polyphony 0.50.1 → 0.53.2

data/ext/polyphony/backend_io_uring_context.c

@@ -10,6 +10,7 @@ const char *op_type_to_str(enum op_type type) {
   case OP_WRITE: return "WRITE";
   case OP_RECV: return "RECV";
   case OP_SEND: return "SEND";
+  case OP_SPLICE: return "SPLICE";
   case OP_TIMEOUT: return "TIMEOUT";
   case OP_POLL: return "POLL";
   case OP_ACCEPT: return "ACCEPT";

data/ext/polyphony/backend_io_uring_context.h

@@ -10,6 +10,7 @@ enum op_type {
   OP_WRITE,
   OP_RECV,
   OP_SEND,
+  OP_SPLICE,
   OP_TIMEOUT,
   OP_POLL,
   OP_ACCEPT,

data/ext/polyphony/backend_libev.c

@@ -1,10 +1,52 @@
+/*
+# Libev-based blocking ops backend for Polyphony
+
+## Backend initialization
+
+The backend is initialized by creating an event loop. For the main thread the
+default event loop is used, but we since we don't need to handle any signals
+(see the waitpid implementation below) we might as well use a non-default event
+loop for the main thread at some time in the future.
+
+In addition, we create an async watcher that is used for interrupting the #poll
+method from another thread.
+
+## Blocking operations
+
+I/O operations start by making sure the io has been set to non-blocking
+operation (O_NONBLOCK). That way, if the syscall would block, we'd get an
+EWOULDBLOCK or EAGAIN instead of blocking.
+
+Once the OS has indicated that the operation would block, we start a watcher
+(its type corresponding to the desired operation), and call ev_xxxx_start. in We
+then call Thread_switch_fiber and switch to another fiber while waiting for the
+watcher to be triggered.
+
+## Polling for events
+
+Backend_poll is called either once the corresponding thread has no more work to
+do (no runnable fibers) or periodically while the thread is scheduling fibers in
+order to prevent event starvation.
+
+## Behaviour of waitpid
+
+On Linux 5.3+, pidfd_open will be used, otherwise a libev child watcher will be
+used. Note that if a child watcher is used, waitpid will only work from the main
+thread.
+
+*/
+
 #ifdef POLYPHONY_BACKEND_LIBEV
 
+#ifdef POLYPHONY_LINUX
+#define _GNU_SOURCE 1
+#endif
+
+#include <fcntl.h>
 #include <netdb.h>
 #include <sys/socket.h>
 #include <sys/uio.h>
 #include <unistd.h>
-#include <fcntl.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
 #include <stdnoreturn.h>
@@ -16,6 +58,9 @@
 #include "ruby/io.h"
 
 VALUE SYM_libev;
+VALUE SYM_send;
+VALUE SYM_splice;
+VALUE SYM_write;
 
 ID ID_ivar_is_nonblocking;
 
@@ -76,17 +121,27 @@ void break_async_callback(struct ev_loop *ev_loop, struct ev_async *ev_async, in
   // of a *blocking* event loop (waking it up) in a thread-safe, signal-safe manner
 }
 
+inline struct ev_loop *libev_new_loop() {
+  #ifdef POLYPHONY_USE_PIDFD_OPEN
+    return ev_loop_new(EVFLAG_NOSIGMASK);
+  #else
+    int is_main_thread = (rb_thread_current() == rb_thread_main());
+    return is_main_thread ? EV_DEFAULT : ev_loop_new(EVFLAG_NOSIGMASK);
+  #endif
+}
+
 static VALUE Backend_initialize(VALUE self) {
   Backend_t *backend;
-  VALUE thread = rb_thread_current();
-  int is_main_thread = (thread == rb_thread_main());
-
+  
   GetBackend(self, backend);
-  backend->ev_loop = is_main_thread ? EV_DEFAULT : ev_loop_new(EVFLAG_NOSIGMASK);
+  backend->ev_loop = libev_new_loop();
 
