Experiment: zero-copy TCP send

docs/params/network.yml

@@ -30,6 +30,18 @@
     будут использоваться обычные синхронные системные вызовы send/recv. Для OSD
     это бессмысленно, так как OSD в любом случае нуждается в io_uring, но, в
     принципе, это может применяться для клиентов со старыми версиями ядра.
+- name: use_zerocopy_send
+  type: bool
+  default: false
+  info: |
+    If true, OSDs and clients will attempt to use TCP zero-copy send
+    (MSG_ZEROCOPY) for big buffers. It's recommended to raise net.ipv4.tcp_wmem
+    and net.core.wmem_max sysctls when using this mode.
+  info_ru: |
+    Если установлено в true, то OSD и клиенты будут стараться использовать
+    TCP-отправку без копирования (MSG_ZEROCOPY) для больших буферов данных.
+    Рекомендуется поднять значения sysctl net.ipv4.tcp_wmem и net.core.wmem_max
+    при использовании этого режима.
 - name: use_rdma
   type: bool
   default: true

mon/mon.js

@@ -64,6 +64,7 @@ const etcd_tree = {
             // client and osd
             tcp_header_buffer_size: 65536,
             use_sync_send_recv: false,
+            use_zerocopy_send: false,
             use_rdma: true,
             rdma_device: null, // for example, "rocep5s0f0"
             rdma_port_num: 1,

src/messenger.cpp

@@ -39,6 +39,12 @@ void osd_messenger_t::init()
             handle_rdma_events();
         }
     }
+#endif
+#ifndef SO_ZEROCOPY
+    if (log_level > 0)
+    {
+        fprintf(stderr, "Zero-copy TCP send is not supported in this build, ignoring\n");
+    }
 #endif
     keepalive_timer_id = tfd->set_timer(1000, true, [this](int)
     {
@@ -162,6 +168,8 @@ void osd_messenger_t::parse_config(const json11::Json & config)
         this->receive_buffer_size = 65536;
     this->use_sync_send_recv = config["use_sync_send_recv"].bool_value() ||
         config["use_sync_send_recv"].uint64_value();
+    this->use_zerocopy_send = config["use_zerocopy_send"].bool_value() ||
+        config["use_zerocopy_send"].uint64_value();
     this->peer_connect_interval = config["peer_connect_interval"].uint64_value();
     if (!this->peer_connect_interval)
         this->peer_connect_interval = 5;
@@ -288,8 +296,7 @@ void osd_messenger_t::handle_connect_epoll(int peer_fd)
         on_connect_peer(peer_osd, -result);
         return;
     }
-    int one = 1;
+    set_socket_options(cl);
-    setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
     cl->peer_state = PEER_CONNECTED;
     tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
     {
@@ -299,6 +306,23 @@ void osd_messenger_t::handle_connect_epoll(int peer_fd)
     check_peer_config(cl);
 }
 
+void osd_messenger_t::set_socket_options(osd_client_t *cl)
+{
+    int one = 1;
+    setsockopt(cl->peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
+#ifdef SO_ZEROCOPY
+    if (!use_zerocopy_send)
+        cl->zerocopy_send = false;
+    else if (setsockopt(cl->peer_fd, SOL_SOCKET, SO_ZEROCOPY, &one, sizeof(one)) != 0)
+    {
+        if (log_level > 0)
+            fprintf(stderr, "[OSD %lu] Failed to enable zero-copy send for client %d: %s\n", this->osd_num, cl->peer_fd, strerror(errno));
+    }
+    else
+        cl->zerocopy_send = true;
+#endif
+}
+
 void osd_messenger_t::handle_peer_epoll(int peer_fd, int epoll_events)
 {
     // Mark client as ready (i.e. some data is available)
@@ -493,14 +517,13 @@ void osd_messenger_t::accept_connections(int listen_fd)
         fprintf(stderr, "[OSD %lu] new client %d: connection from %s\n", this->osd_num, peer_fd,
             addr_to_string(addr).c_str());
         fcntl(peer_fd, F_SETFL, fcntl(peer_fd, F_GETFL, 0) | O_NONBLOCK);
-        int one = 1;
+        auto cl = clients[peer_fd] = new osd_client_t();
-        setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
+        cl->peer_addr = addr;
-        clients[peer_fd] = new osd_client_t();
+        cl->peer_port = ntohs(((sockaddr_in*)&addr)->sin_port);
-        clients[peer_fd]->peer_addr = addr;
+        cl->peer_fd = peer_fd;
-        clients[peer_fd]->peer_port = ntohs(((sockaddr_in*)&addr)->sin_port);
+        cl->peer_state = PEER_CONNECTED;
-        clients[peer_fd]->peer_fd = peer_fd;
+        cl->in_buf = malloc_or_die(receive_buffer_size);
-        clients[peer_fd]->peer_state = PEER_CONNECTED;
+        set_socket_options(cl);
-        clients[peer_fd]->in_buf = malloc_or_die(receive_buffer_size);
         // Add FD to epoll
         tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
         {

