Test SQ poll threads. Unstable and in fact slower :(

blockstore_flush.cpp

@@ -287,8 +287,9 @@ resume_1:
             data->iov = (struct iovec){ it->buf, (size_t)it->len };
             data->callback = simple_callback_w;
             my_uring_prep_writev(
-                sqe, bs->data_fd, &data->iov, 1, bs->data_offset + clean_loc + it->offset
+                sqe, bs->data_fd_index, &data->iov, 1, bs->data_offset + clean_loc + it->offset
             );
+            sqe->flags |= IOSQE_FIXED_FILE;
             wait_count++;
         }
         // Sync data before writing metadata
@@ -320,8 +321,9 @@ resume_1:
             data->iov = (struct iovec){ meta_old.buf, bs->meta_block_size };
             data->callback = simple_callback_w;
             my_uring_prep_writev(
-                sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + meta_old.sector
+                sqe, bs->meta_fd_index, &data->iov, 1, bs->meta_offset + meta_old.sector
             );
+            sqe->flags |= IOSQE_FIXED_FILE;
             wait_count++;
         }
         if (has_delete)
@@ -342,8 +344,9 @@ resume_1:
         data->iov = (struct iovec){ meta_new.buf, bs->meta_block_size };
         data->callback = simple_callback_w;
         my_uring_prep_writev(
-            sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + meta_new.sector
+            sqe, bs->meta_fd_index, &data->iov, 1, bs->meta_offset + meta_new.sector
         );
+        sqe->flags |= IOSQE_FIXED_FILE;
         wait_count++;
     resume_7:
         if (wait_count > 0)
@@ -405,7 +408,8 @@ resume_1:
                 ((journal_entry_start*)flusher->journal_superblock)->crc32 = je_crc32((journal_entry*)flusher->journal_superblock);
                 data->iov = (struct iovec){ flusher->journal_superblock, bs->journal_block_size };
                 data->callback = simple_callback_w;
-                my_uring_prep_writev(sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset);
+                my_uring_prep_writev(sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset);
+                sqe->flags |= IOSQE_FIXED_FILE;
                 wait_count++;
             resume_13:
                 if (wait_count > 0)
@@ -485,8 +489,9 @@ bool journal_flusher_co::scan_dirty(int wait_base)
                             data->iov = (struct iovec){ v.back().buf, (size_t)submit_len };
                             data->callback = simple_callback_r;
                             my_uring_prep_readv(
-                                sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset + submit_offset
+                                sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset + submit_offset
                             );
+                            sqe->flags |= IOSQE_FIXED_FILE;
                             wait_count++;
                         }
                     }
@@ -568,8 +573,9 @@ bool journal_flusher_co::modify_meta_read(uint64_t meta_loc, flusher_meta_write_
         data->callback = simple_callback_r;
         wr.submitted = true;
         my_uring_prep_readv(
-            sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + wr.sector
+            sqe, bs->meta_fd_index, &data->iov, 1, bs->meta_offset + wr.sector
         );
+        sqe->flags |= IOSQE_FIXED_FILE;
         wait_count++;
     }
     else
@@ -638,7 +644,8 @@ bool journal_flusher_co::fsync_batch(bool fsync_meta, int wait_base)
             await_sqe(0);
             data->iov = { 0 };
             data->callback = simple_callback_w;
-            my_uring_prep_fsync(sqe, fsync_meta ? bs->meta_fd : bs->data_fd, IORING_FSYNC_DATASYNC);
+            my_uring_prep_fsync(sqe, fsync_meta ? bs->meta_fd_index : bs->data_fd_index, IORING_FSYNC_DATASYNC);
+            sqe->flags |= IOSQE_FIXED_FILE;
             cur_sync->state = 1;
             wait_count++;
         resume_1:

blockstore_impl.h

@@ -216,6 +216,7 @@ class blockstore_impl_t
     int data_fd;
     uint64_t meta_size, meta_area, meta_len;
     uint64_t data_size, data_len;
+    int meta_fd_index, data_fd_index, journal_fd_index;
 
