Move all sources to subdirs

src/nfs/CMakeLists.txt

@@ -0,0 +1,32 @@
+cmake_minimum_required(VERSION 2.8.12)
+
+project(vitastor)
+
+# vitastor-nfs
+add_executable(vitastor-nfs
+	nfs_proxy.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
+	../util/sha256.c
+	proto/xdr_impl.cpp
+	proto/rpc_xdr.cpp
+	proto/portmap_xdr.cpp
+	proto/nfs_xdr.cpp
+)
+target_link_libraries(vitastor-nfs
+	vitastor_client
+	vitastor_kv
+)

src/nfs/nfs_block.h

@@ -0,0 +1,61 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over Vitastor block images - header
+// Presents all images as files
+// Keeps image/file list in memory and is thus unsuitable for a large number of files
+
+#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
+{
+    std::string name_prefix;
+
+    // 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, json11::Json cfg);
+};
+
+nfsstat3 vitastor_nfs_map_err(int err);

src/nfs/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 "proto/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/nfs_fsstat.cpp

@@ -0,0 +1,129 @@
+// 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
+        bool_t x = FALSE;
+        *reply = (FSINFO3res){
+            .status = NFS3_OK,
+            .resok = (FSINFO3resok){
+                .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,
+                    //.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
+        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,
+                    //.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/nfs_kv.cpp

@@ -0,0 +1,337 @@
+// 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;
+}
+
+static std::string timespec_to_str(timespec t)
+{
+    char buf[64];
+    snprintf(buf, sizeof(buf), "%ju.%09ju", t.tv_sec, t.tv_nsec);
+    int l = strlen(buf);
+    while (l > 0 && buf[l-1] == '0')
+        l--;
+    if (l > 0 && buf[l-1] == '.')
+        l--;
+    buf[l] = 0;
+    return buf;
+}
+
+std::string nfstime_to_str(nfstime3 t)
+{
+    return timespec_to_str((timespec){ .tv_sec = t.seconds, .tv_nsec = t.nseconds });
+}
+
+std::string nfstime_now_str()
+{
+    timespec t;
+    clock_gettime(CLOCK_REALTIME, &t);
+    return timespec_to_str(t);
+}
+
+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());
+    nfstime3 ctime = attrs["ctime"].is_null() ? mtime : nfstime_from_str(attrs["ctime"].string_value());
+    // In theory we could store the binary structure itself, but JSON is simpler :-)
+    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()),
+        // FIXME Counting actual used file size would require reworking statistics
+        .used = (type == NF3DIR ? 4096 : attrs["size"].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 = ctime,
+    };
+}
+
+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-%jx/", 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[32] = { 0 };
+    snprintf(key, sizeof(key), "i%x", INODE_POOL(ino));
+    int n = strnlen(key, sizeof(key)-1);
+    snprintf(key+n+1, sizeof(key)-n-1, "%jx", INODE_NO_POOL(ino));
+    int m = strnlen(key+n+1, sizeof(key)-n-2);
+    key[n] = 'G'+m;
+    return std::string(key);
+}
+
+std::string kv_fh(uint64_t ino)
+{
+    char key[32] = { 0 };
+    snprintf(key, sizeof(key), "S%jx", ino);
+    return key;
+}
+
+uint64_t kv_fh_inode(const std::string & fh)
+{
+    if (fh == NFS_ROOT_HANDLE)
+    {
+        return 1;
+    }
+    else if (fh[0] == 'S')
+    {
+        uint64_t ino = 0;
+        int r = sscanf(fh.c_str()+1, "%jx", &ino);
+        if (r == 1)
+            return ino;
+    }
+    return 0;
+}
+
+bool kv_fh_valid(const std::string & fh)
+{
+    return fh == NFS_ROOT_HANDLE || fh[0] == 'S';
+}
+
+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]);
+    }
+}
+
+void kv_fs_state_t::init(nfs_proxy_t *proxy, json11::Json cfg)
+{
+    this->proxy = proxy;
+    auto & pool_cfg = proxy->cli->st_cli.pool_config.at(proxy->default_pool_id);
+    fs_kv_inode = cfg["fs"].uint64_value();
+    if (fs_kv_inode)
+    {
+        if (!INODE_POOL(fs_kv_inode))
+        {
+            fprintf(stderr, "FS metadata inode number must include pool\n");
+            exit(1);
+        }
+    }
+    else
+    {
+        for (auto & ic: proxy->cli->st_cli.inode_config)
+        {
+            if (ic.second.name == cfg["fs"].string_value())
+            {
+                fs_kv_inode = ic.first;
+                break;
+            }
+        }
+        if (!fs_kv_inode)
+        {
+            fprintf(stderr, "FS metadata image \"%s\" does not exist\n", cfg["fs"].string_value().c_str());
+            exit(1);
+        }
+    }
+    if (proxy->cli->st_cli.inode_config.find(fs_kv_inode) != proxy->cli->st_cli.inode_config.end())
+    {
+        auto & name = proxy->cli->st_cli.inode_config.at(fs_kv_inode).name;
+        if (pool_cfg.used_for_fs != name)
+        {
+            fprintf(stderr, "Please mark pool as used for this file system with `vitastor-cli modify-pool --used-for-fs %s %s`\n",
+                name.c_str(), cfg["fs"].string_value().c_str());
+            exit(1);
+        }
+    }
+    auto img_it = proxy->cli->st_cli.inode_config.lower_bound(INODE_WITH_POOL(proxy->default_pool_id+1, 0));
+    if (img_it != proxy->cli->st_cli.inode_config.begin())
+    {
+        img_it--;
+        if (img_it != proxy->cli->st_cli.inode_config.begin() && INODE_POOL(img_it->first) == proxy->default_pool_id)
+        {
+            idgen[proxy->default_pool_id].min_id = INODE_NO_POOL(img_it->first) + 1;
+        }
+    }
+    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;
+    touch_interval = cfg["touch_interval"].uint64_value();
+    if (touch_interval < 100) // ms
+        touch_interval = 100;
+    pool_block_size = pool_cfg.pg_stripe_size;
+    pool_alignment = pool_cfg.bitmap_granularity;
+    // Open DB and wait
+    int open_res = 0;
+    bool open_done = false;
+    proxy->db = new kv_dbw_t(proxy->cli);
+    std::map<std::string, std::string> kv_cfg;
+    for (auto & kv: cfg.object_items())
+    {
+        kv_cfg[kv.first] = kv.second.as_string();
+    }
+    proxy->db->open(fs_kv_inode, kv_cfg, [&](int res)
+    {
+        open_done = true;
+        open_res = res;
+    });
+    while (!open_done)
+    {
+        proxy->ringloop->loop();
+        if (open_done)
+            break;
+        proxy->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_inode_count = ((uint64_t)1 << (64-POOL_ID_BITS)) - 1;
+    shared_inode_threshold = pool_block_size;
+    if (!cfg["shared_inode_threshold"].is_null())
+    {
+        shared_inode_threshold = cfg["shared_inode_threshold"].uint64_value();
+    }
+    zero_block.resize(pool_block_size < 1048576 ? 1048576 : pool_block_size);
+    scrap_block.resize(pool_block_size < 1048576 ? 1048576 : pool_block_size);
+    touch_timer_id = proxy->epmgr->tfd->set_timer(touch_interval, true, [this](int){ touch_inodes(); });
+}
+
+kv_fs_state_t::~kv_fs_state_t()
+{
+    if (proxy && touch_timer_id >= 0)
+    {
+        proxy->epmgr->tfd->clear_timer(touch_timer_id);
+        touch_timer_id = -1;
+    }
+}
+
+static void touch_inode(nfs_proxy_t *proxy, inode_t ino, bool allow_cache)
+{
+    kv_read_inode(proxy, ino, [proxy, ino](int res, const std::string & value, json11::Json attrs)
+    {
+        if (!res)
+        {
+            auto ientry = attrs.object_items();
+            ientry["mtime"] = ientry["ctime"] = nfstime_now_str();
+            ientry.erase("verf");
+            // FIXME: Use "update" query
+            bool *found = new bool;
+            *found = true;
+            proxy->db->set(kv_inode_key(ino), json11::Json(ientry).dump(), [proxy, ino, found](int res)
+            {
+                if (!*found)
+                    res = -ENOENT;
+                delete found;
+                if (res == -EAGAIN)
+                    touch_inode(proxy, ino, false);
+            }, [value, found](int res, const std::string & old_value)
+            {
+                *found = res == 0;
+                return res == 0 && old_value == value;
+            });
+        }
+    }, allow_cache);
+}
+
+void kv_fs_state_t::touch_inodes()
+{
+    std::set<inode_t> q = std::move(touch_queue);
+    for (auto ino: q)
+    {
+        touch_inode(proxy, ino, true);
+    }
+}

