WIP Implement RDMA v2 based on IBV_WR_RDMA_WRITE with remote buffer management

One BIG FIXME remaining - handling large operations :))

src/CMakeLists.txt

@@ -98,7 +98,7 @@ endif (${WITH_FIO})
 # libvitastor_common.a
 set(MSGR_RDMA "")
 if (IBVERBS_LIBRARIES)
-	set(MSGR_RDMA "msgr_rdma.cpp")
+	set(MSGR_RDMA msgr_rdma.cpp freelist.cpp allocator.cpp)
 endif (IBVERBS_LIBRARIES)
 add_library(vitastor_common STATIC
 	epoll_manager.cpp etcd_state_client.cpp messenger.cpp addr_util.cpp
@@ -278,6 +278,11 @@ add_executable(test_allocator EXCLUDE_FROM_ALL test_allocator.cpp allocator.cpp)
 add_dependencies(build_tests test_allocator)
 add_test(NAME test_allocator COMMAND test_allocator)
 
+# test_freelist
+add_executable(test_freelist EXCLUDE_FROM_ALL test_freelist.cpp)
+add_dependencies(build_tests test_freelist)
+add_test(NAME test_freelist COMMAND test_freelist)
+
 # test_cas
 add_executable(test_cas
 	test_cas.cpp

src/freelist.cpp

@@ -0,0 +1,63 @@
+// Copyright (c) Vitaliy Filippov, 2023+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <assert.h>
+#include "freelist.h"
+
+uint64_t freelist_allocator_t::alloc(uint64_t data_size)
+{
+    for (int i = 0; i < freelist.size(); i++)
+    {
+        if (freelist[i].size >= data_size)
+        {
+            uint64_t r = freelist[i].start;
+            freelist[i].start += data_size;
+            freelist[i].size -= data_size;
+            return r;
+        }
+    }
+    return UINT64_MAX;
+}
+
+void freelist_allocator_t::free(uint64_t start, uint64_t size)
+{
+    int min = 0, max = freelist.size();
+    if (max && freelist[freelist.size()-1].start < start)
+    {
+        min = max;
+    }
+    if (max && freelist[0].start >= start)
+    {
+        max = 0;
+    }
+    while (max-min > 1)
+    {
+        int mid = (min+max)/2;
+        if (freelist[mid].start >= start)
+            max = mid;
+        else
+            min = mid;
+    }
+    // max = the first item where freelist[max].start >= start
+    if (max > 0 && freelist[max-1].start+freelist[max-1].size >= start)
+    {
+        assert(freelist[max-1].start+freelist[max-1].size == start);
+        freelist[max-1].size += size;
+    }
+    else if (max < freelist.size() && freelist[max].start <= size+start)
+    {
+        assert(freelist[max].start == size+start);
+        freelist[max].start -= size;
+        freelist[max].size += size;
+    }
+    else
+    {
+        freelist.insert(freelist.begin()+min, (freelist_item_t){ .start = start, .size = size });
+        max = min; // to skip the if below
+    }
+    if (min != max && max < freelist.size() && freelist[max].start == freelist[min].start+freelist[min].size)
+    {
+        freelist[min].size += freelist[max].size;
+        freelist.erase(freelist.begin()+max, freelist.begin()+max+1);
+    }
+}

src/freelist.h

@@ -0,0 +1,23 @@
+// Copyright (c) Vitaliy Filippov, 2023+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#pragma once
+
+#include <stdint.h>
+#include <vector>
+
+struct freelist_item_t
+{
+    uint64_t start, size;
+};
+
+// Really trivial freelist allocator
+// Should be fine for remote RDMA memory management because
+// most of the time fragmentation shouldn't be an issue as all
+// memory regions are short-lived
+struct freelist_allocator_t
+{
+    std::vector<freelist_item_t> freelist;
+    uint64_t alloc(uint64_t data_size);
+    void free(uint64_t start, uint64_t size);
+};

