bind(2) fails on 13.0-STABLE when sin_family is 0

[Bakul Shah]

ttcp runs fine on 13.0-RELEASE but fails on -stable.

The culprit seems to be bind(2). Running ttcp under gdb:

$ gdb a.out
Reading symbols from a.out...
(gdb) b 295
Breakpoint 1 at 0x203127: file ttcp.c, line 295.
(gdb) run -s -r
Starting program: /usr/ports/benchmarks/ttcp/work/ttcp-1.12_2/a.out -s -r
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket

Breakpoint 1, main (argc=3, argv=0x7fffffffd9b0) at ttcp.c:295
295 if (bind(fd, (struct sockaddr *) &sinme, sizeof(sinme)) < 0)
(gdb) p/x sinme
$1 = {sin_len = 0x0, sin_family = 0x0, sin_port = 0x8913, sin_addr = {
s_addr = 0x0}, sin_zero = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}}
(gdb) n
296 err("bind");
(gdb) p errno
$2 = 47


$ errno 47
#define EAFNOSUPPORT 47 /* Address family not supported by protocol family */

Did something change post 13.0-RELEASE that requires specifying sin_family?
Thanks!

-- Bakul

[Mark Johnston]

Post by Bakul Shah
ttcp runs fine on 13.0-RELEASE but fails on -stable.
$ gdb a.out
Reading symbols from a.out...
(gdb) b 295
Breakpoint 1 at 0x203127: file ttcp.c, line 295.
(gdb) run -s -r
Starting program: /usr/ports/benchmarks/ttcp/work/ttcp-1.12_2/a.out -s -r
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
Breakpoint 1, main (argc=3, argv=0x7fffffffd9b0) at ttcp.c:295
295 if (bind(fd, (struct sockaddr *) &sinme, sizeof(sinme)) < 0)
(gdb) p/x sinme
$1 = {sin_len = 0x0, sin_family = 0x0, sin_port = 0x8913, sin_addr = {
s_addr = 0x0}, sin_zero = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}}
(gdb) n
296 err("bind");
(gdb) p errno
$2 = 47
$ errno 47
#define EAFNOSUPPORT 47 /* Address family not supported by protocol family */
Did something change post 13.0-RELEASE that requires specifying sin_family?
Thanks!

Yes, some changes were made recently to make sockaddr validation
stricter. Several other operating systems also have this requirement.
Linux seems to be a bit more relaxed in that AF_UNSPEC (0) is permitted
if and only if the bind address is INADDR_ANY, which is the case here.

Since 2001 the benchmarks/ttcp port has carried a patch to specify
sin_family. Is there some reason it cannot be used here? I don't
object to re-allowing ttcp's unpatched behaviour if necessary.

[Bakul Shah]

Post by Mark Johnston

Post by Bakul Shah
ttcp runs fine on 13.0-RELEASE but fails on -stable.
$ gdb a.out
Reading symbols from a.out...
(gdb) b 295
Breakpoint 1 at 0x203127: file ttcp.c, line 295.
(gdb) run -s -r
Starting program: /usr/ports/benchmarks/ttcp/work/ttcp-1.12_2/a.out -s -r
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
Breakpoint 1, main (argc=3, argv=0x7fffffffd9b0) at ttcp.c:295
295 if (bind(fd, (struct sockaddr *) &sinme, sizeof(sinme)) < 0)
(gdb) p/x sinme
$1 = {sin_len = 0x0, sin_family = 0x0, sin_port = 0x8913, sin_addr = {
s_addr = 0x0}, sin_zero = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}}
(gdb) n
296 err("bind");
(gdb) p errno
$2 = 47
$ errno 47
#define EAFNOSUPPORT 47 /* Address family not supported by protocol family */
Did something change post 13.0-RELEASE that requires specifying sin_family?
Thanks!

Yes, some changes were made recently to make sockaddr validation
stricter. Several other operating systems also have this requirement.
Linux seems to be a bit more relaxed in that AF_UNSPEC (0) is permitted
if and only if the bind address is INADDR_ANY, which is the case here.
Since 2001 the benchmarks/ttcp port has carried a patch to specify
sin_family. Is there some reason it cannot be used here? I don't
object to re-allowing ttcp's unpatched behaviour if necessary.

The patch was only for the transmit case. Making it unconditional fixed it.
I asked mainly because I was surprised and wanted to make sure this was an
intended change.

Thanks for your response!

-- Bakul

[Mark Johnston]

Post by Bakul Shah

Post by Mark Johnston

Post by Bakul Shah
ttcp runs fine on 13.0-RELEASE but fails on -stable.
$ gdb a.out
Reading symbols from a.out...
(gdb) b 295
Breakpoint 1 at 0x203127: file ttcp.c, line 295.
(gdb) run -s -r
Starting program: /usr/ports/benchmarks/ttcp/work/ttcp-1.12_2/a.out -s -r
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
Breakpoint 1, main (argc=3, argv=0x7fffffffd9b0) at ttcp.c:295
295 if (bind(fd, (struct sockaddr *) &sinme, sizeof(sinme)) < 0)
(gdb) p/x sinme
$1 = {sin_len = 0x0, sin_family = 0x0, sin_port = 0x8913, sin_addr = {
s_addr = 0x0}, sin_zero = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}}
(gdb) n
296 err("bind");
(gdb) p errno
$2 = 47
$ errno 47
#define EAFNOSUPPORT 47 /* Address family not supported by protocol family */
Did something change post 13.0-RELEASE that requires specifying sin_family?
Thanks!

Yes, some changes were made recently to make sockaddr validation
stricter. Several other operating systems also have this requirement.
Linux seems to be a bit more relaxed in that AF_UNSPEC (0) is permitted
if and only if the bind address is INADDR_ANY, which is the case here.
Since 2001 the benchmarks/ttcp port has carried a patch to specify
sin_family. Is there some reason it cannot be used here? I don't
object to re-allowing ttcp's unpatched behaviour if necessary.

The patch was only for the transmit case. Making it unconditional fixed it.
I asked mainly because I was surprised and wanted to make sure this was an
intended change.
Thanks for your response!

I see, thanks. In that case it is probably best to restore some level
of compatibility: https://reviews.freebsd.org/D30539