Return error on failed shrink

.gitea/workflows/buildenv.Dockerfile

@@ -22,7 +22,7 @@ RUN apt-get update
 RUN apt-get -y install etcd qemu-system-x86 qemu-block-extra qemu-utils fio libasan5 \
     liburing1 liburing-dev libgoogle-perftools-dev devscripts libjerasure-dev cmake libibverbs-dev libisal-dev
 RUN apt-get -y build-dep fio qemu=`dpkg -s qemu-system-x86|grep ^Version:|awk '{print $2}'`
-RUN apt-get -y install jq lp-solve sudo
+RUN apt-get -y install jq lp-solve sudo nfs-common
 RUN apt-get --download-only source fio qemu=`dpkg -s qemu-system-x86|grep ^Version:|awk '{print $2}'`
 
 RUN set -ex; \

.gitea/workflows/test.yml

@@ -856,3 +856,21 @@ jobs:
           echo ""
         done
 
+  test_nfs:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_nfs.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+

docs/usage/cli.en.md

@@ -267,6 +267,7 @@ Optional parameters:
 | `--immediate_commit none`      | Put pool only on OSDs with this or larger immediate_commit (none < small < all) |
 | `--primary_affinity_tags tags` | Prefer to put primary copies on OSDs with all specified tags               |
 | `--scrub_interval <time>`      | Enable regular scrubbing for this pool. Format: number + unit s/m/h/d/M/y  |
+| `--no_inode_stats 1`           | Disable per-inode statistics for this pool (use for VitastorFS pools)      |
 | `--pg_stripe_size <number>`    | Increase object grouping stripe                                            |
 | `--max_osd_combinations 10000` | Maximum number of random combinations for LP solver input                  |
 | `--wait`                       | Wait for the new pool to come online                                       |
@@ -288,7 +289,7 @@ Modify an existing pool. Modifiable parameters:
 
 ```
 [-s|--pg_size <number>] [--pg_minsize <number>] [-n|--pg_count <count>]
-[--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>]
+[--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>] [--no_inode_stats 0|1]
 [--max_osd_combinations <number>] [--primary_affinity_tags <tags>] [--scrub_interval <time>]
 ```
 

mon/mon.js

@@ -1737,8 +1737,11 @@ class Mon
                 for (const inode_num in this.state.osd.space[osd_num][pool_id])
                 {
                     const u = BigInt(this.state.osd.space[osd_num][pool_id][inode_num]||0);
-                    inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();
-                    inode_stats[pool_id][inode_num].raw_used += u;
+                    if (inode_num)
+                    {
+                        inode_stats[pool_id][inode_num] = inode_stats[pool_id][inode_num] || inode_stub();
+                        inode_stats[pool_id][inode_num].raw_used += u;
+                    }
                     this.state.pool.stats[pool_id].used_raw_tb += u;
                 }
             }

src/CMakeLists.txt

@@ -185,29 +185,48 @@ target_link_libraries(vitastor-nbd
 	vitastor_client
 )
 
-# vitastor-kv
-add_executable(vitastor-kv
-	kv_cli.cpp
+# libvitastor_kv.so
+add_library(vitastor_kv SHARED
 	kv_db.cpp
 	kv_db.h
 )
-target_link_libraries(vitastor-kv
+target_link_libraries(vitastor_kv
 	vitastor_client
 )
+set_target_properties(vitastor_kv PROPERTIES VERSION ${VERSION} SOVERSION 0)
+
+# vitastor-kv
+add_executable(vitastor-kv
+	kv_cli.cpp
+)
+target_link_libraries(vitastor-kv
+	vitastor_kv
+)
 
 add_executable(vitastor-kv-stress
 	kv_stress.cpp
-	kv_db.cpp
-	kv_db.h
 )
 target_link_libraries(vitastor-kv-stress
-	vitastor_client
+	vitastor_kv
 )
 
 # vitastor-nfs
 add_executable(vitastor-nfs
 	nfs_proxy.cpp
-	nfs_conn.cpp
+	nfs_block.cpp
+	nfs_kv.cpp
+	nfs_kv_create.cpp
+	nfs_kv_getattr.cpp
+	nfs_kv_link.cpp
+	nfs_kv_lookup.cpp
+	nfs_kv_read.cpp
+	nfs_kv_readdir.cpp
+	nfs_kv_remove.cpp
+	nfs_kv_rename.cpp
+	nfs_kv_setattr.cpp
+	nfs_kv_write.cpp
+	nfs_fsstat.cpp
+	nfs_mount.cpp
 	nfs_portmap.cpp
 	sha256.c
 	nfs/xdr_impl.cpp
@@ -217,6 +236,7 @@ add_executable(vitastor-nfs
 )
 target_link_libraries(vitastor-nfs
 	vitastor_client
+	vitastor_kv
 )
 
 # vitastor-cli

src/blockstore.cpp

@@ -82,3 +82,8 @@ uint32_t blockstore_t::get_bitmap_granularity()
 {
     return impl->get_bitmap_granularity();
 }
+
+void blockstore_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
+{
+    impl->set_no_inode_stats(pool_ids);
+}

src/blockstore.h

@@ -216,6 +216,9 @@ public:
     // Get per-inode space usage statistics
     std::map<uint64_t, uint64_t> & get_inode_space_stats();
 
+    // Set per-pool no_inode_stats
+    void set_no_inode_stats(const std::vector<uint64_t> & pool_ids);
+
     // Print diagnostics to stdout
     void dump_diagnostics();
 

src/blockstore_impl.cpp

@@ -733,3 +733,86 @@ void blockstore_impl_t::disk_error_abort(const char *op, int retval, int expecte
     fprintf(stderr, "Disk %s failed: result is %d, expected %d. Can't continue, sorry :-(\n", op, retval, expected);
     exit(1);
 }
+
+void blockstore_impl_t::set_no_inode_stats(const std::vector<uint64_t> & pool_ids)
+{
+    for (auto & np: no_inode_stats)
+    {
+        np.second = 2;
+    }
+    for (auto pool_id: pool_ids)
+    {
+        if (!no_inode_stats[pool_id])
+            recalc_inode_space_stats(pool_id, false);
+        no_inode_stats[pool_id] = 1;
+    }
+    for (auto np_it = no_inode_stats.begin(); np_it != no_inode_stats.end(); )
+    {
+        if (np_it->second == 2)
+        {
+            recalc_inode_space_stats(np_it->first, true);
+            no_inode_stats.erase(np_it++);
+        }
+        else
+            np_it++;
+    }
+}
+
+void blockstore_impl_t::recalc_inode_space_stats(uint64_t pool_id, bool per_inode)
+{
+    auto sp_begin = inode_space_stats.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    auto sp_end = inode_space_stats.lower_bound(((pool_id+1) << (64-POOL_ID_BITS)));
+    inode_space_stats.erase(sp_begin, sp_end);
+    auto sh_it = clean_db_shards.lower_bound((pool_id << (64-POOL_ID_BITS)));
+    while (sh_it != clean_db_shards.end() &&
+        (sh_it->first >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        for (auto & pair: sh_it->second)
+        {
+            uint64_t space_id = per_inode ? pair.first.inode : (pool_id << (64-POOL_ID_BITS));
+            inode_space_stats[space_id] += dsk.data_block_size;
+        }
+        sh_it++;
+    }
+    object_id last_oid = {};
+    bool last_exists = false;
+    auto dirty_it = dirty_db.lower_bound((obj_ver_id){ .oid = { .inode = (pool_id << (64-POOL_ID_BITS)) } });
+    while (dirty_it != dirty_db.end() && (dirty_it->first.oid.inode >> (64-POOL_ID_BITS)) == pool_id)
+    {
+        if (IS_STABLE(dirty_it->second.state) && (IS_BIG_WRITE(dirty_it->second.state) || IS_DELETE(dirty_it->second.state)))
+        {
+            bool exists = false;
+            if (last_oid == dirty_it->first.oid)
+            {
+                exists = last_exists;
+            }
+            else
+            {
+                auto & clean_db = clean_db_shard(dirty_it->first.oid);
+                auto clean_it = clean_db.find(dirty_it->first.oid);
+                exists = clean_it != clean_db.end();
+            }
+            uint64_t space_id = per_inode ? dirty_it->first.oid.inode : (pool_id << (64-POOL_ID_BITS));
+            if (IS_BIG_WRITE(dirty_it->second.state))
+            {
+                if (!exists)
+                    inode_space_stats[space_id] += dsk.data_block_size;
+                last_exists = true;
+            }
+            else
+            {
+                if (exists)
+                {
+                    auto & sp = inode_space_stats[space_id];
+                    if (sp > dsk.data_block_size)
+                        sp -= dsk.data_block_size;
+                    else
+                        inode_space_stats.erase(space_id);
+                }
+                last_exists = false;
+            }
+            last_oid = dirty_it->first.oid;
+        }
+        dirty_it++;
+    }
+}

src/blockstore_impl.h

@@ -272,6 +272,7 @@ class blockstore_impl_t
 
     std::map<pool_id_t, pool_shard_settings_t> clean_db_settings;
     std::map<pool_pg_id_t, blockstore_clean_db_t> clean_db_shards;
+    std::map<uint64_t, int> no_inode_stats;
     uint8_t *clean_bitmaps = NULL;
     blockstore_dirty_db_t dirty_db;
     std::vector<blockstore_op_t*> submit_queue;
@@ -318,6 +319,7 @@ class blockstore_impl_t
 
     blockstore_clean_db_t& clean_db_shard(object_id oid);
     void reshard_clean_db(pool_id_t pool_id, uint32_t pg_count, uint32_t pg_stripe_size);
+    void recalc_inode_space_stats(uint64_t pool_id, bool per_inode);
 
     // Journaling
     void prepare_journal_sector_write(int sector, blockstore_op_t *op);
@@ -428,6 +430,9 @@ public:
     // Space usage statistics
     std::map<uint64_t, uint64_t> inode_space_stats;
 
+    // Set per-pool no_inode_stats
+    void set_no_inode_stats(const std::vector<uint64_t> & pool_ids);
+
     // Print diagnostics to stdout
     void dump_diagnostics();
 

src/blockstore_stable.cpp

@@ -487,18 +487,24 @@ void blockstore_impl_t::mark_stable(obj_ver_id v, bool forget_dirty)
                     }
                     if (!exists)
                     {
-                        inode_space_stats[dirty_it->first.oid.inode] += dsk.data_block_size;
+                        uint64_t space_id = dirty_it->first.oid.inode;
+                        if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
+                            space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
+                        inode_space_stats[space_id] += dsk.data_block_size;
                         used_blocks++;
                     }
                     big_to_flush++;
                 }
                 else if (IS_DELETE(dirty_it->second.state))
                 {
-                    auto & sp = inode_space_stats[dirty_it->first.oid.inode];
+                    uint64_t space_id = dirty_it->first.oid.inode;
+                    if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
+                        space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
+                    auto & sp = inode_space_stats[space_id];
                     if (sp > dsk.data_block_size)
                         sp -= dsk.data_block_size;
                     else
-                        inode_space_stats.erase(dirty_it->first.oid.inode);
+                        inode_space_stats.erase(space_id);
                     used_blocks--;
                     big_to_flush++;
                 }

src/cli.cpp

@@ -131,6 +131,7 @@ static const char* help_text =
     "    --immediate_commit none       Put pool only on OSDs with this or larger immediate_commit (none < small < all)\n"
     "    --primary_affinity_tags tags  Prefer to put primary copies on OSDs with all specified tags\n"
     "    --scrub_interval <time>       Enable regular scrubbing for this pool. Format: number + unit s/m/h/d/M/y\n"
+    "    --no_inode_stats 1            Disable per-inode statistics for this pool (use for VitastorFS pools)\n"
     "    --pg_stripe_size <number>     Increase object grouping stripe\n"
     "    --max_osd_combinations 10000  Maximum number of random combinations for LP solver input\n"
     "    --wait                        Wait for the new pool to come online\n"
@@ -142,7 +143,7 @@ static const char* help_text =
     "vitastor-cli modify-pool|pool-modify <id|name> [--name <new_name>] [PARAMETERS...]\n"
     "  Modify an existing pool. Modifiable parameters:\n"
     "    [-s|--pg_size <number>] [--pg_minsize <number>] [-n|--pg_count <count>]\n"
-    "    [--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>]\n"
+    "    [--failure_domain <level>] [--root_node <node>] [--osd_tags <tags>] [--no_inode_stats 0|1]\n"
     "    [--max_osd_combinations <number>] [--primary_affinity_tags <tags>] [--scrub_interval <time>]\n"
     "  Non-modifiable parameters (changing them WILL lead to data loss):\n"
     "    [--block_size <size>] [--bitmap_granularity <size>]\n"

src/cli_common.cpp

@@ -153,6 +153,7 @@ void cli_tool_t::loop_and_wait(std::function<bool(cli_result_t &)> loop_cb, std:
             ringloop->unregister_consumer(&looper->consumer);
             looper->loop_cb = NULL;
             looper->complete_cb(looper->result);
+            ringloop->submit();
             delete looper;
             return;
         }

src/cli_pool_cfg.cpp

@@ -81,6 +81,11 @@ std::string validate_pool_config(json11::Json::object & new_cfg, json11::Json ol
             }
             value = value.uint64_value();
         }
+        else if (key == "no_inode_stats" && value.bool_value())
+        {
+            // Leave true, remove false
+            value = true;
+        }
         else if (key == "name" || key == "scheme" || key == "immediate_commit" ||
             key == "failure_domain" || key == "root_node" || key == "scrub_interval")
         {
@@ -248,7 +253,7 @@ std::string validate_pool_config(json11::Json::object & new_cfg, json11::Json ol
     // immediate_commit
     if (!cfg["immediate_commit"].is_null() && !etcd_state_client_t::parse_immediate_commit(cfg["immediate_commit"].string_value()))
     {
-        return "immediate_commit must be one of \"all\", \"small\", or \"none\", but it is "+cfg["scrub_interval"].as_string();
+        return "immediate_commit must be one of \"all\", \"small\", or \"none\", but it is "+cfg["immediate_commit"].as_string();
     }
 
     // scrub_interval

src/cli_pool_ls.cpp

@@ -529,6 +529,8 @@ resume_3:
                     st["block_size_fmt"] = format_size(st["block_size"].uint64_value());
                 if (st["bitmap_granularity"].uint64_value())
                     st["bitmap_granularity_fmt"] = format_size(st["bitmap_granularity"].uint64_value());
+                if (st["no_inode_stats"].bool_value())
+                    st["inode_stats_fmt"] = "disabled";
             }
             // All pool parameters are only displayed in the "detailed" mode
             // because there's too many of them to show them in table
@@ -547,6 +549,7 @@ resume_3:
                 { "bitmap_granularity_fmt", "Bitmap granularity" },
                 { "immediate_commit", "Immediate commit" },
                 { "scrub_interval", "Scrub interval" },
+                { "inode_stats_fmt", "Per-inode stats" },
                 { "pg_stripe_size", "PG stripe size" },
                 { "max_osd_combinations", "Max OSD combinations" },
                 { "total_fmt", "Total" },

src/epoll_manager.cpp

@@ -101,7 +101,7 @@ void epoll_manager_t::handle_uring_event()
     my_uring_prep_poll_add(sqe, epoll_fd, POLLIN);
     data->callback = [this](ring_data_t *data)
     {
-        if (data->res < 0)
+        if (data->res < 0 && data->res != -ECANCELED)
         {
             throw std::runtime_error(std::string("epoll failed: ") + strerror(-data->res));
         }

src/etcd_state_client.cpp

@@ -863,6 +863,8 @@ void etcd_state_client_t::parse_state(const etcd_kv_t & kv)
             pc.scrub_interval = parse_time(pool_item.second["scrub_interval"].string_value());
             if (!pc.scrub_interval)
                 pc.scrub_interval = 0;
+            // Disable per-inode stats
+            pc.no_inode_stats = pool_item.second["no_inode_stats"].bool_value();
             // Immediate Commit Mode
             pc.immediate_commit = pool_item.second["immediate_commit"].is_string()
                 ? parse_immediate_commit(pool_item.second["immediate_commit"].string_value())

src/etcd_state_client.h

@@ -60,6 +60,7 @@ struct pool_config_t
     uint64_t pg_stripe_size;
     std::map<pg_num_t, pg_config_t> pg_config;
     uint64_t scrub_interval;
+    bool no_inode_stats;
 };
 
 struct inode_config_t

src/nbd_proxy.cpp

@@ -146,7 +146,7 @@ public:
             "    Note that nbd_timeout, nbd_max_devices and nbd_max_part options may also be specified\n"
             "    in /etc/vitastor/vitastor.conf or in other configuration file specified with --config_file.\n"
             "  --logfile /path/to/log/file.txt\n"
-            "    Wite log messages to the specified file instead of dropping them (in background mode)\n"
+            "    Write log messages to the specified file instead of dropping them (in background mode)\n"
             "    or printing them to the standard output (in foreground mode).\n"
             "  --dev_num N\n"
             "    Use the specified device /dev/nbdN instead of automatic selection.\n"
@@ -298,7 +298,7 @@ public:
                 }
             }
         }
-        if (cfg["logfile"].is_string())
+        if (cfg["logfile"].string_value() != "")
         {
             logfile = cfg["logfile"].string_value();
         }

src/nfs_block.cpp

@@ -1,23 +1,18 @@
 // Copyright (c) Vitaliy Filippov, 2019+
 // License: VNPL-1.1 (see README.md for details)
 //
-// NFS connection handler for NFS proxy
+// NFS proxy over Vitastor block images
 