src/messenger.cpp

@@ -157,13 +157,16 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         this->rdma_max_sge = 128;
     this->rdma_max_send = config["rdma_max_send"].uint64_value();
     if (!this->rdma_max_send)
-        this->rdma_max_send = 1;
+        this->rdma_max_send = 128;
     this->rdma_max_recv = config["rdma_max_recv"].uint64_value();
     if (!this->rdma_max_recv)
         this->rdma_max_recv = 128;
-    this->rdma_max_msg = config["rdma_max_msg"].uint64_value();
-    if (!this->rdma_max_msg || this->rdma_max_msg > 128*1024*1024)
-        this->rdma_max_msg = 129*1024;
+    this->rdma_op_slots = config["rdma_op_slots"].uint64_value();
+    if (!this->rdma_op_slots || this->rdma_op_slots >= 1024*1024)
+        this->rdma_op_slots = 4096;
+    this->rdma_op_memory = config["rdma_op_memory"].uint64_value();
+    if (!this->rdma_op_memory || this->rdma_op_memory >= 1024*1024*1024)
+        this->rdma_op_memory = 16*1024*1024;
 #endif
     this->receive_buffer_size = (uint32_t)config["tcp_header_buffer_size"].uint64_value();
     if (!this->receive_buffer_size || this->receive_buffer_size > 1024*1024*1024)
@@ -388,12 +391,16 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
 #ifdef WITH_RDMA
     if (rdma_context)
     {
-        cl->rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_max_msg);
+        cl->rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_op_slots, rdma_op_memory);
         if (cl->rdma_conn)
         {
+            clients_by_qp[cl->rdma_conn->qp->qp_num] = cl->peer_fd;
             json11::Json payload = json11::Json::object {
                 { "connect_rdma", cl->rdma_conn->addr.to_string() },
-                { "rdma_max_msg", cl->rdma_conn->max_msg },
+                { "rdma_data_rkey", (uint64_t)cl->rdma_conn->in_data_mr->rkey },
+                { "rdma_op_rkey", (uint64_t)cl->rdma_conn->in_op_mr->rkey },
+                { "rdma_op_slots", cl->rdma_conn->op_slots },
+                { "rdma_op_memory", cl->rdma_conn->op_memory },
             };
             std::string payload_str = payload.dump();
             op->req.show_conf.json_len = payload_str.size();
@@ -453,12 +460,14 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
         {
             msgr_rdma_address_t addr;
             if (!msgr_rdma_address_t::from_string(config["rdma_address"].string_value().c_str(), &addr) ||
+                config["rdma_op_memory"].uint64_value() == 0 ||
                 cl->rdma_conn->connect(&addr) != 0)
             {
                 fprintf(
                     stderr, "Failed to connect to OSD %lu (address %s) using RDMA\n",
                     cl->osd_num, config["rdma_address"].string_value().c_str()
                 );
+                clients_by_qp.erase(cl->rdma_conn->qp->qp_num);
                 delete cl->rdma_conn;
                 cl->rdma_conn = NULL;
                 // FIXME: Keep TCP connection in this case
@@ -470,11 +479,12 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
             }
             else
             {
-                uint64_t server_max_msg = config["rdma_max_msg"].uint64_value();
-                if (cl->rdma_conn->max_msg > server_max_msg)
-                {
-                    cl->rdma_conn->max_msg = server_max_msg;
-                }
+                cl->rdma_conn->set_out_capacity(
+                    config["rdma_data_rkey"].uint64_value(),
+                    config["rdma_op_rkey"].uint64_value(),
+                    config["rdma_op_slots"].uint64_value(),
+                    config["rdma_op_memory"].uint64_value()
+                );
                 if (log_level > 0)
                 {
                     fprintf(stderr, "Connected to OSD %lu using RDMA\n", cl->osd_num);

src/messenger.h

@@ -37,6 +37,7 @@
 
 #define MSGR_SENDP_HDR 1
 #define MSGR_SENDP_FREE 2
+#define MSGR_SENDP_LAST 4
 
 struct msgr_sendp_t
 {
@@ -131,9 +132,10 @@ protected:
     bool use_rdma = true;
     std::string rdma_device;
     uint64_t rdma_port_num = 1, rdma_gid_index = 0, rdma_mtu = 0;
-    msgr_rdma_context_t *rdma_context = NULL;
     uint64_t rdma_max_sge = 0, rdma_max_send = 0, rdma_max_recv = 0;
-    uint64_t rdma_max_msg = 0;
+    uint64_t rdma_op_slots = 0, rdma_op_memory = 0;
+    msgr_rdma_context_t *rdma_context = NULL;
+    std::map<uint32_t, int> clients_by_qp;
 #endif
 
     std::vector<int> read_ready_clients;
@@ -170,7 +172,8 @@ public:
 
 #ifdef WITH_RDMA
     bool is_rdma_enabled();
-    bool connect_rdma(int peer_fd, std::string rdma_address, uint64_t client_max_msg);
+    bool connect_rdma(int peer_fd, std::string rdma_address,
+        uint32_t out_data_rkey, uint32_t out_op_rkey, uint64_t out_op_slots, uint64_t out_op_memory);
 #endif
 
 protected:
@@ -191,12 +194,13 @@ protected:
     bool handle_read_buffer(osd_client_t *cl, void *curbuf, int remain);
     bool handle_finished_read(osd_client_t *cl);
     void handle_op_hdr(osd_client_t *cl);
-    bool handle_reply_hdr(osd_client_t *cl);
+    bool handle_reply_hdr(void *reply_hdr, osd_client_t *cl);
     void handle_reply_ready(osd_op_t *op);
 
 #ifdef WITH_RDMA
     bool try_send_rdma(osd_client_t *cl);
     bool try_recv_rdma(osd_client_t *cl);
     void handle_rdma_events();
+    bool rdma_handle_op(osd_client_t *cl, uint32_t op_slot);
 #endif
 };

src/msgr_rdma.cpp

@@ -46,9 +46,20 @@ msgr_rdma_connection_t::~msgr_rdma_connection_t()
     ctx->used_max_cqe -= max_send+max_recv;
     if (qp)
         ibv_destroy_qp(qp);
-    if (recv_buffers.size())
-        for (auto b: recv_buffers)
-            free(b);
+    if (in_data_mr)
+        ibv_dereg_mr(in_data_mr);
+    if (in_op_mr)
+        ibv_dereg_mr(in_op_mr);
+    if (in_data_buf)
+        free(in_data_buf);
+    if (in_ops)
+        free(in_ops);
+    if (out_op_alloc)
+        delete out_op_alloc;
+    if (out_slot_data)
+        free(out_slot_data);
+    if (out_slot_ops)
+        free(out_slot_ops);
 }
 
 msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t ib_port, uint8_t gid_index, uint32_t mtu, int log_level)
@@ -149,7 +160,7 @@ msgr_rdma_context_t *msgr_rdma_context_t::create(const char *ib_devname, uint8_t
     ctx->mr = ibv_reg_mr(ctx->pd, NULL, SIZE_MAX, IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_ON_DEMAND);
     if (!ctx->mr)
     {
-        fprintf(stderr, "Couldn't register RDMA memory region\n");
+        fprintf(stderr, "Couldn't register global RDMA memory region: %s\n", strerror(errno));
         goto cleanup;
     }
 
