On Mon, Aug 29, 2022 at 9:21 AM Kuniyuki Iwashima [off-list ref] wrote:

The more sockets we have in the hash table, the longer we spend looking
up the socket.  While running a number of small workloads on the same
host, they penalise each other and cause performance degradation.>

...

+static int proc_tcp_child_ehash_entries(struct ctl_table *table, int write,
+                                       void *buffer, size_t *lenp, loff_t *ppos)
+{
+       unsigned int tcp_child_ehash_entries;
+       int ret;
+
+       ret = proc_douintvec_minmax(table, write, buffer, lenp, ppos);
+       if (!write || ret)
+               return ret;
+
+       tcp_child_ehash_entries = READ_ONCE(*(unsigned int *)table->data);
+       if (tcp_child_ehash_entries)
+               tcp_child_ehash_entries = roundup_pow_of_two(tcp_child_ehash_entries);

This is not thread safe.

You could simply perform the roundup_pow_of_two() elsewhere,
eg in tcp_set_hashinfo() (and leave the sysctl as set by the user)

+
+       WRITE_ONCE(*(unsigned int *)table->data, tcp_child_ehash_entries);
+
+       return 0;
+}
+