WIP NFS RDMA support

src/CMakeLists.txt

@@ -61,6 +61,10 @@ pkg_check_modules(ISAL libisal)
 if (ISAL_LIBRARIES)
 	add_definitions(-DWITH_ISAL)
 endif (ISAL_LIBRARIES)
+pkg_check_modules(RDMACM librdmacm)
+if (RDMACM_LIBRARIES)
+	add_definitions(-DWITH_RDMACM)
+endif (RDMACM_LIBRARIES)
 
 add_custom_target(build_tests)
 add_custom_target(test

src/nfs/nfs_block.cpp

@@ -315,8 +315,7 @@ static int block_nfs3_read_proc(void *opaque, rpc_op_t *rop)
     if (aligned_count % alignment)
         aligned_count = aligned_count + alignment - (aligned_count % alignment);
     aligned_count -= aligned_offset;
-    void *buf = malloc_or_die(aligned_count);
-    xdr_add_malloc(rop->xdrs, buf);
+    void *buf = self->malloc_or_rdma(rop, aligned_count);
     cluster_op_t *op = new cluster_op_t;
     op->opcode = OSD_OP_READ;
     op->inode = ino_it->second;

src/nfs/nfs_proxy.cpp

@@ -499,6 +499,20 @@ void nfs_proxy_t::check_default_pool()
     }
 }
 
+nfs_client_t *nfs_proxy_t::create_client()
+{
+    auto cli = new nfs_client_t();
+    cli->parent = this;
+    if (kvfs)
+        nfs_kv_procs(cli);
+    else
+        nfs_block_procs(cli);
+    for (auto & fn: pmap.proc_table)
+        cli->proc_table.insert(fn);
+    rpc_clients.insert(cli);
+    return cli;
+}
+
 void nfs_proxy_t::do_accept(int listen_fd)
 {
     struct sockaddr_storage addr;
@@ -512,18 +526,8 @@ void nfs_proxy_t::do_accept(int listen_fd)
         fcntl(nfs_fd, F_SETFL, fcntl(nfs_fd, F_GETFL, 0) | O_NONBLOCK);
         int one = 1;
         setsockopt(nfs_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-        auto cli = new nfs_client_t();
-        if (kvfs)
-            nfs_kv_procs(cli);
-        else
-            nfs_block_procs(cli);
-        cli->parent = this;
+        auto cli = this->create_client();
         cli->nfs_fd = nfs_fd;
-        for (auto & fn: pmap.proc_table)
-        {
-            cli->proc_table.insert(fn);
-        }
-        rpc_clients[nfs_fd] = cli;
         epmgr->tfd->set_fd_handler(nfs_fd, true, [cli](int nfs_fd, int epoll_events)
         {
             // Handle incoming event
@@ -781,7 +785,7 @@ void nfs_client_t::stop()
     if (refs <= 0)
     {
         auto parent = this->parent;
-        parent->rpc_clients.erase(nfs_fd);
+        parent->rpc_clients.erase(this);
         parent->active_connections--;
         parent->epmgr->tfd->set_fd_handler(nfs_fd, true, NULL);
         close(nfs_fd);
@@ -885,6 +889,8 @@ void rpc_queue_reply(rpc_op_t *rop)
         r = xdr_encode(rop->xdrs, rop->reply_fn, rop->reply);
         assert(r);
     }
+    if (!self->rdma_conn)
+    {
         xdr_encode_finish(rop->xdrs, &iov_list, &iov_count);
         assert(iov_count > 0);
         rop->reply_marker = 0;
@@ -904,6 +910,11 @@ void rpc_queue_reply(rpc_op_t *rop)
         }
         to_outbox[to_outbox.size()-1] = rop;
         self->submit_send();
+    }
+    else
+    {
+        self->rdma_queue_reply(rop);
+    }
 }
 
 int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg_len)
@@ -968,6 +979,18 @@ int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg
         // Incoming buffer isn't needed to handle request, so return 0
         return 0;
     }
+    auto rop = create_rpc_op(xdrs, inmsg, NULL);
+    if (!rop)
+    {
+        // No such procedure
+        return 0;
+    }
+    rop->buffer = (uint8_t*)base_buf;
+    return handle_rpc_op(rop);
+}
+
+rpc_op_t *nfs_client_t::create_rpc_op(XDR *xdrs, rpc_msg *inmsg, rdma_msg *rmsg)
+{
     // Find decoder for the request
     auto proc_it = proc_table.find((rpc_service_proc_t){
         .prog = inmsg->body.cbody.prog,
@@ -1019,7 +1042,7 @@ int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg
         };
         rpc_queue_reply(rop);
         // Incoming buffer isn't needed to handle request, so return 0
-        return 0;
+        return NULL;
     }
     // Allocate memory
     rpc_op_t *rop = (rpc_op_t*)malloc_or_die(
@@ -1028,7 +1051,6 @@ int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg
     rpc_reply_stat x = RPC_MSG_ACCEPTED;
     *rop = (rpc_op_t){
         .client = this,
-        .buffer = (uint8_t*)base_buf,
         .xdrs = xdrs,
         .out_msg = (rpc_msg){
             .xid = inmsg->xid,
@@ -1045,10 +1067,25 @@ int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg
         .request = ((uint8_t*)rop) + sizeof(rpc_op_t),
         .reply = ((uint8_t*)rop) + sizeof(rpc_op_t) + proc_it->req_size,
     };
+    // FIXME: malloc and avoid copy?
     memcpy(&rop->in_msg, inmsg, sizeof(rpc_msg));
+    if (rmsg)
+    {
+        memcpy(&rop->in_rdma_msg, rmsg, sizeof(rdma_msg));
+    }
+    return rop;
+}
+
+int nfs_client_t::handle_rpc_op(rpc_op_t *rop)
+{
     // Try to decode the request
     // req_fn may be NULL, that means function has no arguments
-    if (proc_it->req_fn && !proc_it->req_fn(xdrs, rop->request))
+    auto proc_it = proc_table.find((rpc_service_proc_t){
+        .prog = rop->in_msg.body.cbody.prog,
+        .vers = rop->in_msg.body.cbody.vers,
+        .proc = rop->in_msg.body.cbody.proc,
+    });
+    if (proc_it == proc_table.end() || proc_it->req_fn && !proc_it->req_fn(rop->xdrs, rop->request))
     {
         // Invalid request
         rop->out_msg.body.rbody.areply.reply_data.stat = RPC_GARBAGE_ARGS;
@@ -1059,17 +1096,31 @@ int nfs_client_t::handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg
     rop->out_msg.body.rbody.areply.reply_data.stat = RPC_SUCCESS;
     rop->reply_fn = proc_it->resp_fn;
     int ref = proc_it->handler_fn(proc_it->opaque, rop);
-    rop->referenced = ref ? 1 : 0;
+    if (ref)
+        rop->referenced = 1;
     return ref;
 }
 
+void *nfs_client_t::malloc_or_rdma(rpc_op_t *rop, size_t size)
+{
+    if (!rdma_conn)
+    {
+        void *buf = malloc_or_die(size);
+        xdr_add_malloc(rop->xdrs, buf);
+        return buf;
+    }
+    void *buf = rdma_malloc(size);
+    rop->rdma_buf = buf;
+    return buf;
+}
+
 void nfs_proxy_t::daemonize()
 {
     // Stop all clients because client I/O sometimes breaks during daemonize
     // I.e. the new process stops receiving events on the old FD
     // It doesn't happen if we call sleep(1) here, but we don't want to call sleep(1)...
-    for (auto & clp: rpc_clients)
-        clp.second->stop();
+    for (auto & cli: rpc_clients)
+        cli->stop();
     if (fork())
         exit(0);
     setsid();

src/nfs/nfs_proxy.h

@@ -55,7 +55,7 @@ public:
     vitastorkv_dbw_t *db = NULL;
     kv_fs_state_t *kvfs = NULL;
     block_fs_state_t *blockfs = NULL;
-    std::map<int, nfs_client_t*> rpc_clients;
+    std::set<nfs_client_t*> rpc_clients;
 
     std::vector<XDR*> xdr_pool;
 
@@ -72,6 +72,7 @@ public:
     void watch_stats();
     void parse_stats(etcd_kv_t & kv);
     void check_default_pool();
+    nfs_client_t* create_client();
     void do_accept(int listen_fd);
     void daemonize();
     void write_pid();
@@ -101,15 +102,20 @@ struct rpc_free_buffer_t
     unsigned size;
 };
 