@@ -180,7 +191,7 @@ cleanup:
 }
 
 msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx, uint32_t max_send,
-    uint32_t max_recv, uint32_t max_sge, uint32_t max_msg)
+    uint32_t max_recv, uint32_t max_sge, uint64_t op_slots, uint64_t op_memory)
 {
     msgr_rdma_connection_t *conn = new msgr_rdma_connection_t;
 
@@ -190,7 +201,6 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
     conn->max_send = max_send;
     conn->max_recv = max_recv;
     conn->max_sge = max_sge;
-    conn->max_msg = max_msg;
 
     ctx->used_max_cqe += max_send+max_recv;
     if (ctx->used_max_cqe > ctx->max_cqe)
@@ -211,6 +221,30 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
         ctx->max_cqe = new_max_cqe;
     }
 
+    conn->op_memory = op_memory;
+    conn->in_data_buf = memalign_or_die(MEM_ALIGNMENT, op_memory);
+    conn->in_data_mr = ibv_reg_mr(ctx->pd, conn->in_data_buf, op_memory,
+        IBV_ACCESS_ZERO_BASED | IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_ON_DEMAND);
+    if (!conn->in_data_mr)
+    {
+        fprintf(stderr, "Couldn't register %lu MB RDMA memory region for incoming data: %s\n",
+            (op_memory+1024*1024-1)/1024/1024, strerror(errno));
+        delete conn;
+        return NULL;
+    }
+
+    conn->op_slots = op_slots;
+    conn->in_ops = (msgr_rdma_cmd_t *)malloc_or_die(sizeof(msgr_rdma_cmd_t) * op_slots);
+    conn->in_op_mr = ibv_reg_mr(ctx->pd, conn->in_ops, sizeof(msgr_rdma_cmd_t) * op_slots,
+        IBV_ACCESS_ZERO_BASED | IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_ON_DEMAND);
+    if (!conn->in_op_mr)
+    {
+        fprintf(stderr, "Couldn't register %lu KB RDMA memory region for incoming operation headers: %s\n",
+            (sizeof(msgr_rdma_cmd_t) * op_slots + 1023)/1024, strerror(errno));
+        delete conn;
+        return NULL;
+    }
+
     ibv_qp_init_attr init_attr = {
         .send_cq = ctx->cq,
         .recv_cq = ctx->cq,
@@ -237,7 +271,7 @@ msgr_rdma_connection_t *msgr_rdma_connection_t::create(msgr_rdma_context_t *ctx,
 
     ibv_qp_attr attr = {
         .qp_state        = IBV_QPS_INIT,
-        .qp_access_flags = 0,
+        .qp_access_flags = IBV_ACCESS_REMOTE_WRITE,
         .pkey_index      = 0,
         .port_num        = ctx->ib_port,
     };
@@ -265,6 +299,19 @@ static ibv_mtu mtu_to_ibv_mtu(uint32_t mtu)
     return IBV_MTU_4096;
 }
 
+void msgr_rdma_connection_t::set_out_capacity(uint32_t out_data_rkey, uint32_t out_op_rkey, uint64_t out_op_slots, uint64_t out_op_memory)
+{
+    assert(!out_op_alloc);
+    this->out_data_rkey = out_data_rkey;
+    this->out_op_rkey = out_op_rkey;
+    this->out_op_slots = out_op_slots;
+    this->out_op_memory = out_op_memory;
+    out_op_alloc = new allocator(out_op_slots);
+    out_data_alloc.free(0, out_op_memory);
+    out_slot_data = (msgr_rdma_out_pos_t *)malloc_or_die(sizeof(msgr_rdma_out_pos_t) * out_op_slots);
+    out_slot_ops = (osd_op_t **)malloc_or_die(sizeof(osd_op_t *) * out_op_slots);
+}
+
 int msgr_rdma_connection_t::connect(msgr_rdma_address_t *dest)
 {
     auto conn = this;
@@ -311,17 +358,14 @@ int msgr_rdma_connection_t::connect(msgr_rdma_address_t *dest)
     return 0;
 }
 
-bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address, uint64_t client_max_msg)
+bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address,
+    uint32_t out_data_rkey, uint32_t out_op_rkey, uint64_t out_op_slots, uint64_t out_op_memory)
 {
     // Try to connect to the peer using RDMA
     msgr_rdma_address_t addr;
     if (msgr_rdma_address_t::from_string(rdma_address.c_str(), &addr))
     {
-        if (client_max_msg > rdma_max_msg)
-        {
-            client_max_msg = rdma_max_msg;
-        }
-        auto rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, client_max_msg);
+        auto rdma_conn = msgr_rdma_connection_t::create(rdma_context, rdma_max_send, rdma_max_recv, rdma_max_sge, rdma_op_slots, rdma_op_memory);
         if (rdma_conn)
         {
             int r = rdma_conn->connect(&addr);
@@ -336,6 +380,8 @@ bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address, uint64
             else
             {
                 // Remember connection, but switch to RDMA only after sending the configuration response
+                clients_by_qp[rdma_conn->qp->qp_num] = peer_fd;
+                rdma_conn->set_out_capacity(out_data_rkey, out_op_rkey, out_op_slots, out_op_memory);
                 auto cl = clients.at(peer_fd);
                 cl->rdma_conn = rdma_conn;
                 cl->peer_state = PEER_RDMA_CONNECTING;
@@ -346,66 +392,161 @@ bool osd_messenger_t::connect_rdma(int peer_fd, std::string rdma_address, uint64
     return false;
 }
 
