Test: use submit_and_wait()

Makefile

@@ -2,7 +2,7 @@ BLOCKSTORE_OBJS := allocator.o blockstore.o blockstore_impl.o blockstore_init.o
 	blockstore_write.o blockstore_sync.o blockstore_stable.o blockstore_rollback.o blockstore_flush.o crc32c.o ringloop.o timerfd_interval.o
 # -fsanitize=address
 CXXFLAGS := -g -O3 -Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fPIC -fdiagnostics-color=always
-all: $(BLOCKSTORE_OBJS) libfio_blockstore.so osd libfio_sec_osd.so test_blockstore stub_osd osd_test
+all: $(BLOCKSTORE_OBJS) libfio_blockstore.so osd libfio_sec_osd.so test_blockstore stub_osd stub_bench osd_test
 clean:
 	rm -f *.o
 
@@ -14,7 +14,7 @@ allocator.o: allocator.cpp allocator.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 ringloop.o: ringloop.cpp ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
-timerfd_interval.o: timerfd_interval.cpp timerfd_interval.h
+timerfd_interval.o: timerfd_interval.cpp timerfd_interval.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 
 %.o: %.cpp allocator.h blockstore_flush.h blockstore.h blockstore_impl.h blockstore_init.h blockstore_journal.h crc32c.h ringloop.h timerfd_interval.h object_id.h
@@ -26,13 +26,13 @@ libfio_blockstore.so: ./libblockstore.so fio_engine.cpp json11.o
 	g++ $(CXXFLAGS) -shared -o libfio_blockstore.so fio_engine.cpp json11.o ./libblockstore.so -ltcmalloc_minimal -luring
 
 OSD_OBJS := osd.o osd_secondary.o osd_receive.o osd_send.o osd_peering.o osd_peering_pg.o osd_primary.o osd_rmw.o json11.o timerfd_interval.o
-osd_secondary.o: osd_secondary.cpp osd.h osd_ops.h
+osd_secondary.o: osd_secondary.cpp osd.h osd_ops.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
-osd_receive.o: osd_receive.cpp osd.h osd_ops.h
+osd_receive.o: osd_receive.cpp osd.h osd_ops.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
-osd_send.o: osd_send.cpp osd.h osd_ops.h
+osd_send.o: osd_send.cpp osd.h osd_ops.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
-osd_peering.o: osd_peering.cpp osd.h osd_ops.h osd_peering_pg.h
+osd_peering.o: osd_peering.cpp osd.h osd_ops.h osd_peering_pg.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 osd_peering_pg.o: osd_peering_pg.cpp object_id.h osd_peering_pg.h
 	g++ $(CXXFLAGS) -c -o $@ $<
@@ -40,14 +40,16 @@ osd_rmw.o: osd_rmw.cpp osd_rmw.h xor.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 osd_rmw_test: osd_rmw_test.cpp osd_rmw.cpp osd_rmw.h xor.h
 	g++ $(CXXFLAGS) -o $@ $<
-osd_primary.o: osd_primary.cpp osd.h osd_ops.h osd_peering_pg.h xor.h
+osd_primary.o: osd_primary.cpp osd.h osd_ops.h osd_peering_pg.h xor.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
-osd.o: osd.cpp osd.h osd_ops.h osd_peering_pg.h
+osd.o: osd.cpp osd.h osd_ops.h osd_peering_pg.h ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 osd: ./libblockstore.so osd_main.cpp osd.h osd_ops.h $(OSD_OBJS)
 	g++ $(CXXFLAGS) -o osd osd_main.cpp $(OSD_OBJS) ./libblockstore.so -ltcmalloc_minimal -luring
 stub_osd: stub_osd.cpp osd_ops.h rw_blocking.o
 	g++ $(CXXFLAGS) -o stub_osd stub_osd.cpp rw_blocking.o -ltcmalloc_minimal
+stub_bench: stub_bench.cpp osd_ops.h rw_blocking.o
+	g++ $(CXXFLAGS) -o stub_bench stub_bench.cpp rw_blocking.o -ltcmalloc_minimal
 rw_blocking.o: rw_blocking.cpp rw_blocking.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 osd_test: osd_test.cpp osd_ops.h rw_blocking.o

blockstore_impl.cpp

@@ -196,11 +196,6 @@ void blockstore_impl_t::loop()
         {
             flusher->loop();
         }
-        int ret = ringloop->submit();
-        if (ret < 0)
-        {
-            throw std::runtime_error(std::string("io_uring_submit: ") + strerror(-ret));
-        }
         if ((initial_ring_space - ringloop->space_left()) > 0)
         {
             live = true;

fio_sec_osd.cpp

@@ -261,15 +261,18 @@ static enum fio_q_status sec_queue(struct thread_data *td, struct io_u *io)
     bsd->op_n++;
     bsd->queue[n] = io;
 