 #include <sys/time.h>
 
 #include "str_util.h"
 
 #include "nfs_proxy.h"
-
+#include "nfs_common.h"
+#include "nfs_block.h"
 #include "nfs/nfs.h"
-
 #include "cli.h"
 
-#define TRUE 1
-#define FALSE 0
-
-#define MAX_REQUEST_SIZE 128*1024*1024
-
 static unsigned len_pad4(unsigned len)
 {
     return len + (len&3 ? 4-(len&3) : 0);
@@ -28,10 +23,10 @@ static std::string get_inode_name(nfs_client_t *self, diropargs3 & what)
     // Get name
     std::string dirhash = what.dir;
     std::string dir;
-    if (dirhash != "roothandle")
+    if (dirhash != NFS_ROOT_HANDLE)
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
             dir = dir_it->second;
         else
             return "";
@@ -42,24 +37,9 @@ static std::string get_inode_name(nfs_client_t *self, diropargs3 & what)
         : self->parent->name_prefix+name);
 }
 
-static nfsstat3 vitastor_nfs_map_err(int err)
-{
-    return (err == EINVAL ? NFS3ERR_INVAL
-        : (err == ENOENT ? NFS3ERR_NOENT
-        : (err == ENOSPC ? NFS3ERR_NOSPC
-        : (err == EEXIST ? NFS3ERR_EXIST
-        : (err == EIO ? NFS3ERR_IO : (err ? NFS3ERR_IO : NFS3_OK))))));
-}
-
-static int nfs3_null_proc(void *opaque, rpc_op_t *rop)
-{
-    rpc_queue_reply(rop);
-    return 0;
-}
-
 static fattr3 get_dir_attributes(nfs_client_t *self, std::string dir)
 {
-    auto & dinf = self->parent->dir_info.at(dir);
+    auto & dinf = self->parent->blockfs->dir_info.at(dir);
     return (fattr3){
         .type = NF3DIR,
         .mode = 0755,
@@ -108,7 +88,7 @@ static fattr3 get_file_attributes(nfs_client_t *self, inode_t inode_num)
     };
 }
 
-static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     GETATTR3args *args = (GETATTR3args*)rop->request;
@@ -116,12 +96,12 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     bool is_dir = false;
     std::string dirhash = args->object;
     std::string dir;
-    if (args->object == "roothandle")
+    if (args->object == NFS_ROOT_HANDLE)
         is_dir = true;
     else
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
         {
             is_dir = true;
             dir = dir_it->second;
@@ -140,8 +120,8 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     else
     {
         uint64_t inode_num = 0;
-        auto inode_num_it = self->parent->inode_by_hash.find(dirhash);
-        if (inode_num_it != self->parent->inode_by_hash.end())
+        auto inode_num_it = self->parent->blockfs->inode_by_hash.find(dirhash);
+        if (inode_num_it != self->parent->blockfs->inode_by_hash.end())
             inode_num = inode_num_it->second;
         auto inode_it = self->parent->cli->st_cli.inode_config.find(inode_num);
         if (inode_num && inode_it != self->parent->cli->st_cli.inode_config.end())
@@ -179,16 +159,16 @@ static int nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     SETATTR3args *args = (SETATTR3args*)rop->request;
     SETATTR3res *reply = (SETATTR3res*)rop->reply;
     std::string handle = args->object;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
-        if (handle == "roothandle" || self->parent->dir_by_hash.find(handle) != self->parent->dir_by_hash.end())
+        if (handle == NFS_ROOT_HANDLE || self->parent->blockfs->dir_by_hash.find(handle) != self->parent->blockfs->dir_by_hash.end())
         {
             if (args->new_attributes.size.set_it)
             {
@@ -228,7 +208,7 @@ static int nfs3_setattr_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     LOOKUP3args *args = (LOOKUP3args*)rop->request;
@@ -255,8 +235,8 @@ static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
                 return 0;
             }
         }
-        auto dir_it = self->parent->dir_info.find(full_name);
-        if (dir_it != self->parent->dir_info.end())
+        auto dir_it = self->parent->blockfs->dir_info.find(full_name);
+        if (dir_it != self->parent->blockfs->dir_info.end())
         {
             *reply = (LOOKUP3res){
                 .status = NFS3_OK,
@@ -277,7 +257,7 @@ static int nfs3_lookup_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_access_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_access_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     ACCESS3args *args = (ACCESS3args*)rop->request;
@@ -292,7 +272,7 @@ static int nfs3_access_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     //READLINK3args *args = (READLINK3args*)rop->request;
@@ -303,14 +283,14 @@ static int nfs3_readlink_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_read_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_read_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     READ3args *args = (READ3args*)rop->request;
     READ3res *reply = (READ3res*)rop->reply;
     std::string handle = args->file;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
         *reply = (READ3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -367,14 +347,14 @@ static int nfs3_read_proc(void *opaque, rpc_op_t *rop)
 
 static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode, uint64_t new_size, uint64_t offset, uint64_t count, void *buf);
 
-static int nfs3_write_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_write_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     WRITE3args *args = (WRITE3args*)rop->request;
     WRITE3res *reply = (WRITE3res*)rop->reply;
     std::string handle = args->file;
-    auto ino_it = self->parent->inode_by_hash.find(handle);
-    if (ino_it == self->parent->inode_by_hash.end())
+    auto ino_it = self->parent->blockfs->inode_by_hash.find(handle);
+    if (ino_it == self->parent->blockfs->inode_by_hash.end())
     {
         *reply = (WRITE3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -480,8 +460,8 @@ static void complete_extend_write(nfs_client_t *self, rpc_op_t *rop, inode_t ino
 
 static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size, int err)
 {
-    auto ext_it = self->extend_writes.lower_bound((extend_size_t){ .inode = inode, .new_size = 0 });
-    while (ext_it != self->extend_writes.end() &&
+    auto ext_it = self->parent->blockfs->extend_writes.lower_bound((extend_size_t){ .inode = inode, .new_size = 0 });
+    while (ext_it != self->parent->blockfs->extend_writes.end() &&
         ext_it->first.inode == inode &&
         ext_it->first.new_size <= new_size)
     {
@@ -490,7 +470,7 @@ static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t n
         {
             complete_extend_write(self, ext_it->second.rop, inode, ext_it->second.write_res < 0
                 ? ext_it->second.write_res : ext_it->second.resize_res);
-            self->extend_writes.erase(ext_it++);
+            self->parent->blockfs->extend_writes.erase(ext_it++);
         }
         else
             ext_it++;
@@ -500,7 +480,7 @@ static void complete_extend_inode(nfs_client_t *self, uint64_t inode, uint64_t n
 static void extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size)
 {
     // Send an extend request
-    auto & ext = self->extends[inode];
+    auto & ext = self->parent->blockfs->extends[inode];
     ext.cur_extend = new_size;
     auto inode_it = self->parent->cli->st_cli.inode_config.find(inode);
     if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
@@ -514,10 +494,10 @@ static void extend_inode(nfs_client_t *self, uint64_t inode, uint64_t new_size)
             { "force_size", true },
         }), [=](const cli_result_t & r)
         {
-            auto & ext = self->extends[inode];
+            auto & ext = self->parent->blockfs->extends[inode];
             if (r.err)
             {
-                fprintf(stderr, "Error extending inode %ju to %ju bytes: %s\n", inode, new_size, r.text.c_str());
+                fprintf(stderr, "Error extending inode %lu to %lu bytes: %s\n", inode, new_size, r.text.c_str());
             }
             if (r.err == EAGAIN || ext.next_extend > ext.cur_extend)
             {
@@ -548,7 +528,7 @@ static void nfs_do_write(nfs_client_t *self, std::multimap<extend_size_t, extend
     {
         auto inode = op->inode;
         int write_res = op->retval < 0 ? op->retval : (op->retval != op->len ? -ERANGE : 0);
-        if (ewr_it == self->extend_writes.end())
+        if (ewr_it == self->parent->blockfs->extend_writes.end())
         {
             complete_extend_write(self, rop, inode, write_res);
         }
@@ -558,7 +538,7 @@ static void nfs_do_write(nfs_client_t *self, std::multimap<extend_size_t, extend
             if (ewr_it->second.resize_res <= 0)
             {
                 complete_extend_write(self, rop, inode, write_res < 0 ? write_res : ewr_it->second.resize_res);
-                self->extend_writes.erase(ewr_it);
+                self->parent->blockfs->extend_writes.erase(ewr_it);
             }
         }
     };
@@ -572,7 +552,7 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
     if (inode_it != self->parent->cli->st_cli.inode_config.end() &&
         inode_it->second.size < new_size)
     {
-        auto ewr_it = self->extend_writes.emplace((extend_size_t){
+        auto ewr_it = self->parent->blockfs->extend_writes.emplace((extend_size_t){
             .inode = inode,
             .new_size = new_size,
         }, (extend_write_t){
@@ -580,7 +560,7 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
             .resize_res = 1,
             .write_res = 1,
         });
-        auto & ext = self->extends[inode];
+        auto & ext = self->parent->blockfs->extends[inode];
         if (ext.cur_extend > 0)
         {
             // Already resizing, just wait
@@ -595,11 +575,11 @@ static void nfs_resize_write(nfs_client_t *self, rpc_op_t *rop, uint64_t inode,
     }
     else
     {
-        nfs_do_write(self, self->extend_writes.end(), rop, inode, offset, count, buf);
+        nfs_do_write(self, self->parent->blockfs->extend_writes.end(), rop, inode, offset, count, buf);
     }
 }
 
-static int nfs3_create_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_create_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     CREATE3args *args = (CREATE3args*)rop->request;
@@ -650,7 +630,7 @@ static int nfs3_create_proc(void *opaque, rpc_op_t *rop)
     return 1;
 }
 
-static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     MKDIR3args *args = (MKDIR3args*)rop->request;
@@ -669,19 +649,19 @@ static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    auto dir_id_it = self->parent->dir_info.find(full_name);
-    if (dir_id_it != self->parent->dir_info.end())
+    auto dir_id_it = self->parent->blockfs->dir_info.find(full_name);
+    if (dir_id_it != self->parent->blockfs->dir_info.end())
     {
         *reply = (MKDIR3res){ .status = NFS3ERR_EXIST };
         rpc_queue_reply(rop);
         return 0;
     }
     // FIXME: Persist empty directories in some etcd keys, like /vitastor/dir/...
-    self->parent->dir_info[full_name] = (nfs_dir_t){
-        .id = self->parent->next_dir_id++,
+    self->parent->blockfs->dir_info[full_name] = (nfs_dir_t){
+        .id = self->parent->blockfs->next_dir_id++,
         .mod_rev = 0,
     };
-    self->parent->dir_by_hash["S"+base64_encode(sha256(full_name))] = full_name;
+    self->parent->blockfs->dir_by_hash["S"+base64_encode(sha256(full_name))] = full_name;
     *reply = (MKDIR3res){
         .status = NFS3_OK,
         .resok = (MKDIR3resok){
@@ -700,7 +680,7 @@ static int nfs3_mkdir_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
 {
 //    nfs_client_t *self = (nfs_client_t*)opaque;
 //    SYMLINK3args *args = (SYMLINK3args*)rop->request;
@@ -711,7 +691,7 @@ static int nfs3_symlink_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
 {
 //    nfs_client_t *self = (nfs_client_t*)opaque;
 //    MKNOD3args *args = (MKNOD3args*)rop->request;
@@ -722,7 +702,7 @@ static int nfs3_mknod_proc(void *opaque, rpc_op_t *rop)
     return 0;
 }
 
-static int nfs3_remove_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_remove_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     REMOVE3res *reply = (REMOVE3res*)rop->reply;
@@ -752,7 +732,7 @@ static int nfs3_remove_proc(void *opaque, rpc_op_t *rop)
     return 1;
 }
 
-static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     RMDIR3args *args = (RMDIR3args*)rop->request;
@@ -764,8 +744,8 @@ static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    auto dir_it = self->parent->dir_info.find(full_name);
-    if (dir_it == self->parent->dir_info.end())
+    auto dir_it = self->parent->blockfs->dir_info.find(full_name);
+    if (dir_it == self->parent->blockfs->dir_info.end())
     {
         *reply = (RMDIR3res){ .status = NFS3ERR_NOENT };
         rpc_queue_reply(rop);
@@ -781,8 +761,8 @@ static int nfs3_rmdir_proc(void *opaque, rpc_op_t *rop)
             return 0;
         }
     }
-    self->parent->dir_by_hash.erase("S"+base64_encode(sha256(full_name)));
-    self->parent->dir_info.erase(dir_it);
+    self->parent->blockfs->dir_by_hash.erase("S"+base64_encode(sha256(full_name)));
+    self->parent->blockfs->dir_info.erase(dir_it);
     *reply = (RMDIR3res){ .status = NFS3_OK };
     rpc_queue_reply(rop);
     return 0;
@@ -811,12 +791,12 @@ static int continue_dir_rename(nfs_dir_rename_state *rename_st)
     if (!rename_st->items.size())
     {
         // old dir
-        auto old_info = self->parent->dir_info.at(rename_st->old_name);
-        self->parent->dir_info.erase(rename_st->old_name);
-        self->parent->dir_by_hash.erase("S"+base64_encode(sha256(rename_st->old_name)));
+        auto old_info = self->parent->blockfs->dir_info.at(rename_st->old_name);
+        self->parent->blockfs->dir_info.erase(rename_st->old_name);
+        self->parent->blockfs->dir_by_hash.erase("S"+base64_encode(sha256(rename_st->old_name)));
         // new dir
-        self->parent->dir_info[rename_st->new_name] = old_info;
-        self->parent->dir_by_hash["S"+base64_encode(sha256(rename_st->new_name))] = rename_st->new_name;
+        self->parent->blockfs->dir_info[rename_st->new_name] = old_info;
+        self->parent->blockfs->dir_by_hash["S"+base64_encode(sha256(rename_st->new_name))] = rename_st->new_name;
         RENAME3res *reply = (RENAME3res*)rename_st->rop->reply;
         *reply = (RENAME3res){
             .status = NFS3_OK,
@@ -853,7 +833,7 @@ static int continue_dir_rename(nfs_dir_rename_state *rename_st)
 
 static void nfs_do_rename(nfs_client_t *self, rpc_op_t *rop, std::string old_name, std::string new_name);
 
-static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_rename_proc(void *opaque, rpc_op_t *rop)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     RENAME3args *args = (RENAME3args*)rop->request;
@@ -866,8 +846,8 @@ static int nfs3_rename_proc(void *opaque, rpc_op_t *rop)
         rpc_queue_reply(rop);
         return 0;
     }
-    bool old_is_dir = self->parent->dir_info.find(old_name) != self->parent->dir_info.end();
-    bool new_is_dir = self->parent->dir_info.find(new_name) != self->parent->dir_info.end();
+    bool old_is_dir = self->parent->blockfs->dir_info.find(old_name) != self->parent->blockfs->dir_info.end();
+    bool new_is_dir = self->parent->blockfs->dir_info.find(new_name) != self->parent->blockfs->dir_info.end();
     bool old_is_file = false, new_is_file = false;
     for (auto & ic: self->parent->cli->st_cli.inode_config)
     {
@@ -948,7 +928,7 @@ static void nfs_do_rename(nfs_client_t *self, rpc_op_t *rop, std::string old_nam
     });
 }
 
-static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_link_proc(void *opaque, rpc_op_t *rop)
 {
     //nfs_client_t *self = (nfs_client_t*)opaque;
     //LINK3args *args = (LINK3args*)rop->request;
@@ -962,7 +942,7 @@ static int nfs3_link_proc(void *opaque, rpc_op_t *rop)
 static void fill_dir_entry(nfs_client_t *self, rpc_op_t *rop,
     std::map<std::string, nfs_dir_t>::iterator dir_id_it, struct entryplus3 *entry, bool is_plus)
 {
-    if (dir_id_it == self->parent->dir_info.end())
+    if (dir_id_it == self->parent->blockfs->dir_info.end())
     {
         return;
     }
@@ -980,7 +960,7 @@ static void fill_dir_entry(nfs_client_t *self, rpc_op_t *rop,
     }
 }
 
