_ino, st->filename), [st](int res, const std::string & value)
+    {
+        st->res = res;
+        st->direntry_text = value;
+        nfs_kv_continue_delete(st, 1);
+    });
+    return;
+resume_1:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    {
+        std::string err;
+        st->direntry = json11::Json::parse(st->direntry_text, err);
+        if (err != "")
+        {
+            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
+        {
+            st->ino = st->direntry["ino"].uint64_value();
+        }
+    }
+    // Get inode
+    st->self->parent->db->get(kv_inode_key(st->ino), [st](int res, const std::string & value)
+    {
+        st->res = res;
+        st->ientry_text = value;
+        nfs_kv_continue_delete(st, 2);
+    });
+    return;
+resume_2:
+    if (st->res < 0)
+    {
+        fprintf(stderr, "error reading inode %s: %s (code %d)\n",
+            kv_inode_key(st->ino).c_str(), strerror(-st->res), st->res);
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    {
+        std::string err;
+        st->ientry = json11::Json::parse(st->ientry_text, err);
+        if (err != "")
+        {
+            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 = kv_map_type(st->ientry["type"].string_value());
+    if (st->type == -1 || st->is_rmdir != (st->type == NF3DIR))
+    {
+        auto cb = std::move(st->cb);
+        cb(st->is_rmdir ? -ENOTDIR : -EISDIR);
+        return;
+    }
+    // (3) Delete direntry with CAS
+    st->self->parent->db->del(kv_direntry_key(st->dir_ino, st->filename), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_delete(st, 3);
+    }, [st](int res, const std::string & value)
+    {
+        return value == st->direntry_text;
+    });
+    return;
+resume_3:
+    if (st->res == -EAGAIN)
+    {
+        // CAS failure, restart from the beginning
+        nfs_kv_continue_delete(st, 0);
+        return;
+    }
+    else if (st->res < 0 && st->res != -ENOENT)
+    {
+        fprintf(stderr, "failed to remove direntry %s: %s (code %d)\n",
+            kv_direntry_key(st->dir_ino, st->filename).c_str(), strerror(-st->res), st->res);
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (!st->ino)
+    {
+        // direntry contained invalid JSON and was deleted, finish
+        auto cb = std::move(st->cb);
+        cb(0);
+        return;
+    }
+    if (st->is_rmdir)
+    {
+        // (4) Check if directory actually is not empty
+        st->list_handle = st->self->parent->db->list_start(kv_direntry_key(st->ino, ""));
+        st->self->parent->db->list_next(st->list_handle, [st](int res, const std::string & key, const std::string & value)
+        {
+            st->res = res;
+            st->list_key = key;
+            st->self->parent->db->list_close(st->list_handle);
+            nfs_kv_continue_delete(st, 4);
+        });
+        return;
+resume_4:
+        st->prefix = kv_direntry_key(st->ino, "");
+        if (st->res == -ENOENT || st->list_key.size() < st->prefix.size() || st->list_key.substr(0, st->prefix.size()) != st->prefix)
+        {
+            // OK, directory is empty
+        }
+        else
+        {
+            // Not OK, restore direntry
+            st->self->parent->db->del(kv_direntry_key(st->dir_ino, st->filename), [st](int res)
+            {
+                st->res2 = res;
+                nfs_kv_continue_delete(st, 5);
+            }, [st](int res, const std::string & value)
+            {
+                return res == -ENOENT;
+            });
+            return;
+resume_5:
+            if (st->res2 < 0)
+            {
+                fprintf(stderr, "failed to restore direntry %s (%s): %s (code %d)",
+                    kv_direntry_key(st->dir_ino, st->filename).c_str(), st->direntry_text.c_str(), strerror(-st->res2), st->res2);
+                fprintf(stderr, " - inode %lu may be left as garbage\n", st->ino);
+            }
+            if (st->res < 0)
+            {
+                fprintf(stderr, "failed to list entries from %s: %s (code %d)\n",
+                    kv_direntry_key(st->ino, "").c_str(), strerror(-st->res), st->res);
+            }
+            auto cb = std::move(st->cb);
+            cb(st->res < 0 ? st->res : -ENOTEMPTY);
+            return;
+        }
+    }
+    // (5) Reduce inode refcount by 1 or delete inode
+    if (st->ientry["nlink"].uint64_value() > 1)
+    {
+        auto copy = st->ientry.object_items();
+        copy["nlink"] = st->ientry["nlink"].uint64_value()-1;
+        st->self->parent->db->set(kv_inode_key(st->ino), json11::Json(copy).dump(), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_delete(st, 6);
+        }, [st](int res, const std::string & old_value)
+        {
+            return old_value == st->ientry_text;
+        });
+    }
+    else
+    {
+        st->self->parent->db->del(kv_inode_key(st->ino), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_delete(st, 6);
+        }, [st](int res, const std::string & old_value)
+        {
+            return old_value == st->ientry_text;
+        });
+    }
+    return;
+resume_6:
+    if (st->res < 0)
+    {
+        // Assume EAGAIN is OK, maybe someone created a hard link in the meantime
+        auto cb = std::move(st->cb);
+        cb(st->res == -EAGAIN ? 0 : st->res);
+        return;
+    }
+    // (6) If regular file and inode is deleted: delete data
+    if ((!st->type || st->type == NF3REG) && st->ientry["nlink"].uint64_value() <= 1 &&
+        !st->ientry["shared_inode"].uint64_value())
+    {
+        // Remove data
+        st->self->parent->cmd->loop_and_wait(st->self->parent->cmd->start_rm_data(json11::Json::object {
+            { "inode", INODE_NO_POOL(st->self->parent->fs_base_inode + st->ino) },
+            { "pool", (uint64_t)INODE_POOL(st->self->parent->fs_base_inode + st->ino) },
+        }), [st](const cli_result_t & r)
+        {
+            if (r.err)
+            {
+                fprintf(stderr, "Failed to remove inode %lx data: %s (code %d)\n",
+                    st->ino, r.text.c_str(), r.err);
+            }
+            st->res = r.err;
+            nfs_kv_continue_delete(st, 7);
+        });
+        return;
+resume_7:
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    auto cb = std::move(st->cb);
+    cb(0);
+}
+
+int kv_nfs3_remove_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_del_state *st = new kv_del_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    REMOVE3res *reply = (REMOVE3res*)rop->reply;
+    REMOVE3args *args = (REMOVE3args*)rop->request;
+    st->dir_ino = kv_fh_inode(args->object.dir);
+    st->filename = args->object.name;
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] REMOVE %lu/%s\n", st->self->nfs_fd, st->dir_ino, st->filename.c_str());
+    if (!st->dir_ino)
+    {
+        *reply = (REMOVE3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->cb = [st](int res)
+    {
+        *((REMOVE3res*)st->rop->reply) = (REMOVE3res){
+            .status = vitastor_nfs_map_err(res),
+        };
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_delete(st, 0);
+    return 1;
+}
+
+int kv_nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
+{
+    kv_del_state *st = new kv_del_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    RMDIR3args *args = (RMDIR3args*)rop->request;
+    RMDIR3res *reply = (RMDIR3res*)rop->reply;
+    st->dir_ino = kv_fh_inode(args->object.dir);
+    st->filename = args->object.name;
+    st->is_rmdir = true;
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] RMDIR %lu/%s\n", st->self->nfs_fd, st->dir_ino, st->filename.c_str());
+    if (!st->dir_ino)
+    {
+        *reply = (RMDIR3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->cb = [st](int res)
+    {
+        *((RMDIR3res*)st->rop->reply) = (RMDIR3res){
+            .status = vitastor_nfs_map_err(res),
+        };
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_delete(st, 0);
+    return 1;
+}