-    if (write(bsd->connect_fd, op.buf, OSD_PACKET_SIZE) != OSD_PACKET_SIZE)
-    {
-        perror("write");
-        exit(1);
-    }
+    iovec iov[2] = { { .iov_base = op.buf, .iov_len = OSD_PACKET_SIZE } };
+    int iovcnt = 1, wtotal = OSD_PACKET_SIZE;
     if (io->ddir == DDIR_WRITE)
     {
-        // Send data
-        write_blocking(bsd->connect_fd, io->xfer_buf, io->xfer_buflen);
+        iov[1] = { .iov_base = io->xfer_buf, .iov_len = io->xfer_buflen };
+        wtotal += io->xfer_buflen;
+        iovcnt++;
+    }
+    if (writev_blocking(bsd->connect_fd, iov, iovcnt) != wtotal)
+    {
+        perror("writev");
+        exit(1);
     }
 
     if (io->error != 0)

osd.cpp

@@ -50,7 +50,7 @@ osd_t::osd_t(blockstore_config_t & config, blockstore_t *bs, ring_loop_t *ringlo
         }
         if (send_stat_count != 0)
         {
-            printf("avg latency to send subops with data: %ld us\n", send_stat_sum/send_stat_count);
+            printf("avg latency to send stabilize subop: %ld us\n", send_stat_sum/send_stat_count);
             send_stat_count = 0;
             send_stat_sum = 0;
         }
@@ -175,11 +175,26 @@ void osd_t::loop()
     handle_peers();
     read_requests();
     send_replies();
-    ringloop->submit();
 }
 
 void osd_t::handle_epoll_events()
 {
+    io_uring_sqe *sqe = ringloop->get_sqe();
+    if (!sqe)
+    {
+        throw std::runtime_error("can't get SQE, will fall out of sync with EPOLLET");
+    }
+    ring_data_t *data = ((ring_data_t*)sqe->user_data);
+    my_uring_prep_poll_add(sqe, epoll_fd, POLLIN);
+    data->callback = [this](ring_data_t *data)
+    {
+        if (data->res < 0)
+        {
+            throw std::runtime_error(std::string("epoll failed: ") + strerror(-data->res));
+        }
+        handle_epoll_events();
+    };
+    ringloop->submit();
     int nfds;
     epoll_event events[MAX_EPOLL_EVENTS];
 restart:
@@ -196,6 +211,7 @@ restart:
             {
                 char peer_str[256];
                 printf("osd: new client %d: connection from %s port %d\n", peer_fd, inet_ntop(AF_INET, &addr.sin_addr, peer_str, 256), ntohs(addr.sin_port));
+                fcntl(peer_fd, F_SETFL, fcntl(listen_fd, F_GETFL, 0) | O_NONBLOCK);
                 int one = 1;
                 setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
                 clients[peer_fd] = {
@@ -207,7 +223,7 @@ restart:
                 // Add FD to epoll
                 epoll_event ev;
                 ev.data.fd = peer_fd;
-                ev.events = EPOLLET | EPOLLRDHUP;
+                ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
                 if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, peer_fd, &ev) < 0)
                 {
                     throw std::runtime_error(std::string("epoll_ctl: ") + strerror(errno));
@@ -250,21 +266,6 @@ restart:
     {
         goto restart;
     }
-    io_uring_sqe *sqe = ringloop->get_sqe();
-    if (!sqe)
-    {
-        throw std::runtime_error("can't get SQE, will fall out of sync with EPOLLET");
-    }
-    ring_data_t *data = ((ring_data_t*)sqe->user_data);
-    my_uring_prep_poll_add(sqe, epoll_fd, POLLIN);
-    data->callback = [this](ring_data_t *data)
-    {
-        if (data->res < 0)
-        {
-            throw std::runtime_error(std::string("epoll failed: ") + strerror(-data->res));
-        }
-        handle_epoll_events();
-    };
 }
 