+struct nfs_rdma_conn_t;
+
 class nfs_client_t
 {
 public:
     nfs_proxy_t *parent = NULL;
-    int nfs_fd;
-    int epoll_events = 0;
     int refs = 0;
     bool stopped = false;
     std::set<rpc_service_proc_t> proc_table;
+    nfs_rdma_conn_t *rdma_conn = NULL;
+
+    // <TCP>
+    int nfs_fd;
+    int epoll_events = 0;
 
     // Read state
     rpc_cur_buffer_t cur_buffer = { 0 };
@@ -130,7 +136,14 @@ public:
     void submit_send();
     void handle_send(int result);
     int handle_rpc_message(void *base_buf, void *msg_buf, uint32_t msg_len);
+    // </TCP>
 
+    rpc_op_t *create_rpc_op(XDR *xdrs, rpc_msg *inmsg, rdma_msg *rmsg);
+    int handle_rpc_op(rpc_op_t *rop);
     bool deref();
     void stop();
+    void *malloc_or_rdma(rpc_op_t *rop, size_t size);
+    void *rdma_malloc(size_t size);
+    void rdma_free(void *ptr);
+    void rdma_queue_reply(rpc_op_t *rop);
 };

src/nfs/nfs_proxy_rdma.cpp

@@ -0,0 +1,726 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// NFS RDMA support
+
+#define _XOPEN_SOURCE
+#include <limits.h>
+
+#include <netinet/tcp.h>
+#include <sys/epoll.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <signal.h>
+
+#include "proto/nfs.h"
+#include "proto/rpc.h"
+#include "proto/portmap.h"
+
+#include "addr_util.h"
+#include "str_util.h"
+#include "json_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"
+
+#include "rdma_alloc.h"
+
+#define NFS_RDMACM_PRIVATE_DATA_MAGIC_LE 0x180eabf6
+
+struct __attribute__((__packed__)) nfs_rdmacm_private
+{
+    uint32_t format_identifier; // magic, should be 0xf6ab0e18 in big endian
+    uint8_t version;            // version, 1
+    uint8_t remote_invalidate;  // remote invalidation flag (1 or 0)
+    uint8_t max_send_size;      // maximum RDMA Send operation size / 1024 - 1 (i.e. 0 is 1 KB, 255 is 256 KB)
+    uint8_t max_recv_size;      // maximum RDMA Receive operation size / 1024 - 1 (i.e. 0 is 1 KB, 255 is 256 KB)
+};
+
+struct nfs_rdma_buf_t
+{
+    void *buf = NULL;
+    size_t len = 0;
+    ibv_mr *mr = NULL;
+};
+
+struct nfs_rdma_context_t
+{
+    std::string bind_address;
+    int rdmacm_port = 0;
+    int max_send = 8, max_recv = 8; --- FIXME max_send and max_recv should probably be equal
+    uint64_t rdma_malloc_round_to = 1048576, rdma_max_unused_buffers = 500*1048576;
+    uint64_t max_send_size = 256*1024, max_recv_size = 256*1024;
+
+    nfs_proxy_t *proxy = NULL;
+    epoll_manager_t *epmgr = NULL;
+
+    int max_cqe = 0, used_max_cqe = 0;
+    rdma_event_channel *rdmacm_evch = NULL;
+    rdma_cm_id *listener_id = NULL;
+    ibv_comp_channel *channel = NULL;
+    ibv_cq *cq = NULL;
+    rdma_allocator_t *alloc = NULL;
+};
+
+struct nfs_rdma_conn_t
+{
+    nfs_client_t *client = NULL;
+    rdma_cm_id *id = NULL;
+    int max_send = 8, max_recv = 8;
+    bool established = false;
+    std::vector<nfs_rdma_buf_t> recv_buffers;
+    int next_recv_buf = 0;
+    std::vector<nfs_rdma_buf_t> send_buffers;
+    std::vector<rpc_op_t*> outbox;
+    int outbox_pos = 0;
+};
+
+nfs_rdma_context_t* nfs_proxy_t::create_rdma(const std::string & bind_address, int rdmacm_port)
+{
+    nfs_rdma_context_t* self = new nfs_rdma_context_t;
+    self->proxy = this;
+    self->epmgr = epmgr;
+    self->bind_address = bind_address;
+    self->rdmacm_port = rdmacm_port;
+    self->rdmacm_evch = rdma_create_event_channel();
+    if (!self->rdmacm_evch)
+    {
+        fprintf(stderr, "Failed to initialize RDMA-CM event channel: %s (code %d)\n", strerror(errno), errno);
+        delete self;
+        return NULL;
+    }
+    fcntl(self->rdmacm_evch->fd, F_SETFL, fcntl(self->rdmacm_evch->fd, F_GETFL, 0) | O_NONBLOCK);
+    epmgr->tfd->set_fd_handler(self->rdmacm_evch->fd, false, [this](int rdmacm_eventfd, int epoll_events)
+    {
+        self->handle_rdmacm_events();
+    });
+    int r = rdma_create_id(self->rdmacm_evch, &self->listener_id, NULL, RDMA_PS_TCP);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to create RDMA-CM ID: %s (code %d)\n", strerror(errno), errno);
+        delete self;
+        return NULL;
+    }
+    sockaddr_storage addr;
+    if (!string_to_addr(bind_address, 0, rdmacm_port, &addr))
+    {
+        fprintf(stderr, "Server address: %s is not valid\n", bind_address.c_str());
+        delete self;
+        return NULL;
+    }
+    r = rdma_bind_addr(self->listener_id, (sockaddr*)&addr);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to bind RDMA-CM to %s:%d: %s (code %d)\n", bind_address.c_str(), rdmacm_port, strerror(errno), errno);
+        delete self;
+        return NULL;
+    }
+    r = rdma_listen(self->listener_id, 128);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to listen RDMA-CM: %s (code %d)\n", strerror(errno), errno);
+        delete self;
+        return NULL;
+    }
+    self->channel = ibv_create_comp_channel(self->listener_id->ibv_context);
+    if (!self->channel)
+    {
+        fprintf(stderr, "Couldn't create RDMA completion channel\n");
+        delete self;
+        return NULL;
+    }
+    fcntl(self->channel->fd, F_SETFL, fcntl(self->channel->fd, F_GETFL, 0) | O_NONBLOCK);
+    epmgr->tfd->set_fd_handler(self->channel->fd, false, [this](int channel_eventfd, int epoll_events)
+    {
+        handle_io();
+    });
+    self->max_cqe = 4096;
+    self->cq = ibv_create_cq(self->listener_id->ibv_context, self->max_cqe, NULL, self->channel, 0);
+    if (!self->cq)
+    {
+        fprintf(stderr, "Couldn't create RDMA completion queue\n");
+        delete self;
+        return NULL;
+    }
+    self->alloc = rdma_malloc_create(self->listener_id->pd, rdma_malloc_round_to, rdma_max_unused_buffers, IBV_ACCESS_LOCAL_WRITE);
+    return self;
+}
+
+nfs_rdma_context_t::~nfs_rdma_context_t()
+{
+    if (listener_id)
+    {
+        int r = rdma_destroy_id(listener_id);
+        if (r != 0)
+            fprintf(stderr, "Failed to destroy RDMA-CM ID: %s (code %d)\n", strerror(errno), errno);
+        else
+            listener_id = NULL;
+    }
+    if (rdmacm_evch)
+    {
+        epmgr->tfd->set_fd_handler(rdmacm_evch->fd, false, NULL);
+        rdma_destroy_event_channel(rdmacm_evch);
+        rdmacm_evch = NULL;
+    }
+    if (cq)
+    {
+        ibv_destroy_cq(cq);
+        cq = NULL;
+    }
+    if (channel)
+    {
+        ibv_destroy_comp_channel(channel);
+        channel = NULL;
+    }
+    if (alloc)
+    {
+        rdma_malloc_destroy(alloc);
+        alloc = NULL;
+    }
+    //if (mr)
+    //    ibv_dereg_mr(mr);
+    //if (pd)
+    //    ibv_dealloc_pd(pd);
+    //if (context)
+    //    ibv_close_device(context);
+}
+
+void nfs_proxy_t::handle_rdmacm_events()
+{
+    rdma_cm_event *ev = NULL;
+    while (1)
+    {
+        int r = rdma_get_cm_event(rdmacm_evch, &ev);
+        if (r != 0)
+        {
+            if (errno == EAGAIN || errno == EINTR)
+                break;
+            fprintf(stderr, "Failed to get RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
+            exit(1);
+        }
+        if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
+        {
+            rdmacm_accept(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR ||
+            ev->event == RDMA_CM_EVENT_REJECTED ||
+            ev->event == RDMA_CM_EVENT_DISCONNECTED ||
+            ev->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
+        {
+            auto event_type_name = ev->event == RDMA_CM_EVENT_CONNECT_ERROR ? "RDMA_CM_EVENT_CONNECT_ERROR" : (
+                ev->event == RDMA_CM_EVENT_REJECTED ? "RDMA_CM_EVENT_REJECTED" : (
+                ev->event == RDMA_CM_EVENT_DISCONNECTED ? "RDMA_CM_EVENT_DISCONNECTED" : "RDMA_CM_EVENT_DEVICE_REMOVAL"));
+            auto conn_it = rdma_connections.find(ev->id);
+            if (conn_it == rdma_connections.end())
+            {
+                fprintf(stderr, "Received %s event for an unknown connection 0x%lx - ignoring\n",
+                    event_type_name, (uint64_t)ev->id);
+            }
+            else
+            {
+                fprintf(stderr, "Received %s event for connection 0x%lx - closing it\n",
+                    event_type_name, (uint64_t)ev->id);
+                auto conn = conn_it->second;
+                delete conn;
+            }
+        }
+        else if (ev->event == RDMA_CM_EVENT_ESTABLISHED)
+        {
+            rdmacm_established(ev);
+        }
+        else if (ev->event == RDMA_CM_EVENT_ADDR_CHANGE || ev->event == RDMA_CM_EVENT_TIMEWAIT_EXIT)
