unix: use relaxed loads/stores for feature checks

Make ThreadSanitizer stop complaining about the static variables that
libuv uses to record the presence (or lack) of system calls and other
kernel features.

Fixes: https://github.com/libuv/libuv/issues/2884
PR-URL: https://github.com/libuv/libuv/pull/2886
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: Jameson Nash <vtjnash@gmail.com>

src/unix/core.c

@@ -216,15 +216,23 @@ int uv__getiovmax(void) {
 #if defined(IOV_MAX)
   return IOV_MAX;
 #elif defined(_SC_IOV_MAX)
-  static int iovmax = -1;
+  static int iovmax_cached = -1;
-  if (iovmax == -1) {
+  int iovmax;
-    iovmax = sysconf(_SC_IOV_MAX);
+
+  iovmax = uv__load_relaxed(&iovmax_cached);
+  if (iovmax != -1)
+    return iovmax;
+
   /* On some embedded devices (arm-linux-uclibc based ip camera),
    * sysconf(_SC_IOV_MAX) can not get the correct value. The return
    * value is -1 and the errno is EINPROGRESS. Degrade the value to 1.
    */
-    if (iovmax == -1) iovmax = 1;
+  iovmax = sysconf(_SC_IOV_MAX);
-  }
+  if (iovmax == -1)
+    iovmax = 1;
+
+  uv__store_relaxed(&iovmax_cached, iovmax);
+
   return iovmax;
 #else
   return 1024;
@@ -658,7 +666,7 @@ ssize_t uv__recvmsg(int fd, struct msghdr* msg, int flags) {
   int* end;
 #if defined(__linux__)
   static int no_msg_cmsg_cloexec;
-  if (no_msg_cmsg_cloexec == 0) {
+  if (0 == uv__load_relaxed(&no_msg_cmsg_cloexec)) {
     rc = recvmsg(fd, msg, flags | 0x40000000);  /* MSG_CMSG_CLOEXEC */
     if (rc != -1)
       return rc;
@@ -667,7 +675,7 @@ ssize_t uv__recvmsg(int fd, struct msghdr* msg, int flags) {
     rc = recvmsg(fd, msg, flags);
     if (rc == -1)
       return UV__ERR(errno);
-    no_msg_cmsg_cloexec = 1;
+    uv__store_relaxed(&no_msg_cmsg_cloexec, 1);
   } else {
     rc = recvmsg(fd, msg, flags);
   }

src/unix/fs.c

@@ -312,7 +312,7 @@ static int uv__fs_mkstemp(uv_fs_t* req) {
   uv_once(&once, uv__mkostemp_initonce);
 
 #ifdef O_CLOEXEC
-  if (no_cloexec_support == 0 && uv__mkostemp != NULL) {
+  if (uv__load_relaxed(&no_cloexec_support) == 0 && uv__mkostemp != NULL) {
     r = uv__mkostemp(path, O_CLOEXEC);
 
     if (r >= 0)
@@ -325,7 +325,7 @@ static int uv__fs_mkstemp(uv_fs_t* req) {
 
     /* We set the static variable so that next calls don't even
        try to use mkostemp. */
-    no_cloexec_support = 1;
+    uv__store_relaxed(&no_cloexec_support, 1);
   }
 #endif  /* O_CLOEXEC */
 
@@ -456,7 +456,7 @@ static ssize_t uv__fs_read(uv_fs_t* req) {
     result = preadv(req->file, (struct iovec*) req->bufs, req->nbufs, req->off);
 #else
 # if defined(__linux__)
-    if (no_preadv) retry:
+    if (uv__load_relaxed(&no_preadv)) retry:
 # endif
     {
       result = uv__fs_preadv(req->file, req->bufs, req->nbufs, req->off);
@@ -468,7 +468,7 @@ static ssize_t uv__fs_read(uv_fs_t* req) {
                           req->nbufs,
                           req->off);
       if (result == -1 && errno == ENOSYS) {
-        no_preadv = 1;
+        uv__store_relaxed(&no_preadv, 1);
         goto retry;
       }
     }
@@ -1351,7 +1351,7 @@ static int uv__fs_statx(int fd,
   int mode;
   int rc;
 