 void osd_t::cancel_osd_ops(osd_client_t & cl)
@@ -348,7 +349,7 @@ void osd_t::stop_client(int peer_fd)
 
 void osd_t::exec_op(osd_op_t *cur_op)
 {
-    gettimeofday(&cur_op->tv_begin, NULL);
+    clock_gettime(CLOCK_REALTIME, &cur_op->tv_begin);
     if (stopping)
     {
         // Throw operation away

osd.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <sys/types.h>
-#include <sys/time.h>
+#include <time.h>
 #include <sys/ioctl.h>
 #include <sys/stat.h>
 #include <fcntl.h>
@@ -93,8 +93,8 @@ struct osd_primary_op_data_t;
 
 struct osd_op_t
 {
-    timeval tv_begin;
-    timeval tv_send;
+    timespec tv_begin;
+    timespec tv_send;
     int op_type = OSD_OP_IN;
     int peer_fd;
     osd_any_op_t req;
@@ -131,8 +131,8 @@ struct osd_client_t
     int read_ready = 0;
     osd_op_t *read_op = NULL;
     int read_reply_id = 0;
-    iovec read_iov = { 0 };
-    msghdr read_msg = { 0 };
+    iovec read_iov;
+    msghdr read_msg;
     void *read_buf = NULL;
     int read_remaining = 0;
     int read_state = 0;
@@ -215,11 +215,11 @@ class osd_t
     // event loop, socket read/write
     void loop();
     void handle_epoll_events();
-    bool try_receive(osd_client_t & cl);
     void read_requests();
     void handle_read(ring_data_t *data, int peer_fd);
     void handle_op_hdr(osd_client_t *cl);
     void handle_reply_hdr(osd_client_t *cl);
+    bool try_send(osd_client_t & cl);
     void send_replies();
     void handle_send(ring_data_t *data, int peer_fd);
     void outbox_push(osd_client_t & cl, osd_op_t *op);

osd_peering.cpp

@@ -89,7 +89,7 @@ void osd_t::connect_peer(osd_num_t osd_num, const char *peer_host, int peer_port
     // Add FD to epoll (EPOLLOUT for tracking connect() result)
     epoll_event ev;
     ev.data.fd = peer_fd;
-    ev.events = EPOLLOUT | EPOLLRDHUP | EPOLLET;
+    ev.events = EPOLLOUT | EPOLLIN | EPOLLRDHUP | EPOLLET;
     if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, peer_fd, &ev) < 0)
     {
         throw std::runtime_error(std::string("epoll_ctl: ") + strerror(errno));
@@ -115,13 +115,12 @@ void osd_t::handle_connect_result(int peer_fd)
     }
     int one = 1;
     setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
-    fcntl(peer_fd, F_SETFL, fcntl(peer_fd, F_GETFL, 0) & ~O_NONBLOCK);
     // Disable EPOLLOUT on this fd
     cl.connect_callback = NULL;
     cl.peer_state = PEER_CONNECTED;
     epoll_event ev;
     ev.data.fd = peer_fd;
-    ev.events = EPOLLRDHUP | EPOLLET;
+    ev.events = EPOLLIN | EPOLLRDHUP | EPOLLET;
     if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, peer_fd, &ev) < 0)
     {
         throw std::runtime_error(std::string("epoll_ctl: ") + strerror(errno));

osd_receive.cpp

@@ -1,46 +1,40 @@
 #include "osd.h"
 