-static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
-{
-    ibv_send_wr *bad_wr = NULL;
-    ibv_send_wr wr = {
-        .wr_id = (uint64_t)(cl->peer_fd*2+1),
-        .sg_list = sge,
-        .num_sge = op_sge,
-        .opcode = IBV_WR_SEND,
-        .send_flags = IBV_SEND_SIGNALED,
-    };
-    int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
-    if (err || bad_wr)
-    {
-        fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
-        exit(1);
-    }
-    cl->rdma_conn->cur_send++;
-}
-
 bool osd_messenger_t::try_send_rdma(osd_client_t *cl)
 {
     auto rc = cl->rdma_conn;
-    if (!cl->send_list.size() || rc->cur_send > 0)
+    if (!cl->send_list.size() && !rc->in_slots_freed.size() || rc->cur_send >= rc->max_send)
     {
-        // Only send one batch at a time
         return true;
     }
-    uint64_t op_size = 0, op_sge = 0;
-    ibv_sge sge[rc->max_sge];
-    while (rc->send_pos < cl->send_list.size())
+    int i = 0;
+    while (i < rc->in_slots_freed.size())
     {
-        iovec & iov = cl->send_list[rc->send_pos];
-        if (op_size >= rc->max_msg || op_sge >= rc->max_sge)
-        {
-            try_send_rdma_wr(cl, sge, op_sge);
-            op_sge = 0;
-            op_size = 0;
-            if (rc->cur_send >= rc->max_send)
-            {
-                break;
-            }
-        }
-        uint32_t len = (uint32_t)(op_size+iov.iov_len-rc->send_buf_pos < rc->max_msg
-            ? iov.iov_len-rc->send_buf_pos : rc->max_msg-op_size);
-        sge[op_sge++] = {
-            .addr = (uintptr_t)((uint8_t*)iov.iov_base+rc->send_buf_pos),
-            .length = len,
-            .lkey = rc->ctx->mr->lkey,
+        auto op_slot = rc->in_slots_freed[i++];
+        assert(op_slot < 0x80000000);
+        ibv_send_wr *bad_wr = NULL;
+        ibv_send_wr wr = {
+            .wr_id = 0,
+            .opcode = IBV_WR_RDMA_WRITE_WITH_IMM,
+            .imm_data = 0x80000000 | op_slot,
         };
-        op_size += len;
-        rc->send_buf_pos += len;
-        if (rc->send_buf_pos >= iov.iov_len)
+        int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
+        if (err || bad_wr)
         {
-            rc->send_pos++;
-            rc->send_buf_pos = 0;
+            fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
+            exit(1);
+        }
+        rc->cur_send++;
+        if (rc->cur_send >= rc->max_send)
+        {
+            break;
         }
     }
-    if (op_sge > 0)
+    rc->in_slots_freed.erase(rc->in_slots_freed.begin(), rc->in_slots_freed.begin()+i);
+    if (!cl->send_list.size() || rc->cur_send >= rc->max_send)
     {
-        try_send_rdma_wr(cl, sge, op_sge);
+        return true;
+    }
+    ibv_sge sge[rc->max_sge];
+    int op_start = 0;
+    while (op_start < cl->send_list.size())
+    {
+        uint64_t op_data_size = 0;
+        int op_end = op_start;
+        while (!(cl->outbox[op_end].flags & MSGR_SENDP_LAST))
+        {
+            op_data_size += cl->send_list[op_end].iov_len;
+            op_end++;
+        }
+        op_data_size += cl->send_list[op_end].iov_len;
+        op_end++;
+        op_data_size -= cl->send_list[op_start].iov_len;
+        // Operation boundaries in send_list: op_start..op_end, first iovec is the header
+        uint64_t op_slot = rc->out_op_alloc->find_free();
+        if (op_slot == UINT64_MAX)
+        {
+            // op queue is full
+            return true;
+        }
+        uint64_t data_pos = UINT64_MAX;
+        if (op_data_size >= 0)
+        {
+            if (rc->cur_send > rc->max_send-1-(op_end-op_start-1+rc->max_sge)/rc->max_sge)
+            {
+                // RDMA queue is full
+                return true;
+            }
+            // FIXME: Oops, and what if op data is larger than the whole buffer... :)
+            data_pos = rc->out_data_alloc.alloc(op_data_size);
+            if (data_pos == UINT64_MAX)
+            {
+                // data buffers are full
+                return true;
+            }
+            int cur_sge = 0;
+            for (int data_sent = 1; data_sent < op_end; data_sent++)
+            {
+                sge[cur_sge++] = {
+                    .addr = (uintptr_t)cl->send_list[data_sent].iov_base,
+                    .length = (uint32_t)cl->send_list[data_sent].iov_len,
+                    .lkey = rc->ctx->mr->lkey,
+                };
+                if (data_sent == op_end-1 || cur_sge >= rc->max_sge)
+                {
+                    ibv_send_wr *bad_wr = NULL;
+                    ibv_send_wr wr = {
+                        .wr_id = op_slot,
+                        .next = NULL,
+                        .sg_list = sge,
+                        .num_sge = cur_sge,
+                        .opcode = IBV_WR_RDMA_WRITE,
+                        .send_flags = 0,
+                        .wr = {
+                            .rdma = {
+                                .remote_addr = data_pos,
+                                .rkey = rc->out_data_rkey,
+                            },
+                        },
+                    };
+                    int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
+                    if (err || bad_wr)
+                    {
+                        fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
+                        exit(1);
+                    }
+                    rc->cur_send++;
+                    cur_sge = 0;
+                }
+            }
+        }
+        if (rc->cur_send > rc->max_send-1)
+        {
+            // RDMA queue is full
+            return true;
+        }
+        rc->out_op_alloc->set(op_slot, true);
+        assert(cl->send_list[op_start].iov_len == OSD_PACKET_SIZE);
+        sge[0] = {
+            .addr = (uintptr_t)cl->send_list[op_start].iov_base,
+            .length = (uint32_t)cl->send_list[op_start].iov_len,
+            .lkey = rc->ctx->mr->lkey,
+        };
+        rc->out_slot_data[op_slot] = { .data_pos = data_pos, .data_size = op_data_size };
+        rc->out_slot_ops[op_slot] = (cl->outbox[op_end-1].flags & MSGR_SENDP_FREE)
+            ? cl->outbox[op_end-1].op : NULL;
+        sge[1] = {
+            .addr = (uintptr_t)(rc->out_slot_data+op_slot),
+            .length = sizeof(rc->out_slot_data[op_slot]),
+            .lkey = rc->ctx->mr->lkey,
+        };
+        ibv_send_wr *bad_wr = NULL;
+        ibv_send_wr wr = {
+            .wr_id = op_slot,
+            .next = NULL,
+            .sg_list = sge,
+            .num_sge = 2,
+            .opcode = IBV_WR_RDMA_WRITE_WITH_IMM,
+            .send_flags = IBV_SEND_SIGNALED,
+            .imm_data = (uint32_t)op_slot,
+            .wr = {
+                .rdma = {
+                    .remote_addr = op_slot*sizeof(msgr_rdma_cmd_t),
+                    .rkey = rc->out_op_rkey,
+                },
+            },
+        };
+        int err = ibv_post_send(cl->rdma_conn->qp, &wr, &bad_wr);
+        if (err || bad_wr)
+        {
+            fprintf(stderr, "RDMA send failed: %s\n", strerror(err));
+            exit(1);
+        }
+        rc->cur_send++;
+        op_start = op_end;
+    }
+    if (op_start > 0)
+    {
+        cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+op_start);
     }
     return true;
 }
@@ -427,23 +568,87 @@ static void try_recv_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
     cl->rdma_conn->cur_recv++;
 }
 