-  if (no_statx)
+  if (uv__load_relaxed(&no_statx))
     return UV_ENOSYS;
 
   dirfd = AT_FDCWD;
@@ -1384,7 +1384,7 @@ static int uv__fs_statx(int fd,
      * implemented, rc might return 1 with 0 set as the error code in which
      * case we return ENOSYS.
      */
-    no_statx = 1;
+    uv__store_relaxed(&no_statx, 1);
     return UV_ENOSYS;
   }
 

src/unix/kqueue.c

@@ -82,7 +82,7 @@ int uv__io_fork(uv_loop_t* loop) {
        process. So we sidestep the issue by pretending like we never
        started it in the first place.
     */
-    uv__has_forked_with_cfrunloop = 1;
+    uv__store_relaxed(&uv__has_forked_with_cfrunloop, 1);
     uv__free(loop->cf_state);
     loop->cf_state = NULL;
   }
@@ -487,7 +487,7 @@ int uv_fs_event_start(uv_fs_event_t* handle,
   if (!(statbuf.st_mode & S_IFDIR))
     goto fallback;
 
-  if (!uv__has_forked_with_cfrunloop) {
+  if (0 == uv__load_relaxed(&uv__has_forked_with_cfrunloop)) {
     int r;
     /* The fallback fd is no longer needed */
     uv__close_nocheckstdio(fd);
@@ -522,7 +522,8 @@ int uv_fs_event_stop(uv_fs_event_t* handle) {
   uv__handle_stop(handle);
 
 #if defined(__APPLE__) && MAC_OS_X_VERSION_MAX_ALLOWED >= 1070
-  if (!uv__has_forked_with_cfrunloop && handle->cf_cb != NULL)
+  if (0 == uv__load_relaxed(&uv__has_forked_with_cfrunloop))
+    if (handle->cf_cb != NULL)
       r = uv__fsevents_close(handle);
 #endif
 

src/unix/linux-core.c

@@ -198,8 +198,10 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
    * that being the largest value I have seen in the wild (and only once.)
    */
   static const int max_safe_timeout = 1789569;
-  static int no_epoll_pwait;
+  static int no_epoll_pwait_cached;
-  static int no_epoll_wait;
+  static int no_epoll_wait_cached;
+  int no_epoll_pwait;
+  int no_epoll_wait;
   struct epoll_event events[1024];
   struct epoll_event* pe;
   struct epoll_event e;
@@ -271,6 +273,15 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
   count = 48; /* Benchmarks suggest this gives the best throughput. */
   real_timeout = timeout;
 
+  /* You could argue there is a dependency between these two but
+   * ultimately we don't care about their ordering with respect
+   * to one another. Worst case, we make a few system calls that
+   * could have been avoided because another thread already knows
+   * they fail with ENOSYS. Hardly the end of the world.
+   */
+  no_epoll_pwait = uv__load_relaxed(&no_epoll_pwait_cached);
+  no_epoll_wait = uv__load_relaxed(&no_epoll_wait_cached);
+
   for (;;) {
     /* See the comment for max_safe_timeout for an explanation of why
      * this is necessary.  Executive summary: kernel bug workaround.
@@ -288,16 +299,20 @@ void uv__io_poll(uv_loop_t* loop, int timeout) {
                          ARRAY_SIZE(events),
                          timeout,
                          &sigset);
-      if (nfds == -1 && errno == ENOSYS)
+      if (nfds == -1 && errno == ENOSYS) {
+        uv__store_relaxed(&no_epoll_pwait_cached, 1);
         no_epoll_pwait = 1;
+      }
     } else {
       nfds = epoll_wait(loop->backend_fd,
                         events,
                         ARRAY_SIZE(events),
                         timeout);
-      if (nfds == -1 && errno == ENOSYS)
+      if (nfds == -1 && errno == ENOSYS) {
+        uv__store_relaxed(&no_epoll_wait_cached, 1);
         no_epoll_wait = 1;
       }
+    }
 