-bool osd_t::try_receive(osd_client_t & cl)
-{
-    int peer_fd = cl.peer_fd;
-    io_uring_sqe* sqe = ringloop->get_sqe();
-    if (!sqe)
-    {
-        return false;
-    }
-    ring_data_t* data = ((ring_data_t*)sqe->user_data);
-    if (!cl.read_buf)
-    {
-        // no reads in progress
-        // so this is either a new command or a reply to a previously sent command
-        if (!cl.read_op)
-        {
-            cl.read_op = new osd_op_t;
-            cl.read_op->peer_fd = peer_fd;
-        }
-        cl.read_op->op_type = OSD_OP_IN;
-        cl.read_buf = &cl.read_op->req.buf;
-        cl.read_remaining = OSD_PACKET_SIZE;
-        cl.read_state = CL_READ_OP;
-    }
-    cl.read_iov.iov_base = cl.read_buf;
-    cl.read_iov.iov_len = cl.read_remaining;
-    cl.read_msg.msg_iov = &cl.read_iov;
-    cl.read_msg.msg_iovlen = 1;
-    data->callback = [this, peer_fd](ring_data_t *data) { handle_read(data, peer_fd); };
-    my_uring_prep_recvmsg(sqe, peer_fd, &cl.read_msg, 0);
-    return true;
-}
-
 void osd_t::read_requests()
 {
-    for (auto & p: clients)
+    for (int i = 0; i < read_ready_clients.size(); i++)
     {
-        if (p.second.peer_state == PEER_CONNECTED && p.second.read_iov.iov_len == 0)
+        int peer_fd = read_ready_clients[i];
+        auto & cl = clients[peer_fd];
+        io_uring_sqe* sqe = ringloop->get_sqe();
+        if (!sqe)
         {
-            try_receive(p.second);
+            read_ready_clients.erase(read_ready_clients.begin(), read_ready_clients.begin() + i);
+            return;
         }
+        ring_data_t* data = ((ring_data_t*)sqe->user_data);
+        if (!cl.read_buf)
+        {
+            // no reads in progress
+            // so this is either a new command or a reply to a previously sent command
+            if (!cl.read_op)
+            {
+                cl.read_op = new osd_op_t;
+                cl.read_op->peer_fd = peer_fd;
+            }
+            cl.read_op->op_type = OSD_OP_IN;
+            cl.read_buf = &cl.read_op->req.buf;
+            cl.read_remaining = OSD_PACKET_SIZE;
+            cl.read_state = CL_READ_OP;
+        }
+        cl.read_iov.iov_base = cl.read_buf;
+        cl.read_iov.iov_len = cl.read_remaining;
+        cl.read_msg.msg_iov = &cl.read_iov;
+        cl.read_msg.msg_iovlen = 1;
+        data->callback = [this, peer_fd](ring_data_t *data) { handle_read(data, peer_fd); };
+        my_uring_prep_recvmsg(sqe, peer_fd, &cl.read_msg, 0);
     }
+    read_ready_clients.clear();
 }
 
 void osd_t::handle_read(ring_data_t *data, int peer_fd)
@@ -49,9 +43,11 @@ void osd_t::handle_read(ring_data_t *data, int peer_fd)
     if (cl_it != clients.end())
     {
         auto & cl = cl_it->second;
-        cl.read_iov.iov_len = 0;
         if (data->res == -EAGAIN)
         {
+            cl.read_ready--;
+            if (cl.read_ready > 0)
+                read_ready_clients.push_back(peer_fd);
             return;
         }
         else if (data->res < 0)
@@ -61,6 +57,7 @@ void osd_t::handle_read(ring_data_t *data, int peer_fd)
             stop_client(peer_fd);
             return;
         }
+        read_ready_clients.push_back(peer_fd);
         if (data->res > 0)
         {
             cl.read_remaining -= data->res;
@@ -95,12 +92,12 @@ void osd_t::handle_read(ring_data_t *data, int peer_fd)
                     cl.read_reply_id = 0;
                     cl.read_state = 0;
                     // Measure subop latency
-                    timeval tv_end;
-                    gettimeofday(&tv_end, NULL);
+                    timespec tv_end;
+                    clock_gettime(CLOCK_REALTIME, &tv_end);
                     subop_stat_count[request->req.hdr.opcode]++;
                     subop_stat_sum[request->req.hdr.opcode] += (
                         (tv_end.tv_sec - request->tv_begin.tv_sec)*1000000 +
-                        tv_end.tv_usec - request->tv_begin.tv_usec
+                        (tv_end.tv_nsec - request->tv_begin.tv_nsec)/1000
                     );
                     request->callback(request);
                 }
@@ -195,12 +192,12 @@ void osd_t::handle_reply_hdr(osd_client_t *cl)
         cl->read_state = 0;
         cl->sent_ops.erase(req_it);
         // Measure subop latency
-        timeval tv_end;
-        gettimeofday(&tv_end, NULL);
+        timespec tv_end;
+        clock_gettime(CLOCK_REALTIME, &tv_end);
         subop_stat_count[op->req.hdr.opcode]++;
         subop_stat_sum[op->req.hdr.opcode] += (
             (tv_end.tv_sec - op->tv_begin.tv_sec)*1000000 +
-            tv_end.tv_usec - op->tv_begin.tv_usec
+            (tv_end.tv_nsec - op->tv_begin.tv_nsec)/1000
         );
         op->callback(op);
     }