+static void copy_data_to_recv_list(uint8_t *data_buf, uint64_t data_size, osd_client_t *cl)
+{
+    uint64_t pos = 0;
+    while (cl->recv_list.done < cl->recv_list.count)
+    {
+        uint64_t cur = cl->recv_list.buf[cl->recv_list.done].iov_len;
+        assert(cur <= data_size-pos);
+        memcpy(cl->recv_list.buf[cl->recv_list.done].iov_base, data_buf+pos, cur);
+        pos += cur;
+    }
+    cl->recv_list.reset();
+}
+
 bool osd_messenger_t::try_recv_rdma(osd_client_t *cl)
 {
     auto rc = cl->rdma_conn;
     while (rc->cur_recv < rc->max_recv)
     {
-        void *buf = malloc_or_die(rc->max_msg);
-        rc->recv_buffers.push_back(buf);
-        ibv_sge sge = {
-            .addr = (uintptr_t)buf,
-            .length = (uint32_t)rc->max_msg,
-            .lkey = rc->ctx->mr->lkey,
-        };
-        try_recv_rdma_wr(cl, &sge, 1);
+        try_recv_rdma_wr(cl, NULL, 0);
     }
     return true;
 }
 
+bool osd_messenger_t::rdma_handle_op(osd_client_t *cl, uint32_t op_slot)
+{
+    auto rc = cl->rdma_conn;
+    if (op_slot >= rc->in_op_cap)
+    {
+        // Invalid incoming index
+        fprintf(stderr, "Client %d invalid incoming RDMA op slot: %u, dropping connection\n", cl->peer_fd, op_slot);
+        stop_client(cl->peer_fd);
+        return false;
+    }
+    osd_op_header_t *hdr = (osd_op_header_t *)rc->in_ops[op_slot].header;
+    uint8_t *data_buf = (uint8_t*)rc->in_data_buf + rc->in_ops[op_slot].pos.data_pos;
+    uint64_t data_size = rc->in_ops[op_slot].pos.data_size;
+    if (hdr->magic == SECONDARY_OSD_REPLY_MAGIC)
+    {
+        // Reply
+        if (cl->read_op)
+        {
+            delete cl->read_op;
+            cl->read_op = NULL;
+        }
+        if (!handle_reply_hdr(rc->in_ops[op_slot].header, cl))
+            return false;
+        if (cl->read_state == CL_READ_REPLY_DATA)
+        {
+            // copy reply data to cl->recv_list
+            copy_data_to_recv_list(data_buf, data_size, cl);
+            // and handle reply with data
+            handle_reply_ready(cl->read_op);
+            cl->read_op = NULL;
+            cl->read_state = 0;
+            cl->read_remaining = 0;
+        }
+    }
+    else
+    {
+        // Operation
+        cl->read_op = new osd_op_t;
+        cl->read_op->peer_fd = cl->peer_fd;
+        cl->read_op->op_type = OSD_OP_IN;
+        memcpy(&cl->read_op->req, hdr, OSD_PACKET_SIZE);
+        handle_op_hdr(cl);
+        if (cl->read_state == CL_READ_DATA)
+        {
+            copy_data_to_recv_list(data_buf, data_size, cl);
+            // And handle the incoming op with data
+            cl->received_ops.push_back(cl->read_op);
+            set_immediate.push_back([this, op = cl->read_op]() { exec_op(op); });
+            cl->read_op = NULL;
+            cl->read_state = 0;
+        }
+    }
+    // We don't need the incoming data buffer anymore, notify peer about it
+    // FIXME: Allow to pass memory to the internal layer without copying and notify after handling it
+    rc->in_slots_freed.push_back(op_slot);
+    return true;
+}
+
 #define RDMA_EVENTS_AT_ONCE 32
 