+  // start async watcher used for breaking a poll op (from another thread)
   ev_async_init(&backend->break_async, break_async_callback);
   ev_async_start(backend->ev_loop, &backend->break_async);
-  ev_unref(backend->ev_loop); // don't count the break_async watcher
+  // the break_async watcher is unreferenced, in order for Backend_poll to not
+  // block when no other watcher is active
+  ev_unref(backend->ev_loop);
 
   backend->currently_polling = 0;
   backend->pending_count = 0;
@@ -334,6 +389,58 @@ error:
   return RAISE_EXCEPTION(switchpoint_result);
 }
 
+VALUE Backend_feed_loop(VALUE self, VALUE io, VALUE receiver, VALUE method) {
+  Backend_t *backend;
+  struct libev_io watcher;
+  rb_io_t *fptr;
+  VALUE str;
+  long total;
+  long len = 8192;
+  int shrinkable;
+  char *buf;
+  VALUE switchpoint_result = Qnil;
+  VALUE underlying_io = rb_ivar_get(io, ID_ivar_io);
+  ID method_id = SYM2ID(method);
+
+  READ_LOOP_PREPARE_STR();
+
+  GetBackend(self, backend);
+  if (underlying_io != Qnil) io = underlying_io;
+  GetOpenFile(io, fptr);
+  rb_io_check_byte_readable(fptr);
+  io_set_nonblock(fptr, io);
+  rectify_io_file_pos(fptr);
+  watcher.fiber = Qnil;
+
+  while (1) {
+    ssize_t n = read(fptr->fd, buf, len);
+    if (n < 0) {
+      int e = errno;
+      if ((e != EWOULDBLOCK && e != EAGAIN)) rb_syserr_fail(e, strerror(e));
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, fptr->fd, &watcher, EV_READ);
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+    }
+    else {
+      switchpoint_result = backend_snooze();
+
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+
+      if (n == 0) break; // EOF
+      total = n;
+      READ_LOOP_PASS_STR_TO_RECEIVER(receiver, method_id);
+    }
+  }
+
+  RB_GC_GUARD(str);
+  RB_GC_GUARD(watcher.fiber);
+  RB_GC_GUARD(switchpoint_result);
+
+  return io;
+error:
+  return RAISE_EXCEPTION(switchpoint_result);
+}
+
 VALUE Backend_write(VALUE self, VALUE io, VALUE str) {
   Backend_t *backend;
   struct libev_io watcher;
@@ -620,6 +727,167 @@ error:
   return RAISE_EXCEPTION(switchpoint_result);
 }
 