src/nfs/nfs_kv.h

@@ -0,0 +1,134 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - header
+
+#pragma once
+
+#include "proto/nfs.h"
+
+#define KV_ROOT_INODE 1
+#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;
+};
+
+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_idgen_t
+{
+    uint64_t next_id = 1, allocated_id = 0;
+    uint64_t min_id = 1;
+    std::vector<uint64_t> unallocated_ids;
+};
+
+struct kv_fs_state_t
+{
+    nfs_proxy_t *proxy = NULL;
+    int touch_timer_id = -1;
+
+    uint64_t fs_kv_inode = 0;
+    uint64_t fs_inode_count = 0;
+    int readdir_getattr_parallel = 8, id_alloc_batch_size = 200;
+    uint64_t pool_block_size = 0;
+    uint64_t pool_alignment = 0;
+    uint64_t shared_inode_threshold = 0;
+    uint64_t touch_interval = 1000;
+
+    std::map<list_cookie_t, list_cookie_val_t> list_cookies;
+    std::map<pool_id_t, kv_idgen_t> idgen;
+    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;
+    std::set<inode_t> touch_queue;
+
+    std::vector<uint8_t> zero_block;
+    std::vector<uint8_t> scrap_block;
+
+    void init(nfs_proxy_t *proxy, json11::Json cfg);
+    void touch_inodes();
+    ~kv_fs_state_t();
+};
+
+struct shared_file_header_t
+{
+    uint64_t magic = 0;
+    uint64_t inode = 0;
+    uint64_t alloc = 0;
+};
+
+struct nfs_rmw_t
+{
+    nfs_proxy_t *parent = NULL;
+    uint64_t ino = 0;
+    uint64_t offset = 0;
+    uint8_t *buf = NULL;
+    uint64_t size = 0;
+    uint8_t *part_buf = NULL;
+    uint64_t version = 0;
+    nfs_rmw_t *other = NULL;
+    std::function<void(nfs_rmw_t *)> cb;
+    int res = 0;
+};
+
+nfsstat3 vitastor_nfs_map_err(int err);
+nfstime3 nfstime_from_str(const std::string & s);
+std::string nfstime_to_str(nfstime3 t);
+std::string nfstime_now_str();
+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, pool_id_t pool_id, std::function<void(int res, uint64_t new_id)> cb);
+void kv_read_inode(nfs_proxy_t *proxy, 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);
+void nfs_do_rmw(nfs_rmw_t *rmw);
+
+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/nfs_kv_create.cpp