-static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
+static void block_nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
 {
     nfs_client_t *self = (nfs_client_t*)opaque;
     READDIRPLUS3args plus_args;
@@ -999,10 +979,10 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     }
     std::string dirhash = args->dir;
     std::string dir;
-    if (dirhash != "roothandle")
+    if (dirhash != NFS_ROOT_HANDLE)
     {
-        auto dir_it = self->parent->dir_by_hash.find(dirhash);
-        if (dir_it != self->parent->dir_by_hash.end())
+        auto dir_it = self->parent->blockfs->dir_by_hash.find(dirhash);
+        if (dir_it != self->parent->blockfs->dir_by_hash.end())
             dir = dir_it->second;
     }
     std::string prefix = dir.size() ? dir+"/" : self->parent->name_prefix;
@@ -1043,12 +1023,12 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
         }
         else
         {
-            // skip directories, they will be added from dir_info
+            // skip directories, they will be added from blockfs->dir_info
         }
     }
-    // Add directories from dir_info
-    for (auto dir_id_it = self->parent->dir_info.lower_bound(prefix);
-        dir_id_it != self->parent->dir_info.end(); dir_id_it++)
+    // Add directories from blockfs->dir_info
+    for (auto dir_id_it = self->parent->blockfs->dir_info.lower_bound(prefix);
+        dir_id_it != self->parent->blockfs->dir_info.end(); dir_id_it++)
     {
         if (prefix != "" && dir_id_it->first.substr(0, prefix.size()) != prefix)
             break;
@@ -1061,12 +1041,12 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     }
     // Add . and ..
     {
-        auto dir_id_it = self->parent->dir_info.find(dir);
+        auto dir_id_it = self->parent->blockfs->dir_info.find(dir);
         fill_dir_entry(self, rop, dir_id_it, &entries["."], is_plus);
         auto sl = dir.rfind("/");
         if (sl != std::string::npos)
         {
-            auto dir_id_it = self->parent->dir_info.find(dir.substr(0, sl));
+            auto dir_id_it = self->parent->blockfs->dir_info.find(dir.substr(0, sl));
             fill_dir_entry(self, rop, dir_id_it, &entries[".."], is_plus);
         }
     }
@@ -1147,7 +1127,7 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     {
         READDIRPLUS3res *reply = (READDIRPLUS3res*)rop->reply;
         *reply = { .status = NFS3_OK };
-        *(uint64_t*)(reply->resok.cookieverf) = self->parent->dir_info.at(dir).mod_rev;
+        *(uint64_t*)(reply->resok.cookieverf) = self->parent->blockfs->dir_info.at(dir).mod_rev;
         reply->resok.reply.entries = entries.size() ? &entries.begin()->second : NULL;
         reply->resok.reply.eof = eof;
     }
@@ -1155,250 +1135,123 @@ static void nfs3_readdir_common(void *opaque, rpc_op_t *rop, bool is_plus)
     {
         READDIR3res *reply = (READDIR3res*)rop->reply;
         *reply = { .status = NFS3_OK };
-        *(uint64_t*)(reply->resok.cookieverf) = self->parent->dir_info.at(dir).mod_rev;
+        *(uint64_t*)(reply->resok.cookieverf) = self->parent->blockfs->dir_info.at(dir).mod_rev;
         reply->resok.reply.entries = entries.size() ? (entry3*)&entries.begin()->second : NULL;
         reply->resok.reply.eof = eof;
     }
     rpc_queue_reply(rop);
 }
 
-static int nfs3_readdir_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readdir_proc(void *opaque, rpc_op_t *rop)
 {
-    nfs3_readdir_common(opaque, rop, false);
+    block_nfs3_readdir_common(opaque, rop, false);
     return 0;
 }
 
-static int nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop)
+static int block_nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop)
 {
-    nfs3_readdir_common(opaque, rop, true);
+    block_nfs3_readdir_common(opaque, rop, true);
     return 0;
 }
 
-// Get file system statistics
-static int nfs3_fsstat_proc(void *opaque, rpc_op_t *rop)
+void block_fs_state_t::init(nfs_proxy_t *proxy)
 {
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    //FSSTAT3args *args = (FSSTAT3args*)rop->request;
-    FSSTAT3res *reply = (FSSTAT3res*)rop->reply;
-    uint64_t tbytes = 0, fbytes = 0;
-    auto pst_it = self->parent->pool_stats.find(self->parent->default_pool_id);
-    if (pst_it != self->parent->pool_stats.end())
-    {
-        auto ttb = pst_it->second["total_raw_tb"].number_value();
-        auto ftb = (pst_it->second["total_raw_tb"].number_value() - pst_it->second["used_raw_tb"].number_value());
-        tbytes = ttb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
-        fbytes = ftb / pst_it->second["raw_to_usable"].number_value() * ((uint64_t)2<<40);
-    }
-    *reply = (FSSTAT3res){
-        .status = NFS3_OK,
-        .resok = (FSSTAT3resok){
-            .obj_attributes = {
-                .attributes_follow = 1,
-                .attributes = get_dir_attributes(self, ""),
-            },
-            .tbytes = tbytes, // total bytes
-            .fbytes = fbytes, // free bytes
-            .abytes = fbytes, // available bytes
-            .tfiles = (size3)(1 << 31), // maximum total files
-            .ffiles = (size3)(1 << 31), // free files
-            .afiles = (size3)(1 << 31), // available files
-            .invarsec = 0,
-        },
+    // We need inode name hashes for NFS handles to remain stateless and <= 64 bytes long
+    dir_info[""] = (nfs_dir_t){
+        .id = 1,
+        .mod_rev = 0,
     };
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_fsinfo_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    FSINFO3args *args = (FSINFO3args*)rop->request;
-    FSINFO3res *reply = (FSINFO3res*)rop->reply;
-    if (args->fsroot != "roothandle")
+    clock_gettime(CLOCK_REALTIME, &dir_info[""].mtime);
+    assert(proxy->cli->st_cli.on_inode_change_hook == NULL);
+    proxy->cli->st_cli.on_inode_change_hook = [this, proxy](inode_t changed_inode, bool removed)
     {
-        // Example error
-        *reply = (FSINFO3res){ .status = NFS3ERR_INVAL };
-    }
-    else
-    {
-        // Fill info
-        *reply = (FSINFO3res){
-            .status = NFS3_OK,
-            .resok = (FSINFO3resok){
-                .obj_attributes = {
-                    .attributes_follow = 1,
-                    .attributes = get_dir_attributes(self, ""),
-                },
-                .rtmax = 128*1024*1024,
-                .rtpref = 128*1024*1024,
-                .rtmult = 4096,
-                .wtmax = 128*1024*1024,
-                .wtpref = 128*1024*1024,
-                .wtmult = 4096,
-                .dtpref = 128,
-                .maxfilesize = 0x7fffffffffffffff,
-                .time_delta = {
-                    .seconds = 1,
-                    .nseconds = 0,
-                },
-                .properties = FSF3_SYMLINK | FSF3_HOMOGENEOUS,
-            },
-        };
-    }
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_pathconf_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    PATHCONF3args *args = (PATHCONF3args*)rop->request;
-    PATHCONF3res *reply = (PATHCONF3res*)rop->reply;
-    if (args->object != "roothandle")
-    {
-        // Example error
-        *reply = (PATHCONF3res){ .status = NFS3ERR_INVAL };
-    }
-    else
-    {
-        // Fill info
-        bool_t x = FALSE;
-        *reply = (PATHCONF3res){
-            .status = NFS3_OK,
-            .resok = (PATHCONF3resok){
-                .obj_attributes = {
-                    // Without at least one reference to a non-constant value (local variable or something else),
-                    // with gcc 8 we get "internal compiler error: side-effects element in no-side-effects CONSTRUCTOR" here
-                    // FIXME: get rid of this after raising compiler requirement
-                    .attributes_follow = x,
-                },
-                .linkmax = 0,
-                .name_max = 255,
-                .no_trunc = TRUE,
-                .chown_restricted = FALSE,
-                .case_insensitive = FALSE,
-                .case_preserving = TRUE,
-            },
-        };
-    }
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int nfs3_commit_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    //COMMIT3args *args = (COMMIT3args*)rop->request;
-    cluster_op_t *op = new cluster_op_t;
-    // fsync. we don't know how to fsync a single inode, so just fsync everything
-    op->opcode = OSD_OP_SYNC;
-    op->callback = [self, rop](cluster_op_t *op)
-    {
-        COMMIT3res *reply = (COMMIT3res*)rop->reply;
-        *reply = (COMMIT3res){ .status = vitastor_nfs_map_err(op->retval) };
-        *(uint64_t*)reply->resok.verf = self->parent->server_id;
-        rpc_queue_reply(rop);
+        auto inode_cfg_it = proxy->cli->st_cli.inode_config.find(changed_inode);
+        if (inode_cfg_it == proxy->cli->st_cli.inode_config.end())
+        {
+            return;
+        }
+        auto & inode_cfg = inode_cfg_it->second;
+        std::string full_name = inode_cfg.name;
+        if (proxy->name_prefix != "" && full_name.substr(0, proxy->name_prefix.size()) != proxy->name_prefix)
+        {
+            return;
+        }
+        // Calculate directory modification time and revision (used as "cookie verifier")
+        timespec now;
+        clock_gettime(CLOCK_REALTIME, &now);
+        dir_info[""].mod_rev = dir_info[""].mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dir_info[""].mod_rev;
+        dir_info[""].mtime = now;
+        int pos = full_name.find('/', proxy->name_prefix.size());
+        while (pos >= 0)
+        {
+            std::string dir = full_name.substr(0, pos);
+            auto & dinf = dir_info[dir];
+            if (!dinf.id)
+                dinf.id = next_dir_id++;
+            dinf.mod_rev = dinf.mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dinf.mod_rev;
+            dinf.mtime = now;
+            dir_by_hash["S"+base64_encode(sha256(dir))] = dir;
+            pos = full_name.find('/', pos+1);
+        }
+        // Alter inode_by_hash
+        if (removed)
+        {
+            auto ino_it = hash_by_inode.find(changed_inode);
+            if (ino_it != hash_by_inode.end())
+            {
+                inode_by_hash.erase(ino_it->second);
+                hash_by_inode.erase(ino_it);
+            }
+        }
+        else
+        {
+            std::string hash = "S"+base64_encode(sha256(full_name));
+            auto hbi_it = hash_by_inode.find(changed_inode);
+            if (hbi_it != hash_by_inode.end() && hbi_it->second != hash)
+            {
+                // inode had a different name, remove old hash=>inode pointer
+                inode_by_hash.erase(hbi_it->second);
+            }
+            inode_by_hash[hash] = changed_inode;
+            hash_by_inode[changed_inode] = hash;
+        }
     };
-    self->parent->cli->execute(op);
-    return 1;
 }
 
-static int mount3_mnt_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    //nfs_dirpath *args = (nfs_dirpath*)rop->request;
-    nfs_mountres3 *reply = (nfs_mountres3*)rop->reply;
-    u_int flavor = RPC_AUTH_NONE;
-    reply->fhs_status = MNT3_OK;
-    reply->mountinfo.fhandle = xdr_copy_string(rop->xdrs, "roothandle");
-    reply->mountinfo.auth_flavors.auth_flavors_len = 1;
-    reply->mountinfo.auth_flavors.auth_flavors_val = (u_int*)xdr_copy_string(rop->xdrs, (char*)&flavor, sizeof(u_int)).data;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_dump_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    nfs_mountlist *reply = (nfs_mountlist*)rop->reply;
-    *reply = (struct nfs_mountbody*)malloc_or_die(sizeof(struct nfs_mountbody));
-    xdr_add_malloc(rop->xdrs, *reply);
-    (*reply)->ml_hostname = xdr_copy_string(rop->xdrs, "127.0.0.1");
-    (*reply)->ml_directory = xdr_copy_string(rop->xdrs, self->parent->export_root);
-    (*reply)->ml_next = NULL;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_umnt_proc(void *opaque, rpc_op_t *rop)
-{
-    //nfs_client_t *self = (nfs_client_t*)opaque;
-    //nfs_dirpath *arg = (nfs_dirpath*)rop->request;
-    // do nothing
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_umntall_proc(void *opaque, rpc_op_t *rop)
-{
-    // do nothing
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-static int mount3_export_proc(void *opaque, rpc_op_t *rop)
-{
-    nfs_client_t *self = (nfs_client_t*)opaque;
-    nfs_exports *reply = (nfs_exports*)rop->reply;
-    *reply = (struct nfs_exportnode*)calloc_or_die(1, sizeof(struct nfs_exportnode) + sizeof(struct nfs_groupnode));
-    xdr_add_malloc(rop->xdrs, *reply);
-    (*reply)->ex_dir = xdr_copy_string(rop->xdrs, self->parent->export_root);
-    (*reply)->ex_groups = (struct nfs_groupnode*)(reply+1);
-    (*reply)->ex_groups->gr_name = xdr_copy_string(rop->xdrs, "127.0.0.1");
-    (*reply)->ex_groups->gr_next = NULL;
-    (*reply)->ex_next = NULL;
-    rpc_queue_reply(rop);
-    return 0;
-}
-
-nfs_client_t::nfs_client_t()
+void nfs_block_procs(nfs_client_t *self)
 {
     struct rpc_service_proc_t pt[] = {
-        {NFS_PROGRAM, NFS_V3, NFS3_NULL,        nfs3_null_proc,        NULL,                            0,                        NULL,                           0,                       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,     nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,     nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,      nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,      nfs3_access_proc,      (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,    nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READ,        nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        this},
-        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,       nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,      nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,       nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,     nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,       nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,      nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,       nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       this},
-        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,      nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_LINK,        nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,     nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     this},
-        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS, nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), this},
-        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,      nfs3_fsstat_proc,      (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,      nfs3_fsinfo_proc,      (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      this},
-        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,    nfs3_pathconf_proc,    (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    this},
-        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,      nfs3_commit_proc,      (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,      NULL,                            0,                        NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,     (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,    NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,    (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc, NULL,                            0,                        NULL,                         0,                         this},
-        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,  NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       this},
+        {NFS_PROGRAM, NFS_V3, NFS3_NULL,          nfs3_null_proc,              NULL,                            0,                        NULL,                           0,                       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,       block_nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,       block_nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,        block_nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,        block_nfs3_access_proc,      (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,      block_nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READ,          block_nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,         block_nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,        block_nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,         block_nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,       block_nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,         block_nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,        block_nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,         block_nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,        block_nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LINK,          block_nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,       block_nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS,   block_nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,        nfs3_fsstat_proc,            (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,        nfs3_fsinfo_proc,            (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,      nfs3_pathconf_proc,          (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,        nfs3_commit_proc,            (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,              NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,             (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,            NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,            (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc,         NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,          NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       self},
     };
     for (int i = 0; i < sizeof(pt)/sizeof(pt[0]); i++)
     {
-        proc_table.insert(pt[i]);
+        self->proc_table.insert(pt[i]);
     }
 }
-
-nfs_client_t::~nfs_client_t()
-{
-}