src/nfs_kv_rename.cpp

@@ -0,0 +1,206 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - RENAME
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_kv_rename_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t old_dir_ino = 0, new_dir_ino = 0;
+    std::string old_name, new_name;
+    std::string old_direntry_text;
+    std::string ientry_text;
+    json11::Json direntry, ientry;
+    int res = 0, res2 = 0;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_rename(nfs_kv_rename_state *st, int state)
+{
+    // Simplified algorithm (non-atomic and without ENOTDIR/EISDIR):
+    // 1) Check if the target directory exists
+    // 2) Delete & save (using CAS) the source direntry
+    // 3) Write the target direntry using CAS, fail if it already exists
+    // 4) Restore the source direntry on failure
+    // Atomic version would require something like a journal
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    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(stderr, "BUG: invalid state in nfs_kv_continue_rename()");
+        abort();
+    }
+    kv_read_inode(st->self, st->new_dir_ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res == 0 ? (attrs["type"].string_value() == "dir" ? 0 : -ENOTDIR) : res;
+        nfs_kv_continue_rename(st, 1);
+    });
+    return;
+resume_1:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    // Read & delete the old direntry
+    st->self->parent->db->del(kv_direntry_key(st->old_dir_ino, st->old_name), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_rename(st, 2);
+    }, [=](int res, const std::string & old_value)
+    {
+        st->res = res;
+        st->old_direntry_text = old_value;
+        return true;
+    });
+    return;
+resume_2:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        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"].string_value() == "dir" &&
+        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, const std::string & value, json11::Json ientry)
+        {
+            st->res = res;
+            st->ientry_text = value;
+            st->ientry = ientry;
+            nfs_kv_continue_rename(st, 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->old_direntry_text, [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_rename(st, 5);
+    }, [st](int res, const std::string & old_value)
+    {
+        return res == -ENOENT;
+    });
+    return;
+resume_5:
+    if (st->res < 0)
+    {
+        if (st->res == -EAGAIN)
+            st->res = -EEXIST;
+        st->res2 = st->res;
+        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, 6);
+        }, [st](int res, const std::string & old_value)
+        {
+            return res == -ENOENT;
+        });
+        return;
+resume_6:
+        if (st->res < 0)
+        {
+            if (st->res == -EAGAIN)
+                st->res = -EEXIST;
+            fprintf(stderr, "error restoring %s = %s after failed rename: %s (code %d)\n",
+                kv_direntry_key(st->old_dir_ino, st->old_name).c_str(), st->old_direntry_text.c_str(),
+                strerror(-st->res), st->res);
+        }
+        auto cb = std::move(st->cb);
+        cb(st->res2);
+        return;
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+int kv_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);
+    st->old_name = args->from.name;
+    st->new_name = args->to.name;
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] RENAME %lu/%s -> %lu/%s\n", st->self->nfs_fd, st->old_dir_ino, st->old_name.c_str(), st->new_dir_ino, st->new_name.c_str());
+    if (!st->old_dir_ino || !st->new_dir_ino || st->old_name == "" || st->new_name == "")
+    {
+        RENAME3res *reply = (RENAME3res*)rop->reply;
+        *reply = (RENAME3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    if (st->old_dir_ino == st->new_dir_ino && st->old_name == st->new_name)
+    {
+        RENAME3res *reply = (RENAME3res*)rop->reply;
+        *reply = (RENAME3res){ .status = NFS3_OK };
+        rpc_queue_reply(st->rop);
+        delete st;
+        return 0;
+    }
+    st->cb = [st](int res)
+    {
+        RENAME3res *reply = (RENAME3res*)st->rop->reply;
+        *reply = (RENAME3res){ .status = vitastor_nfs_map_err(res) };
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_rename(st, 0);
+    return 1;
+}

src/nfs_kv_setattr.cpp

@@ -0,0 +1,183 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - SETATTR
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+#include "cli.h"
+
+struct nfs_kv_setattr_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t old_size = 0, new_size = 0;
+    json11::Json::object set_attrs;
+    int res = 0, cas_res = 0;
+    std::string ientry_text;
+    json11::Json ientry;
+    json11::Json::object new_attrs;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_setattr(nfs_kv_setattr_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_setattr()");
+        abort();
+    }
+resume_0:
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        nfs_kv_continue_setattr(st, 1);
+    });
+    return;
+resume_1:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (st->ientry["type"].string_value() == "link" ||
+        st->ientry["type"].string_value() != "file" &&
+        st->ientry["type"].string_value() != "" &&
+        !st->set_attrs["size"].is_null())
+    {
+        auto cb = std::move(st->cb);
+        cb(-EINVAL);
+        return;
+    }
+    // Now we can update it
+    st->new_attrs = st->ientry.object_items();
+    st->old_size = st->ientry["size"].uint64_value();
+    for (auto & kv: st->set_attrs)
+    {
+        if (kv.first == "size")
+        {
+            st->new_size = kv.second.uint64_value();
+        }
+        st->new_attrs[kv.first] = kv.second;
+    }
+    st->self->parent->db->set(kv_inode_key(st->ino), json11::Json(st->new_attrs).dump(), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_setattr(st, 2);
+    }, [st](int res, const std::string & cas_value)
+    {
+        st->cas_res = res;
+        return (res == 0 || res == -ENOENT && st->ino == KV_ROOT_INODE) && cas_value == st->ientry_text;
+    });
+    return;
+resume_2:
+    if (st->cas_res == -ENOENT)
+    {
+        st->res = -ENOENT;
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Retry
+        fprintf(stderr, "CAS failure during setattr, retrying\n");
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        fprintf(stderr, "Failed to update inode %lu: %s (code %d)\n", st->ino, strerror(-st->res), st->res);
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (!st->set_attrs["size"].is_null() &&
+        st->ientry["size"].uint64_value() > st->set_attrs["size"].uint64_value())
+    {
+        // Delete extra data when downsizing
+        st->self->parent->cmd->loop_and_wait(st->self->parent->cmd->start_rm_data(json11::Json::object {
+            { "inode", INODE_NO_POOL(st->self->parent->fs_base_inode + st->ino) },
+            { "pool", (uint64_t)INODE_POOL(st->self->parent->fs_base_inode + st->ino) },
+            { "min_offset", st->set_attrs["size"].uint64_value() },
+        }), [st](const cli_result_t & r)
+        {
+            if (r.err)
+            {
+                fprintf(stderr, "Failed to truncate inode %lu: %s (code %d)\n",
+                    st->ino, r.text.c_str(), r.err);
+            }
+            st->res = r.err;
+            nfs_kv_continue_setattr(st, 3);
+        });
+        return;
+    }
+resume_3:
+    auto cb = std::move(st->cb);
+    cb(0);
+}
+
+int kv_nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_kv_setattr_state *st = new nfs_kv_setattr_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    auto args = (SETATTR3args*)rop->request;
+    auto reply = (SETATTR3res*)rop->reply;
+    std::string fh = args->object;
+    if (!kv_fh_valid(fh))
+    {
+        *reply = (SETATTR3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->ino = kv_fh_inode(fh);
+    if (args->new_attributes.size.set_it)
+        st->set_attrs["size"] = args->new_attributes.size.size;
+    if (args->new_attributes.mode.set_it)
+        st->set_attrs["mode"] = (uint64_t)args->new_attributes.mode.mode;
+    if (args->new_attributes.uid.set_it)
+        st->set_attrs["uid"] = (uint64_t)args->new_attributes.uid.uid;
+    if (args->new_attributes.gid.set_it)
+        st->set_attrs["gid"] = (uint64_t)args->new_attributes.gid.gid;
+    if (args->new_attributes.atime.set_it)
+        st->set_attrs["atime"] = nfstime_to_str(args->new_attributes.atime.atime);
+    if (args->new_attributes.mtime.set_it)
+        st->set_attrs["mtime"] = nfstime_to_str(args->new_attributes.mtime.mtime);
+    fprintf(stderr, "SETATTR %lu ATTRS %s\n", st->ino, json11::Json(st->set_attrs).dump().c_str());
+    st->cb = [st](int res)
+    {
+        auto reply = (SETATTR3res*)st->rop->reply;
+        if (res < 0)
+        {
+            *reply = (SETATTR3res){
+                .status = vitastor_nfs_map_err(res),
+            };
+        }
+        else
+        {
+            *reply = (SETATTR3res){
+                .status = NFS3_OK,
+                .resok = (SETATTR3resok){
+                    .obj_wcc = (wcc_data){
+                        .after = (post_op_attr){
+                            .attributes_follow = 1,
+                            .attributes = get_kv_attributes(st->self, st->ino, st->new_attrs),
+                        },
+                    },
+                },
+            };
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_setattr(st, 0);
+    return 1;
+}

