Use EPOLLOUT in no-io_uring mode

.gitea/workflows/test.yml

@@ -324,6 +324,24 @@ jobs:
           echo ""
         done
 
+  test_no_uring:
+    runs-on: ubuntu-latest
+    needs: build
+    container: ${{env.TEST_IMAGE}}:${{github.sha}}
+    steps:
+    - name: Run test
+      id: test
+      timeout-minutes: 3
+      run: /root/vitastor/tests/test_no_uring.sh
+    - name: Print logs
+      if: always() && steps.test.outcome == 'failure'
+      run: |
+        for i in /root/vitastor/testdata/*.log /root/vitastor/testdata/*.txt; do
+          echo "-------- $i --------"
+          cat $i
+          echo ""
+        done
+
   test_interrupted_rebalance:
     runs-on: ubuntu-latest
     needs: build

src/client/messenger.cpp

@@ -504,18 +504,21 @@ void osd_messenger_t::handle_connect_epoll(int peer_fd)
     int one = 1;
     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)
+    // EPOLLOUT is only required without io_uring, to continue partially sent messages
+    tfd->set_fd_handler(peer_fd, use_sync_send_recv, [this](int peer_fd, int epoll_events)
     {
         handle_peer_epoll(peer_fd, epoll_events);
     });
     // Check OSD number
-    init_tls_client(cl);
+    if (!init_tls_client(cl))
+    {
+        return;
+    }
     check_peer_config(cl);
 }
 
 void osd_messenger_t::handle_peer_epoll(int peer_fd, int epoll_events)
 {
-    // Mark client as ready (i.e. some data is available)
     auto cl = clients_by_fd.at(peer_fd);
     if (epoll_events & EPOLLRDHUP)
     {
@@ -525,8 +528,9 @@ void osd_messenger_t::handle_peer_epoll(int peer_fd, int epoll_events)
             fprintf(stderr, "[OSD %ju] client %ju disconnected\n", this->osd_num, cl->client_id);
         }
         stop_client(cl->client_id);
+        return;
     }
-    else if (epoll_events & EPOLLIN)
+    if (epoll_events & EPOLLIN)
     {
         // Mark client as ready (i.e. some data is available)
         cl->read_ready++;
@@ -539,6 +543,23 @@ void osd_messenger_t::handle_peer_epoll(int peer_fd, int epoll_events)
                 read_requests();
         }
     }
+    if (epoll_events & EPOLLOUT)
+    {
+        // The socket became writable again - continue the interrupted use_sync_send_recv
+        // read_requests() above may have stopped the client, so look it up again
+        auto cl_it = clients.find(cl->client_id);
+        if (cl_it == clients.end())
+        {
+            return;
+        }
+        cl = cl_it->second;
+        cl->write_ready = true;
+        // send_list is only left non-empty when we're waiting for EPOLLOUT
+        if (cl->send_list.size() && cl->peer_state != PEER_RDMA)
+        {
+            try_send(cl);
+        }
+    }
 }
 
 void osd_messenger_t::on_connect_peer(osd_num_t peer_osd, int errcode, uint64_t client_id)
@@ -770,9 +791,12 @@ void osd_messenger_t::accept_connections(int listen_fd)
         cl->peer_fd = peer_fd;
         cl->peer_state = PEER_CONNECTED;
         cl->in_buf = (uint8_t*)malloc_or_die(receive_buffer_size);
-        init_tls_client(cl);
-        // Add FD to epoll
-        tfd->set_fd_handler(peer_fd, false, [this](int peer_fd, int epoll_events)
+        if (!init_tls_client(cl))
+        {
+            continue;
+        }
+        // Add FD to epoll. EPOLLOUT is only required without io_uring
+        tfd->set_fd_handler(peer_fd, use_sync_send_recv, [this](int peer_fd, int epoll_events)
         {
             handle_peer_epoll(peer_fd, epoll_events);
         });

