freebsd,linux: add recvmmsg() + sendmmsg() udp implementation

This commits adds support for recvmmsg() and sendmmsg() extensions to
recvmsg() and sendmsg() that allows the caller to receive and send
multiple message from a socket using a single system call. This has
performance benefits for some applications.

Co-authored-by: Ondřej Surý <ondrej@sury.org>
Co-authored-by: Witold Kręcicki <wpk@culm.net>

PR-URL: https://github.com/libuv/libuv/pull/2532
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
Reviewed-By: Saúl Ibarra Corretgé <s@saghul.net>
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>

docs/src/udp.rst

@@ -42,6 +42,12 @@ Data types
             * any traffic, in effect "stealing" the port from the previous listener.
             */
             UV_UDP_REUSEADDR = 4
+            /*
+             * Indicates that the message was received by recvmmsg and that it's not at
+             * the beginning of the buffer allocated by alloc_cb - so the buffer provided
+             * must not be freed by the recv_cb callback.
+             */
+            UV_UDP_MMSG_CHUNK = 8
         };
 
 .. c:type:: void (*uv_udp_send_cb)(uv_udp_send_t* req, int status)
@@ -62,12 +68,13 @@ Data types
     * `buf`: :c:type:`uv_buf_t` with the received data.
     * `addr`: ``struct sockaddr*`` containing the address of the sender.
       Can be NULL. Valid for the duration of the callback only.
-    * `flags`: One or more or'ed UV_UDP_* constants. Right now only
+    * `flags`: One or more or'ed UV_UDP_* constants.
-      ``UV_UDP_PARTIAL`` is used.
 
     The callee is responsible for freeing the buffer, libuv does not reuse it.
     The buffer may be a null buffer (where `buf->base` == NULL and `buf->len` == 0)
-    on error.
+    on error. Don't free the buffer when the UV_UDP_MMSG_CHUNK flag is set.
+    The final callback receives the whole buffer (containing the first chunk)
+    with the UV_UDP_MMSG_CHUNK flag cleared.
 
     .. note::
         The receive callback will be called with `nread` == 0 and `addr` == NULL when there is

include/uv.h

@@ -605,7 +605,13 @@ enum uv_udp_flags {
    * (provided they all set the flag) but only the last one to bind will receive
    * any traffic, in effect "stealing" the port from the previous listener.
    */
-  UV_UDP_REUSEADDR = 4
+  UV_UDP_REUSEADDR = 4,
+  /*
+   * Indicates that the message was received by recvmmsg and that it's not at
+   * the beginning of the buffer allocated by alloc_cb - so the buffer provided
+   * must not be freed by the recv_cb callback.
+   */
+  UV_UDP_MMSG_CHUNK = 8
 };
 
 typedef void (*uv_udp_send_cb)(uv_udp_send_t* req, int status);

src/unix/freebsd.c

@@ -288,3 +288,28 @@ int uv_cpu_info(uv_cpu_info_t** cpu_infos, int* count) {
   uv__free(cp_times);
   return 0;
 }
+
+
+int uv__sendmmsg(int fd,
+                 struct uv__mmsghdr* mmsg,
+                 unsigned int vlen,
+                 unsigned int flags) {
+#if __FreeBSD__ >= 11
+  return sendmmsg(fd, mmsg, vlen, flags);
+#else
+  return errno = ENOSYS, -1;
+#endif
+}
+
+
+int uv__recvmmsg(int fd,
+                 struct uv__mmsghdr* mmsg,
+                 unsigned int vlen,
+                 unsigned int flags,
+                 struct timespec* timeout) {
+#if __FreeBSD__ >= 11
+  return recvmmsg(fd, mmsg, vlen, flags, timeout);
+#else
+  return errno = ENOSYS, -1;
+#endif
+}

src/unix/internal.h

@@ -31,6 +31,7 @@
 #include <fcntl.h>  /* O_CLOEXEC and O_NONBLOCK, if supported. */
 #include <stdio.h>
 #include <errno.h>
+#include <sys/socket.h>
 
 #if defined(__STRICT_ANSI__)
 # define inline __inline
@@ -327,4 +328,27 @@ int uv__getsockpeername(const uv_handle_t* handle,
                         struct sockaddr* name,
                         int* namelen);
 
