5 files changed, 338 insertions, 3 deletions

diff --git a/tests/accept_filter_probe.py b/tests/accept_filter_probe.py
new file mode 100644
index 0000000..22f585a
--- /dev/null
+++ b/tests/accept_filter_probe.py
@@ -0,0 +1,110 @@
+"""What a deferred accept does to connections that never send anything.
+
+Those connections never reach opentracker, so its own client timeout cannot
+clean them up: whatever bound exists is the kernel's. This measures it, and
+then asks the question that follows from it -- whether enough of them keep
+the tracker from answering anybody else.
+
+    make opentracker
+    python3 tests/accept_filter_probe.py
+
+On FreeBSD the filter does nothing unless accf_data is loaded, so run it both
+ways to see whether the numbers differ:
+
+    kldload accf_data
+"""
+
+import os
+import socket
+import signal
+import subprocess
+import sys
+import tempfile
+import time
+import urllib.error
+import urllib.request
+
+TRACKER = os.environ.get("OPENTRACKER", "./opentracker")
+IDLE = int(os.environ.get("IDLE", "300"))
+PATIENCE = float(os.environ.get("PATIENCE", "90"))
+HASH = "".join("%%%02x" % b for b in range(1, 21))
+
+
+def start(port):
+    directory = tempfile.mkdtemp(prefix="ot-accf-")
+    with open(os.path.join(directory, "conf"), "w") as handle:
+        handle.write("listen.tcp 127.0.0.1:%d\nlisten.udp 127.0.0.1:%d\n" % (port, port))
+    log = open(os.path.join(directory, "log"), "w+b")
+    tracker = subprocess.Popen([TRACKER, "-f", os.path.join(directory, "conf")], stdout=log, stderr=log)
+    for _ in range(300):
+        try:
+            socket.create_connection(("127.0.0.1", port), timeout=0.2).close()
+            return tracker, log
+        except OSError:
+            time.sleep(0.02)
+    tracker.kill()
+    raise SystemExit("tracker never listened on %d" % port)
+
+
+def announce(port, timeout=5):
+    """One ordinary request, so we can see whether the tracker still serves."""
+    started = time.time()
+    try:
+        urllib.request.urlopen(
+            "http://127.0.0.1:%d/announce?info_hash=%s&peer_id=-OT0001-probe0000001&port=6881&left=0" % (port, HASH),
+            timeout=timeout,
+        ).read()
+        return time.time() - started, None
+    except Exception as problem:  # noqa: BLE001 - reported, not swallowed
+        return time.time() - started, problem
+
+
+def main():
+    port = int(os.environ.get("PORT", "28000"))
+    tracker, log = start(port)
+    try:
+        print("  how long a connection that says nothing is held")
+        quiet = socket.create_connection(("127.0.0.1", port), timeout=5)
+        quiet.settimeout(PATIENCE)
+        started = time.time()
+        try:
+            held = "dropped after %.1f s" % (time.time() - started) if quiet.recv(1) == b"" else "sent us something"
+        except socket.timeout:
+            held = "still open after %.0f s" % PATIENCE
+        except OSError as problem:
+            held = "ended after %.1f s: %s" % (time.time() - started, problem)
+        quiet.close()
+        print("    %s" % held)
+
+        print("  and whether %d of them keep anybody else out" % IDLE)
+        before, problem = announce(port)
+        print("    a request before: %.0f ms%s" % (before * 1000, "" if problem is None else " (%s)" % problem))
+
+        idlers = []
+        for _ in range(IDLE):
+            try:
+                idlers.append(socket.create_connection(("127.0.0.1", port), timeout=2))
+            except OSError as problem:
+                print("    could only open %d of them: %s" % (len(idlers), problem))
+                break
+        print("    holding %d connections that have sent nothing" % len(idlers))
+
+        during, problem = announce(port, timeout=10)
+        print("    a request during: %.0f ms%s" % (during * 1000, "" if problem is None else "  REFUSED: %s" % problem))
+
+        for idler in idlers:
+            idler.close()
+        after, problem = announce(port)
+        print("    a request after:  %.0f ms%s" % (after * 1000, "" if problem is None else " (%s)" % problem))
+    finally:
+        tracker.send_signal(signal.SIGINT)
+        try:
+            tracker.wait(timeout=10)
+        except subprocess.TimeoutExpired:
+            tracker.kill()
+            tracker.wait()
+        log.close()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/integration/test_expiry.py b/tests/integration/test_expiry.py
index 5a21e57..5e6dc0a 100644
--- a/tests/integration/test_expiry.py
+++ b/tests/integration/test_expiry.py
@@ -27,6 +27,11 @@ class ExpirySweep(TrackerTestCase):
     def test_the_sweep_takes_the_idle_and_leaves_the_answered(self):
         idle = socket.create_connection(("127.0.0.1", self.tracker.port), timeout=5)
         try:
+            # Enough for a kernel that holds a connection back until data
+            # arrives to hand it over, and far too little to be answered, so
+            # the tracker has accepted it and is waiting for the rest
+            idle.sendall(b"GET /ann")
+
             self.assertEqual(self.tracker.announce(self.hash_for("expiry-sweep"))[b"complete"], 1)
 