src/client/messenger.h

@@ -147,6 +147,8 @@ struct osd_client_t
     size_t write_op_pos = 0;
     msghdr write_msg = { 0 };
     int write_state = 0;
+    // Socket is writable. Only tracked with use_sync_send_recv, io_uring doesn't need it
+    bool write_ready = true;
     std::vector<iovec> send_list;
     size_t send_list_size = 0;
     std::deque<osd_op_t*> send_free_ops;
@@ -248,7 +250,7 @@ protected:
 
     void init_tls();
     void destroy_tls();
-    void init_tls_client(osd_client_t *cl);
+    bool init_tls_client(osd_client_t *cl);
     bool do_tls_handshake(osd_client_t *cl, bool from_recv = false);
     bool derive_aes_keys(osd_client_t *cl, bool update_my, bool update_peer);
 
@@ -343,7 +345,8 @@ protected:
     void cancel_op(osd_op_t *op);
 
     bool try_send(osd_client_t *cl);
-    void handle_send(int result, bool prev, bool more, osd_client_t *cl);
+    bool wakeup_send(osd_client_t *cl);
+    bool handle_send(int result, bool prev, bool more, osd_client_t *cl);
     static void post_send_free(osd_op_t *op);
     bool op_write_to(osd_client_t *cl, msgr_op_writer_t & wr);
     void next_write_op(osd_client_t *cl);
@@ -370,7 +373,7 @@ protected:
     void op_decrypt_free(osd_client_t* cl);
 
 #ifdef WITH_RDMA
-    void try_send_rdma(osd_client_t *cl);
+    bool try_send_rdma(osd_client_t *cl);
     bool init_recv_rdma(osd_client_t *cl);
     void handle_rdma_events(msgr_rdma_context_t *rdma_context);
     msgr_rdma_context_t* choose_rdma_context(osd_client_t *cl);

src/client/msgr_encrypt.cpp

@@ -616,7 +616,7 @@ void osd_messenger_t::init_tls()
     }
 }
 
-void osd_messenger_t::init_tls_client(osd_client_t *cl)
+bool osd_messenger_t::init_tls_client(osd_client_t *cl)
 {
     if (gcm_enabled)
     {
@@ -625,13 +625,13 @@ void osd_messenger_t::init_tls_client(osd_client_t *cl)
         cl->hs->init(cl->is_incoming);
         if (cl->hs->out_size())
         {
-            if (cl->write_state == 0)
+            if (!wakeup_send(cl))
             {
-                cl->write_state = CL_WRITE_READY;
-                write_ready_clients.push_back(cl->client_id);
+                return false;
             }
         }
     }
+    return true;
 }
 
 void osd_messenger_t::destroy_tls()

src/client/msgr_rdma.cpp

@@ -571,7 +571,7 @@ static void try_send_rdma_wr(osd_client_t *cl, ibv_sge *sge, int op_sge)
     cl->rdma_conn->cur_send++;
 }
 
-void osd_messenger_t::try_send_rdma(osd_client_t *cl)
+bool osd_messenger_t::try_send_rdma(osd_client_t *cl)
 {
     auto rc = cl->rdma_conn;
     if (!rc->send_out_size)
@@ -602,7 +602,7 @@ void osd_messenger_t::try_send_rdma(osd_client_t *cl)
         if (cl->io_error)
         {
             stop_client(cl->client_id);
-            return;
+            return false;
         }
         if (copied > 0)
         {
@@ -622,6 +622,7 @@ void osd_messenger_t::try_send_rdma(osd_client_t *cl)
             cl->send_free_ops.push_back(NULL); // end marker
         }
     }
+    return true;
 }
 
 static void try_recv_rdma_wr(osd_client_t *cl, void *buf)
@@ -750,6 +751,7 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
             }
         }
     } while (event_count > 0);
+    handle_immediate_ops();
 }
 
 void osd_messenger_t::destroy_rdma_conn(msgr_rdma_connection_t *rdma_conn)