src/messenger.h

@@ -45,6 +45,12 @@ struct msgr_sendp_t
     int flags;
 };
 
+struct msgr_zc_not_t
+{
+    osd_op_t *op;
+    uint32_t nsend;
+};
+
 struct osd_client_t
 {
     int refs = 0;
@@ -57,6 +63,7 @@ struct osd_client_t
     int ping_time_remaining = 0;
     int idle_time_remaining = 0;
     osd_num_t osd_num = 0;
+    bool zerocopy_send = false;
 
     void *in_buf = NULL;
 
@@ -87,6 +94,12 @@ struct osd_client_t
     int write_state = 0;
     std::vector<iovec> send_list, next_send_list;
     std::vector<msgr_sendp_t> outbox, next_outbox;
+    std::vector<msgr_zc_not_t> zerocopy_sent;
+    uint64_t outbox_size = 0, next_outbox_size = 0;
+    uint32_t zerocopy_notification_idx = 0;
+    uint32_t zerocopy_notification_prev = 0;
+    uint8_t zerocopy_notification_buf[256];
+    struct msghdr zerocopy_notification_msg;
 
     ~osd_client_t()
     {
@@ -127,6 +140,7 @@ protected:
     int osd_ping_timeout = 0;
     int log_level = 0;
     bool use_sync_send_recv = false;
+    bool use_zerocopy_send = false;
 
 #ifdef WITH_RDMA
     bool use_rdma = true;
@@ -181,10 +195,12 @@ protected:
     void check_peer_config(osd_client_t *cl);
     void cancel_osd_ops(osd_client_t *cl);
     void cancel_op(osd_op_t *op);
+    void set_socket_options(osd_client_t *cl);
 
     bool try_send(osd_client_t *cl);
     void measure_exec(osd_op_t *cur_op);
     void handle_send(int result, osd_client_t *cl);
+    void handle_zerocopy_notification(osd_client_t *cl, int res);
 
     bool handle_read(int result, osd_client_t *cl);
     bool handle_read_buffer(osd_client_t *cl, void *curbuf, int remain);

src/msgr_send.cpp

@@ -6,6 +6,12 @@
 
 #include "messenger.h"
 