             time.sleep(CLIENT_TIMEOUT + 3 * CHECK_INTERVAL)
@@ -34,6 +39,6 @@ class ExpirySweep(TrackerTestCase):
             self.assertEqual(self.tracker.torrent_count(), 1, "the tracker stopped serving while sweeping")
 
             idle.settimeout(5)
-            self.assertEqual(idle.recv(64), b"", "a connection that never sent a request was not swept")
+            self.assertEqual(idle.recv(64), b"", "a connection that went quiet mid-request was not swept")
         finally:
             idle.close()
diff --git a/tests/syscount.c b/tests/syscount.c
new file mode 100644
index 0000000..ec45c15
--- /dev/null
+++ b/tests/syscount.c
@@ -0,0 +1,144 @@
+/* Counts the system calls a run makes, by standing in front of libc.
+
+     cc -O2 -shared -fPIC -o syscount.so tests/syscount.c -ldl
+     LD_PRELOAD=./syscount.so ./opentracker -f conf      (DYLD_INSERT_LIBRARIES on macOS)
+
+   The tally goes to stderr when the process exits. Reading the source tells
+   you which calls a path should make; this tells you which it does make,
+   which is not always the same thing. */
+
+#define _GNU_SOURCE
+#include <dlfcn.h>
+#include <stdarg.h>
+#include <stdatomic.h>
+#include <stdio.h>
+#include <sys/socket.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#ifdef __linux__
+#include <sys/epoll.h>
+#endif
+
+#define COUNTER(name) static _Atomic long g_##name
+#define BUMP(name)    atomic_fetch_add_explicit(&g_##name, 1, memory_order_relaxed)
+#define REPORT(name)  fprintf(stderr, "  %-14s %10ld\n", #name, atomic_load(&g_##name))
+
+COUNTER(accept);
+COUNTER(accept4);
+COUNTER(read);
+COUNTER(write);
+COUNTER(close);
+COUNTER(fcntl);
+COUNTER(recvfrom);
+COUNTER(sendto);
+COUNTER(epoll_ctl);
+COUNTER(epoll_wait);
+
+typedef int (*fn_accept)(int, struct sockaddr *, socklen_t *);
+typedef int (*fn_accept4)(int, struct sockaddr *, socklen_t *, int);
+typedef ssize_t (*fn_read)(int, void *, size_t);
+typedef ssize_t (*fn_write)(int, const void *, size_t);
+typedef int (*fn_close)(int);
+typedef int (*fn_fcntl)(int, int, ...);
+typedef ssize_t (*fn_recvfrom)(int, void *, size_t, int, struct sockaddr *, socklen_t *);
+typedef ssize_t (*fn_sendto)(int, const void *, size_t, int, const struct sockaddr *, socklen_t);
+
+static fn_accept   real_accept;
+static fn_accept4  real_accept4;
+static fn_read     real_read;
+static fn_write    real_write;
+static fn_close    real_close;
+static fn_fcntl    real_fcntl;
+static fn_recvfrom real_recvfrom;
+static fn_sendto   real_sendto;
+
+/* Resolving lazily keeps the library out of the way until something is
+   actually called, which matters because dlsym itself allocates */
+#define WRAP(name)                                                                                                     \