src/nfs_kv_write.cpp

@@ -0,0 +1,796 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - WRITE
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_rmw_t
+{
+    nfs_kv_write_state *st = NULL;
+    int continue_state = 0;
+    uint64_t ino = 0;
+    uint64_t offset = 0;
+    uint8_t *buf = NULL;
+    uint64_t size = 0;
+    uint8_t *part_buf = NULL;
+};
+
+struct nfs_kv_write_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t offset = 0, size = 0;
+    bool stable = false;
+    uint8_t *buf = NULL;
+    std::function<void(int res)> cb;
+    // state
+    bool allow_cache = true;
+    int res = 0, res2 = 0;
+    int waiting = 0;
+    std::string ientry_text;
+    json11::Json ientry;
+    uint64_t new_size = 0;
+    uint64_t aligned_size = 0;
+    uint8_t *aligned_buf = NULL;
+    uint64_t shared_inode = 0, shared_offset = 0;
+    bool was_immediate = false;
+    nfs_rmw_t rmw[2];
+    kv_inode_extend_t *ext = NULL;
+
+    ~nfs_kv_write_state()
+    {
+        if (aligned_buf)
+        {
+            free(aligned_buf);
+            aligned_buf = NULL;
+        }
+    }
+};
+
+static void nfs_kv_continue_write(nfs_kv_write_state *st, int state);
+
+static void finish_allocate_shared(nfs_client_t *self, int res)
+{
+    std::vector<shared_alloc_queue_t> waiting;
+    waiting.swap(self->parent->kvfs->allocating_shared);
+    for (auto & w: waiting)
+    {
+        w.st->res = res;
+        if (res == 0)
+        {
+            w.st->shared_inode = self->parent->kvfs->cur_shared_inode;
+            w.st->shared_offset = self->parent->kvfs->cur_shared_offset;
+            self->parent->kvfs->cur_shared_offset += (w.size + self->parent->pool_alignment-1) & ~(self->parent->pool_alignment-1);
+        }
+        nfs_kv_continue_write(w.st, w.state);
+    }
+}
+
+static void allocate_shared_inode(nfs_kv_write_state *st, int state, uint64_t size)
+{
+    if (st->self->parent->kvfs->cur_shared_inode == 0)
+    {
+        st->self->parent->kvfs->allocating_shared.push_back({ st, state, size });
+        if (st->self->parent->kvfs->allocating_shared.size() > 1)
+        {
+            return;
+        }
+        allocate_new_id(st->self, [st](int res, uint64_t new_id)
+        {
+            if (res < 0)
+            {
+                finish_allocate_shared(st->self, res);
+                return;
+            }
+            st->self->parent->kvfs->cur_shared_inode = new_id;
+            st->self->parent->kvfs->cur_shared_offset = 0;
+            st->self->parent->db->set(
+                kv_inode_key(new_id), json11::Json(json11::Json::object{ { "type", "shared" } }).dump(),
+                [st](int res)
+                {
+                    if (res < 0)
+                    {
+                        st->self->parent->kvfs->cur_shared_inode = 0;
+                    }
+                    finish_allocate_shared(st->self, res);
+                },
+                [](int res, const std::string & old_value)
+                {
+                    return res == -ENOENT;
+                }
+            );
+        });
+    }
+}
+
+uint64_t align_shared_size(nfs_client_t *self, uint64_t size)
+{
+    return (size + sizeof(shared_file_header_t) + self->parent->pool_alignment-1)
+        & ~(self->parent->pool_alignment-1);
+}
+
+static void nfs_do_write(uint64_t ino, uint64_t offset, uint8_t *buf, uint64_t size, nfs_kv_write_state *st, int state)
+{
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_WRITE;
+    op->inode = st->self->parent->fs_base_inode + ino;
+    op->offset = offset;
+    op->len = size;
+    op->iov.push_back(buf, size);
+    st->waiting++;
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        if (op->retval != op->len)
+        {
+            st->res = op->retval >= 0 ? -EIO : op->retval;
+        }
+        delete op;
+        st->waiting--;
+        if (!st->waiting)
+        {
+            nfs_kv_continue_write(st, state);
+        }
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static void nfs_do_shared_write(nfs_kv_write_state *st, int state)
+{
+    nfs_do_write(st->shared_inode, st->shared_offset, st->aligned_buf, st->aligned_size, st, state);
+}
+
+static void nfs_do_unshare_write(nfs_kv_write_state *st, int state)
+{
+    nfs_do_write(st->ino, 0, st->aligned_buf + sizeof(shared_file_header_t),
+        st->aligned_size - sizeof(shared_file_header_t), st, state);
+}
+
+static void nfs_do_rmw(nfs_rmw_t *rmw)
+{
+    auto parent = rmw->st->self->parent;
+    auto align = parent->pool_alignment;
+    bool is_begin = (rmw->offset % align);
+    bool is_end = ((rmw->offset+rmw->size) % align);
+    // RMW either only at beginning or only at end and within a single block
+    assert(is_begin != is_end);
+    assert((rmw->offset/parent->pool_block_size) == ((rmw->offset+rmw->size-1)/parent->pool_block_size));
+    if (!rmw->part_buf)
+    {
+        rmw->part_buf = (uint8_t*)malloc_or_die(align);
+    }
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_READ;
+    op->inode = parent->fs_base_inode + rmw->ino;
+    op->offset = (rmw->offset + (is_begin ? 0 : rmw->size)) & ~(align-1);
+    op->len = align;
+    op->iov.push_back(rmw->part_buf, op->len);
+    rmw->st->waiting++;
+    op->callback = [rmw](cluster_op_t *rd_op)
+    {
+        if (rd_op->retval != rd_op->len)
+        {
+            free(rmw->part_buf);
+            rmw->part_buf = NULL;
+            rmw->st->res = rd_op->retval >= 0 ? -EIO : rd_op->retval;
+            rmw->st->waiting--;
+            if (!rmw->st->waiting)
+            {
+                nfs_kv_continue_write(rmw->st, rmw->continue_state);
+            }
+        }
+        else
+        {
+            auto parent = rmw->st->self->parent;
+            auto align = parent->pool_alignment;
+            bool is_begin = (rmw->offset % align);
+            auto op = new cluster_op_t;
+            op->opcode = OSD_OP_WRITE;
+            op->inode = rmw->st->self->parent->fs_base_inode + rmw->ino;
+            op->offset = rmw->offset & ~(align-1);
+            op->len = (rmw->size + align-1) & ~(align-1);
+            op->version = rd_op->version+1;
+            if (is_begin)
+            {
+                op->iov.push_back(rmw->part_buf, rmw->offset % align);
+            }
+            op->iov.push_back(rmw->buf, rmw->size);
+            if (!is_begin)
+            {
+                auto tail = ((rmw->offset+rmw->size) % align);
+                op->iov.push_back(rmw->part_buf + tail, align - tail);
+            }
+            op->callback = [rmw](cluster_op_t *op)
+            {
+                if (op->retval == -EAGAIN)
+                {
+                    // CAS failure - retry
+                    rmw->st->waiting--;
+                    nfs_do_rmw(rmw);
+                }
+                else
+                {
+                    free(rmw->part_buf);
+                    rmw->part_buf = NULL;
+                    if (op->retval != op->len)
+                    {
+                        rmw->st->res = (op->retval >= 0 ? -EIO : op->retval);
+                    }
+                    rmw->st->waiting--;
+                    if (!rmw->st->waiting)
+                    {
+                        nfs_kv_continue_write(rmw->st, rmw->continue_state);
+                    }
+                }
+                delete op;
+            };
+            parent->cli->execute(op);
+        }
+        delete rd_op;
+    };
+    parent->cli->execute(op);
+}
+
+static void nfs_do_shared_read(nfs_kv_write_state *st, int state)
+{
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_READ;
+    op->inode = st->self->parent->fs_base_inode + st->ientry["shared_ino"].uint64_value();
+    op->offset = st->ientry["shared_offset"].uint64_value();
+    op->len = align_shared_size(st->self, st->ientry["size"].uint64_value());
+    op->iov.push_back(st->aligned_buf, op->len);
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        st->res = op->retval == op->len ? 0 : op->retval;
+        delete op;
+        nfs_kv_continue_write(st, state);
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static void nfs_do_fsync(nfs_kv_write_state *st, int state)
+{
+    // Client requested a stable write. Add an fsync
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_SYNC;
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        delete op;