src/nfs_block.h

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

src/nfs_common.h

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

src/nfs_fsstat.cpp

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

src/nfs_kv.cpp

@@ -0,0 +1,192 @@
+// 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());
+    // FIXME In theory we could store the binary structure itself instead of JSON
+    return (fattr3){
+        .type = (type == 0 ? NF3REG : (ftype3)type),
+        .mode = (attrs["mode"].is_null() ? (type == NF3DIR ? 0755 : 0644) : (uint32_t)mode),
+        .nlink = (nlink == 0 ? 1 : (uint32_t)nlink),
+        .uid = (uint32_t)attrs["uid"].uint64_value(),
+        .gid = (uint32_t)attrs["gid"].uint64_value(),
+        .size = (type == NF3DIR ? 4096 : attrs["size"].uint64_value()),
+        .used = (type == NF3DIR ? 4096 : attrs["alloc"].uint64_value()),
+        .rdev = (type == NF3BLK || type == NF3CHR
+            ? (specdata3){ (uint32_t)attrs["major"].uint64_value(), (uint32_t)attrs["minor"].uint64_value() }
+            : (specdata3){}),
+        .fsid = self->parent->fsid,
+        .fileid = ino,
+        .atime = atime,
+        .mtime = mtime,
+        .ctime = mtime,
+    };
+}
+
+std::string kv_direntry_key(uint64_t dir_ino, const std::string & filename)
+{
+    // encode as: d <length> <hex dir_ino> / <filename>
+    char key[24] = { 0 };
+    snprintf(key, sizeof(key), "d-%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[24] = { 0 };
+    snprintf(key, sizeof(key), "i-%jx", ino);
+    int n = strnlen(key, sizeof(key)-1) - 2;
+    if (n < 10)
+        key[1] = '0'+n;
+    else
+        key[1] = 'A'+(n-10);
+    return std::string(key, n+2);
+}
+
+std::string kv_fh(uint64_t ino)
+{
+    return "S"+std::string((char*)&ino, 8);
+}
+
+uint64_t kv_fh_inode(const std::string & fh)
+{
+    if (fh.size() == 1 && fh[0] == 'R')
+    {
+        return 1;
+    }
+    else if (fh.size() == 9 && fh[0] == 'S')
+    {
+        return *(uint64_t*)&fh[1];
+    }
+    else if (fh.size() > 17 && fh[0] == 'I')
+    {
+        return *(uint64_t*)&fh[fh.size()-8];
+    }
+    return 0;
+}
+
+bool kv_fh_valid(const std::string & fh)
+{
+    return fh == NFS_ROOT_HANDLE || fh.size() == 9 && fh[0] == 'S' || fh.size() > 17 && fh[0] == 'I';
+}
+
+void nfs_kv_procs(nfs_client_t *self)
+{
+    struct rpc_service_proc_t pt[] = {
+        {NFS_PROGRAM, NFS_V3, NFS3_NULL,        nfs3_null_proc,           NULL,                            0,                        NULL,                           0,                       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_GETATTR,     kv_nfs3_getattr_proc,     (xdrproc_t)xdr_GETATTR3args,     sizeof(GETATTR3args),     (xdrproc_t)xdr_GETATTR3res,     sizeof(GETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SETATTR,     kv_nfs3_setattr_proc,     (xdrproc_t)xdr_SETATTR3args,     sizeof(SETATTR3args),     (xdrproc_t)xdr_SETATTR3res,     sizeof(SETATTR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LOOKUP,      kv_nfs3_lookup_proc,      (xdrproc_t)xdr_LOOKUP3args,      sizeof(LOOKUP3args),      (xdrproc_t)xdr_LOOKUP3res,      sizeof(LOOKUP3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_ACCESS,      nfs3_access_proc,         (xdrproc_t)xdr_ACCESS3args,      sizeof(ACCESS3args),      (xdrproc_t)xdr_ACCESS3res,      sizeof(ACCESS3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READLINK,    kv_nfs3_readlink_proc,    (xdrproc_t)xdr_READLINK3args,    sizeof(READLINK3args),    (xdrproc_t)xdr_READLINK3res,    sizeof(READLINK3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READ,        kv_nfs3_read_proc,        (xdrproc_t)xdr_READ3args,        sizeof(READ3args),        (xdrproc_t)xdr_READ3res,        sizeof(READ3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_WRITE,       kv_nfs3_write_proc,       (xdrproc_t)xdr_WRITE3args,       sizeof(WRITE3args),       (xdrproc_t)xdr_WRITE3res,       sizeof(WRITE3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_CREATE,      kv_nfs3_create_proc,      (xdrproc_t)xdr_CREATE3args,      sizeof(CREATE3args),      (xdrproc_t)xdr_CREATE3res,      sizeof(CREATE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKDIR,       kv_nfs3_mkdir_proc,       (xdrproc_t)xdr_MKDIR3args,       sizeof(MKDIR3args),       (xdrproc_t)xdr_MKDIR3res,       sizeof(MKDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_SYMLINK,     kv_nfs3_symlink_proc,     (xdrproc_t)xdr_SYMLINK3args,     sizeof(SYMLINK3args),     (xdrproc_t)xdr_SYMLINK3res,     sizeof(SYMLINK3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_MKNOD,       kv_nfs3_mknod_proc,       (xdrproc_t)xdr_MKNOD3args,       sizeof(MKNOD3args),       (xdrproc_t)xdr_MKNOD3res,       sizeof(MKNOD3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_REMOVE,      kv_nfs3_remove_proc,      (xdrproc_t)xdr_REMOVE3args,      sizeof(REMOVE3args),      (xdrproc_t)xdr_REMOVE3res,      sizeof(REMOVE3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RMDIR,       kv_nfs3_rmdir_proc,       (xdrproc_t)xdr_RMDIR3args,       sizeof(RMDIR3args),       (xdrproc_t)xdr_RMDIR3res,       sizeof(RMDIR3res),       self},
+        {NFS_PROGRAM, NFS_V3, NFS3_RENAME,      kv_nfs3_rename_proc,      (xdrproc_t)xdr_RENAME3args,      sizeof(RENAME3args),      (xdrproc_t)xdr_RENAME3res,      sizeof(RENAME3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_LINK,        kv_nfs3_link_proc,        (xdrproc_t)xdr_LINK3args,        sizeof(LINK3args),        (xdrproc_t)xdr_LINK3res,        sizeof(LINK3res),        self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIR,     kv_nfs3_readdir_proc,     (xdrproc_t)xdr_READDIR3args,     sizeof(READDIR3args),     (xdrproc_t)xdr_READDIR3res,     sizeof(READDIR3res),     self},
+        {NFS_PROGRAM, NFS_V3, NFS3_READDIRPLUS, kv_nfs3_readdirplus_proc, (xdrproc_t)xdr_READDIRPLUS3args, sizeof(READDIRPLUS3args), (xdrproc_t)xdr_READDIRPLUS3res, sizeof(READDIRPLUS3res), self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSSTAT,      nfs3_fsstat_proc,         (xdrproc_t)xdr_FSSTAT3args,      sizeof(FSSTAT3args),      (xdrproc_t)xdr_FSSTAT3res,      sizeof(FSSTAT3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_FSINFO,      nfs3_fsinfo_proc,         (xdrproc_t)xdr_FSINFO3args,      sizeof(FSINFO3args),      (xdrproc_t)xdr_FSINFO3res,      sizeof(FSINFO3res),      self},
+        {NFS_PROGRAM, NFS_V3, NFS3_PATHCONF,    nfs3_pathconf_proc,       (xdrproc_t)xdr_PATHCONF3args,    sizeof(PATHCONF3args),    (xdrproc_t)xdr_PATHCONF3res,    sizeof(PATHCONF3res),    self},
+        {NFS_PROGRAM, NFS_V3, NFS3_COMMIT,      nfs3_commit_proc,         (xdrproc_t)xdr_COMMIT3args,      sizeof(COMMIT3args),      (xdrproc_t)xdr_COMMIT3res,      sizeof(COMMIT3res),      self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_NULL,    nfs3_null_proc,         NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_MNT,     mount3_mnt_proc,        (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      (xdrproc_t)xdr_nfs_mountres3, sizeof(nfs_mountres3),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_DUMP,    mount3_dump_proc,       NULL,                            0,                        (xdrproc_t)xdr_nfs_mountlist, sizeof(nfs_mountlist),     self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNT,    mount3_umnt_proc,       (xdrproc_t)xdr_nfs_dirpath,      sizeof(nfs_dirpath),      NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_UMNTALL, mount3_umntall_proc,    NULL,                            0,                        NULL,                         0,                         self},
+        {MOUNT_PROGRAM, MOUNT_V3, MOUNT3_EXPORT,  mount3_export_proc,     NULL,                            0,                        (xdrproc_t)xdr_nfs_exports,   sizeof(nfs_exports),       self},
+    };
+    for (int i = 0; i < sizeof(pt)/sizeof(pt[0]); i++)
+    {
+        self->proc_table.insert(pt[i]);
+    }
+}

src/nfs_kv.h

@@ -0,0 +1,97 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - header
+
+#pragma once
+
+#include "nfs/nfs.h"
+
+#define KV_ROOT_INODE 1
+#define KV_NEXT_ID_KEY "id"
+#define SHARED_FILE_MAGIC_V1 0x711A5158A6EDF17E
+
+struct nfs_kv_write_state;
+
+struct list_cookie_t
+{
+    uint64_t dir_ino, cookieverf, cookie;
+};
+
+inline bool operator < (const list_cookie_t & a, const list_cookie_t & b)
+{
+    return a.dir_ino < b.dir_ino || a.dir_ino == b.dir_ino &&
+        (a.cookieverf < b.cookieverf || a.cookieverf == b.cookieverf && a.cookie < b.cookie);
+};
+
+struct list_cookie_val_t
+{
+    std::string key;
+};
+
+struct shared_alloc_queue_t
+{
+    nfs_kv_write_state *st;
+    int state;
+    uint64_t size;
+};
+
+struct kv_inode_extend_t
+{
+    int refcnt = 0;
+    uint64_t cur_extend = 0, next_extend = 0, done_extend = 0;
+    std::vector<std::function<void()>> waiters;
+};
+
+struct kv_fs_state_t
+{
+    std::map<list_cookie_t, list_cookie_val_t> list_cookies;
+    uint64_t fs_next_id = 1, fs_allocated_id = 0;
+    std::vector<uint64_t> unallocated_ids;
+    std::vector<shared_alloc_queue_t> allocating_shared;
+    uint64_t cur_shared_inode = 0, cur_shared_offset = 0;
+    std::map<inode_t, kv_inode_extend_t> extends;
+    std::vector<uint8_t> zero_block;
+};
+
+struct shared_file_header_t
+{
+    uint64_t magic = 0;
+    uint64_t inode = 0;
+    uint64_t alloc = 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, std::function<void(int res, uint64_t new_id)> cb);
+void kv_read_inode(nfs_client_t *self, uint64_t ino,
+    std::function<void(int res, const std::string & value, json11::Json ientry)> cb,
+    bool allow_cache = false);
+uint64_t align_shared_size(nfs_client_t *self, uint64_t size);
+
+int kv_nfs3_getattr_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_setattr_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_lookup_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readlink_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_read_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_write_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_create_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_mkdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_symlink_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_mknod_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_remove_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_rmdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_rename_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_link_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readdir_proc(void *opaque, rpc_op_t *rop);
+int kv_nfs3_readdirplus_proc(void *opaque, rpc_op_t *rop);

src/nfs_kv_create.cpp

@@ -0,0 +1,343 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - CREATE, MKDIR, SYMLINK, MKNOD
+
+#include <sys/time.h>
+
+#include "str_util.h"
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+void allocate_new_id(nfs_client_t *self, std::function<void(int res, uint64_t new_id)> cb)
+{
+    if (self->parent->kvfs->fs_next_id <= self->parent->kvfs->fs_allocated_id)
+    {
+        cb(0, self->parent->kvfs->fs_next_id++);
+        return;
+    }
+    else if (self->parent->kvfs->fs_next_id > self->parent->fs_inode_count)
+    {
+        cb(-ENOSPC, 0);
+        return;
+    }
+    self->parent->db->get(KV_NEXT_ID_KEY, [=](int res, const std::string & prev_str)
+    {
+        if (res < 0 && res != -ENOENT)
+        {
+            cb(res, 0);
+            return;
+        }
+        uint64_t prev_val = stoull_full(prev_str);
+        if (prev_val >= self->parent->fs_inode_count)
+        {
+            cb(-ENOSPC, 0);
+            return;
+        }
+        if (prev_val < 1)
+        {
+            prev_val = 1;
+        }
+        uint64_t new_val = prev_val + self->parent->id_alloc_batch_size;
+        if (new_val >= self->parent->fs_inode_count)
+        {
+            new_val = self->parent->fs_inode_count;
+        }
+        self->parent->db->set(KV_NEXT_ID_KEY, std::to_string(new_val), [=](int res)
+        {
+            if (res == -EAGAIN)
+            {
+                // CAS failure - retry
+                allocate_new_id(self, cb);
+            }
+            else if (res < 0)
+            {
+                cb(res, 0);
+            }
+            else
+            {
+                self->parent->kvfs->fs_next_id = prev_val+2;
+                self->parent->kvfs->fs_allocated_id = new_val;
+                cb(0, prev_val+1);
+            }
+        }, [prev_val](int res, const std::string & value)
+        {
+            // FIXME: Allow to modify value from CAS callback? ("update" query)
+            return res < 0 || stoull_full(value) == prev_val;
+        });
+    });
+}
+
+struct kv_create_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    bool exclusive = false;
+    uint64_t verf = 0;
+    uint64_t dir_ino = 0;
+    std::string filename;
+    int res = 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;
+    }
+    if (st->attrobj.find("mtime") == st->attrobj.end())
+        st->attrobj["mtime"] = nfstime_now_str();
+    if (st->attrobj.find("atime") == st->attrobj.end())
+        st->attrobj["atime"] = st->attrobj["mtime"];
+    st->attrs = std::move(st->attrobj);
+resume_1:
+    // Generate inode ID
+    allocate_new_id(st->self, [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
+        {
+            st->self->parent->kvfs->unallocated_ids.push_back(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;
+    }
+    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_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_client_t *self, uint64_t ino,
+    std::function<void(int res, const std::string & value, json11::Json ientry)> cb,
+    bool allow_cache)
+{
+    auto key = kv_inode_key(ino);
+    self->parent->db->get(key, [=](int res, const std::string & value)
+    {
+        if (ino == KV_ROOT_INODE && res == -ENOENT)
+        {
+            // Allow root inode to not exist
+            cb(0, "", json11::Json(json11::Json::object{ { "type", "dir" } }));
+            return;
+        }
+        if (res < 0)
+        {
+            if (res != -ENOENT)
+                fprintf(stderr, "Error reading inode %s: %s (code %d)\n", kv_inode_key(ino).c_str(), strerror(-res), res);
+            cb(res, "", json11::Json());
+            return;
+        }
+        std::string err;
+        auto attrs = json11::Json::parse(value, err);
+        if (err != "")
+        {
+            fprintf(stderr, "Invalid JSON in inode %s = %s: %s\n", kv_inode_key(ino).c_str(), value.c_str(), err.c_str());
+            res = -EIO;
+        }
+        cb(res, value, attrs);
+    }, allow_cache);
+}
+
+int kv_nfs3_getattr_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_client_t *self = (nfs_client_t*)opaque;
+    GETATTR3args *args = (GETATTR3args*)rop->request;
+    GETATTR3res *reply = (GETATTR3res*)rop->reply;
+    std::string fh = args->object;
+    auto ino = kv_fh_inode(fh);
+    if (self->parent->trace)
+        fprintf(stderr, "[%d] GETATTR %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, 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_kv_link.cpp

@@ -0,0 +1,188 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - LINK
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_kv_link_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t dir_ino = 0;
+    std::string filename;
+    std::string ientry_text;
+    json11::Json ientry;
+    bool retrying = false;
+    int wait = 0;
+    int res = 0, res2 = 0;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_link(nfs_kv_link_state *st, int state)
+{
+    // 1) Read the source inode
+    // 2) If it's a directory - fail with -EISDIR
+    // 3) Create the new direntry with the same inode reference
+    // 4) Update the inode entry with refcount++
+    // 5) Retry update if CAS failed but the inode exists
+    // 6) Otherwise fail and remove the new direntry
+    // Yeah we may leave a bad direntry if we crash
+    // But the other option is to possibly leave an inode with too big refcount
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else if (state == 4) goto resume_4;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_link()");
+        abort();
+    }
+resume_0:
+    // Check that the source inode exists and is not a directory
+    st->wait = st->retrying ? 1 : 2;
+    st->res2 = 0;
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res == 0 ? (attrs["type"].string_value() == "dir" ? -EISDIR : 0) : res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        if (!--st->wait)
+            nfs_kv_continue_link(st, 1);
+    });
+    if (!st->retrying)
+    {
+        // Check that the new directory exists
+        kv_read_inode(st->self, st->dir_ino, [st](int res, const std::string & value, json11::Json attrs)
+        {
+            st->res2 = res == 0 ? (attrs["type"].string_value() == "dir" ? 0 : -ENOTDIR) : res;
+            if (!--st->wait)
+                nfs_kv_continue_link(st, 1);
+        });
+    }
+    return;
+resume_1:
+    if (st->res < 0 || st->res2 < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res < 0 ? st->res : st->res2);
+        return;
+    }
+    // Write the new direntry
+    if (!st->retrying)
+    {
+        st->self->parent->db->set(kv_direntry_key(st->dir_ino, st->filename),
+            json11::Json(json11::Json::object{ { "ino", st->ino } }).dump(), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_link(st, 2);
+        }, [st](int res, const std::string & old_value)
+        {
+            return res == -ENOENT;
+        });
+        return;
+resume_2:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+    }
+    // Increase inode refcount
+    {
+        auto new_ientry = st->ientry.object_items();
+        auto nlink = new_ientry["nlink"].uint64_value();
+        new_ientry["nlink"] = nlink ? nlink+1 : 2;
+        st->ientry = new_ientry;
+    }
+    st->self->parent->db->set(kv_inode_key(st->ino), st->ientry.dump(), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_link(st, 3);
+    }, [st](int res, const std::string & old_value)
+    {
+        st->res2 = res;
+        return res == 0 && old_value == st->ientry_text;
+    });
+    return;
+resume_3:
+    if (st->res2 == -ENOENT)
+    {
+        st->res = -ENOENT;
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Re-read inode and retry
+        st->retrying = true;
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        // Maybe inode was deleted in the meantime, delete our direntry
+        st->self->parent->db->del(kv_direntry_key(st->dir_ino, st->filename), [st](int res)
+        {
+            st->res2 = res;
+            nfs_kv_continue_link(st, 4);
+        });
+        return;
+resume_4:
+        if (st->res2 < 0)
+        {
+            fprintf(stderr, "Warning: failed to delete new linked direntry %ju/%s: %s (code %d)\n",
+                st->dir_ino, st->filename.c_str(), strerror(-st->res2), st->res2);
+        }
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+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_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, 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, kv_fh_inode(args->symlink), [=](int res, const std::string & value, json11::Json attrs)
+    {
+        if (res < 0)
+        {
+            *reply = (READLINK3res){ .status = vitastor_nfs_map_err(-res) };
+        }
+        else if (attrs["type"] != "link")
+        {
+            *reply = (READLINK3res){ .status = NFS3ERR_INVAL };
+        }
+        else
+        {
+            *reply = (READLINK3res){
+                .status = NFS3_OK,
+                .resok = (READLINK3resok){
+                    .data = xdr_copy_string(rop->xdrs, attrs["symlink"].string_value()),
+                },
+            };
+        }
+        rpc_queue_reply(rop);
+    });
+    return 1;
+}