 void osd_messenger_t::handle_rdma_events()
@@ -468,9 +673,9 @@ void osd_messenger_t::handle_rdma_events()
         event_count = ibv_poll_cq(rdma_context->cq, RDMA_EVENTS_AT_ONCE, wc);
         for (int i = 0; i < event_count; i++)
         {
-            int client_id = wc[i].wr_id >> 1;
-            bool is_send = wc[i].wr_id & 1;
-            auto cl_it = clients.find(client_id);
+            auto cqp_it = clients_by_qp.find(wc[i].qp_num);
+            int peer_fd = cqp_it != clients_by_qp.end() ? cqp_it->second : -1;
+            auto cl_it = clients.find(peer_fd);
             if (cl_it == clients.end())
             {
                 continue;
@@ -478,55 +683,51 @@ void osd_messenger_t::handle_rdma_events()
             osd_client_t *cl = cl_it->second;
             if (wc[i].status != IBV_WC_SUCCESS)
             {
-                fprintf(stderr, "RDMA work request failed for client %d", client_id);
+                fprintf(stderr, "RDMA work request failed for client %d", peer_fd);
                 if (cl->osd_num)
-                {
                     fprintf(stderr, " (OSD %lu)", cl->osd_num);
-                }
                 fprintf(stderr, " with status: %s, stopping client\n", ibv_wc_status_str(wc[i].status));
-                stop_client(client_id);
+                if (peer_fd >= 0)
+                    stop_client(peer_fd);
                 continue;
             }
-            if (!is_send)
+            auto rc = cl->rdma_conn;
+            if (wc[i].opcode == IBV_WC_RDMA_WRITE)
             {
-                cl->rdma_conn->cur_recv--;
-                if (!handle_read_buffer(cl, cl->rdma_conn->recv_buffers[0], wc[i].byte_len))
+                // Operation or reply is sent, we can free it
+                auto & op = rc->out_slot_ops[wc[i].wr_id];
+                if (op)
                 {
-                    // handle_read_buffer may stop the client
-                    continue;
+                    delete op;
+                    op = NULL;
                 }
-                free(cl->rdma_conn->recv_buffers[0]);
-                cl->rdma_conn->recv_buffers.erase(cl->rdma_conn->recv_buffers.begin(), cl->rdma_conn->recv_buffers.begin()+1);
-                try_recv_rdma(cl);
+                rc->cur_send--;
+                try_send_rdma(cl);
             }
-            else
+            else if (wc[i].opcode == IBV_WC_RECV)
             {
-                cl->rdma_conn->cur_send--;
-                if (!cl->rdma_conn->cur_send)
+                if (!(wc[i].imm_data & 0x80000000))
                 {
-                    // Wait for the whole batch
-                    for (int i = 0; i < cl->rdma_conn->send_pos; i++)
+                    // Operation or reply received. Handle it
+                    if (!rdma_handle_op(cl, wc[i].imm_data))
                     {
-                        if (cl->outbox[i].flags & MSGR_SENDP_FREE)
-                        {
-                            // Reply fully sent
-                            delete cl->outbox[i].op;
-                        }
+                        // false means that the client is stopped due to invalid operation
+                        continue;
                     }
-                    if (cl->rdma_conn->send_pos > 0)
-                    {
-                        cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+cl->rdma_conn->send_pos);
-                        cl->outbox.erase(cl->outbox.begin(), cl->outbox.begin()+cl->rdma_conn->send_pos);
-                        cl->rdma_conn->send_pos = 0;
-                    }
-                    if (cl->rdma_conn->send_buf_pos > 0)
-                    {
-                        cl->send_list[0].iov_base = (uint8_t*)cl->send_list[0].iov_base + cl->rdma_conn->send_buf_pos;
-                        cl->send_list[0].iov_len -= cl->rdma_conn->send_buf_pos;
-                        cl->rdma_conn->send_buf_pos = 0;
-                    }
-                    try_send_rdma(cl);
+                    rc->cur_recv--;
+                    try_recv_rdma(cl);
                 }
+                else
+                {
+                    // Outbox slot is marked as free (the remote side doesn't need it anymore)
+                    uint32_t op_slot = wc[i].imm_data & 0x7FFFFFFF;
+                    auto & pos = rc->in_ops[op_slot].pos;
+                    if (pos.data_size > 0)
+                        rc->out_data_alloc.free(pos.data_pos, pos.data_size);
+                    rc->out_op_alloc->set(op_slot, false);
+                }
+                // Try to continue sending
+                try_send_rdma(cl);
             }
         }
     } while (event_count > 0);

src/msgr_rdma.h

@@ -5,6 +5,11 @@
 #include <infiniband/verbs.h>
 #include <string>
 #include <vector>
+#include "allocator.h"
+#include "freelist.h"
+#include "osd_ops.h"
+
+struct osd_op_t;
 
 struct msgr_rdma_address_t
 {
@@ -39,6 +44,17 @@ struct msgr_rdma_context_t
     ~msgr_rdma_context_t();
 };
 