+        {
+            // Do nothing
+        }
+        else
+        {
+            // Other events are unexpected
+            fprintf(stderr, "Unexpected RDMA-CM event type: %d\n", ev->event);
+        }
+        r = rdma_ack_cm_event(ev);
+        if (r != 0)
+        {
+            fprintf(stderr, "Failed to ack (free) RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
+            exit(1);
+        }
+    }
+}
+
+void nfs_rdma_context_t::rdmacm_accept(rdma_cm_event *ev)
+{
+    ctx->used_max_cqe += max_send+max_recv;
+    if (ctx->used_max_cqe > ctx->max_cqe)
+    {
+        // Resize CQ
+        int new_max_cqe = ctx->max_cqe;
+        while (ctx->used_max_cqe > new_max_cqe)
+        {
+            new_max_cqe *= 2;
+        }
+        if (ibv_resize_cq(ctx->cq, new_max_cqe) != 0)
+        {
+            fprintf(stderr, "Couldn't resize RDMA completion queue to %d entries\n", new_max_cqe);
+            delete conn;
+            return NULL;
+        }
+        ctx->max_cqe = new_max_cqe;
+    }
+    ibv_qp_init_attr init_attr = {
+        .send_cq = ctx->cq,
+        .recv_cq = ctx->cq,
+        .cap     = {
+            .max_send_wr  = max_send*2, // ?????? тут большой вопрос сколько на самом деле
+            .max_recv_wr  = max_recv,
+            .max_send_sge = max_sge,
+            .max_recv_sge = max_sge,
+        },
+        .qp_type = IBV_QPT_RC,
+    };
+    r = rdma_create_qp(ev->id, NULL, &init_attr);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to create a queue pair via RDMA-CM: %s (code %d)\n", strerror(errno), errno);
+        exit(1);
+    }
+    nfs_rdmacm_private private_data = {
+        .format_identifier = NFS_RDMACM_PRIVATE_DATA_MAGIC_LE,
+        .version = 1,
+        .remote_invalidate = ?,
+        .max_send_size = (max_send_size <= 256*1024 ? max_send_size/1024 - 1 : 255),
+        .max_recv_size = (max_recv_size <= 256*1024 ? max_recv_size/1024 - 1 : 255),
+    };
+    rdma_conn_param conn_params = {
+        .private_data = &private_data,
+        .private_data_len = sizeof(private_data),
+        //.responder_resources = max_qp_rd_atom of the device?,
+        //.initiator_depth = max_qp_init_rd_atom of the device?,
+        .rnr_retry_count = 7,
+        //.qp_num = manually created QP number?,
+    };
+    r = rdma_accept(ev->id, &conn_params);
+    if (r != 0)
+    {
+        fprintf(stderr, "Failed to accept RDMA-CM connection: %s (code %d)\n", strerror(errno), errno);
+        rdma_destroy_qp(ev->id);
+        rdma_destroy_id(ev->id);
+    }
+    else
+    {
+        auto conn = new nfs_rdma_conn_t();
+        conn->id = ctx->id;
+        rdma_connections[ctx->id] = conn;
+        rdma_connections_by_qp[conn->id->qp->qp_num];
+        auto cli = this->proxy->create_client();
+        conn->client = cli;
+        cli->rdma_conn = conn;
+    }
+}
+
+nfs_rdma_conn_t::~nfs_rdma_conn_t()
+{
+    if (id)
+    {
+        parent->rdma_connections.erase(id);
+        if (id->qp)
+        {
+            parent->rdma_connections_by_qp.erase(id->qp->qp_num);
+            rdma_destroy_qp(id);
+        }
+        rdma_destroy_id(id);
+    }
+}
+
+void nfs_rdma_context_t::rdmacm_established(rdma_cm_event *ev)
+{
+    auto conn_it = rdma_connections.find(ev->id);
+    if (conn_it == rdma_connections.end())
+    {
+        fprintf(stderr, "Received RDMA_CM_EVENT_ESTABLISHED event for an unknown connection 0x%lx - ignoring\n", (uint64_t)ev->id);
+        return;
+    }
+    fprintf(stderr, "Received RDMA_CM_EVENT_ESTABLISHED event for connection 0x%lx - connection established\n", (uint64_t)ev->id);
+    auto conn = conn_it->second;
+    conn->established = true;
+    // Handle NFS private_data
+    if (ev->private_data_len >= sizeof(nfs_rdmacm_private))
+    {
+        nfs_rdmacm_private *private_data = (nfs_rdmacm_private *)ev->private_data;
+        if (private_data->format_identifier == NFS_RDMACM_PRIVATE_DATA_MAGIC_LE &&
+            private_data->version == 1)
+        {
+            conn->remote_invalidate = private_data->remote_invalidate;
+            conn->remote_max_send = (private_data->max_send_size+1) * 1024;
+            conn->remote_max_recv = (private_data->max_recv_size+1) * 1024;
+            if (conn->remote_max_recv < conn->max_send)
+                conn->max_send = conn->remote_max_recv;
+        }
+    }
+    // Post initial receive requests
+    conn->post_initial_receives();
+}
+
+void nfs_rdma_conn_t::post_initial_receives()
+{
+    while (cur_recv < max_recv)
+    {
+        auto b = create_buf(max_recv_size);
+        recv_buffers.push_back(b);
+        post_recv(b);
+    }
+}
+
+nfs_rdma_buf_t nfs_rdma_conn_t::create_buf(size_t len)
+{
+    nfs_rdma_buf_t b;
+    b.buf = malloc_or_die(len);
+    b.len = len;
+    b.mr = ibv_reg_mr(id->pd, b.buf, len, IBV_ACCESS_LOCAL_WRITE);
+    if (!b.mr)
+    {
+        fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno));
+        exit(1);
+    }
+    return b;
+}
+
+void nfs_rdma_conn_t::post_recv(nfs_rdma_buf_t b)
+{
+    ibv_sge sge = {
+        .addr = (uintptr_t)b.buf,
+        .length = (uint32_t)b.len,
+        .lkey = b.mr->lkey,
+    };
+    ibv_recv_wr *bad_wr = NULL;
+    ibv_recv_wr wr = {
+        .wr_id   = 1, // 1 is any read, 2 is any write :)
+        .sg_list = &sge,
+        .num_sge = 1,
+    };
+    int err = ibv_post_recv(id->qp, &wr, &bad_wr);
+    if (err || bad_wr)
+    {
+        fprintf(stderr, "RDMA receive failed: %s\n", strerror(err));
+        exit(1);
+    }
+    cur_recv++;
+}
+
+void nfs_client_t::rdma_queue_reply(rpc_op_t *rop)
+{
+    rdma_conn->outbox.push_back(rop);
+    rdma_conn->post_send();
+}
+
+void nfs_rdma_conn_t::post_send()
+{
+    while (outbox.size() > outbox_pos)
+    {
+        auto rop = outbox[outbox_pos];
+        // Check that exactly 1 write chunk is provided for READ3 and READLINK3
+        if (rop->in_msg.body.cbody.prog == NFS_PROGRAM &&
+            (rop->in_msg.body.cbody.proc == NFS3_READ || rop->in_msg.body.cbody.proc == NFS3_READLINK) &&
+            (!rop->in_rdma_msg.rdma_body.rdma_msg.rdma_writes ||
+            rop->in_rdma_msg.rdma_body.rdma_msg.rdma_writes->next))
+        {
+            rop->rdma_error = ERR_CHUNK;
+            ...
+        }
+        iovec *iov_list = NULL;
+        unsigned iov_count = 0;
+        xdr_encode_finish(rop->xdrs, &iov_list, &iov_count);
+        assert(iov_count > 0);
+        nfs_rdma_buf_t b;
+        if (send_buffers.size())
+        {
+            b = send_buffers.back();
+            send_buffers.pop_back();
+        }
+        else
+        {
+            b = create_buf(max_send_size);
+        }
+        // READ3resok and READLINK3resok - extract last byte buffer from iovecs and send it in a "write chunk"
+        iovec *chunk_iov = NULL;
+        if (rop->in_msg.body.cbody.prog == NFS_PROGRAM &&
+            (rop->in_msg.body.cbody.proc == NFS3_READ && ((READ3res*)rop->reply)->status == NFS3_OK ||
+            rop->in_msg.body.cbody.proc == NFS3_READLINK && ((READLINK3res*)rop->reply)->status == NFS3_OK))
+        {
+            assert(iov_count > 1);
+            iov_count--;
+            chunk_iov = &iov_list[iov_count];
+        }
+        // FIXME: Avoid extra copy - to do that we have to initially encode into nfs_rdma_buf_t
+        size_t pos = 0;
+        for (unsigned i = 0; i < iov_count; i++)
+        {
+            assert(pos + iov_list[i].iov_len <= b.len);
+            memcpy(b.buf + pos, iov_list[i].iov_base, iov_list[i].iov_len);
+            pos += iov_list[i].iov_len;
+        }
+        ibv_sge chunk_sge;
+        ibv_send_wr chunk_wr;
+        ibv_sge sge = {
+            .addr = (uintptr_t)b.buf,
+            .length = (uint32_t)pos,
+            .lkey = b.mr->lkey,
+        };
+        ibv_send_wr *bad_wr = NULL;
+        ibv_send_wr wr = {
+            .wr_id = 2, // 2 is send
+            .sg_list = &sge,
+            .num_sge = 1,
+            .opcode = IBV_WR_SEND,
+            .send_flags = IBV_SEND_SIGNALED,
+        };
+        ibv_send_wr *send_wr = &wr;
+        if (chunk_iov != NULL)
+        {
+            auto & wr_chunk = rop->in_rdma_msg.rdma_body.rdma_msg.rdma_writes->entry.target;
+            chunk_sge = {
+                .addr = (uintptr_t)chunk_iov->iov_base,
+                .length = (uint32_t)chunk_iov->iov_len,
