kernel/module.c: don't allow modprobe to hang forever on a module load
cover kernel/module.c: don't allow modprobe to hang forever on a module load
After commit 6e6de3dee51a ("kernel/module.c: Only return -EEXIST for modules that have finished loading")
the simple test leads to hanged modprobe process (INTERRUPTIBLE):
1. Make sure nft_ct module is not used by your firewall rules.
2. Run 3 copies of
# export i=0; while true; if [[ $(($i % 100)) -eq 0 ]] ; then \
echo "i=$i"; fi; do modprobe nft_ct; i=$(($i + 1)); done3. Run 2 copies of
# while true; do rmmod nft_ct; done
Hanged "modprobe" process will appear in ~ 10 seconds.
# cat /proc/30184/stack [<0>] load_module+0x53f/0x2060 [<0>] __do_sys_finit_module+0xd2/0x100 [<0>] do_syscall_64+0x5b/0x1c0 [<0>] entry_SYSCALL_64_after_hwframe+0x44/0xa9
Konstantin Khorenko (1):
kernel/module.c: wakeup processes in module_wq on module unload
kernel/module.c | 2 ++ 1 file changed, 2 insertions(+)
Signature · 1 line
2.15.1
PATCH 1/1 kernel/module.c: wakeup processes in module_wq on module unload
Fix the race between load and unload a kernel module.
sys_delete_module()
try_stop_module()
mod->state = _GOING
add_unformed_module()
old = find_module_all() (old->state == _GOING =>
wait_event_interruptible())
During pre-condition finished_loading() rets 0 schedule() (never gets waken up later)
free_module()
mod->state = _UNFORMED
list_del_rcu(&mod->list)
(dels mod from "modules" list)
return
The race above leads to modprobe hanging forever on loading a module.
Error paths on loading module call wake_up_all(&module_wq) after freeing module, so let's do the same on straight module unload.
for modules that have finished loading")
--- kernel/module.c | 2 ++ 1 file changed, 2 insertions(+)
kernel/module.c +2 -0
@@ -1033,6 +1033,8 @@ SYSCALL_DEFINE2(delete_module, const char __user *, name_user,strlcpy (last_unloaded_module, mod->name,sizeof (last_unloaded_module));free_module (mod); +/* someone could wait for the module in add_unformed_module() */ +wake_up_all (&module_wq);return 0 ; out:mutex_unlock (&module_mutex);
Signature · 1 line
2.15.1
Re: [PATCH 1/1] kernel/module.c: wakeup processes in module_wq on module unload
Quoted 40 lines · Konstantin Khorenko
Fix the race between load and unload a kernel module.
sys_delete_module()
try_stop_module()
mod->state=_GOING
add_unformed_module()
old =find_module_all()(old->state==_GOING=>
wait_event_interruptible())During pre-condition
finished_loading()rets 0schedule()(never gets waken up later)
free_module()
mod->state=_UNFORMED
list_del_rcu(&mod->list)
(dels mod from "modules" list)return
The race above leads to modprobe hanging forever on loading a module.
Error paths on loading module call
wake_up_all(&module_wq)after freeing module, so let's do the same on straight module unload.Fixes 6e6de3dee51a ("kernel/module.c: Only return -EEXISTfor modules that have finished loading")
Signed-off-by Konstantin Khorenko <[email protected]>--- kernel/module.c | 2 ++ 1 file changed, 2 insertions(+)kernel/module.c +0
-0
kernel/module.c SYSCALL_DEFINE2(delete_module, const char __user *, name_user, · line 1037
Quoted 5 lines
kernel/module.c +2
-0@@ -1033,6 +1033,8 @@ SYSCALL_DEFINE2(delete_module, const char __user *, name_user,strlcpy (last_unloaded_module, mod->name,sizeof (last_unloaded_module));free_module (mod); +/* someone could wait for the module in add_unformed_module() */ +wake_up_all (&module_wq);
free_module() acts on that *this* module so it makes sense to include the wake_up_all(&module_wq) call in delete_module. IIRC I did try a module load & unload stress test but it looks like my testing wasn't hard enough :(
In any case,
P.
Quoted 4 lines
kernel/module.c +0
-0return 0 ; out:mutex_unlock (&module_mutex);
Re: [PATCH 1/1] kernel/module.c: wakeup processes in module_wq on module unload
In reply to Konstantin Khorenko+++ Konstantin Khorenko [13/11/19 12:29 +0300]:
Quoted 24 lines
Fix the race between load and unload a kernel module.
sys_delete_module()try_stop_module()
mod->state=_GOING
add_unformed_module()
old =find_module_all()(old->state==_GOING=>
wait_event_interruptible())During pre-condition
finished_loading()rets 0schedule()(never gets waken up later)free_module()
mod->state=_UNFORMED
list_del_rcu(&mod->list)
(dels mod from "modules" list)return
The race above leads to modprobe hanging forever on loading a module.
Quoted 7 lines
Error paths on loading module call
wake_up_all(&module_wq)after freeing module, so let's do the same on straight module unload.Fixes 6e6de3dee51a ("kernel/module.c: Only return -EEXISTfor modules that have finished loading")
Signed-off-by Konstantin Khorenko <[email protected]>
Thanks Konstantin for catching this. I've applied this to modules-next.
Jessica