+VALUE Backend_send(VALUE self, VALUE io, VALUE str, VALUE flags) {
+  Backend_t *backend;
+  struct libev_io watcher;
+  rb_io_t *fptr;
+  VALUE switchpoint_result = Qnil;
+  VALUE underlying_io;
+  char *buf = StringValuePtr(str);
+  long len = RSTRING_LEN(str);
+  long left = len;
+  int flags_int = NUM2INT(flags);
+
+  underlying_io = rb_ivar_get(io, ID_ivar_io);
+  if (underlying_io != Qnil) io = underlying_io;
+  GetBackend(self, backend);
+  io = rb_io_get_write_io(io);
+  GetOpenFile(io, fptr);
+  io_set_nonblock(fptr, io);
+  watcher.fiber = Qnil;
+
+  while (left > 0) {
+    ssize_t n = send(fptr->fd, buf, left, flags_int);
+    if (n < 0) {
+      int e = errno;
+      if ((e != EWOULDBLOCK && e != EAGAIN)) rb_syserr_fail(e, strerror(e));
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, fptr->fd, &watcher, EV_WRITE);
+
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+    }
+    else {
+      buf += n;
+      left -= n;
+    }
+  }
+
+  if (watcher.fiber == Qnil) {
+    switchpoint_result = backend_snooze();
+
+    if (TEST_EXCEPTION(switchpoint_result)) goto error;
+  }
+
+  RB_GC_GUARD(watcher.fiber);
+  RB_GC_GUARD(switchpoint_result);
+
+  return INT2NUM(len);
+error:
+  return RAISE_EXCEPTION(switchpoint_result);
+}
+
+#ifdef POLYPHONY_LINUX
+VALUE Backend_splice(VALUE self, VALUE src, VALUE dest, VALUE maxlen) {
+  Backend_t *backend;
+  struct libev_io watcher;
+  VALUE switchpoint_result = Qnil;
+  VALUE underlying_io;
+  rb_io_t *src_fptr;
+  rb_io_t *dest_fptr;
+  int len;
+
+  GetBackend(self, backend);
+
+  underlying_io = rb_ivar_get(src, ID_ivar_io);
+  if (underlying_io != Qnil) src = underlying_io;
+  GetOpenFile(src, src_fptr);
+  io_set_nonblock(src_fptr, src);
+
+  underlying_io = rb_ivar_get(dest, ID_ivar_io);
+  if (underlying_io != Qnil) dest = underlying_io;
+  dest = rb_io_get_write_io(dest);
+  GetOpenFile(dest, dest_fptr);
+  io_set_nonblock(dest_fptr, dest);
+
+  watcher.fiber = Qnil;
+  while (1) {
+    len = splice(src_fptr->fd, 0, dest_fptr->fd, 0, NUM2INT(maxlen), 0);
+    if (len < 0) {
+      int e = errno;
+      if ((e != EWOULDBLOCK && e != EAGAIN)) rb_syserr_fail(e, strerror(e));
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, src_fptr->fd, &watcher, EV_READ);
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, dest_fptr->fd, &watcher, EV_WRITE);
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+    }
+    else {
+      break;
+    }
+  }
+
+  if (watcher.fiber == Qnil) {
+    switchpoint_result = backend_snooze();
+    if (TEST_EXCEPTION(switchpoint_result)) goto error;
+  }
+
+  RB_GC_GUARD(watcher.fiber);
+  RB_GC_GUARD(switchpoint_result);
+
+  return INT2NUM(len);
+error:
+  return RAISE_EXCEPTION(switchpoint_result);
+}
+
+VALUE Backend_splice_to_eof(VALUE self, VALUE src, VALUE dest, VALUE maxlen) {
+  Backend_t *backend;
+  struct libev_io watcher;
+  VALUE switchpoint_result = Qnil;
+  VALUE underlying_io;
+  rb_io_t *src_fptr;
+  rb_io_t *dest_fptr;
+  int len;
+  int total = 0;
+
+  GetBackend(self, backend);
+
+  underlying_io = rb_ivar_get(src, ID_ivar_io);
+  if (underlying_io != Qnil) src = underlying_io;
+  GetOpenFile(src, src_fptr);
+  io_set_nonblock(src_fptr, src);
+
+  underlying_io = rb_ivar_get(dest, ID_ivar_io);
+  if (underlying_io != Qnil) dest = underlying_io;
+  dest = rb_io_get_write_io(dest);
+  GetOpenFile(dest, dest_fptr);
+  io_set_nonblock(dest_fptr, dest);
+
+  watcher.fiber = Qnil;
+  while (1) {
+    len = splice(src_fptr->fd, 0, dest_fptr->fd, 0, NUM2INT(maxlen), 0);
+    if (len < 0) {
+      int e = errno;
+      if ((e != EWOULDBLOCK && e != EAGAIN)) rb_syserr_fail(e, strerror(e));
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, src_fptr->fd, &watcher, EV_READ);
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+
+      switchpoint_result = libev_wait_fd_with_watcher(backend, dest_fptr->fd, &watcher, EV_WRITE);
+      if (TEST_EXCEPTION(switchpoint_result)) goto error;
+    }
+    else if (len == 0) {
+      break;
+    }
+    else {
+      total += len;
+    }
+  }
+
+  if (watcher.fiber == Qnil) {
+    switchpoint_result = backend_snooze();
+    if (TEST_EXCEPTION(switchpoint_result)) goto error;
+  }
+
+  RB_GC_GUARD(watcher.fiber);
+  RB_GC_GUARD(switchpoint_result);
+
+  return INT2NUM(total);
+error:
+  return RAISE_EXCEPTION(switchpoint_result);
+}
+#endif
+
 VALUE Backend_wait_io(VALUE self, VALUE io, VALUE write) {
   Backend_t *backend;
   rb_io_t *fptr;
@@ -747,6 +1015,7 @@ VALUE Backend_timeout(int argc,VALUE *argv, VALUE self) {
   return result;
 }
 