+struct msgr_rdma_out_pos_t
+{
+    uint64_t data_pos, data_size;
+};
+
+struct msgr_rdma_cmd_t
+{
+    uint8_t header[OSD_PACKET_SIZE];
+    msgr_rdma_out_pos_t pos;
+};
+
 struct msgr_rdma_connection_t
 {
     msgr_rdma_context_t *ctx = NULL;
@@ -46,13 +62,24 @@ struct msgr_rdma_connection_t
     msgr_rdma_address_t addr;
     int max_send = 0, max_recv = 0, max_sge = 0;
     int cur_send = 0, cur_recv = 0;
-    uint64_t max_msg = 0;
+    uint64_t op_slots = 0, op_memory = 0;
 
-    int send_pos = 0, send_buf_pos = 0;
-    int recv_pos = 0, recv_buf_pos = 0;
-    std::vector<void*> recv_buffers;
+    ibv_mr *in_data_mr = NULL, *in_op_mr = NULL;
+    msgr_rdma_cmd_t *in_ops = NULL;
+    int in_op_cap = 0;
+    void *in_data_buf = NULL;
+    std::vector<uint32_t> in_slots_freed;
+
+    uint32_t out_data_rkey = 0, out_op_rkey = 0;
+    uint64_t out_op_slots = 0, out_op_memory = 0;
+    allocator *out_op_alloc = NULL;
+    freelist_allocator_t out_data_alloc;
+    msgr_rdma_out_pos_t *out_slot_data = NULL;
+    osd_op_t **out_slot_ops = NULL;
 
     ~msgr_rdma_connection_t();
-    static msgr_rdma_connection_t *create(msgr_rdma_context_t *ctx, uint32_t max_send, uint32_t max_recv, uint32_t max_sge, uint32_t max_msg);
+    static msgr_rdma_connection_t *create(msgr_rdma_context_t *ctx, uint32_t max_send,
+        uint32_t max_recv, uint32_t max_sge, uint64_t op_slots, uint64_t op_memory);
     int connect(msgr_rdma_address_t *dest);
+    void set_out_capacity(uint32_t out_data_rkey, uint32_t out_op_rkey, uint64_t out_op_slots, uint64_t out_op_memory);
 };

src/msgr_receive.cpp

@@ -172,7 +172,7 @@ bool osd_messenger_t::handle_finished_read(osd_client_t *cl)
     if (cl->read_state == CL_READ_HDR)
     {
         if (cl->read_op->req.hdr.magic == SECONDARY_OSD_REPLY_MAGIC)
-            return handle_reply_hdr(cl);
+            return handle_reply_hdr(cl->read_op->req.buf, cl);
         else if (cl->read_op->req.hdr.magic == SECONDARY_OSD_OP_MAGIC)
             handle_op_hdr(cl);
         else
@@ -286,7 +286,7 @@ void osd_messenger_t::handle_op_hdr(osd_client_t *cl)
     }
 }
 
-bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
+bool osd_messenger_t::handle_reply_hdr(void *reply_hdr, osd_client_t *cl)
 {
     auto req_it = cl->sent_ops.find(cl->read_op->req.hdr.id);
     if (req_it == cl->sent_ops.end())
@@ -297,7 +297,7 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
         return false;
     }
     osd_op_t *op = req_it->second;
-    memcpy(op->reply.buf, cl->read_op->req.buf, OSD_PACKET_SIZE);
+    memcpy(op->reply.buf, reply_hdr, OSD_PACKET_SIZE);
     cl->sent_ops.erase(req_it);
     if (op->reply.hdr.opcode == OSD_OP_SEC_READ || op->reply.hdr.opcode == OSD_OP_READ)
     {
@@ -328,14 +328,16 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
         {
             goto reuse;
         }
-        delete cl->read_op;
+        if (cl->read_op)
+            delete cl->read_op;
         cl->read_op = op;
         cl->read_state = CL_READ_REPLY_DATA;
     }
     else if (op->reply.hdr.opcode == OSD_OP_SEC_LIST && op->reply.hdr.retval > 0)
     {
         assert(!op->iov.count);
-        delete cl->read_op;
+        if (cl->read_op)
+            delete cl->read_op;
         cl->read_op = op;
         cl->read_state = CL_READ_REPLY_DATA;
         cl->read_remaining = sizeof(obj_ver_id) * op->reply.hdr.retval;
@@ -345,7 +347,8 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
     else if (op->reply.hdr.opcode == OSD_OP_SEC_READ_BMP && op->reply.hdr.retval > 0)
     {
         assert(!op->iov.count);
-        delete cl->read_op;
+        if (cl->read_op)
+            delete cl->read_op;
         cl->read_op = op;
         cl->read_state = CL_READ_REPLY_DATA;
         cl->read_remaining = op->reply.hdr.retval;
@@ -355,7 +358,8 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
     }
     else if (op->reply.hdr.opcode == OSD_OP_SHOW_CONFIG && op->reply.hdr.retval > 0)
     {
-        delete cl->read_op;
+        if (cl->read_op)
+            delete cl->read_op;
         cl->read_op = op;
         cl->read_state = CL_READ_REPLY_DATA;
         cl->read_remaining = op->reply.hdr.retval;
@@ -368,7 +372,8 @@ bool osd_messenger_t::handle_reply_hdr(osd_client_t *cl)
 reuse:
         // It's fine to reuse cl->read_op for the next reply
         handle_reply_ready(op);
-        cl->recv_list.push_back(cl->read_op->req.buf, OSD_PACKET_SIZE);
+        if (cl->read_op)
+            cl->recv_list.push_back(cl->read_op->req.buf, OSD_PACKET_SIZE);
         cl->read_remaining = OSD_PACKET_SIZE;
         cl->read_state = CL_READ_HDR;
     }