src/client/msgr_rdmacm.cpp

@@ -503,7 +503,6 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
     cl->osd_num = peer_osd;
     cl->in_buf = (uint8_t*)malloc_or_die(receive_buffer_size);
     cl->rdma_conn = rc;
-    init_tls_client(cl);
     clients[conn->client_id] = cl;
     if (conn->timeout_id >= 0)
     {
@@ -517,6 +516,10 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
         fprintf(stderr, "Successfully connected with OSD %ju using RDMA-CM\n", peer_osd);
     }
     // Add initial receive request(s)
+    if (!init_tls_client(cl))
+    {
+        return;
+    }
     init_recv_rdma(cl);
     if (peer_osd)
     {

src/client/msgr_receive.cpp

@@ -458,8 +458,7 @@ void osd_messenger_t::read_requests()
             {
                 result = -errno;
             }
-            // like set_immediate
-            tfd->set_timer_us(0, false, [this, result, cl](int){ handle_read(result, cl); });
+            handle_read(result, cl);
         }
     }
     read_ready_clients.clear();
@@ -490,8 +489,10 @@ void osd_messenger_t::handle_read(int result, osd_client_t *cl)
         }
         stop_client(cl->client_id);
 out_wakeup:
-        if (set_immediate_ops.size())
+        if (ringloop && set_immediate_ops.size())
+        {
             ringloop->wakeup();
+        }
         return;
     }
     bool full_read = false;
@@ -501,7 +502,9 @@ out_wakeup:
         {
             full_read = result >= cl->read_iov.iov_len;
             if (!handle_read_buffer(cl, cl->in_buf, result))
+            {
                 goto out_wakeup;
+            }
         }
         else
         {
@@ -597,13 +600,16 @@ bool osd_messenger_t::handle_read_buffer(osd_client_t *cl, uint8_t *curbuf, size
             bufsize -= done;
             if (cl->hs->out_size())
             {
-                if (cl->write_state == 0)
+                if (!wakeup_send(cl))
                 {
-                    cl->write_state = CL_WRITE_READY;
-                    write_ready_clients.push_back(cl->client_id);
+                    return false;
                 }
             }
-            if (cl->hs->done() && !cl->hs->out_size())
+            if (!cl->hs)
+            {
+                // Deleted by wakeup_send with use_sync_send_recv
+            }
+            else if (cl->hs->done() && !cl->hs->out_size())
             {
                 // Delete hs when done and nothing to send
                 delete cl->hs;

src/client/msgr_send.cpp

@@ -17,6 +17,10 @@
 
 #define GCM_TMP_BUF_SIZE 4096
 
+// How many times to retry sendmsg() in place before falling back to EPOLLOUT.
+// Only used with use_sync_send_recv - io_uring waits for the socket by itself
+#define MSGR_SEND_RETRIES 4
+
 class msgr_op_writer_t
 {
 public:
@@ -467,36 +471,60 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
         return;
     }
 #endif
-    if (!ringloop)
+    try_send(cl);
+}
+
+bool osd_messenger_t::wakeup_send(osd_client_t *cl)
+{
+    if (cl->peer_state == PEER_RDMA)
     {
-        // FIXME: It's worse because it doesn't allow batching
-        while (cl->write_op || cl->write_ops.size())
-        {
-            try_send(cl);
-        }
+        return try_send_rdma(cl);
     }
-    else
+    if (use_sync_send_recv)
     {
-        if (!try_send(cl) && cl->write_state == 0)
-        {
-            cl->write_state = CL_WRITE_READY;
-            write_ready_clients.push_back(cl->client_id);
-        }
-        ringloop->wakeup();
+        // FIXME: Batch send/recv without io_uring too.
+        return try_send(cl);
     }
+    if (cl->write_state == 0)
+    {
+        cl->write_state = CL_WRITE_READY;
+        write_ready_clients.push_back(cl->client_id);
+    }
+    ringloop->wakeup();
+    return true;
 }
 
