[PATCH] No need to protect current->group_info in sys_getgroups(), in_group_p() and in_egroup_p()

From: Eric Dumazet
Date: Wed Mar 15 2006 - 04:34:55 EST



While doing some benchmarks of an Apache/PHP SMP server, I noticed high oprofile numbers in in_group_p() and _atomic_dec_and_lock().

rank percent
1 4.8911 % __link_path_walk
2 4.8503 % __d_lookup
*3 4.2911 % _atomic_dec_and_lock
4 3.9307 % __copy_to_user_ll
5 4.9004 % sysenter_past_esp
*6 3.3248 % in_group_p

It appears that in_group_p() does an uncessary

get_group_info(current->group_info); /* atomic_inc() */
... /* access current->group_info */
put_group_info(current->group_info); /* _atomic_dec_and_lock */


It is not necessary to do this, because the current task holds a reference on its own group_info, and this reference cannot change during the lookup.

This patch deletes the get_group_info()/put_group_info() pair from sys_getgroups(), in_group_p() and in_egroup_p() functions.

Signed-off-by: Eric Dumazet <dada1@xxxxxxxxxxxxx>

--- a/kernel/sys.c 2006-03-15 10:14:37.000000000 +0100
+++ b/kernel/sys.c 2006-03-15 10:15:55.000000000 +0100
@@ -1433,7 +1433,6 @@
return -EINVAL;

/* no need to grab task_lock here; it cannot change */
- get_group_info(current->group_info);
i = current->group_info->ngroups;
if (gidsetsize) {
if (i > gidsetsize) {
@@ -1446,7 +1445,6 @@
}
}
out:
- put_group_info(current->group_info);
return i;
}

@@ -1487,9 +1485,7 @@
{
int retval = 1;
if (grp != current->fsgid) {
- get_group_info(current->group_info);
retval = groups_search(current->group_info, grp);
- put_group_info(current->group_info);
}
return retval;
}
@@ -1500,9 +1496,7 @@
{
int retval = 1;
if (grp != current->egid) {
- get_group_info(current->group_info);
retval = groups_search(current->group_info, grp);
- put_group_info(current->group_info);
}
return retval;
}