@@ -0,0 +1,365 @@
+// 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, pool_id_t pool_id, std::function<void(int res, uint64_t new_id)> cb)
+{
+    auto & idgen = self->parent->kvfs->idgen[pool_id];
+    if (idgen.unallocated_ids.size())
+    {
+        auto new_id = idgen.unallocated_ids.back();
+        idgen.unallocated_ids.pop_back();
+        cb(0, INODE_WITH_POOL(pool_id, new_id));
+        return;
+    }
+    else if (idgen.next_id <= idgen.allocated_id)
+    {
+        idgen.next_id++;
+        cb(0, INODE_WITH_POOL(pool_id, idgen.next_id-1));
+        return;
+    }
+    // FIXME: Maybe allow FS and block volumes to cohabitate in the same pool, but with different ID ranges
+    else if (idgen.next_id >= ((uint64_t)1 << (64-POOL_ID_BITS)))
+    {
+        cb(-ENOSPC, 0);
+        return;
+    }
+    self->parent->db->get((pool_id ? "id"+std::to_string(pool_id) : "id"), [=](int res, const std::string & prev_str)
+    {
+        auto & idgen = self->parent->kvfs->idgen[pool_id];
+        if (res < 0 && res != -ENOENT)
+        {
+            cb(res, 0);
+            return;
+        }
+        uint64_t prev_val = stoull_full(prev_str);
+        if (prev_val >= ((uint64_t)1 << (64-POOL_ID_BITS)))
+        {
+            cb(-ENOSPC, 0);
+            return;
+        }
+        if (prev_val < idgen.min_id)
+        {
+            prev_val = idgen.min_id;
+        }
+        uint64_t new_val = prev_val + self->parent->kvfs->id_alloc_batch_size;
+        if (new_val >= self->parent->kvfs->fs_inode_count)
+        {
+            new_val = self->parent->kvfs->fs_inode_count;
+        }
+        self->parent->db->set((pool_id ? "id"+std::to_string(pool_id) : "id"), std::to_string(new_val), [=](int res)
+        {
+            if (res == -EAGAIN)
+            {
+                // CAS failure - retry
+                allocate_new_id(self, pool_id, cb);
+            }
+            else if (res < 0)
+            {
+                cb(res, 0);
+            }
+            else
+            {
+                auto & idgen = self->parent->kvfs->idgen[pool_id];
+                idgen.next_id = prev_val+2;
+                idgen.allocated_id = new_val;
+                cb(0, INODE_WITH_POOL(pool_id, 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;
+    // state
+    int res = 0;
+    pool_id_t pool_id = 0;
+    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_continue_create(kv_create_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;
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] CREATE %ju/%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;
+    }
+    st->attrobj["ctime"] = nfstime_now_str();
+    if (st->attrobj.find("mtime") == st->attrobj.end())
+        st->attrobj["mtime"] = st->attrobj["ctime"];
+    st->attrs = std::move(st->attrobj);
+resume_1:
+    // Generate inode ID
+    // Directories and special files don't need pool
+    st->pool_id = kv_map_type(st->attrs["type"].string_value()) == NF3REG
+        ? st->self->parent->default_pool_id
+        : 0;
+    allocate_new_id(st->self, st->pool_id, [st](int res, uint64_t new_id)
+    {
+        st->res = res;
+        st->new_id = new_id;
+        kv_continue_create(st, 2);
+    });
+    return;
+resume_2:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    st->self->parent->db->set(kv_inode_key(st->new_id), st->attrs.dump().c_str(), [st](int res)
+    {
+        st->res = res;
+        kv_continue_create(st, 3);
+    }, [st](int res, const std::string & value)
+    {
+        return res == -ENOENT;
+    });
+    return;
+resume_3:
+    if (st->res == -EAGAIN)
+    {
+        // Inode ID generator failure - retry
+        goto resume_1;
+    }
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    {
+        auto direntry = json11::Json::object{ { "ino", st->new_id } };
+        if (st->attrs["type"].string_value() == "dir")
+        {
+            direntry["type"] = "dir";
+        }
+        st->direntry_text = json11::Json(direntry).dump().c_str();
+    }
+    // Set direntry
+    st->dup_ino = 0;
+    st->self->parent->db->set(kv_direntry_key(st->dir_ino, st->filename), st->direntry_text, [st](int res)
+    {
+        st->res = res;
+        kv_continue_create(st, 4);
+    }, [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;
+    });
+    return;
+resume_4:
+    if (st->res == -EAGAIN)
+    {
+        // Direntry already exists
+        st->self->parent->db->del(kv_inode_key(st->new_id), [st](int res)
+        {
+            st->res = res;
+            kv_continue_create(st, 5);
+        });
+resume_5:
+        if (st->res < 0)
+        {
+            fprintf(stderr, "failed to delete duplicate inode %ju left from create %s (code %d)\n", st->new_id, strerror(-st->res), st->res);
+        }
+        else
+        {
+            auto & idgen = st->self->parent->kvfs->idgen[INODE_POOL(st->new_id)];
+            idgen.unallocated_ids.push_back(INODE_NO_POOL(st->new_id));
+        }
+        if (st->dup_ino)
+        {
+            // Successfully created by the previous "exclusive" request
+            st->new_id = st->dup_ino;
+        }
+        st->res = st->dup_ino ? 0 : -EEXIST;
+    }
+    if (!st->res)
+    {
+        st->self->parent->kvfs->touch_queue.insert(st->dir_ino);
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+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_continue_create(st, 0);
+    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_continue_create(st, 0);
+    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_continue_create(st, 0);
+    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_continue_create(st, 0);
+    return 1;
+}

src/nfs/nfs_kv_getattr.cpp

@@ -0,0 +1,78 @@
+// 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_proxy_t *proxy, 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);
+    proxy->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 %ju\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->parent, ino, [=](int res, const std::string & value, json11::Json attrs)
+    {
+        if (self->parent->trace)
+            fprintf(stderr, "[%d] GETATTR %ju -> %s\n", self->nfs_fd, ino, value.c_str());
+        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/nfs_kv_link.cpp