+                .lkey = parent->get_rdma_data_lkey(chunk_iov->iov_base),
+            };
+            chunk_wr = {
+                .wr_id = 4, // 2 is chunk write
+                .sg_list = &chunk_sge,
+                .num_sge = 1,
+                .opcode = IBV_WR_RDMA_WRITE,
+                .wr = {
+                    .rdma = {
+                        .remote_addr = wr_chunk.offset,
+                        .rkey = wr_chunk.handle,
+                    },
+                },
+            };
+            // send chunk_wr first, then normal wr
+            chunk_wr.next = &wr;
+            send_wr = &chunk_wr;
+        }
+        int err = ibv_post_send(id->qp, send_wr, &bad_wr);
+        if (err || bad_wr)
+        {
+            fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
+            exit(1);
+        }
+        cl->rdma_conn->cur_send++;
+    }
+}
+
+#define RDMA_EVENTS_AT_ONCE 32
+
+void nfs_rdma_context_t::handle_io()
+{
+    // Request next notification
+    ibv_cq *ev_cq;
+    void *ev_ctx;
+    // FIXME: This is inefficient as it calls read()...
+    if (ibv_get_cq_event(channel, &ev_cq, &ev_ctx) == 0)
+    {
+        ibv_ack_cq_events(cq, 1);
+    }
+    if (ibv_req_notify_cq(cq, 0) != 0)
+    {
+        fprintf(stderr, "Failed to request RDMA completion notification, exiting\n");
+        exit(1);
+    }
+    ibv_wc wc[RDMA_EVENTS_AT_ONCE];
+    int event_count;
+    do
+    {
+        event_count = ibv_poll_cq(cq, RDMA_EVENTS_AT_ONCE, wc);
+        for (int i = 0; i < event_count; i++)
+        {
+            auto conn_it = rdma_connections_by_qp.find(wc[i].qp_num);
+            if (conn_it == rdma_connections_by_qp.end())
+            {
+                continue;
+            }
+            auto conn = conn_it->second;
+            if (wc[i].status != IBV_WC_SUCCESS)
+            {
+                fprintf(stderr, "RDMA work request failed for queue %d with status: %s, stopping client\n", wc[i].qp_num, ibv_wc_status_str(wc[i].status));
+                delete conn;
+                continue;
+            }
+            //auto read_buf_it = conn->buffers.find(wc[i].wr_id);
+            //auto read_buf = read_buf_it != conn->buffers.end() ? read_buf_it->second : NULL;
+            auto is_send = wc[i].wr_id == 2;//conn->sends.at(wc[i].wr_id);
+            if (!is_send)
+            {
+                conn->cur_recv--;
+                auto & b = conn->recv_buffers[conn->next_recv_buf];
+                auto is_continued = conn->handle_recv(b.buf, wc[i].byte_len);
+                if (is_continued)
+                {
+                    // Buffer is required to handle request
+                    // Due to the credit-based flow control in RPC-RDMA, we can just remove that buffer and reuse it later
+                    used_buffers[b.buf] = b;
+                    conn->recv_buffers.erase(conn->recv_buffers.begin()+conn->next_recv_buf, conn->recv_buffers.begin()+conn->next_recv_buf+1);
+                }
+                else
+                {
+                    // Buffer is not required to handle request and can be reused immediately
+                    conn->post_recv(b);
+                    conn->next_recv_buf = (conn->next_recv_buf+1) % conn->recv_buffers.size();
+                }
+            }
+            else
+            {
+                auto rop = conn->outbox[0];
+                conn->outbox.erase(conn->outbox.begin(), conn->outbox.begin()+1);
+                xdr_reset(rop->xdrs);
+                parent->xdr_pool.push_back(rop->xdrs);
+                if (rop->buffer && rop->referenced)
+                {
+                    // Reuse the buffer
+                    auto & ub = conn->used_buffers.at(rop->buffer);
+                    conn->recv_buffers.push_back(ub);
+                    conn->post_recv(ub);
+                }
+                if (rop->rdma_buf)
+                {
+                    rdma_malloc_free(alloc, rop->rdma_buf);
+                    rop->rdma_buf = NULL;
+                }
+                free(rop);
+                conn->post_send();
+            }
+        }
+    } while (event_count > 0);
+}
+
+// returns false if handling is done, returns true if handling is continued asynchronously
+bool nfs_rdma_conn_t::handle_recv(uint8_t *buf, size_t len)
+{
+    // Take an XDR object from the pool
+    XDR *xdrs;
+    if (parent->xdr_pool.size())
+    {
+        xdrs = parent->xdr_pool.back();
+        parent->xdr_pool.pop_back();
+    }
+    else
+    {
+        xdrs = xdr_create();
+    }
+    // Decode the RDMA-RPC header
+    rdma_msg rmsg;
+    if (!xdr_decode(xdrs, buf, len, (xdrproc_t)xdr_rdma_msg, &rmsg))
+    {
+        // Invalid message, ignore it
+        xdr_reset(xdrs);
+        parent->xdr_pool.push_back(xdrs);
+        return 0;
+    }
+    if (rmsg.rdma_vers != 1 || rmsg.rdma_body.proc != RDMA_MSG /*&& rmsg.rdma_body.proc != RDMA_NOMSG*/)
+    {
+        // Bad RDMA-RPC version
+        rpc_op_t *rop = (rpc_op_t*)malloc_or_die(sizeof(rpc_op_t));
+        *rop = (rpc_op_t){
+            .client = this,
+            .xdrs = xdrs,
+            .rdma_error = ERR_VERS,
+/*
+        uint32_t x = 1;
+            .out_rdma_msg = (rdma_msg){
+                .rdma_xid = rmsg.rdma_xid,
+                .rdma_vers = rmsg.rdma_vers,
+                .rdma_credit = rmsg.rdma_credit,
+                .rdma_body = (rdma_body){
+                    .proc = RDMA_ERROR,
+                    .rdma_error = (rpc_rdma_error){
+                        .err = ERR_VERS,
+                        .range = (rpc_rdma_errvers){
+                            .rdma_vers_low = x,
+                            .rdma_vers_high = x,
+                        },
+                    },
+                },
+            },
+*/
+        };
+        rpc_queue_reply(rop);
+        // Incoming buffer isn't needed to handle request, so return 0
+        return 0;
+    }
+    rpc_msg inmsg = { .xid = rmsg.rdma_xid };
+    if (!xdr_rpc_msg_body(xdrs, &inmsg.body) || inmsg.body.dir != RPC_CALL)
+    {
+        // Invalid message, ignore it
+        xdr_reset(xdrs);
+        parent->xdr_pool.push_back(xdrs);
+        return 0;
+    }
+    // Check that exactly 1 read chunk is provided for WRITE3 and SYMLINK3
+    if (inmsg.body.cbody.prog == NFS_PROGRAM &&
+        (inmsg.body.cbody.proc == NFS3_WRITE || inmsg.body.cbody.proc == NFS3_SYMLINK) &&
+        (!rmsg.rdma_body.rdma_msg.rdma_reads || rmsg.rdma_body.rdma_msg.rdma_reads->next))
+    {
+        rpc_op_t *rop = (rpc_op_t*)malloc_or_die(sizeof(rpc_op_t));
+        *rop = (rpc_op_t){
+            .client = this,
+            .xdrs = xdrs,
+            .rdma_error = ERR_CHUNK,
+        };
+        rpc_queue_reply(rop);
+        return 0;
+    }
+    // Read that chunk
+    rpc_op_t *rop = client->create_rpc_op(xdrs, inmsg);
+    if (!rop)
+    {
+        // No such procedure
+        return 0;
+    }
+    rop->buffer = buf;
+    if (inmsg.body.cbody.prog == NFS_PROGRAM && inmsg.body.cbody.proc == NFS3_WRITE)
+    {
+        auto & rd_chunk = rmsg.rdma_body.rdma_msg.rdma_reads->entry.target;
+        void *buf = rdma_malloc_alloc(ctx->alloc, rd_chunk.length);
+        ibv_sge chunk_sge = {
+            .addr = (uintptr_t)buf,
+            .length = rd_chunk.length,
+            .lkey = parent->get_rdma_data_lkey(buf),
+        };
+        ibv_recv_wr *bad_wr = NULL;
+        ibv_recv_wr wr = {
+            .wr_id = 3, // 3 is chunk read
+            .sg_list = &chunk_sge,
+            .num_sge = 1,
+            .opcode = IBV_WR_RDMA_READ,
+            .wr = {
+                .rdma = {
+                    .remote_addr = rd_chunk.offset,
+                    .rkey = rd_chunk.handle,
+                },
+            },
+        };
+        int err = ibv_post_recv(id->qp, &wr, &bad_wr);
+        if (err || bad_wr)
+        {
+            fprintf(stderr, "RDMA receive failed: %s\n", strerror(err));
+            exit(1);
+        }
+        cur_recv++;
+        rop->rdma_buf = buf; // вот это надо присрать
+        rop->referenced = 1;
+        chunk_inbox.push_back(rop);
+        return 1;
+    }
+    return client->handle_rpc_op(rop);
+}
+
+void *nfs_client_t::rdma_malloc(size_t size)
+{
+    return rdma_malloc_alloc(parent->rdma_context->alloc, size);
+}
+
+void *nfs_client_t::rdma_free(void *ptr)
+{
+    return rdma_malloc_free(parent->rdma_context->alloc, ptr);
+}