     void *metadata_buffer = NULL;
 

blockstore_init.cpp

@@ -55,7 +55,8 @@ int blockstore_init_meta::loop()
                 bs->meta_len - metadata_read > bs->metadata_buf_size ? bs->metadata_buf_size : bs->meta_len - metadata_read,
             };
             data->callback = [this](ring_data_t *data) { handle_event(data); };
-            my_uring_prep_readv(sqe, bs->meta_fd, &data->iov, 1, bs->meta_offset + metadata_read);
+            my_uring_prep_readv(sqe, bs->meta_fd_index, &data->iov, 1, bs->meta_offset + metadata_read);
+            sqe->flags |= IOSQE_FIXED_FILE;
             bs->ringloop->submit();
             submitted = (prev == 1 ? 2 : 1);
             prev = submitted;
@@ -216,7 +217,8 @@ int blockstore_init_journal::loop()
     data = ((ring_data_t*)sqe->user_data);
     data->iov = { submitted_buf, bs->journal.block_size };
     data->callback = simple_callback;
-    my_uring_prep_readv(sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset);
+    my_uring_prep_readv(sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset);
+    sqe->flags |= IOSQE_FIXED_FILE;
     bs->ringloop->submit();
     wait_count = 1;
 resume_1:
@@ -254,7 +256,8 @@ resume_1:
             GET_SQE();
             data->iov = (struct iovec){ submitted_buf, 2*bs->journal.block_size };
             data->callback = simple_callback;
-            my_uring_prep_writev(sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset);
+            my_uring_prep_writev(sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset);
+            sqe->flags |= IOSQE_FIXED_FILE;
             wait_count++;
             bs->ringloop->submit();
         resume_6:
@@ -266,7 +269,8 @@ resume_1:
             if (!bs->disable_journal_fsync)
             {
                 GET_SQE();
-                my_uring_prep_fsync(sqe, bs->journal.fd, IORING_FSYNC_DATASYNC);
+                my_uring_prep_fsync(sqe, bs->journal_fd_index, IORING_FSYNC_DATASYNC);
+                sqe->flags |= IOSQE_FIXED_FILE;
                 data->iov = { 0 };
                 data->callback = simple_callback;
                 wait_count++;
@@ -325,7 +329,8 @@ resume_1:
                     end - journal_pos < JOURNAL_BUFFER_SIZE ? end - journal_pos : JOURNAL_BUFFER_SIZE,
                 };
                 data->callback = [this](ring_data_t *data1) { handle_event(data1); };
-                my_uring_prep_readv(sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset + journal_pos);
+                my_uring_prep_readv(sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset + journal_pos);
+                sqe->flags |= IOSQE_FIXED_FILE;
                 bs->ringloop->submit();
             }
             while (done.size() > 0)
@@ -340,7 +345,8 @@ resume_1:
                         GET_SQE();
                         data->iov = { init_write_buf, bs->journal.block_size };
                         data->callback = simple_callback;
-                        my_uring_prep_writev(sqe, bs->journal.fd, &data->iov, 1, bs->journal.offset + init_write_sector);
+                        my_uring_prep_writev(sqe, bs->journal_fd_index, &data->iov, 1, bs->journal.offset + init_write_sector);
+                        sqe->flags |= IOSQE_FIXED_FILE;
                         wait_count++;
                         bs->ringloop->submit();
                     resume_7:
@@ -354,7 +360,8 @@ resume_1:
                             GET_SQE();
                             data->iov = { 0 };
                             data->callback = simple_callback;
-                            my_uring_prep_fsync(sqe, bs->journal.fd, IORING_FSYNC_DATASYNC);
+                            my_uring_prep_fsync(sqe, bs->journal_fd_index, IORING_FSYNC_DATASYNC);
+                            sqe->flags |= IOSQE_FIXED_FILE;
                             wait_count++;
                             bs->ringloop->submit();
                         }