@@ -0,0 +1,193 @@
+// 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->parent, 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->parent, 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;
+        new_ientry["ctime"] = nfstime_now_str();
+        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 %ju/%s: %s (code %d)\n",
+                st->dir_ino, st->filename.c_str(), strerror(-st->res2), st->res2);
+        }
+    }
+    if (!st->res)
+    {
+        st->self->parent->kvfs->touch_queue.insert(st->dir_ino);
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+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 %ju -> %ju/%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/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 %ju/%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->parent, 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 %ju\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->parent, 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/nfs_kv_read.cpp

@@ -0,0 +1,198 @@
+// 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;
+    int eof = 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;
+};
+
+#define align_down(size) ((size) & ~(st->self->parent->kvfs->pool_alignment-1))
+#define align_up(size) (((size) + st->self->parent->kvfs->pool_alignment-1) & ~(st->self->parent->kvfs->pool_alignment-1))
+
+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();
+    }
+resume_0:
+    if (st->offset + sizeof(shared_file_header_t) < st->self->parent->kvfs->shared_inode_threshold)
+    {
+        kv_read_inode(st->self->parent, 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)
+        {
+            if (st->offset >= st->ientry["size"].uint64_value())
+            {
+                st->size = 0;
+                st->eof = 1;
+                auto cb = std::move(st->cb);
+                cb(0);
+                return;
+            }
+            st->op = new cluster_op_t;
+            {
+                st->op->opcode = OSD_OP_READ;
+                st->op->inode = st->ientry["shared_ino"].uint64_value();
+                // Always read including header to react if the file was possibly moved away
+                auto read_offset = st->ientry["shared_offset"].uint64_value();
+                st->op->offset = align_down(read_offset);
+                if (st->op->offset < read_offset)
+                {
+                    st->op->iov.push_back(st->self->parent->kvfs->scrap_block.data(),
+                        read_offset-st->op->offset);
+                }
+                auto read_size = st->offset+st->size;
+                if (read_size > st->ientry["size"].uint64_value())
+                {
+                    st->eof = 1;
+                    st->size = st->ientry["size"].uint64_value()-st->offset;
+                    read_size = st->ientry["size"].uint64_value();
+                }
+                read_size += sizeof(shared_file_header_t);
+                assert(!st->aligned_buf);
+                st->aligned_buf = (uint8_t*)malloc_or_die(read_size);
+                st->buf = st->aligned_buf + sizeof(shared_file_header_t) + st->offset;
+                st->op->iov.push_back(st->aligned_buf, read_size);
+                st->op->len = align_up(read_offset+read_size) - st->op->offset;
+                if (read_offset+read_size < st->op->offset+st->op->len)
+                {
+                    st->op->iov.push_back(st->self->parent->kvfs->scrap_block.data(),
+                        st->op->offset+st->op->len - (read_offset+read_size));
+                }
+            }
+            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)
+            {
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                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)
+            {
+                // Got unrelated data - retry from the beginning
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                st->allow_cache = false;
+                goto resume_0;
+            }
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+    }
+    st->aligned_offset = align_down(st->offset);
+    st->aligned_size = align_up(st->offset+st->size) - st->aligned_offset;
+    assert(!st->aligned_buf);
+    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->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:
+    if (st->res < 0)
+    {
+        free(st->aligned_buf);
+        st->aligned_buf = NULL;
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res < 0 ? st->res : 0);
+    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 = st->eof;
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] READ %ju %ju+%ju\n", st->self->nfs_fd, st->ino, st->offset, st->size);
+    nfs_kv_continue_read(st, 0);
+    return 1;
+}

src/nfs/nfs_kv_readdir.cpp

@@ -0,0 +1,375 @@
+// 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->kvfs->readdir_getattr_parallel)
+    {
+        auto idx = st->getattr_cur++;
+        st->getattr_running++;
+        kv_read_inode(st->self->parent, 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->parent, 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->parent, 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 = st->prefix+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);
+        if (st->self->parent->trace)
+        {
+            fprintf(stderr, "[%d] READDIR %ju %ju %s\n",
+                st->self->nfs_fd, st->dir_ino, st->offset, name.c_str());
+        }
+        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 = 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);
+        }
+    }
+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 %ju VERF %jx OFFSET %ju LIMIT %ju\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/nfs_kv_remove.cpp

@@ -0,0 +1,321 @@
+// 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;
+    bool allow_cache = true;
+    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();
+    }
+resume_0:
+    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);
+    }, st->allow_cache);
+    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);
+    }, st->allow_cache);
+    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
+        st->allow_cache = false;
+        goto resume_0;
+    }
+    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 %ju 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;
+        copy["ctime"] = nfstime_now_str();
+        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->kvfs->touch_queue.erase(st->ino);
+        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_ino"].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->ino) },
+            { "pool", (uint64_t)INODE_POOL(st->ino) },
+        }), [st](const cli_result_t & r)
+        {
+            if (r.err)
+            {
+                fprintf(stderr, "Failed to remove inode %jx 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;
+    }
+    if (!st->res)
+    {
+        st->self->parent->kvfs->touch_queue.insert(st->dir_ino);
+    }
+    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 %ju/%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 %ju/%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/nfs_kv_rename.cpp

