With CONFIG_CPUMASK_OFFSTACK=y, the 'bind_writers' buffer is allocated via
alloc_cpumask_var() in param_set_cpumask(). But it is not freed, when
setting the module parameter multiple times by sysfs interface or removing
module.
Below kmemleak trace is seen for this issue:
unreferenced object 0xffff888100aabff8 (size 8):
comm "bash", pid 323, jiffies
4295059233
hex dump (first 8 bytes):
07 00 00 00 00 00 00 00 ........
backtrace (crc
ac50919):
__kmalloc_node_noprof+0x2e5/0x420
alloc_cpumask_var_node+0x1f/0x30
param_set_cpumask+0x26/0xb0 [locktorture]
param_attr_store+0x93/0x100
module_attr_store+0x1b/0x30
kernfs_fop_write_iter+0x114/0x1b0
vfs_write+0x300/0x410
ksys_write+0x60/0xd0
do_syscall_64+0xa4/0x260
entry_SYSCALL_64_after_hwframe+0x77/0x7f
This issue can be reproduced by:
insmod locktorture.ko bind_writers=1
rmmod locktorture
or:
insmod locktorture.ko bind_writers=1
echo 2 > /sys/module/locktorture/parameters/bind_writers
Considering that setting the module parameter 'bind_writers' or
'bind_readers' by sysfs interface has no real effect, set the parameter
permissions to 0444. To fix the memory leak when removing module, free
'bind_writers' and 'bind_readers' memory in lock_torture_cleanup().
Fixes: 73e341242483 ("locktorture: Add readers_bind and writers_bind module parameters")
Suggested-by: Zhang Changzhong <zhangchangzhong@huawei.com>
Signed-off-by: Wang Liang <wangliang74@huawei.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
.get = param_get_cpumask,
};
-module_param_cb(bind_readers, <_bind_ops, &bind_readers, 0644);
-module_param_cb(bind_writers, <_bind_ops, &bind_writers, 0644);
+module_param_cb(bind_readers, <_bind_ops, &bind_readers, 0444);
+module_param_cb(bind_writers, <_bind_ops, &bind_writers, 0444);
long torture_sched_setaffinity(pid_t pid, const struct cpumask *in_mask, bool dowarn);
cxt.cur_ops->exit();
cxt.init_called = false;
}
+
+ free_cpumask_var(bind_readers);
+ free_cpumask_var(bind_writers);
+
torture_cleanup_end();
}