blockstore_journal.cpp

@@ -126,8 +126,9 @@ void prepare_journal_sector_write(journal_t & journal, int cur_sector, io_uring_
     };
     data->callback = cb;
     my_uring_prep_writev(
-        sqe, journal.fd, &data->iov, 1, journal.offset + journal.sector_info[cur_sector].offset
+        sqe, journal.fd_index, &data->iov, 1, journal.offset + journal.sector_info[cur_sector].offset
     );
+    sqe->flags |= IOSQE_FIXED_FILE;
 }
 
 journal_t::~journal_t()

blockstore_journal.h

@@ -130,7 +130,7 @@ struct journal_sector_info_t
 
 struct journal_t
 {
-    int fd;
+    int fd, fd_index;
     uint64_t device_size;
     bool inmemory = false;
     void *buffer = NULL;

blockstore_open.cpp

@@ -140,6 +140,10 @@ void blockstore_impl_t::parse_config(blockstore_config_t & config)
 
 void blockstore_impl_t::calc_lengths()
 {
+    // register fds
+    data_fd_index = ringloop->register_fd(data_fd);
+    meta_fd_index = meta_fd == data_fd ? data_fd_index : ringloop->register_fd(meta_fd);
+    journal.fd_index = journal_fd_index = journal.fd == meta_fd ? meta_fd_index : ringloop->register_fd(journal.fd);
     // data
     data_len = data_size - data_offset;
     if (data_fd == meta_fd && data_offset < meta_offset)

blockstore_read.cpp

@@ -31,10 +31,11 @@ int blockstore_impl_t::fulfill_read_push(blockstore_op_t *op, void *buf, uint64_
     PRIV(op)->pending_ops++;
     my_uring_prep_readv(
         sqe,
-        IS_JOURNAL(item_state) ? journal.fd : data_fd,
+        IS_JOURNAL(item_state) ? journal_fd_index : data_fd_index,
         &data->iov, 1,
         (IS_JOURNAL(item_state) ? journal.offset : data_offset) + offset
     );
+    sqe->flags |= IOSQE_FIXED_FILE;
     data->callback = [this, op](ring_data_t *data) { handle_read_event(data, op); };
     return 1;
 }

blockstore_sync.cpp

@@ -78,7 +78,8 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
         if (!disable_data_fsync)
         {
             BS_SUBMIT_GET_SQE(sqe, data);
-            my_uring_prep_fsync(sqe, data_fd, IORING_FSYNC_DATASYNC);
+            my_uring_prep_fsync(sqe, data_fd_index, IORING_FSYNC_DATASYNC);
+            sqe->flags |= IOSQE_FIXED_FILE;
             data->iov = { 0 };
             data->callback = cb;
             PRIV(op)->min_used_journal_sector = PRIV(op)->max_used_journal_sector = 0;
@@ -161,7 +162,8 @@ int blockstore_impl_t::continue_sync(blockstore_op_t *op)
         if (!disable_journal_fsync)
         {
             BS_SUBMIT_GET_SQE(sqe, data);
-            my_uring_prep_fsync(sqe, journal.fd, IORING_FSYNC_DATASYNC);
+            my_uring_prep_fsync(sqe, journal_fd_index, IORING_FSYNC_DATASYNC);
+            sqe->flags |= IOSQE_FIXED_FILE;
             data->iov = { 0 };
             data->callback = cb;
             PRIV(op)->pending_ops = 1;

blockstore_write.cpp

@@ -122,8 +122,9 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
         data->iov.iov_len = op->len + stripe_offset + stripe_end; // to check it in the callback
         data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
         my_uring_prep_writev(
-            sqe, data_fd, PRIV(op)->iov_zerofill, vcnt, data_offset + (loc << block_order) + op->offset - stripe_offset
+            sqe, data_fd_index, PRIV(op)->iov_zerofill, vcnt, data_offset + (loc << block_order) + op->offset - stripe_offset
         );
+        sqe->flags |= IOSQE_FIXED_FILE;
         PRIV(op)->pending_ops = 1;
         PRIV(op)->min_used_journal_sector = PRIV(op)->max_used_journal_sector = 0;
         // Remember big write as unsynced
@@ -198,8 +199,9 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
             data2->iov = (struct iovec){ op->buf, op->len };
             data2->callback = cb;
             my_uring_prep_writev(
-                sqe2, journal.fd, &data2->iov, 1, journal.offset + journal.next_free
+                sqe2, journal_fd_index, &data2->iov, 1, journal.offset + journal.next_free
             );
+            sqe2->flags |= IOSQE_FIXED_FILE;
             PRIV(op)->pending_ops++;
         }
         else