src/nfs_kv_read.cpp

@@ -0,0 +1,163 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - READ
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_kv_read_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    bool allow_cache = true;
+    inode_t ino = 0;
+    uint64_t offset = 0, size = 0;
+    std::function<void(int)> cb;
+    // state
+    int res = 0;
+    json11::Json ientry;
+    uint64_t aligned_size = 0, aligned_offset = 0;
+    uint8_t *aligned_buf = NULL;
+    cluster_op_t *op = NULL;
+    uint8_t *buf = NULL;
+};
+
+static void nfs_kv_continue_read(nfs_kv_read_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_read()");
+        abort();
+    }
+    if (st->offset + sizeof(shared_file_header_t) < st->self->parent->shared_inode_threshold)
+    {
+        kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+        {
+            st->res = res;
+            st->ientry = attrs;
+            nfs_kv_continue_read(st, 1);
+        }, st->allow_cache);
+        return;
+resume_1:
+        if (st->res < 0 || kv_map_type(st->ientry["type"].string_value()) != NF3REG)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res < 0 ? st->res : -EINVAL);
+            return;
+        }
+        if (st->ientry["shared_ino"].uint64_value() != 0)
+        {
+            st->aligned_size = align_shared_size(st->self, st->offset+st->size);
+            st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+            st->buf = st->aligned_buf + sizeof(shared_file_header_t) + st->offset;
+            st->op = new cluster_op_t;
+            st->op->opcode = OSD_OP_READ;
+            st->op->inode = st->self->parent->fs_base_inode + st->ientry["shared_ino"].uint64_value();
+            st->op->offset = st->ientry["shared_offset"].uint64_value();
+            if (st->offset+st->size > st->ientry["size"].uint64_value())
+            {
+                st->op->len = align_shared_size(st->self, st->ientry["size"].uint64_value());
+                memset(st->aligned_buf+st->op->len, 0, st->aligned_size-st->op->len);
+            }
+            else
+                st->op->len = st->aligned_size;
+            st->op->iov.push_back(st->aligned_buf, st->op->len);
+            st->op->callback = [st, state](cluster_op_t *op)
+            {
+                st->res = op->retval == op->len ? 0 : op->retval;
+                delete op;
+                nfs_kv_continue_read(st, 2);
+            };
+            st->self->parent->cli->execute(st->op);
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+            auto hdr = ((shared_file_header_t*)st->aligned_buf);
+            if (hdr->magic != SHARED_FILE_MAGIC_V1 || hdr->inode != st->ino)
+            {
+                // Got unrelated data - retry from the beginning
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                st->allow_cache = false;
+                nfs_kv_continue_read(st, 0);
+                return;
+            }
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+    }
+    st->aligned_offset = (st->offset & ~(st->self->parent->pool_alignment-1));
+    st->aligned_size = ((st->offset + st->size + st->self->parent->pool_alignment-1) &
+        ~(st->self->parent->pool_alignment-1)) - st->aligned_offset;
+    st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+    st->buf = st->aligned_buf + st->offset - st->aligned_offset;
+    st->op = new cluster_op_t;
+    st->op->opcode = OSD_OP_READ;
+    st->op->inode = st->self->parent->fs_base_inode + st->ino;
+    st->op->offset = st->aligned_offset;
+    st->op->len = st->aligned_size;
+    st->op->iov.push_back(st->aligned_buf, st->aligned_size);
+    st->op->callback = [st](cluster_op_t *op)
+    {
+        st->res = op->retval;
+        delete op;
+        nfs_kv_continue_read(st, 3);
+    };
+    st->self->parent->cli->execute(st->op);
+    return;
+resume_3:
+    auto cb = std::move(st->cb);
+    cb(st->res < 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 = 0;
+        }
+        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_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->readdir_getattr_parallel)
+    {
+        auto idx = st->getattr_cur++;
+        st->getattr_running++;
+        kv_read_inode(st->self, st->entries[idx].fileid, [st, idx](int res, const std::string & value, json11::Json ientry)
+        {
+            if (res == 0)
+            {
+                st->entries[idx].name_attributes = (post_op_attr){
+                    // FIXME: maybe do not read parent attributes and leave them to a GETATTR?
+                    .attributes_follow = 1,
+                    .attributes = get_kv_attributes(st->self, st->entries[idx].fileid, ientry),
+                };
+            }
+            st->getattr_running--;
+            kv_getattr_next(st);
+            if (st->getattr_running == 0 && !st->list_handle)
+            {
+                nfs_kv_continue_readdir(st, 4);
+            }
+        });
+    }
+}
+
+static void nfs_kv_continue_readdir(nfs_kv_readdir_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else if (state == 4) goto resume_4;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_readdir()");
+        abort();
+    }
+    // Limit results based on maximum reply size
+    // Sadly we have to calculate reply size by hand
+    // reply without entries is 4+4+(dir_attributes ? sizeof(fattr3) : 0)+8+4 bytes
+    st->reply_size = 20;
+    if (st->reply_size > st->maxcount)
+    {
+        // Error, too small max reply size
+        auto cb = std::move(st->cb);
+        cb(-NFS3ERR_TOOSMALL);
+        return;
+    }
+    // Add . and ..
+    if (st->cookie <= 1)
+    {
+        kv_read_inode(st->self, st->dir_ino, [st](int res, const std::string & value, json11::Json ientry)
+        {
+            st->res = res;
+            st->ientry_text = value;
+            st->ientry = ientry;
+            nfs_kv_continue_readdir(st, 1);
+        });
+        return;
+resume_1:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        if (st->cookie == 0)
+        {
+            auto fh = kv_fh(st->dir_ino);
+            auto entry_size = 20 + 4/*len_pad4(".")*/ + (st->is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+            if (st->reply_size + entry_size > st->maxcount)
+            {
+                auto cb = std::move(st->cb);
+                cb(-NFS3ERR_TOOSMALL);
+                return;
+            }
+            entryplus3 dot = {};
+            dot.name = xdr_copy_string(st->rop->xdrs, ".");
+            dot.fileid = st->dir_ino;
+            dot.name_attributes = (post_op_attr){
+                .attributes_follow = 1,
+                .attributes = get_kv_attributes(st->self, st->dir_ino, st->ientry),
+            };
+            dot.name_handle = (post_op_fh3){
+                .handle_follows = 1,
+                .handle = xdr_copy_string(st->rop->xdrs, fh),
+            };
+            st->entries.push_back(dot);
+            st->reply_size += entry_size;
+        }
+        st->parent_ino = st->ientry["parent_ino"].uint64_value();
+        if (st->parent_ino)
+        {
+            kv_read_inode(st->self, st->ientry["parent_ino"].uint64_value(), [st](int res, const std::string & value, json11::Json ientry)
+            {
+                st->res = res;
+                st->parent_ientry_text = value;
+                st->parent_ientry = ientry;
+                nfs_kv_continue_readdir(st, 2);
+            });
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+        }
+        auto fh = kv_fh(st->parent_ino);
+        auto entry_size = 20 + 4/*len_pad4("..")*/ + (st->is_plus ? 8 + 88 + len_pad4(fh.size()) : 0);
+        if (st->reply_size + entry_size > st->maxcount)
+        {
+            st->eof = false;
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+        entryplus3 dotdot = {};
+        dotdot.name = xdr_copy_string(st->rop->xdrs, "..");
+        dotdot.fileid = st->dir_ino;
+        dotdot.name_attributes = (post_op_attr){
+            // FIXME: maybe do not read parent attributes and leave them to a GETATTR?
+            .attributes_follow = 1,
+            .attributes = get_kv_attributes(st->self,
+                st->parent_ino ? st->parent_ino : st->dir_ino,
+                st->parent_ino ? st->parent_ientry : st->ientry),
+        };
+        dotdot.name_handle = (post_op_fh3){
+            .handle_follows = 1,
+            .handle = xdr_copy_string(st->rop->xdrs, fh),
+        };
+        st->entries.push_back(dotdot);
+        st->reply_size += entry_size;
+    }
+    st->prefix = kv_direntry_key(st->dir_ino, "");
+    st->eof = true;
+    st->start = st->prefix;
+    if (st->cookie > 1)
+    {
+        auto lc_it = st->self->parent->kvfs->list_cookies.find((list_cookie_t){ st->dir_ino, st->cookieverf, st->cookie });
+        if (lc_it != st->self->parent->kvfs->list_cookies.end())
+        {
+            st->start = 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 %lu %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_kv_remove.cpp

@@ -0,0 +1,315 @@
+// 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;
+        st->self->parent->db->set(kv_inode_key(st->ino), json11::Json(copy).dump(), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_delete(st, 6);
+        }, [st](int res, const std::string & old_value)
+        {
+            return old_value == st->ientry_text;
+        });
+    }
+    else
+    {
+        st->self->parent->db->del(kv_inode_key(st->ino), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_delete(st, 6);
+        }, [st](int res, const std::string & old_value)
+        {
+            return old_value == st->ientry_text;
+        });
+    }
+    return;
+resume_6:
+    if (st->res < 0)
+    {
+        // Assume EAGAIN is OK, maybe someone created a hard link in the meantime
+        auto cb = std::move(st->cb);
+        cb(st->res == -EAGAIN ? 0 : st->res);
+        return;
+    }
+    // (6) If regular file and inode is deleted: delete data
+    if ((!st->type || st->type == NF3REG) && st->ientry["nlink"].uint64_value() <= 1 &&
+        !st->ientry["shared_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->self->parent->fs_base_inode + st->ino) },
+            { "pool", (uint64_t)INODE_POOL(st->self->parent->fs_base_inode + st->ino) },
+        }), [st](const cli_result_t & r)
+        {
+            if (r.err)
+            {
+                fprintf(stderr, "Failed to remove inode %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;
+    }
+    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_kv_rename.cpp

@@ -0,0 +1,392 @@
+// 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, 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, 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;
+                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->self->parent->fs_base_inode + st->new_direntry["ino"].uint64_value()) },
+                    { "pool", (uint64_t)INODE_POOL(st->self->parent->fs_base_inode + 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, 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;
+            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;
+        }
+    }
+    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_kv_setattr.cpp

@@ -0,0 +1,189 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - SETATTR
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+#include "cli.h"
+
+struct nfs_kv_setattr_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t old_size = 0, new_size = 0;
+    json11::Json::object set_attrs;
+    int res = 0, cas_res = 0;
+    std::string ientry_text;
+    json11::Json ientry;
+    json11::Json::object new_attrs;
+    std::function<void(int)> cb;
+};
+
+static void nfs_kv_continue_setattr(nfs_kv_setattr_state *st, int state)
+{
+    // 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();
+    }
+resume_0:
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        nfs_kv_continue_setattr(st, 1);
+    });
+    return;
+resume_1:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (st->ientry["type"].string_value() != "file" &&
+        st->ientry["type"].string_value() != "" &&
+        !st->set_attrs["size"].is_null())
+    {
+        auto cb = std::move(st->cb);
+        cb(-EINVAL);
+        return;
+    }
+    // Now we can update it
+    st->new_attrs = st->ientry.object_items();
+    st->old_size = st->ientry["size"].uint64_value();
+    for (auto & kv: st->set_attrs)
+    {
+        if (kv.first == "size")
+        {
+            st->new_size = kv.second.uint64_value();
+        }
+        st->new_attrs[kv.first] = kv.second;
+    }
+    st->self->parent->db->set(kv_inode_key(st->ino), json11::Json(st->new_attrs).dump(), [st](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_setattr(st, 2);
+    }, [st](int res, const std::string & cas_value)
+    {
+        st->cas_res = res;
+        return (res == 0 || res == -ENOENT && st->ino == KV_ROOT_INODE) && cas_value == st->ientry_text;
+    });
+    return;
+resume_2:
+    if (st->cas_res == -ENOENT)
+    {
+        st->res = -ENOENT;
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Retry
+        fprintf(stderr, "CAS failure during setattr, retrying\n");
+        goto resume_0;
+    }
+    if (st->res < 0)
+    {
+        fprintf(stderr, "Failed to update inode %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->self->parent->fs_base_inode + st->ino) },
+            { "pool", (uint64_t)INODE_POOL(st->self->parent->fs_base_inode + st->ino) },
+            { "min_offset", st->set_attrs["size"].uint64_value() },
+        }), [st](const cli_result_t & r)
+        {
+            if (r.err)
+            {
+                fprintf(stderr, "Failed to truncate inode %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->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_kv_write.cpp