+        nfs_kv_continue_write(st, state);
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static bool nfs_do_shared_readmodify(nfs_kv_write_state *st, int base_state, int state, bool unshare)
+{
+    assert(state <= base_state);
+    if (state < base_state)       {}
+    else if (state == base_state) goto resume_0;
+    assert(!st->aligned_buf);
+    st->aligned_size = unshare
+        ? sizeof(shared_file_header_t) + (st->new_size + st->self->parent->pool_alignment-1) & ~(st->self->parent->pool_alignment-1)
+        : align_shared_size(st->self, st->new_size);
+    st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+    memset(st->aligned_buf + sizeof(shared_file_header_t), 0, st->offset);
+    if (st->ientry["shared_ino"].uint64_value() != 0)
+    {
+        // Read old data if shared non-empty
+        nfs_do_shared_read(st, base_state);
+        return false;
+resume_0:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return true;
+        }
+        auto hdr = ((shared_file_header_t*)st->aligned_buf);
+        if (hdr->magic != SHARED_FILE_MAGIC_V1 || hdr->inode != st->ino ||
+            align_shared_size(st->self, hdr->size) > align_shared_size(st->self, st->ientry["size"].uint64_value()))
+        {
+            // Got unrelated data - retry from the beginning
+            st->allow_cache = false;
+            nfs_kv_continue_write(st, 0);
+            return false;
+        }
+    }
+    *((shared_file_header_t*)st->aligned_buf) = {
+        .magic = SHARED_FILE_MAGIC_V1,
+        .inode = st->ino,
+        .size = st->new_size,
+    };
+    memcpy(st->aligned_buf + sizeof(shared_file_header_t) + st->offset, st->buf, st->size);
+    memset(st->aligned_buf + sizeof(shared_file_header_t) + st->offset + st->size, 0,
+        st->aligned_size - sizeof(shared_file_header_t) - st->offset - st->size);
+    return true;
+}
+
+static void nfs_do_align_write(nfs_kv_write_state *st, uint64_t ino, uint64_t offset, int state)
+{
+    auto alignment = st->self->parent->pool_alignment;
+    uint8_t *good_buf = st->buf;
+    uint64_t good_offset = offset;
+    uint64_t good_size = st->size;
+    st->waiting++;
+    if (offset % alignment)
+    {
+        // Requires read-modify-write in the beginning
+        auto s = (alignment - (offset % alignment));
+        if (good_size > s)
+        {
+            good_buf += s;
+            good_offset += s;
+            good_size -= s;
+        }
+        else
+            good_size = 0;
+        s = s > st->size ? st->size : s;
+        st->rmw[0] = {
+            .st = st,
+            .continue_state = state,
+            .ino = ino,
+            .offset = offset,
+            .buf = st->buf,
+            .size = s,
+        };
+        nfs_do_rmw(&st->rmw[0]);
+    }
+    if ((offset+st->size-1) % alignment)
+    {
+        // Requires read-modify-write in the end
+        auto s = ((offset+st->size-1) % alignment);
+        if (good_size > s)
+            good_size -= s;
+        else
+            good_size = 0;
+        if (((offset+st->size-1) / alignment) > (offset / alignment))
+        {
+            st->rmw[1] = {
+                .st = st,
+                .continue_state = state,
+                .ino = ino,
+                .offset = offset + st->size-s,
+                .buf = st->buf + st->size-s,
+                .size = s,
+            };
+            nfs_do_rmw(&st->rmw[1]);
+        }
+    }
+    if (good_size > 0)
+    {
+        // Normal write
+        nfs_do_write(ino, good_offset, good_buf, good_size, st, state);
+    }
+    st->waiting--;
+    if (!st->waiting)
+    {
+        nfs_kv_continue_write(st, state);
+    }
+}
+
+static std::string new_normal_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni.erase("shared_ino");
+    ni.erase("shared_offset");
+    ni.erase("shared_alloc");
+    ni.erase("shared_ver");
+    ni["size"] = st->ext->cur_extend;
+    return json11::Json(ni).dump();
+}
+
+static std::string new_moved_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni["shared_ino"] = st->shared_inode;
+    ni["shared_offset"] = st->shared_offset;
+    ni["shared_alloc"] = st->aligned_size;
+    ni.erase("shared_ver");
+    ni["size"] = st->new_size;
+    return json11::Json(ni).dump();
+}
+
+static std::string new_shared_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni["size"] = st->new_size;
+    ni["shared_ver"] = ni["shared_ver"].uint64_value()+1;
+    return json11::Json(ni).dump();
+}
+
+static void nfs_kv_extend_inode(nfs_kv_write_state *st, int state)
+{
+    if (state == 1)
+    {
+        goto resume_1;
+    }
+    st->ext->cur_extend = st->ext->next_extend;
+    st->ext->next_extend = 0;
+    st->res2 = -EAGAIN;
+    st->self->parent->db->set(kv_inode_key(st->ino), new_normal_ientry(st), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_write(st, 13);
+    }, [st](int res, const std::string & old_value)
+    {
+        if (res != 0)
+        {
+            return false;
+        }
+        if (old_value == st->ientry_text)
+        {
+            return true;
+        }
+        std::string err;
+        auto ientry = json11::Json::parse(old_value, err).object_items();
+        if (err != "")
+        {
+            st->res2 = -EINVAL;
+            return false;
+        }
+        else if (ientry.size() == st->ientry.object_items().size())
+        {
+            for (auto & kv: st->ientry.object_items())
+            {
+                if (kv.first != "size" && ientry[kv.first] != kv.second)
+                {
+                    // Something except size changed
+                    return false;
+                }
+            }
+            // OK, only size changed
+            if (ientry["size"] >= st->new_size)
+            {
+                // Already extended
+                st->res2 = 0;
+                return false;
+            }
+            // size is different but can still be extended, other parameters don't differ
+            return true;
+        }
+        return false;
+    });
+    return;
+resume_1:
+    if (st->res == -EAGAIN)
+    {
+        // EAGAIN may be OK in fact (see above)
+        st->res = st->res2;
+    }
+    if (st->res == 0)
+    {
+        st->ext->done_extend = st->ext->cur_extend;
+    }
+    st->ext->cur_extend = 0;
+    // Wake up other extenders anyway
+    auto waiters = std::move(st->ext->waiters);
+    for (auto & cb: waiters)
+    {
+        cb();
+    }
+}
+
+// Packing small files into "shared inodes". Insane algorithm...
+// Write:
+// - If (offset+size <= threshold):
+//   - Read inode from cache
+//   - If inode does not exist - stop with -ENOENT
+//   - If inode is not a regular file - stop with -EINVAL
+//   - If it's empty (size == 0 || empty == true):
+//     - If preset size is larger than threshold:
+//       - Write data into non-shared inode
+//       - In parallel: clear empty flag
+//         - If CAS failure: re-read inode and restart
+//     - Otherwise:
+//       - Allocate/take a shared inode
+//       - Allocate space in its end
+//       - Write data into shared inode
+//         - If CAS failure: allocate another shared inode and retry
+//       - Write shared inode reference, set size
+//         - If CAS failure: free allocated shared space, re-read inode and restart
+//   - If it's not empty:
+//     - If non-shared:
+//       - Write data into non-shared inode
+//       - In parallel: check if data fits into inode size and extend if it doesn't
+//         - If CAS failure: re-read inode and retry to extend the size
+//     - If shared:
+//       - Read whole file from shared inode
+//         - If the file header in data doesn't match: re-read inode and restart
+//       - If data doesn't fit into the same shared inode:
+//         - Allocate space in a new shared inode
+//         - Write data into the new shared inode
+//           - If CAS failure: allocate another shared inode and retry
+//         - Update inode metadata (set new size and new shared inode)
+//           - If CAS failure: free allocated shared space, re-read inode and restart
+//       - If it fits:
+//         - Write updated data into the shared inode
+//         - Update inode entry in any case to block parallel non-shared writes
+//           - If CAS failure: re-read inode and restart
+// - Otherwise:
+//   - Write data into non-shared inode
+//   - Read inode in parallel
+//     - If not a regular file:
+//       - Remove data
+//       - Stop with -EINVAL
+//     - If shared:
+//       - Read whole file from shared inode
+//       - Write data into non-shared inode
+//         - If CAS failure (block should not exist): restart
+//       - Update inode metadata (make non-shared, update size)
+//         - If CAS failure: restart
+//       - Zero out the shared inode header
+//         - If CAS failure: restart
+//     - Check if size fits
+//       - Extend if it doesn't
+// Read:
+// - If (offset+size <= threshold):
+//   - Read inode from cache
+//   - If empty: return zeroes
+//   - If shared:
+//     - Read the whole file from shared inode, or at least data and shared inode header
+//     - If the file header in data doesn't match: re-read inode and restart
+//   - If non-shared:
+//     - Read data from non-shared inode
+// - Otherwise:
+//   - Read data from non-shared inode
+
+static void nfs_kv_continue_write(nfs_kv_write_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    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 if (state == 7) goto resume_7;
+    else if (state == 8) goto resume_8;
+    else if (state == 9) goto resume_9;
+    else if (state == 10) goto resume_10;
+    else if (state == 11) goto resume_11;
+    else if (state == 12) goto resume_12;
+    else if (state == 13) goto resume_13;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_write()");
+        abort();
+    }
+resume_0:
+    if (!st->size)
+    {
+        auto cb = std::move(st->cb);
+        cb(0);
+        return;
+    }
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        nfs_kv_continue_write(st, 1);
+    }, st->allow_cache);
+    return;
+resume_1:
+    if (st->res < 0 ||
+        st->ientry["type"].uint64_value() != 0 &&
+        st->ientry["type"].uint64_value() != NF3REG)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res == 0 ? -EINVAL : st->res);
+        return;
+    }
+    st->was_immediate = st->self->parent->cli->get_immediate_commit(st->self->parent->fs_base_inode + st->ino);
+    st->new_size = st->ientry["size"].uint64_value();
+    if (st->new_size < st->offset + st->size)
+    {
+        st->new_size = st->offset + st->size;
+    }
+    if (st->offset + st->size + sizeof(shared_file_header_t) < st->self->parent->shared_inode_threshold)
+    {
+        if (st->ientry["size"].uint64_value() == 0 ||
+            st->ientry["empty"].bool_value() &&
+            st->ientry["size"].uint64_value() + sizeof(shared_file_header_t) < st->self->parent->shared_inode_threshold ||
+            st->ientry["shared_ino"].uint64_value() != 0 &&
+            st->ientry["size"].uint64_value() < st->offset+st->size &&
+            st->ientry["shared_alloc"].uint64_value() < align_shared_size(st->self, st->offset+st->size))
+        {
+            // Either empty, or shared and requires moving into a larger place (redirect-write)
+            allocate_shared_inode(st, 2, st->new_size);
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+resume_3:
+            if (!nfs_do_shared_readmodify(st, 3, state, false))
+                return;
+            nfs_do_shared_write(st, 4); // FIXME assemble from parts, do not copy?
+            return;
+resume_4:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+            st->self->parent->db->set(kv_inode_key(st->ino), new_moved_ientry(st), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_write(st, 5);
+            }, [st](int res, const std::string & old_value)
+            {
+                return res == 0 && old_value == st->ientry_text;
+            });
+            return;
+resume_5:
+            if (st->res < 0)
+            {
+                st->res2 = st->res;
+                memset(st->aligned_buf, 0, st->aligned_size);
+                nfs_do_shared_write(st, 6);
+                return;
+resume_6:
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                if (st->res2 == -EAGAIN)
+                {
+                    goto resume_0;
+                }
+                else
+                {
+                    auto cb = std::move(st->cb);
+                    cb(st->res2);
+                    return;
+                }
+            }
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+        else if (st->ientry["shared_ino"].uint64_value() > 0)
+        {
+            // Non-empty, shared, can be updated in-place
+            nfs_do_align_write(st, st->ientry["shared_ino"].uint64_value(),
+                st->ientry["shared_offset"].uint64_value() + sizeof(shared_file_header_t) + st->offset, 7);
+            return;
+resume_7:
+            if (st->res == 0 && st->stable && !st->was_immediate)
+            {
+                nfs_do_fsync(st, 8);
+                return;
+            }
+            // We always have to change inode entry on shared writes
+            st->self->parent->db->set(kv_inode_key(st->ino), new_shared_ientry(st), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_write(st, 8);
+            }, [st](int res, const std::string & old_value)
+            {
+                return res == 0 && old_value == st->ientry_text;
+            });
+            return;
+resume_8:
+            if (st->res == -EAGAIN)
+            {
+                goto resume_0;
+            }
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        // Fall through for non-shared
+    }
+    // Non-shared write
+    if (st->ientry["shared_ino"].uint64_value() != 0)
+    {
+        // Unshare
+resume_9:
+        if (!nfs_do_shared_readmodify(st, 9, state, true))
+            return;
+        nfs_do_unshare_write(st, 10);
+        return;
+    }
+    else
+    {
+        // Just write
+        nfs_do_align_write(st, st->ino, st->offset, 10);
+    }
+resume_10:
+    if (st->res == 0 && st->stable && !st->was_immediate)
+    {
+        nfs_do_fsync(st, 11);
+        return;
+    }
+resume_11:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (st->ientry["empty"].bool_value() ||
+        st->ientry["size"].uint64_value() < st->new_size ||
+        st->ientry["shared_ino"].uint64_value() != 0)
+    {
+        st->ext = &st->self->parent->kvfs->extends[st->ino];
+        st->ext->refcnt++;
+resume_12:
+        if (st->ext->next_extend < st->new_size)
+        {
+            // Aggregate inode extension requests
+            st->ext->next_extend = st->new_size;
+        }
+        if (st->ext->cur_extend > 0)
+        {
+            // Wait for current extend which is already in progress
+            st->ext->waiters.push_back([st](){ nfs_kv_continue_write(st, 12); });
+            return;
+        }
+        if (st->ext->done_extend < st->new_size)
+        {
+            nfs_kv_extend_inode(st, 0);
+            return;
+resume_13:
+            nfs_kv_extend_inode(st, 1);
+        }
+        st->ext->refcnt--;
+        assert(st->ext->refcnt >= 0);
+        if (st->ext->refcnt == 0)
+        {
+            st->self->parent->kvfs->extends.erase(st->ino);
+        }
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Restart
+        goto resume_0;
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+int kv_nfs3_write_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_kv_write_state *st = new nfs_kv_write_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    WRITE3args *args = (WRITE3args*)rop->request;
+    WRITE3res *reply = (WRITE3res*)rop->reply;
+    st->ino = kv_fh_inode(args->file);
+    st->offset = args->offset;
+    st->size = (args->count > args->data.size ? args->data.size : args->count);
+    if (!st->ino || st->size > MAX_REQUEST_SIZE)
+    {
+        *reply = (WRITE3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->buf = (uint8_t*)args->data.data;
+    st->stable = (args->stable != UNSTABLE);
+    st->cb = [st](int res)
+    {
+        WRITE3res *reply = (WRITE3res*)st->rop->reply;
+        *reply = (WRITE3res){ .status = vitastor_nfs_map_err(res) };
+        if (res == 0)
+        {
+            reply->resok.count = (unsigned)st->size;
+            reply->resok.committed = st->stable || st->was_immediate ? FILE_SYNC : UNSTABLE;
+            *(uint64_t*)reply->resok.verf = st->self->parent->server_id;
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_write(st, 0);
+    return 1;
+}