+  if (!real_##name)                                                                                                    \
+    real_##name = (fn_##name)dlsym(RTLD_NEXT, #name);                                                                  \
+  BUMP(name)
+
+int accept(int s, struct sockaddr *addr, socklen_t *len) {
+  WRAP(accept);
+  return real_accept(s, addr, len);
+}
+
+#ifdef SOCK_NONBLOCK
+int accept4(int s, struct sockaddr *addr, socklen_t *len, int flags) {
+  WRAP(accept4);
+  return real_accept4(s, addr, len, flags);
+}
+#endif
+
+ssize_t read(int fd, void *buf, size_t count) {
+  WRAP(read);
+  return real_read(fd, buf, count);
+}
+
+ssize_t write(int fd, const void *buf, size_t count) {
+  WRAP(write);
+  return real_write(fd, buf, count);
+}
+
+int close(int fd) {
+  WRAP(close);
+  return real_close(fd);
+}
+
+int fcntl(int fd, int cmd, ...) {
+  va_list args;
+  long    arg;
+
+  va_start(args, cmd);
+  arg = va_arg(args, long);
+  va_end(args);
+  WRAP(fcntl);
+  return real_fcntl(fd, cmd, arg);
+}
+
+ssize_t recvfrom(int s, void *buf, size_t len, int flags, struct sockaddr *from, socklen_t *fromlen) {
+  WRAP(recvfrom);
+  return real_recvfrom(s, buf, len, flags, from, fromlen);
+}
+
+ssize_t sendto(int s, const void *buf, size_t len, int flags, const struct sockaddr *to, socklen_t tolen) {
+  WRAP(sendto);
+  return real_sendto(s, buf, len, flags, to, tolen);
+}
+
+#ifdef __linux__
+typedef int (*fn_epoll_ctl)(int, int, int, struct epoll_event *);
+typedef int (*fn_epoll_wait)(int, struct epoll_event *, int, int);
+static fn_epoll_ctl  real_epoll_ctl;
+static fn_epoll_wait real_epoll_wait;
+
+int epoll_ctl(int ep, int op, int fd, struct epoll_event *ev) {
+  WRAP(epoll_ctl);
+  return real_epoll_ctl(ep, op, fd, ev);
+}
+
+int epoll_wait(int ep, struct epoll_event *events, int max, int timeout) {
+  WRAP(epoll_wait);
+  return real_epoll_wait(ep, events, max, timeout);
+}
+#endif
+
+__attribute__((destructor)) static void report(void) {
+  fprintf(stderr, "\n  system calls this run\n");
+  REPORT(accept);
+#ifdef SOCK_NONBLOCK
+  REPORT(accept4);
+#endif
+  REPORT(read);
+  REPORT(write);
+  REPORT(close);
+  REPORT(fcntl);
+  REPORT(recvfrom);
+  REPORT(sendto);
+#ifdef __linux__
+  REPORT(epoll_ctl);
+  REPORT(epoll_wait);
+#endif
+}
diff --git a/tests/t_event.c b/tests/t_event.c
index 5509f19..ad3948f 100644
--- a/tests/t_event.c
+++ b/tests/t_event.c
@@ -4,6 +4,7 @@
    $id$ */
 
 /* System */
+#include <fcntl.h>
 #include <string.h>
 #include <sys/socket.h>
 #include <sys/time.h>
@@ -205,6 +206,41 @@ static void case_unknown_descriptor(void) {
   CHECK_INT(event_del(4096), -1);
 }
 
