Hi,
I'm recently working on SYSCTLs features. While testing the sysctl
knob `net.inet.tcp.reass.maxsegments` I got something weird. When
I adjust `kern.ipc.nmbclusters`, the `net.inet.tcp.reass.maxsegments`
got 'auto tuned' regardless what I have set (the kernel env) in
`/boot/loader.conf`.
The relevant code:
```
static int tcp_reass_maxseg = 0;
SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
&tcp_reass_maxseg, 0,
"Global maximum number of TCP Segments in Reassembly Queue");
tcp_reass_global_init(void)
{
tcp_reass_maxseg = nmbclusters / 16;
TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
&tcp_reass_maxseg);
tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
/* Set the zone limit and read back the effective value. */
tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
tcp_reass_maxseg);
...
EVENTHANDLER_REGISTER(nmbclusters_change,
tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
}
/* Initialize TCP reassembly queue */
static void
tcp_reass_zone_change(void *tag)
{
/* Set the zone limit and read back the effective value. */
tcp_reass_maxseg = nmbclusters / 16;
tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
tcp_reass_maxseg);
}
```
The above code shows clearly the current behavior.
What I expected is, kernel env has priority to the 'auto tuning', unless
the admin requested.
So comes some questions.
1. Is the above behavior expected ?
2. Is it desirable to make `net.inet.tcp.reass.maxsegments` write tunable ?
3. Is it valid to set 0 to `net.inet.tcp.reass.maxsegments` ? IIUC that will
effectually disable TCP reassembly.
4. Then can we introduce a constant `-1` to explicitly enable the 'auto tuning'
feature ?
Best regards,
Zhenlei
