Hello world, the attached patch fixes the problem reported in PR 66756: When opeing a file, the main lock for all units was acquired, the unit lock was acquired, and then the main lock was released and re-aquired. To the thread sanitizer, this is a lock-order inversion.
One option would have been to simply close the bug, because this only occurs in opening a file, when the gfc_unit has not yet had a chance to escape to another thread. However, it appears that this causes trouble debugging parallel applications, hence this patch. What this patch does is to change the assumptions for insert_unit: Previously, this used to lock the newly created unit, and the caller had to unlock. Now, gfc_get_unit can do the locking after releasing the global lock. This gets rid of the thread sanitizer issue; the thread sanitizer output is clean. However, I would appreciate feedback about whether this approach (and my code) is correct. Regression-tested. Comments? Suggestions for improvements/other approaches? Close the PR as WONTFIX instead? OK for trunk? Regards Thomas 2017-09-28 Thomas Koenig <tkoe...@gcc.gnu.org> PR fortran/66756 * io/fbuf.c (fbuf_destroy): Lock unit before freeing the buffer. * io/unit.c (insert_unit): Do not create lock and lock, move to (gfc_get_unit): here; lock after insert_unit has succeded. (init_units): Do not unlock unit locks for stdin, stdout and stderr.
Index: io/fbuf.c =================================================================== --- io/fbuf.c (Revision 253162) +++ io/fbuf.c (Arbeitskopie) @@ -50,9 +50,11 @@ fbuf_destroy (gfc_unit *u) { if (u->fbuf == NULL) return; + __gthread_mutex_lock (&u->lock); free (u->fbuf->buf); free (u->fbuf); u->fbuf = NULL; + __gthread_mutex_unlock (&u->lock); } Index: io/unit.c =================================================================== --- io/unit.c (Revision 253162) +++ io/unit.c (Arbeitskopie) @@ -221,23 +221,14 @@ insert (gfc_unit *new, gfc_unit *t) return t; } +/* insert_unit()-- Create a new node, insert it into the treap. It is assumed + that the caller holds unit_lock. */ -/* insert_unit()-- Create a new node, insert it into the treap. */ - static gfc_unit * insert_unit (int n) { gfc_unit *u = xcalloc (1, sizeof (gfc_unit)); u->unit_number = n; -#ifdef __GTHREAD_MUTEX_INIT - { - __gthread_mutex_t tmp = __GTHREAD_MUTEX_INIT; - u->lock = tmp; - } -#else - __GTHREAD_MUTEX_INIT_FUNCTION (&u->lock); -#endif - __gthread_mutex_lock (&u->lock); u->priority = pseudo_random (); unit_root = insert (u, unit_root); return u; @@ -361,9 +352,20 @@ retry: if (created) { - /* Newly created units have their lock held already - from insert_unit. Just unlock UNIT_LOCK and return. */ +#ifdef __GTHREAD_MUTEX_INIT + { + __gthread_mutex_t tmp = __GTHREAD_MUTEX_INIT; + p->lock = tmp; + } +#else + __GTHREAD_MUTEX_INIT_FUNCTION (&p->lock); +#endif __gthread_mutex_unlock (&unit_lock); + + /* Nobody outside this address has seen this unit yet. We could safely + keep it unlocked until now. */ + + __gthread_mutex_lock (&p->lock); return p; } @@ -618,8 +620,6 @@ init_units (void) u->filename = strdup (stdin_name); fbuf_init (u, 0); - - __gthread_mutex_unlock (&u->lock); } if (options.stdout_unit >= 0) @@ -649,8 +649,6 @@ init_units (void) u->filename = strdup (stdout_name); fbuf_init (u, 0); - - __gthread_mutex_unlock (&u->lock); } if (options.stderr_unit >= 0) @@ -680,8 +678,6 @@ init_units (void) fbuf_init (u, 256); /* 256 bytes should be enough, probably not doing any kind of exotic formatting to stderr. */ - - __gthread_mutex_unlock (&u->lock); } /* Calculate the maximum file offset in a portable manner.