https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128875
Bug ID: 128875
Summary: REWIND that fails to seek leaves the unit locked: next
I/O fails with "Recursive I/O not allowed" or hangs
Product: gcc
Version: 16.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libfortran
Assignee: unassigned at gcc dot gnu.org
Reporter: federico.perini at gmail dot com
Target Milestone: ---
Created attachment 66222
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66222&action=edit
Patch
Reproducer:
program p
implicit none
integer :: ios
rewind (6, iostat=ios)
print *, 'after REWIND, iostat =', ios
end program p
$ gfortran p.f90 && ./a.out | cat
At line 5 of file p.f90 (unit = 6, file = 'stdout')
Fortran runtime error: Recursive I/O not allowed
Also check on Compiler Explorer: https://godbolt.org/z/TrWrYjWEM
Expected: "after REWIND, iostat = 29" (ESPIPE), as with the unit
unlocked. With "backspace (6, iostat=ios)" in place of the PRINT the
program hangs. With standard output redirected to a regular file, the
REWIND succeeds and the program prints the line.
Seen with GCC 16.2.0 (Homebrew, aarch64-apple-darwin) and with
master (75dfd8b97da1, 2026-11-18) on x86_64-pc-linux-gnu. With GCC
15.3.0 (x86_64-pc-linux-gnu) the PRINT hangs instead: the "Recursive
I/O not allowed" check comes from PR 119136, which is in GCC 16 and
reached the gcc-15 branch only after 15.3.
When the seek in st_rewind (libgfortran/io/file_pos.c) fails, st_rewind
calls generate_error and library_end and returns without unlock_unit (u):
if (sseek (u->s, 0, SEEK_SET) < 0)
{
generate_error (&fpp->common, LIBERROR_OS, NULL);
library_end ();
return;
}
Without IOSTAT= or ERR= the program stops there, so nothing else
happens. With IOSTAT= or ERR= it goes on, and the unit stays locked:
- the next data transfer statement on the unit fails with
"Fortran runtime error: Recursive I/O not allowed", because
check_for_recursive finds the unit held by the same thread;
- the next file positioning statement on the unit (REWIND, BACKSPACE,
ENDFILE) waits forever for the lock in find_unit.
The seek fails on every unit that is neither buffered nor seekable.
fd_to_stream buffers only regular files, so this includes a
preconnected unit on a pipe or a terminal.
Also in st_rewind, needs_unlock starts as true, where st_backspace,
st_endfile and st_flush start it as false. On a direct-access unit
with asynchronous I/O, the "Cannot REWIND a file opened for DIRECT
access" path therefore unlocks u->au->io_lock without having taken it.
Found while testing the fix for PR 128486: with that fix, the
preconnected units of a Windows process started without standard
handles fail REWIND with EBADF and reach this path. The bug itself
does not depend on that fix, or on Windows: it shows up whenever a
REWIND cannot seek, as in the reproducer above, which fails with
GCC 16.2.0 without either patch.
A patch follows.
Thank you,
Federico
Assisted by: Claude Opus 5.5