@@ -0,0 +1,401 @@
+// 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"
+#include "cli.h"
+
+struct nfs_kv_rename_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    // params:
+    uint64_t old_dir_ino = 0, new_dir_ino = 0;
+    std::string old_name, new_name;
+    // state:
+    bool allow_cache = true;
+    std::string old_direntry_text, old_ientry_text, new_direntry_text, new_ientry_text;
+    json11::Json old_direntry, old_ientry, new_direntry, new_ientry;
+    std::string new_dir_prefix;
+    void *list_handle = NULL;
+    bool new_exists = false;
+    bool rm_dest_data = false;
+    int res = 0, res2 = 0;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_rename(nfs_kv_rename_state *st, int state)
+{
+    // Algorithm (non-atomic of course):
+    // 1) Read source direntry
+    // 2) Read destination direntry
+    // 3) If destination exists:
+    // 3.1) Check file/folder compatibility (EISDIR/ENOTDIR)
+    // 3.2) Check if destination is empty if it's a folder
+    // 4) If not:
+    // 4.1) Check that the destination directory is actually a directory
+    // 5) Overwrite destination direntry, restart from beginning if CAS failure
+    // 6) Delete source direntry, restart from beginning if CAS failure
+    // 7) If the moved direntry was a regular file:
+    // 7.1) Read inode
+    // 7.2) Delete inode if its link count <= 1
+    // 7.3) Delete inode data if its link count <= 1 and it's a regular non-shared file
+    // 7.4) Reduce link count by 1 if it's > 1
+    // 8) If the moved direntry is a directory:
+    // 8.1) Change parent_ino reference in its inode
+    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
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_rename()");
+        abort();
+    }
+resume_0:
+    // Read the old direntry
+    st->self->parent->db->get(kv_direntry_key(st->old_dir_ino, st->old_name), [=](int res, const std::string & value)
+    {
+        st->res = res;
+        st->old_direntry_text = value;
+        nfs_kv_continue_rename(st, 1);
+    }, st->allow_cache);
+    return;
+resume_1:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    {
+        std::string err;
+        st->old_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());
+            auto cb = std::move(st->cb);
+            cb(-EIO);
+            return;
+        }
+    }
+    // Read the new direntry
+    st->self->parent->db->get(kv_direntry_key(st->new_dir_ino, st->new_name), [=](int res, const std::string & value)
+    {
+        st->res = res;
+        st->new_direntry_text = value;
+        nfs_kv_continue_rename(st, 2);
+    }, st->allow_cache);
+    return;
+resume_2:
+    if (st->res < 0 && st->res != -ENOENT)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (st->res == 0)
+    {
+        std::string err;
+        st->new_direntry = json11::Json::parse(st->new_direntry_text, err);
+        if (err != "")
+        {
+            fprintf(stderr, "Invalid JSON in direntry %s = %s: %s\n",
+                kv_direntry_key(st->new_dir_ino, st->new_name).c_str(),
+                st->new_direntry_text.c_str(), err.c_str());
+            auto cb = std::move(st->cb);
+            cb(-EIO);
+            return;
+        }
+    }
+    st->new_exists = st->res == 0;
+    if (st->new_exists)
+    {
+        // Check file/folder compatibility (EISDIR/ENOTDIR)
+        if ((st->old_direntry["type"] == "dir") != (st->new_direntry["type"] == "dir"))
+        {
+            auto cb = std::move(st->cb);
+            cb((st->new_direntry["type"] == "dir") ? -ENOTDIR : -EISDIR);
+            return;
+        }
+        if (st->new_direntry["type"] == "dir")
+        {
+            // Check that the destination directory is empty
+            st->new_dir_prefix = kv_direntry_key(st->new_direntry["ino"].uint64_value(), "");
+            st->list_handle = st->self->parent->db->list_start(st->new_dir_prefix);
+            st->self->parent->db->list_next(st->list_handle, [st](int res, const std::string & key, const std::string & value)
+            {
+                st->res = res;
+                nfs_kv_continue_rename(st, 3);
+            });
+            return;
+resume_3:
+            st->self->parent->db->list_close(st->list_handle);
+            if (st->res != -ENOENT)
+            {
+                auto cb = std::move(st->cb);
+                cb(-ENOTEMPTY);
+                return;
+            }
+        }
+    }
+    else
+    {
+        // Check that the new directory is actually a directory
+        kv_read_inode(st->self->parent, 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, 4);
+        });
+        return;
+resume_4:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+    }
+    // Write the new direntry
+    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 st->new_exists ? (old_value == st->new_direntry_text) : (res == -ENOENT);
+    });
+    return;
+resume_5:
+    if (st->res == -EAGAIN)
+    {
+        // CAS failure
+        st->allow_cache = false;
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    // 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, 6);
+    }, [=](int res, const std::string & old_value)
+    {
+        return res == 0 && old_value == st->old_direntry_text;
+    });
+    return;
+resume_6:
+    if (st->res == -EAGAIN)
+    {
+        // CAS failure
+        st->allow_cache = false;
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    st->allow_cache = true;
+resume_7again:
+    if (st->new_exists && st->new_direntry["type"].string_value() != "dir")
+    {
+        // (Maybe) delete old destination file data
+        kv_read_inode(st->self->parent, st->new_direntry["ino"].uint64_value(), [st](int res, const std::string & value, json11::Json attrs)
+        {
+            st->res = res;
+            st->new_ientry_text = value;
+            st->new_ientry = attrs;
+            nfs_kv_continue_rename(st, 7);
+        }, st->allow_cache);
+        return;
+resume_7:
+        if (st->res == 0)
+        {
+            // (5) Reduce inode refcount by 1 or delete inode
+            if (st->new_ientry["nlink"].uint64_value() > 1)
+            {
+                auto copy = st->new_ientry.object_items();
+                copy["nlink"] = st->new_ientry["nlink"].uint64_value()-1;
+                copy["ctime"] = nfstime_now_str();
+                copy.erase("verf");
+                st->self->parent->db->set(kv_inode_key(st->new_direntry["ino"].uint64_value()), json11::Json(copy).dump(), [st](int res)
+                {
+                    st->res = res;
+                    nfs_kv_continue_rename(st, 8);
+                }, [st](int res, const std::string & old_value)
+                {
+                    return old_value == st->new_ientry_text;
+                });
+            }
+            else
+            {
+                st->rm_dest_data = kv_map_type(st->new_ientry["type"].string_value()) == NF3REG
+                    && !st->new_ientry["shared_ino"].uint64_value();
+                st->self->parent->db->del(kv_inode_key(st->new_direntry["ino"].uint64_value()), [st](int res)
+                {
+                    st->res = res;
+                    nfs_kv_continue_rename(st, 8);
+                }, [st](int res, const std::string & old_value)
+                {
+                    return old_value == st->new_ientry_text;
+                });
+            }
+            return;
+resume_8:
+            if (st->res == -EAGAIN)
+            {
+                // CAS failure - re-read inode
+                st->allow_cache = false;
+                goto resume_7again;
+            }
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+            // Delete inode data if required
+            if (st->rm_dest_data)
+            {
+                st->self->parent->cmd->loop_and_wait(st->self->parent->cmd->start_rm_data(json11::Json::object {
+                    { "inode", INODE_NO_POOL(st->new_direntry["ino"].uint64_value()) },
+                    { "pool", (uint64_t)INODE_POOL(st->new_direntry["ino"].uint64_value()) },
+                }), [st](const cli_result_t & r)
+                {
+                    if (r.err)
+                    {
+                        fprintf(stderr, "Failed to remove inode %jx data: %s (code %d)\n",
+                            st->new_direntry["ino"].uint64_value(), r.text.c_str(), r.err);
+                    }
+                    st->res = r.err;
+                    nfs_kv_continue_rename(st, 9);
+                });
+                return;
+resume_9:
+                if (st->res < 0)
+                {
+                    auto cb = std::move(st->cb);
+                    cb(st->res);
+                    return;
+                }
+            }
+        }
+    }
+    if (st->old_direntry["type"].string_value() == "dir" && st->new_dir_ino != st->old_dir_ino)
+    {
+        // Change parent_ino in old ientry
+        st->allow_cache = true;
+resume_10:
+        kv_read_inode(st->self->parent, st->old_direntry["ino"].uint64_value(), [st](int res, const std::string & value, json11::Json ientry)
+        {
+            st->res = res;
+            st->old_ientry_text = value;
+            st->old_ientry = ientry;
+            nfs_kv_continue_rename(st, 11);
+        }, st->allow_cache);
+        return;
+resume_11:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        {
+            auto ientry_new = st->old_ientry.object_items();
+            ientry_new["parent_ino"] = st->new_dir_ino;
+            ientry_new["ctime"] = nfstime_now_str();
+            ientry_new.erase("verf");
+            st->self->parent->db->set(kv_inode_key(st->old_direntry["ino"].uint64_value()), json11::Json(ientry_new).dump(), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_rename(st, 12);
+            }, [st](int res, const std::string & old_value)
+            {
+                return old_value == st->old_ientry_text;
+            });
+        }
+        return;
+resume_12:
+        if (st->res == -EAGAIN)
+        {
+            // CAS failure - try again
+            st->allow_cache = false;
+            goto resume_10;
+        }
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+    }
+    if (!st->res)
+    {
+        st->self->parent->kvfs->touch_queue.insert(st->old_dir_ino);
+        st->self->parent->kvfs->touch_queue.insert(st->new_dir_ino);
+    }
+    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 %ju/%s -> %ju/%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/nfs_kv_setattr.cpp