src/msgr_send.cpp

@@ -96,6 +96,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len });
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
     }
+    to_outbox[to_outbox.size()-1].flags |= MSGR_SENDP_LAST;
     if (cur_op->op_type == OSD_OP_IN)
     {
         to_outbox[to_outbox.size()-1].flags |= MSGR_SENDP_FREE;

src/msgr_stop.cpp

@@ -129,6 +129,7 @@ void osd_messenger_t::stop_client(int peer_fd, bool force, bool force_delete)
 #ifdef WITH_RDMA
     if (cl->rdma_conn)
     {
+        clients_by_qp.erase(cl->rdma_conn->qp->qp_num);
         delete cl->rdma_conn;
     }
 #endif

src/osd_secondary.cpp

@@ -166,15 +166,20 @@ void osd_t::exec_show_config(osd_op_t *cur_op)
     {
         // Indicate that RDMA is enabled
         wire_config["rdma_enabled"] = true;
-        if (req_json["connect_rdma"].is_string())
+        if (req_json["connect_rdma"].is_string() && req_json["rdma_op_memory"].uint64_value() != 0)
         {
             // Peer is trying to connect using RDMA, try to satisfy him
-            bool ok = msgr.connect_rdma(cur_op->peer_fd, req_json["connect_rdma"].string_value(), req_json["rdma_max_msg"].uint64_value());
+            bool ok = msgr.connect_rdma(cur_op->peer_fd, req_json["connect_rdma"].string_value(),
+                req_json["rdma_data_rkey"].uint64_value(), req_json["rdma_op_rkey"].uint64_value(),
+                req_json["rdma_op_slots"].uint64_value(), req_json["rdma_op_memory"].uint64_value());
             if (ok)
             {
                 auto rc = msgr.clients.at(cur_op->peer_fd)->rdma_conn;
                 wire_config["rdma_address"] = rc->addr.to_string();
-                wire_config["rdma_max_msg"] = rc->max_msg;
+                wire_config["rdma_data_rkey"] = (uint64_t)rc->in_data_mr->rkey;
+                wire_config["rdma_op_rkey"] = (uint64_t)rc->in_op_mr->rkey;
+                wire_config["rdma_op_slots"] = rc->op_slots;
+                wire_config["rdma_op_memory"] = rc->op_memory;
             }
         }
     }

src/test_freelist.cpp

@@ -0,0 +1,64 @@
+// Copyright (c) Vitaliy Filippov, 2023+
+// License: VNPL-1.1 or GNU GPL-2.0+ (see README.md for details)
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdexcept>
+#include "freelist.cpp"
+
+inline bool operator == (const freelist_item_t & a, const freelist_item_t & b)
+{
+    return a.start == b.start && a.size == b.size;
+}
+
+void dump(std::vector<freelist_item_t> & freelist)
+{
+    printf("free: ");
+    for (auto & item: freelist)
+    {
+        printf("%lx+%lx ", item.start, item.size);
+    }
+    printf("\n");
+}
+
+void dump(freelist_allocator_t &alloc)
+{
+    dump(alloc.freelist);
+}
+
+uint64_t test_alloc(freelist_allocator_t &alloc, uint64_t size)
+{
+    uint64_t r = alloc.alloc(size);
+    printf("alloc %lx: %lx\n", size, r);
+    return r;
+}
+
+void assert_eq(freelist_allocator_t &alloc, std::vector<freelist_item_t> v)
+{
+    if (alloc.freelist != v)
+    {
+        printf("expected ");
+        dump(v);
+        printf("got ");
+        dump(alloc);
+        throw std::runtime_error("test failed");
+    }
+    dump(alloc);
+}
+
+int main(int narg, char *args[])
+{
+    freelist_allocator_t alloc;
+    alloc.free(0, 0x1000000);
+    assert_eq(alloc, { { 0, 0x1000000 } });
+    assert(test_alloc(alloc, 0x1000) == 0);
+    assert_eq(alloc, { { 0x1000, 0xfff000 } });
+    assert(test_alloc(alloc, 0x4000) == 0x1000);
+    alloc.free(0x1000000, 0x4000);
+    assert_eq(alloc, { { 0x5000, 0xfff000 } });
+    alloc.free(0, 0x1000);
+    assert_eq(alloc, { { 0, 0x1000 }, { 0x5000, 0xfff000 } });
+    alloc.free(0x1000, 0x4000);
+    assert_eq(alloc, { { 0, 0x1004000 } });
+    return 0;
+}