+/* event_close exists so that the mechanisms which drop a closed descriptor
+   by themselves are not asked to do it twice. The one that keeps its own
+   list has to be told, and forgetting would leave the entry there to be
+   reported for ever, so what is checked here is the silence afterwards. */
+static void case_close_takes_it_out(void) {
+  int other[2], closed, count, i, strays = 0;
+
+  test_case("a closed descriptor is gone from the set and really closed");
+  pair_open();
+  CHECK(sv[0] != -1);
+  CHECK_INT(socketpair(AF_UNIX, SOCK_STREAM, 0, other), 0);
+  CHECK_INT(event_add(sv[0], EVENT_IN, 1), 0);
+  CHECK_INT(event_add(other[0], EVENT_IN, 2), 0);
+
+  closed = sv[0];
+  CHECK_INT(event_close(closed), 0);
+  sv[0] = -1;
+  CHECK_INT(fcntl(closed, F_GETFD), -1);
+
+  /* The survivor is made ready so the wait has something to come back with,
+     and anything left over from the closed one shows up beside it */
+  CHECK_INT((int)write(other[1], "x", 1), 1);
+  count = event_wait(ready, EVENTS, 200);
+  CHECK(count >= 1);
+  for (i = 0; i < count; ++i)
+    if (ready[i].fd != other[0])
+      ++strays;
+  CHECK_INT(strays, 0);
+
+  event_del(other[0]);
+  close(other[0]);
+  close(other[1]);
+  pair_close();
+}
+
 void test_event(void) {
   CHECK_INT(event_init(), 0);
   CHECK(event_mechanism() != NULL);
@@ -217,6 +253,7 @@ void test_event(void) {
   case_delete_from_the_middle();
   case_timeout();
   case_unknown_descriptor();
+  case_close_takes_it_out();
 
   event_deinit();
 }
diff --git a/tests/t_socket.c b/tests/t_socket.c
index ab4b785..20a0c0a 100644
--- a/tests/t_socket.c
+++ b/tests/t_socket.c
@@ -78,15 +78,17 @@ static int sock_flag(int s, int level, int option) {
 }
 
 /* Wait for one event, so a test never blocks forever on a lost packet */
-static int wait_for(int s, short events) {
+static int wait_for_ms(int s, short events, int timeout_ms) {
   struct pollfd pfd;
 
   pfd.fd      = s;
   pfd.events  = events;
   pfd.revents = 0;
-  return poll(&pfd, 1, 2000) == 1;
+  return poll(&pfd, 1, timeout_ms) == 1;
 }
 
+static int wait_for(int s, short events) { return wait_for_ms(s, events, 2000); }
+
 /* ------------------------------------------------------------ invariants */
 
 /* Asked for the wildcard address, because a socket bound to one specific
@@ -200,6 +202,42 @@ static void case_a_signal_ends_a_blocking_read(void) {
   close(reader_socket);
 }
 
+/* What a deferred accept is for is a descriptor that already has the request
+   in it, so what is checked is the silence before: a client that has
+   connected but said nothing must not make the listener readable. Kernels
+   that cannot do this say so, and then there is nothing to check. */
+static void case_deferred_accept(void) {
+  struct sockaddr_in6 address;
+  char                ip[16];
+  uint16_t            port = 0;
+  int                 listener, client;
+
+  test_case("a connection that has sent nothing is held back from accept");
+  listener = socket_tcp_listen(ip_loopback, 0, SOCKET_EXCLUSIVE);
+  CHECK(listener != -1);
+  if (socket_defer_accept(listener) == -1) {
+    close(listener);
+    return;
+  }
+  bound_address(listener, ip, &port);
+
+  client = socket(AF_INET6, SOCK_STREAM, 0);
+  CHECK(client != -1);
+  memset(&address, 0, sizeof(address));
+  address.sin6_family = AF_INET6;
+  address.sin6_port   = htons(port);
+  memcpy(&address.sin6_addr, ip_loopback, 16);
+  CHECK_INT(connect(client, (struct sockaddr *)&address, sizeof(address)), 0);
+
+  CHECK_INT(wait_for_ms(listener, POLLIN, 300), 0);
+
+  CHECK_INT((int)write(client, "GET / HTTP/1.0\r\n\r\n", 18), 18);
+  CHECK_INT(wait_for(listener, POLLIN), 1);
+
+  close(client);
+  close(listener);
+}
+
 /* An IPv4 peer must arrive in the v4 mapped form opentracker stores, however
    the kernel chose to report it. Linux fills a v4 mapped sockaddr_in6 here
    and the BSDs fill a plain sockaddr_in, so this is the case that tells the
@@ -352,6 +390,7 @@ void test_socket(void) {
   case_a_signal_ends_a_blocking_read();
   case_udp_v4mapped();
   case_tcp_roundtrip();
+  case_deferred_accept();
   case_connect_refused();
   case_bind_failure();
 #ifdef WANT_SYNC_LIVE