osd.cpp

@@ -110,6 +110,7 @@ osd_t::osd_t(blockstore_config_t & config, blockstore_t *bs, ring_loop_t *ringlo
         close(listen_fd);
         throw std::runtime_error(std::string("epoll_create: ") + strerror(errno));
     }
+    epoll_fd_index = ringloop->register_fd(epoll_fd);
 
     epoll_event ev;
     ev.data.fd = listen_fd;
@@ -186,16 +187,19 @@ void osd_t::handle_epoll_events()
         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->allow_cancel = true;
+    my_uring_prep_poll_add(sqe, epoll_fd_index, POLLIN);
+    sqe->flags |= IOSQE_FIXED_FILE;
     data->callback = [this](ring_data_t *data)
     {
-        if (data->res < 0)
+        if (data->res < 0 && data->res != -ECANCELED)
         {
             throw std::runtime_error(std::string("epoll failed: ") + strerror(-data->res));
         }
         handle_epoll_events();
     };
     ringloop->submit();
+    // FIXME With SQ thread we have no guarantee that epoll request will be submitted right here...
     int nfds;
     epoll_event events[MAX_EPOLL_EVENTS];
 restart:
@@ -219,12 +223,13 @@ restart:
                     .peer_addr = addr,
                     .peer_port = ntohs(addr.sin_port),
                     .peer_fd = peer_fd,
+                    .peer_fd_index = ringloop->register_fd(peer_fd),
                     .peer_state = PEER_CONNECTED,
                 };
                 // Add FD to epoll
                 epoll_event ev;
                 ev.data.fd = peer_fd;
-                ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
+                ev.events = 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));
@@ -263,7 +268,7 @@ restart:
             }
         }
     }
-    if (nfds == MAX_EPOLL_EVENTS)
+    if (nfds > 0)
     {
         goto restart;
     }

osd.h

@@ -122,7 +122,7 @@ struct osd_client_t
 {
     sockaddr_in peer_addr;
     int peer_port;
-    int peer_fd;
+    int peer_fd, peer_fd_index;
     int peer_state;
     std::function<void(osd_num_t, int)> connect_callback;
     osd_num_t osd_num = 0;
@@ -196,7 +196,7 @@ class osd_t
     timerfd_interval *tick_tfd;
 
     int wait_state = 0;
-    int epoll_fd = 0;
+    int epoll_fd = 0, epoll_fd_index = -1;
     int listen_fd = 0;
     ring_consumer_t consumer;
 

osd_peering.cpp

@@ -116,6 +116,7 @@ void osd_t::handle_connect_result(int peer_fd)
     int one = 1;
     setsockopt(peer_fd, SOL_TCP, TCP_NODELAY, &one, sizeof(one));
     // Disable EPOLLOUT on this fd
+    cl.peer_fd_index = ringloop->register_fd(peer_fd);
     cl.connect_callback = NULL;
     cl.peer_state = PEER_CONNECTED;
     epoll_event ev;

osd_receive.cpp

@@ -32,36 +32,38 @@ void osd_t::read_requests()
         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);
+        my_uring_prep_recvmsg(sqe, cl.peer_fd_index, &cl.read_msg, 0);
+        sqe->flags |= IOSQE_FIXED_FILE;
     }
     read_ready_clients.clear();
 }
 