+#if defined(__linux__)            ||                                      \
+    defined(__FreeBSD__)          ||                                      \
+    defined(__FreeBSD_kernel__)
+#define HAVE_MMSG 1
+struct uv__mmsghdr {
+  struct msghdr msg_hdr;
+  unsigned int msg_len;
+};
+
+int uv__recvmmsg(int fd,
+                 struct uv__mmsghdr* mmsg,
+                 unsigned int vlen,
+                 unsigned int flags,
+                 struct timespec* timeout);
+int uv__sendmmsg(int fd,
+                 struct uv__mmsghdr* mmsg,
+                 unsigned int vlen,
+                 unsigned int flags);
+#else
+#define HAVE_MMSG 0
+#endif
+
+
 #endif /* UV_UNIX_INTERNAL_H_ */

src/unix/linux-syscalls.c

@@ -126,6 +126,7 @@
 # endif
 #endif /* __NR_getrandom */
 
+struct uv__mmsghdr;
 
 int uv__sendmmsg(int fd,
                  struct uv__mmsghdr* mmsg,

src/unix/linux-syscalls.h

@@ -61,20 +61,6 @@ struct uv__statx {
   uint64_t unused1[14];
 };
 
-struct uv__mmsghdr {
-  struct msghdr msg_hdr;
-  unsigned int msg_len;
-};
-
-int uv__recvmmsg(int fd,
-                 struct uv__mmsghdr* mmsg,
-                 unsigned int vlen,
-                 unsigned int flags,
-                 struct timespec* timeout);
-int uv__sendmmsg(int fd,
-                 struct uv__mmsghdr* mmsg,
-                 unsigned int vlen,
-                 unsigned int flags);
 ssize_t uv__preadv(int fd, const struct iovec *iov, int iovcnt, int64_t offset);
 ssize_t uv__pwritev(int fd, const struct iovec *iov, int iovcnt, int64_t offset);
 int uv__dup3(int oldfd, int newfd, int flags);

src/unix/udp.c

@@ -32,6 +32,8 @@
 #endif
 #include <sys/un.h>
 
+#define UV__UDP_DGRAM_MAXSIZE (64 * 1024)
+
 #if defined(IPV6_JOIN_GROUP) && !defined(IPV6_ADD_MEMBERSHIP)
 # define IPV6_ADD_MEMBERSHIP IPV6_JOIN_GROUP
 #endif
@@ -49,6 +51,36 @@ static int uv__udp_maybe_deferred_bind(uv_udp_t* handle,
                                        int domain,
                                        unsigned int flags);
 
+#if HAVE_MMSG
+
+#define UV__MMSG_MAXWIDTH 20
+
+static int uv__udp_recvmmsg(uv_udp_t* handle, uv_buf_t* buf);
+static void uv__udp_sendmmsg(uv_udp_t* handle);
+
+static int uv__recvmmsg_avail;
+static int uv__sendmmsg_avail;
+static uv_once_t once = UV_ONCE_INIT;
+
+static void uv__udp_mmsg_init(void) {
+  int ret;
+  int s;
+  s = uv__socket(AF_INET, SOCK_DGRAM, 0);
+  if (s < 0)
+    return;
+  ret = uv__sendmmsg(s, NULL, 0, 0);
+  if (ret == 0 || errno != ENOSYS) {
+    uv__sendmmsg_avail = 1;
+    uv__recvmmsg_avail = 1;
+  } else {
+    ret = uv__recvmmsg(s, NULL, 0, 0, NULL);
+    if (ret == 0 || errno != ENOSYS)
+      uv__recvmmsg_avail = 1;
+  }
+  uv__close(s);
+}
+
+#endif
 
 void uv__udp_close(uv_udp_t* handle) {
   uv__io_close(handle->loop, &handle->io_watcher);
@@ -148,6 +180,60 @@ static void uv__udp_io(uv_loop_t* loop, uv__io_t* w, unsigned int revents) {
   }
 }
 