 bool osd_messenger_t::try_send(osd_client_t *cl)
 {
+    assert(cl->peer_state != PEER_RDMA);
+restart:
     if (cl->peer_state == PEER_STOPPED || cl->peer_fd < 0)
     {
         return true;
     }
-    if (cl->write_msg.msg_iovlen > 0 || !ringloop->space_left() && !use_sync_send_recv)
+    if (cl->write_msg.msg_iovlen > 0)
     {
-        return false;
+        return true;
+    }
+    if (!use_sync_send_recv)
+    {
+        if (!ringloop->space_left())
+        {
+            if (cl->write_state == 0)
+            {
+                cl->write_state = CL_WRITE_READY;
+                write_ready_clients.push_back(cl->client_id);
+            }
+            ringloop->wakeup();
+            return true;
+        }
+    }
+    else if (!cl->write_ready)
+    {
+        // Socket buffer is full, the unsent tail is still in send_list.
+        // Do not copy more data into it, just wait for EPOLLOUT
+        return true;
     }
-    assert(cl->peer_state != PEER_RDMA);
     if (cl->hs)
     {
         // Send handshake message
@@ -523,11 +551,10 @@ copy_ops:
     if (cl->io_error)
     {
         stop_client(cl->client_id);
-        return true;
+        return false;
     }
     if (!cl->send_list.size())
     {
-        cl->write_state = 0;
         return true;
     }
     // get_op_writer_t already guarantees that it's under IOV_MAX
@@ -575,16 +602,29 @@ copy_ops:
     }
     else
     {
-        cl->write_msg.msg_iov = cl->send_list.data();
-        cl->write_msg.msg_iovlen = cl->send_list.size();
-        cl->refs++;
-        int result = sendmsg(cl->peer_fd, &cl->write_msg, MSG_NOSIGNAL);
-        if (result < 0)
+        // Try to first send with a busy loop, then postpone to the EPOLLOUT handler
+        int send_retries = MSGR_SEND_RETRIES;
+        while (cl->send_list.size())
         {
-            result = -errno;
+            cl->write_msg.msg_iov = cl->send_list.data();
+            cl->write_msg.msg_iovlen = cl->send_list.size();
+            cl->refs++;
+            int result = sendmsg(cl->peer_fd, &cl->write_msg, MSG_NOSIGNAL);
+            if (result < 0)
+            {
+                result = -errno;
+            }
+            if (!handle_send(result, false, false, cl))
+            {
+                return false;
+            }
+            if (cl->send_list.size() && --send_retries <= 0)
+            {
+                cl->write_ready = false;
+                return true;
+            }
         }
-        // like set_immediate
-        tfd->set_timer_us(0, false, [this, result, cl](int){ handle_send(result, false, false, cl); });
+        goto restart;
     }
     return true;
 }
@@ -664,21 +704,42 @@ void osd_messenger_t::send_replies()
     {
         uint64_t client_id = write_ready_clients[i];
         auto cl_it = clients.find(client_id);
-        if (cl_it != clients.end() && cl_it->second->peer_state != PEER_RDMA && !try_send(cl_it->second))
+        if (cl_it != clients.end() && cl_it->second->peer_state != PEER_RDMA)
         {
-            write_ready_clients.erase(write_ready_clients.begin(), write_ready_clients.begin() + i);
-            return;
+            if (!ringloop->space_left())
+            {
+                write_ready_clients.erase(write_ready_clients.begin(), write_ready_clients.begin() + i);
+                return;
+            }
+            cl_it->second->write_state = 0;
+            try_send(cl_it->second);
         }
     }
     write_ready_clients.clear();
 }
 