src/nfs/proto/nfs.x

@@ -168,7 +168,7 @@ struct WRITE3args {
 	offset3     offset;
 	count3      count;
 	stable_how  stable;
-	opaque      data<>;
+	opaque      data<>; /* RDMA DDP-eligible */
 };
 
 typedef opaque writeverf3[NFS3_WRITEVERFSIZE];
@@ -409,7 +409,7 @@ struct READ3resok {
 	post_op_attr   file_attributes;
 	count3         count;
 	bool           eof;
-	opaque         data<>;
+	opaque         data<>; /* RDMA DDP-eligible */
 };
 
 struct READ3resfail {
@@ -514,7 +514,7 @@ typedef string nfspath3<>;
 
 struct symlinkdata3 {
 	sattr3    symlink_attributes;
-	nfspath3  symlink_data;
+	nfspath3  symlink_data; /* RDMA DDP-eligible */
 };
 
 struct SYMLINK3args {
@@ -546,7 +546,7 @@ struct READLINK3args {
 
 struct READLINK3resok {
 	post_op_attr   symlink_attributes;
-	nfspath3       data;
+	nfspath3       data; /* RDMA DDP-eligible */
 };
 
 struct READLINK3resfail {

src/nfs/proto/rpc_impl.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "rpc.h"
+#include "rpc_rdma.h"
 
 struct rpc_op_t;
 
@@ -27,14 +28,18 @@ inline bool operator < (const rpc_service_proc_t & a, const rpc_service_proc_t &
     return a.prog < b.prog || a.prog == b.prog && (a.vers < b.vers || a.vers == b.vers && a.proc < b.proc);
 }
 
+struct rdma_msg;
+
 struct rpc_op_t
 {
     void *client;
     uint8_t *buffer;
     XDR *xdrs;
     rpc_msg in_msg, out_msg;
+    rdma_msg in_rdma_msg;
     void *request;
     void *reply;
+    void *rdma_buf = NULL;
     xdrproc_t reply_fn;
     uint32_t reply_marker;
     bool referenced;

src/nfs/proto/rpc_rdma.h

@@ -0,0 +1,144 @@
+/*
+ * Please do not edit this file.
+ * It was generated using rpcgen.
+ */
+
+#ifndef _RPC_RDMA_H_RPCGEN
+#define _RPC_RDMA_H_RPCGEN
+
+#include "xdr_impl.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+struct xdr_rdma_segment {
+	uint32_t handle;
+	uint32_t length;
+	uint64_t offset;
+};
+typedef struct xdr_rdma_segment xdr_rdma_segment;
+
+struct xdr_read_chunk {
+	uint32_t position;
+	struct xdr_rdma_segment target;
+};
+typedef struct xdr_read_chunk xdr_read_chunk;
+
+struct xdr_read_list {
+	struct xdr_read_chunk entry;
+	struct xdr_read_list *next;
+};
+typedef struct xdr_read_list xdr_read_list;
+
+struct xdr_write_chunk {
+	struct {
+		u_int target_len;
+		struct xdr_rdma_segment *target_val;
+	} target;
+};
+typedef struct xdr_write_chunk xdr_write_chunk;
+
+struct xdr_write_list {
+	struct xdr_write_chunk entry;
+	struct xdr_write_list *next;
+};
+typedef struct xdr_write_list xdr_write_list;
+
+struct rpc_rdma_header {
+	struct xdr_read_list *rdma_reads;
+	struct xdr_write_list *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+};
+typedef struct rpc_rdma_header rpc_rdma_header;
+
+struct rpc_rdma_header_nomsg {
+	struct xdr_read_list *rdma_reads;
+	struct xdr_write_list *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+};
+typedef struct rpc_rdma_header_nomsg rpc_rdma_header_nomsg;
+
+struct rpc_rdma_header_padded {
+	uint32_t rdma_align;
+	uint32_t rdma_thresh;
+	struct xdr_read_list *rdma_reads;
+	struct xdr_write_list *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+};
+typedef struct rpc_rdma_header_padded rpc_rdma_header_padded;
+
+enum rpc_rdma_errcode {
+	ERR_VERS = 1,
+	ERR_CHUNK = 2,
+};
+typedef enum rpc_rdma_errcode rpc_rdma_errcode;
+
+struct rpc_rdma_errvers {
+	uint32_t rdma_vers_low;
+	uint32_t rdma_vers_high;
+};
+typedef struct rpc_rdma_errvers rpc_rdma_errvers;
+
+struct rpc_rdma_error {
+	rpc_rdma_errcode err;
+	union {
+		rpc_rdma_errvers range;
+	};
+};
+typedef struct rpc_rdma_error rpc_rdma_error;
+
+enum rdma_proc {
+	RDMA_MSG = 0,
+	RDMA_NOMSG = 1,
+	RDMA_MSGP = 2,
+	RDMA_DONE = 3,
+	RDMA_ERROR = 4,
+};
+typedef enum rdma_proc rdma_proc;
+
+struct rdma_body {
+	rdma_proc proc;
+	union {
+		rpc_rdma_header rdma_msg;
+		rpc_rdma_header_nomsg rdma_nomsg;
+		rpc_rdma_header_padded rdma_msgp;
+		rpc_rdma_error rdma_error;
+	};
+};
+typedef struct rdma_body rdma_body;
+
+struct rdma_msg {
+	uint32_t rdma_xid;
+	uint32_t rdma_vers;
+	uint32_t rdma_credit;
+	rdma_body rdma_body;
+};
+typedef struct rdma_msg rdma_msg;
+
+/* the xdr functions */
+
+
+extern  bool_t xdr_xdr_rdma_segment (XDR *, xdr_rdma_segment*);
+extern  bool_t xdr_xdr_read_chunk (XDR *, xdr_read_chunk*);
+extern  bool_t xdr_xdr_read_list (XDR *, xdr_read_list*);
+extern  bool_t xdr_xdr_write_chunk (XDR *, xdr_write_chunk*);
+extern  bool_t xdr_xdr_write_list (XDR *, xdr_write_list*);
+extern  bool_t xdr_rpc_rdma_header (XDR *, rpc_rdma_header*);
+extern  bool_t xdr_rpc_rdma_header_nomsg (XDR *, rpc_rdma_header_nomsg*);
+extern  bool_t xdr_rpc_rdma_header_padded (XDR *, rpc_rdma_header_padded*);
+extern  bool_t xdr_rpc_rdma_errcode (XDR *, rpc_rdma_errcode*);
+extern  bool_t xdr_rpc_rdma_errvers (XDR *, rpc_rdma_errvers*);
+extern  bool_t xdr_rpc_rdma_error (XDR *, rpc_rdma_error*);
+extern  bool_t xdr_rdma_proc (XDR *, rdma_proc*);
+extern  bool_t xdr_rdma_body (XDR *, rdma_body*);
+extern  bool_t xdr_rdma_msg (XDR *, rdma_msg*);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* !_RPC_RDMA_H_RPCGEN */