osd_send.cpp

@@ -5,15 +5,52 @@ void osd_t::outbox_push(osd_client_t & cl, osd_op_t *cur_op)
     assert(cur_op->peer_fd);
     if (cur_op->op_type == OSD_OP_OUT)
     {
-        gettimeofday(&cur_op->tv_begin, NULL);
-    }
-    if (cl.write_state == 0)
-    {
-        cl.write_state = CL_WRITE_READY;
-        write_ready_clients.push_back(cur_op->peer_fd);
+        clock_gettime(CLOCK_REALTIME, &cur_op->tv_begin);
     }
     cl.outbox.push_back(cur_op);
-    ringloop->wakeup();
+    if (cl.write_op || cl.outbox.size() > 1 || !try_send(cl))
+    {
+        if (cl.write_state == 0)
+        {
+            cl.write_state = CL_WRITE_READY;
+            write_ready_clients.push_back(cur_op->peer_fd);
+        }
+        ringloop->wakeup();
+    }
+}
+
+bool osd_t::try_send(osd_client_t & cl)
+{
+    int peer_fd = cl.peer_fd;
+    io_uring_sqe* sqe = ringloop->get_sqe();
+    if (!sqe)
+    {
+        return false;
+    }
+    ring_data_t* data = ((ring_data_t*)sqe->user_data);
+    if (!cl.write_op)
+    {
+        // pick next command
+        cl.write_op = cl.outbox.front();
+        cl.outbox.pop_front();
+        cl.write_state = CL_WRITE_REPLY;
+        clock_gettime(CLOCK_REALTIME, &cl.write_op->tv_send);
+        if (cl.write_op->op_type == OSD_OP_IN)
+        {
+            // Measure execution latency
+            timespec tv_end = cl.write_op->tv_send;
+            op_stat_count[cl.write_op->req.hdr.opcode]++;
+            op_stat_sum[cl.write_op->req.hdr.opcode] += (
+                (tv_end.tv_sec - cl.write_op->tv_begin.tv_sec)*1000000 +
+                (tv_end.tv_nsec - cl.write_op->tv_begin.tv_nsec)/1000
+            );
+        }
+    }
+    cl.write_msg.msg_iov = cl.write_op->send_list.get_iovec();
+    cl.write_msg.msg_iovlen = cl.write_op->send_list.get_size();
+    data->callback = [this, peer_fd](ring_data_t *data) { handle_send(data, peer_fd); };
+    my_uring_prep_sendmsg(sqe, peer_fd, &cl.write_msg, 0);
+    return true;
 }
 