@@ -0,0 +1,913 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS proxy over VitastorKV database - WRITE
+
+#include <sys/time.h>
+
+#include "nfs_proxy.h"
+#include "nfs_kv.h"
+
+struct nfs_rmw_t
+{
+    nfs_kv_write_state *st = NULL;
+    int continue_state = 0;
+    uint64_t ino = 0;
+    uint64_t offset = 0;
+    uint8_t *buf = NULL;
+    uint64_t size = 0;
+    uint8_t *part_buf = NULL;
+    uint64_t version = 0;
+};
+
+struct nfs_kv_write_state
+{
+    nfs_client_t *self = NULL;
+    rpc_op_t *rop = NULL;
+    uint64_t ino = 0;
+    uint64_t offset = 0, size = 0;
+    bool stable = false;
+    uint8_t *buf = NULL;
+    std::function<void(int res)> cb;
+    // state
+    bool allow_cache = true;
+    int res = 0, res2 = 0;
+    int waiting = 0;
+    std::string ientry_text;
+    json11::Json ientry;
+    uint64_t new_size = 0;
+    uint64_t aligned_size = 0;
+    uint8_t *aligned_buf = NULL;
+    uint64_t shared_inode = 0, shared_offset = 0;
+    bool was_immediate = false;
+    nfs_rmw_t rmw[2];
+    shared_file_header_t shdr;
+    kv_inode_extend_t *ext = NULL;
+
+    ~nfs_kv_write_state()
+    {
+        if (aligned_buf)
+        {
+            free(aligned_buf);
+            aligned_buf = NULL;
+        }
+    }
+};
+
+static void nfs_kv_continue_write(nfs_kv_write_state *st, int state);
+
+static void finish_allocate_shared(nfs_client_t *self, int res)
+{
+    std::vector<shared_alloc_queue_t> waiting;
+    waiting.swap(self->parent->kvfs->allocating_shared);
+    for (auto & w: waiting)
+    {
+        w.st->res = res;
+        if (res == 0)
+        {
+            w.st->shared_inode = self->parent->kvfs->cur_shared_inode;
+            w.st->shared_offset = self->parent->kvfs->cur_shared_offset;
+            self->parent->kvfs->cur_shared_offset += (w.size + self->parent->pool_alignment-1) & ~(self->parent->pool_alignment-1);
+        }
+        nfs_kv_continue_write(w.st, w.state);
+    }
+}
+
+static void allocate_shared_inode(nfs_kv_write_state *st, int state, uint64_t size)
+{
+    if (st->self->parent->kvfs->cur_shared_inode == 0)
+    {
+        st->self->parent->kvfs->allocating_shared.push_back({ st, state, size });
+        if (st->self->parent->kvfs->allocating_shared.size() > 1)
+        {
+            return;
+        }
+        allocate_new_id(st->self, [st](int res, uint64_t new_id)
+        {
+            if (res < 0)
+            {
+                finish_allocate_shared(st->self, res);
+                return;
+            }
+            st->self->parent->kvfs->cur_shared_inode = new_id;
+            st->self->parent->kvfs->cur_shared_offset = 0;
+            st->self->parent->db->set(
+                kv_inode_key(new_id), json11::Json(json11::Json::object{ { "type", "shared" } }).dump(),
+                [st](int res)
+                {
+                    if (res < 0)
+                    {
+                        st->self->parent->kvfs->cur_shared_inode = 0;
+                    }
+                    finish_allocate_shared(st->self, res);
+                },
+                [](int res, const std::string & old_value)
+                {
+                    return res == -ENOENT;
+                }
+            );
+        });
+    }
+    else
+    {
+        st->res = 0;
+        st->shared_inode = st->self->parent->kvfs->cur_shared_inode;
+        st->shared_offset = st->self->parent->kvfs->cur_shared_offset;
+        st->self->parent->kvfs->cur_shared_offset += (size + st->self->parent->pool_alignment-1) & ~(st->self->parent->pool_alignment-1);
+        nfs_kv_continue_write(st, state);
+    }
+}
+
+uint64_t align_shared_size(nfs_client_t *self, uint64_t size)
+{
+    return (size + sizeof(shared_file_header_t) + self->parent->pool_alignment-1)
+        & ~(self->parent->pool_alignment-1);
+}
+
+static void nfs_do_write(uint64_t ino, uint64_t offset, uint64_t size, std::function<void(cluster_op_t *op)> prepare, nfs_kv_write_state *st, int state)
+{
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_WRITE;
+    op->inode = st->self->parent->fs_base_inode + ino;
+    op->offset = offset;
+    op->len = size;
+    prepare(op);
+    st->waiting++;
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        if (op->retval != op->len)
+        {
+            st->res = op->retval >= 0 ? -EIO : op->retval;
+        }
+        delete op;
+        st->waiting--;
+        if (!st->waiting)
+        {
+            nfs_kv_continue_write(st, state);
+        }
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static void nfs_do_unshare_write(nfs_kv_write_state *st, int state)
+{
+    uint64_t unshare_size = (st->ientry["size"].uint64_value() + st->self->parent->pool_alignment-1)
+        & ~(st->self->parent->pool_alignment-1);
+    nfs_do_write(st->ino, 0, unshare_size, [&](cluster_op_t *op)
+    {
+        op->iov.push_back(st->aligned_buf + sizeof(shared_file_header_t), unshare_size);
+    }, st, state);
+}
+
+static void nfs_do_rmw(nfs_rmw_t *rmw)
+{
+    auto parent = rmw->st->self->parent;
+    auto align = parent->pool_alignment;
+    assert(rmw->size < align);
+    assert((rmw->offset/parent->pool_block_size) == ((rmw->offset+rmw->size-1)/parent->pool_block_size));
+    if (!rmw->part_buf)
+    {
+        rmw->part_buf = (uint8_t*)malloc_or_die(align);
+    }
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_READ;
+    op->inode = parent->fs_base_inode + rmw->ino;
+    op->offset = rmw->offset & ~(align-1);
+    op->len = align;
+    op->iov.push_back(rmw->part_buf, op->len);
+    rmw->st->waiting++;
+    op->callback = [rmw](cluster_op_t *rd_op)
+    {
+        if (rd_op->retval != rd_op->len)
+        {
+            free(rmw->part_buf);
+            rmw->part_buf = NULL;
+            rmw->st->res = rd_op->retval >= 0 ? -EIO : rd_op->retval;
+            rmw->st->waiting--;
+            if (!rmw->st->waiting)
+            {
+                nfs_kv_continue_write(rmw->st, rmw->continue_state);
+            }
+        }
+        else
+        {
+            if (!rmw->version)
+            {
+                auto st = rmw->st;
+                rmw->version = rd_op->version+1;
+                if (st->rmw[0].st && st->rmw[1].st &&
+                    st->rmw[0].offset/st->self->parent->pool_block_size == st->rmw[1].offset/st->self->parent->pool_block_size)
+                {
+                    // Same block... RMWs should be sequential
+                    int other = rmw == &st->rmw[0] ? 1 : 0;
+                    st->rmw[other].version = rmw->version+1;
+                }
+            }
+            auto parent = rmw->st->self->parent;
+            auto align = parent->pool_alignment;
+            bool is_begin = (rmw->offset % align);
+            bool is_end = ((rmw->offset+rmw->size) % align);
+            auto op = new cluster_op_t;
+            op->opcode = OSD_OP_WRITE;
+            op->inode = rmw->st->self->parent->fs_base_inode + rmw->ino;
+            op->offset = rmw->offset & ~(align-1);
+            op->len = align;
+            op->version = rmw->version;
+            if (is_begin)
+            {
+                op->iov.push_back(rmw->part_buf, rmw->offset % align);
+            }
+            op->iov.push_back(rmw->buf, rmw->size);
+            if (is_end)
+            {
+                op->iov.push_back(rmw->part_buf + (rmw->offset % align) + rmw->size, align - (rmw->offset % align) - rmw->size);
+            }
+            op->callback = [rmw](cluster_op_t *op)
+            {
+                if (op->retval == -EINTR)
+                {
+                    // CAS failure - retry
+                    rmw->version = 0;
+                    rmw->st->waiting--;
+                    nfs_do_rmw(rmw);
+                }
+                else
+                {
+                    free(rmw->part_buf);
+                    rmw->part_buf = NULL;
+                    if (op->retval != op->len)
+                    {
+                        rmw->st->res = (op->retval >= 0 ? -EIO : op->retval);
+                    }
+                    rmw->st->waiting--;
+                    if (!rmw->st->waiting)
+                    {
+                        nfs_kv_continue_write(rmw->st, rmw->continue_state);
+                    }
+                }
+                delete op;
+            };
+            parent->cli->execute(op);
+        }
+        delete rd_op;
+    };
+    parent->cli->execute(op);
+}
+
+static void nfs_do_shared_read(nfs_kv_write_state *st, int state)
+{
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_READ;
+    op->inode = st->self->parent->fs_base_inode + st->ientry["shared_ino"].uint64_value();
+    op->offset = st->ientry["shared_offset"].uint64_value();
+    op->len = align_shared_size(st->self, st->ientry["size"].uint64_value());
+    op->iov.push_back(st->aligned_buf, op->len);
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        st->res = op->retval == op->len ? 0 : op->retval;
+        delete op;
+        nfs_kv_continue_write(st, state);
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static void nfs_do_fsync(nfs_kv_write_state *st, int state)
+{
+    // Client requested a stable write. Add an fsync
+    auto op = new cluster_op_t;
+    op->opcode = OSD_OP_SYNC;
+    op->callback = [st, state](cluster_op_t *op)
+    {
+        delete op;
+        nfs_kv_continue_write(st, state);
+    };
+    st->self->parent->cli->execute(op);
+}
+
+static bool nfs_do_shared_readmodify(nfs_kv_write_state *st, int base_state, int state, bool unshare)
+{
+    assert(state <= base_state);
+    if (state < base_state)       {}
+    else if (state == base_state) goto resume_0;
+    assert(!st->aligned_buf);
+    st->aligned_size = unshare
+        ? sizeof(shared_file_header_t) + ((st->new_size + st->self->parent->pool_alignment-1) & ~(st->self->parent->pool_alignment-1))
+        : align_shared_size(st->self, st->new_size);
+    st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+    // FIXME do not allocate zeroes if we only need zeroes
+    memset(st->aligned_buf + sizeof(shared_file_header_t), 0, st->offset);
+    memset(st->aligned_buf + sizeof(shared_file_header_t) + st->offset + st->size, 0,
+        st->aligned_size - sizeof(shared_file_header_t) - st->offset - st->size);
+    if (st->ientry["shared_ino"].uint64_value() != 0 &&
+        st->ientry["size"].uint64_value() != 0)
+    {
+        // Read old data if shared non-empty
+        nfs_do_shared_read(st, base_state);
+        return false;
+resume_0:
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return false;
+        }
+        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
+            st->allow_cache = false;
+            free(st->aligned_buf);
+            st->aligned_buf = NULL;
+            nfs_kv_continue_write(st, 0);
+            return false;
+        }
+    }
+    // FIXME put shared_file_header_t after data to not break alignment
+    *((shared_file_header_t*)st->aligned_buf) = {
+        .magic = SHARED_FILE_MAGIC_V1,
+        .inode = st->ino,
+        .alloc = st->aligned_size,
+    };
+    return true;
+}
+
+static void nfs_do_shared_write(nfs_kv_write_state *st, int state, bool only_aligned)
+{
+    nfs_do_write(st->shared_inode, st->shared_offset, st->aligned_size, [&](cluster_op_t *op)
+    {
+        if (only_aligned)
+            op->iov.push_back(st->aligned_buf, st->aligned_size);
+        else
+        {
+            op->iov.push_back(st->aligned_buf, sizeof(shared_file_header_t) + st->offset);
+            op->iov.push_back(st->buf, st->size);
+            op->iov.push_back(
+                st->aligned_buf + sizeof(shared_file_header_t) + st->offset + st->size,
+                st->aligned_size - (sizeof(shared_file_header_t) + st->offset + st->size)
+            );
+        }
+    }, st, state);
+}
+
+static void nfs_do_align_write(nfs_kv_write_state *st, uint64_t ino, uint64_t offset, uint64_t shared_alloc, int state)
+{
+    auto alignment = st->self->parent->pool_alignment;
+    uint64_t end = (offset+st->size);
+    uint8_t *good_buf = st->buf;
+    uint64_t good_offset = offset;
+    uint64_t good_size = st->size;
+    bool begin_shdr = false;
+    uint64_t end_pad = 0;
+    st->waiting++;
+    st->rmw[0].st = NULL;
+    st->rmw[1].st = NULL;
+    if (offset % alignment)
+    {
+        if (shared_alloc && st->offset == 0 && (offset % alignment) == sizeof(shared_file_header_t))
+        {
+            // RMW can be skipped at shared beginning
+            st->shdr = {
+                .magic = SHARED_FILE_MAGIC_V1,
+                .inode = st->ino,
+                .alloc = shared_alloc,
+            };
+            begin_shdr = true;
+            good_offset -= sizeof(shared_file_header_t);
+            offset = 0;
+        }
+        else
+        {
+            // Requires read-modify-write in the beginning
+            auto s = (alignment - (offset % alignment));
+            if (good_size > s)
+            {
+                good_buf += s;
+                good_offset += s;
+                good_size -= s;
+            }
+            else
+                good_size = 0;
+            s = s > st->size ? st->size : s;
+            st->rmw[0] = {
+                .st = st,
+                .continue_state = state,
+                .ino = ino,
+                .offset = offset,
+                .buf = st->buf,
+                .size = s,
+            };
+            nfs_do_rmw(&st->rmw[0]);
+        }
+    }
+    if ((end % alignment) &&
+        (offset == 0 || end/alignment > (offset-1)/alignment))
+    {
+        // Requires read-modify-write in the end
+        assert(st->offset+st->size <= st->new_size);
+        if (st->offset+st->size == st->new_size)
+        {
+            // rmw can be skipped at end - we can just zero pad the request
+            end_pad = alignment - (end % alignment);
+        }
+        else
+        {
+            auto s = (end % alignment);
+            if (good_size > s)
+                good_size -= s;
+            else
+                good_size = 0;
+            st->rmw[1] = {
+                .st = st,
+                .continue_state = state,
+                .ino = ino,
+                .offset = end - s,
+                .buf = st->buf + st->size - s,
+                .size = s,
+            };
+            nfs_do_rmw(&st->rmw[1]);
+        }
+    }
+    if (good_size > 0 || end_pad > 0 || begin_shdr)
+    {
+        // Normal write
+        nfs_do_write(ino, good_offset, (begin_shdr ? sizeof(shared_file_header_t) : 0)+good_size+end_pad, [&](cluster_op_t *op)
+        {
+            if (begin_shdr)
+                op->iov.push_back(&st->shdr, sizeof(shared_file_header_t));
+            op->iov.push_back(good_buf, good_size);
+            if (end_pad)
+                op->iov.push_back(st->self->parent->kvfs->zero_block.data(), end_pad);
+        }, st, state);
+    }
+    st->waiting--;
+    if (!st->waiting)
+    {
+        nfs_kv_continue_write(st, state);
+    }
+}
+
+static std::string new_normal_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni.erase("shared_ino");
+    ni.erase("shared_offset");
+    ni.erase("shared_alloc");
+    ni.erase("shared_ver");
+    ni["size"] = st->ext->cur_extend;
+    return json11::Json(ni).dump();
+}
+
+static std::string new_moved_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni["shared_ino"] = st->shared_inode;
+    ni["shared_offset"] = st->shared_offset;
+    ni["shared_alloc"] = st->aligned_size;
+    ni.erase("shared_ver");
+    ni["size"] = st->new_size;
+    return json11::Json(ni).dump();
+}
+
+static std::string new_shared_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni["size"] = st->new_size;
+    ni["shared_ver"] = ni["shared_ver"].uint64_value()+1;
+    return json11::Json(ni).dump();
+}
+
+static std::string new_unshared_ientry(nfs_kv_write_state *st)
+{
+    auto ni = st->ientry.object_items();
+    ni.erase("empty");
+    ni.erase("shared_ino");
+    ni.erase("shared_offset");
+    ni.erase("shared_alloc");
+    ni.erase("shared_ver");
+    return json11::Json(ni).dump();
+}
+
+static void nfs_kv_extend_inode(nfs_kv_write_state *st, int state, int base_state)
+{
+    if (state == base_state+1)
+        goto resume_1;
+    st->ext->cur_extend = st->ext->next_extend;
+    st->ext->next_extend = 0;
+    st->res2 = -EAGAIN;
+    st->self->parent->db->set(kv_inode_key(st->ino), new_normal_ientry(st), [st, base_state](int res)
+    {
+        st->res = res;
+        nfs_kv_continue_write(st, base_state+1);
+    }, [st](int res, const std::string & old_value)
+    {
+        if (res != 0)
+        {
+            return false;
+        }
+        if (old_value == st->ientry_text)
+        {
+            return true;
+        }
+        std::string err;
+        auto ientry = json11::Json::parse(old_value, err).object_items();
+        if (err != "")
+        {
+            fprintf(stderr, "Invalid JSON in inode %lu = %s: %s\n", st->ino, old_value.c_str(), err.c_str());
+            st->res2 = -EINVAL;
+            return false;
+        }
+        else if (ientry.size() == st->ientry.object_items().size())
+        {
+            for (auto & kv: st->ientry.object_items())
+            {
+                if (kv.first != "size" && ientry[kv.first] != kv.second)
+                {
+                    // Something except size changed
+                    return false;
+                }
+            }
+            // OK, only size changed
+            if (ientry["size"] >= st->new_size)
+            {
+                // Already extended
+                st->res2 = 0;
+                return false;
+            }
+            // size is different but can still be extended, other parameters don't differ
+            return true;
+        }
+        return false;
+    });
+    return;
+resume_1:
+    if (st->res == -EAGAIN)
+    {
+        // EAGAIN may be OK in fact (see above)
+        st->res = st->res2;
+    }
+    if (st->res == 0)
+    {
+        st->ext->done_extend = st->ext->cur_extend;
+    }
+    st->ext->cur_extend = 0;
+    // Wake up other extenders anyway
+    auto waiters = std::move(st->ext->waiters);
+    for (auto & cb: waiters)
+    {
+        cb();
+    }
+}
+
+// Packing small files into "shared inodes". Insane algorithm...
+// Write:
+// - If (offset+size <= threshold):
+//   - Read inode from cache
+//   - If inode does not exist - stop with -ENOENT
+//   - If inode is not a regular file - stop with -EINVAL
+//   - If it's empty (size == 0 || empty == true):
+//     - If preset size is larger than threshold:
+//       - Write data into non-shared inode
+//       - In parallel: clear empty flag
+//         - If CAS failure: re-read inode and restart
+//     - Otherwise:
+//       - Allocate/take a shared inode
+//       - Allocate space in its end
+//       - Write data into shared inode
+//         - If CAS failure: allocate another shared inode and retry
+//       - Write shared inode reference, set size
+//         - If CAS failure: free allocated shared space, re-read inode and restart
+//   - If it's not empty:
+//     - If non-shared:
+//       - Write data into non-shared inode
+//       - In parallel: check if data fits into inode size and extend if it doesn't
+//         - If CAS failure: re-read inode and retry to extend the size
+//     - If shared:
+//       - If data doesn't fit into the same shared inode:
+//         - Allocate space in a new shared inode
+//         - Read whole file from shared inode
+//         - Write data into the new shared inode
+//           - If CAS failure: allocate another shared inode and retry
+//         - Update inode metadata (set new size and new shared inode)
+//           - If CAS failure: free allocated shared space, re-read inode and restart
+//       - If it fits:
+//         - Update shared inode data in-place
+//         - Update inode entry in any case to block parallel non-shared writes
+//           - If CAS failure: re-read inode and restart
+// - Otherwise:
+//   - Read inode
+//     - If not a regular file - stop with -EINVAL
+//     - If shared:
+//       - Read whole file from shared inode
+//       - Write data into non-shared inode
+//         - If CAS failure (block should not exist): restart
+//       - Update inode metadata (make non-shared, update size)
+//         - If CAS failure: restart
+//       - Zero out the shared inode header
+//   - Write data into non-shared inode
+//   - Check if size fits
+//     - Extend if it doesn't
+// Read:
+// - If (offset+size <= threshold):
+//   - Read inode from cache
+//   - If empty: return zeroes
+//   - If shared:
+//     - Read the whole file from shared inode, or at least data and shared inode header
+//     - If the file header in data doesn't match: re-read inode and restart
+//   - If non-shared:
+//     - Read data from non-shared inode
+// - Otherwise:
+//   - Read data from non-shared inode
+
+static void nfs_kv_continue_write(nfs_kv_write_state *st, int state)
+{
+    if (state == 0)      {}
+    else if (state == 1) goto resume_1;
+    else if (state == 2) goto resume_2;
+    else if (state == 3) goto resume_3;
+    else if (state == 4) goto resume_4;
+    else if (state == 5) goto resume_5;
+    else if (state == 6) goto resume_6;
+    else if (state == 7) goto resume_7;
+    else if (state == 8) goto resume_8;
+    else if (state == 9) goto resume_9;
+    else if (state == 10) goto resume_10;
+    else if (state == 11) goto resume_11;
+    else if (state == 12) goto resume_12;
+    else if (state == 13) goto resume_13;
+    else if (state == 14) goto resume_14;
+    else if (state == 15) goto resume_15;
+    else if (state == 16) goto resume_16;
+    else
+    {
+        fprintf(stderr, "BUG: invalid state in nfs_kv_continue_write()");
+        abort();
+    }
+resume_0:
+    if (!st->size)
+    {
+        auto cb = std::move(st->cb);
+        cb(0);
+        return;
+    }
+    kv_read_inode(st->self, st->ino, [st](int res, const std::string & value, json11::Json attrs)
+    {
+        st->res = res;
+        st->ientry_text = value;
+        st->ientry = attrs;
+        nfs_kv_continue_write(st, 1);
+    }, st->allow_cache);
+    return;
+resume_1:
+    if (st->res < 0 || kv_map_type(st->ientry["type"].string_value()) != NF3REG)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res == 0 ? -EINVAL : st->res);
+        return;
+    }
+    st->was_immediate = st->self->parent->cli->get_immediate_commit(st->self->parent->fs_base_inode + st->ino);
+    st->new_size = st->ientry["size"].uint64_value();
+    if (st->new_size < st->offset + st->size)
+    {
+        st->new_size = st->offset + st->size;
+    }
+    if (st->offset + st->size + sizeof(shared_file_header_t) < st->self->parent->shared_inode_threshold)
+    {
+        if (st->ientry["size"].uint64_value() == 0 &&
+            st->ientry["shared_ino"].uint64_value() == 0 ||
+            st->ientry["empty"].bool_value() &&
+            (st->ientry["size"].uint64_value() + sizeof(shared_file_header_t)) < st->self->parent->shared_inode_threshold ||
+            st->ientry["shared_ino"].uint64_value() != 0 &&
+            st->ientry["shared_alloc"].uint64_value() < sizeof(shared_file_header_t)+st->offset+st->size)
+        {
+            // Either empty, or shared and requires moving into a larger place (redirect-write)
+            allocate_shared_inode(st, 2, align_shared_size(st->self, st->new_size));
+            return;
+resume_2:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+resume_3:
+            if (!nfs_do_shared_readmodify(st, 3, state, false))
+                return;
+            nfs_do_shared_write(st, 4, false);
+            return;
+resume_4:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+            st->self->parent->db->set(kv_inode_key(st->ino), new_moved_ientry(st), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_write(st, 5);
+            }, [st](int res, const std::string & old_value)
+            {
+                return res == 0 && old_value == st->ientry_text;
+            });
+            return;
+resume_5:
+            if (st->res < 0)
+            {
+                st->res2 = st->res;
+                memset(st->aligned_buf, 0, st->aligned_size);
+                nfs_do_shared_write(st, 6, true);
+                return;
+resume_6:
+                free(st->aligned_buf);
+                st->aligned_buf = NULL;
+                if (st->res2 == -EAGAIN)
+                {
+                    goto resume_0;
+                }
+                else
+                {
+                    auto cb = std::move(st->cb);
+                    cb(st->res2);
+                    return;
+                }
+            }
+            auto cb = std::move(st->cb);
+            cb(0);
+            return;
+        }
+        else if (st->ientry["shared_ino"].uint64_value() != 0)
+        {
+            // Non-empty, shared, can be updated in-place
+            nfs_do_align_write(st, st->ientry["shared_ino"].uint64_value(),
+                st->ientry["shared_offset"].uint64_value() + sizeof(shared_file_header_t) + st->offset,
+                st->ientry["shared_alloc"].uint64_value(), 7);
+            return;
+resume_7:
+            if (st->res == 0 && st->stable && !st->was_immediate)
+            {
+                nfs_do_fsync(st, 8);
+                return;
+            }
+resume_8:
+            // We always have to change inode entry on shared writes
+            st->self->parent->db->set(kv_inode_key(st->ino), new_shared_ientry(st), [st](int res)
+            {
+                st->res = res;
+                nfs_kv_continue_write(st, 9);
+            }, [st](int res, const std::string & old_value)
+            {
+                return res == 0 && old_value == st->ientry_text;
+            });
+            return;
+resume_9:
+            if (st->res == -EAGAIN)
+            {
+                goto resume_0;
+            }
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        // Fall through for non-shared
+    }
+    // Unshare?
+    if (st->ientry["shared_ino"].uint64_value() != 0)
+    {
+        if (st->ientry["size"].uint64_value() != 0)
+        {
+            assert(!st->aligned_buf);
+            st->aligned_size = align_shared_size(st->self, st->ientry["size"].uint64_value());
+            st->aligned_buf = (uint8_t*)malloc_or_die(st->aligned_size);
+            nfs_do_shared_read(st, 10);
+            return;
+resume_10:
+            nfs_do_unshare_write(st, 11);
+            return;
+resume_11:
+            if (st->res < 0)
+            {
+                auto cb = std::move(st->cb);
+                cb(st->res);
+                return;
+            }
+        }
+        st->self->parent->db->set(kv_inode_key(st->ino), new_unshared_ientry(st), [st](int res)
+        {
+            st->res = res;
+            nfs_kv_continue_write(st, 12);
+        }, [st](int res, const std::string & old_value)
+        {
+            return res == 0 && old_value == st->ientry_text;
+        });
+        return;
+resume_12:
+        if (st->res == -EAGAIN)
+        {
+            // Restart
+            goto resume_0;
+        }
+        if (st->res < 0)
+        {
+            auto cb = std::move(st->cb);
+            cb(st->res);
+            return;
+        }
+        st->ientry_text = new_unshared_ientry(st);
+    }
+    // Non-shared write
+    nfs_do_align_write(st, st->ino, st->offset, 0, 13);
+    return;
+resume_13:
+    if (st->res == 0 && st->stable && !st->was_immediate)
+    {
+        nfs_do_fsync(st, 14);
+        return;
+    }
+resume_14:
+    if (st->res < 0)
+    {
+        auto cb = std::move(st->cb);
+        cb(st->res);
+        return;
+    }
+    if (st->ientry["empty"].bool_value() ||
+        st->ientry["size"].uint64_value() < st->new_size ||
+        st->ientry["shared_ino"].uint64_value() != 0)
+    {
+        st->ext = &st->self->parent->kvfs->extends[st->ino];
+        st->ext->refcnt++;
+resume_15:
+        if (st->ext->next_extend < st->new_size)
+        {
+            // Aggregate inode extension requests
+            st->ext->next_extend = st->new_size;
+        }
+        if (st->ext->cur_extend > 0)
+        {
+            // Wait for current extend which is already in progress
+            st->ext->waiters.push_back([st](){ nfs_kv_continue_write(st, 15); });
+            return;
+        }
+        if (st->ext->done_extend < st->new_size)
+        {
+            nfs_kv_extend_inode(st, 15, 15);
+            return;
+resume_16:
+            nfs_kv_extend_inode(st, 16, 15);
+        }
+        st->ext->refcnt--;
+        assert(st->ext->refcnt >= 0);
+        if (st->ext->refcnt == 0)
+        {
+            st->self->parent->kvfs->extends.erase(st->ino);
+        }
+    }
+    if (st->res == -EAGAIN)
+    {
+        // Restart
+        goto resume_0;
+    }
+    auto cb = std::move(st->cb);
+    cb(st->res);
+}
+
+int kv_nfs3_write_proc(void *opaque, rpc_op_t *rop)
+{
+    nfs_kv_write_state *st = new nfs_kv_write_state;
+    st->self = (nfs_client_t*)opaque;
+    st->rop = rop;
+    WRITE3args *args = (WRITE3args*)rop->request;
+    WRITE3res *reply = (WRITE3res*)rop->reply;
+    st->ino = kv_fh_inode(args->file);
+    st->offset = args->offset;
+    st->size = (args->count > args->data.size ? args->data.size : args->count);
+    if (st->self->parent->trace)
+        fprintf(stderr, "[%d] WRITE %ju %ju+%ju\n", st->self->nfs_fd, st->ino, st->offset, st->size);
+    if (!st->ino || st->size > MAX_REQUEST_SIZE)
+    {
+        *reply = (WRITE3res){ .status = NFS3ERR_INVAL };
+        rpc_queue_reply(rop);
+        delete st;
+        return 0;
+    }
+    st->buf = (uint8_t*)args->data.data;
+    st->stable = (args->stable != UNSTABLE);
+    st->cb = [st](int res)
+    {
+        WRITE3res *reply = (WRITE3res*)st->rop->reply;
+        *reply = (WRITE3res){ .status = vitastor_nfs_map_err(res) };
+        if (res == 0)
+        {
+            reply->resok.count = (unsigned)st->size;
+            reply->resok.committed = st->stable || st->was_immediate ? FILE_SYNC : UNSTABLE;
+            *(uint64_t*)reply->resok.verf = st->self->parent->server_id;
+        }
+        rpc_queue_reply(st->rop);
+        delete st;
+    };
+    nfs_kv_continue_write(st, 0);
+    return 1;
+}