+#if HAVE_MMSG
+static int uv__udp_recvmmsg(uv_udp_t* handle, uv_buf_t* buf) {
+  struct sockaddr_in6 peers[UV__MMSG_MAXWIDTH];
+  struct iovec iov[UV__MMSG_MAXWIDTH];
+  struct uv__mmsghdr msgs[UV__MMSG_MAXWIDTH];
+  ssize_t nread;
+  uv_buf_t chunk_buf;
+  size_t chunks;
+  int flags;
+  size_t k;
+
+  /* prepare structures for recvmmsg */
+  chunks = buf->len / UV__UDP_DGRAM_MAXSIZE;
+  if (chunks > ARRAY_SIZE(iov))
+    chunks = ARRAY_SIZE(iov);
+  for (k = 0; k < chunks; ++k) {
+    iov[k].iov_base = buf->base + k * UV__UDP_DGRAM_MAXSIZE;
+    iov[k].iov_len = UV__UDP_DGRAM_MAXSIZE;
+    msgs[k].msg_hdr.msg_iov = iov + k;
+    msgs[k].msg_hdr.msg_iovlen = 1;
+    msgs[k].msg_hdr.msg_name = peers + k;
+    msgs[k].msg_hdr.msg_namelen = sizeof(peers[0]);
+  }
+
+  do
+    nread = uv__recvmmsg(handle->io_watcher.fd, msgs, chunks, 0, NULL);
+  while (nread == -1 && errno == EINTR);
+
+  if (nread < 1) {
+    if (nread == 0 || errno == EAGAIN || errno == EWOULDBLOCK)
+      handle->recv_cb(handle, 0, buf, NULL, 0);
+    else
+      handle->recv_cb(handle, UV__ERR(errno), buf, NULL, 0);
+  } else {
+    /* count to zero, so the buffer base comes last */
+    for (k = nread; k > 0 && handle->recv_cb != NULL;) {
+      k--;
+      flags = 0;
+      if (msgs[k].msg_hdr.msg_flags & MSG_TRUNC)
+        flags |= UV_UDP_PARTIAL;
+      if (k != 0)
+        flags |= UV_UDP_MMSG_CHUNK;
+
+      chunk_buf = uv_buf_init(iov[k].iov_base, iov[k].iov_len);
+      handle->recv_cb(handle,
+                      msgs[k].msg_len,
+                      &chunk_buf,
+                      msgs[k].msg_hdr.msg_name,
+                      flags);
+    }
+  }
+  return nread;
+}
+#endif
 
 static void uv__udp_recvmsg(uv_udp_t* handle) {
   struct sockaddr_storage peer;
@@ -167,13 +253,27 @@ static void uv__udp_recvmsg(uv_udp_t* handle) {
 
   do {
     buf = uv_buf_init(NULL, 0);
-    handle->alloc_cb((uv_handle_t*) handle, 64 * 1024, &buf);
+    handle->alloc_cb((uv_handle_t*) handle, UV__UDP_DGRAM_MAXSIZE, &buf);
     if (buf.base == NULL || buf.len == 0) {
       handle->recv_cb(handle, UV_ENOBUFS, &buf, NULL, 0);
       return;
     }
     assert(buf.base != NULL);
 
+#if HAVE_MMSG
+    uv_once(&once, uv__udp_mmsg_init);
+    if (uv__recvmmsg_avail) {
+      /* Returned space for more than 1 datagram, use it to receive
+       * multiple datagrams. */
+      if (buf.len >= 2 * UV__UDP_DGRAM_MAXSIZE) {
+        nread = uv__udp_recvmmsg(handle, &buf);
+        if (nread > 0)
+          count -= nread;
+        continue;
+      }
+    }
+#endif
+
     memset(&h, 0, sizeof(h));
     memset(&peer, 0, sizeof(peer));
     h.msg_name = &peer;
@@ -199,21 +299,120 @@ static void uv__udp_recvmsg(uv_udp_t* handle) {
 
       handle->recv_cb(handle, nread, &buf, (const struct sockaddr*) &peer, flags);
     }
+    count--;
   }
   /* recv_cb callback may decide to pause or close the handle */
   while (nread != -1
-      && count-- > 0
+      && count > 0
       && handle->io_watcher.fd != -1
       && handle->recv_cb != NULL);
 }
 