src/nfs_mount.cpp

@@ -0,0 +1,119 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy - common NULL, ACCESS, COMMIT, DUMP, EXPORT, MNT, UMNT, UMNTALL
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs/nfs.h"
+
+nfsstat3 vitastor_nfs_map_err(int err)
+{
+    return (err == EINVAL ? NFS3ERR_INVAL
+        : (err == ENOENT ? NFS3ERR_NOENT
+        : (err == ENOSPC ? NFS3ERR_NOSPC
+        : (err == EEXIST ? NFS3ERR_EXIST
+        : (err == EIO ? NFS3ERR_IO : (err ? NFS3ERR_IO : NFS3_OK))))));
+}
+
+int nfs3_null_proc(void *opaque, rpc_op_t *rop)
+{
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int nfs3_access_proc(void *opaque, rpc_op_t *rop)
+{
+    //nfs_client_t *self = (nfs_client_t*)opaque;
+    ACCESS3args *args = (ACCESS3args*)rop->request;
+    ACCESS3res *reply = (ACCESS3res*)rop->reply;
+    *reply = (ACCESS3res){
+        .status = NFS3_OK,
+        .resok = (ACCESS3resok){
+            .access = args->access,
+        },
+    };
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int nfs3_commit_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    //COMMIT3args *args = (COMMIT3args*)rop->request;
+    cluster_op_t *op = new cluster_op_t;
+    // fsync. we don't know how to fsync a single inode, so just fsync everything
+    op->opcode = OSD_OP_SYNC;
+    op->callback = [self, rop](cluster_op_t *op)
+    {
+        COMMIT3res *reply = (COMMIT3res*)rop->reply;
+        *reply = (COMMIT3res){ .status = vitastor_nfs_map_err(op->retval) };
+        *(uint64_t*)reply->resok.verf = self->parent->server_id;
+        rpc_queue_reply(rop);
+    };
+    self->parent->cli->execute(op);
+    return 1;
+}
+
+int mount3_mnt_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    //nfs_dirpath *args = (nfs_dirpath*)rop->request;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] MNT\n", self->nfs_fd);
+    nfs_mountres3 *reply = (nfs_mountres3*)rop->reply;
+    u_int flavor = RPC_AUTH_NONE;
+    reply->fhs_status = MNT3_OK;
+    reply->mountinfo.fhandle = xdr_copy_string(rop->xdrs, NFS_ROOT_HANDLE);
+    reply->mountinfo.auth_flavors.auth_flavors_len = 1;
+    reply->mountinfo.auth_flavors.auth_flavors_val = (u_int*)xdr_copy_string(rop->xdrs, (char*)&flavor, sizeof(u_int)).data;
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int mount3_dump_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] DUMP\n", self->nfs_fd);
+    nfs_mountlist *reply = (nfs_mountlist*)rop->reply;
+    *reply = (struct nfs_mountbody*)malloc_or_die(sizeof(struct nfs_mountbody));
+    xdr_add_malloc(rop->xdrs, *reply);
+    (*reply)->ml_hostname = xdr_copy_string(rop->xdrs, "127.0.0.1");
+    (*reply)->ml_directory = xdr_copy_string(rop->xdrs, self->parent->export_root);
+    (*reply)->ml_next = NULL;
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int mount3_umnt_proc(void *opaque, rpc_op_t *rop)
+{
+    //nfs_client_t *self = (nfs_client_t*)opaque;
+    //nfs_dirpath *arg = (nfs_dirpath*)rop->request;
+    // do nothing
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int mount3_umntall_proc(void *opaque, rpc_op_t *rop)
+{
+    // do nothing
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+int mount3_export_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    nfs_exports *reply = (nfs_exports*)rop->reply;
+    *reply = (struct nfs_exportnode*)calloc_or_die(1, sizeof(struct nfs_exportnode) + sizeof(struct nfs_groupnode));
+    xdr_add_malloc(rop->xdrs, *reply);
+    (*reply)->ex_dir = xdr_copy_string(rop->xdrs, self->parent->export_root);
+    (*reply)->ex_groups = (struct nfs_groupnode*)(reply+1);
+    (*reply)->ex_groups->gr_name = xdr_copy_string(rop->xdrs, "127.0.0.1");
+    (*reply)->ex_groups->gr_next = NULL;
+    (*reply)->ex_next = NULL;
+    rpc_queue_reply(rop);
+    return 0;
+}