src/nfs_mount.cpp

@@ -0,0 +1,126 @@
+// 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 "nfs/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_proxy.cpp

@@ -21,6 +21,9 @@
 #include "addr_util.h"
 #include "str_util.h"
 #include "nfs_proxy.h"
+#include "nfs_kv.h"
+#include "nfs_block.h"
+#include "nfs_common.h"
 #include "http_client.h"
 #include "cli.h"
 
@@ -31,6 +34,8 @@ const char *exe_name = NULL;
 
 nfs_proxy_t::~nfs_proxy_t()
 {
+    if (db)
+        delete db;
     if (cmd)
         delete cmd;
     if (cli)
@@ -57,12 +62,14 @@ json11::Json::object nfs_proxy_t::parse_args(int narg, const char *args[])
                 "\n"
                 "USAGE:\n"
                 "  %s [STANDARD OPTIONS] [OTHER OPTIONS]\n"
+                "  --fs <META>       mount VitastorFS with metadata in image <META>\n"
                 "  --subdir <DIR>    export images prefixed <DIR>/ (default empty - export all images)\n"
                 "  --portmap 0       do not listen on port 111 (portmap/rpcbind, requires root)\n"
                 "  --bind <IP>       bind service to <IP> address (default 0.0.0.0)\n"
                 "  --nfspath <PATH>  set NFS export path to <PATH> (default is /)\n"
                 "  --port <PORT>     use port <PORT> for NFS services (default is 2049)\n"
                 "  --pool <POOL>     use <POOL> as default pool for new files (images)\n"
+                "  --logfile <FILE>  log to the specified file\n"
                 "  --foreground 1    stay in foreground, do not daemonize\n"
                 "\n"
                 "NFS proxy is stateless if you use immediate_commit=all in your cluster and if\n"
@@ -92,6 +99,9 @@ void nfs_proxy_t::run(json11::Json cfg)
     srand48(tv.tv_sec*1000000000 + tv.tv_nsec);
     server_id = (uint64_t)lrand48() | ((uint64_t)lrand48() << 31) | ((uint64_t)lrand48() << 62);
     // Parse options
+    if (cfg["logfile"].string_value() != "")
+        logfile = cfg["logfile"].string_value();
+    trace = cfg["log_level"].uint64_value() > 5 || cfg["trace"].uint64_value() > 0;
     bind_address = cfg["bind"].string_value();
     if (bind_address == "")
         bind_address = "0.0.0.0";
@@ -131,67 +141,12 @@ void nfs_proxy_t::run(json11::Json cfg)
     cmd->ringloop = ringloop;
     cmd->epmgr = epmgr;
     cmd->cli = cli;
-    // We need inode name hashes for NFS handles to remain stateless and <= 64 bytes long
-    dir_info[""] = (nfs_dir_t){
-        .id = 1,
-        .mod_rev = 0,
-    };
-    clock_gettime(CLOCK_REALTIME, &dir_info[""].mtime);
     watch_stats();
-    assert(cli->st_cli.on_inode_change_hook == NULL);
-    cli->st_cli.on_inode_change_hook = [this](inode_t changed_inode, bool removed)
+    if (!fs_kv_inode)
     {
-        auto inode_cfg_it = cli->st_cli.inode_config.find(changed_inode);
-        if (inode_cfg_it == cli->st_cli.inode_config.end())
-        {
-            return;
-        }
-        auto & inode_cfg = inode_cfg_it->second;
-        std::string full_name = inode_cfg.name;
-        if (name_prefix != "" && full_name.substr(0, name_prefix.size()) != name_prefix)
-        {
-            return;
-        }
-        // Calculate directory modification time and revision (used as "cookie verifier")
-        timespec now;
-        clock_gettime(CLOCK_REALTIME, &now);
-        dir_info[""].mod_rev = dir_info[""].mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dir_info[""].mod_rev;
-        dir_info[""].mtime = now;
-        int pos = full_name.find('/', name_prefix.size());
-        while (pos >= 0)
-        {
-            std::string dir = full_name.substr(0, pos);
-            auto & dinf = dir_info[dir];
-            if (!dinf.id)
-                dinf.id = next_dir_id++;
-            dinf.mod_rev = dinf.mod_rev < inode_cfg.mod_revision ? inode_cfg.mod_revision : dinf.mod_rev;
-            dinf.mtime = now;
-            dir_by_hash["S"+base64_encode(sha256(dir))] = dir;
-            pos = full_name.find('/', pos+1);
-        }
-        // Alter inode_by_hash
-        if (removed)
-        {
-            auto ino_it = hash_by_inode.find(changed_inode);
-            if (ino_it != hash_by_inode.end())
-            {
-                inode_by_hash.erase(ino_it->second);
-                hash_by_inode.erase(ino_it);
-            }
-        }
-        else
-        {
-            std::string hash = "S"+base64_encode(sha256(full_name));
-            auto hbi_it = hash_by_inode.find(changed_inode);
-            if (hbi_it != hash_by_inode.end() && hbi_it->second != hash)
-            {
-                // inode had a different name, remove old hash=>inode pointer
-                inode_by_hash.erase(hbi_it->second);
-            }
-            inode_by_hash[hash] = changed_inode;
-            hash_by_inode[changed_inode] = hash;
-        }
-    };
+        blockfs = new block_fs_state_t();
+        blockfs->init(this);
+    }
     // Load image metadata
     while (!cli->is_ready())
     {
@@ -202,6 +157,72 @@ void nfs_proxy_t::run(json11::Json cfg)
     }
     // Check default pool
     check_default_pool();
+    // Check if we're using VitastorFS
+    fs_kv_inode = cfg["fs"].uint64_value();
+    if (fs_kv_inode)
+    {
+        if (!INODE_POOL(fs_kv_inode))
+        {
+            fprintf(stderr, "FS metadata inode number must include pool\n");
+            exit(1);
+        }
+    }
+    else if (cfg["fs"].is_string())
+    {
+        for (auto & ic: cli->st_cli.inode_config)
+        {
+            if (ic.second.name == cfg["fs"].string_value())
+            {
+                fs_kv_inode = ic.first;
+                break;
+            }
+        }
+        if (!fs_kv_inode)
+        {
+            fprintf(stderr, "FS metadata image \"%s\" does not exist\n", cfg["fs"].string_value().c_str());
+            exit(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;
+    if (fs_kv_inode)
+    {
+        // Open DB and wait
+        int open_res = 0;
+        bool open_done = false;
+        db = new kv_dbw_t(cli);
+        db->open(fs_kv_inode, cfg, [&](int res)
+        {
+            open_done = true;
+            open_res = res;
+        });
+        while (!open_done)
+        {
+            ringloop->loop();
+            if (open_done)
+                break;
+            ringloop->wait();
+        }
+        if (open_res < 0)
+        {
+            fprintf(stderr, "Failed to open key/value filesystem metadata index: %s (code %d)\n",
+                strerror(-open_res), open_res);
+            exit(1);
+        }
+        fs_base_inode = ((uint64_t)default_pool_id << (64-POOL_ID_BITS));
+        fs_inode_count = ((uint64_t)1 << (64-POOL_ID_BITS)) - 1;
+        shared_inode_threshold = pool_block_size;
+        if (!cfg["shared_inode_threshold"].is_null())
+        {
+            shared_inode_threshold = cfg["shared_inode_threshold"].uint64_value();
+        }
+        kvfs = new kv_fs_state_t;
+        kvfs->zero_block.resize(pool_block_size);
+    }
     // Self-register portmap and NFS
     pmap.reg_ports.insert((portmap_id_t){
         .prog = PMAP_PROGRAM,
@@ -275,9 +296,13 @@ void nfs_proxy_t::run(json11::Json cfg)
     }
     // Destroy the client
     cli->flush();
+    delete kvfs;
+    delete db;
     delete cli;
     delete epmgr;
     delete ringloop;
+    kvfs = NULL;
+    db = NULL;
     cli = NULL;
     epmgr = NULL;
     ringloop = NULL;
@@ -382,8 +407,11 @@ void nfs_proxy_t::check_default_pool()
     {
         if (cli->st_cli.pool_config.size() == 1)
         {
-            default_pool = cli->st_cli.pool_config.begin()->second.name;
-            default_pool_id = cli->st_cli.pool_config.begin()->first;
+            auto pool_it = cli->st_cli.pool_config.begin();
+            default_pool_id = pool_it->first;
+            default_pool = pool_it->second.name;
+            pool_block_size = pool_it->second.pg_stripe_size;
+            pool_alignment = pool_it->second.bitmap_granularity;
         }
         else
         {
@@ -398,6 +426,8 @@ void nfs_proxy_t::check_default_pool()
             if (p.second.name == default_pool)
             {
                 default_pool_id = p.first;
+                pool_block_size = p.second.pg_stripe_size;
+                pool_alignment = p.second.bitmap_granularity;
                 break;
             }
         }
@@ -421,6 +451,10 @@ void nfs_proxy_t::do_accept(int listen_fd)
         int one = 1;
         setsockopt(nfs_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
         auto cli = new nfs_client_t();
+        if (fs_kv_inode)
+            nfs_kv_procs(cli);
+        else
+            nfs_block_procs(cli);
         cli->parent = this;
         cli->nfs_fd = nfs_fd;
         for (auto & fn: pmap.proc_table)
@@ -968,8 +1002,8 @@ void nfs_proxy_t::daemonize()
     close(1);
     close(2);
     open("/dev/null", O_RDONLY);
-    open("/dev/null", O_WRONLY);
-    open("/dev/null", O_WRONLY);
+    open(logfile.c_str(), O_WRONLY|O_APPEND|O_CREAT, 0666);
+    open(logfile.c_str(), O_WRONLY|O_APPEND|O_CREAT, 0666);
 }
 