 void osd_t::send_replies()
@@ -21,40 +58,11 @@ void osd_t::send_replies()
     for (int i = 0; i < write_ready_clients.size(); i++)
     {
         int peer_fd = write_ready_clients[i];
-        auto & cl = clients[peer_fd];
-        io_uring_sqe* sqe = ringloop->get_sqe();
-        if (!sqe)
+        if (!try_send(clients[peer_fd]))
         {
             write_ready_clients.erase(write_ready_clients.begin(), write_ready_clients.begin() + i);
             return;
         }
-        ring_data_t* data = ((ring_data_t*)sqe->user_data);
-        if (!cl.write_op)
-        {
-            // pick next command
-            cl.write_op = cl.outbox.front();
-            cl.outbox.pop_front();
-            cl.write_state = CL_WRITE_REPLY;
-            if (cl.write_op->op_type == OSD_OP_OUT)
-            {
-                gettimeofday(&cl.write_op->tv_send, NULL);
-            }
-            else
-            {
-                // Measure execution latency
-                timeval tv_end;
-                gettimeofday(&tv_end, NULL);
-                op_stat_count[cl.write_op->req.hdr.opcode]++;
-                op_stat_sum[cl.write_op->req.hdr.opcode] += (
-                    (tv_end.tv_sec - cl.write_op->tv_begin.tv_sec)*1000000 +
-                    tv_end.tv_usec - cl.write_op->tv_begin.tv_usec
-                );
-            }
-        }
-        cl.write_msg.msg_iov = cl.write_op->send_list.get_iovec();
-        cl.write_msg.msg_iovlen = cl.write_op->send_list.get_size();
-        data->callback = [this, peer_fd](ring_data_t *data) { handle_send(data, peer_fd); };
-        my_uring_prep_sendmsg(sqe, peer_fd, &cl.write_msg, 0);
     }
     write_ready_clients.clear();
 }
@@ -93,24 +101,22 @@ void osd_t::handle_send(ring_data_t *data, int peer_fd)
             if (cur_op->send_list.sent >= cur_op->send_list.count)
             {
                 // Done
+                if (cur_op->req.hdr.opcode == OSD_OP_SECONDARY_STABILIZE)
+                {
+                    timespec tv_end;
+                    clock_gettime(CLOCK_REALTIME, &tv_end);
+                    send_stat_count++;
+                    send_stat_sum += (
+                        (tv_end.tv_sec - cl.write_op->tv_send.tv_sec)*1000000 +
+                        (tv_end.tv_nsec - cl.write_op->tv_send.tv_nsec)/1000
+                    );
+                }
                 if (cur_op->op_type == OSD_OP_IN)
                 {
                     delete cur_op;
                 }
                 else
                 {
-                    // Measure subops with data
-                    if (cur_op->req.hdr.opcode == OSD_OP_SECONDARY_STABILIZE ||
-                        cur_op->req.hdr.opcode == OSD_OP_SECONDARY_WRITE)
-                    {
-                        timeval tv_end;
-                        gettimeofday(&tv_end, NULL);
-                        send_stat_count++;
-                        send_stat_sum += (
-                            (tv_end.tv_sec - cl.write_op->tv_send.tv_sec)*1000000 +
-                            tv_end.tv_usec - cl.write_op->tv_send.tv_usec
-                        );
-                    }
                     cl.sent_ops[cl.write_op->req.hdr.id] = cl.write_op;
                 }
                 cl.write_op = NULL;

ringloop.h

@@ -146,8 +146,7 @@ public:
     }
     inline int wait()
     {
-        struct io_uring_cqe *cqe;
-        return io_uring_wait_cqe(&ring, &cqe);
+        return io_uring_submit_and_wait(&ring, 1);
     }
     inline unsigned space_left()
     {

rw_blocking.cpp

@@ -50,3 +50,37 @@ int write_blocking(int fd, void *write_buf, size_t remaining)
     }
     return done;
 }
+
+int writev_blocking(int fd, iovec *iov, int iovcnt)
+{
+    int v = 0;
+    int done = 0;
+    while (v < iovcnt)
+    {
+        ssize_t r = writev(fd, iov, iovcnt);
+        if (r < 0)
+        {
+            if (errno != EAGAIN && errno != EPIPE)
+            {
+                perror("writev");
+                exit(1);
+            }
+            continue;
+        }
+        while (v < iovcnt)
+        {
+            if (iov[v].iov_len > r)
+            {
+                iov[v].iov_len -= r;
+                iov[v].iov_base += r;
+                break;
+            }
+            else
+            {
+                v++;
+            }
+        }
+        done += r;
+    }
+    return done;
+}

rw_blocking.h

@@ -1,6 +1,8 @@
 #pragma once
 
 #include <unistd.h>
+#include <sys/uio.h>
 
 int read_blocking(int fd, void *read_buf, size_t remaining);
 int write_blocking(int fd, void *write_buf, size_t remaining);
+int writev_blocking(int fd, iovec *iov, int iovcnt);