+#include <linux/errqueue.h>
+
+#ifndef MSG_ZEROCOPY
+#define MSG_ZEROCOPY 0
+#endif
+
 void osd_messenger_t::outbox_push(osd_op_t *cur_op)
 {
     assert(cur_op->peer_fd);
@@ -36,6 +42,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
     }
     auto & to_send_list = cl->write_msg.msg_iovlen ? cl->next_send_list : cl->send_list;
     auto & to_outbox = cl->write_msg.msg_iovlen ? cl->next_outbox : cl->outbox;
+    auto & to_size = cl->write_msg.msg_iovlen ? cl->next_outbox_size : cl->outbox_size;
     if (cur_op->op_type == OSD_OP_IN)
     {
         measure_exec(cur_op);
@@ -46,6 +53,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
         to_send_list.push_back((iovec){ .iov_base = cur_op->req.buf, .iov_len = OSD_PACKET_SIZE });
         cl->sent_ops[cur_op->req.hdr.id] = cur_op;
     }
+    to_size += OSD_PACKET_SIZE;
     to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = MSGR_SENDP_HDR });
     // Bitmap
     if (cur_op->op_type == OSD_OP_IN &&
@@ -57,6 +65,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             .iov_len = cur_op->reply.sec_rw.attr_len,
         });
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        to_size += cur_op->reply.sec_rw.attr_len;
     }
     else if (cur_op->op_type == OSD_OP_OUT &&
         (cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE || cur_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE) &&
@@ -67,6 +76,7 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             .iov_len = cur_op->req.sec_rw.attr_len,
         });
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+        to_size += cur_op->req.sec_rw.attr_len;
     }
     // Operation data
     if ((cur_op->op_type == OSD_OP_IN
@@ -86,14 +96,21 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
             assert(cur_op->iov.buf[i].iov_base);
             to_send_list.push_back(cur_op->iov.buf[i]);
             to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
+            to_size += cur_op->iov.buf[i].iov_len;
         }
     }
     if (cur_op->req.hdr.opcode == OSD_OP_SEC_READ_BMP)
     {
         if (cur_op->op_type == OSD_OP_IN && cur_op->reply.hdr.retval > 0)
+        {
             to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->reply.hdr.retval });
+            to_size += cur_op->reply.hdr.retval;
+        }
         else if (cur_op->op_type == OSD_OP_OUT && cur_op->req.sec_read_bmp.len > 0)
+        {
             to_send_list.push_back((iovec){ .iov_base = cur_op->buf, .iov_len = (size_t)cur_op->req.sec_read_bmp.len });
+            to_size += cur_op->req.sec_read_bmp.len;
+        }
         to_outbox.push_back((msgr_sendp_t){ .op = cur_op, .flags = 0 });
     }
     if (cur_op->op_type == OSD_OP_IN)
@@ -177,17 +194,19 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
         }
         cl->write_msg.msg_iov = cl->send_list.data();
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+        cl->write_msg.msg_flags = (cl->zerocopy_send && (cl->outbox_size/cl->send_list.size()) >= 4096 ? MSG_ZEROCOPY : 0);
         cl->refs++;
         ring_data_t* data = ((ring_data_t*)sqe->user_data);
         data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, cl); };
-        my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, 0);
+        my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, cl->write_msg.msg_flags);
     }
     else
     {
         cl->write_msg.msg_iov = cl->send_list.data();
         cl->write_msg.msg_iovlen = cl->send_list.size() < IOV_MAX ? cl->send_list.size() : IOV_MAX;
+        cl->write_msg.msg_flags = (cl->zerocopy_send && (cl->outbox_size/cl->send_list.size()) >= 4096 ? MSG_ZEROCOPY : 0);
         cl->refs++;
-        int result = sendmsg(peer_fd, &cl->write_msg, MSG_NOSIGNAL);
+        int result = sendmsg(peer_fd, &cl->write_msg, MSG_NOSIGNAL | cl->write_msg.msg_flags);
         if (result < 0)
         {
             result = -errno;
@@ -197,6 +216,62 @@ bool osd_messenger_t::try_send(osd_client_t *cl)
     return true;
 }
 
