unix: add flag for blocking SIGPROF during poll

Add a per-event loop flag for blocking SIGPROF signals when polling for
events.

The motivation for this addition is to reduce the number of wakeups and
subsequent clock_gettime() system calls when using a sampling profiler.

On Linux, this switches from epoll_wait() to epoll_pwait() when enabled.
Other platforms bracket the poll syscall with pthread_sigmask() calls.

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

PR-URL: https://github.com/libuv/libuv/pull/15
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>

src/unix/internal.h

@@ -111,6 +111,11 @@ enum {
   UV_TCP_SINGLE_ACCEPT = 0x400  /* Only accept() when idle. */
 };
 
+/* loop flags */
+enum {
+  UV_LOOP_BLOCK_SIGPROF = 1
+};
+
 /* core */
 int uv__nonblock(int fd, int set);
 int uv__cloexec(int fd, int set);

src/unix/kqueue.c

@@ -56,9 +56,11 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   unsigned int nevents;
   unsigned int revents;
   ngx_queue_t* q;
+  uv__io_t* w;
+  sigset_t* pset;
+  sigset_t set;
   uint64_t base;
   uint64_t diff;
-  uv__io_t* w;
   int filter;
   int fflags;
   int count;
@@ -118,6 +120,13 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
     w->events = w->pevents;
   }
 
+  pset = NULL;
+  if (loop->flags & UV_LOOP_BLOCK_SIGPROF) {
+    pset = &set;
+    sigemptyset(pset);
+    sigaddset(pset, SIGPROF);
+  }
+
   assert(timeout >= -1);
   base = loop->time;
   count = 48; /* Benchmarks suggest this gives the best throughput. */
@@ -128,6 +137,9 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
       spec.tv_nsec = (timeout % 1000) * 1000000;
     }
 
+    if (pset != NULL)
+      pthread_sigmask(SIG_BLOCK, pset, NULL);
+
     nfds = kevent(loop->backend_fd,
                   events,
                   nevents,
@@ -135,6 +147,9 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
                   ARRAY_SIZE(events),
                   timeout == -1 ? NULL : &spec);
 
+    if (pset != NULL)
+      pthread_sigmask(SIG_UNBLOCK, pset, NULL);
+
     /* 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
      * operating system didn't reschedule our process while in the syscall.

src/unix/linux-core.c

@@ -33,6 +33,7 @@
 #include <sys/prctl.h>
 #include <sys/sysinfo.h>
 #include <unistd.h>
+#include <signal.h>
 #include <fcntl.h>
 #include <time.h>
 
@@ -130,6 +131,8 @@ 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* pset;
+  sigset_t set;
   uint64_t base;
   uint64_t diff;
   int nevents;
@@ -180,12 +183,25 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
     w->events = w->pevents;
   }
 
+  pset = NULL;
+  if (loop->flags & UV_LOOP_BLOCK_SIGPROF) {
+    pset = &set;
+    sigemptyset(pset);
+    sigaddset(pset, SIGPROF);
+  }
+
   assert(timeout >= -1);
   base = loop->time;
   count = 48; /* Benchmarks suggest this gives the best throughput. */
 
   for (;;) {
-    if (!no_epoll_wait) {
+    if (no_epoll_wait || pset != NULL) {
+      nfds = uv__epoll_pwait(loop->backend_fd,
+                             events,
+                             ARRAY_SIZE(events),
+                             timeout,
+                             pset);
+    } else {
       nfds = uv__epoll_wait(loop->backend_fd,
                             events,
                             ARRAY_SIZE(events),
@@ -194,12 +210,6 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
         no_epoll_wait = 1;
         continue;
       }
-    } else {
-      nfds = uv__epoll_pwait(loop->backend_fd,
-                             events,
-                             ARRAY_SIZE(events),
-                             timeout,
-                             NULL);
     }
 
     /* Update loop->time unconditionally. It's tempting to skip the update when

src/unix/sunos.c

@@ -112,6 +112,8 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   struct timespec spec;
   ngx_queue_t* q;
   uv__io_t* w;
+  sigset_t* pset;
+  sigset_t set;
   uint64_t base;
   uint64_t diff;
   unsigned int nfds;
@@ -119,6 +121,7 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   int saved_errno;
   int nevents;
   int count;
+  int err;
   int fd;
 
   if (loop->nfds == 0) {
@@ -140,6 +143,13 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
     w->events = w->pevents;
   }
 
+  pset = NULL;
+  if (loop->flags & UV_LOOP_BLOCK_SIGPROF) {
+    pset = &set;
+    sigemptyset(pset);
+    sigaddset(pset, SIGPROF);
+  }
+
   assert(timeout >= -1);
   base = loop->time;
   count = 48; /* Benchmarks suggest this gives the best throughput. */
@@ -155,11 +165,20 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
 
     nfds = 1;
     saved_errno = 0;
-    if (port_getn(loop->backend_fd,
+
-                  events,
+    if (pset != NULL)
-                  ARRAY_SIZE(events),
+      pthread_sigmask(SIG_BLOCK, pset, NULL);
-                  &nfds,
+
-                  timeout == -1 ? NULL : &spec)) {
+    err = port_getn(loop->backend_fd,
+                    events,
+                    ARRAY_SIZE(events),
+                    &nfds,
+                    timeout == -1 ? NULL : &spec);
+
+    if (pset != NULL)
+      pthread_sigmask(SIG_UNBLOCK, pset, NULL);
+
+    if (err) {
       /* Work around another kernel bug: port_getn() may return events even
        * on error.
        */