@@ -0,0 +1,204 @@
+// 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;
+    std::string expected_ctime;
+    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)
+{
+    // FIXME: NFS client does a lot of setattr calls, so maybe process them asynchronously
+    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();
+    }
+    st->self->parent->kvfs->touch_queue.erase(st->ino);
+resume_0:
+    kv_read_inode(st->self->parent, 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() != "file" &&
+        st->ientry["type"].string_value() != "" &&
+        !st->set_attrs["size"].is_null())
+    {
+        auto cb = std::move(st->cb);
+        cb(-EINVAL);
+        return;
+    }
+    if (st->expected_ctime != "")
+    {
+        auto actual_ctime = (st->ientry["ctime"].is_null() ? st->ientry["mtime"] : st->ientry["ctime"]);
+        if (actual_ctime != st->expected_ctime)
+        {
+            auto cb = std::move(st->cb);
+            cb(NFS3ERR_NOT_SYNC);
+            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->new_attrs.erase("verf");
+    st->new_attrs["ctime"] = nfstime_now_str();
+    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
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        fprintf(stderr, "Failed to update inode %ju: %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() &&
+        !st->ientry["shared_ino"].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->ino) },
+            { "pool", (uint64_t)INODE_POOL(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 %ju: %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(st->res);
+}
+
+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->guard.check)
+        st->expected_ctime = nfstime_to_str(args->guard.obj_ctime);
+    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 == SET_TO_SERVER_TIME)
+        st->set_attrs["atime"] = nfstime_now_str();
+    else if (args->new_attributes.atime.set_it == SET_TO_CLIENT_TIME)
+        st->set_attrs["atime"] = nfstime_to_str(args->new_attributes.atime.atime);
+    if (args->new_attributes.mtime.set_it == SET_TO_SERVER_TIME)
+        st->set_attrs["mtime"] = nfstime_now_str();
+    else if (args->new_attributes.mtime.set_it == SET_TO_CLIENT_TIME)
+        st->set_attrs["mtime"] = nfstime_to_str(args->new_attributes.mtime.mtime);
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] SETATTR %ju ATTRS %s\n", st->self->nfs_fd, 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/nfs_mount.cpp

@@ -0,0 +1,126 @@
+// 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 "proto/nfs.h"
+
+nfsstat3 vitastor_nfs_map_err(int err)
+{
+    if (err < 0)
+    {
+        err = -err;
+    }
+    return (err == EINVAL ? NFS3ERR_INVAL
+        : (err == ENOENT ? NFS3ERR_NOENT
+        : (err == ENOSPC ? NFS3ERR_NOSPC
+        : (err == EEXIST ? NFS3ERR_EXIST
+        : (err == EISDIR ? NFS3ERR_ISDIR
+        : (err == ENOTDIR ? NFS3ERR_NOTDIR
+        : (err == ENOTEMPTY ? NFS3ERR_NOTEMPTY
+        : (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/nfs_portmap.cpp