-void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t *cl)
+static void eat_send_list(osd_client_t *cl, size_t n)
+{
+    int i = 0;
+    while (i < cl->send_list.size() && n >= cl->send_list[i].iov_len)
+    {
+        n -= cl->send_list[i].iov_len;
+        i++;
+    }
+    if (i > 0)
+        cl->send_list.erase(cl->send_list.begin(), cl->send_list.begin()+i);
+    if (cl->send_list.size())
+    {
+        cl->send_list[0].iov_base = (uint8_t*)cl->send_list[0].iov_base + n;
+        cl->send_list[0].iov_len -= n;
+    }
+}
+
+bool osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t *cl)
 {
     if (!prev)
     {
         cl->write_msg.msg_iovlen = 0;
-        cl->send_list.clear();
     }
     if (!more)
     {
@@ -690,14 +751,15 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
         {
             destroy_client(cl);
         }
-        return;
+        // FIXME: Return error codes instead of true/false (?)
+        return false;
     }
     if (result < 0 && result != -EAGAIN && result != -EINTR)
     {
         // this is a client socket, so don't panic. just disconnect it
         fprintf(stderr, "Client %ju socket write error: %d (%s). Disconnecting client\n", cl->client_id, -result, strerror(-result));
         stop_client(cl->client_id);
-        return;
+        return false;
     }
     if (result >= 0)
     {
@@ -709,14 +771,26 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
                 post_send_free(cl->zc_free_list[i]);
             if (i > 0)
                 cl->zc_free_list.erase(cl->zc_free_list.begin(), cl->zc_free_list.begin()+i+1);
-            return;
+            return true;
         }
-        if (cl->send_list_size > result)
+        if (cl->send_list_size <= result)
+        {
+            cl->send_list.clear();
+            cl->send_list_size = 0;
+        }
+        else if (!use_sync_send_recv)
         {
             fprintf(stderr, "Client %ju socket write error: expected to send "
                 "%zu bytes with MSG_WAITALL but sent %u. Disconnecting client\n", cl->client_id, cl->send_list_size, result);
             stop_client(cl->client_id);
-            return;
+            return false;
+        }
+        else
+        {
+            // Partially sent
+            eat_send_list(cl, result);
+            cl->send_list_size -= result;
+            return true;
         }
         for (auto op: cl->send_free_ops)
         {
@@ -728,9 +802,6 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
         if (more)
             cl->zc_free_list.push_back(NULL); // end marker
         cl->send_free_ops.clear();
-        cl->write_state = 0;
-        if (cl->write_op || cl->write_ops.size())
-            cl->write_state = CL_WRITE_READY;
         if ((cl->proto_csum_status & MSGR_CSUM_NEG) && !cl->write_op && !cl->write_ops.size())
         {
             // Checksums negotiated, enable
@@ -750,10 +821,16 @@ void osd_messenger_t::handle_send(int result, bool prev, bool more, osd_client_t
         }
 #endif
     }
-    if (cl->write_state != 0)
+    if ((cl->write_op || cl->write_ops.size() || cl->hs && cl->hs->out_size()) && !use_sync_send_recv)
     {
-        write_ready_clients.push_back(cl->client_id);
+        if (cl->write_state == 0)
+        {
+            cl->write_state = CL_WRITE_READY;
+            write_ready_clients.push_back(cl->client_id);
+        }
+        ringloop->wakeup();
     }
+    return true;
 }
 
 static inline bool op_has_data_for_ssl(osd_op_t *op)

src/test/mock/messenger.cpp

@@ -73,6 +73,11 @@ void osd_messenger_t::accept_connections(int listen_fd)
 {
 }
 
+bool osd_messenger_t::wakeup_send(osd_client_t *cl)
+{
+    return true;
+}
+
 #ifdef WITH_RDMA
 json11::Json osd_messenger_t::connect_rdma(uint64_t client_id, std::string rdma_address, uint64_t client_max_msg)
 {

tests/run_tests.sh

@@ -27,6 +27,8 @@ ETCD_SCHEME=https ./test_snapshot.sh
 
 ./test_img_online_resize.sh
 
+./test_no_uring.sh
+
 ./test_interrupted_rebalance.sh
 IMMEDIATE_COMMIT=1 ./test_interrupted_rebalance.sh
 SCHEME=ec ./test_interrupted_rebalance.sh

tests/test_no_uring.sh

@@ -0,0 +1,8 @@
+#!/bin/bash -ex
+
+. `dirname $0`/run_3osds.sh
+
+$VITASTOR_FIO -bs=128k -direct=1 -numjobs=1 -iodepth=4 \
+    -rw=write -pool=1 -inode=1 -size=128M -runtime=10 -no_io_uring=1
+
+format_green OK