src/nfs_proxy.cpp

@@ -21,6 +21,9 @@
 #include "addr_util.h"
 #include "str_util.h"
 #include "nfs_proxy.h"
+#include "nfs_kv.h"
+#include "nfs_block.h"
+#include "nfs_common.h"
 #include "http_client.h"
 #include "cli.h"
 
@@ -31,6 +34,8 @@ const char *exe_name = NULL;
 
 nfs_proxy_t::~nfs_proxy_t()
 {
+    if (db)
+        delete db;
     if (cmd)
         delete cmd;
     if (cli)
@@ -57,6 +62,7 @@ json11::Json::object nfs_proxy_t::parse_args(int narg, const char *args[])
                 "\n"
                 "USAGE:\n"
                 "  %s [STANDARD OPTIONS] [OTHER OPTIONS]\n"
+                "  --fs <META>       mount VitastorFS with metadata in image <META>\n"
                 "  --subdir <DIR>    export images prefixed <DIR>/ (default empty - export all images)\n"
                 "  --portmap 0       do not listen on port 111 (portmap/rpcbind, requires root)\n"
                 "  --bind <IP>       bind service to <IP> address (default 0.0.0.0)\n"
@@ -92,6 +98,7 @@ void nfs_proxy_t::run(json11::Json cfg)
     srand48(tv.tv_sec*1000000000 + tv.tv_nsec);
     server_id = (uint64_t)lrand48() | ((uint64_t)lrand48() << 31) | ((uint64_t)lrand48() << 62);
     // Parse options
+    trace = cfg["log_level"].uint64_value() > 5;
     bind_address = cfg["bind"].string_value();
     if (bind_address == "")
         bind_address = "0.0.0.0";
@@ -131,67 +138,12 @@ void nfs_proxy_t::run(json11::Json cfg)
     cmd->ringloop = ringloop;
     cmd->epmgr = epmgr;
     cmd->cli = cli;
-    // We need inode name hashes for NFS handles to remain stateless and <= 64 bytes long
-    dir_info[""] = (nfs_dir_t){
-        .id = 1,
-        .mod_rev = 0,
-    };
-    clock_gettime(CLOCK_REALTIME, &dir_info[""].mtime);
     watch_stats();
-    assert(cli->st_cli.on_inode_change_hook == NULL);
-    cli->st_cli.on_inode_change_hook = [this](inode_t changed_inode, bool removed)
+    if (!fs_kv_inode)
     {
-        auto inode_cfg_it = cli->st_cli.inode_config.find(changed_inode);
-        if (inode_cfg_it == cli->st_cli.inode_config.end())
-        {
-            return;
-        }
-        auto & inode_cfg = inode_cfg_it->second;
-        std::string full_name = inode_cfg.name;
-        if (name_prefix != "" && full_name.substr(0, name_prefix.size()) != name_prefix)
-        {
-            return;
-        }
-        // Calculate directory modification time and revision (used as "cookie verifier")
-        timespec now;
-        clock_gettime(CLOCK_REALTIME, &now);
-        dir_info[""].mod_rev = dir_info[""].mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dir_info[""].mod_rev;
-        dir_info[""].mtime = now;
-        int pos = full_name.find('/', name_prefix.size());
-        while (pos >= 0)
-        {
-            std::string dir = full_name.substr(0, pos);
-            auto & dinf = dir_info[dir];
-            if (!dinf.id)
-                dinf.id = next_dir_id++;
-            dinf.mod_rev = dinf.mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dinf.mod_rev;
-            dinf.mtime = now;
-            dir_by_hash["S"+base64_encode(sha256(dir))] = dir;
-            pos = full_name.find('/', pos+1);
-        }
-        // Alter inode_by_hash
-        if (removed)
-        {
-            auto ino_it = hash_by_inode.find(changed_inode);
-            if (ino_it != hash_by_inode.end())
-            {
-                inode_by_hash.erase(ino_it->second);
-                hash_by_inode.erase(ino_it);
-            }
-        }
-        else
-        {
-            std::string hash = "S"+base64_encode(sha256(full_name));
-            auto hbi_it = hash_by_inode.find(changed_inode);
-            if (hbi_it != hash_by_inode.end() && hbi_it->second != hash)
-            {
-                // inode had a different name, remove old hash=>inode pointer
-                inode_by_hash.erase(hbi_it->second);
-            }
-            inode_by_hash[hash] = changed_inode;
-            hash_by_inode[changed_inode] = hash;
-        }
-    };
+        blockfs = new block_fs_state_t();
+        blockfs->init(this);
+    }
     // Load image metadata
     while (!cli->is_ready())
     {
@@ -202,6 +154,54 @@ void nfs_proxy_t::run(json11::Json cfg)
     }
     // Check default pool
     check_default_pool();
+    // Check if we're using VitastorFS
+    fs_kv_inode = cfg["fs"].uint64_value();
+    if (!fs_kv_inode && cfg["fs"].is_string())
+    {
+        for (auto & ic: cli->st_cli.inode_config)
+        {
+            if (ic.second.name == cfg["fs"].string_value())
+            {
+                fs_kv_inode = ic.first;
+                break;
+            }
+        }
+    }
+    readdir_getattr_parallel = cfg["readdir_getattr_parallel"].uint64_value();
+    if (!readdir_getattr_parallel)
+        readdir_getattr_parallel = 8;
+    id_alloc_batch_size = cfg["id_alloc_batch_size"].uint64_value();
+    if (!id_alloc_batch_size)
+        id_alloc_batch_size = 200;
+    if (fs_kv_inode)
+    {
+        // Open DB and wait
+        int open_res = 0;
+        bool open_done = false;
+        db = new kv_dbw_t(cli);
+        db->open(fs_kv_inode, cfg, [&](int res)
+        {
+            open_done = true;
+            open_res = res;
+        });
+        while (!open_done)
+        {
+            ringloop->loop();
+            if (open_done)
+                break;
+            ringloop->wait();
+        }
+        if (open_res < 0)
+        {
+            fprintf(stderr, "Failed to open key/value filesystem metadata index: %s (code %d)\n",
+                strerror(-open_res), open_res);
+            exit(1);
+        }
+        fs_base_inode = ((uint64_t)default_pool_id << (64-POOL_ID_BITS));
+        fs_inode_count = ((uint64_t)1 << (64-POOL_ID_BITS)) - 1;
+        shared_inode_threshold = pool_block_size;
+        kvfs = new kv_fs_state_t;
+    }
     // Self-register portmap and NFS
     pmap.reg_ports.insert((portmap_id_t){
         .prog = PMAP_PROGRAM,
@@ -275,9 +275,13 @@ void nfs_proxy_t::run(json11::Json cfg)
     }
     // Destroy the client
     cli->flush();
+    delete kvfs;
+    delete db;
     delete cli;
     delete epmgr;
     delete ringloop;
+    kvfs = NULL;
+    db = NULL;
     cli = NULL;
     epmgr = NULL;
     ringloop = NULL;
@@ -382,8 +386,11 @@ void nfs_proxy_t::check_default_pool()
     {
         if (cli->st_cli.pool_config.size() == 1)
         {
-            default_pool = cli->st_cli.pool_config.begin()->second.name;
-            default_pool_id = cli->st_cli.pool_config.begin()->first;
+            auto pool_it = cli->st_cli.pool_config.begin();
+            default_pool_id = pool_it->first;
+            default_pool = pool_it->second.name;
+            pool_block_size = pool_it->second.pg_stripe_size;
+            pool_alignment = pool_it->second.bitmap_granularity;
         }
         else
         {
@@ -398,6 +405,8 @@ void nfs_proxy_t::check_default_pool()
             if (p.second.name == default_pool)
             {
                 default_pool_id = p.first;
+                pool_block_size = p.second.pg_stripe_size;
+                pool_alignment = p.second.bitmap_granularity;
                 break;
             }
         }
@@ -421,6 +430,10 @@ void nfs_proxy_t::do_accept(int listen_fd)
         int one = 1;
         setsockopt(nfs_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
         auto cli = new nfs_client_t();
+        if (fs_kv_inode)
+            nfs_kv_procs(cli);
+        else
+            nfs_block_procs(cli);
         cli->parent = this;
         cli->nfs_fd = nfs_fd;
         for (auto & fn: pmap.proc_table)