src/nfs/proto/rpc_rdma.x

@@ -0,0 +1,166 @@
+/* RFC 8166 - Remote Direct Memory Access Transport for Remote Procedure Call Version 1 */
+
+/*
+ * Copyright (c) 2010-2017 IETF Trust and the persons
+ * identified as authors of the code.  All rights reserved.
+ *
+ * The authors of the code are:
+ * B. Callaghan, T. Talpey, and C. Lever
+ *
+ * Redistribution and use in source and binary forms, with
+ * or without modification, are permitted provided that the
+ * following conditions are met:
+ *
+ * - Redistributions of source code must retain the above
+ *   copyright notice, this list of conditions and the
+ *   following disclaimer.
+ *
+ * - Redistributions in binary form must reproduce the above
+ *   copyright notice, this list of conditions and the
+ *   following disclaimer in the documentation and/or other
+ *   materials provided with the distribution.
+ *
+ * - Neither the name of Internet Society, IETF or IETF
+ *   Trust, nor the names of specific contributors, may be
+ *   used to endorse or promote products derived from this
+ *   software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS
+ *   AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED
+ *   WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ *   IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ *   FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO
+ *   EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+ *   LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ *   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
+ *   NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ *   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ *   INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ *   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ *   OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
+ *   IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
+ *   ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/*
+ * Plain RDMA segment (Section 3.4.3)
+ */
+struct xdr_rdma_segment {
+	uint32_t handle;           /* Registered memory handle */
+	uint32_t length;           /* Length of the chunk in bytes */
+	uint64_t offset;           /* Chunk virtual address or offset */
+};
+
+/*
+ * RDMA read segment (Section 3.4.5)
+ */
+struct xdr_read_chunk {
+	uint32_t position;        /* Position in XDR stream */
+	struct xdr_rdma_segment target;
+};
+
+/*
+ * Read list (Section 4.3.1)
+ */
+struct xdr_read_list {
+	struct xdr_read_chunk entry;
+	struct xdr_read_list  *next;
+};
+
+/*
+ * Write chunk (Section 3.4.6)
+ */
+struct xdr_write_chunk {
+	struct xdr_rdma_segment target<>;
+};
+
+/*
+ * Write list (Section 4.3.2)
+ */
+struct xdr_write_list {
+	struct xdr_write_chunk entry;
+	struct xdr_write_list  *next;
+};
+
+/*
+ * Chunk lists (Section 4.3)
+ */
+struct rpc_rdma_header {
+	struct xdr_read_list   *rdma_reads;
+	struct xdr_write_list  *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+	/* rpc body follows */
+};
+
+struct rpc_rdma_header_nomsg {
+	struct xdr_read_list   *rdma_reads;
+	struct xdr_write_list  *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+};
+
+/* Not to be used */
+struct rpc_rdma_header_padded {
+	uint32_t               rdma_align;
+	uint32_t               rdma_thresh;
+	struct xdr_read_list   *rdma_reads;
+	struct xdr_write_list  *rdma_writes;
+	struct xdr_write_chunk *rdma_reply;
+	/* rpc body follows */
+};
+
+/*
+ * Error handling (Section 4.5)
+ */
+enum rpc_rdma_errcode {
+	ERR_VERS = 1,       /* Value fixed for all versions */
+	ERR_CHUNK = 2
+};
+
+/* Structure fixed for all versions */
+struct rpc_rdma_errvers {
+	uint32_t rdma_vers_low;
+	uint32_t rdma_vers_high;
+};
+
+union rpc_rdma_error switch (rpc_rdma_errcode err) {
+	case ERR_VERS:
+		rpc_rdma_errvers range;
+	case ERR_CHUNK:
+		void;
+};
+
+/*
+ * Procedures (Section 4.2.4)
+ */
+enum rdma_proc {
+	RDMA_MSG = 0,     /* Value fixed for all versions */
+	RDMA_NOMSG = 1,   /* Value fixed for all versions */
+	RDMA_MSGP = 2,    /* Not to be used */
+	RDMA_DONE = 3,    /* Not to be used */
+	RDMA_ERROR = 4    /* Value fixed for all versions */
+};
+
+/* The position of the proc discriminator field is
+ * fixed for all versions */
+union rdma_body switch (rdma_proc proc) {
+	case RDMA_MSG:
+		rpc_rdma_header rdma_msg;
+	case RDMA_NOMSG:
+		rpc_rdma_header_nomsg rdma_nomsg;
+	case RDMA_MSGP:   /* Not to be used */
+		rpc_rdma_header_padded rdma_msgp;
+	case RDMA_DONE:   /* Not to be used */
+		void;
+	case RDMA_ERROR:
+		rpc_rdma_error rdma_error;
+};
+
+/*
+ * Fixed header fields (Section 4.2)
+ */
+struct rdma_msg {
+	uint32_t  rdma_xid;      /* Position fixed for all versions */
+	uint32_t  rdma_vers;     /* Position fixed for all versions */
+	uint32_t  rdma_credit;   /* Position fixed for all versions */
+	rdma_body rdma_body;
+};

