aboutsummaryrefslogtreecommitdiff
path: root/butl/fdstream
diff options
context:
space:
mode:
Diffstat (limited to 'butl/fdstream')
-rw-r--r--butl/fdstream20
1 files changed, 20 insertions, 0 deletions
diff --git a/butl/fdstream b/butl/fdstream
index 8d01b2f..8d6385f 100644
--- a/butl/fdstream
+++ b/butl/fdstream
@@ -486,6 +486,9 @@ namespace butl
// process' umask, so effective permissions are permissions & ~umask. On
// Windows permissions other than ru and wu are unlikelly to have effect.
//
+ // Also note that on POSIX the FD_CLOEXEC flag is set for the file descriptor
+ // to prevent its leakage into child processes.
+ //
LIBBUTL_EXPORT auto_fd
fdopen (const char*,
fdopen_mode,
@@ -510,6 +513,13 @@ namespace butl
// Duplicate an open file descriptor. Throw ios::failure on the underlying
// OS error.
//
+ // Note that on POSIX the FD_CLOEXEC flag is set for the new descriptor if it
+ // is present for the source one. That's in contrast to POSIX dup() that
+ // doesn't copy file descriptor flags. Also note that duplicating descriptor
+ // and setting the flag is not an atomic operation.
+ //
+ // @@ Should we copy HANDLE_FLAG_INHERIT flag on Windows as well?
+ //
LIBBUTL_EXPORT auto_fd
fddup (int fd);
@@ -563,6 +573,9 @@ namespace butl
// closed. One difference, however, would be if you were to both write to
// and read from the descriptor.
//
+ // Note that on POSIX the FD_CLOEXEC flag is set for the file descriptor to
+ // prevent its leakage into child processes.
+ //
// @@ You may have noticed that this interface doesn't match fdopen() (it
// does not throw and return int instead of auto_fd). The reason for this
// is process and its need to translate everything to process_error).
@@ -591,6 +604,13 @@ namespace butl
// In particular, the process class that uses fdpipe underneath makes the
// appropriate end of pipe (the one being passed to the child) inheritable.
//
+ // Note that on POSIX the FD_CLOEXEC flag is set for both ends, so they get
+ // automatically closed by the child process to prevent undesired behaviors
+ // (such as child deadlock on read from a pipe due to the write-end leakage
+ // into the child process). Opening pipe and setting the flag is not an
+ // atomic operation. Also note that you don't need to reset the flag for a
+ // pipe end being passed to the process class ctor.
+ //
LIBBUTL_EXPORT fdpipe
fdopen_pipe (fdopen_mode = fdopen_mode::none);
}