 void osd_t::handle_read(ring_data_t *data, int peer_fd)
 {
+    int res = data->res;
     auto cl_it = clients.find(peer_fd);
     if (cl_it != clients.end())
     {
         auto & cl = cl_it->second;
-        if (data->res == -EAGAIN)
+        if (res == -EAGAIN)
         {
             cl.read_ready--;
             if (cl.read_ready > 0)
                 read_ready_clients.push_back(peer_fd);
             return;
         }
-        else if (data->res < 0)
+        else if (res < 0)
         {
             // this is a client socket, so don't panic. just disconnect it
-            printf("Client %d socket read error: %d (%s). Disconnecting client\n", peer_fd, -data->res, strerror(-data->res));
+            printf("Client %d socket read error: %d (%s). Disconnecting client\n", peer_fd, -res, strerror(-res));
             stop_client(peer_fd);
             return;
         }
         read_ready_clients.push_back(peer_fd);
-        if (data->res > 0)
+        if (res > 0)
         {
-            cl.read_remaining -= data->res;
-            cl.read_buf += data->res;
+            cl.read_remaining -= res;
+            cl.read_buf += res;
             if (cl.read_remaining <= 0)
             {
                 cl.read_buf = NULL;

osd_send.cpp

@@ -49,7 +49,8 @@ bool osd_t::try_send(osd_client_t & cl)
     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);
+    my_uring_prep_sendmsg(sqe, cl.peer_fd_index, &cl.write_msg, 0);
+    sqe->flags |= IOSQE_FIXED_FILE;
     return true;
 }
 
@@ -69,32 +70,33 @@ void osd_t::send_replies()
 
 void osd_t::handle_send(ring_data_t *data, int peer_fd)
 {
+    int res = data->res;
     auto cl_it = clients.find(peer_fd);
     if (cl_it != clients.end())
     {
         auto & cl = cl_it->second;
-        if (data->res < 0 && data->res != -EAGAIN)
+        if (res < 0 && res != -EAGAIN)
         {
             // this is a client socket, so don't panic. just disconnect it
-            printf("Client %d socket write error: %d (%s). Disconnecting client\n", peer_fd, -data->res, strerror(-data->res));
+            printf("Client %d socket write error: %d (%s). Disconnecting client\n", peer_fd, -res, strerror(-res));
             stop_client(peer_fd);
             return;
         }
-        if (data->res >= 0)
+        if (res >= 0)
         {
             osd_op_t *cur_op = cl.write_op;
-            while (data->res > 0 && cur_op->send_list.sent < cur_op->send_list.count)
+            while (res > 0 && cur_op->send_list.sent < cur_op->send_list.count)
             {
                 iovec & iov = cur_op->send_list.buf[cur_op->send_list.sent];
-                if (iov.iov_len <= data->res)
+                if (iov.iov_len <= res)
                 {
-                    data->res -= iov.iov_len;
+                    res -= iov.iov_len;
                     cur_op->send_list.sent++;
                 }
                 else
                 {
-                    iov.iov_len -= data->res;
-                    iov.iov_base += data->res;
+                    iov.iov_len -= res;
+                    iov.iov_base += res;
                     break;
                 }
             }

ringloop.cpp

@@ -1,14 +1,19 @@
+#include <set>
+
 #include "ringloop.h"
 
 ring_loop_t::ring_loop_t(int qd)
 {
-    int ret = io_uring_queue_init(qd, &ring, 0);
+    io_uring_params params = { 0 };
+    params.flags = IORING_SETUP_SQPOLL;
+    params.sq_thread_idle = 10;
+    int ret = io_uring_queue_init_params(qd, &ring, &params);
     if (ret < 0)
     {
         throw std::runtime_error(std::string("io_uring_queue_init: ") + strerror(-ret));
     }
-    free_ring_data_ptr = *ring.cq.kring_entries;
-    ring_datas = (struct ring_data_t*)malloc(sizeof(ring_data_t) * free_ring_data_ptr);
+    ring_data_total = free_ring_data_ptr = *ring.cq.kring_entries;
+    ring_datas = (ring_data_t*)malloc(sizeof(ring_data_t) * free_ring_data_ptr);
     free_ring_data = (int*)malloc(sizeof(int) * free_ring_data_ptr);
     if (!ring_datas || !free_ring_data)
     {
@@ -16,6 +21,7 @@ ring_loop_t::ring_loop_t(int qd)
     }
     for (int i = 0; i < free_ring_data_ptr; i++)
     {
+        ring_datas[i] = { 0 };
         free_ring_data[i] = i;
     }
 }
@@ -27,6 +33,105 @@ ring_loop_t::~ring_loop_t()
     io_uring_queue_exit(&ring);
 }
 
