linux: Normalize and return timeout on select (BZ #27651)

The commit 2433d39b697, which added time64 support to select, changed
the function to use __NR_pselect6 (or __NR_pelect6_time64) on all
architectures.  However, on architectures where the symbol was
implemented with __NR_select the kernel normalizes the passed timeout
instead of return EINVAL.  For instance, the input timeval
{ 0, 5000000 } is interpreted as { 5, 0 }.

And as indicated by BZ #27651, this semantic seems to be expected
and changing it results in some performance issues (most likely
the program does not check the return code and keeps issuing
select with unormalized tv_usec argument).

To avoid a different semantic depending whether which syscall the
architecture used to issue, select now always normalize the timeout
input.  This is a slight change for some ABIs (for instance aarch64).

Checked on x86_64-linux-gnu and i686-linux-gnu.

diff --git a/include/time.h b/include/time.h
index caf2af5e74ed6179e182e0ab7ce2012b2c421a10..e0636132a6ee3fe97d6bcfaa29b6e5ff48885d59 100644 (file)
--- a/include/time.h
+++ b/include/time.h
@@ -502,6 +502,11 @@ time_now (void)
   __clock_gettime (TIME_CLOCK_GETTIME_CLOCKID, &ts);
   return ts.tv_sec;
 }
+
+#define NSEC_PER_SEC    1000000000L  /* Nanoseconds per second.  */
+#define USEC_PER_SEC    1000000L     /* Microseconds per second.  */
+#define NSEC_PER_USEC   1000L        /* Nanoseconds per microsecond.  */
+
 #endif
 
 #endif

diff --git a/misc/tst-select.c b/misc/tst-select.c
index 5ad057cd51d0f45c3c95addff96b24abeff9953f..534105b50020fa44921cd40abf3c1c964f3a9f7b 100644 (file)
--- a/misc/tst-select.c
+++ b/misc/tst-select.c
@@ -19,6 +19,7 @@
 #include <errno.h>
 #include <support/capture_subprocess.h>
 #include <support/check.h>
+#include <support/support.h>
 #include <support/timespec.h>
 #include <support/xunistd.h>
 #include <support/xtime.h>
@@ -47,6 +48,12 @@ do_test_child (void *clousure)
   int r = select (args->fds[0][0] + 1, &rfds, NULL, NULL, &args->tmo);
   TEST_COMPARE (r, 0);
 
+  if (support_select_modifies_timeout ())
+    {
+      TEST_COMPARE (args->tmo.tv_sec, 0);
+      TEST_COMPARE (args->tmo.tv_usec, 0);
+    }
+
   TEST_TIMESPEC_NOW_OR_AFTER (CLOCK_REALTIME, ts);
 
   xwrite (args->fds[1][1], "foo", 3);
@@ -69,6 +76,16 @@ do_test (void)
                                      sc_allow_none);
   }
 
+  if (support_select_normalizes_timeout ())
+    {
+      /* This is handled as 1 second instead of failing with EINVAL.  */
+      args.tmo = (struct timeval) { .tv_sec = 0, .tv_usec = 1000000 };
+      struct support_capture_subprocess result;
+      result = support_capture_subprocess (do_test_child, &args);
+      support_capture_subprocess_check (&result, "tst-select-child", 0,
+                                       sc_allow_none);
+    }
+
   /* Same as before, but simulating polling.  */
   args.tmo = (struct timeval) { .tv_sec = 0, .tv_usec = 0 };
   {

diff --git a/sunrpc/svcauth_des.c b/sunrpc/svcauth_des.c
index 7607abc8186813f6c2e1795aecb07f7d2585e60e..25a85c90970d4cedd9da724180bfc11016d7fec1 100644 (file)
--- a/sunrpc/svcauth_des.c
+++ b/sunrpc/svcauth_des.c
@@ -58,7 +58,6 @@
 
 #define debug(msg)             /*printf("svcauth_des: %s\n", msg) */
 
-#define USEC_PER_SEC ((uint32_t) 1000000L)
 #define BEFORE(t1, t2) timercmp(t1, t2, <)
 
 /*

diff --git a/sysdeps/unix/sysv/linux/select.c b/sysdeps/unix/sysv/linux/select.c
index 415aa87d3c87d70c2ad8aeeb034f738a20420c7b..3b67ff44765ededf20320398ad54721574112c0b 100644 (file)
--- a/sysdeps/unix/sysv/linux/select.c
+++ b/sysdeps/unix/sysv/linux/select.c
@@ -33,12 +33,34 @@ int
 __select64 (int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
            struct __timeval64 *timeout)
 {
-  struct __timespec64 ts64, *pts64 = NULL;
-  if (timeout != NULL)
+  __time64_t s = timeout != NULL ? timeout->tv_sec : 0;
+  int32_t us = timeout != NULL ? timeout->tv_usec : 0;
+  int32_t ns;
+
+  if (s < 0 || us < 0)
+    return INLINE_SYSCALL_ERROR_RETURN_VALUE (EINVAL);
+
+  /* Normalize the timeout, as legacy Linux __NR_select and __NR__newselect.
+     Different than syscall, it also handle possible overflow.  */
+  if (us / USEC_PER_SEC > INT64_MAX - s)
     {
-      ts64 = timeval64_to_timespec64 (*timeout);
-      pts64 = &ts64;
+      s = INT64_MAX;
+      ns = NSEC_PER_SEC - 1;
     }
+  else
+    {
+      s += us / USEC_PER_SEC;
+      us = us % USEC_PER_SEC;
+      ns = us * NSEC_PER_USEC;
+    }
+
+  struct __timespec64 ts64, *pts64 = NULL;
+   if (timeout != NULL)
+     {
+       ts64.tv_sec = s;
+       ts64.tv_nsec = ns;
+       pts64 = &ts64;
+     }
 
 #ifndef __NR_pselect6_time64
 # define __NR_pselect6_time64 __NR_pselect6
@@ -52,10 +74,10 @@ __select64 (int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
          (though the pselect() glibc call suppresses this behavior).
          Since select() on Linux has the same behavior as the pselect6
          syscall, we update the timeout here.  */
-      if (r == 0 || errno != ENOSYS)
+      if (r >= 0 || errno != ENOSYS)
        {
          if (timeout != NULL)
-           TIMEVAL_TO_TIMESPEC (timeout, &ts64);
+           TIMESPEC_TO_TIMEVAL (timeout, &ts64);
          return r;
        }
 
@@ -64,14 +86,15 @@ __select64 (int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
 
 #ifndef __ASSUME_TIME64_SYSCALLS
   struct timespec ts32, *pts32 = NULL;
-  if (timeout != NULL)
+  if (pts64 != NULL)
     {
-      if (! in_time_t_range (timeout->tv_sec))
+      if (! in_time_t_range (pts64->tv_sec))
        {
          __set_errno (EINVAL);
          return -1;
        }
-      ts32 = valid_timespec64_to_timespec (ts64);
+      ts32.tv_sec = s;
+      ts32.tv_nsec = ns;
       pts32 = &ts32;
     }
 # ifndef __ASSUME_PSELECT