+#ifdef POLYPHONY_USE_PIDFD_OPEN
 VALUE Backend_waitpid(VALUE self, VALUE pid) {
   int pid_int = NUM2INT(pid);
   int fd = pidfd_open(pid_int, 0);
@@ -759,18 +1028,53 @@ VALUE Backend_waitpid(VALUE self, VALUE pid) {
     RAISE_IF_EXCEPTION(resume_value);
     RB_GC_GUARD(resume_value);
   }
+  else {
+    int e = errno;
+    rb_syserr_fail(e, strerror(e));
+  }
 
   int status = 0;
   pid_t ret = waitpid(pid_int, &status, WNOHANG);
   if (ret < 0) {
     int e = errno;
-    if (e == ECHILD)
-      ret = pid_int;
-    else
-      rb_syserr_fail(e, strerror(e));
+    rb_syserr_fail(e, strerror(e));
   }
   return rb_ary_new_from_args(2, INT2NUM(ret), INT2NUM(WEXITSTATUS(status)));
 }
+#else
+struct libev_child {
+  struct ev_child child;
+  VALUE fiber;
+};
+
+void Backend_child_callback(EV_P_ ev_child *w, int revents) {
+  struct libev_child *watcher = (struct libev_child *)w;
+  int exit_status = WEXITSTATUS(w->rstatus);
+  VALUE status;
+
+  status = rb_ary_new_from_args(2, INT2NUM(w->rpid), INT2NUM(exit_status));
+  Fiber_make_runnable(watcher->fiber, status);
+}
+
+VALUE Backend_waitpid(VALUE self, VALUE pid) {
+  Backend_t *backend;
+  struct libev_child watcher;
+  VALUE switchpoint_result = Qnil;
+  GetBackend(self, backend);
+
+  watcher.fiber = rb_fiber_current();
+  ev_child_init(&watcher.child, Backend_child_callback, NUM2INT(pid), 0);
+  ev_child_start(backend->ev_loop, &watcher.child);
+
+  switchpoint_result = backend_await(backend);
+
+  ev_child_stop(backend->ev_loop, &watcher.child);
+  RAISE_IF_EXCEPTION(switchpoint_result);
+  RB_GC_GUARD(watcher.fiber);
+  RB_GC_GUARD(switchpoint_result);
+  return switchpoint_result;
+}
+#endif
 
 void Backend_async_callback(EV_P_ ev_async *w, int revents) { }
 
@@ -796,10 +1100,35 @@ VALUE Backend_kind(VALUE self) {
   return SYM_libev;
 }
 
+VALUE Backend_chain(int argc,VALUE *argv, VALUE self) {
+  VALUE result = Qnil;
+  if (argc == 0) return result;
+
+  for (int i = 0; i < argc; i++) {
+    VALUE op = argv[i];
+    VALUE op_type = RARRAY_AREF(op, 0);
+    VALUE op_len = RARRAY_LEN(op);
+
+    if (op_type == SYM_write && op_len == 3)
+      result = Backend_write(self, RARRAY_AREF(op, 1), RARRAY_AREF(op, 2));
+    else if (op_type == SYM_send && op_len == 4)
+      result = Backend_send(self, RARRAY_AREF(op, 1), RARRAY_AREF(op, 2), RARRAY_AREF(op, 3));
+    #ifdef POLYPHONY_LINUX
+    else if (op_type == SYM_splice && op_len == 4)
+      result = Backend_splice(self, RARRAY_AREF(op, 1), RARRAY_AREF(op, 2), RARRAY_AREF(op, 3));
+    #endif
+    else
+      rb_raise(rb_eRuntimeError, "Invalid op specified or bad op arity");
+  }
+  
+  RB_GC_GUARD(result);
+  return result;
+}
+
 void Init_Backend() {
   ev_set_allocator(xrealloc);
 