+void osd_messenger_t::handle_zerocopy_notification(osd_client_t *cl, int res)
+{
+    cl->refs--;
+    if (cl->peer_state == PEER_STOPPED)
+    {
+        if (cl->refs <= 0)
+        {
+            delete cl;
+        }
+        return;
+    }
+    if (res != 0)
+    {
+        return;
+    }
+    if (cl->zerocopy_notification_msg.msg_flags & MSG_CTRUNC)
+    {
+        fprintf(stderr, "zero-copy send notification truncated on client socket %d\n", cl->peer_fd);
+        return;
+    }
+    for (struct cmsghdr *cm = CMSG_FIRSTHDR(&cl->zerocopy_notification_msg); cm; cm = CMSG_NXTHDR(&cl->zerocopy_notification_msg, cm))
+    {
+        if (cm->cmsg_level == SOL_IP && cm->cmsg_type == IP_RECVERR)
+        {
+            struct sock_extended_err *serr = (struct sock_extended_err*)CMSG_DATA(cm);
+            if (serr->ee_errno == 0 && serr->ee_origin == SO_EE_ORIGIN_ZEROCOPY)
+            {
+                // completed sends numbered serr->ee_info .. serr->ee_data
+                int start = 0;
+                while (start < cl->zerocopy_sent.size() && cl->zerocopy_sent[start].nsend < serr->ee_info)
+                    start++;
+                int end = start;
+                if (serr->ee_data < serr->ee_info)
+                {
+                    // counter has wrapped around
+                    while (end < cl->zerocopy_sent.size() && cl->zerocopy_sent[end].nsend >= cl->zerocopy_sent[start].nsend)
+                        end++;
+                }
+                while (end < cl->zerocopy_sent.size() && cl->zerocopy_sent[end].nsend <= serr->ee_data)
+                    end++;
+                if (end > start)
+                {
+                    for (int i = start; i < end; i++)
+                    {
+                        delete cl->zerocopy_sent[i].op;
+                    }
+                    cl->zerocopy_sent.erase(
+                        cl->zerocopy_sent.begin() + start,
+                        cl->zerocopy_sent.begin() + end
+                    );
+                }
+            }
+        }
+    }
+}
+
 void osd_messenger_t::send_replies()
 {
     for (int i = 0; i < write_ready_clients.size(); i++)
@@ -224,16 +299,19 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
         }
         return;
     }
-    if (result < 0 && result != -EAGAIN && result != -EINTR)
+    if (result < 0 && result != -EAGAIN && result != -EINTR && result != -ENOBUFS)
     {
         // this is a client socket, so don't panic. just disconnect it
         fprintf(stderr, "Client %d socket write error: %d (%s). Disconnecting client\n", cl->peer_fd, -result, strerror(-result));
         stop_client(cl->peer_fd);
         return;
     }
+    bool used_zerocopy = false;
     if (result >= 0)
     {
+        used_zerocopy = (cl->write_msg.msg_flags & MSG_ZEROCOPY) ? true : false;
         int done = 0;
+        int bytes_written = result;
         while (result > 0 && done < cl->send_list.size())
         {
             iovec & iov = cl->send_list[done];
@@ -242,8 +320,20 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 if (cl->outbox[done].flags & MSGR_SENDP_FREE)
                 {
                     // Reply fully sent
+                    if (!used_zerocopy)
+                    {
                         delete cl->outbox[done].op;
                     }
+                    else
+                    {
+                        // With zero-copy send the difference is that we must keep the buffer (i.e. the operation)
+                        // allocated until we get send notification from MSG_ERRQUEUE
+                        cl->zerocopy_sent.push_back((msgr_zc_not_t){
+                            .op = cl->outbox[done].op,
+                            .nsend = cl->zerocopy_notification_idx,
+                        });
+                    }
+                }
                 result -= iov.iov_len;
                 done++;
             }
@@ -254,6 +344,11 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
                 break;
             }
         }
+        if (used_zerocopy)
+        {
+            cl->zerocopy_notification_idx++;
+        }
+        cl->outbox_size -= bytes_written;
         if (done > 0)
         {
             cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+done);
@@ -263,8 +358,10 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
         {
             cl->send_list.insert(cl->send_list.end(), cl->next_send_list.begin(), cl->next_send_list.end());
             cl->outbox.insert(cl->outbox.end(), cl->next_outbox.begin(), cl->next_outbox.end());
+            cl->outbox_size += cl->next_outbox_size;
             cl->next_send_list.clear();
             cl->next_outbox.clear();
+            cl->next_outbox_size = 0;
         }
         cl->write_state = cl->outbox.size() > 0 ? CL_WRITE_READY : 0;
 #ifdef WITH_RDMA
@@ -287,4 +384,34 @@ void osd_messenger_t::handle_send(int result, osd_client_t *cl)
     {
         write_ready_clients.push_back(cl->peer_fd);
     }
+    if (used_zerocopy && (cl->zerocopy_notification_idx-cl->zerocopy_notification_prev) >= 16 &&
+        cl->zerocopy_sent.size() > 0)
+    {
+        cl->zerocopy_notification_prev = cl->zerocopy_notification_idx;
+        cl->zerocopy_notification_msg = {
+            .msg_control = cl->zerocopy_notification_buf,
+            .msg_controllen = sizeof(cl->zerocopy_notification_buf),
+        };
+        cl->refs++;
+        io_uring_sqe* sqe = NULL;
+        if (ringloop && !use_sync_send_recv)
+        {
+            sqe = ringloop->get_sqe();
+        }
+        if (!sqe)
+        {
+            int res = recvmsg(cl->peer_fd, &cl->zerocopy_notification_msg, MSG_ERRQUEUE|MSG_DONTWAIT);
+            if (res < 0)
+            {
+                res = -errno;
+            }
+            handle_zerocopy_notification(cl, res);
+        }
+        else
+        {
+            ring_data_t* data = ((ring_data_t*)sqe->user_data);
+            data->callback = [this, cl](ring_data_t *data) { handle_zerocopy_notification(cl, data->res); };
+            my_uring_prep_recvmsg(sqe, cl->peer_fd, &cl->zerocopy_notification_msg, MSG_ERRQUEUE);
+        }
+    }
 }