homepage

To test bpo-32375, I just recompiled the master branch of Python twice using these flags:
* -D_FORTIFY_SOURCE=2 -Og
* -D_FORTIFY_SOURCE=2 -O3
I got a few warnings, the same that I get without FORTIFY SOURCE.
Using -D_FORTIFY_SOURCE=2 -O3, I get one warning, but it's no longer from getpath.c:
In file included from /usr/include/string.h:494,
 from ./Include/Python.h:30,
 from /home/vstinner/python/master/Modules/socketmodule.c:103:
In function ‘memset’,
 inlined from ‘getsockaddrarg’ at /home/vstinner/python/master/Modules/socketmodule.c:2331:9,
 inlined from ‘sock_bind’ at /home/vstinner/python/master/Modules/socketmodule.c:3113:10:
/usr/include/bits/string_fortified.h:71:10: warning: ‘__builtin_memset’ offset [17, 88] from the object at ‘addrbuf’ is out of the bounds of referenced subobject ‘sa’ with type ‘struct sockaddr’ at offset 0 [-Warray-bounds]
 71 | return __builtin___memset_chk (__dest, __ch, __len, __bos0 (__dest));
 | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The warning comes from the code:
 case AF_ALG:
 ...
 struct sockaddr_alg *sa;
 sa = (struct sockaddr_alg *)addr_ret;
 memset(sa, 0, sizeof(*sa));
called from:
static PyObject *
sock_bind(PySocketSockObject *s, PyObject *addro)
{
 sock_addr_t addrbuf;
 int addrlen;
 int res;
 if (!getsockaddrarg(s, addro, SAS2SA(&addrbuf), &addrlen, "bind")) {
 return NULL;
 }

Similar when building Python on Fedora 30 (GCC version 9.2.1) with gcc -O3:
In function ‘getsockaddrarg’,
 inlined from ‘sock_bind’ at /home/vstinner/python/master/Modules/socketmodule.c:3113:10:
/home/vstinner/python/master/Modules/socketmodule.c:2331:9: warning: ‘memset’ offset [17, 88] from the object at ‘addrbuf’ is out of the bounds of referenced subobject ‘sa’ with type ‘struct sockaddr’ at offset 0 [-Warray-bounds]
 2331 | memset(sa, 0, sizeof(*sa));
 | ^~~~~~~~~~~~~~~~~~~~~~~~~~

New changeset d565fb9828ee9c494bb7a80057a08e4738273e30 by Victor Stinner in branch 'master':
bpo-38282: Rewrite getsockaddrarg() helper function (GH-16698)
https://github.com/python/cpython/commit/d565fb9828ee9c494bb7a80057a08e4738273e30

The failures on FreeBSD are:
/usr/home/buildbot/python/3.x.koobs-freebsd10.nondebug/build/Modules/socketmodule.c:1931:50: error: no member named 'bt_l2' in 'union sock_addr'
 struct sockaddr_l2 *addr = &addrbuf->bt_l2;
 ~~~~~~~ ^
/usr/home/buildbot/python/3.x.koobs-freebsd10.nondebug/build/Modules/socketmodule.c:1950:50: error: no member named 'bt_rc' in 'union sock_addr'
 struct sockaddr_rc *addr = &addrbuf->bt_rc;
 ~~~~~~~ ^
/usr/home/buildbot/python/3.x.koobs-freebsd10.nondebug/build/Modules/socketmodule.c:1967:51: error: no member named 'bt_hci' in 'union sock_addr'
 struct sockaddr_hci *addr = &addrbuf->bt_hci;
 ~~~~~~~ ^
3 errors generated.
This is likely due to the fact that those fields are protected by:
typedef union sock_addr {
#ifdef HAVE_BLUETOOTH_BLUETOOTH_H
 struct sockaddr_l2 bt_l2;
 struct sockaddr_rc bt_rc;
 struct sockaddr_sco bt_sco;
 struct sockaddr_hci bt_hci;
#elif defined(MS_WINDOWS)
}
and in those machines HAVE_BLUETOOTH_BLUETOOTH_H is absent.

PR 16738 fixes the FreeBSD buildbots by correctly handling the Bluetooth L2CAP socket structure in FreeBSD if only bluetooth.h is available. Check out https://man.cx/ng_btsocket(4) for example for more info in the ng_btsocket protocol for FreeBSD.

New changeset 27b33fb41a7c64a6211d73d14804aa0cd6defccb by Pablo Galindo in branch 'master':
bpo-38282: Correctly manage the Bluetooth L2CAP socket structure in FreeBSD (GH-16738)
https://github.com/python/cpython/commit/27b33fb41a7c64a6211d73d14804aa0cd6defccb

I close the issue.
I planned to backport the change to 3.8, but 3.8.0 final is supposed to be released today. I prefer to avoid putting a last minute regression on some platforms like DragonflyBSD in 3.8. Backporting the change is not strictly required, it's just a compiler warning, it's a false alarm, and it's only when building with _FORTIFY_SOURCE defined.