-  VALUE cBackend = rb_define_class_under(mPolyphony, "Backend", rb_cData);
+  VALUE cBackend = rb_define_class_under(mPolyphony, "Backend", rb_cObject);
   rb_define_alloc_func(cBackend, Backend_allocate);
 
   rb_define_method(cBackend, "initialize", Backend_initialize, 0);
@@ -808,27 +1137,40 @@ void Init_Backend() {
 
   rb_define_method(cBackend, "poll", Backend_poll, 3);
   rb_define_method(cBackend, "break", Backend_wakeup, 0);
+  rb_define_method(cBackend, "kind", Backend_kind, 0);
+  rb_define_method(cBackend, "chain", Backend_chain, -1);
 
-  rb_define_method(cBackend, "read", Backend_read, 4);
-  rb_define_method(cBackend, "read_loop", Backend_read_loop, 1);
-  rb_define_method(cBackend, "write", Backend_write_m, -1);
   rb_define_method(cBackend, "accept", Backend_accept, 2);
   rb_define_method(cBackend, "accept_loop", Backend_accept_loop, 2);
   rb_define_method(cBackend, "connect", Backend_connect, 3);
+  rb_define_method(cBackend, "feed_loop", Backend_feed_loop, 3);
+  rb_define_method(cBackend, "read", Backend_read, 4);
+  rb_define_method(cBackend, "read_loop", Backend_read_loop, 1);
   rb_define_method(cBackend, "recv", Backend_recv, 3);
   rb_define_method(cBackend, "recv_loop", Backend_read_loop, 1);
-  rb_define_method(cBackend, "send", Backend_write, 2);
-  rb_define_method(cBackend, "wait_io", Backend_wait_io, 2);
+  rb_define_method(cBackend, "recv_feed_loop", Backend_feed_loop, 3);
+  rb_define_method(cBackend, "send", Backend_send, 3);
+  rb_define_method(cBackend, "sendv", Backend_sendv, 3);
   rb_define_method(cBackend, "sleep", Backend_sleep, 1);
-  rb_define_method(cBackend, "timer_loop", Backend_timer_loop, 1);
+
+  #ifdef POLYPHONY_LINUX
+  rb_define_method(cBackend, "splice", Backend_splice, 3);
+  rb_define_method(cBackend, "splice_to_eof", Backend_splice_to_eof, 3);
+  #endif
+
   rb_define_method(cBackend, "timeout", Backend_timeout, -1);
-  rb_define_method(cBackend, "waitpid", Backend_waitpid, 1);
+  rb_define_method(cBackend, "timer_loop", Backend_timer_loop, 1);
   rb_define_method(cBackend, "wait_event", Backend_wait_event, 1);
-
-  rb_define_method(cBackend, "kind", Backend_kind, 0);
+  rb_define_method(cBackend, "wait_io", Backend_wait_io, 2);
+  rb_define_method(cBackend, "waitpid", Backend_waitpid, 1);
+  rb_define_method(cBackend, "write", Backend_write_m, -1);
 
   ID_ivar_is_nonblocking = rb_intern("@is_nonblocking");
   SYM_libev = ID2SYM(rb_intern("libev"));
+
+  SYM_send = ID2SYM(rb_intern("send"));
+  SYM_splice = ID2SYM(rb_intern("splice"));
+  SYM_write = ID2SYM(rb_intern("write"));
 }
 
 #endif // POLYPHONY_BACKEND_LIBEV