+void ring_loop_t::drain_events(void *completions_ptr)
+{
+    std::set<ring_data_t*> & completions = *((std::set<ring_data_t*> *)completions_ptr);
+    if (free_ring_data_ptr < ring_data_total)
+    {
+        // Try to cancel requests that are allowed to be canceled by the caller (epoll, timerfd and similar)
+        for (int i = 0; i < ring_data_total; i++)
+        {
+            if (ring_datas[i].allow_cancel)
+            {
+                // allow_cancel may only be true while the operation is inflight
+                io_uring_sqe *sqe = get_sqe();
+                if (!sqe)
+                {
+                    throw std::runtime_error("can't get SQE to cancel operation");
+                }
+                ring_data_t *data = (ring_data_t*)sqe->user_data;
+                data->callback = NULL;
+                ring_datas[i].res = -ECANCELED;
+                my_uring_prep_cancel(sqe, &ring_datas[i], 0);
+                // It seems (FIXME) cancel operations don't always get completions
+                completions.insert(data);
+            }
+        }
+        if (completions.size() > 0)
+        {
+            submit();
+        }
+    }
+    int inflight = ring_data_total - free_ring_data_ptr;
+    while (completions.size() < inflight)
+    {
+        io_uring_cqe *cqe;
+        while (!io_uring_peek_cqe(&ring, &cqe))
+        {
+            ring_data_t *d = (ring_data_t*)cqe->user_data;
+            d->res = cqe->res;
+            d->allow_cancel = false;
+            completions.insert(d);
+            io_uring_cqe_seen(&ring, cqe);
+        }
+        if (completions.size() < inflight)
+        {
+            wait();
+        }
+    }
+}
+
+void ring_loop_t::run_completions(void *completions_ptr)
+{
+    std::set<ring_data_t*> & completions = *((std::set<ring_data_t*> *)completions_ptr);
+    // Call event callbacks
+    for (ring_data_t *d: completions)
+    {
+        free_ring_data[free_ring_data_ptr++] = d - ring_datas;
+        if (d->callback)
+            d->callback(d);
+    }
+}
+
+int ring_loop_t::register_fd(int fd)
+{
+    std::set<ring_data_t*> completions;
+    drain_events((void*)&completions);
+    // Modify registered files
+    int idx = reg_fds.size();
+    reg_fds.push_back(fd);
+    if (registered)
+    {
+        io_uring_unregister_files(&ring);
+    }
+    int ret = io_uring_register_files(&ring, reg_fds.data(), reg_fds.size());
+    if (ret != 0)
+    {
+        throw std::runtime_error(std::string("io_uring_register_files_update: ") + strerror(-ret));
+    }
+    registered = 1;
+    run_completions((void*)&completions);
+    return idx;
+}
+
+void ring_loop_t::unregister_fd(int fd_index)
+{
+    std::set<ring_data_t*> completions;
+    drain_events((void*)&completions);
+    // Modify registered files
+    reg_fds.erase(reg_fds.begin()+fd_index, reg_fds.begin()+fd_index+1);
+    if (registered)
+    {
+        io_uring_unregister_files(&ring);
+    }
+    int ret = io_uring_register_files(&ring, reg_fds.data(), reg_fds.size());
+    if (ret != 0)
+    {
+        throw std::runtime_error(std::string("io_uring_register_files_update: ") + strerror(-ret));
+    }
+    run_completions((void*)&completions);
+}
+
 int ring_loop_t::register_consumer(ring_consumer_t & consumer)
 {
     consumer.number = consumers.size();
@@ -50,17 +155,16 @@ void ring_loop_t::unregister_consumer(ring_consumer_t & consumer)
 
