Hi,

I noticed a change in the value returned by the getsockopt for the 
TCP_DEFER_ACCEPT option with a 2.6.32 kernel. The value retrieved with 
the getsockopt is different from the one specified with the setsockopt. 
Is it an expected behaviour ?

I saw there were  changes around the TCP_DEFER_ACCEPT option with the 
number of attempts converted to a number of seconds.

The following program is working fine with a 2.6.31 but fails with a 
2.6.32 kernel.

Thanks
  -- Daniel

#include <stdio.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <netinet/tcp.h>

int main(int argc, char *argv[])
{
    int val1 = 12, val2;
    socklen_t len = sizeof(val2);
    int fd;
   
    fd = socket(PF_INET, SOCK_STREAM, 0);
    if (fd < 0) {
        perror("socket");
        return -1;
    }
   
    if (setsockopt(fd, SOL_TCP, TCP_DEFER_ACCEPT, &val1, sizeof(val1))) {
        perror("setsockopt");
        return -1;
    }
   
    if (getsockopt(fd, SOL_TCP, TCP_DEFER_ACCEPT, &val2, &len)) {
        perror("getsockopt");
        return -1;
    }

    if (val1 != val2) {
        fprintf(stderr, "error %d != %d\n", val1, val2);
        return -1;
    }

    return 0;
}

--

[ message continues ]

Added Cc.

I don't think this change was intentional. ...However, is this difference 
particularly significant besides failing such a test program? The actual 
value now returned by the getsockopt is more accurate than what the
userspace initially provided.

In general, I wonder if there's something that mandates that a set/get 
pair of value should be equal?

-- 
 i.
--

[ message continues ]

Nothing... really... we can round the value, and we indeed round it in 2.6.32

defer value is given in second by user, and converted to number of retransmits by kernel.

Program assumption is wrong.

--

[ message continues ]

BTW, previous kernels were rounding too :

1 -> 3
2 -> 3
3 -> 3
4 -> 6
5 -> 6
6 -> 6
7 -> 12
8 -> 12
9 -> 12
10 -> 12
11 -> 12
12 -> 12
13 -> 24
14 -> 24
15 -> 24
16 -> 24
17 -> 24
18 -> 24
19 -> 24

New kernel (or other sysctl settings, I didnot check) :

1 -> 3
2 -> 3
3 -> 3
4 -> 9
5 -> 9
6 -> 9
7 -> 9
8 -> 9
9 -> 9
10 -> 21
11 -> 21
12 -> 21
13 -> 21
14 -> 21
15 -> 21
16 -> 21
17 -> 21
18 -> 21
19 -> 21

#include <stdio.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <netinet/tcp.h>

int main(int argc, char *argv[])
{
   int i, val1 = 12, val2;
   socklen_t len = sizeof(val2);
   int fd;
     fd = socket(PF_INET, SOCK_STREAM, 0);
   if (fd < 0) {
       perror("socket");
       return -1;
   }
   for (i = 1; i < 20 ; i++) {
     val1 = i;
     if (setsockopt(fd, SOL_TCP, TCP_DEFER_ACCEPT, &val1, sizeof(val1))) {
       perror("setsockopt");
       return -1;
     }
     if (getsockopt(fd, SOL_TCP, TCP_DEFER_ACCEPT, &val2, &len)) {
       perror("getsockopt");
       return -1;
     }
   printf("%d -> %d\n", i, val2);
   }
   return 0;
}

--

[ message continues ]

It's not problem if the set / get values are not same, but I was asking 
because I am working with a test suite checking if a checkpoint / 
restart solution is correct.  One of these tests, sets the 
TCP_DEFER_ACCEPT value to 12, checkpoints / restarts, and reads the 
value in order to check if it was correctly restored. The value 12 was 
chosen because it is not rounded, so we were able to safely do the test. 
But with the 2.6.32, the behaviour changed, so I preferred to report it 
in case that is something not expected.


--

[ message continues ]

12 happened to be rounded to 12 with previous kernels,
but with recent kernels we have another conversion table :

1 -> 3
2 -> 3
3 -> 3
4 -> 9
5 -> 9
6 -> 9
7 -> 9
8 -> 9
9 -> 9
10 -> 21
...
21 -> 21
22 -> 45
...
45 -> 45
46 -> 93
...
93 -> 93
94 -> 189
...
189 -> 189
190 -> 309
...
--

[ message continues ]

Thanks Eric, I will change the test program.

  -- Daniel
--

[ message continues ]

From: "Ilpo Järvinen" <ilpo.jarvinen@helsinki.fi>

There is no such requirement, we've been violating that premise since
day one for socket receive and send queue buffer limit socket options.

The kernel is always allowed to add fuzz or overhead adjustments to
whatever the user gives it.  If the user wants to know what the kernel
actually ended up using, it get getsockopt() to find out.
--

[ message continues ]