stub_bench.cpp

@@ -0,0 +1,148 @@
+/**
+ * Stub benchmarker
+ */
+
+#include <sys/types.h>
+#include <time.h>
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <arpa/inet.h>
+#include <string.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <signal.h>
+
+#include <stdexcept>
+
+#include "rw_blocking.h"
+#include "osd_ops.h"
+
+int connect_stub(const char *server_address, int server_port);
+
+void run_bench(int peer_fd);
+
+static uint64_t write_sum = 0, write_count = 0;
+static uint64_t sync_sum = 0, sync_count = 0;
+
+void handle_sigint(int sig)
+{
+    printf("4k randwrite: %lu us avg\n", write_sum/write_count);
+    printf("sync: %lu us avg\n", sync_sum/sync_count);
+    exit(0);
+}
+
+int main(int narg, char *args[])
+{
+    signal(SIGINT, handle_sigint);
+    int peer_fd = connect_stub("127.0.0.1", 11203);
+    run_bench(peer_fd);
+    close(peer_fd);
+    return 0;
+}
+
+int connect_stub(const char *server_address, int server_port)
+{
+    struct sockaddr_in addr;
+    int r;
+    if ((r = inet_pton(AF_INET, server_address, &addr.sin_addr)) != 1)
+    {
+        fprintf(stderr, "server address: %s%s\n", server_address, r == 0 ? " is not valid" : ": no ipv4 support");
+        return -1;
+    }
+    addr.sin_family = AF_INET;
+    addr.sin_port = htons(server_port);
+    int connect_fd = socket(AF_INET, SOCK_STREAM, 0);
+    if (connect_fd < 0)
+    {
+        perror("socket");
+        return -1;
+    }
+    if (connect(connect_fd, (sockaddr*)&addr, sizeof(addr)) < 0)
+    {
+        perror("connect");
+        return -1;
+    }
+    int one = 1;
+    setsockopt(connect_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
+    return connect_fd;
+}
+
+bool check_reply(int r, osd_any_op_t & op, osd_any_reply_t & reply, int expected)
+{
+    if (r != OSD_PACKET_SIZE)
+    {
+        printf("read failed\n");
+        return false;
+    }
+    if (reply.hdr.magic != SECONDARY_OSD_REPLY_MAGIC ||
+        reply.hdr.id != op.hdr.id || reply.hdr.opcode != op.hdr.opcode)
+    {
+        printf("bad reply: magic, id or opcode does not match request\n");
+        return false;
+    }
+    if (reply.hdr.retval != expected)
+    {
+        printf("operation failed, retval=%ld (%s)\n", reply.hdr.retval, strerror(-reply.hdr.retval));
+        return false;
+    }
+    return true;
+}
+
+void run_bench(int peer_fd)
+{
+    osd_any_op_t op;
+    osd_any_reply_t reply;
+    void *buf = NULL;
+    int r;
+    timespec tv_begin, tv_end;
+    clock_gettime(CLOCK_REALTIME, &tv_begin);
+    while (1)
+    {
+        // write
+        op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
+        op.hdr.id = 1;
+        op.hdr.opcode = OSD_OP_SECONDARY_WRITE;
+        op.sec_rw.oid.inode = 3;
+        op.sec_rw.oid.stripe = (rand() << 17) % (1 << 29); // 512 MB
+        op.sec_rw.version = 0;
+        op.sec_rw.len = 4096;
+        op.sec_rw.offset = (rand() * op.sec_rw.len) % (1 << 17);
+        buf = malloc(op.sec_rw.len);
+        memset(buf, rand() % 255, op.sec_rw.len);
+        r = write_blocking(peer_fd, op.buf, OSD_PACKET_SIZE) == OSD_PACKET_SIZE;
+        if (r)
+            r = write_blocking(peer_fd, buf, op.sec_rw.len) == op.sec_rw.len;
+        free(buf);
+        if (!r)
+            break;
+        r = read_blocking(peer_fd, reply.buf, OSD_PACKET_SIZE);
+        if (!check_reply(r, op, reply, op.sec_rw.len))
+            break;
+        clock_gettime(CLOCK_REALTIME, &tv_end);
+        write_count++;
+        write_sum += (
+            (tv_end.tv_sec - tv_begin.tv_sec)*1000000 +
+            tv_end.tv_nsec/1000 - tv_begin.tv_nsec/1000
+        );
+        // sync/stab
+        op.hdr.magic = SECONDARY_OSD_OP_MAGIC;
+        op.hdr.id = 1;
+        op.hdr.opcode = OSD_OP_TEST_SYNC_STAB_ALL;
+        r = write_blocking(peer_fd, op.buf, OSD_PACKET_SIZE) == OSD_PACKET_SIZE;
+        if (!r)
+            break;
+        r = read_blocking(peer_fd, reply.buf, OSD_PACKET_SIZE);
+        if (!check_reply(r, op, reply, 0))
+            break;
+        clock_gettime(CLOCK_REALTIME, &tv_begin);
+        sync_count++;
+        sync_sum += (
+            (tv_begin.tv_sec - tv_end.tv_sec)*1000000 +
+            tv_begin.tv_nsec/1000 - tv_end.tv_nsec/1000
+        );
+    }
+}