+#if HAVE_MMSG
+static void uv__udp_sendmmsg(uv_udp_t* handle) {
+  uv_udp_send_t* req;
+  struct uv__mmsghdr h[UV__MMSG_MAXWIDTH];
+  struct uv__mmsghdr *p;
+  QUEUE* q;
+  ssize_t npkts;
+  size_t pkts;
+  size_t i;
+
+  if (QUEUE_EMPTY(&handle->write_queue))
+    return;
+
+write_queue_drain:
+  for (pkts = 0, q = QUEUE_HEAD(&handle->write_queue);
+       pkts < UV__MMSG_MAXWIDTH && q != &handle->write_queue;
+       ++pkts, q = QUEUE_HEAD(q)) {
+    assert(q != NULL);
+    req = QUEUE_DATA(q, uv_udp_send_t, queue);
+    assert(req != NULL);
+
+    p = &h[pkts];
+    memset(p, 0, sizeof(*p));
+    if (req->addr.ss_family == AF_UNSPEC) {
+      p->msg_hdr.msg_name = NULL;
+      p->msg_hdr.msg_namelen = 0;
+    } else {
+      p->msg_hdr.msg_name = &req->addr;
+      if (req->addr.ss_family == AF_INET6)
+        p->msg_hdr.msg_namelen = sizeof(struct sockaddr_in6);
+      else if (req->addr.ss_family == AF_INET)
+        p->msg_hdr.msg_namelen = sizeof(struct sockaddr_in);
+      else if (req->addr.ss_family == AF_UNIX)
+        p->msg_hdr.msg_namelen = sizeof(struct sockaddr_un);
+      else {
+        assert(0 && "unsupported address family");
+        abort();
+      }
+    }
+    h[pkts].msg_hdr.msg_iov = (struct iovec*) req->bufs;
+    h[pkts].msg_hdr.msg_iovlen = req->nbufs;
+  }
+
+  do
+    npkts = uv__sendmmsg(handle->io_watcher.fd, h, pkts, 0);
+  while (npkts == -1 && errno == EINTR);
+
+  if (npkts < 1) {
+    if (errno == EAGAIN || errno == EWOULDBLOCK || errno == ENOBUFS)
+      return;
+    for (i = 0, q = QUEUE_HEAD(&handle->write_queue);
+         i < pkts && q != &handle->write_queue;
+         ++i, q = QUEUE_HEAD(q)) {
+      assert(q != NULL);
+      req = QUEUE_DATA(q, uv_udp_send_t, queue);
+      assert(req != NULL);
+
+      req->status = UV__ERR(errno);
+      QUEUE_REMOVE(&req->queue);
+      QUEUE_INSERT_TAIL(&handle->write_completed_queue, &req->queue);
+    }
+    uv__io_feed(handle->loop, &handle->io_watcher);
+    return;
+  }
+
+  for (i = 0, q = QUEUE_HEAD(&handle->write_queue);
+       i < pkts && q != &handle->write_queue;
+       ++i, q = QUEUE_HEAD(&handle->write_queue)) {
+    assert(q != NULL);
+    req = QUEUE_DATA(q, uv_udp_send_t, queue);
+    assert(req != NULL);
+
+    req->status = req->bufs[0].len;
+
+    /* Sending a datagram is an atomic operation: either all data
+     * is written or nothing is (and EMSGSIZE is raised). That is
+     * why we don't handle partial writes. Just pop the request
+     * off the write queue and onto the completed queue, done.
+     */
+    QUEUE_REMOVE(&req->queue);
+    QUEUE_INSERT_TAIL(&handle->write_completed_queue, &req->queue);
+  }
+
+  /* couldn't batch everything, continue sending (jump to avoid stack growth) */
+  if (!QUEUE_EMPTY(&handle->write_queue))
+    goto write_queue_drain;
+  uv__io_feed(handle->loop, &handle->io_watcher);
+  return;
+}
+#endif
 
 static void uv__udp_sendmsg(uv_udp_t* handle) {
   uv_udp_send_t* req;
-  QUEUE* q;
   struct msghdr h;
+  QUEUE* q;
   ssize_t size;
 
+#if HAVE_MMSG
+  uv_once(&once, uv__udp_mmsg_init);
+  if (uv__sendmmsg_avail) {
+    uv__udp_sendmmsg(handle);
+    return;
+  }
+#endif
+
   while (!QUEUE_EMPTY(&handle->write_queue)) {
     q = QUEUE_HEAD(&handle->write_queue);
     assert(q != NULL);
@@ -263,7 +462,6 @@ static void uv__udp_sendmsg(uv_udp_t* handle) {
   }
 }
 
-
 /* On the BSDs, SO_REUSEPORT implies SO_REUSEADDR but with some additional
  * refinements for programs that use multicast.
  *