src/nfs/proto/rpc_rdma_xdr.cpp

@@ -0,0 +1,200 @@
+/*
+ * Please do not edit this file.
+ * It was generated using rpcgen.
+ */
+
+#include "rpc_rdma.h"
+#include "xdr_impl_inline.h"
+
+bool_t
+xdr_xdr_rdma_segment (XDR *xdrs, xdr_rdma_segment *objp)
+{
+	
+	 if (!xdr_uint32_t (xdrs, &objp->handle))
+		 return FALSE;
+	 if (!xdr_uint32_t (xdrs, &objp->length))
+		 return FALSE;
+	 if (!xdr_uint64_t (xdrs, &objp->offset))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_xdr_read_chunk (XDR *xdrs, xdr_read_chunk *objp)
+{
+	
+	 if (!xdr_uint32_t (xdrs, &objp->position))
+		 return FALSE;
+	 if (!xdr_xdr_rdma_segment (xdrs, &objp->target))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_xdr_read_list (XDR *xdrs, xdr_read_list *objp)
+{
+	
+	 if (!xdr_xdr_read_chunk (xdrs, &objp->entry))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->next, sizeof (xdr_read_list), (xdrproc_t) xdr_xdr_read_list))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_xdr_write_chunk (XDR *xdrs, xdr_write_chunk *objp)
+{
+	
+	 if (!xdr_array (xdrs, (char **)&objp->target.target_val, (u_int *) &objp->target.target_len, ~0,
+		sizeof (xdr_rdma_segment), (xdrproc_t) xdr_xdr_rdma_segment))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_xdr_write_list (XDR *xdrs, xdr_write_list *objp)
+{
+	
+	 if (!xdr_xdr_write_chunk (xdrs, &objp->entry))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->next, sizeof (xdr_write_list), (xdrproc_t) xdr_xdr_write_list))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_header (XDR *xdrs, rpc_rdma_header *objp)
+{
+	
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reads, sizeof (xdr_read_list), (xdrproc_t) xdr_xdr_read_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_writes, sizeof (xdr_write_list), (xdrproc_t) xdr_xdr_write_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reply, sizeof (xdr_write_chunk), (xdrproc_t) xdr_xdr_write_chunk))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_header_nomsg (XDR *xdrs, rpc_rdma_header_nomsg *objp)
+{
+	
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reads, sizeof (xdr_read_list), (xdrproc_t) xdr_xdr_read_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_writes, sizeof (xdr_write_list), (xdrproc_t) xdr_xdr_write_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reply, sizeof (xdr_write_chunk), (xdrproc_t) xdr_xdr_write_chunk))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_header_padded (XDR *xdrs, rpc_rdma_header_padded *objp)
+{
+	
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_align))
+		 return FALSE;
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_thresh))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reads, sizeof (xdr_read_list), (xdrproc_t) xdr_xdr_read_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_writes, sizeof (xdr_write_list), (xdrproc_t) xdr_xdr_write_list))
+		 return FALSE;
+	 if (!xdr_pointer (xdrs, (char **)&objp->rdma_reply, sizeof (xdr_write_chunk), (xdrproc_t) xdr_xdr_write_chunk))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_errcode (XDR *xdrs, rpc_rdma_errcode *objp)
+{
+	
+	 if (!xdr_enum (xdrs, (enum_t *) objp))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_errvers (XDR *xdrs, rpc_rdma_errvers *objp)
+{
+	
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_vers_low))
+		 return FALSE;
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_vers_high))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rpc_rdma_error (XDR *xdrs, rpc_rdma_error *objp)
+{
+	
+	 if (!xdr_rpc_rdma_errcode (xdrs, &objp->err))
+		 return FALSE;
+	switch (objp->err) {
+	case ERR_VERS:
+		 if (!xdr_rpc_rdma_errvers (xdrs, &objp->range))
+			 return FALSE;
+		break;
+	case ERR_CHUNK:
+		break;
+	default:
+		return FALSE;
+	}
+	return TRUE;
+}
+
+bool_t
+xdr_rdma_proc (XDR *xdrs, rdma_proc *objp)
+{
+	
+	 if (!xdr_enum (xdrs, (enum_t *) objp))
+		 return FALSE;
+	return TRUE;
+}
+
+bool_t
+xdr_rdma_body (XDR *xdrs, rdma_body *objp)
+{
+	
+	 if (!xdr_rdma_proc (xdrs, &objp->proc))
+		 return FALSE;
+	switch (objp->proc) {
+	case RDMA_MSG:
+		 if (!xdr_rpc_rdma_header (xdrs, &objp->rdma_msg))
+			 return FALSE;
+		break;
+	case RDMA_NOMSG:
+		 if (!xdr_rpc_rdma_header_nomsg (xdrs, &objp->rdma_nomsg))
+			 return FALSE;
+		break;
+	case RDMA_MSGP:
+		 if (!xdr_rpc_rdma_header_padded (xdrs, &objp->rdma_msgp))
+			 return FALSE;
+		break;
+	case RDMA_DONE:
+		break;
+	case RDMA_ERROR:
+		 if (!xdr_rpc_rdma_error (xdrs, &objp->rdma_error))
+			 return FALSE;
+		break;
+	default:
+		return FALSE;
+	}
+	return TRUE;
+}
+
+bool_t
+xdr_rdma_msg (XDR *xdrs, rdma_msg *objp)
+{
+	
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_xid))
+		 return FALSE;
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_vers))
+		 return FALSE;
+	 if (!xdr_uint32_t (xdrs, &objp->rdma_credit))
+		 return FALSE;
+	 if (!xdr_rdma_body (xdrs, &objp->rdma_body))
+		 return FALSE;
+	return TRUE;
+}

src/nfs/proto/run-rpcgen.sh

@@ -46,3 +46,4 @@ run_rpcgen() {
 run_rpcgen nfs
 run_rpcgen rpc
 run_rpcgen portmap
+run_rpcgen rpc_rdma

src/nfs/proto/xdr_impl_inline.h

@@ -28,6 +28,19 @@
 // RPC over TCP:
 //
 // BE 32bit length, then rpc_msg, then the procedure message itself
+//
+// RPC over RDMA:
+// RFC 8166 - Remote Direct Memory Access Transport for Remote Procedure Call Version 1
+// RFC 8267 - Network File System (NFS) Upper-Layer Binding to RPC-over-RDMA Version 1
+// RFC 8797 - Remote Direct Memory Access - Connection Manager (RDMA-CM) Private Data for RPC-over-RDMA Version 1
+// message is received in an RDMA Receive operation
+// message: list of read chunks, list of write chunks, optional reply write chunk, then actual RPC body if present
+// read chunk: BE 32bit position, BE 32bit registered memory key, BE 32bit length, BE 64bit offset
+// write chunk: BE 32bit registered memory key, BE 32bit length, BE 64bit offset
+// in reality for NFS 3.0: only 1 read chunk in write3 and symlink3, only 1 write chunk in read3 and readlink3
+// read chunk is read by the server using RDMA Read from the client memory after receiving RPC request
+// write chunk is pushed by the server using RDMA Write to the client memory before sending RPC reply
+// connection is established using RDMA-CM at default port 20049
 
 #pragma once
 
@@ -61,6 +74,7 @@ struct xdr_linked_list_t
 struct XDR
 {
     int x_op;
+    bool rdma = false;
 
     // For decoding:
     uint8_t *buf = NULL;
@@ -106,13 +120,19 @@ inline int xdr_opaque(XDR *xdrs, void *data, uint32_t len)
     return 1;
 }
 
-inline int xdr_bytes(XDR *xdrs, xdr_string_t *data, uint32_t maxlen)
+inline int xdr_bytes(XDR *xdrs, xdr_string_t *data, uint32_t maxlen, bool rdma_chunk = false)
 {
     if (xdrs->x_op == XDR_DECODE)
     {
         if (xdrs->avail < 4)
             return 0;
         uint32_t len = be32toh(*((uint32_t*)xdrs->buf));
+        if (rdma_chunk && xdrs->rdma)
+        {
+            // Skip RDMA chunks while decoding
+            data->size = len;
+            return 1;
+        }
         uint32_t padded = len_pad4(len);
         if (xdrs->avail < 4+padded)
             return 0;
@@ -123,7 +143,8 @@ inline int xdr_bytes(XDR *xdrs, xdr_string_t *data, uint32_t maxlen)
     }
     else
     {
-        if (data->size < XDR_COPY_LENGTH)
+        // Always encode RDMA chunks as separate iovecs
+        if (data->size < XDR_COPY_LENGTH && (!rdma_chunk || !xdrs->rdma))
         {
             unsigned old = xdrs->cur_out.size();
             xdrs->cur_out.resize(old + 4+data->size);
@@ -146,8 +167,9 @@ inline int xdr_bytes(XDR *xdrs, xdr_string_t *data, uint32_t maxlen)
                 .iov_len = data->size,
             });
         }
-        if (data->size & 3)
+        if ((data->size & 3) && (!rdma_chunk || !xdrs->rdma))
         {
+            // No padding for RDMA chunks
             int pad = 4-(data->size & 3);
             unsigned old = xdrs->cur_out.size();
             xdrs->cur_out.resize(old+pad);
@@ -158,9 +180,9 @@ inline int xdr_bytes(XDR *xdrs, xdr_string_t *data, uint32_t maxlen)
     return 1;
 }
 
