deps: reduce cpu profiler overhead

* Block SIGPROF when in the epoll_pwait() system call so the event loop
  doesn't keep waking up on signal delivery.  The clock_gettime() system
  call that libuv does after EINTR is very expensive on virtualized
  systems.

* Replace sched_yield() with nanosleep() in V8's tick event processor
  thread.  The former only yields the CPU when there is another process
  scheduled on the same CPU.

* Fix a bug in the epoll_pwait() system call wrapper in libuv,
  see libuv/libuv#4.

Refs strongloop/strong-agent#3 and strongloop-internal/scrum-cs#37.

Author: bnoordhuis

deps/uv/src/unix/linux-core.c

@@ -130,6 +130,7 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   struct uv__epoll_event e;
   ngx_queue_t* q;
   uv__io_t* w;
+  sigset_t set;
   uint64_t base;
   uint64_t diff;
   int nevents;
@@ -138,7 +139,6 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   int fd;
   int op;
   int i;
-  static int no_epoll_wait;
 
   if (loop->nfds == 0) {
     assert(ngx_queue_empty(&loop->watcher_queue));
@@ -184,23 +184,15 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   base = loop->time;
   count = 48; /* Benchmarks suggest this gives the best throughput. */
 
+  sigemptyset(&set);
+  sigaddset(&set, SIGPROF);
+
   for (;;) {
-    if (!no_epoll_wait) {
-      nfds = uv__epoll_wait(loop->backend_fd,
-                            events,
-                            ARRAY_SIZE(events),
-                            timeout);
-      if (nfds == -1 && errno == ENOSYS) {
-        no_epoll_wait = 1;
-        continue;
-      }
-    } else {
-      nfds = uv__epoll_pwait(loop->backend_fd,
-                             events,
-                             ARRAY_SIZE(events),
-                             timeout,
-                             NULL);
-    }
+    nfds = uv__epoll_pwait(loop->backend_fd,
+                           events,
+                           ARRAY_SIZE(events),
+                           timeout,
+                           &set);
 
     /* Update loop->time unconditionally. It's tempting to skip the update when
      * timeout == 0 (i.e. non-blocking poll) but there is no guarantee that the

deps/uv/src/unix/linux-syscalls.c

@@ -21,6 +21,7 @@
 
 #include "linux-syscalls.h"
 #include <unistd.h>
+#include <signal.h>
 #include <sys/syscall.h>
 #include <sys/types.h>
 #include <errno.h>
@@ -298,7 +299,7 @@ int uv__epoll_pwait(int epfd,
                  nevents,
                  timeout,
                  sigmask,
-                 sizeof(*sigmask));
+                 _NSIG / sizeof(long));
 #else
   return errno = ENOSYS, -1;
 #endif

deps/v8/src/platform-linux.cc

@@ -37,6 +37,7 @@
 #include <sys/syscall.h>
 #include <sys/types.h>
 #include <stdlib.h>
+#include <time.h>
 
 // Ubuntu Dapper requires memory pages to be marked as
 // executable. Otherwise, OS raises an exception when executing code
@@ -813,7 +814,8 @@ void Thread::SetThreadLocal(LocalStorageKey key, void* value) {
 
 
 void Thread::YieldCPU() {
-  sched_yield();
+  const timespec delay = { 0, 1 };
+  nanosleep(&delay, NULL);
 }