src/nfs_proxy.h

@@ -4,17 +4,18 @@
 #include "epoll_manager.h"
 #include "nfs_portmap.h"
 #include "nfs/xdr_impl.h"
+#include "kv_db.h"
 
+#define NFS_ROOT_HANDLE "R"
 #define RPC_INIT_BUF_SIZE 32768
+#define MAX_REQUEST_SIZE 128*1024*1024
+#define TRUE 1
+#define FALSE 0
 
 class cli_tool_t;
 
-struct nfs_dir_t
-{
-    uint64_t id;
-    uint64_t mod_rev;
-    timespec mtime;
-};
+struct kv_fs_state_t;
+struct block_fs_state_t;
 
 class nfs_proxy_t
 {
@@ -27,28 +28,28 @@ public:
     std::string export_root;
     bool portmap_enabled;
     unsigned nfs_port;
+    uint64_t fs_kv_inode = 0;
+    uint64_t fs_base_inode = 0;
+    uint64_t fs_inode_count = 0;
+    int readdir_getattr_parallel = 8, id_alloc_batch_size = 200;
+    int trace = 0;
 
     pool_id_t default_pool_id;
+    uint64_t pool_block_size = 0;
+    uint64_t pool_alignment = 0;
+    uint64_t shared_inode_threshold = 0;
 
     portmap_service_t pmap;
     ring_loop_t *ringloop = NULL;
     epoll_manager_t *epmgr = NULL;
     cluster_client_t *cli = NULL;
     cli_tool_t *cmd = NULL;
+    kv_dbw_t *db = NULL;
+    kv_fs_state_t *kvfs = NULL;
+    block_fs_state_t *blockfs = NULL;
 
     std::vector<XDR*> xdr_pool;
 
-    // filehandle = "S"+base64(sha256(full name with prefix)) or "roothandle" for mount root)
-
-    uint64_t next_dir_id = 2;
-    // filehandle => dir with name_prefix
-    std::map<std::string, std::string> dir_by_hash;
-    // dir with name_prefix => dir info
-    std::map<std::string, nfs_dir_t> dir_info;
-    // filehandle => inode ID
-    std::map<std::string, inode_t> inode_by_hash;
-    // inode ID => filehandle
-    std::map<inode_t, std::string> hash_by_inode;
     // inode ID => statistics
     std::map<inode_t, json11::Json> inode_stats;
     // pool ID => statistics
@@ -86,28 +87,6 @@ struct rpc_free_buffer_t
     unsigned size;
 };
 
-struct extend_size_t
-{
-    inode_t inode;
-    uint64_t new_size;
-};
-
-inline bool operator < (const extend_size_t &a, const extend_size_t &b)
-{
-    return a.inode < b.inode || a.inode == b.inode && a.new_size < b.new_size;
-}
-
-struct extend_write_t
-{
-    rpc_op_t *rop;
-    int resize_res, write_res; // 1 = started, 0 = completed OK, -errno = completed with error
-};
-
-struct extend_inode_t
-{
-    uint64_t cur_extend = 0, next_extend = 0;
-};
-
 class nfs_client_t
 {
 public:
@@ -122,8 +101,6 @@ public:
     rpc_cur_buffer_t cur_buffer = { 0 };
     std::map<uint8_t*, rpc_used_buffer_t> used_buffers;
     std::vector<rpc_free_buffer_t> free_buffers;
-    std::map<inode_t, extend_inode_t> extends;
-    std::multimap<extend_size_t, extend_write_t> extend_writes;
 
     iovec read_iov;
     msghdr read_msg = { 0 };
@@ -133,9 +110,6 @@ public:
     std::vector<iovec> send_list, next_send_list;
     std::vector<rpc_op_t*> outbox, next_outbox;
 
-    nfs_client_t();
-    ~nfs_client_t();
-
     void select_read_buffer(unsigned wanted_size);
     void submit_read(unsigned wanted_size);
     void handle_read(int result);