@@ -0,0 +1,184 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Portmap service for NFS proxy
+
+#include <netinet/in.h>
+#include <string.h>
+
+#include "proto/portmap.h"
+#include "proto/xdr_impl_inline.h"
+
+#include "malloc_or_die.h"
+#include "nfs_portmap.h"
+#include "sha256.h"
+#include "str_util.h"
+
+/*
+ * The NULL procedure. All protocols/versions must provide a NULL procedure
+ * as index 0.
+ * It is used by clients, and rpcinfo, to "ping" a service and verify that
+ * the service is available and that it does support the indicated version.
+ */
+static int pmap2_null_proc(struct rpc_context *rpc, rpc_op_t *rop)
+{
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+/*
+ * v2 GETPORT.
+ * This is the lookup function for portmapper version 2.
+ * A client provides program, version and protocol (tcp or udp)
+ * and portmapper returns which port that service is available on,
+ * (or 0 if no such program is registered.)
+ */
+static int pmap2_getport_proc(portmap_service_t *self, rpc_op_t *rop)
+{
+    PMAP2GETPORTargs *args = (PMAP2GETPORTargs *)rop->request;
+    uint32_t *reply = (uint32_t *)rop->reply;
+    auto it = self->reg_ports.lower_bound((portmap_id_t){
+        .prog = args->prog,
+        .vers = args->vers,
+        .udp = args->prot == IPPROTO_UDP,
+        .ipv6 = false,
+    });
+    if (it != self->reg_ports.end() &&
+        it->prog == args->prog && it->vers == args->vers &&
+        it->udp == (args->prot == IPPROTO_UDP))
+    {
+        *reply = it->port;
+    }
+    else
+    {
+        *reply = 0;
+    }
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+/*
+ * v2 DUMP.
+ * This RPC returns a list of all endpoints that are registered with
+ * portmapper.
+ */
+static int pmap2_dump_proc(portmap_service_t *self, rpc_op_t *rop)
+{
+    pmap2_mapping_list *list = (pmap2_mapping_list*)malloc_or_die(sizeof(pmap2_mapping_list) * self->reg_ports.size());
+    xdr_add_malloc(rop->xdrs, list);
+    PMAP2DUMPres *reply = (PMAP2DUMPres *)rop->reply;
+    int i = 0;
+    for (auto it = self->reg_ports.begin(); it != self->reg_ports.end(); it++)
+    {
+        if (it->ipv6)
+            continue;
+        list[i] = {
+            .map = {
+                .prog = it->prog,
+                .vers = it->vers,
+                .prot = it->udp ? IPPROTO_UDP : IPPROTO_TCP,
+                .port = it->port,
+            },
+            .next = list+i+1,
+        };
+        i++;
+    }
+    list[i-1].next = NULL;
+    // Send reply
+    reply->list = list;
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+/*
+ * v3 GETADDR.
+ * This is the lookup function for portmapper version 3.
+ */
+static int pmap3_getaddr_proc(portmap_service_t *self, rpc_op_t *rop)
+{
+    PMAP3GETADDRargs *args = (PMAP3GETADDRargs *)rop->request;
+    PMAP3GETADDRres *reply = (PMAP3GETADDRres *)rop->reply;
+    portmap_id_t ref = (portmap_id_t){
+        .prog = args->prog,
+        .vers = args->vers,
+        .udp = args->netid == "udp" || args->netid == "udp6",
+        .ipv6 = args->netid == "tcp6" || args->netid == "udp6",
+    };
+    auto it = self->reg_ports.lower_bound(ref);
+    if (it != self->reg_ports.end() &&
+        it->prog == ref.prog && it->vers == ref.vers &&
+        it->udp == ref.udp && it->ipv6 == ref.ipv6)
+    {
+        reply->addr = xdr_copy_string(rop->xdrs, it->addr);
+    }
+    else
+    {
+        reply->addr = {};
+    }
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+/*
+ * v3 DUMP.
+ * This RPC returns a list of all endpoints that are registered with
+ * portmapper.
+ */
+static std::string netid_udp = "udp";
+static std::string netid_udp6 = "udp6";
+static std::string netid_tcp = "tcp";
+static std::string netid_tcp6 = "tcp6";
+static int pmap3_dump_proc(portmap_service_t *self, rpc_op_t *rop)
+{
+    PMAP3DUMPres *reply = (PMAP3DUMPres *)rop->reply;
+    pmap3_mapping_list *list = (pmap3_mapping_list*)malloc_or_die(sizeof(pmap3_mapping_list*) * self->reg_ports.size());
+    xdr_add_malloc(rop->xdrs, list);
+    int i = 0;
+    for (auto it = self->reg_ports.begin(); it != self->reg_ports.end(); it++)
+    {
+        list[i] = (pmap3_mapping_list){
+            .map = (pmap3_mapping){
+                .prog  = it->prog,
+                .vers  = it->vers,
+                .netid = xdr_copy_string(rop->xdrs, it->ipv6
+                    ? (it->udp ? netid_udp6 : netid_tcp6)
+                    : (it->udp ? netid_udp : netid_tcp)),
+                .addr  = xdr_copy_string(rop->xdrs, it->addr), // 0.0.0.0.port
+                .owner = xdr_copy_string(rop->xdrs, it->owner),
+            },
+            .next = list+i+1,
+        };
+        i++;
+    }
+    list[i-1].next = NULL;
+    reply->list = list;
+    rpc_queue_reply(rop);
+    return 0;
+}
+
+portmap_service_t::portmap_service_t()
+{
+    struct rpc_service_proc_t pt[] = {
+        {PMAP_PROGRAM, PMAP_V2, PMAP2_NULL, (rpc_handler_t)pmap2_null_proc, NULL, 0, NULL, 0, this},
+        {PMAP_PROGRAM, PMAP_V2, PMAP2_GETPORT, (rpc_handler_t)pmap2_getport_proc, (xdrproc_t)xdr_PMAP2GETPORTargs, sizeof(PMAP2GETPORTargs), (xdrproc_t)xdr_u_int, sizeof(u_int), this},
+        {PMAP_PROGRAM, PMAP_V2, PMAP2_DUMP, (rpc_handler_t)pmap2_dump_proc, NULL, 0, (xdrproc_t)xdr_PMAP2DUMPres, sizeof(PMAP2DUMPres), this},
+        {PMAP_PROGRAM, PMAP_V3, PMAP3_NULL, (rpc_handler_t)pmap2_null_proc, NULL, 0, NULL, 0, this},
+        {PMAP_PROGRAM, PMAP_V3, PMAP3_GETADDR, (rpc_handler_t)pmap3_getaddr_proc, (xdrproc_t)xdr_PMAP3GETADDRargs, sizeof(PMAP3GETADDRargs), (xdrproc_t)xdr_string, sizeof(xdr_string_t), this},
+        {PMAP_PROGRAM, PMAP_V3, PMAP3_DUMP, (rpc_handler_t)pmap3_dump_proc, NULL, 0, (xdrproc_t)xdr_PMAP3DUMPres, sizeof(PMAP3DUMPres), this},
+    };
+    for (int i = 0; i < sizeof(pt)/sizeof(pt[0]); i++)
+    {
+        proc_table.push_back(pt[i]);
+    }
+}
+
+std::string sha256(const std::string & str)
+{
+    std::string hash;
+    hash.resize(32);
+    SHA256_CTX ctx;
+    sha256_init(&ctx);
+    sha256_update(&ctx, (uint8_t*)str.data(), str.size());
+    sha256_final(&ctx, (uint8_t*)hash.data());
+    return hash;
+}