 void ring_loop_t::loop()
 {
-    struct io_uring_cqe *cqe;
+    io_uring_cqe *cqe;
     while (!io_uring_peek_cqe(&ring, &cqe))
     {
-        struct ring_data_t *d = (struct ring_data_t*)cqe->user_data;
+        ring_data_t *d = (ring_data_t*)cqe->user_data;
+        d->res = cqe->res;
+        d->allow_cancel = false;
+        io_uring_cqe_seen(&ring, cqe);
+        free_ring_data[free_ring_data_ptr++] = d - ring_datas;
         if (d->callback)
-        {
-            d->res = cqe->res;
             d->callback(d);
-        }
-        free_ring_data[free_ring_data_ptr++] = d - ring_datas;
-        io_uring_cqe_seen(&ring, cqe);
     }
     do
     {

ringloop.h

@@ -107,6 +107,7 @@ static inline void my_uring_prep_cancel(struct io_uring_sqe *sqe, void *user_dat
 struct ring_data_t
 {
     struct iovec iov; // for single-entry read/write operations
+    bool allow_cancel;
     int res;
     std::function<void(ring_data_t*)> callback;
 };
@@ -122,22 +123,35 @@ class ring_loop_t
     std::vector<ring_consumer_t> consumers;
     struct ring_data_t *ring_datas;
     int *free_ring_data;
-    unsigned free_ring_data_ptr;
+    unsigned free_ring_data_ptr, ring_data_total;
     bool loop_again;
     struct io_uring ring;
+    int registered = 0;
+    std::vector<int> reg_fds;
+    void drain_events(void *completions_ptr);
+    void run_completions(void *completions_ptr);
+
 public:
     ring_loop_t(int qd);
     ~ring_loop_t();
     int register_consumer(ring_consumer_t & consumer);
     void unregister_consumer(ring_consumer_t & consumer);
 
+    int register_fd(int fd);
+    void unregister_fd(int fd_index);
+
     inline struct io_uring_sqe* get_sqe()
     {
         if (free_ring_data_ptr == 0)
+        {
             return NULL;
+        }
         struct io_uring_sqe* sqe = io_uring_get_sqe(&ring);
         if (sqe)
-            io_uring_sqe_set_data(sqe, ring_datas + free_ring_data[--free_ring_data_ptr]);
+        {
+            ring_data_t *data = ring_datas + free_ring_data[--free_ring_data_ptr];
+            io_uring_sqe_set_data(sqe, data);
+        }
         return sqe;
     }
     inline int submit()

timerfd_interval.cpp

@@ -21,6 +21,7 @@ timerfd_interval::timerfd_interval(ring_loop_t *ringloop, int seconds, std::func
     }
     consumer.loop = [this]() { loop(); };
     ringloop->register_consumer(consumer);
+    timerfd_index = ringloop->register_fd(timerfd);
     this->ringloop = ringloop;
     this->callback = cb;
 }
@@ -44,10 +45,12 @@ void timerfd_interval::loop()
         return;
     }
     struct ring_data_t *data = ((ring_data_t*)sqe->user_data);
-    my_uring_prep_poll_add(sqe, timerfd, POLLIN);
+    my_uring_prep_poll_add(sqe, timerfd_index, POLLIN);
+    sqe->flags |= IOSQE_FIXED_FILE;
+    data->allow_cancel = true;
     data->callback = [&](ring_data_t *data)
     {
-        if (data->res < 0)
+        if (data->res < 0 && data->res != -ECANCELED)
         {
             throw std::runtime_error(std::string("waiting for timer failed: ") + strerror(-data->res));
         }

timerfd_interval.h

@@ -5,7 +5,7 @@
 class timerfd_interval
 {
     int wait_state;
-    int timerfd;
+    int timerfd, timerfd_index;
     int status;
     ring_loop_t *ringloop;
     ring_consumer_t consumer;