vcl: nest vcl_mq_epfd to support epoll_wait without high CPU usage

Now, libc epfd and vls epfd are independent and can only epoll_wait independently without timeout, then app calling epoll_wait will occupy high CPU. So we nest vcl_mq_epfd into libc epfd when using eventfd with VPP, and then we can only epoll_wait libc epfd with specified timeout.

Type: feature
Signed-off-by: hanlin <hanlin_wang@163.com>
Change-Id: I6b6e0f501c769e186714bfbc187cfaed2533b4c2
Signed-off-by: hanlin <hanlin_wang@163.com>

3 files changed, 125 insertions, 2 deletions

diff --git a/src/vcl/ldp.c b/src/vcl/ldp.c
index 091c7e5cff4..e3e6dfac1c6 100644
--- a/src/vcl/ldp.c
+++ b/src/vcl/ldp.c
@@ -80,6 +80,7 @@ typedef struct ldp_worker_ctx_
    * Epoll state
    */
   u8 epoll_wait_vcl;
+  u8 mq_epfd_added;
   int vcl_mq_epfd;
 
 } ldp_worker_ctx_t;
@@ -2344,17 +2345,132 @@ done:
   return rv;
 }
 
+static inline int
+ldp_epoll_pwait_eventfd (int epfd, struct epoll_event *events,
+			 int maxevents, int timeout, const sigset_t * sigmask)
+{
+  ldp_worker_ctx_t *ldpw = ldp_worker_get_current ();
+  int libc_epfd, rv = 0, num_ev;
+  vls_handle_t ep_vlsh;
+
+  if ((errno = -ldp_init ()))
+    return -1;
+
+  if (PREDICT_FALSE (!events || (timeout < -1)))
+    {
+      errno = EFAULT;
+      return -1;
+    }
+
+  if (epfd == ldpw->vcl_mq_epfd)
+    return libc_epoll_pwait (epfd, events, maxevents, timeout, sigmask);
+
+  ep_vlsh = ldp_fd_to_vlsh (epfd);
+  if (PREDICT_FALSE (ep_vlsh == VLS_INVALID_HANDLE))
+    {
+      LDBG (0, "epfd %d: bad ep_vlsh %d!", epfd, ep_vlsh);
+      errno = EBADFD;
+      return -1;
+    }
+
+  libc_epfd = vls_attr (ep_vlsh, VPPCOM_ATTR_GET_LIBC_EPFD, 0, 0);
+  if (PREDICT_FALSE (!libc_epfd))
+    {
+      u32 size = sizeof (epfd);
+
+      LDBG (1, "epfd %d, vep_vlsh %d calling libc_epoll_create1: "
+	    "EPOLL_CLOEXEC", epfd, ep_vlsh);
+      libc_epfd = libc_epoll_create1 (EPOLL_CLOEXEC);
+      if (libc_epfd < 0)
+	{
+	  rv = libc_epfd;
+	  goto done;
+	}
+
+      rv = vls_attr (ep_vlsh, VPPCOM_ATTR_SET_LIBC_EPFD, &libc_epfd, &size);
+      if (rv < 0)
+	{
+	  errno = -rv;
+	  rv = -1;
+	  goto done;
+	}
+    }
+  if (PREDICT_FALSE (libc_epfd <= 0))
+    {
+      errno = -libc_epfd;
+      rv = -1;
+      goto done;
+    }
+
+  if (PREDICT_FALSE (!ldpw->mq_epfd_added))
+    {
+      struct epoll_event e = { 0 };
+      e.events = EPOLLIN;
+      e.data.fd = ldpw->vcl_mq_epfd;
+      if (libc_epoll_ctl (libc_epfd, EPOLL_CTL_ADD, ldpw->vcl_mq_epfd, &e) <
+	  0)
+	{
+	  LDBG (0, "epfd %d, add libc mq epoll fd %d to libc epoll fd %d",
+		epfd, ldpw->vcl_mq_epfd, libc_epfd);
+	  rv = -1;
+	  goto done;
+	}
+      ldpw->mq_epfd_added = 1;
+    }
+
+  rv = vls_epoll_wait (ep_vlsh, events, maxevents, 0);
+  if (rv > 0)
+    goto done;
+  else if (rv < 0)
+    {
+      errno = -rv;
+      rv = -1;
+      goto done;
+    }
+
+  rv = libc_epoll_pwait (libc_epfd, events, maxevents, timeout, sigmask);
+  if (rv <= 0)
+    goto done;
+  for (int i = 0; i < rv; i++)
+    {
+      if (events[i].data.fd == ldpw->vcl_mq_epfd)
+	{
+	  /* We should remove mq epoll fd from events. */
+	  rv--;
+	  if (i != rv)
+	    {
+	      events[i].events = events[rv].events;
+	      events[i].data.u64 = events[rv].data.u64;
+	    }
+	  num_ev = vls_epoll_wait (ep_vlsh, &events[rv], maxevents - rv, 0);
+	  if (PREDICT_TRUE (num_ev > 0))
+	    rv += num_ev;
+	  break;
+	}
+    }
+
+done:
+  return rv;
+}
+
 int
 epoll_pwait (int epfd, struct epoll_event *events,
 	     int maxevents, int timeout, const sigset_t * sigmask)
 {
-  return ldp_epoll_pwait (epfd, events, maxevents, timeout, sigmask);
+  if (vls_use_eventfd ())
+    return ldp_epoll_pwait_eventfd (epfd, events, maxevents, timeout,
+				    sigmask);
+  else
+    return ldp_epoll_pwait (epfd, events, maxevents, timeout, sigmask);
 }
 
 int
 epoll_wait (int epfd, struct epoll_event *events, int maxevents, int timeout)
 {
-  return ldp_epoll_pwait (epfd, events, maxevents, timeout, NULL);
+  if (vls_use_eventfd ())
+    return ldp_epoll_pwait_eventfd (epfd, events, maxevents, timeout, NULL);
+  else
+    return ldp_epoll_pwait (epfd, events, maxevents, timeout, NULL);
 }
 
 int
diff --git a/src/vcl/vcl_locked.c b/src/vcl/vcl_locked.c
index ae9c9762dc1..19522c05d61 100644
--- a/src/vcl/vcl_locked.c
+++ b/src/vcl/vcl_locked.c
@@ -1381,6 +1381,12 @@ vls_app_create (char *app_name)
   return VPPCOM_OK;
 }
 
+unsigned char
+vls_use_eventfd (void)
+{
+  return vcm->cfg.use_mq_eventfd;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vcl/vcl_locked.h b/src/vcl/vcl_locked.h
index 2c8f2115230..18b341b11ca 100644
--- a/src/vcl/vcl_locked.h
+++ b/src/vcl/vcl_locked.h
@@ -50,6 +50,7 @@ vcl_session_handle_t vlsh_to_sh (vls_handle_t vlsh);
 vcl_session_handle_t vlsh_to_session_index (vls_handle_t vlsh);
 vls_handle_t vls_session_index_to_vlsh (uint32_t session_index);
 int vls_app_create (char *app_name);
+unsigned char vls_use_eventfd (void);
 
 #endif /* SRC_VCL_VCL_LOCKED_H_ */