     if (sigmask != 0 && no_epoll_pwait != 0)
       if (pthread_sigmask(SIG_UNBLOCK, &sigset, NULL))

src/unix/os390-syscalls.c

@@ -33,7 +33,6 @@
 #pragma linkage(BPX4CTW, OS)
 #pragma linkage(BPX1CTW, OS)
 
-static int number_of_epolls;
 static QUEUE global_epoll_queue;
 static uv_mutex_t global_epoll_lock;
 static uv_once_t once = UV_ONCE_INIT;

src/unix/pthread-fixes.c

@@ -30,6 +30,8 @@
 */
 
 /* Android versions < 4.1 have a broken pthread_sigmask. */
+#include "uv-common.h"
+
 #include <errno.h>
 #include <pthread.h>
 #include <signal.h>
@@ -38,13 +40,13 @@ int uv__pthread_sigmask(int how, const sigset_t* set, sigset_t* oset) {
   static int workaround;
   int err;
 
-  if (workaround) {
+  if (uv__load_relaxed(&workaround)) {
     return sigprocmask(how, set, oset);
   } else {
     err = pthread_sigmask(how, set, oset);
     if (err) {
       if (err == EINVAL && sigprocmask(how, set, oset) == 0) {
-        workaround = 1;
+        uv__store_relaxed(&workaround, 1);
         return 0;
       } else {
         return -1;

src/unix/tcp.c

@@ -326,16 +326,19 @@ int uv_tcp_close_reset(uv_tcp_t* handle, uv_close_cb close_cb) {
 
 
 int uv_tcp_listen(uv_tcp_t* tcp, int backlog, uv_connection_cb cb) {
-  static int single_accept = -1;
+  static int single_accept_cached = -1;
   unsigned long flags;
+  int single_accept;
   int err;
 
   if (tcp->delayed_error)
     return tcp->delayed_error;
 
+  single_accept = uv__load_relaxed(&single_accept_cached);
   if (single_accept == -1) {
     const char* val = getenv("UV_TCP_SINGLE_ACCEPT");
     single_accept = (val != NULL && atoi(val) != 0);  /* Off by default. */
+    uv__store_relaxed(&single_accept_cached, single_accept);
   }
 
   if (single_accept)

src/uv-common.c

@@ -859,11 +859,11 @@ __attribute__((destructor))
 void uv_library_shutdown(void) {
   static int was_shutdown;
 
-  if (was_shutdown)
+  if (uv__load_relaxed(&was_shutdown))
     return;
 
   uv__process_title_cleanup();
   uv__signal_cleanup();
   uv__threadpool_cleanup();
-  was_shutdown = 1;
+  uv__store_relaxed(&was_shutdown, 1);
 }

src/uv-common.h

@@ -60,6 +60,14 @@ extern int snprintf(char*, size_t, const char*, ...);
 #define STATIC_ASSERT(expr)                                                   \
   void uv__static_assert(int static_assert_failed[1 - 2 * !(expr)])
 
+#ifdef __GNUC__
+#define uv__load_relaxed(p) __atomic_load_n(p, __ATOMIC_RELAXED)
+#define uv__store_relaxed(p, v) __atomic_store_n(p, v, __ATOMIC_RELAXED)
+#else
+#define uv__load_relaxed(p) (*p)
+#define uv__store_relaxed(p, v) do *p = v; while (0)
+#endif
+
 /* Handle flags. Some flags are specific to Windows or UNIX. */
 enum {
   /* Used by all handles. */