-inline int xdr_string(XDR *xdrs, xdr_string_t *data, uint32_t maxlen)
+inline int xdr_string(XDR *xdrs, xdr_string_t *data, uint32_t maxlen, bool rdma_chunk = false)
 {
-    return xdr_bytes(xdrs, data, maxlen);
+    return xdr_bytes(xdrs, data, maxlen, rdma_chunk);
 }
 
 inline int xdr_u_int(XDR *xdrs, void *data)

src/nfs/rdma_alloc.cpp

@@ -0,0 +1,192 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Simple & stupid RDMA-enabled memory allocator (allocates buffers within ibv_mr's)
+
+#include <stdio.h>
+#include <assert.h>
+#include <map>
+#include <set>
+#include "rdma_alloc.h"
+#include "malloc_or_die.h"
+
+struct rdma_buf_t
+{
+    void *buf = NULL;
+    size_t len = 0;
+    ibv_mr *mr = NULL;
+};
+
+struct rdma_frag_t
+{
+    rdma_buf_t *buf = NULL;
+    size_t len = 0;
+    bool is_free = false;
+};
+
+struct rdma_allocator_t
+{
+    size_t rdma_alloc_size = 1048576;
+    size_t rdma_max_unused = 500*1048576;
+    int rdma_access = IBV_ACCESS_LOCAL_WRITE;
+    ibv_pd *pd = NULL;
+
+    std::set<rdma_buf_t*> buffers;
+    std::map<void*, rdma_frag_t> bufferfrags;
+    std::set<std::pair<size_t, void*>> freelist;
+    size_t freebuffers = 0;
+};
+
+rdma_allocator_t *rdma_malloc_create(ibv_pd *pd, size_t rdma_alloc_size, size_t rdma_max_unused, int rdma_access)
+{
+    rdma_allocator_t *self = (rdma_allocator_t*)malloc_or_die(sizeof(rdma_allocator_t));
+    self->pd = pd;
+    self->rdma_alloc_size = rdma_alloc_size ? rdma_alloc_size : 1048576;
+    self->rdma_max_unused = rdma_max_unused ? rdma_max_unused : 500*1048576;
+    self->rdma_access = rdma_access;
+    return self;
+}
+
+static void rdma_malloc_free_unused_buffers(rdma_allocator_t *self, size_t max_unused, bool force)
+{
+    auto free_it = self->freelist.end();
+    if (free_it == self->freelist.begin())
+        return;
+    free_it--;
+    do
+    {
+        auto frag_it = self->bufferfrags.find(free_it->second);
+        assert(frag_it != self->bufferfrags.end());
+        if (frag_it->second.len != frag_it->second.buf->len)
+        {
+            if (force)
+            {
+                fprintf(stderr, "BUG: Attempt to destroy RDMA allocator while buffers are not freed yet\n");
+                abort();
+            }
+            break;
+        }
+        self->freebuffers -= frag_it->second.buf->len;
+        ibv_dereg_mr(frag_it->second.buf->mr);
+        free(frag_it->second.buf);
+        self->buffers.erase(frag_it->second.buf);
+        self->bufferfrags.erase(frag_it);
+        self->freelist.erase(free_it--);
+    } while (free_it != self->freelist.begin() && self->freebuffers > max_unused);
+}
+
+void rdma_malloc_destroy(rdma_allocator_t *self)
+{
+    rdma_malloc_free_unused_buffers(self, 0, true);
+    assert(!self->freebuffers);
+    assert(!self->buffers.size());
+    assert(!self->bufferfrags.size());
+    assert(!self->freelist.size());
+    free(self);
+}
+
+void *rdma_malloc_alloc(rdma_allocator_t *self, size_t size)
+{
+    auto it = self->freelist.lower_bound(std::pair<size_t, void*>(size, 0));
+    if (it == self->freelist.end())
+    {
+        // round size up to rdma_malloc_size (1 MB)
+        size_t alloc_size = ((size + self->rdma_alloc_size - 1) / self->rdma_alloc_size) * self->rdma_alloc_size;
+        rdma_buf_t *b = (rdma_buf_t*)malloc_or_die(alloc_size + sizeof(rdma_buf_t));
+        b->buf = b+1;
+        b->len = alloc_size;
+        b->mr = ibv_reg_mr(self->pd, b->buf, b->len, self->rdma_access);
+        if (!b->mr)
+        {
+            fprintf(stderr, "Failed to register RDMA memory region: %s\n", strerror(errno));
+            exit(1);
+        }
+        self->buffers.insert(b);
+        self->bufferfrags[b->buf] = (rdma_frag_t){ .buf = b, .len = alloc_size, .is_free = true };
+        it = self->freelist.insert(std::pair<size_t, void*>(alloc_size, b->buf)).first;
+    }
+    void *ptr = it->second;
+    auto & frag = self->bufferfrags.at(ptr);
+    assert(frag.len >= size && frag.is_free);
+    if (frag.len == size)
+    {
+        frag.is_free = false;
+        self->freelist.erase(it);
+    }
+    else
+    {
+        ptr = (uint8_t*)ptr + frag.len - size;
+        frag.len -= size;
+        self->bufferfrags[ptr] = (rdma_frag_t){ .buf = frag.buf, .len = size, .is_free = false };
+    }
+    return ptr;
+}
+
+void rdma_malloc_free(rdma_allocator_t *self, void *buf)
+{
+    auto frag_it = self->bufferfrags.find(buf);
+    if (frag_it == self->bufferfrags.end())
+    {
+        fprintf(stderr, "BUG: Attempt to double-free RDMA buffer fragment 0x%jx\n", (size_t)buf);
+        return;
+    }
+    auto prev_it = frag_it, next_it = frag_it;
+    if (frag_it != self->bufferfrags.begin())
+        prev_it--;
+    next_it++;
+    bool merge_back = prev_it != frag_it &&
+        prev_it->second.is_free &&
+        prev_it->second.buf == frag_it->second.buf &&
+        (uint8_t*)prev_it->first+prev_it->second.len == frag_it->first;
+    bool merge_next = next_it != self->bufferfrags.end() &&
+        next_it->second.is_free &&
+        next_it->second.buf == frag_it->second.buf &&
+        next_it->first == (uint8_t*)frag_it->first+frag_it->second.len;
+    if (merge_back && merge_next)
+    {
+        prev_it->second.len += frag_it->second.len + next_it->second.len;
+        self->freelist.erase(std::pair<size_t, void*>(next_it->second.len, next_it->first));
+        self->bufferfrags.erase(next_it);
+        self->bufferfrags.erase(frag_it);
+        frag_it = prev_it;
+    }
+    else if (merge_back)
+    {
+        prev_it->second.len += frag_it->second.len;
+        self->bufferfrags.erase(frag_it);
+        frag_it = prev_it;
+    }
+    else if (merge_next)
+    {
+        frag_it->second.is_free = true;
+        frag_it->second.len += next_it->second.len;
+        self->freelist.erase(std::pair<size_t, void*>(next_it->second.len, next_it->first));
+        self->bufferfrags.erase(next_it);
+    }
+    else
+    {
+        frag_it->second.is_free = true;
+        self->freelist.insert(std::pair<size_t, void*>(frag_it->second.len, frag_it->first));
+    }
+    assert(frag_it->second.len <= frag_it->second.buf->len);
+    if (frag_it->second.len == frag_it->second.buf->len)
+    {
+        // The whole buffer is freed
+        self->freebuffers += frag_it->second.buf->len;
+        if (self->freebuffers > self->rdma_max_unused)
+        {
+            rdma_malloc_free_unused_buffers(self, self->rdma_max_unused, false);
+        }
+    }
+}
+
+uint32_t rdma_malloc_get_lkey(rdma_allocator_t *self, void *buf)
+{
+    auto frag_it = self->bufferfrags.find(buf);
+    if (frag_it == self->bufferfrags.end())
+    {
+        fprintf(stderr, "BUG: Attempt to use an unknown RDMA buffer fragment 0x%zx\n", (size_t)buf);
+        abort();
+    }
+    return frag_it->second.buf->mr->lkey;
+}

src/nfs/rdma_alloc.h

@@ -0,0 +1,17 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.1 (see README.md for details)
+//
+// Simple & stupid RDMA-enabled memory allocator (allocates buffers within ibv_mr's)
+
+#pragma once
+
+#include <infiniband/verbs.h>
+#include <stdint.h>
+
+struct rdma_allocator_t;
+
+rdma_allocator_t *rdma_malloc_create(ibv_pd *pd, size_t rdma_alloc_size, size_t rdma_max_unused, int rdma_access);
+void rdma_malloc_destroy(rdma_allocator_t *self);
+void *rdma_malloc_alloc(rdma_allocator_t *self, size_t size);
+void rdma_malloc_free(rdma_allocator_t *self, void *buf);
+uint32_t rdma_malloc_get_lkey(rdma_allocator_t *self, void *buf);