src/nfs/nfs_portmap.h

@@ -0,0 +1,39 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Portmap service for NFS proxy
+
+#pragma once
+
+#include <string>
+#include <set>
+#include <vector>
+
+#include "proto/rpc_impl.h"
+
+struct portmap_id_t
+{
+    unsigned prog, vers;
+    bool udp;
+    bool ipv6;
+    unsigned port;
+    std::string owner;
+    std::string addr;
+};
+
+class portmap_service_t
+{
+public:
+    std::set<portmap_id_t> reg_ports;
+    std::vector<rpc_service_proc_t> proc_table;
+    portmap_service_t();
+};
+
+inline bool operator < (const portmap_id_t &a, const portmap_id_t &b)
+{
+    return a.prog < b.prog || a.prog == b.prog && a.vers < b.vers ||
+        a.prog == b.prog && a.vers == b.vers && a.udp < b.udp ||
+        a.prog == b.prog && a.vers == b.vers && a.udp == b.udp && a.ipv6 < b.ipv6;
+}
+
+std::string sha256(const std::string & str);

src/nfs/nfs_proxy.h

@@ -0,0 +1,133 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Simplified NFS proxy - main entrypoint header
+
+#pragma once
+
+#include "cluster_client.h"
+#include "epoll_manager.h"
+#include "nfs_portmap.h"
+#include "proto/xdr_impl.h"
+#include "vitastor_kv.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 kv_fs_state_t;
+struct block_fs_state_t;
+
+class nfs_proxy_t
+{
+public:
+    std::string bind_address;
+    uint64_t fsid = 1;
+    uint64_t server_id = 0;
+    // FIXME: Maybe allow to create files in different pools?
+    std::string default_pool;
+    std::string export_root;
+    bool portmap_enabled;
+    unsigned nfs_port;
+    int trace = 0;
+    std::string logfile = "/dev/null";
+    std::string pidfile;
+    bool exit_on_umount = false;
+    std::string mountpoint;
+    std::string mountopts;
+    std::string fsname;
+
+    int active_connections = 0;
+    bool finished = false;
+    int listening_port = 0;
+    pool_id_t default_pool_id = 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;
+
+    // inode ID => statistics
+    std::map<inode_t, json11::Json> inode_stats;
+    // pool ID => statistics
+    std::map<pool_id_t, json11::Json> pool_stats;
+
+    ~nfs_proxy_t();
+
+    static json11::Json::object parse_args(int narg, const char *args[]);
+    void run(json11::Json cfg);
+    void watch_stats();
+    void parse_stats(etcd_kv_t & kv);
+    void check_default_pool();
+    void do_accept(int listen_fd);
+    void daemonize();
+    void write_pid();
+    void mount_fs();
+    void check_already_mounted();
+    void check_exit();
+};
+
+struct rpc_cur_buffer_t
+{
+    uint8_t *buf;
+    unsigned size;
+    unsigned read_pos;
+    unsigned parsed_pos;
+    int refs;
+};
+
+struct rpc_used_buffer_t
+{
+    unsigned size;
+    int refs;
+};
+
+struct rpc_free_buffer_t
+{
+    uint8_t *buf;
+    unsigned size;
+};
+
+class nfs_client_t
+{
+public:
+    nfs_proxy_t *parent = NULL;
+    int nfs_fd;
+    int epoll_events = 0;
+    int refs = 0;
+    bool stopped = false;
+    std::set<rpc_service_proc_t> proc_table;
+
+    // Read state
+    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;
+
+    iovec read_iov;
+    msghdr read_msg = { 0 };
+
+    // Write state
+    msghdr write_msg = { 0 };
+    std::vector<iovec> send_list, next_send_list;
+    std::vector<rpc_op_t*> outbox, next_outbox;
+
+    void select_read_buffer(unsigned wanted_size);
+    void submit_read(unsigned wanted_size);
+    void handle_read(int result);
+    void submit_send();
+    void handle_send(int result);
+    int handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg_len);
+
+    bool deref();
+    void stop();
+};

src/nfs/proto/xdr_impl_inline.h

@@ -37,7 +37,7 @@
 #include <endian.h>
 #include <vector>
 
-#include "../malloc_or_die.h"
+#include "malloc_or_die.h"
 
 #define FALSE 0
 #define TRUE 1