 int main(int narg, const char *args[])

src/nfs_proxy.h

@@ -4,17 +4,18 @@
 #include "epoll_manager.h"
 #include "nfs_portmap.h"
 #include "nfs/xdr_impl.h"
+#include "kv_db.h"
 
+#define NFS_ROOT_HANDLE "R"
 #define RPC_INIT_BUF_SIZE 32768
+#define MAX_REQUEST_SIZE 128*1024*1024
+#define TRUE 1
+#define FALSE 0
 
 class cli_tool_t;
 
-struct nfs_dir_t
-{
-    uint64_t id;
-    uint64_t mod_rev;
-    timespec mtime;
-};
+struct kv_fs_state_t;
+struct block_fs_state_t;
 
 class nfs_proxy_t
 {
@@ -27,28 +28,29 @@ public:
     std::string export_root;
     bool portmap_enabled;
     unsigned nfs_port;
+    uint64_t fs_kv_inode = 0;
+    uint64_t fs_base_inode = 0;
+    uint64_t fs_inode_count = 0;
+    int readdir_getattr_parallel = 8, id_alloc_batch_size = 200;
+    int trace = 0;
+    std::string logfile = "/dev/null";
 
     pool_id_t default_pool_id;
+    uint64_t pool_block_size = 0;
+    uint64_t pool_alignment = 0;
+    uint64_t shared_inode_threshold = 0;
 
     portmap_service_t pmap;
     ring_loop_t *ringloop = NULL;
     epoll_manager_t *epmgr = NULL;
     cluster_client_t *cli = NULL;
     cli_tool_t *cmd = NULL;
+    kv_dbw_t *db = NULL;
+    kv_fs_state_t *kvfs = NULL;
+    block_fs_state_t *blockfs = NULL;
 
     std::vector<XDR*> xdr_pool;
 
-    // filehandle = "S"+base64(sha256(full name with prefix)) or "roothandle" for mount root)
-
-    uint64_t next_dir_id = 2;
-    // filehandle => dir with name_prefix
-    std::map<std::string, std::string> dir_by_hash;
-    // dir with name_prefix => dir info
-    std::map<std::string, nfs_dir_t> dir_info;
-    // filehandle => inode ID
-    std::map<std::string, inode_t> inode_by_hash;
-    // inode ID => filehandle
-    std::map<inode_t, std::string> hash_by_inode;
     // inode ID => statistics
     std::map<inode_t, json11::Json> inode_stats;
     // pool ID => statistics
@@ -86,28 +88,6 @@ struct rpc_free_buffer_t
     unsigned size;
 };
 
-struct extend_size_t
-{
-    inode_t inode;
-    uint64_t new_size;
-};
-
-inline bool operator < (const extend_size_t &a, const extend_size_t &b)
-{
-    return a.inode < b.inode || a.inode == b.inode && a.new_size < b.new_size;
-}
-
-struct extend_write_t
-{
-    rpc_op_t *rop;
-    int resize_res, write_res; // 1 = started, 0 = completed OK, -errno = completed with error
-};
-
-struct extend_inode_t
-{
-    uint64_t cur_extend = 0, next_extend = 0;
-};
-
 class nfs_client_t
 {
 public:
@@ -122,8 +102,6 @@ public:
     rpc_cur_buffer_t cur_buffer = { 0 };
     std::map<uint8_t*, rpc_used_buffer_t> used_buffers;
     std::vector<rpc_free_buffer_t> free_buffers;
-    std::map<inode_t, extend_inode_t> extends;
-    std::multimap<extend_size_t, extend_write_t> extend_writes;
 
     iovec read_iov;
     msghdr read_msg = { 0 };
@@ -133,9 +111,6 @@ public:
     std::vector<iovec> send_list, next_send_list;
     std::vector<rpc_op_t*> outbox, next_outbox;
 
-    nfs_client_t();
-    ~nfs_client_t();
-
     void select_read_buffer(unsigned wanted_size);
     void submit_read(unsigned wanted_size);
     void handle_read(int result);

src/osd.h

@@ -239,6 +239,7 @@ class osd_t
     void report_statistics();
     void report_pg_state(pg_t & pg);
     void report_pg_states();
+    void apply_no_inode_stats();
     void apply_pg_count();
     void apply_pg_config();
 

src/osd_cluster.cpp

@@ -390,7 +390,16 @@ void osd_t::on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes
     }
     if (run_primary)
     {
-        apply_pg_count();
+        bool pools = changes.find(st_cli.etcd_prefix+"/config/pools") != changes.end();
+        bool pgs = changes.find(st_cli.etcd_prefix+"/config/pgs") != changes.end();
+        if (pools)
+        {
+            apply_no_inode_stats();
+        }
+        if (pools || pgs)
+        {
+            apply_pg_count();
+        }
         apply_pg_config();
     }
 }
@@ -602,11 +611,32 @@ void osd_t::on_load_pgs_hook(bool success)
     else
     {
         peering_state &= ~OSD_LOADING_PGS;
-        apply_pg_count();
-        apply_pg_config();
+        if (run_primary)
+        {
+            apply_no_inode_stats();
+            apply_pg_count();
+            apply_pg_config();
+        }
     }
 }
 
+void osd_t::apply_no_inode_stats()
+{
+    if (!bs)
+    {
+        return;
+    }
+    std::vector<uint64_t> no_inode_stats;
+    for (auto & pool_item: st_cli.pool_config)
+    {
+        if (pool_item.second.no_inode_stats)
+        {
+            no_inode_stats.push_back(pool_item.first);
+        }
+    }
+    bs->set_no_inode_stats(no_inode_stats);
+}
+
 void osd_t::apply_pg_count()
 {
     for (auto & pool_item: st_cli.pool_config)

tests/run_tests.sh

@@ -68,3 +68,5 @@ SCHEME=xor ./test_scrub.sh
 PG_SIZE=3 ./test_scrub.sh
 PG_SIZE=6 PG_MINSIZE=4 OSD_COUNT=6 SCHEME=ec ./test_scrub.sh
 SCHEME=ec ./test_scrub.sh
+
+./test_nfs.sh

tests/test_nfs.sh

@@ -0,0 +1,167 @@
+#!/bin/bash -ex
+
+PG_COUNT=16
+. `dirname $0`/run_3osds.sh
+
+build/src/vitastor-cli --etcd_address $ETCD_URL create -s 10G fsmeta
+build/src/vitastor-nfs --fs fsmeta --etcd_address $ETCD_URL --portmap 0 --port 2050 --foreground 1 --trace 1 >>./testdata/nfs.log 2>&1 &
+NFS_PID=$!
+
+mkdir -p testdata/nfs
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+MNT=$(pwd)/testdata/nfs
+trap "sudo umount -f $MNT"' || true; kill -9 $(jobs -p)' EXIT
+
+# write small file
+ls -l ./testdata/nfs
+dd if=/dev/urandom of=./testdata/f1 bs=100k count=1
+cp testdata/f1 ./testdata/nfs/
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep f1
+diff ./testdata/f1 ./testdata/nfs/f1
+format_green "100K file ok"
+
+# overwrite it inplace
+dd if=/dev/urandom of=./testdata/f1_90k bs=90k count=1
+cp testdata/f1_90k ./testdata/nfs/f1
+sudo umount ./testdata/nfs/
+format_green "inplace overwrite 90K ok"
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep f1
+# create another copy
+dd if=./testdata/f1_90k of=./testdata/nfs/f1_nfs bs=1M
+diff ./testdata/f1_90k ./testdata/nfs/f1_nfs
+sudo umount ./testdata/nfs/
+format_green "another copy 90K ok"
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep f1
+cp ./testdata/nfs/f1 ./testdata/f1_nfs
+diff ./testdata/f1_90k ./testdata/nfs/f1
+format_green "90K data ok"
+
+# test partial shared overwrite
+dd if=/dev/urandom of=./testdata/f1_90k bs=9317 count=1 seek=5 conv=notrunc
+dd if=./testdata/f1_90k of=./testdata/nfs/f1 bs=9317 count=1 skip=5 seek=5 conv=notrunc
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+diff ./testdata/f1_90k ./testdata/nfs/f1
+format_green "partial inplace shared overwrite ok"
+
+# move it to a larger shared space
+dd if=/dev/urandom of=./testdata/f1_110k bs=110k count=1
+cp testdata/f1_110k ./testdata/nfs/f1
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep f1
+diff ./testdata/f1_110k ./testdata/nfs/f1
+format_green "move shared 90K -> 110K ok"
+
+# extend it to large file + rm
+dd if=/dev/urandom of=./testdata/f1_2M bs=2M count=1
+cp ./testdata/f1_2M ./testdata/nfs/f1
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep f1
+cp ./testdata/nfs/f1 ./testdata/f1_nfs
+diff ./testdata/f1_2M ./testdata/nfs/f1
+rm ./testdata/nfs/f1
+format_green "extend to 2M + rm ok"
+
+# mkdir
+mkdir -p ./testdata/nfs/dir1/dir2
+echo abcdef > ./testdata/nfs/dir1/dir2/hnpfls
+# rename dir
+mv ./testdata/nfs/dir1 ./testdata/nfs/dir3
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep dir3
+ls -l ./testdata/nfs/dir3 | grep dir2
+ls -l ./testdata/nfs/dir3/dir2 | grep hnpfls
+echo abcdef > ./testdata/hnpfls
+diff ./testdata/hnpfls ./testdata/nfs/dir3/dir2/hnpfls
+format_green "rename dir with file ok"
+
+# touch
+touch -t 202401011404 ./testdata/nfs/dir3/dir2/hnpfls
+sudo chown 65534:65534 ./testdata/nfs/dir3/dir2/hnpfls
+sudo chmod 755 ./testdata/nfs/dir3/dir2/hnpfls
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+T=`stat -c '%a %u %g %y' ./testdata/nfs/dir3/dir2/hnpfls | perl -pe 's/(:\d+)(.*)/$1/'`
+[[ "$T" = "755 65534 65534 2024-01-01 14:04" ]]
+format_green "set attrs ok"
+
+# move dir
+mv ./testdata/nfs/dir3/dir2 ./testdata/nfs/
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+ls -l ./testdata/nfs | grep dir3
+ls -l ./testdata/nfs | grep dir2
+format_green "move dir ok"
+
+# symlink, readlink
+ln -s dir2 ./testdata/nfs/sym2
+[[ "`stat -c '%A' ./testdata/nfs/sym2`" = "lrwxrwxrwx" ]]
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+[[ "`stat -c '%A' ./testdata/nfs/sym2`" = "lrwxrwxrwx" ]]
+[[ "`readlink ./testdata/nfs/sym2`" = "dir2" ]]
+format_green "symlink, readlink ok"
+
+# mknod: chr, blk, sock, fifo + remove
+sudo mknod ./testdata/nfs/nod_chr c 1 5
+sudo mknod ./testdata/nfs/nod_blk b 2 6
+mkfifo ./testdata/nfs/nod_fifo
+perl -e 'use Socket; socket($sock, PF_UNIX, SOCK_STREAM, undef) || die $!; bind($sock, sockaddr_un("./testdata/nfs/nod_sock")) || die $!;'
+chmod 777 ./testdata/nfs/nod_*
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+[[ "`ls testdata|wc -l`" -ge 4 ]]
+[[ "`stat -c '%A' ./testdata/nfs/nod_blk`" = "brwxrwxrwx" ]]
+[[ "`stat -c '%A' ./testdata/nfs/nod_chr`" = "crwxrwxrwx" ]]
+[[ "`stat -c '%A' ./testdata/nfs/nod_fifo`" = "prwxrwxrwx" ]]
+[[ "`stat -c '%A' ./testdata/nfs/nod_sock`" = "srwxrwxrwx" ]]
+sudo rm ./testdata/nfs/nod_*
+format_green "mknod + rm ok"
+
+# hardlink
+echo ABCDEF > ./testdata/nfs/linked1
+i=`stat -c '%i' ./testdata/nfs/linked1`
+ln ./testdata/nfs/linked1 ./testdata/nfs/linked2
+[[ "`stat -c '%i' ./testdata/nfs/linked2`" -eq $i ]]
+echo BABABA > ./testdata/nfs/linked2
+diff ./testdata/nfs/linked2 ./testdata/nfs/linked1
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+diff ./testdata/nfs/linked2 ./testdata/nfs/linked1
+[[ "`cat ./testdata/nfs/linked2`" = "BABABA" ]]
+rm ./testdata/nfs/linked2
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+[[ "`cat ./testdata/nfs/linked1`" = "BABABA" ]]
+format_green "hardlink ok"
+
+# rm small
+ls -l ./testdata/nfs
+dd if=/dev/urandom of=./testdata/nfs/smallfile bs=100k count=1
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+rm ./testdata/nfs/smallfile
+if ls ./testdata/nfs | grep smallfile; then false; fi
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+if ls ./testdata/nfs | grep smallfile; then false; fi
+format_green "rm small ok"
+
+# rename over existing
+echo ZXCVBN > ./testdata/nfs/over1
+mv ./testdata/nfs/over1 ./testdata/nfs/linked2
+sudo umount ./testdata/nfs/
+sudo mount localhost:/ ./testdata/nfs -o port=2050,mountport=2050,nfsvers=3,soft,nolock,tcp
+if ls ./testdata/nfs | grep over1; then false; fi
+[[ "`cat ./testdata/nfs/linked2`" = "ZXCVBN" ]]
+[[ "`cat ./testdata/nfs/linked1`" = "BABABA" ]]
+format_green "rename over existing file ok"
+
+format_green OK